Linux Audio

Check our new training course

Yocto distribution development and maintenance

Need a Yocto distribution for your embedded project?
Loading...
Note: File does not exist in v3.5.6.
   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/fs_context.h>
  43#include <linux/namei.h>
  44#include <linux/pagemap.h>
  45#include <linux/proc_fs.h>
  46#include <linux/rcupdate.h>
  47#include <linux/sched.h>
  48#include <linux/sched/deadline.h>
  49#include <linux/sched/mm.h>
  50#include <linux/sched/task.h>
  51#include <linux/seq_file.h>
  52#include <linux/security.h>
  53#include <linux/slab.h>
  54#include <linux/spinlock.h>
  55#include <linux/stat.h>
  56#include <linux/string.h>
  57#include <linux/time.h>
  58#include <linux/time64.h>
  59#include <linux/backing-dev.h>
  60#include <linux/sort.h>
  61#include <linux/oom.h>
  62#include <linux/sched/isolation.h>
  63#include <linux/uaccess.h>
  64#include <linux/atomic.h>
  65#include <linux/mutex.h>
  66#include <linux/cgroup.h>
  67#include <linux/wait.h>
  68
  69DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
  70DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
  71
  72/* See "Frequency meter" comments, below. */
  73
  74struct fmeter {
  75	int cnt;		/* unprocessed events count */
  76	int val;		/* most recent output value */
  77	time64_t time;		/* clock (secs) when val computed */
  78	spinlock_t lock;	/* guards read or write of above */
  79};
  80
  81struct cpuset {
  82	struct cgroup_subsys_state css;
  83
  84	unsigned long flags;		/* "unsigned long" so bitops work */
  85
  86	/*
  87	 * On default hierarchy:
  88	 *
  89	 * The user-configured masks can only be changed by writing to
  90	 * cpuset.cpus and cpuset.mems, and won't be limited by the
  91	 * parent masks.
  92	 *
  93	 * The effective masks is the real masks that apply to the tasks
  94	 * in the cpuset. They may be changed if the configured masks are
  95	 * changed or hotplug happens.
  96	 *
  97	 * effective_mask == configured_mask & parent's effective_mask,
  98	 * and if it ends up empty, it will inherit the parent's mask.
  99	 *
 100	 *
 101	 * On legacy hierarchy:
 102	 *
 103	 * The user-configured masks are always the same with effective masks.
 104	 */
 105
 106	/* user-configured CPUs and Memory Nodes allow to tasks */
 107	cpumask_var_t cpus_allowed;
 108	nodemask_t mems_allowed;
 109
 110	/* effective CPUs and Memory Nodes allow to tasks */
 111	cpumask_var_t effective_cpus;
 112	nodemask_t effective_mems;
 113
 114	/*
 115	 * CPUs allocated to child sub-partitions (default hierarchy only)
 116	 * - CPUs granted by the parent = effective_cpus U subparts_cpus
 117	 * - effective_cpus and subparts_cpus are mutually exclusive.
 118	 *
 119	 * effective_cpus contains only onlined CPUs, but subparts_cpus
 120	 * may have offlined ones.
 121	 */
 122	cpumask_var_t subparts_cpus;
 123
 124	/*
 125	 * This is old Memory Nodes tasks took on.
 126	 *
 127	 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
 128	 * - A new cpuset's old_mems_allowed is initialized when some
 129	 *   task is moved into it.
 130	 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
 131	 *   cpuset.mems_allowed and have tasks' nodemask updated, and
 132	 *   then old_mems_allowed is updated to mems_allowed.
 133	 */
 134	nodemask_t old_mems_allowed;
 135
 136	struct fmeter fmeter;		/* memory_pressure filter */
 137
 138	/*
 139	 * Tasks are being attached to this cpuset.  Used to prevent
 140	 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
 141	 */
 142	int attach_in_progress;
 143
 144	/* partition number for rebuild_sched_domains() */
 145	int pn;
 146
 147	/* for custom sched domain */
 148	int relax_domain_level;
 149
 150	/* number of CPUs in subparts_cpus */
 151	int nr_subparts_cpus;
 152
 153	/* partition root state */
 154	int partition_root_state;
 155
 156	/*
 157	 * Default hierarchy only:
 158	 * use_parent_ecpus - set if using parent's effective_cpus
 159	 * child_ecpus_count - # of children with use_parent_ecpus set
 160	 */
 161	int use_parent_ecpus;
 162	int child_ecpus_count;
 163};
 164
 165/*
 166 * Partition root states:
 167 *
 168 *   0 - not a partition root
 169 *
 170 *   1 - partition root
 171 *
 172 *  -1 - invalid partition root
 173 *       None of the cpus in cpus_allowed can be put into the parent's
 174 *       subparts_cpus. In this case, the cpuset is not a real partition
 175 *       root anymore.  However, the CPU_EXCLUSIVE bit will still be set
 176 *       and the cpuset can be restored back to a partition root if the
 177 *       parent cpuset can give more CPUs back to this child cpuset.
 178 */
 179#define PRS_DISABLED		0
 180#define PRS_ENABLED		1
 181#define PRS_ERROR		-1
 182
 183/*
 184 * Temporary cpumasks for working with partitions that are passed among
 185 * functions to avoid memory allocation in inner functions.
 186 */
 187struct tmpmasks {
 188	cpumask_var_t addmask, delmask;	/* For partition root */
 189	cpumask_var_t new_cpus;		/* For update_cpumasks_hier() */
 190};
 191
 192static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
 193{
 194	return css ? container_of(css, struct cpuset, css) : NULL;
 195}
 196
 197/* Retrieve the cpuset for a task */
 198static inline struct cpuset *task_cs(struct task_struct *task)
 199{
 200	return css_cs(task_css(task, cpuset_cgrp_id));
 201}
 202
 203static inline struct cpuset *parent_cs(struct cpuset *cs)
 204{
 205	return css_cs(cs->css.parent);
 206}
 207
 208/* bits in struct cpuset flags field */
 209typedef enum {
 210	CS_ONLINE,
 211	CS_CPU_EXCLUSIVE,
 212	CS_MEM_EXCLUSIVE,
 213	CS_MEM_HARDWALL,
 214	CS_MEMORY_MIGRATE,
 215	CS_SCHED_LOAD_BALANCE,
 216	CS_SPREAD_PAGE,
 217	CS_SPREAD_SLAB,
 218} cpuset_flagbits_t;
 219
 220/* convenient tests for these bits */
 221static inline bool is_cpuset_online(struct cpuset *cs)
 222{
 223	return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
 224}
 225
 226static inline int is_cpu_exclusive(const struct cpuset *cs)
 227{
 228	return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
 229}
 230
 231static inline int is_mem_exclusive(const struct cpuset *cs)
 232{
 233	return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
 234}
 235
 236static inline int is_mem_hardwall(const struct cpuset *cs)
 237{
 238	return test_bit(CS_MEM_HARDWALL, &cs->flags);
 239}
 240
 241static inline int is_sched_load_balance(const struct cpuset *cs)
 242{
 243	return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
 244}
 245
 246static inline int is_memory_migrate(const struct cpuset *cs)
 247{
 248	return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
 249}
 250
 251static inline int is_spread_page(const struct cpuset *cs)
 252{
 253	return test_bit(CS_SPREAD_PAGE, &cs->flags);
 254}
 255
 256static inline int is_spread_slab(const struct cpuset *cs)
 257{
 258	return test_bit(CS_SPREAD_SLAB, &cs->flags);
 259}
 260
 261static inline int is_partition_root(const struct cpuset *cs)
 262{
 263	return cs->partition_root_state > 0;
 264}
 265
 266static struct cpuset top_cpuset = {
 267	.flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
 268		  (1 << CS_MEM_EXCLUSIVE)),
 269	.partition_root_state = PRS_ENABLED,
 270};
 271
 272/**
 273 * cpuset_for_each_child - traverse online children of a cpuset
 274 * @child_cs: loop cursor pointing to the current child
 275 * @pos_css: used for iteration
 276 * @parent_cs: target cpuset to walk children of
 277 *
 278 * Walk @child_cs through the online children of @parent_cs.  Must be used
 279 * with RCU read locked.
 280 */
 281#define cpuset_for_each_child(child_cs, pos_css, parent_cs)		\
 282	css_for_each_child((pos_css), &(parent_cs)->css)		\
 283		if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
 284
 285/**
 286 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
 287 * @des_cs: loop cursor pointing to the current descendant
 288 * @pos_css: used for iteration
 289 * @root_cs: target cpuset to walk ancestor of
 290 *
 291 * Walk @des_cs through the online descendants of @root_cs.  Must be used
 292 * with RCU read locked.  The caller may modify @pos_css by calling
 293 * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
 294 * iteration and the first node to be visited.
 295 */
 296#define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)	\
 297	css_for_each_descendant_pre((pos_css), &(root_cs)->css)		\
 298		if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
 299
 300/*
 301 * There are two global locks guarding cpuset structures - cpuset_mutex and
 302 * callback_lock. We also require taking task_lock() when dereferencing a
 303 * task's cpuset pointer. See "The task_lock() exception", at the end of this
 304 * comment.
 305 *
 306 * A task must hold both locks to modify cpusets.  If a task holds
 307 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
 308 * is the only task able to also acquire callback_lock and be able to
 309 * modify cpusets.  It can perform various checks on the cpuset structure
 310 * first, knowing nothing will change.  It can also allocate memory while
 311 * just holding cpuset_mutex.  While it is performing these checks, various
 312 * callback routines can briefly acquire callback_lock to query cpusets.
 313 * Once it is ready to make the changes, it takes callback_lock, blocking
 314 * everyone else.
 315 *
 316 * Calls to the kernel memory allocator can not be made while holding
 317 * callback_lock, as that would risk double tripping on callback_lock
 318 * from one of the callbacks into the cpuset code from within
 319 * __alloc_pages().
 320 *
 321 * If a task is only holding callback_lock, then it has read-only
 322 * access to cpusets.
 323 *
 324 * Now, the task_struct fields mems_allowed and mempolicy may be changed
 325 * by other task, we use alloc_lock in the task_struct fields to protect
 326 * them.
 327 *
 328 * The cpuset_common_file_read() handlers only hold callback_lock across
 329 * small pieces of code, such as when reading out possibly multi-word
 330 * cpumasks and nodemasks.
 331 *
 332 * Accessing a task's cpuset should be done in accordance with the
 333 * guidelines for accessing subsystem state in kernel/cgroup.c
 334 */
 335
 336DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
 337
 338void cpuset_read_lock(void)
 339{
 340	percpu_down_read(&cpuset_rwsem);
 341}
 342
 343void cpuset_read_unlock(void)
 344{
 345	percpu_up_read(&cpuset_rwsem);
 346}
 347
 348static DEFINE_SPINLOCK(callback_lock);
 349
 350static struct workqueue_struct *cpuset_migrate_mm_wq;
 351
 352/*
 353 * CPU / memory hotplug is handled asynchronously.
 354 */
 355static void cpuset_hotplug_workfn(struct work_struct *work);
 356static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
 357
 358static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
 359
 360/*
 361 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
 362 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
 363 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
 364 * With v2 behavior, "cpus" and "mems" are always what the users have
 365 * requested and won't be changed by hotplug events. Only the effective
 366 * cpus or mems will be affected.
 367 */
 368static inline bool is_in_v2_mode(void)
 369{
 370	return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
 371	      (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
 372}
 373
 374/*
 375 * Return in pmask the portion of a cpusets's cpus_allowed that
 376 * are online.  If none are online, walk up the cpuset hierarchy
 377 * until we find one that does have some online cpus.
 378 *
 379 * One way or another, we guarantee to return some non-empty subset
 380 * of cpu_online_mask.
 381 *
 382 * Call with callback_lock or cpuset_mutex held.
 383 */
 384static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
 385{
 386	while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
 387		cs = parent_cs(cs);
 388		if (unlikely(!cs)) {
 389			/*
 390			 * The top cpuset doesn't have any online cpu as a
 391			 * consequence of a race between cpuset_hotplug_work
 392			 * and cpu hotplug notifier.  But we know the top
 393			 * cpuset's effective_cpus is on its way to be
 394			 * identical to cpu_online_mask.
 395			 */
 396			cpumask_copy(pmask, cpu_online_mask);
 397			return;
 398		}
 399	}
 400	cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
 401}
 402
 403/*
 404 * Return in *pmask the portion of a cpusets's mems_allowed that
 405 * are online, with memory.  If none are online with memory, walk
 406 * up the cpuset hierarchy until we find one that does have some
 407 * online mems.  The top cpuset always has some mems online.
 408 *
 409 * One way or another, we guarantee to return some non-empty subset
 410 * of node_states[N_MEMORY].
 411 *
 412 * Call with callback_lock or cpuset_mutex held.
 413 */
 414static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
 415{
 416	while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
 417		cs = parent_cs(cs);
 418	nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
 419}
 420
 421/*
 422 * update task's spread flag if cpuset's page/slab spread flag is set
 423 *
 424 * Call with callback_lock or cpuset_mutex held.
 425 */
 426static void cpuset_update_task_spread_flag(struct cpuset *cs,
 427					struct task_struct *tsk)
 428{
 429	if (is_spread_page(cs))
 430		task_set_spread_page(tsk);
 431	else
 432		task_clear_spread_page(tsk);
 433
 434	if (is_spread_slab(cs))
 435		task_set_spread_slab(tsk);
 436	else
 437		task_clear_spread_slab(tsk);
 438}
 439
 440/*
 441 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
 442 *
 443 * One cpuset is a subset of another if all its allowed CPUs and
 444 * Memory Nodes are a subset of the other, and its exclusive flags
 445 * are only set if the other's are set.  Call holding cpuset_mutex.
 446 */
 447
 448static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
 449{
 450	return	cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
 451		nodes_subset(p->mems_allowed, q->mems_allowed) &&
 452		is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
 453		is_mem_exclusive(p) <= is_mem_exclusive(q);
 454}
 455
 456/**
 457 * alloc_cpumasks - allocate three cpumasks for cpuset
 458 * @cs:  the cpuset that have cpumasks to be allocated.
 459 * @tmp: the tmpmasks structure pointer
 460 * Return: 0 if successful, -ENOMEM otherwise.
 461 *
 462 * Only one of the two input arguments should be non-NULL.
 463 */
 464static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
 465{
 466	cpumask_var_t *pmask1, *pmask2, *pmask3;
 467
 468	if (cs) {
 469		pmask1 = &cs->cpus_allowed;
 470		pmask2 = &cs->effective_cpus;
 471		pmask3 = &cs->subparts_cpus;
 472	} else {
 473		pmask1 = &tmp->new_cpus;
 474		pmask2 = &tmp->addmask;
 475		pmask3 = &tmp->delmask;
 476	}
 477
 478	if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
 479		return -ENOMEM;
 480
 481	if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
 482		goto free_one;
 483
 484	if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
 485		goto free_two;
 486
 487	return 0;
 488
 489free_two:
 490	free_cpumask_var(*pmask2);
 491free_one:
 492	free_cpumask_var(*pmask1);
 493	return -ENOMEM;
 494}
 495
 496/**
 497 * free_cpumasks - free cpumasks in a tmpmasks structure
 498 * @cs:  the cpuset that have cpumasks to be free.
 499 * @tmp: the tmpmasks structure pointer
 500 */
 501static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
 502{
 503	if (cs) {
 504		free_cpumask_var(cs->cpus_allowed);
 505		free_cpumask_var(cs->effective_cpus);
 506		free_cpumask_var(cs->subparts_cpus);
 507	}
 508	if (tmp) {
 509		free_cpumask_var(tmp->new_cpus);
 510		free_cpumask_var(tmp->addmask);
 511		free_cpumask_var(tmp->delmask);
 512	}
 513}
 514
 515/**
 516 * alloc_trial_cpuset - allocate a trial cpuset
 517 * @cs: the cpuset that the trial cpuset duplicates
 518 */
 519static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
 520{
 521	struct cpuset *trial;
 522
 523	trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
 524	if (!trial)
 525		return NULL;
 526
 527	if (alloc_cpumasks(trial, NULL)) {
 528		kfree(trial);
 529		return NULL;
 530	}
 531
 532	cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
 533	cpumask_copy(trial->effective_cpus, cs->effective_cpus);
 534	return trial;
 535}
 536
 537/**
 538 * free_cpuset - free the cpuset
 539 * @cs: the cpuset to be freed
 540 */
 541static inline void free_cpuset(struct cpuset *cs)
 542{
 543	free_cpumasks(cs, NULL);
 544	kfree(cs);
 545}
 546
 547/*
 548 * validate_change() - Used to validate that any proposed cpuset change
 549 *		       follows the structural rules for cpusets.
 550 *
 551 * If we replaced the flag and mask values of the current cpuset
 552 * (cur) with those values in the trial cpuset (trial), would
 553 * our various subset and exclusive rules still be valid?  Presumes
 554 * cpuset_mutex held.
 555 *
 556 * 'cur' is the address of an actual, in-use cpuset.  Operations
 557 * such as list traversal that depend on the actual address of the
 558 * cpuset in the list must use cur below, not trial.
 559 *
 560 * 'trial' is the address of bulk structure copy of cur, with
 561 * perhaps one or more of the fields cpus_allowed, mems_allowed,
 562 * or flags changed to new, trial values.
 563 *
 564 * Return 0 if valid, -errno if not.
 565 */
 566
 567static int validate_change(struct cpuset *cur, struct cpuset *trial)
 568{
 569	struct cgroup_subsys_state *css;
 570	struct cpuset *c, *par;
 571	int ret;
 572
 573	rcu_read_lock();
 574
 575	/* Each of our child cpusets must be a subset of us */
 576	ret = -EBUSY;
 577	cpuset_for_each_child(c, css, cur)
 578		if (!is_cpuset_subset(c, trial))
 579			goto out;
 580
 581	/* Remaining checks don't apply to root cpuset */
 582	ret = 0;
 583	if (cur == &top_cpuset)
 584		goto out;
 585
 586	par = parent_cs(cur);
 587
 588	/* On legacy hierarchy, we must be a subset of our parent cpuset. */
 589	ret = -EACCES;
 590	if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
 591		goto out;
 592
 593	/*
 594	 * If either I or some sibling (!= me) is exclusive, we can't
 595	 * overlap
 596	 */
 597	ret = -EINVAL;
 598	cpuset_for_each_child(c, css, par) {
 599		if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
 600		    c != cur &&
 601		    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
 602			goto out;
 603		if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
 604		    c != cur &&
 605		    nodes_intersects(trial->mems_allowed, c->mems_allowed))
 606			goto out;
 607	}
 608
 609	/*
 610	 * Cpusets with tasks - existing or newly being attached - can't
 611	 * be changed to have empty cpus_allowed or mems_allowed.
 612	 */
 613	ret = -ENOSPC;
 614	if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
 615		if (!cpumask_empty(cur->cpus_allowed) &&
 616		    cpumask_empty(trial->cpus_allowed))
 617			goto out;
 618		if (!nodes_empty(cur->mems_allowed) &&
 619		    nodes_empty(trial->mems_allowed))
 620			goto out;
 621	}
 622
 623	/*
 624	 * We can't shrink if we won't have enough room for SCHED_DEADLINE
 625	 * tasks.
 626	 */
 627	ret = -EBUSY;
 628	if (is_cpu_exclusive(cur) &&
 629	    !cpuset_cpumask_can_shrink(cur->cpus_allowed,
 630				       trial->cpus_allowed))
 631		goto out;
 632
 633	ret = 0;
 634out:
 635	rcu_read_unlock();
 636	return ret;
 637}
 638
 639#ifdef CONFIG_SMP
 640/*
 641 * Helper routine for generate_sched_domains().
 642 * Do cpusets a, b have overlapping effective cpus_allowed masks?
 643 */
 644static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
 645{
 646	return cpumask_intersects(a->effective_cpus, b->effective_cpus);
 647}
 648
 649static void
 650update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
 651{
 652	if (dattr->relax_domain_level < c->relax_domain_level)
 653		dattr->relax_domain_level = c->relax_domain_level;
 654	return;
 655}
 656
 657static void update_domain_attr_tree(struct sched_domain_attr *dattr,
 658				    struct cpuset *root_cs)
 659{
 660	struct cpuset *cp;
 661	struct cgroup_subsys_state *pos_css;
 662
 663	rcu_read_lock();
 664	cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
 665		/* skip the whole subtree if @cp doesn't have any CPU */
 666		if (cpumask_empty(cp->cpus_allowed)) {
 667			pos_css = css_rightmost_descendant(pos_css);
 668			continue;
 669		}
 670
 671		if (is_sched_load_balance(cp))
 672			update_domain_attr(dattr, cp);
 673	}
 674	rcu_read_unlock();
 675}
 676
 677/* Must be called with cpuset_mutex held.  */
 678static inline int nr_cpusets(void)
 679{
 680	/* jump label reference count + the top-level cpuset */
 681	return static_key_count(&cpusets_enabled_key.key) + 1;
 682}
 683
 684/*
 685 * generate_sched_domains()
 686 *
 687 * This function builds a partial partition of the systems CPUs
 688 * A 'partial partition' is a set of non-overlapping subsets whose
 689 * union is a subset of that set.
 690 * The output of this function needs to be passed to kernel/sched/core.c
 691 * partition_sched_domains() routine, which will rebuild the scheduler's
 692 * load balancing domains (sched domains) as specified by that partial
 693 * partition.
 694 *
 695 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
 696 * for a background explanation of this.
 697 *
 698 * Does not return errors, on the theory that the callers of this
 699 * routine would rather not worry about failures to rebuild sched
 700 * domains when operating in the severe memory shortage situations
 701 * that could cause allocation failures below.
 702 *
 703 * Must be called with cpuset_mutex held.
 704 *
 705 * The three key local variables below are:
 706 *    cp - cpuset pointer, used (together with pos_css) to perform a
 707 *	   top-down scan of all cpusets. For our purposes, rebuilding
 708 *	   the schedulers sched domains, we can ignore !is_sched_load_
 709 *	   balance cpusets.
 710 *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
 711 *	   that need to be load balanced, for convenient iterative
 712 *	   access by the subsequent code that finds the best partition,
 713 *	   i.e the set of domains (subsets) of CPUs such that the
 714 *	   cpus_allowed of every cpuset marked is_sched_load_balance
 715 *	   is a subset of one of these domains, while there are as
 716 *	   many such domains as possible, each as small as possible.
 717 * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
 718 *	   the kernel/sched/core.c routine partition_sched_domains() in a
 719 *	   convenient format, that can be easily compared to the prior
 720 *	   value to determine what partition elements (sched domains)
 721 *	   were changed (added or removed.)
 722 *
 723 * Finding the best partition (set of domains):
 724 *	The triple nested loops below over i, j, k scan over the
 725 *	load balanced cpusets (using the array of cpuset pointers in
 726 *	csa[]) looking for pairs of cpusets that have overlapping
 727 *	cpus_allowed, but which don't have the same 'pn' partition
 728 *	number and gives them in the same partition number.  It keeps
 729 *	looping on the 'restart' label until it can no longer find
 730 *	any such pairs.
 731 *
 732 *	The union of the cpus_allowed masks from the set of
 733 *	all cpusets having the same 'pn' value then form the one
 734 *	element of the partition (one sched domain) to be passed to
 735 *	partition_sched_domains().
 736 */
 737static int generate_sched_domains(cpumask_var_t **domains,
 738			struct sched_domain_attr **attributes)
 739{
 740	struct cpuset *cp;	/* top-down scan of cpusets */
 741	struct cpuset **csa;	/* array of all cpuset ptrs */
 742	int csn;		/* how many cpuset ptrs in csa so far */
 743	int i, j, k;		/* indices for partition finding loops */
 744	cpumask_var_t *doms;	/* resulting partition; i.e. sched domains */
 745	struct sched_domain_attr *dattr;  /* attributes for custom domains */
 746	int ndoms = 0;		/* number of sched domains in result */
 747	int nslot;		/* next empty doms[] struct cpumask slot */
 748	struct cgroup_subsys_state *pos_css;
 749	bool root_load_balance = is_sched_load_balance(&top_cpuset);
 750
 751	doms = NULL;
 752	dattr = NULL;
 753	csa = NULL;
 754
 755	/* Special case for the 99% of systems with one, full, sched domain */
 756	if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
 757		ndoms = 1;
 758		doms = alloc_sched_domains(ndoms);
 759		if (!doms)
 760			goto done;
 761
 762		dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
 763		if (dattr) {
 764			*dattr = SD_ATTR_INIT;
 765			update_domain_attr_tree(dattr, &top_cpuset);
 766		}
 767		cpumask_and(doms[0], top_cpuset.effective_cpus,
 768			    housekeeping_cpumask(HK_FLAG_DOMAIN));
 769
 770		goto done;
 771	}
 772
 773	csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
 774	if (!csa)
 775		goto done;
 776	csn = 0;
 777
 778	rcu_read_lock();
 779	if (root_load_balance)
 780		csa[csn++] = &top_cpuset;
 781	cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
 782		if (cp == &top_cpuset)
 783			continue;
 784		/*
 785		 * Continue traversing beyond @cp iff @cp has some CPUs and
 786		 * isn't load balancing.  The former is obvious.  The
 787		 * latter: All child cpusets contain a subset of the
 788		 * parent's cpus, so just skip them, and then we call
 789		 * update_domain_attr_tree() to calc relax_domain_level of
 790		 * the corresponding sched domain.
 791		 *
 792		 * If root is load-balancing, we can skip @cp if it
 793		 * is a subset of the root's effective_cpus.
 794		 */
 795		if (!cpumask_empty(cp->cpus_allowed) &&
 796		    !(is_sched_load_balance(cp) &&
 797		      cpumask_intersects(cp->cpus_allowed,
 798					 housekeeping_cpumask(HK_FLAG_DOMAIN))))
 799			continue;
 800
 801		if (root_load_balance &&
 802		    cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
 803			continue;
 804
 805		if (is_sched_load_balance(cp) &&
 806		    !cpumask_empty(cp->effective_cpus))
 807			csa[csn++] = cp;
 808
 809		/* skip @cp's subtree if not a partition root */
 810		if (!is_partition_root(cp))
 811			pos_css = css_rightmost_descendant(pos_css);
 812	}
 813	rcu_read_unlock();
 814
 815	for (i = 0; i < csn; i++)
 816		csa[i]->pn = i;
 817	ndoms = csn;
 818
 819restart:
 820	/* Find the best partition (set of sched domains) */
 821	for (i = 0; i < csn; i++) {
 822		struct cpuset *a = csa[i];
 823		int apn = a->pn;
 824
 825		for (j = 0; j < csn; j++) {
 826			struct cpuset *b = csa[j];
 827			int bpn = b->pn;
 828
 829			if (apn != bpn && cpusets_overlap(a, b)) {
 830				for (k = 0; k < csn; k++) {
 831					struct cpuset *c = csa[k];
 832
 833					if (c->pn == bpn)
 834						c->pn = apn;
 835				}
 836				ndoms--;	/* one less element */
 837				goto restart;
 838			}
 839		}
 840	}
 841
 842	/*
 843	 * Now we know how many domains to create.
 844	 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
 845	 */
 846	doms = alloc_sched_domains(ndoms);
 847	if (!doms)
 848		goto done;
 849
 850	/*
 851	 * The rest of the code, including the scheduler, can deal with
 852	 * dattr==NULL case. No need to abort if alloc fails.
 853	 */
 854	dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
 855			      GFP_KERNEL);
 856
 857	for (nslot = 0, i = 0; i < csn; i++) {
 858		struct cpuset *a = csa[i];
 859		struct cpumask *dp;
 860		int apn = a->pn;
 861
 862		if (apn < 0) {
 863			/* Skip completed partitions */
 864			continue;
 865		}
 866
 867		dp = doms[nslot];
 868
 869		if (nslot == ndoms) {
 870			static int warnings = 10;
 871			if (warnings) {
 872				pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
 873					nslot, ndoms, csn, i, apn);
 874				warnings--;
 875			}
 876			continue;
 877		}
 878
 879		cpumask_clear(dp);
 880		if (dattr)
 881			*(dattr + nslot) = SD_ATTR_INIT;
 882		for (j = i; j < csn; j++) {
 883			struct cpuset *b = csa[j];
 884
 885			if (apn == b->pn) {
 886				cpumask_or(dp, dp, b->effective_cpus);
 887				cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
 888				if (dattr)
 889					update_domain_attr_tree(dattr + nslot, b);
 890
 891				/* Done with this partition */
 892				b->pn = -1;
 893			}
 894		}
 895		nslot++;
 896	}
 897	BUG_ON(nslot != ndoms);
 898
 899done:
 900	kfree(csa);
 901
 902	/*
 903	 * Fallback to the default domain if kmalloc() failed.
 904	 * See comments in partition_sched_domains().
 905	 */
 906	if (doms == NULL)
 907		ndoms = 1;
 908
 909	*domains    = doms;
 910	*attributes = dattr;
 911	return ndoms;
 912}
 913
 914static void update_tasks_root_domain(struct cpuset *cs)
 915{
 916	struct css_task_iter it;
 917	struct task_struct *task;
 918
 919	css_task_iter_start(&cs->css, 0, &it);
 920
 921	while ((task = css_task_iter_next(&it)))
 922		dl_add_task_root_domain(task);
 923
 924	css_task_iter_end(&it);
 925}
 926
 927static void rebuild_root_domains(void)
 928{
 929	struct cpuset *cs = NULL;
 930	struct cgroup_subsys_state *pos_css;
 931
 932	percpu_rwsem_assert_held(&cpuset_rwsem);
 933	lockdep_assert_cpus_held();
 934	lockdep_assert_held(&sched_domains_mutex);
 935
 936	rcu_read_lock();
 937
 938	/*
 939	 * Clear default root domain DL accounting, it will be computed again
 940	 * if a task belongs to it.
 941	 */
 942	dl_clear_root_domain(&def_root_domain);
 943
 944	cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
 945
 946		if (cpumask_empty(cs->effective_cpus)) {
 947			pos_css = css_rightmost_descendant(pos_css);
 948			continue;
 949		}
 950
 951		css_get(&cs->css);
 952
 953		rcu_read_unlock();
 954
 955		update_tasks_root_domain(cs);
 956
 957		rcu_read_lock();
 958		css_put(&cs->css);
 959	}
 960	rcu_read_unlock();
 961}
 962
 963static void
 964partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
 965				    struct sched_domain_attr *dattr_new)
 966{
 967	mutex_lock(&sched_domains_mutex);
 968	partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
 969	rebuild_root_domains();
 970	mutex_unlock(&sched_domains_mutex);
 971}
 972
 973/*
 974 * Rebuild scheduler domains.
 975 *
 976 * If the flag 'sched_load_balance' of any cpuset with non-empty
 977 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
 978 * which has that flag enabled, or if any cpuset with a non-empty
 979 * 'cpus' is removed, then call this routine to rebuild the
 980 * scheduler's dynamic sched domains.
 981 *
 982 * Call with cpuset_mutex held.  Takes get_online_cpus().
 983 */
 984static void rebuild_sched_domains_locked(void)
 985{
 986	struct cgroup_subsys_state *pos_css;
 987	struct sched_domain_attr *attr;
 988	cpumask_var_t *doms;
 989	struct cpuset *cs;
 990	int ndoms;
 991
 992	lockdep_assert_cpus_held();
 993	percpu_rwsem_assert_held(&cpuset_rwsem);
 994
 995	/*
 996	 * If we have raced with CPU hotplug, return early to avoid
 997	 * passing doms with offlined cpu to partition_sched_domains().
 998	 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
 999	 *
1000	 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1001	 * should be the same as the active CPUs, so checking only top_cpuset
1002	 * is enough to detect racing CPU offlines.
1003	 */
1004	if (!top_cpuset.nr_subparts_cpus &&
1005	    !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1006		return;
1007
1008	/*
1009	 * With subpartition CPUs, however, the effective CPUs of a partition
1010	 * root should be only a subset of the active CPUs.  Since a CPU in any
1011	 * partition root could be offlined, all must be checked.
1012	 */
1013	if (top_cpuset.nr_subparts_cpus) {
1014		rcu_read_lock();
1015		cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1016			if (!is_partition_root(cs)) {
1017				pos_css = css_rightmost_descendant(pos_css);
1018				continue;
1019			}
1020			if (!cpumask_subset(cs->effective_cpus,
1021					    cpu_active_mask)) {
1022				rcu_read_unlock();
1023				return;
1024			}
1025		}
1026		rcu_read_unlock();
1027	}
1028
1029	/* Generate domain masks and attrs */
1030	ndoms = generate_sched_domains(&doms, &attr);
1031
1032	/* Have scheduler rebuild the domains */
1033	partition_and_rebuild_sched_domains(ndoms, doms, attr);
1034}
1035#else /* !CONFIG_SMP */
1036static void rebuild_sched_domains_locked(void)
1037{
1038}
1039#endif /* CONFIG_SMP */
1040
1041void rebuild_sched_domains(void)
1042{
1043	get_online_cpus();
1044	percpu_down_write(&cpuset_rwsem);
1045	rebuild_sched_domains_locked();
1046	percpu_up_write(&cpuset_rwsem);
1047	put_online_cpus();
1048}
1049
1050/**
1051 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1052 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1053 *
1054 * Iterate through each task of @cs updating its cpus_allowed to the
1055 * effective cpuset's.  As this function is called with cpuset_mutex held,
1056 * cpuset membership stays stable.
1057 */
1058static void update_tasks_cpumask(struct cpuset *cs)
1059{
1060	struct css_task_iter it;
1061	struct task_struct *task;
1062
1063	css_task_iter_start(&cs->css, 0, &it);
1064	while ((task = css_task_iter_next(&it)))
1065		set_cpus_allowed_ptr(task, cs->effective_cpus);
1066	css_task_iter_end(&it);
1067}
1068
1069/**
1070 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1071 * @new_cpus: the temp variable for the new effective_cpus mask
1072 * @cs: the cpuset the need to recompute the new effective_cpus mask
1073 * @parent: the parent cpuset
1074 *
1075 * If the parent has subpartition CPUs, include them in the list of
1076 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1077 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1078 * to mask those out.
1079 */
1080static void compute_effective_cpumask(struct cpumask *new_cpus,
1081				      struct cpuset *cs, struct cpuset *parent)
1082{
1083	if (parent->nr_subparts_cpus) {
1084		cpumask_or(new_cpus, parent->effective_cpus,
1085			   parent->subparts_cpus);
1086		cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1087		cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1088	} else {
1089		cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1090	}
1091}
1092
1093/*
1094 * Commands for update_parent_subparts_cpumask
1095 */
1096enum subparts_cmd {
1097	partcmd_enable,		/* Enable partition root	 */
1098	partcmd_disable,	/* Disable partition root	 */
1099	partcmd_update,		/* Update parent's subparts_cpus */
1100};
1101
1102/**
1103 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1104 * @cpuset:  The cpuset that requests change in partition root state
1105 * @cmd:     Partition root state change command
1106 * @newmask: Optional new cpumask for partcmd_update
1107 * @tmp:     Temporary addmask and delmask
1108 * Return:   0, 1 or an error code
1109 *
1110 * For partcmd_enable, the cpuset is being transformed from a non-partition
1111 * root to a partition root. The cpus_allowed mask of the given cpuset will
1112 * be put into parent's subparts_cpus and taken away from parent's
1113 * effective_cpus. The function will return 0 if all the CPUs listed in
1114 * cpus_allowed can be granted or an error code will be returned.
1115 *
1116 * For partcmd_disable, the cpuset is being transofrmed from a partition
1117 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1118 * parent's subparts_cpus will be taken away from that cpumask and put back
1119 * into parent's effective_cpus. 0 should always be returned.
1120 *
1121 * For partcmd_update, if the optional newmask is specified, the cpu
1122 * list is to be changed from cpus_allowed to newmask. Otherwise,
1123 * cpus_allowed is assumed to remain the same. The cpuset should either
1124 * be a partition root or an invalid partition root. The partition root
1125 * state may change if newmask is NULL and none of the requested CPUs can
1126 * be granted by the parent. The function will return 1 if changes to
1127 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1128 * Error code should only be returned when newmask is non-NULL.
1129 *
1130 * The partcmd_enable and partcmd_disable commands are used by
1131 * update_prstate(). The partcmd_update command is used by
1132 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1133 * newmask set.
1134 *
1135 * The checking is more strict when enabling partition root than the
1136 * other two commands.
1137 *
1138 * Because of the implicit cpu exclusive nature of a partition root,
1139 * cpumask changes that violates the cpu exclusivity rule will not be
1140 * permitted when checked by validate_change(). The validate_change()
1141 * function will also prevent any changes to the cpu list if it is not
1142 * a superset of children's cpu lists.
1143 */
1144static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1145					  struct cpumask *newmask,
1146					  struct tmpmasks *tmp)
1147{
1148	struct cpuset *parent = parent_cs(cpuset);
1149	int adding;	/* Moving cpus from effective_cpus to subparts_cpus */
1150	int deleting;	/* Moving cpus from subparts_cpus to effective_cpus */
1151	int new_prs;
1152	bool part_error = false;	/* Partition error? */
1153
1154	percpu_rwsem_assert_held(&cpuset_rwsem);
1155
1156	/*
1157	 * The parent must be a partition root.
1158	 * The new cpumask, if present, or the current cpus_allowed must
1159	 * not be empty.
1160	 */
1161	if (!is_partition_root(parent) ||
1162	   (newmask && cpumask_empty(newmask)) ||
1163	   (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1164		return -EINVAL;
1165
1166	/*
1167	 * Enabling/disabling partition root is not allowed if there are
1168	 * online children.
1169	 */
1170	if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1171		return -EBUSY;
1172
1173	/*
1174	 * Enabling partition root is not allowed if not all the CPUs
1175	 * can be granted from parent's effective_cpus or at least one
1176	 * CPU will be left after that.
1177	 */
1178	if ((cmd == partcmd_enable) &&
1179	   (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1180	     cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1181		return -EINVAL;
1182
1183	/*
1184	 * A cpumask update cannot make parent's effective_cpus become empty.
1185	 */
1186	adding = deleting = false;
1187	new_prs = cpuset->partition_root_state;
1188	if (cmd == partcmd_enable) {
1189		cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1190		adding = true;
1191	} else if (cmd == partcmd_disable) {
1192		deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1193				       parent->subparts_cpus);
1194	} else if (newmask) {
1195		/*
1196		 * partcmd_update with newmask:
1197		 *
1198		 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1199		 * addmask = newmask & parent->effective_cpus
1200		 *		     & ~parent->subparts_cpus
1201		 */
1202		cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1203		deleting = cpumask_and(tmp->delmask, tmp->delmask,
1204				       parent->subparts_cpus);
1205
1206		cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1207		adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1208					parent->subparts_cpus);
1209		/*
1210		 * Return error if the new effective_cpus could become empty.
1211		 */
1212		if (adding &&
1213		    cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1214			if (!deleting)
1215				return -EINVAL;
1216			/*
1217			 * As some of the CPUs in subparts_cpus might have
1218			 * been offlined, we need to compute the real delmask
1219			 * to confirm that.
1220			 */
1221			if (!cpumask_and(tmp->addmask, tmp->delmask,
1222					 cpu_active_mask))
1223				return -EINVAL;
1224			cpumask_copy(tmp->addmask, parent->effective_cpus);
1225		}
1226	} else {
1227		/*
1228		 * partcmd_update w/o newmask:
1229		 *
1230		 * addmask = cpus_allowed & parent->effective_cpus
1231		 *
1232		 * Note that parent's subparts_cpus may have been
1233		 * pre-shrunk in case there is a change in the cpu list.
1234		 * So no deletion is needed.
1235		 */
1236		adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1237				     parent->effective_cpus);
1238		part_error = cpumask_equal(tmp->addmask,
1239					   parent->effective_cpus);
1240	}
1241
1242	if (cmd == partcmd_update) {
1243		int prev_prs = cpuset->partition_root_state;
1244
1245		/*
1246		 * Check for possible transition between PRS_ENABLED
1247		 * and PRS_ERROR.
1248		 */
1249		switch (cpuset->partition_root_state) {
1250		case PRS_ENABLED:
1251			if (part_error)
1252				new_prs = PRS_ERROR;
1253			break;
1254		case PRS_ERROR:
1255			if (!part_error)
1256				new_prs = PRS_ENABLED;
1257			break;
1258		}
1259		/*
1260		 * Set part_error if previously in invalid state.
1261		 */
1262		part_error = (prev_prs == PRS_ERROR);
1263	}
1264
1265	if (!part_error && (new_prs == PRS_ERROR))
1266		return 0;	/* Nothing need to be done */
1267
1268	if (new_prs == PRS_ERROR) {
1269		/*
1270		 * Remove all its cpus from parent's subparts_cpus.
1271		 */
1272		adding = false;
1273		deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1274				       parent->subparts_cpus);
1275	}
1276
1277	if (!adding && !deleting && (new_prs == cpuset->partition_root_state))
1278		return 0;
1279
1280	/*
1281	 * Change the parent's subparts_cpus.
1282	 * Newly added CPUs will be removed from effective_cpus and
1283	 * newly deleted ones will be added back to effective_cpus.
1284	 */
1285	spin_lock_irq(&callback_lock);
1286	if (adding) {
1287		cpumask_or(parent->subparts_cpus,
1288			   parent->subparts_cpus, tmp->addmask);
1289		cpumask_andnot(parent->effective_cpus,
1290			       parent->effective_cpus, tmp->addmask);
1291	}
1292	if (deleting) {
1293		cpumask_andnot(parent->subparts_cpus,
1294			       parent->subparts_cpus, tmp->delmask);
1295		/*
1296		 * Some of the CPUs in subparts_cpus might have been offlined.
1297		 */
1298		cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1299		cpumask_or(parent->effective_cpus,
1300			   parent->effective_cpus, tmp->delmask);
1301	}
1302
1303	parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1304
1305	if (cpuset->partition_root_state != new_prs)
1306		cpuset->partition_root_state = new_prs;
1307	spin_unlock_irq(&callback_lock);
1308
1309	return cmd == partcmd_update;
1310}
1311
1312/*
1313 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1314 * @cs:  the cpuset to consider
1315 * @tmp: temp variables for calculating effective_cpus & partition setup
1316 *
1317 * When configured cpumask is changed, the effective cpumasks of this cpuset
1318 * and all its descendants need to be updated.
1319 *
1320 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1321 *
1322 * Called with cpuset_mutex held
1323 */
1324static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1325{
1326	struct cpuset *cp;
1327	struct cgroup_subsys_state *pos_css;
1328	bool need_rebuild_sched_domains = false;
1329	int new_prs;
1330
1331	rcu_read_lock();
1332	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1333		struct cpuset *parent = parent_cs(cp);
1334
1335		compute_effective_cpumask(tmp->new_cpus, cp, parent);
1336
1337		/*
1338		 * If it becomes empty, inherit the effective mask of the
1339		 * parent, which is guaranteed to have some CPUs.
1340		 */
1341		if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1342			cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1343			if (!cp->use_parent_ecpus) {
1344				cp->use_parent_ecpus = true;
1345				parent->child_ecpus_count++;
1346			}
1347		} else if (cp->use_parent_ecpus) {
1348			cp->use_parent_ecpus = false;
1349			WARN_ON_ONCE(!parent->child_ecpus_count);
1350			parent->child_ecpus_count--;
1351		}
1352
1353		/*
1354		 * Skip the whole subtree if the cpumask remains the same
1355		 * and has no partition root state.
1356		 */
1357		if (!cp->partition_root_state &&
1358		    cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1359			pos_css = css_rightmost_descendant(pos_css);
1360			continue;
1361		}
1362
1363		/*
1364		 * update_parent_subparts_cpumask() should have been called
1365		 * for cs already in update_cpumask(). We should also call
1366		 * update_tasks_cpumask() again for tasks in the parent
1367		 * cpuset if the parent's subparts_cpus changes.
1368		 */
1369		new_prs = cp->partition_root_state;
1370		if ((cp != cs) && new_prs) {
1371			switch (parent->partition_root_state) {
1372			case PRS_DISABLED:
1373				/*
1374				 * If parent is not a partition root or an
1375				 * invalid partition root, clear its state
1376				 * and its CS_CPU_EXCLUSIVE flag.
1377				 */
1378				WARN_ON_ONCE(cp->partition_root_state
1379					     != PRS_ERROR);
1380				new_prs = PRS_DISABLED;
1381
1382				/*
1383				 * clear_bit() is an atomic operation and
1384				 * readers aren't interested in the state
1385				 * of CS_CPU_EXCLUSIVE anyway. So we can
1386				 * just update the flag without holding
1387				 * the callback_lock.
1388				 */
1389				clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1390				break;
1391
1392			case PRS_ENABLED:
1393				if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1394					update_tasks_cpumask(parent);
1395				break;
1396
1397			case PRS_ERROR:
1398				/*
1399				 * When parent is invalid, it has to be too.
1400				 */
1401				new_prs = PRS_ERROR;
1402				break;
1403			}
1404		}
1405
1406		if (!css_tryget_online(&cp->css))
1407			continue;
1408		rcu_read_unlock();
1409
1410		spin_lock_irq(&callback_lock);
1411
1412		cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1413		if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1414			cp->nr_subparts_cpus = 0;
1415			cpumask_clear(cp->subparts_cpus);
1416		} else if (cp->nr_subparts_cpus) {
1417			/*
1418			 * Make sure that effective_cpus & subparts_cpus
1419			 * are mutually exclusive.
1420			 *
1421			 * In the unlikely event that effective_cpus
1422			 * becomes empty. we clear cp->nr_subparts_cpus and
1423			 * let its child partition roots to compete for
1424			 * CPUs again.
1425			 */
1426			cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1427				       cp->subparts_cpus);
1428			if (cpumask_empty(cp->effective_cpus)) {
1429				cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1430				cpumask_clear(cp->subparts_cpus);
1431				cp->nr_subparts_cpus = 0;
1432			} else if (!cpumask_subset(cp->subparts_cpus,
1433						   tmp->new_cpus)) {
1434				cpumask_andnot(cp->subparts_cpus,
1435					cp->subparts_cpus, tmp->new_cpus);
1436				cp->nr_subparts_cpus
1437					= cpumask_weight(cp->subparts_cpus);
1438			}
1439		}
1440
1441		if (new_prs != cp->partition_root_state)
1442			cp->partition_root_state = new_prs;
1443
1444		spin_unlock_irq(&callback_lock);
1445
1446		WARN_ON(!is_in_v2_mode() &&
1447			!cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1448
1449		update_tasks_cpumask(cp);
1450
1451		/*
1452		 * On legacy hierarchy, if the effective cpumask of any non-
1453		 * empty cpuset is changed, we need to rebuild sched domains.
1454		 * On default hierarchy, the cpuset needs to be a partition
1455		 * root as well.
1456		 */
1457		if (!cpumask_empty(cp->cpus_allowed) &&
1458		    is_sched_load_balance(cp) &&
1459		   (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1460		    is_partition_root(cp)))
1461			need_rebuild_sched_domains = true;
1462
1463		rcu_read_lock();
1464		css_put(&cp->css);
1465	}
1466	rcu_read_unlock();
1467
1468	if (need_rebuild_sched_domains)
1469		rebuild_sched_domains_locked();
1470}
1471
1472/**
1473 * update_sibling_cpumasks - Update siblings cpumasks
1474 * @parent:  Parent cpuset
1475 * @cs:      Current cpuset
1476 * @tmp:     Temp variables
1477 */
1478static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1479				    struct tmpmasks *tmp)
1480{
1481	struct cpuset *sibling;
1482	struct cgroup_subsys_state *pos_css;
1483
1484	/*
1485	 * Check all its siblings and call update_cpumasks_hier()
1486	 * if their use_parent_ecpus flag is set in order for them
1487	 * to use the right effective_cpus value.
1488	 */
1489	rcu_read_lock();
1490	cpuset_for_each_child(sibling, pos_css, parent) {
1491		if (sibling == cs)
1492			continue;
1493		if (!sibling->use_parent_ecpus)
1494			continue;
1495
1496		update_cpumasks_hier(sibling, tmp);
1497	}
1498	rcu_read_unlock();
1499}
1500
1501/**
1502 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1503 * @cs: the cpuset to consider
1504 * @trialcs: trial cpuset
1505 * @buf: buffer of cpu numbers written to this cpuset
1506 */
1507static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1508			  const char *buf)
1509{
1510	int retval;
1511	struct tmpmasks tmp;
1512
1513	/* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1514	if (cs == &top_cpuset)
1515		return -EACCES;
1516
1517	/*
1518	 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1519	 * Since cpulist_parse() fails on an empty mask, we special case
1520	 * that parsing.  The validate_change() call ensures that cpusets
1521	 * with tasks have cpus.
1522	 */
1523	if (!*buf) {
1524		cpumask_clear(trialcs->cpus_allowed);
1525	} else {
1526		retval = cpulist_parse(buf, trialcs->cpus_allowed);
1527		if (retval < 0)
1528			return retval;
1529
1530		if (!cpumask_subset(trialcs->cpus_allowed,
1531				    top_cpuset.cpus_allowed))
1532			return -EINVAL;
1533	}
1534
1535	/* Nothing to do if the cpus didn't change */
1536	if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1537		return 0;
1538
1539	retval = validate_change(cs, trialcs);
1540	if (retval < 0)
1541		return retval;
1542
1543#ifdef CONFIG_CPUMASK_OFFSTACK
1544	/*
1545	 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1546	 * to allocated cpumasks.
1547	 */
1548	tmp.addmask  = trialcs->subparts_cpus;
1549	tmp.delmask  = trialcs->effective_cpus;
1550	tmp.new_cpus = trialcs->cpus_allowed;
1551#endif
1552
1553	if (cs->partition_root_state) {
1554		/* Cpumask of a partition root cannot be empty */
1555		if (cpumask_empty(trialcs->cpus_allowed))
1556			return -EINVAL;
1557		if (update_parent_subparts_cpumask(cs, partcmd_update,
1558					trialcs->cpus_allowed, &tmp) < 0)
1559			return -EINVAL;
1560	}
1561
1562	spin_lock_irq(&callback_lock);
1563	cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1564
1565	/*
1566	 * Make sure that subparts_cpus is a subset of cpus_allowed.
1567	 */
1568	if (cs->nr_subparts_cpus) {
1569		cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1570			       cs->cpus_allowed);
1571		cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1572	}
1573	spin_unlock_irq(&callback_lock);
1574
1575	update_cpumasks_hier(cs, &tmp);
1576
1577	if (cs->partition_root_state) {
1578		struct cpuset *parent = parent_cs(cs);
1579
1580		/*
1581		 * For partition root, update the cpumasks of sibling
1582		 * cpusets if they use parent's effective_cpus.
1583		 */
1584		if (parent->child_ecpus_count)
1585			update_sibling_cpumasks(parent, cs, &tmp);
1586	}
1587	return 0;
1588}
1589
1590/*
1591 * Migrate memory region from one set of nodes to another.  This is
1592 * performed asynchronously as it can be called from process migration path
1593 * holding locks involved in process management.  All mm migrations are
1594 * performed in the queued order and can be waited for by flushing
1595 * cpuset_migrate_mm_wq.
1596 */
1597
1598struct cpuset_migrate_mm_work {
1599	struct work_struct	work;
1600	struct mm_struct	*mm;
1601	nodemask_t		from;
1602	nodemask_t		to;
1603};
1604
1605static void cpuset_migrate_mm_workfn(struct work_struct *work)
1606{
1607	struct cpuset_migrate_mm_work *mwork =
1608		container_of(work, struct cpuset_migrate_mm_work, work);
1609
1610	/* on a wq worker, no need to worry about %current's mems_allowed */
1611	do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1612	mmput(mwork->mm);
1613	kfree(mwork);
1614}
1615
1616static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1617							const nodemask_t *to)
1618{
1619	struct cpuset_migrate_mm_work *mwork;
1620
1621	mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1622	if (mwork) {
1623		mwork->mm = mm;
1624		mwork->from = *from;
1625		mwork->to = *to;
1626		INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1627		queue_work(cpuset_migrate_mm_wq, &mwork->work);
1628	} else {
1629		mmput(mm);
1630	}
1631}
1632
1633static void cpuset_post_attach(void)
1634{
1635	flush_workqueue(cpuset_migrate_mm_wq);
1636}
1637
1638/*
1639 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1640 * @tsk: the task to change
1641 * @newmems: new nodes that the task will be set
1642 *
1643 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1644 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1645 * parallel, it might temporarily see an empty intersection, which results in
1646 * a seqlock check and retry before OOM or allocation failure.
1647 */
1648static void cpuset_change_task_nodemask(struct task_struct *tsk,
1649					nodemask_t *newmems)
1650{
1651	task_lock(tsk);
1652
1653	local_irq_disable();
1654	write_seqcount_begin(&tsk->mems_allowed_seq);
1655
1656	nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1657	mpol_rebind_task(tsk, newmems);
1658	tsk->mems_allowed = *newmems;
1659
1660	write_seqcount_end(&tsk->mems_allowed_seq);
1661	local_irq_enable();
1662
1663	task_unlock(tsk);
1664}
1665
1666static void *cpuset_being_rebound;
1667
1668/**
1669 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1670 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1671 *
1672 * Iterate through each task of @cs updating its mems_allowed to the
1673 * effective cpuset's.  As this function is called with cpuset_mutex held,
1674 * cpuset membership stays stable.
1675 */
1676static void update_tasks_nodemask(struct cpuset *cs)
1677{
1678	static nodemask_t newmems;	/* protected by cpuset_mutex */
1679	struct css_task_iter it;
1680	struct task_struct *task;
1681
1682	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
1683
1684	guarantee_online_mems(cs, &newmems);
1685
1686	/*
1687	 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1688	 * take while holding tasklist_lock.  Forks can happen - the
1689	 * mpol_dup() cpuset_being_rebound check will catch such forks,
1690	 * and rebind their vma mempolicies too.  Because we still hold
1691	 * the global cpuset_mutex, we know that no other rebind effort
1692	 * will be contending for the global variable cpuset_being_rebound.
1693	 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1694	 * is idempotent.  Also migrate pages in each mm to new nodes.
1695	 */
1696	css_task_iter_start(&cs->css, 0, &it);
1697	while ((task = css_task_iter_next(&it))) {
1698		struct mm_struct *mm;
1699		bool migrate;
1700
1701		cpuset_change_task_nodemask(task, &newmems);
1702
1703		mm = get_task_mm(task);
1704		if (!mm)
1705			continue;
1706
1707		migrate = is_memory_migrate(cs);
1708
1709		mpol_rebind_mm(mm, &cs->mems_allowed);
1710		if (migrate)
1711			cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1712		else
1713			mmput(mm);
1714	}
1715	css_task_iter_end(&it);
1716
1717	/*
1718	 * All the tasks' nodemasks have been updated, update
1719	 * cs->old_mems_allowed.
1720	 */
1721	cs->old_mems_allowed = newmems;
1722
1723	/* We're done rebinding vmas to this cpuset's new mems_allowed. */
1724	cpuset_being_rebound = NULL;
1725}
1726
1727/*
1728 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1729 * @cs: the cpuset to consider
1730 * @new_mems: a temp variable for calculating new effective_mems
1731 *
1732 * When configured nodemask is changed, the effective nodemasks of this cpuset
1733 * and all its descendants need to be updated.
1734 *
1735 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
1736 *
1737 * Called with cpuset_mutex held
1738 */
1739static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1740{
1741	struct cpuset *cp;
1742	struct cgroup_subsys_state *pos_css;
1743
1744	rcu_read_lock();
1745	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1746		struct cpuset *parent = parent_cs(cp);
1747
1748		nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1749
1750		/*
1751		 * If it becomes empty, inherit the effective mask of the
1752		 * parent, which is guaranteed to have some MEMs.
1753		 */
1754		if (is_in_v2_mode() && nodes_empty(*new_mems))
1755			*new_mems = parent->effective_mems;
1756
1757		/* Skip the whole subtree if the nodemask remains the same. */
1758		if (nodes_equal(*new_mems, cp->effective_mems)) {
1759			pos_css = css_rightmost_descendant(pos_css);
1760			continue;
1761		}
1762
1763		if (!css_tryget_online(&cp->css))
1764			continue;
1765		rcu_read_unlock();
1766
1767		spin_lock_irq(&callback_lock);
1768		cp->effective_mems = *new_mems;
1769		spin_unlock_irq(&callback_lock);
1770
1771		WARN_ON(!is_in_v2_mode() &&
1772			!nodes_equal(cp->mems_allowed, cp->effective_mems));
1773
1774		update_tasks_nodemask(cp);
1775
1776		rcu_read_lock();
1777		css_put(&cp->css);
1778	}
1779	rcu_read_unlock();
1780}
1781
1782/*
1783 * Handle user request to change the 'mems' memory placement
1784 * of a cpuset.  Needs to validate the request, update the
1785 * cpusets mems_allowed, and for each task in the cpuset,
1786 * update mems_allowed and rebind task's mempolicy and any vma
1787 * mempolicies and if the cpuset is marked 'memory_migrate',
1788 * migrate the tasks pages to the new memory.
1789 *
1790 * Call with cpuset_mutex held. May take callback_lock during call.
1791 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1792 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1793 * their mempolicies to the cpusets new mems_allowed.
1794 */
1795static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1796			   const char *buf)
1797{
1798	int retval;
1799
1800	/*
1801	 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1802	 * it's read-only
1803	 */
1804	if (cs == &top_cpuset) {
1805		retval = -EACCES;
1806		goto done;
1807	}
1808
1809	/*
1810	 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1811	 * Since nodelist_parse() fails on an empty mask, we special case
1812	 * that parsing.  The validate_change() call ensures that cpusets
1813	 * with tasks have memory.
1814	 */
1815	if (!*buf) {
1816		nodes_clear(trialcs->mems_allowed);
1817	} else {
1818		retval = nodelist_parse(buf, trialcs->mems_allowed);
1819		if (retval < 0)
1820			goto done;
1821
1822		if (!nodes_subset(trialcs->mems_allowed,
1823				  top_cpuset.mems_allowed)) {
1824			retval = -EINVAL;
1825			goto done;
1826		}
1827	}
1828
1829	if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1830		retval = 0;		/* Too easy - nothing to do */
1831		goto done;
1832	}
1833	retval = validate_change(cs, trialcs);
1834	if (retval < 0)
1835		goto done;
1836
1837	spin_lock_irq(&callback_lock);
1838	cs->mems_allowed = trialcs->mems_allowed;
1839	spin_unlock_irq(&callback_lock);
1840
1841	/* use trialcs->mems_allowed as a temp variable */
1842	update_nodemasks_hier(cs, &trialcs->mems_allowed);
1843done:
1844	return retval;
1845}
1846
1847bool current_cpuset_is_being_rebound(void)
1848{
1849	bool ret;
1850
1851	rcu_read_lock();
1852	ret = task_cs(current) == cpuset_being_rebound;
1853	rcu_read_unlock();
1854
1855	return ret;
1856}
1857
1858static int update_relax_domain_level(struct cpuset *cs, s64 val)
1859{
1860#ifdef CONFIG_SMP
1861	if (val < -1 || val >= sched_domain_level_max)
1862		return -EINVAL;
1863#endif
1864
1865	if (val != cs->relax_domain_level) {
1866		cs->relax_domain_level = val;
1867		if (!cpumask_empty(cs->cpus_allowed) &&
1868		    is_sched_load_balance(cs))
1869			rebuild_sched_domains_locked();
1870	}
1871
1872	return 0;
1873}
1874
1875/**
1876 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1877 * @cs: the cpuset in which each task's spread flags needs to be changed
1878 *
1879 * Iterate through each task of @cs updating its spread flags.  As this
1880 * function is called with cpuset_mutex held, cpuset membership stays
1881 * stable.
1882 */
1883static void update_tasks_flags(struct cpuset *cs)
1884{
1885	struct css_task_iter it;
1886	struct task_struct *task;
1887
1888	css_task_iter_start(&cs->css, 0, &it);
1889	while ((task = css_task_iter_next(&it)))
1890		cpuset_update_task_spread_flag(cs, task);
1891	css_task_iter_end(&it);
1892}
1893
1894/*
1895 * update_flag - read a 0 or a 1 in a file and update associated flag
1896 * bit:		the bit to update (see cpuset_flagbits_t)
1897 * cs:		the cpuset to update
1898 * turning_on: 	whether the flag is being set or cleared
1899 *
1900 * Call with cpuset_mutex held.
1901 */
1902
1903static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1904		       int turning_on)
1905{
1906	struct cpuset *trialcs;
1907	int balance_flag_changed;
1908	int spread_flag_changed;
1909	int err;
1910
1911	trialcs = alloc_trial_cpuset(cs);
1912	if (!trialcs)
1913		return -ENOMEM;
1914
1915	if (turning_on)
1916		set_bit(bit, &trialcs->flags);
1917	else
1918		clear_bit(bit, &trialcs->flags);
1919
1920	err = validate_change(cs, trialcs);
1921	if (err < 0)
1922		goto out;
1923
1924	balance_flag_changed = (is_sched_load_balance(cs) !=
1925				is_sched_load_balance(trialcs));
1926
1927	spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1928			|| (is_spread_page(cs) != is_spread_page(trialcs)));
1929
1930	spin_lock_irq(&callback_lock);
1931	cs->flags = trialcs->flags;
1932	spin_unlock_irq(&callback_lock);
1933
1934	if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1935		rebuild_sched_domains_locked();
1936
1937	if (spread_flag_changed)
1938		update_tasks_flags(cs);
1939out:
1940	free_cpuset(trialcs);
1941	return err;
1942}
1943
1944/*
1945 * update_prstate - update partititon_root_state
1946 * cs: the cpuset to update
1947 * new_prs: new partition root state
1948 *
1949 * Call with cpuset_mutex held.
1950 */
1951static int update_prstate(struct cpuset *cs, int new_prs)
1952{
1953	int err, old_prs = cs->partition_root_state;
1954	struct cpuset *parent = parent_cs(cs);
1955	struct tmpmasks tmpmask;
1956
1957	if (old_prs == new_prs)
1958		return 0;
1959
1960	/*
1961	 * Cannot force a partial or invalid partition root to a full
1962	 * partition root.
1963	 */
1964	if (new_prs && (old_prs == PRS_ERROR))
1965		return -EINVAL;
1966
1967	if (alloc_cpumasks(NULL, &tmpmask))
1968		return -ENOMEM;
1969
1970	err = -EINVAL;
1971	if (!old_prs) {
1972		/*
1973		 * Turning on partition root requires setting the
1974		 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1975		 * cannot be NULL.
1976		 */
1977		if (cpumask_empty(cs->cpus_allowed))
1978			goto out;
1979
1980		err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1981		if (err)
1982			goto out;
1983
1984		err = update_parent_subparts_cpumask(cs, partcmd_enable,
1985						     NULL, &tmpmask);
1986		if (err) {
1987			update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1988			goto out;
1989		}
1990	} else {
1991		/*
1992		 * Turning off partition root will clear the
1993		 * CS_CPU_EXCLUSIVE bit.
1994		 */
1995		if (old_prs == PRS_ERROR) {
1996			update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1997			err = 0;
1998			goto out;
1999		}
2000
2001		err = update_parent_subparts_cpumask(cs, partcmd_disable,
2002						     NULL, &tmpmask);
2003		if (err)
2004			goto out;
2005
2006		/* Turning off CS_CPU_EXCLUSIVE will not return error */
2007		update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2008	}
2009
2010	/*
2011	 * Update cpumask of parent's tasks except when it is the top
2012	 * cpuset as some system daemons cannot be mapped to other CPUs.
2013	 */
2014	if (parent != &top_cpuset)
2015		update_tasks_cpumask(parent);
2016
2017	if (parent->child_ecpus_count)
2018		update_sibling_cpumasks(parent, cs, &tmpmask);
2019
2020	rebuild_sched_domains_locked();
2021out:
2022	if (!err) {
2023		spin_lock_irq(&callback_lock);
2024		cs->partition_root_state = new_prs;
2025		spin_unlock_irq(&callback_lock);
2026	}
2027
2028	free_cpumasks(NULL, &tmpmask);
2029	return err;
2030}
2031
2032/*
2033 * Frequency meter - How fast is some event occurring?
2034 *
2035 * These routines manage a digitally filtered, constant time based,
2036 * event frequency meter.  There are four routines:
2037 *   fmeter_init() - initialize a frequency meter.
2038 *   fmeter_markevent() - called each time the event happens.
2039 *   fmeter_getrate() - returns the recent rate of such events.
2040 *   fmeter_update() - internal routine used to update fmeter.
2041 *
2042 * A common data structure is passed to each of these routines,
2043 * which is used to keep track of the state required to manage the
2044 * frequency meter and its digital filter.
2045 *
2046 * The filter works on the number of events marked per unit time.
2047 * The filter is single-pole low-pass recursive (IIR).  The time unit
2048 * is 1 second.  Arithmetic is done using 32-bit integers scaled to
2049 * simulate 3 decimal digits of precision (multiplied by 1000).
2050 *
2051 * With an FM_COEF of 933, and a time base of 1 second, the filter
2052 * has a half-life of 10 seconds, meaning that if the events quit
2053 * happening, then the rate returned from the fmeter_getrate()
2054 * will be cut in half each 10 seconds, until it converges to zero.
2055 *
2056 * It is not worth doing a real infinitely recursive filter.  If more
2057 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2058 * just compute FM_MAXTICKS ticks worth, by which point the level
2059 * will be stable.
2060 *
2061 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2062 * arithmetic overflow in the fmeter_update() routine.
2063 *
2064 * Given the simple 32 bit integer arithmetic used, this meter works
2065 * best for reporting rates between one per millisecond (msec) and
2066 * one per 32 (approx) seconds.  At constant rates faster than one
2067 * per msec it maxes out at values just under 1,000,000.  At constant
2068 * rates between one per msec, and one per second it will stabilize
2069 * to a value N*1000, where N is the rate of events per second.
2070 * At constant rates between one per second and one per 32 seconds,
2071 * it will be choppy, moving up on the seconds that have an event,
2072 * and then decaying until the next event.  At rates slower than
2073 * about one in 32 seconds, it decays all the way back to zero between
2074 * each event.
2075 */
2076
2077#define FM_COEF 933		/* coefficient for half-life of 10 secs */
2078#define FM_MAXTICKS ((u32)99)   /* useless computing more ticks than this */
2079#define FM_MAXCNT 1000000	/* limit cnt to avoid overflow */
2080#define FM_SCALE 1000		/* faux fixed point scale */
2081
2082/* Initialize a frequency meter */
2083static void fmeter_init(struct fmeter *fmp)
2084{
2085	fmp->cnt = 0;
2086	fmp->val = 0;
2087	fmp->time = 0;
2088	spin_lock_init(&fmp->lock);
2089}
2090
2091/* Internal meter update - process cnt events and update value */
2092static void fmeter_update(struct fmeter *fmp)
2093{
2094	time64_t now;
2095	u32 ticks;
2096
2097	now = ktime_get_seconds();
2098	ticks = now - fmp->time;
2099
2100	if (ticks == 0)
2101		return;
2102
2103	ticks = min(FM_MAXTICKS, ticks);
2104	while (ticks-- > 0)
2105		fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2106	fmp->time = now;
2107
2108	fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2109	fmp->cnt = 0;
2110}
2111
2112/* Process any previous ticks, then bump cnt by one (times scale). */
2113static void fmeter_markevent(struct fmeter *fmp)
2114{
2115	spin_lock(&fmp->lock);
2116	fmeter_update(fmp);
2117	fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2118	spin_unlock(&fmp->lock);
2119}
2120
2121/* Process any previous ticks, then return current value. */
2122static int fmeter_getrate(struct fmeter *fmp)
2123{
2124	int val;
2125
2126	spin_lock(&fmp->lock);
2127	fmeter_update(fmp);
2128	val = fmp->val;
2129	spin_unlock(&fmp->lock);
2130	return val;
2131}
2132
2133static struct cpuset *cpuset_attach_old_cs;
2134
2135/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2136static int cpuset_can_attach(struct cgroup_taskset *tset)
2137{
2138	struct cgroup_subsys_state *css;
2139	struct cpuset *cs;
2140	struct task_struct *task;
2141	int ret;
2142
2143	/* used later by cpuset_attach() */
2144	cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2145	cs = css_cs(css);
2146
2147	percpu_down_write(&cpuset_rwsem);
2148
2149	/* allow moving tasks into an empty cpuset if on default hierarchy */
2150	ret = -ENOSPC;
2151	if (!is_in_v2_mode() &&
2152	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2153		goto out_unlock;
2154
2155	cgroup_taskset_for_each(task, css, tset) {
2156		ret = task_can_attach(task, cs->cpus_allowed);
2157		if (ret)
2158			goto out_unlock;
2159		ret = security_task_setscheduler(task);
2160		if (ret)
2161			goto out_unlock;
2162	}
2163
2164	/*
2165	 * Mark attach is in progress.  This makes validate_change() fail
2166	 * changes which zero cpus/mems_allowed.
2167	 */
2168	cs->attach_in_progress++;
2169	ret = 0;
2170out_unlock:
2171	percpu_up_write(&cpuset_rwsem);
2172	return ret;
2173}
2174
2175static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2176{
2177	struct cgroup_subsys_state *css;
2178
2179	cgroup_taskset_first(tset, &css);
2180
2181	percpu_down_write(&cpuset_rwsem);
2182	css_cs(css)->attach_in_progress--;
2183	percpu_up_write(&cpuset_rwsem);
2184}
2185
2186/*
2187 * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
2188 * but we can't allocate it dynamically there.  Define it global and
2189 * allocate from cpuset_init().
2190 */
2191static cpumask_var_t cpus_attach;
2192
2193static void cpuset_attach(struct cgroup_taskset *tset)
2194{
2195	/* static buf protected by cpuset_mutex */
2196	static nodemask_t cpuset_attach_nodemask_to;
2197	struct task_struct *task;
2198	struct task_struct *leader;
2199	struct cgroup_subsys_state *css;
2200	struct cpuset *cs;
2201	struct cpuset *oldcs = cpuset_attach_old_cs;
2202
2203	cgroup_taskset_first(tset, &css);
2204	cs = css_cs(css);
2205
2206	percpu_down_write(&cpuset_rwsem);
2207
2208	/* prepare for attach */
2209	if (cs == &top_cpuset)
2210		cpumask_copy(cpus_attach, cpu_possible_mask);
2211	else
2212		guarantee_online_cpus(cs, cpus_attach);
2213
2214	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2215
2216	cgroup_taskset_for_each(task, css, tset) {
2217		/*
2218		 * can_attach beforehand should guarantee that this doesn't
2219		 * fail.  TODO: have a better way to handle failure here
2220		 */
2221		WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2222
2223		cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2224		cpuset_update_task_spread_flag(cs, task);
2225	}
2226
2227	/*
2228	 * Change mm for all threadgroup leaders. This is expensive and may
2229	 * sleep and should be moved outside migration path proper.
2230	 */
2231	cpuset_attach_nodemask_to = cs->effective_mems;
2232	cgroup_taskset_for_each_leader(leader, css, tset) {
2233		struct mm_struct *mm = get_task_mm(leader);
2234
2235		if (mm) {
2236			mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2237
2238			/*
2239			 * old_mems_allowed is the same with mems_allowed
2240			 * here, except if this task is being moved
2241			 * automatically due to hotplug.  In that case
2242			 * @mems_allowed has been updated and is empty, so
2243			 * @old_mems_allowed is the right nodesets that we
2244			 * migrate mm from.
2245			 */
2246			if (is_memory_migrate(cs))
2247				cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2248						  &cpuset_attach_nodemask_to);
2249			else
2250				mmput(mm);
2251		}
2252	}
2253
2254	cs->old_mems_allowed = cpuset_attach_nodemask_to;
2255
2256	cs->attach_in_progress--;
2257	if (!cs->attach_in_progress)
2258		wake_up(&cpuset_attach_wq);
2259
2260	percpu_up_write(&cpuset_rwsem);
2261}
2262
2263/* The various types of files and directories in a cpuset file system */
2264
2265typedef enum {
2266	FILE_MEMORY_MIGRATE,
2267	FILE_CPULIST,
2268	FILE_MEMLIST,
2269	FILE_EFFECTIVE_CPULIST,
2270	FILE_EFFECTIVE_MEMLIST,
2271	FILE_SUBPARTS_CPULIST,
2272	FILE_CPU_EXCLUSIVE,
2273	FILE_MEM_EXCLUSIVE,
2274	FILE_MEM_HARDWALL,
2275	FILE_SCHED_LOAD_BALANCE,
2276	FILE_PARTITION_ROOT,
2277	FILE_SCHED_RELAX_DOMAIN_LEVEL,
2278	FILE_MEMORY_PRESSURE_ENABLED,
2279	FILE_MEMORY_PRESSURE,
2280	FILE_SPREAD_PAGE,
2281	FILE_SPREAD_SLAB,
2282} cpuset_filetype_t;
2283
2284static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2285			    u64 val)
2286{
2287	struct cpuset *cs = css_cs(css);
2288	cpuset_filetype_t type = cft->private;
2289	int retval = 0;
2290
2291	get_online_cpus();
2292	percpu_down_write(&cpuset_rwsem);
2293	if (!is_cpuset_online(cs)) {
2294		retval = -ENODEV;
2295		goto out_unlock;
2296	}
2297
2298	switch (type) {
2299	case FILE_CPU_EXCLUSIVE:
2300		retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2301		break;
2302	case FILE_MEM_EXCLUSIVE:
2303		retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2304		break;
2305	case FILE_MEM_HARDWALL:
2306		retval = update_flag(CS_MEM_HARDWALL, cs, val);
2307		break;
2308	case FILE_SCHED_LOAD_BALANCE:
2309		retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2310		break;
2311	case FILE_MEMORY_MIGRATE:
2312		retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2313		break;
2314	case FILE_MEMORY_PRESSURE_ENABLED:
2315		cpuset_memory_pressure_enabled = !!val;
2316		break;
2317	case FILE_SPREAD_PAGE:
2318		retval = update_flag(CS_SPREAD_PAGE, cs, val);
2319		break;
2320	case FILE_SPREAD_SLAB:
2321		retval = update_flag(CS_SPREAD_SLAB, cs, val);
2322		break;
2323	default:
2324		retval = -EINVAL;
2325		break;
2326	}
2327out_unlock:
2328	percpu_up_write(&cpuset_rwsem);
2329	put_online_cpus();
2330	return retval;
2331}
2332
2333static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2334			    s64 val)
2335{
2336	struct cpuset *cs = css_cs(css);
2337	cpuset_filetype_t type = cft->private;
2338	int retval = -ENODEV;
2339
2340	get_online_cpus();
2341	percpu_down_write(&cpuset_rwsem);
2342	if (!is_cpuset_online(cs))
2343		goto out_unlock;
2344
2345	switch (type) {
2346	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2347		retval = update_relax_domain_level(cs, val);
2348		break;
2349	default:
2350		retval = -EINVAL;
2351		break;
2352	}
2353out_unlock:
2354	percpu_up_write(&cpuset_rwsem);
2355	put_online_cpus();
2356	return retval;
2357}
2358
2359/*
2360 * Common handling for a write to a "cpus" or "mems" file.
2361 */
2362static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2363				    char *buf, size_t nbytes, loff_t off)
2364{
2365	struct cpuset *cs = css_cs(of_css(of));
2366	struct cpuset *trialcs;
2367	int retval = -ENODEV;
2368
2369	buf = strstrip(buf);
2370
2371	/*
2372	 * CPU or memory hotunplug may leave @cs w/o any execution
2373	 * resources, in which case the hotplug code asynchronously updates
2374	 * configuration and transfers all tasks to the nearest ancestor
2375	 * which can execute.
2376	 *
2377	 * As writes to "cpus" or "mems" may restore @cs's execution
2378	 * resources, wait for the previously scheduled operations before
2379	 * proceeding, so that we don't end up keep removing tasks added
2380	 * after execution capability is restored.
2381	 *
2382	 * cpuset_hotplug_work calls back into cgroup core via
2383	 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2384	 * operation like this one can lead to a deadlock through kernfs
2385	 * active_ref protection.  Let's break the protection.  Losing the
2386	 * protection is okay as we check whether @cs is online after
2387	 * grabbing cpuset_mutex anyway.  This only happens on the legacy
2388	 * hierarchies.
2389	 */
2390	css_get(&cs->css);
2391	kernfs_break_active_protection(of->kn);
2392	flush_work(&cpuset_hotplug_work);
2393
2394	get_online_cpus();
2395	percpu_down_write(&cpuset_rwsem);
2396	if (!is_cpuset_online(cs))
2397		goto out_unlock;
2398
2399	trialcs = alloc_trial_cpuset(cs);
2400	if (!trialcs) {
2401		retval = -ENOMEM;
2402		goto out_unlock;
2403	}
2404
2405	switch (of_cft(of)->private) {
2406	case FILE_CPULIST:
2407		retval = update_cpumask(cs, trialcs, buf);
2408		break;
2409	case FILE_MEMLIST:
2410		retval = update_nodemask(cs, trialcs, buf);
2411		break;
2412	default:
2413		retval = -EINVAL;
2414		break;
2415	}
2416
2417	free_cpuset(trialcs);
2418out_unlock:
2419	percpu_up_write(&cpuset_rwsem);
2420	put_online_cpus();
2421	kernfs_unbreak_active_protection(of->kn);
2422	css_put(&cs->css);
2423	flush_workqueue(cpuset_migrate_mm_wq);
2424	return retval ?: nbytes;
2425}
2426
2427/*
2428 * These ascii lists should be read in a single call, by using a user
2429 * buffer large enough to hold the entire map.  If read in smaller
2430 * chunks, there is no guarantee of atomicity.  Since the display format
2431 * used, list of ranges of sequential numbers, is variable length,
2432 * and since these maps can change value dynamically, one could read
2433 * gibberish by doing partial reads while a list was changing.
2434 */
2435static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2436{
2437	struct cpuset *cs = css_cs(seq_css(sf));
2438	cpuset_filetype_t type = seq_cft(sf)->private;
2439	int ret = 0;
2440
2441	spin_lock_irq(&callback_lock);
2442
2443	switch (type) {
2444	case FILE_CPULIST:
2445		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2446		break;
2447	case FILE_MEMLIST:
2448		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2449		break;
2450	case FILE_EFFECTIVE_CPULIST:
2451		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2452		break;
2453	case FILE_EFFECTIVE_MEMLIST:
2454		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2455		break;
2456	case FILE_SUBPARTS_CPULIST:
2457		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2458		break;
2459	default:
2460		ret = -EINVAL;
2461	}
2462
2463	spin_unlock_irq(&callback_lock);
2464	return ret;
2465}
2466
2467static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2468{
2469	struct cpuset *cs = css_cs(css);
2470	cpuset_filetype_t type = cft->private;
2471	switch (type) {
2472	case FILE_CPU_EXCLUSIVE:
2473		return is_cpu_exclusive(cs);
2474	case FILE_MEM_EXCLUSIVE:
2475		return is_mem_exclusive(cs);
2476	case FILE_MEM_HARDWALL:
2477		return is_mem_hardwall(cs);
2478	case FILE_SCHED_LOAD_BALANCE:
2479		return is_sched_load_balance(cs);
2480	case FILE_MEMORY_MIGRATE:
2481		return is_memory_migrate(cs);
2482	case FILE_MEMORY_PRESSURE_ENABLED:
2483		return cpuset_memory_pressure_enabled;
2484	case FILE_MEMORY_PRESSURE:
2485		return fmeter_getrate(&cs->fmeter);
2486	case FILE_SPREAD_PAGE:
2487		return is_spread_page(cs);
2488	case FILE_SPREAD_SLAB:
2489		return is_spread_slab(cs);
2490	default:
2491		BUG();
2492	}
2493
2494	/* Unreachable but makes gcc happy */
2495	return 0;
2496}
2497
2498static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2499{
2500	struct cpuset *cs = css_cs(css);
2501	cpuset_filetype_t type = cft->private;
2502	switch (type) {
2503	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2504		return cs->relax_domain_level;
2505	default:
2506		BUG();
2507	}
2508
2509	/* Unreachable but makes gcc happy */
2510	return 0;
2511}
2512
2513static int sched_partition_show(struct seq_file *seq, void *v)
2514{
2515	struct cpuset *cs = css_cs(seq_css(seq));
2516
2517	switch (cs->partition_root_state) {
2518	case PRS_ENABLED:
2519		seq_puts(seq, "root\n");
2520		break;
2521	case PRS_DISABLED:
2522		seq_puts(seq, "member\n");
2523		break;
2524	case PRS_ERROR:
2525		seq_puts(seq, "root invalid\n");
2526		break;
2527	}
2528	return 0;
2529}
2530
2531static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2532				     size_t nbytes, loff_t off)
2533{
2534	struct cpuset *cs = css_cs(of_css(of));
2535	int val;
2536	int retval = -ENODEV;
2537
2538	buf = strstrip(buf);
2539
2540	/*
2541	 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2542	 */
2543	if (!strcmp(buf, "root"))
2544		val = PRS_ENABLED;
2545	else if (!strcmp(buf, "member"))
2546		val = PRS_DISABLED;
2547	else
2548		return -EINVAL;
2549
2550	css_get(&cs->css);
2551	get_online_cpus();
2552	percpu_down_write(&cpuset_rwsem);
2553	if (!is_cpuset_online(cs))
2554		goto out_unlock;
2555
2556	retval = update_prstate(cs, val);
2557out_unlock:
2558	percpu_up_write(&cpuset_rwsem);
2559	put_online_cpus();
2560	css_put(&cs->css);
2561	return retval ?: nbytes;
2562}
2563
2564/*
2565 * for the common functions, 'private' gives the type of file
2566 */
2567
2568static struct cftype legacy_files[] = {
2569	{
2570		.name = "cpus",
2571		.seq_show = cpuset_common_seq_show,
2572		.write = cpuset_write_resmask,
2573		.max_write_len = (100U + 6 * NR_CPUS),
2574		.private = FILE_CPULIST,
2575	},
2576
2577	{
2578		.name = "mems",
2579		.seq_show = cpuset_common_seq_show,
2580		.write = cpuset_write_resmask,
2581		.max_write_len = (100U + 6 * MAX_NUMNODES),
2582		.private = FILE_MEMLIST,
2583	},
2584
2585	{
2586		.name = "effective_cpus",
2587		.seq_show = cpuset_common_seq_show,
2588		.private = FILE_EFFECTIVE_CPULIST,
2589	},
2590
2591	{
2592		.name = "effective_mems",
2593		.seq_show = cpuset_common_seq_show,
2594		.private = FILE_EFFECTIVE_MEMLIST,
2595	},
2596
2597	{
2598		.name = "cpu_exclusive",
2599		.read_u64 = cpuset_read_u64,
2600		.write_u64 = cpuset_write_u64,
2601		.private = FILE_CPU_EXCLUSIVE,
2602	},
2603
2604	{
2605		.name = "mem_exclusive",
2606		.read_u64 = cpuset_read_u64,
2607		.write_u64 = cpuset_write_u64,
2608		.private = FILE_MEM_EXCLUSIVE,
2609	},
2610
2611	{
2612		.name = "mem_hardwall",
2613		.read_u64 = cpuset_read_u64,
2614		.write_u64 = cpuset_write_u64,
2615		.private = FILE_MEM_HARDWALL,
2616	},
2617
2618	{
2619		.name = "sched_load_balance",
2620		.read_u64 = cpuset_read_u64,
2621		.write_u64 = cpuset_write_u64,
2622		.private = FILE_SCHED_LOAD_BALANCE,
2623	},
2624
2625	{
2626		.name = "sched_relax_domain_level",
2627		.read_s64 = cpuset_read_s64,
2628		.write_s64 = cpuset_write_s64,
2629		.private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2630	},
2631
2632	{
2633		.name = "memory_migrate",
2634		.read_u64 = cpuset_read_u64,
2635		.write_u64 = cpuset_write_u64,
2636		.private = FILE_MEMORY_MIGRATE,
2637	},
2638
2639	{
2640		.name = "memory_pressure",
2641		.read_u64 = cpuset_read_u64,
2642		.private = FILE_MEMORY_PRESSURE,
2643	},
2644
2645	{
2646		.name = "memory_spread_page",
2647		.read_u64 = cpuset_read_u64,
2648		.write_u64 = cpuset_write_u64,
2649		.private = FILE_SPREAD_PAGE,
2650	},
2651
2652	{
2653		.name = "memory_spread_slab",
2654		.read_u64 = cpuset_read_u64,
2655		.write_u64 = cpuset_write_u64,
2656		.private = FILE_SPREAD_SLAB,
2657	},
2658
2659	{
2660		.name = "memory_pressure_enabled",
2661		.flags = CFTYPE_ONLY_ON_ROOT,
2662		.read_u64 = cpuset_read_u64,
2663		.write_u64 = cpuset_write_u64,
2664		.private = FILE_MEMORY_PRESSURE_ENABLED,
2665	},
2666
2667	{ }	/* terminate */
2668};
2669
2670/*
2671 * This is currently a minimal set for the default hierarchy. It can be
2672 * expanded later on by migrating more features and control files from v1.
2673 */
2674static struct cftype dfl_files[] = {
2675	{
2676		.name = "cpus",
2677		.seq_show = cpuset_common_seq_show,
2678		.write = cpuset_write_resmask,
2679		.max_write_len = (100U + 6 * NR_CPUS),
2680		.private = FILE_CPULIST,
2681		.flags = CFTYPE_NOT_ON_ROOT,
2682	},
2683
2684	{
2685		.name = "mems",
2686		.seq_show = cpuset_common_seq_show,
2687		.write = cpuset_write_resmask,
2688		.max_write_len = (100U + 6 * MAX_NUMNODES),
2689		.private = FILE_MEMLIST,
2690		.flags = CFTYPE_NOT_ON_ROOT,
2691	},
2692
2693	{
2694		.name = "cpus.effective",
2695		.seq_show = cpuset_common_seq_show,
2696		.private = FILE_EFFECTIVE_CPULIST,
2697	},
2698
2699	{
2700		.name = "mems.effective",
2701		.seq_show = cpuset_common_seq_show,
2702		.private = FILE_EFFECTIVE_MEMLIST,
2703	},
2704
2705	{
2706		.name = "cpus.partition",
2707		.seq_show = sched_partition_show,
2708		.write = sched_partition_write,
2709		.private = FILE_PARTITION_ROOT,
2710		.flags = CFTYPE_NOT_ON_ROOT,
2711	},
2712
2713	{
2714		.name = "cpus.subpartitions",
2715		.seq_show = cpuset_common_seq_show,
2716		.private = FILE_SUBPARTS_CPULIST,
2717		.flags = CFTYPE_DEBUG,
2718	},
2719
2720	{ }	/* terminate */
2721};
2722
2723
2724/*
2725 *	cpuset_css_alloc - allocate a cpuset css
2726 *	cgrp:	control group that the new cpuset will be part of
2727 */
2728
2729static struct cgroup_subsys_state *
2730cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2731{
2732	struct cpuset *cs;
2733
2734	if (!parent_css)
2735		return &top_cpuset.css;
2736
2737	cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2738	if (!cs)
2739		return ERR_PTR(-ENOMEM);
2740
2741	if (alloc_cpumasks(cs, NULL)) {
2742		kfree(cs);
2743		return ERR_PTR(-ENOMEM);
2744	}
2745
2746	set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2747	nodes_clear(cs->mems_allowed);
2748	nodes_clear(cs->effective_mems);
2749	fmeter_init(&cs->fmeter);
2750	cs->relax_domain_level = -1;
2751
2752	return &cs->css;
2753}
2754
2755static int cpuset_css_online(struct cgroup_subsys_state *css)
2756{
2757	struct cpuset *cs = css_cs(css);
2758	struct cpuset *parent = parent_cs(cs);
2759	struct cpuset *tmp_cs;
2760	struct cgroup_subsys_state *pos_css;
2761
2762	if (!parent)
2763		return 0;
2764
2765	get_online_cpus();
2766	percpu_down_write(&cpuset_rwsem);
2767
2768	set_bit(CS_ONLINE, &cs->flags);
2769	if (is_spread_page(parent))
2770		set_bit(CS_SPREAD_PAGE, &cs->flags);
2771	if (is_spread_slab(parent))
2772		set_bit(CS_SPREAD_SLAB, &cs->flags);
2773
2774	cpuset_inc();
2775
2776	spin_lock_irq(&callback_lock);
2777	if (is_in_v2_mode()) {
2778		cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2779		cs->effective_mems = parent->effective_mems;
2780		cs->use_parent_ecpus = true;
2781		parent->child_ecpus_count++;
2782	}
2783	spin_unlock_irq(&callback_lock);
2784
2785	if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2786		goto out_unlock;
2787
2788	/*
2789	 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2790	 * set.  This flag handling is implemented in cgroup core for
2791	 * histrical reasons - the flag may be specified during mount.
2792	 *
2793	 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2794	 * refuse to clone the configuration - thereby refusing the task to
2795	 * be entered, and as a result refusing the sys_unshare() or
2796	 * clone() which initiated it.  If this becomes a problem for some
2797	 * users who wish to allow that scenario, then this could be
2798	 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2799	 * (and likewise for mems) to the new cgroup.
2800	 */
2801	rcu_read_lock();
2802	cpuset_for_each_child(tmp_cs, pos_css, parent) {
2803		if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2804			rcu_read_unlock();
2805			goto out_unlock;
2806		}
2807	}
2808	rcu_read_unlock();
2809
2810	spin_lock_irq(&callback_lock);
2811	cs->mems_allowed = parent->mems_allowed;
2812	cs->effective_mems = parent->mems_allowed;
2813	cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2814	cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2815	spin_unlock_irq(&callback_lock);
2816out_unlock:
2817	percpu_up_write(&cpuset_rwsem);
2818	put_online_cpus();
2819	return 0;
2820}
2821
2822/*
2823 * If the cpuset being removed has its flag 'sched_load_balance'
2824 * enabled, then simulate turning sched_load_balance off, which
2825 * will call rebuild_sched_domains_locked(). That is not needed
2826 * in the default hierarchy where only changes in partition
2827 * will cause repartitioning.
2828 *
2829 * If the cpuset has the 'sched.partition' flag enabled, simulate
2830 * turning 'sched.partition" off.
2831 */
2832
2833static void cpuset_css_offline(struct cgroup_subsys_state *css)
2834{
2835	struct cpuset *cs = css_cs(css);
2836
2837	get_online_cpus();
2838	percpu_down_write(&cpuset_rwsem);
2839
2840	if (is_partition_root(cs))
2841		update_prstate(cs, 0);
2842
2843	if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2844	    is_sched_load_balance(cs))
2845		update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2846
2847	if (cs->use_parent_ecpus) {
2848		struct cpuset *parent = parent_cs(cs);
2849
2850		cs->use_parent_ecpus = false;
2851		parent->child_ecpus_count--;
2852	}
2853
2854	cpuset_dec();
2855	clear_bit(CS_ONLINE, &cs->flags);
2856
2857	percpu_up_write(&cpuset_rwsem);
2858	put_online_cpus();
2859}
2860
2861static void cpuset_css_free(struct cgroup_subsys_state *css)
2862{
2863	struct cpuset *cs = css_cs(css);
2864
2865	free_cpuset(cs);
2866}
2867
2868static void cpuset_bind(struct cgroup_subsys_state *root_css)
2869{
2870	percpu_down_write(&cpuset_rwsem);
2871	spin_lock_irq(&callback_lock);
2872
2873	if (is_in_v2_mode()) {
2874		cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2875		top_cpuset.mems_allowed = node_possible_map;
2876	} else {
2877		cpumask_copy(top_cpuset.cpus_allowed,
2878			     top_cpuset.effective_cpus);
2879		top_cpuset.mems_allowed = top_cpuset.effective_mems;
2880	}
2881
2882	spin_unlock_irq(&callback_lock);
2883	percpu_up_write(&cpuset_rwsem);
2884}
2885
2886/*
2887 * Make sure the new task conform to the current state of its parent,
2888 * which could have been changed by cpuset just after it inherits the
2889 * state from the parent and before it sits on the cgroup's task list.
2890 */
2891static void cpuset_fork(struct task_struct *task)
2892{
2893	if (task_css_is_root(task, cpuset_cgrp_id))
2894		return;
2895
2896	set_cpus_allowed_ptr(task, current->cpus_ptr);
2897	task->mems_allowed = current->mems_allowed;
2898}
2899
2900struct cgroup_subsys cpuset_cgrp_subsys = {
2901	.css_alloc	= cpuset_css_alloc,
2902	.css_online	= cpuset_css_online,
2903	.css_offline	= cpuset_css_offline,
2904	.css_free	= cpuset_css_free,
2905	.can_attach	= cpuset_can_attach,
2906	.cancel_attach	= cpuset_cancel_attach,
2907	.attach		= cpuset_attach,
2908	.post_attach	= cpuset_post_attach,
2909	.bind		= cpuset_bind,
2910	.fork		= cpuset_fork,
2911	.legacy_cftypes	= legacy_files,
2912	.dfl_cftypes	= dfl_files,
2913	.early_init	= true,
2914	.threaded	= true,
2915};
2916
2917/**
2918 * cpuset_init - initialize cpusets at system boot
2919 *
2920 * Description: Initialize top_cpuset
2921 **/
2922
2923int __init cpuset_init(void)
2924{
2925	BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2926
2927	BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2928	BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2929	BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2930
2931	cpumask_setall(top_cpuset.cpus_allowed);
2932	nodes_setall(top_cpuset.mems_allowed);
2933	cpumask_setall(top_cpuset.effective_cpus);
2934	nodes_setall(top_cpuset.effective_mems);
2935
2936	fmeter_init(&top_cpuset.fmeter);
2937	set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2938	top_cpuset.relax_domain_level = -1;
2939
2940	BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2941
2942	return 0;
2943}
2944
2945/*
2946 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2947 * or memory nodes, we need to walk over the cpuset hierarchy,
2948 * removing that CPU or node from all cpusets.  If this removes the
2949 * last CPU or node from a cpuset, then move the tasks in the empty
2950 * cpuset to its next-highest non-empty parent.
2951 */
2952static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2953{
2954	struct cpuset *parent;
2955
2956	/*
2957	 * Find its next-highest non-empty parent, (top cpuset
2958	 * has online cpus, so can't be empty).
2959	 */
2960	parent = parent_cs(cs);
2961	while (cpumask_empty(parent->cpus_allowed) ||
2962			nodes_empty(parent->mems_allowed))
2963		parent = parent_cs(parent);
2964
2965	if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2966		pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2967		pr_cont_cgroup_name(cs->css.cgroup);
2968		pr_cont("\n");
2969	}
2970}
2971
2972static void
2973hotplug_update_tasks_legacy(struct cpuset *cs,
2974			    struct cpumask *new_cpus, nodemask_t *new_mems,
2975			    bool cpus_updated, bool mems_updated)
2976{
2977	bool is_empty;
2978
2979	spin_lock_irq(&callback_lock);
2980	cpumask_copy(cs->cpus_allowed, new_cpus);
2981	cpumask_copy(cs->effective_cpus, new_cpus);
2982	cs->mems_allowed = *new_mems;
2983	cs->effective_mems = *new_mems;
2984	spin_unlock_irq(&callback_lock);
2985
2986	/*
2987	 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2988	 * as the tasks will be migratecd to an ancestor.
2989	 */
2990	if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2991		update_tasks_cpumask(cs);
2992	if (mems_updated && !nodes_empty(cs->mems_allowed))
2993		update_tasks_nodemask(cs);
2994
2995	is_empty = cpumask_empty(cs->cpus_allowed) ||
2996		   nodes_empty(cs->mems_allowed);
2997
2998	percpu_up_write(&cpuset_rwsem);
2999
3000	/*
3001	 * Move tasks to the nearest ancestor with execution resources,
3002	 * This is full cgroup operation which will also call back into
3003	 * cpuset. Should be done outside any lock.
3004	 */
3005	if (is_empty)
3006		remove_tasks_in_empty_cpuset(cs);
3007
3008	percpu_down_write(&cpuset_rwsem);
3009}
3010
3011static void
3012hotplug_update_tasks(struct cpuset *cs,
3013		     struct cpumask *new_cpus, nodemask_t *new_mems,
3014		     bool cpus_updated, bool mems_updated)
3015{
3016	if (cpumask_empty(new_cpus))
3017		cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3018	if (nodes_empty(*new_mems))
3019		*new_mems = parent_cs(cs)->effective_mems;
3020
3021	spin_lock_irq(&callback_lock);
3022	cpumask_copy(cs->effective_cpus, new_cpus);
3023	cs->effective_mems = *new_mems;
3024	spin_unlock_irq(&callback_lock);
3025
3026	if (cpus_updated)
3027		update_tasks_cpumask(cs);
3028	if (mems_updated)
3029		update_tasks_nodemask(cs);
3030}
3031
3032static bool force_rebuild;
3033
3034void cpuset_force_rebuild(void)
3035{
3036	force_rebuild = true;
3037}
3038
3039/**
3040 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3041 * @cs: cpuset in interest
3042 * @tmp: the tmpmasks structure pointer
3043 *
3044 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3045 * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
3046 * all its tasks are moved to the nearest ancestor with both resources.
3047 */
3048static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3049{
3050	static cpumask_t new_cpus;
3051	static nodemask_t new_mems;
3052	bool cpus_updated;
3053	bool mems_updated;
3054	struct cpuset *parent;
3055retry:
3056	wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3057
3058	percpu_down_write(&cpuset_rwsem);
3059
3060	/*
3061	 * We have raced with task attaching. We wait until attaching
3062	 * is finished, so we won't attach a task to an empty cpuset.
3063	 */
3064	if (cs->attach_in_progress) {
3065		percpu_up_write(&cpuset_rwsem);
3066		goto retry;
3067	}
3068
3069	parent = parent_cs(cs);
3070	compute_effective_cpumask(&new_cpus, cs, parent);
3071	nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3072
3073	if (cs->nr_subparts_cpus)
3074		/*
3075		 * Make sure that CPUs allocated to child partitions
3076		 * do not show up in effective_cpus.
3077		 */
3078		cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3079
3080	if (!tmp || !cs->partition_root_state)
3081		goto update_tasks;
3082
3083	/*
3084	 * In the unlikely event that a partition root has empty
3085	 * effective_cpus or its parent becomes erroneous, we have to
3086	 * transition it to the erroneous state.
3087	 */
3088	if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3089	   (parent->partition_root_state == PRS_ERROR))) {
3090		if (cs->nr_subparts_cpus) {
3091			spin_lock_irq(&callback_lock);
3092			cs->nr_subparts_cpus = 0;
3093			cpumask_clear(cs->subparts_cpus);
3094			spin_unlock_irq(&callback_lock);
3095			compute_effective_cpumask(&new_cpus, cs, parent);
3096		}
3097
3098		/*
3099		 * If the effective_cpus is empty because the child
3100		 * partitions take away all the CPUs, we can keep
3101		 * the current partition and let the child partitions
3102		 * fight for available CPUs.
3103		 */
3104		if ((parent->partition_root_state == PRS_ERROR) ||
3105		     cpumask_empty(&new_cpus)) {
3106			update_parent_subparts_cpumask(cs, partcmd_disable,
3107						       NULL, tmp);
3108			spin_lock_irq(&callback_lock);
3109			cs->partition_root_state = PRS_ERROR;
3110			spin_unlock_irq(&callback_lock);
3111		}
3112		cpuset_force_rebuild();
3113	}
3114
3115	/*
3116	 * On the other hand, an erroneous partition root may be transitioned
3117	 * back to a regular one or a partition root with no CPU allocated
3118	 * from the parent may change to erroneous.
3119	 */
3120	if (is_partition_root(parent) &&
3121	   ((cs->partition_root_state == PRS_ERROR) ||
3122	    !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3123	     update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3124		cpuset_force_rebuild();
3125
3126update_tasks:
3127	cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3128	mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3129
3130	if (is_in_v2_mode())
3131		hotplug_update_tasks(cs, &new_cpus, &new_mems,
3132				     cpus_updated, mems_updated);
3133	else
3134		hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3135					    cpus_updated, mems_updated);
3136
3137	percpu_up_write(&cpuset_rwsem);
3138}
3139
3140/**
3141 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3142 *
3143 * This function is called after either CPU or memory configuration has
3144 * changed and updates cpuset accordingly.  The top_cpuset is always
3145 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3146 * order to make cpusets transparent (of no affect) on systems that are
3147 * actively using CPU hotplug but making no active use of cpusets.
3148 *
3149 * Non-root cpusets are only affected by offlining.  If any CPUs or memory
3150 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3151 * all descendants.
3152 *
3153 * Note that CPU offlining during suspend is ignored.  We don't modify
3154 * cpusets across suspend/resume cycles at all.
3155 */
3156static void cpuset_hotplug_workfn(struct work_struct *work)
3157{
3158	static cpumask_t new_cpus;
3159	static nodemask_t new_mems;
3160	bool cpus_updated, mems_updated;
3161	bool on_dfl = is_in_v2_mode();
3162	struct tmpmasks tmp, *ptmp = NULL;
3163
3164	if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3165		ptmp = &tmp;
3166
3167	percpu_down_write(&cpuset_rwsem);
3168
3169	/* fetch the available cpus/mems and find out which changed how */
3170	cpumask_copy(&new_cpus, cpu_active_mask);
3171	new_mems = node_states[N_MEMORY];
3172
3173	/*
3174	 * If subparts_cpus is populated, it is likely that the check below
3175	 * will produce a false positive on cpus_updated when the cpu list
3176	 * isn't changed. It is extra work, but it is better to be safe.
3177	 */
3178	cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3179	mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3180
3181	/*
3182	 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3183	 * we assumed that cpus are updated.
3184	 */
3185	if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3186		cpus_updated = true;
3187
3188	/* synchronize cpus_allowed to cpu_active_mask */
3189	if (cpus_updated) {
3190		spin_lock_irq(&callback_lock);
3191		if (!on_dfl)
3192			cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3193		/*
3194		 * Make sure that CPUs allocated to child partitions
3195		 * do not show up in effective_cpus. If no CPU is left,
3196		 * we clear the subparts_cpus & let the child partitions
3197		 * fight for the CPUs again.
3198		 */
3199		if (top_cpuset.nr_subparts_cpus) {
3200			if (cpumask_subset(&new_cpus,
3201					   top_cpuset.subparts_cpus)) {
3202				top_cpuset.nr_subparts_cpus = 0;
3203				cpumask_clear(top_cpuset.subparts_cpus);
3204			} else {
3205				cpumask_andnot(&new_cpus, &new_cpus,
3206					       top_cpuset.subparts_cpus);
3207			}
3208		}
3209		cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3210		spin_unlock_irq(&callback_lock);
3211		/* we don't mess with cpumasks of tasks in top_cpuset */
3212	}
3213
3214	/* synchronize mems_allowed to N_MEMORY */
3215	if (mems_updated) {
3216		spin_lock_irq(&callback_lock);
3217		if (!on_dfl)
3218			top_cpuset.mems_allowed = new_mems;
3219		top_cpuset.effective_mems = new_mems;
3220		spin_unlock_irq(&callback_lock);
3221		update_tasks_nodemask(&top_cpuset);
3222	}
3223
3224	percpu_up_write(&cpuset_rwsem);
3225
3226	/* if cpus or mems changed, we need to propagate to descendants */
3227	if (cpus_updated || mems_updated) {
3228		struct cpuset *cs;
3229		struct cgroup_subsys_state *pos_css;
3230
3231		rcu_read_lock();
3232		cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3233			if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3234				continue;
3235			rcu_read_unlock();
3236
3237			cpuset_hotplug_update_tasks(cs, ptmp);
3238
3239			rcu_read_lock();
3240			css_put(&cs->css);
3241		}
3242		rcu_read_unlock();
3243	}
3244
3245	/* rebuild sched domains if cpus_allowed has changed */
3246	if (cpus_updated || force_rebuild) {
3247		force_rebuild = false;
3248		rebuild_sched_domains();
3249	}
3250
3251	free_cpumasks(NULL, ptmp);
3252}
3253
3254void cpuset_update_active_cpus(void)
3255{
3256	/*
3257	 * We're inside cpu hotplug critical region which usually nests
3258	 * inside cgroup synchronization.  Bounce actual hotplug processing
3259	 * to a work item to avoid reverse locking order.
3260	 */
3261	schedule_work(&cpuset_hotplug_work);
3262}
3263
3264void cpuset_wait_for_hotplug(void)
3265{
3266	flush_work(&cpuset_hotplug_work);
3267}
3268
3269/*
3270 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3271 * Call this routine anytime after node_states[N_MEMORY] changes.
3272 * See cpuset_update_active_cpus() for CPU hotplug handling.
3273 */
3274static int cpuset_track_online_nodes(struct notifier_block *self,
3275				unsigned long action, void *arg)
3276{
3277	schedule_work(&cpuset_hotplug_work);
3278	return NOTIFY_OK;
3279}
3280
3281static struct notifier_block cpuset_track_online_nodes_nb = {
3282	.notifier_call = cpuset_track_online_nodes,
3283	.priority = 10,		/* ??! */
3284};
3285
3286/**
3287 * cpuset_init_smp - initialize cpus_allowed
3288 *
3289 * Description: Finish top cpuset after cpu, node maps are initialized
3290 */
3291void __init cpuset_init_smp(void)
3292{
3293	cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3294	top_cpuset.mems_allowed = node_states[N_MEMORY];
3295	top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3296
3297	cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3298	top_cpuset.effective_mems = node_states[N_MEMORY];
3299
3300	register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3301
3302	cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3303	BUG_ON(!cpuset_migrate_mm_wq);
3304}
3305
3306/**
3307 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3308 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3309 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3310 *
3311 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3312 * attached to the specified @tsk.  Guaranteed to return some non-empty
3313 * subset of cpu_online_mask, even if this means going outside the
3314 * tasks cpuset.
3315 **/
3316
3317void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3318{
3319	unsigned long flags;
3320
3321	spin_lock_irqsave(&callback_lock, flags);
3322	rcu_read_lock();
3323	guarantee_online_cpus(task_cs(tsk), pmask);
3324	rcu_read_unlock();
3325	spin_unlock_irqrestore(&callback_lock, flags);
3326}
3327
3328/**
3329 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3330 * @tsk: pointer to task_struct with which the scheduler is struggling
3331 *
3332 * Description: In the case that the scheduler cannot find an allowed cpu in
3333 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3334 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3335 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3336 * This is the absolute last resort for the scheduler and it is only used if
3337 * _every_ other avenue has been traveled.
3338 **/
3339
3340void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3341{
3342	rcu_read_lock();
3343	do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3344		task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3345	rcu_read_unlock();
3346
3347	/*
3348	 * We own tsk->cpus_allowed, nobody can change it under us.
3349	 *
3350	 * But we used cs && cs->cpus_allowed lockless and thus can
3351	 * race with cgroup_attach_task() or update_cpumask() and get
3352	 * the wrong tsk->cpus_allowed. However, both cases imply the
3353	 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3354	 * which takes task_rq_lock().
3355	 *
3356	 * If we are called after it dropped the lock we must see all
3357	 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3358	 * set any mask even if it is not right from task_cs() pov,
3359	 * the pending set_cpus_allowed_ptr() will fix things.
3360	 *
3361	 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3362	 * if required.
3363	 */
3364}
3365
3366void __init cpuset_init_current_mems_allowed(void)
3367{
3368	nodes_setall(current->mems_allowed);
3369}
3370
3371/**
3372 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3373 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3374 *
3375 * Description: Returns the nodemask_t mems_allowed of the cpuset
3376 * attached to the specified @tsk.  Guaranteed to return some non-empty
3377 * subset of node_states[N_MEMORY], even if this means going outside the
3378 * tasks cpuset.
3379 **/
3380
3381nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3382{
3383	nodemask_t mask;
3384	unsigned long flags;
3385
3386	spin_lock_irqsave(&callback_lock, flags);
3387	rcu_read_lock();
3388	guarantee_online_mems(task_cs(tsk), &mask);
3389	rcu_read_unlock();
3390	spin_unlock_irqrestore(&callback_lock, flags);
3391
3392	return mask;
3393}
3394
3395/**
3396 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3397 * @nodemask: the nodemask to be checked
3398 *
3399 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3400 */
3401int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3402{
3403	return nodes_intersects(*nodemask, current->mems_allowed);
3404}
3405
3406/*
3407 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3408 * mem_hardwall ancestor to the specified cpuset.  Call holding
3409 * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
3410 * (an unusual configuration), then returns the root cpuset.
3411 */
3412static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3413{
3414	while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3415		cs = parent_cs(cs);
3416	return cs;
3417}
3418
3419/**
3420 * cpuset_node_allowed - Can we allocate on a memory node?
3421 * @node: is this an allowed node?
3422 * @gfp_mask: memory allocation flags
3423 *
3424 * If we're in interrupt, yes, we can always allocate.  If @node is set in
3425 * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
3426 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3427 * yes.  If current has access to memory reserves as an oom victim, yes.
3428 * Otherwise, no.
3429 *
3430 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3431 * and do not allow allocations outside the current tasks cpuset
3432 * unless the task has been OOM killed.
3433 * GFP_KERNEL allocations are not so marked, so can escape to the
3434 * nearest enclosing hardwalled ancestor cpuset.
3435 *
3436 * Scanning up parent cpusets requires callback_lock.  The
3437 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3438 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3439 * current tasks mems_allowed came up empty on the first pass over
3440 * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
3441 * cpuset are short of memory, might require taking the callback_lock.
3442 *
3443 * The first call here from mm/page_alloc:get_page_from_freelist()
3444 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3445 * so no allocation on a node outside the cpuset is allowed (unless
3446 * in interrupt, of course).
3447 *
3448 * The second pass through get_page_from_freelist() doesn't even call
3449 * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
3450 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3451 * in alloc_flags.  That logic and the checks below have the combined
3452 * affect that:
3453 *	in_interrupt - any node ok (current task context irrelevant)
3454 *	GFP_ATOMIC   - any node ok
3455 *	tsk_is_oom_victim   - any node ok
3456 *	GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
3457 *	GFP_USER     - only nodes in current tasks mems allowed ok.
3458 */
3459bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3460{
3461	struct cpuset *cs;		/* current cpuset ancestors */
3462	int allowed;			/* is allocation in zone z allowed? */
3463	unsigned long flags;
3464
3465	if (in_interrupt())
3466		return true;
3467	if (node_isset(node, current->mems_allowed))
3468		return true;
3469	/*
3470	 * Allow tasks that have access to memory reserves because they have
3471	 * been OOM killed to get memory anywhere.
3472	 */
3473	if (unlikely(tsk_is_oom_victim(current)))
3474		return true;
3475	if (gfp_mask & __GFP_HARDWALL)	/* If hardwall request, stop here */
3476		return false;
3477
3478	if (current->flags & PF_EXITING) /* Let dying task have memory */
3479		return true;
3480
3481	/* Not hardwall and node outside mems_allowed: scan up cpusets */
3482	spin_lock_irqsave(&callback_lock, flags);
3483
3484	rcu_read_lock();
3485	cs = nearest_hardwall_ancestor(task_cs(current));
3486	allowed = node_isset(node, cs->mems_allowed);
3487	rcu_read_unlock();
3488
3489	spin_unlock_irqrestore(&callback_lock, flags);
3490	return allowed;
3491}
3492
3493/**
3494 * cpuset_mem_spread_node() - On which node to begin search for a file page
3495 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3496 *
3497 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3498 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3499 * and if the memory allocation used cpuset_mem_spread_node()
3500 * to determine on which node to start looking, as it will for
3501 * certain page cache or slab cache pages such as used for file
3502 * system buffers and inode caches, then instead of starting on the
3503 * local node to look for a free page, rather spread the starting
3504 * node around the tasks mems_allowed nodes.
3505 *
3506 * We don't have to worry about the returned node being offline
3507 * because "it can't happen", and even if it did, it would be ok.
3508 *
3509 * The routines calling guarantee_online_mems() are careful to
3510 * only set nodes in task->mems_allowed that are online.  So it
3511 * should not be possible for the following code to return an
3512 * offline node.  But if it did, that would be ok, as this routine
3513 * is not returning the node where the allocation must be, only
3514 * the node where the search should start.  The zonelist passed to
3515 * __alloc_pages() will include all nodes.  If the slab allocator
3516 * is passed an offline node, it will fall back to the local node.
3517 * See kmem_cache_alloc_node().
3518 */
3519
3520static int cpuset_spread_node(int *rotor)
3521{
3522	return *rotor = next_node_in(*rotor, current->mems_allowed);
3523}
3524
3525int cpuset_mem_spread_node(void)
3526{
3527	if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3528		current->cpuset_mem_spread_rotor =
3529			node_random(&current->mems_allowed);
3530
3531	return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
3532}
3533
3534int cpuset_slab_spread_node(void)
3535{
3536	if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3537		current->cpuset_slab_spread_rotor =
3538			node_random(&current->mems_allowed);
3539
3540	return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
3541}
3542
3543EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3544
3545/**
3546 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3547 * @tsk1: pointer to task_struct of some task.
3548 * @tsk2: pointer to task_struct of some other task.
3549 *
3550 * Description: Return true if @tsk1's mems_allowed intersects the
3551 * mems_allowed of @tsk2.  Used by the OOM killer to determine if
3552 * one of the task's memory usage might impact the memory available
3553 * to the other.
3554 **/
3555
3556int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3557				   const struct task_struct *tsk2)
3558{
3559	return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3560}
3561
3562/**
3563 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3564 *
3565 * Description: Prints current's name, cpuset name, and cached copy of its
3566 * mems_allowed to the kernel log.
3567 */
3568void cpuset_print_current_mems_allowed(void)
3569{
3570	struct cgroup *cgrp;
3571
3572	rcu_read_lock();
3573
3574	cgrp = task_cs(current)->css.cgroup;
3575	pr_cont(",cpuset=");
3576	pr_cont_cgroup_name(cgrp);
3577	pr_cont(",mems_allowed=%*pbl",
3578		nodemask_pr_args(&current->mems_allowed));
3579
3580	rcu_read_unlock();
3581}
3582
3583/*
3584 * Collection of memory_pressure is suppressed unless
3585 * this flag is enabled by writing "1" to the special
3586 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3587 */
3588
3589int cpuset_memory_pressure_enabled __read_mostly;
3590
3591/**
3592 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3593 *
3594 * Keep a running average of the rate of synchronous (direct)
3595 * page reclaim efforts initiated by tasks in each cpuset.
3596 *
3597 * This represents the rate at which some task in the cpuset
3598 * ran low on memory on all nodes it was allowed to use, and
3599 * had to enter the kernels page reclaim code in an effort to
3600 * create more free memory by tossing clean pages or swapping
3601 * or writing dirty pages.
3602 *
3603 * Display to user space in the per-cpuset read-only file
3604 * "memory_pressure".  Value displayed is an integer
3605 * representing the recent rate of entry into the synchronous
3606 * (direct) page reclaim by any task attached to the cpuset.
3607 **/
3608
3609void __cpuset_memory_pressure_bump(void)
3610{
3611	rcu_read_lock();
3612	fmeter_markevent(&task_cs(current)->fmeter);
3613	rcu_read_unlock();
3614}
3615
3616#ifdef CONFIG_PROC_PID_CPUSET
3617/*
3618 * proc_cpuset_show()
3619 *  - Print tasks cpuset path into seq_file.
3620 *  - Used for /proc/<pid>/cpuset.
3621 *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3622 *    doesn't really matter if tsk->cpuset changes after we read it,
3623 *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
3624 *    anyway.
3625 */
3626int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3627		     struct pid *pid, struct task_struct *tsk)
3628{
3629	char *buf;
3630	struct cgroup_subsys_state *css;
3631	int retval;
3632
3633	retval = -ENOMEM;
3634	buf = kmalloc(PATH_MAX, GFP_KERNEL);
3635	if (!buf)
3636		goto out;
3637
3638	css = task_get_css(tsk, cpuset_cgrp_id);
3639	retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3640				current->nsproxy->cgroup_ns);
3641	css_put(css);
3642	if (retval >= PATH_MAX)
3643		retval = -ENAMETOOLONG;
3644	if (retval < 0)
3645		goto out_free;
3646	seq_puts(m, buf);
3647	seq_putc(m, '\n');
3648	retval = 0;
3649out_free:
3650	kfree(buf);
3651out:
3652	return retval;
3653}
3654#endif /* CONFIG_PROC_PID_CPUSET */
3655
3656/* Display task mems_allowed in /proc/<pid>/status file. */
3657void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3658{
3659	seq_printf(m, "Mems_allowed:\t%*pb\n",
3660		   nodemask_pr_args(&task->mems_allowed));
3661	seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3662		   nodemask_pr_args(&task->mems_allowed));
3663}