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 hierachy:
 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 when on default hierarchy or the
 362 * cgroup_v2_mode flag is set.
 363 */
 364static inline bool is_in_v2_mode(void)
 365{
 366	return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
 367	      (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
 368}
 369
 370/*
 371 * Return in pmask the portion of a cpusets's cpus_allowed that
 372 * are online.  If none are online, walk up the cpuset hierarchy
 373 * until we find one that does have some online cpus.
 374 *
 375 * One way or another, we guarantee to return some non-empty subset
 376 * of cpu_online_mask.
 377 *
 378 * Call with callback_lock or cpuset_mutex held.
 379 */
 380static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
 381{
 382	while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
 383		cs = parent_cs(cs);
 384		if (unlikely(!cs)) {
 385			/*
 386			 * The top cpuset doesn't have any online cpu as a
 387			 * consequence of a race between cpuset_hotplug_work
 388			 * and cpu hotplug notifier.  But we know the top
 389			 * cpuset's effective_cpus is on its way to to be
 390			 * identical to cpu_online_mask.
 391			 */
 392			cpumask_copy(pmask, cpu_online_mask);
 393			return;
 394		}
 395	}
 396	cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
 397}
 398
 399/*
 400 * Return in *pmask the portion of a cpusets's mems_allowed that
 401 * are online, with memory.  If none are online with memory, walk
 402 * up the cpuset hierarchy until we find one that does have some
 403 * online mems.  The top cpuset always has some mems online.
 404 *
 405 * One way or another, we guarantee to return some non-empty subset
 406 * of node_states[N_MEMORY].
 407 *
 408 * Call with callback_lock or cpuset_mutex held.
 409 */
 410static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
 411{
 412	while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
 413		cs = parent_cs(cs);
 414	nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
 415}
 416
 417/*
 418 * update task's spread flag if cpuset's page/slab spread flag is set
 419 *
 420 * Call with callback_lock or cpuset_mutex held.
 421 */
 422static void cpuset_update_task_spread_flag(struct cpuset *cs,
 423					struct task_struct *tsk)
 424{
 425	if (is_spread_page(cs))
 426		task_set_spread_page(tsk);
 427	else
 428		task_clear_spread_page(tsk);
 429
 430	if (is_spread_slab(cs))
 431		task_set_spread_slab(tsk);
 432	else
 433		task_clear_spread_slab(tsk);
 434}
 435
 436/*
 437 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
 438 *
 439 * One cpuset is a subset of another if all its allowed CPUs and
 440 * Memory Nodes are a subset of the other, and its exclusive flags
 441 * are only set if the other's are set.  Call holding cpuset_mutex.
 442 */
 443
 444static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
 445{
 446	return	cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
 447		nodes_subset(p->mems_allowed, q->mems_allowed) &&
 448		is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
 449		is_mem_exclusive(p) <= is_mem_exclusive(q);
 450}
 451
 452/**
 453 * alloc_cpumasks - allocate three cpumasks for cpuset
 454 * @cs:  the cpuset that have cpumasks to be allocated.
 455 * @tmp: the tmpmasks structure pointer
 456 * Return: 0 if successful, -ENOMEM otherwise.
 457 *
 458 * Only one of the two input arguments should be non-NULL.
 459 */
 460static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
 461{
 462	cpumask_var_t *pmask1, *pmask2, *pmask3;
 463
 464	if (cs) {
 465		pmask1 = &cs->cpus_allowed;
 466		pmask2 = &cs->effective_cpus;
 467		pmask3 = &cs->subparts_cpus;
 468	} else {
 469		pmask1 = &tmp->new_cpus;
 470		pmask2 = &tmp->addmask;
 471		pmask3 = &tmp->delmask;
 472	}
 473
 474	if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
 475		return -ENOMEM;
 476
 477	if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
 478		goto free_one;
 479
 480	if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
 481		goto free_two;
 482
 483	return 0;
 484
 485free_two:
 486	free_cpumask_var(*pmask2);
 487free_one:
 488	free_cpumask_var(*pmask1);
 489	return -ENOMEM;
 490}
 491
 492/**
 493 * free_cpumasks - free cpumasks in a tmpmasks structure
 494 * @cs:  the cpuset that have cpumasks to be free.
 495 * @tmp: the tmpmasks structure pointer
 496 */
 497static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
 498{
 499	if (cs) {
 500		free_cpumask_var(cs->cpus_allowed);
 501		free_cpumask_var(cs->effective_cpus);
 502		free_cpumask_var(cs->subparts_cpus);
 503	}
 504	if (tmp) {
 505		free_cpumask_var(tmp->new_cpus);
 506		free_cpumask_var(tmp->addmask);
 507		free_cpumask_var(tmp->delmask);
 508	}
 509}
 510
 511/**
 512 * alloc_trial_cpuset - allocate a trial cpuset
 513 * @cs: the cpuset that the trial cpuset duplicates
 514 */
 515static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
 516{
 517	struct cpuset *trial;
 518
 519	trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
 520	if (!trial)
 521		return NULL;
 522
 523	if (alloc_cpumasks(trial, NULL)) {
 524		kfree(trial);
 525		return NULL;
 526	}
 527
 528	cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
 529	cpumask_copy(trial->effective_cpus, cs->effective_cpus);
 530	return trial;
 531}
 532
 533/**
 534 * free_cpuset - free the cpuset
 535 * @cs: the cpuset to be freed
 536 */
 537static inline void free_cpuset(struct cpuset *cs)
 538{
 539	free_cpumasks(cs, NULL);
 540	kfree(cs);
 541}
 542
 543/*
 544 * validate_change() - Used to validate that any proposed cpuset change
 545 *		       follows the structural rules for cpusets.
 546 *
 547 * If we replaced the flag and mask values of the current cpuset
 548 * (cur) with those values in the trial cpuset (trial), would
 549 * our various subset and exclusive rules still be valid?  Presumes
 550 * cpuset_mutex held.
 551 *
 552 * 'cur' is the address of an actual, in-use cpuset.  Operations
 553 * such as list traversal that depend on the actual address of the
 554 * cpuset in the list must use cur below, not trial.
 555 *
 556 * 'trial' is the address of bulk structure copy of cur, with
 557 * perhaps one or more of the fields cpus_allowed, mems_allowed,
 558 * or flags changed to new, trial values.
 559 *
 560 * Return 0 if valid, -errno if not.
 561 */
 562
 563static int validate_change(struct cpuset *cur, struct cpuset *trial)
 564{
 565	struct cgroup_subsys_state *css;
 566	struct cpuset *c, *par;
 567	int ret;
 568
 569	rcu_read_lock();
 570
 571	/* Each of our child cpusets must be a subset of us */
 572	ret = -EBUSY;
 573	cpuset_for_each_child(c, css, cur)
 574		if (!is_cpuset_subset(c, trial))
 575			goto out;
 576
 577	/* Remaining checks don't apply to root cpuset */
 578	ret = 0;
 579	if (cur == &top_cpuset)
 580		goto out;
 581
 582	par = parent_cs(cur);
 583
 584	/* On legacy hiearchy, we must be a subset of our parent cpuset. */
 585	ret = -EACCES;
 586	if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
 587		goto out;
 588
 589	/*
 590	 * If either I or some sibling (!= me) is exclusive, we can't
 591	 * overlap
 592	 */
 593	ret = -EINVAL;
 594	cpuset_for_each_child(c, css, par) {
 595		if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
 596		    c != cur &&
 597		    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
 598			goto out;
 599		if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
 600		    c != cur &&
 601		    nodes_intersects(trial->mems_allowed, c->mems_allowed))
 602			goto out;
 603	}
 604
 605	/*
 606	 * Cpusets with tasks - existing or newly being attached - can't
 607	 * be changed to have empty cpus_allowed or mems_allowed.
 608	 */
 609	ret = -ENOSPC;
 610	if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
 611		if (!cpumask_empty(cur->cpus_allowed) &&
 612		    cpumask_empty(trial->cpus_allowed))
 613			goto out;
 614		if (!nodes_empty(cur->mems_allowed) &&
 615		    nodes_empty(trial->mems_allowed))
 616			goto out;
 617	}
 618
 619	/*
 620	 * We can't shrink if we won't have enough room for SCHED_DEADLINE
 621	 * tasks.
 622	 */
 623	ret = -EBUSY;
 624	if (is_cpu_exclusive(cur) &&
 625	    !cpuset_cpumask_can_shrink(cur->cpus_allowed,
 626				       trial->cpus_allowed))
 627		goto out;
 628
 629	ret = 0;
 630out:
 631	rcu_read_unlock();
 632	return ret;
 633}
 634
 635#ifdef CONFIG_SMP
 636/*
 637 * Helper routine for generate_sched_domains().
 638 * Do cpusets a, b have overlapping effective cpus_allowed masks?
 639 */
 640static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
 641{
 642	return cpumask_intersects(a->effective_cpus, b->effective_cpus);
 643}
 644
 645static void
 646update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
 647{
 648	if (dattr->relax_domain_level < c->relax_domain_level)
 649		dattr->relax_domain_level = c->relax_domain_level;
 650	return;
 651}
 652
 653static void update_domain_attr_tree(struct sched_domain_attr *dattr,
 654				    struct cpuset *root_cs)
 655{
 656	struct cpuset *cp;
 657	struct cgroup_subsys_state *pos_css;
 658
 659	rcu_read_lock();
 660	cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
 661		/* skip the whole subtree if @cp doesn't have any CPU */
 662		if (cpumask_empty(cp->cpus_allowed)) {
 663			pos_css = css_rightmost_descendant(pos_css);
 664			continue;
 665		}
 666
 667		if (is_sched_load_balance(cp))
 668			update_domain_attr(dattr, cp);
 669	}
 670	rcu_read_unlock();
 671}
 672
 673/* Must be called with cpuset_mutex held.  */
 674static inline int nr_cpusets(void)
 675{
 676	/* jump label reference count + the top-level cpuset */
 677	return static_key_count(&cpusets_enabled_key.key) + 1;
 678}
 679
 680/*
 681 * generate_sched_domains()
 682 *
 683 * This function builds a partial partition of the systems CPUs
 684 * A 'partial partition' is a set of non-overlapping subsets whose
 685 * union is a subset of that set.
 686 * The output of this function needs to be passed to kernel/sched/core.c
 687 * partition_sched_domains() routine, which will rebuild the scheduler's
 688 * load balancing domains (sched domains) as specified by that partial
 689 * partition.
 690 *
 691 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
 692 * for a background explanation of this.
 693 *
 694 * Does not return errors, on the theory that the callers of this
 695 * routine would rather not worry about failures to rebuild sched
 696 * domains when operating in the severe memory shortage situations
 697 * that could cause allocation failures below.
 698 *
 699 * Must be called with cpuset_mutex held.
 700 *
 701 * The three key local variables below are:
 702 *    cp - cpuset pointer, used (together with pos_css) to perform a
 703 *	   top-down scan of all cpusets. For our purposes, rebuilding
 704 *	   the schedulers sched domains, we can ignore !is_sched_load_
 705 *	   balance cpusets.
 706 *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
 707 *	   that need to be load balanced, for convenient iterative
 708 *	   access by the subsequent code that finds the best partition,
 709 *	   i.e the set of domains (subsets) of CPUs such that the
 710 *	   cpus_allowed of every cpuset marked is_sched_load_balance
 711 *	   is a subset of one of these domains, while there are as
 712 *	   many such domains as possible, each as small as possible.
 713 * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
 714 *	   the kernel/sched/core.c routine partition_sched_domains() in a
 715 *	   convenient format, that can be easily compared to the prior
 716 *	   value to determine what partition elements (sched domains)
 717 *	   were changed (added or removed.)
 718 *
 719 * Finding the best partition (set of domains):
 720 *	The triple nested loops below over i, j, k scan over the
 721 *	load balanced cpusets (using the array of cpuset pointers in
 722 *	csa[]) looking for pairs of cpusets that have overlapping
 723 *	cpus_allowed, but which don't have the same 'pn' partition
 724 *	number and gives them in the same partition number.  It keeps
 725 *	looping on the 'restart' label until it can no longer find
 726 *	any such pairs.
 727 *
 728 *	The union of the cpus_allowed masks from the set of
 729 *	all cpusets having the same 'pn' value then form the one
 730 *	element of the partition (one sched domain) to be passed to
 731 *	partition_sched_domains().
 732 */
 733static int generate_sched_domains(cpumask_var_t **domains,
 734			struct sched_domain_attr **attributes)
 735{
 736	struct cpuset *cp;	/* top-down scan of cpusets */
 737	struct cpuset **csa;	/* array of all cpuset ptrs */
 738	int csn;		/* how many cpuset ptrs in csa so far */
 739	int i, j, k;		/* indices for partition finding loops */
 740	cpumask_var_t *doms;	/* resulting partition; i.e. sched domains */
 741	struct sched_domain_attr *dattr;  /* attributes for custom domains */
 742	int ndoms = 0;		/* number of sched domains in result */
 743	int nslot;		/* next empty doms[] struct cpumask slot */
 744	struct cgroup_subsys_state *pos_css;
 745	bool root_load_balance = is_sched_load_balance(&top_cpuset);
 746
 747	doms = NULL;
 748	dattr = NULL;
 749	csa = NULL;
 750
 751	/* Special case for the 99% of systems with one, full, sched domain */
 752	if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
 753		ndoms = 1;
 754		doms = alloc_sched_domains(ndoms);
 755		if (!doms)
 756			goto done;
 757
 758		dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
 759		if (dattr) {
 760			*dattr = SD_ATTR_INIT;
 761			update_domain_attr_tree(dattr, &top_cpuset);
 762		}
 763		cpumask_and(doms[0], top_cpuset.effective_cpus,
 764			    housekeeping_cpumask(HK_FLAG_DOMAIN));
 765
 766		goto done;
 767	}
 768
 769	csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
 770	if (!csa)
 771		goto done;
 772	csn = 0;
 773
 774	rcu_read_lock();
 775	if (root_load_balance)
 776		csa[csn++] = &top_cpuset;
 777	cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
 778		if (cp == &top_cpuset)
 779			continue;
 780		/*
 781		 * Continue traversing beyond @cp iff @cp has some CPUs and
 782		 * isn't load balancing.  The former is obvious.  The
 783		 * latter: All child cpusets contain a subset of the
 784		 * parent's cpus, so just skip them, and then we call
 785		 * update_domain_attr_tree() to calc relax_domain_level of
 786		 * the corresponding sched domain.
 787		 *
 788		 * If root is load-balancing, we can skip @cp if it
 789		 * is a subset of the root's effective_cpus.
 790		 */
 791		if (!cpumask_empty(cp->cpus_allowed) &&
 792		    !(is_sched_load_balance(cp) &&
 793		      cpumask_intersects(cp->cpus_allowed,
 794					 housekeeping_cpumask(HK_FLAG_DOMAIN))))
 795			continue;
 796
 797		if (root_load_balance &&
 798		    cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
 799			continue;
 800
 801		if (is_sched_load_balance(cp) &&
 802		    !cpumask_empty(cp->effective_cpus))
 803			csa[csn++] = cp;
 804
 805		/* skip @cp's subtree if not a partition root */
 806		if (!is_partition_root(cp))
 807			pos_css = css_rightmost_descendant(pos_css);
 808	}
 809	rcu_read_unlock();
 810
 811	for (i = 0; i < csn; i++)
 812		csa[i]->pn = i;
 813	ndoms = csn;
 814
 815restart:
 816	/* Find the best partition (set of sched domains) */
 817	for (i = 0; i < csn; i++) {
 818		struct cpuset *a = csa[i];
 819		int apn = a->pn;
 820
 821		for (j = 0; j < csn; j++) {
 822			struct cpuset *b = csa[j];
 823			int bpn = b->pn;
 824
 825			if (apn != bpn && cpusets_overlap(a, b)) {
 826				for (k = 0; k < csn; k++) {
 827					struct cpuset *c = csa[k];
 828
 829					if (c->pn == bpn)
 830						c->pn = apn;
 831				}
 832				ndoms--;	/* one less element */
 833				goto restart;
 834			}
 835		}
 836	}
 837
 838	/*
 839	 * Now we know how many domains to create.
 840	 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
 841	 */
 842	doms = alloc_sched_domains(ndoms);
 843	if (!doms)
 844		goto done;
 845
 846	/*
 847	 * The rest of the code, including the scheduler, can deal with
 848	 * dattr==NULL case. No need to abort if alloc fails.
 849	 */
 850	dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
 851			      GFP_KERNEL);
 852
 853	for (nslot = 0, i = 0; i < csn; i++) {
 854		struct cpuset *a = csa[i];
 855		struct cpumask *dp;
 856		int apn = a->pn;
 857
 858		if (apn < 0) {
 859			/* Skip completed partitions */
 860			continue;
 861		}
 862
 863		dp = doms[nslot];
 864
 865		if (nslot == ndoms) {
 866			static int warnings = 10;
 867			if (warnings) {
 868				pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
 869					nslot, ndoms, csn, i, apn);
 870				warnings--;
 871			}
 872			continue;
 873		}
 874
 875		cpumask_clear(dp);
 876		if (dattr)
 877			*(dattr + nslot) = SD_ATTR_INIT;
 878		for (j = i; j < csn; j++) {
 879			struct cpuset *b = csa[j];
 880
 881			if (apn == b->pn) {
 882				cpumask_or(dp, dp, b->effective_cpus);
 883				cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
 884				if (dattr)
 885					update_domain_attr_tree(dattr + nslot, b);
 886
 887				/* Done with this partition */
 888				b->pn = -1;
 889			}
 890		}
 891		nslot++;
 892	}
 893	BUG_ON(nslot != ndoms);
 894
 895done:
 896	kfree(csa);
 897
 898	/*
 899	 * Fallback to the default domain if kmalloc() failed.
 900	 * See comments in partition_sched_domains().
 901	 */
 902	if (doms == NULL)
 903		ndoms = 1;
 904
 905	*domains    = doms;
 906	*attributes = dattr;
 907	return ndoms;
 908}
 909
 910static void update_tasks_root_domain(struct cpuset *cs)
 911{
 912	struct css_task_iter it;
 913	struct task_struct *task;
 914
 915	css_task_iter_start(&cs->css, 0, &it);
 916
 917	while ((task = css_task_iter_next(&it)))
 918		dl_add_task_root_domain(task);
 919
 920	css_task_iter_end(&it);
 921}
 922
 923static void rebuild_root_domains(void)
 924{
 925	struct cpuset *cs = NULL;
 926	struct cgroup_subsys_state *pos_css;
 927
 928	percpu_rwsem_assert_held(&cpuset_rwsem);
 929	lockdep_assert_cpus_held();
 930	lockdep_assert_held(&sched_domains_mutex);
 931
 932	cgroup_enable_task_cg_lists();
 933
 934	rcu_read_lock();
 935
 936	/*
 937	 * Clear default root domain DL accounting, it will be computed again
 938	 * if a task belongs to it.
 939	 */
 940	dl_clear_root_domain(&def_root_domain);
 941
 942	cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
 943
 944		if (cpumask_empty(cs->effective_cpus)) {
 945			pos_css = css_rightmost_descendant(pos_css);
 946			continue;
 947		}
 948
 949		css_get(&cs->css);
 950
 951		rcu_read_unlock();
 952
 953		update_tasks_root_domain(cs);
 954
 955		rcu_read_lock();
 956		css_put(&cs->css);
 957	}
 958	rcu_read_unlock();
 959}
 960
 961static void
 962partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
 963				    struct sched_domain_attr *dattr_new)
 964{
 965	mutex_lock(&sched_domains_mutex);
 966	partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
 967	rebuild_root_domains();
 968	mutex_unlock(&sched_domains_mutex);
 969}
 970
 971/*
 972 * Rebuild scheduler domains.
 973 *
 974 * If the flag 'sched_load_balance' of any cpuset with non-empty
 975 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
 976 * which has that flag enabled, or if any cpuset with a non-empty
 977 * 'cpus' is removed, then call this routine to rebuild the
 978 * scheduler's dynamic sched domains.
 979 *
 980 * Call with cpuset_mutex held.  Takes get_online_cpus().
 981 */
 982static void rebuild_sched_domains_locked(void)
 983{
 984	struct sched_domain_attr *attr;
 985	cpumask_var_t *doms;
 986	int ndoms;
 987
 988	lockdep_assert_cpus_held();
 989	percpu_rwsem_assert_held(&cpuset_rwsem);
 990
 991	/*
 992	 * We have raced with CPU hotplug. Don't do anything to avoid
 993	 * passing doms with offlined cpu to partition_sched_domains().
 994	 * Anyways, hotplug work item will rebuild sched domains.
 995	 */
 996	if (!top_cpuset.nr_subparts_cpus &&
 997	    !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
 998		return;
 999
1000	if (top_cpuset.nr_subparts_cpus &&
1001	   !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask))
1002		return;
1003
1004	/* Generate domain masks and attrs */
1005	ndoms = generate_sched_domains(&doms, &attr);
1006
1007	/* Have scheduler rebuild the domains */
1008	partition_and_rebuild_sched_domains(ndoms, doms, attr);
1009}
1010#else /* !CONFIG_SMP */
1011static void rebuild_sched_domains_locked(void)
1012{
1013}
1014#endif /* CONFIG_SMP */
1015
1016void rebuild_sched_domains(void)
1017{
1018	get_online_cpus();
1019	percpu_down_write(&cpuset_rwsem);
1020	rebuild_sched_domains_locked();
1021	percpu_up_write(&cpuset_rwsem);
1022	put_online_cpus();
1023}
1024
1025/**
1026 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1027 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1028 *
1029 * Iterate through each task of @cs updating its cpus_allowed to the
1030 * effective cpuset's.  As this function is called with cpuset_mutex held,
1031 * cpuset membership stays stable.
1032 */
1033static void update_tasks_cpumask(struct cpuset *cs)
1034{
1035	struct css_task_iter it;
1036	struct task_struct *task;
1037
1038	css_task_iter_start(&cs->css, 0, &it);
1039	while ((task = css_task_iter_next(&it)))
1040		set_cpus_allowed_ptr(task, cs->effective_cpus);
1041	css_task_iter_end(&it);
1042}
1043
1044/**
1045 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1046 * @new_cpus: the temp variable for the new effective_cpus mask
1047 * @cs: the cpuset the need to recompute the new effective_cpus mask
1048 * @parent: the parent cpuset
1049 *
1050 * If the parent has subpartition CPUs, include them in the list of
1051 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1052 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1053 * to mask those out.
1054 */
1055static void compute_effective_cpumask(struct cpumask *new_cpus,
1056				      struct cpuset *cs, struct cpuset *parent)
1057{
1058	if (parent->nr_subparts_cpus) {
1059		cpumask_or(new_cpus, parent->effective_cpus,
1060			   parent->subparts_cpus);
1061		cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1062		cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1063	} else {
1064		cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1065	}
1066}
1067
1068/*
1069 * Commands for update_parent_subparts_cpumask
1070 */
1071enum subparts_cmd {
1072	partcmd_enable,		/* Enable partition root	 */
1073	partcmd_disable,	/* Disable partition root	 */
1074	partcmd_update,		/* Update parent's subparts_cpus */
1075};
1076
1077/**
1078 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1079 * @cpuset:  The cpuset that requests change in partition root state
1080 * @cmd:     Partition root state change command
1081 * @newmask: Optional new cpumask for partcmd_update
1082 * @tmp:     Temporary addmask and delmask
1083 * Return:   0, 1 or an error code
1084 *
1085 * For partcmd_enable, the cpuset is being transformed from a non-partition
1086 * root to a partition root. The cpus_allowed mask of the given cpuset will
1087 * be put into parent's subparts_cpus and taken away from parent's
1088 * effective_cpus. The function will return 0 if all the CPUs listed in
1089 * cpus_allowed can be granted or an error code will be returned.
1090 *
1091 * For partcmd_disable, the cpuset is being transofrmed from a partition
1092 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1093 * parent's subparts_cpus will be taken away from that cpumask and put back
1094 * into parent's effective_cpus. 0 should always be returned.
1095 *
1096 * For partcmd_update, if the optional newmask is specified, the cpu
1097 * list is to be changed from cpus_allowed to newmask. Otherwise,
1098 * cpus_allowed is assumed to remain the same. The cpuset should either
1099 * be a partition root or an invalid partition root. The partition root
1100 * state may change if newmask is NULL and none of the requested CPUs can
1101 * be granted by the parent. The function will return 1 if changes to
1102 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1103 * Error code should only be returned when newmask is non-NULL.
1104 *
1105 * The partcmd_enable and partcmd_disable commands are used by
1106 * update_prstate(). The partcmd_update command is used by
1107 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1108 * newmask set.
1109 *
1110 * The checking is more strict when enabling partition root than the
1111 * other two commands.
1112 *
1113 * Because of the implicit cpu exclusive nature of a partition root,
1114 * cpumask changes that violates the cpu exclusivity rule will not be
1115 * permitted when checked by validate_change(). The validate_change()
1116 * function will also prevent any changes to the cpu list if it is not
1117 * a superset of children's cpu lists.
1118 */
1119static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1120					  struct cpumask *newmask,
1121					  struct tmpmasks *tmp)
1122{
1123	struct cpuset *parent = parent_cs(cpuset);
1124	int adding;	/* Moving cpus from effective_cpus to subparts_cpus */
1125	int deleting;	/* Moving cpus from subparts_cpus to effective_cpus */
1126	bool part_error = false;	/* Partition error? */
1127
1128	percpu_rwsem_assert_held(&cpuset_rwsem);
1129
1130	/*
1131	 * The parent must be a partition root.
1132	 * The new cpumask, if present, or the current cpus_allowed must
1133	 * not be empty.
1134	 */
1135	if (!is_partition_root(parent) ||
1136	   (newmask && cpumask_empty(newmask)) ||
1137	   (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1138		return -EINVAL;
1139
1140	/*
1141	 * Enabling/disabling partition root is not allowed if there are
1142	 * online children.
1143	 */
1144	if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1145		return -EBUSY;
1146
1147	/*
1148	 * Enabling partition root is not allowed if not all the CPUs
1149	 * can be granted from parent's effective_cpus or at least one
1150	 * CPU will be left after that.
1151	 */
1152	if ((cmd == partcmd_enable) &&
1153	   (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1154	     cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1155		return -EINVAL;
1156
1157	/*
1158	 * A cpumask update cannot make parent's effective_cpus become empty.
1159	 */
1160	adding = deleting = false;
1161	if (cmd == partcmd_enable) {
1162		cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1163		adding = true;
1164	} else if (cmd == partcmd_disable) {
1165		deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1166				       parent->subparts_cpus);
1167	} else if (newmask) {
1168		/*
1169		 * partcmd_update with newmask:
1170		 *
1171		 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1172		 * addmask = newmask & parent->effective_cpus
1173		 *		     & ~parent->subparts_cpus
1174		 */
1175		cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1176		deleting = cpumask_and(tmp->delmask, tmp->delmask,
1177				       parent->subparts_cpus);
1178
1179		cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1180		adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1181					parent->subparts_cpus);
1182		/*
1183		 * Return error if the new effective_cpus could become empty.
1184		 */
1185		if (adding &&
1186		    cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1187			if (!deleting)
1188				return -EINVAL;
1189			/*
1190			 * As some of the CPUs in subparts_cpus might have
1191			 * been offlined, we need to compute the real delmask
1192			 * to confirm that.
1193			 */
1194			if (!cpumask_and(tmp->addmask, tmp->delmask,
1195					 cpu_active_mask))
1196				return -EINVAL;
1197			cpumask_copy(tmp->addmask, parent->effective_cpus);
1198		}
1199	} else {
1200		/*
1201		 * partcmd_update w/o newmask:
1202		 *
1203		 * addmask = cpus_allowed & parent->effectiveb_cpus
1204		 *
1205		 * Note that parent's subparts_cpus may have been
1206		 * pre-shrunk in case there is a change in the cpu list.
1207		 * So no deletion is needed.
1208		 */
1209		adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1210				     parent->effective_cpus);
1211		part_error = cpumask_equal(tmp->addmask,
1212					   parent->effective_cpus);
1213	}
1214
1215	if (cmd == partcmd_update) {
1216		int prev_prs = cpuset->partition_root_state;
1217
1218		/*
1219		 * Check for possible transition between PRS_ENABLED
1220		 * and PRS_ERROR.
1221		 */
1222		switch (cpuset->partition_root_state) {
1223		case PRS_ENABLED:
1224			if (part_error)
1225				cpuset->partition_root_state = PRS_ERROR;
1226			break;
1227		case PRS_ERROR:
1228			if (!part_error)
1229				cpuset->partition_root_state = PRS_ENABLED;
1230			break;
1231		}
1232		/*
1233		 * Set part_error if previously in invalid state.
1234		 */
1235		part_error = (prev_prs == PRS_ERROR);
1236	}
1237
1238	if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1239		return 0;	/* Nothing need to be done */
1240
1241	if (cpuset->partition_root_state == PRS_ERROR) {
1242		/*
1243		 * Remove all its cpus from parent's subparts_cpus.
1244		 */
1245		adding = false;
1246		deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1247				       parent->subparts_cpus);
1248	}
1249
1250	if (!adding && !deleting)
1251		return 0;
1252
1253	/*
1254	 * Change the parent's subparts_cpus.
1255	 * Newly added CPUs will be removed from effective_cpus and
1256	 * newly deleted ones will be added back to effective_cpus.
1257	 */
1258	spin_lock_irq(&callback_lock);
1259	if (adding) {
1260		cpumask_or(parent->subparts_cpus,
1261			   parent->subparts_cpus, tmp->addmask);
1262		cpumask_andnot(parent->effective_cpus,
1263			       parent->effective_cpus, tmp->addmask);
1264	}
1265	if (deleting) {
1266		cpumask_andnot(parent->subparts_cpus,
1267			       parent->subparts_cpus, tmp->delmask);
1268		/*
1269		 * Some of the CPUs in subparts_cpus might have been offlined.
1270		 */
1271		cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1272		cpumask_or(parent->effective_cpus,
1273			   parent->effective_cpus, tmp->delmask);
1274	}
1275
1276	parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1277	spin_unlock_irq(&callback_lock);
1278
1279	return cmd == partcmd_update;
1280}
1281
1282/*
1283 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1284 * @cs:  the cpuset to consider
1285 * @tmp: temp variables for calculating effective_cpus & partition setup
1286 *
1287 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1288 * and all its descendants need to be updated.
1289 *
1290 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1291 *
1292 * Called with cpuset_mutex held
1293 */
1294static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1295{
1296	struct cpuset *cp;
1297	struct cgroup_subsys_state *pos_css;
1298	bool need_rebuild_sched_domains = false;
1299
1300	rcu_read_lock();
1301	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1302		struct cpuset *parent = parent_cs(cp);
1303
1304		compute_effective_cpumask(tmp->new_cpus, cp, parent);
1305
1306		/*
1307		 * If it becomes empty, inherit the effective mask of the
1308		 * parent, which is guaranteed to have some CPUs.
1309		 */
1310		if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1311			cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1312			if (!cp->use_parent_ecpus) {
1313				cp->use_parent_ecpus = true;
1314				parent->child_ecpus_count++;
1315			}
1316		} else if (cp->use_parent_ecpus) {
1317			cp->use_parent_ecpus = false;
1318			WARN_ON_ONCE(!parent->child_ecpus_count);
1319			parent->child_ecpus_count--;
1320		}
1321
1322		/*
1323		 * Skip the whole subtree if the cpumask remains the same
1324		 * and has no partition root state.
1325		 */
1326		if (!cp->partition_root_state &&
1327		    cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1328			pos_css = css_rightmost_descendant(pos_css);
1329			continue;
1330		}
1331
1332		/*
1333		 * update_parent_subparts_cpumask() should have been called
1334		 * for cs already in update_cpumask(). We should also call
1335		 * update_tasks_cpumask() again for tasks in the parent
1336		 * cpuset if the parent's subparts_cpus changes.
1337		 */
1338		if ((cp != cs) && cp->partition_root_state) {
1339			switch (parent->partition_root_state) {
1340			case PRS_DISABLED:
1341				/*
1342				 * If parent is not a partition root or an
1343				 * invalid partition root, clear the state
1344				 * state and the CS_CPU_EXCLUSIVE flag.
1345				 */
1346				WARN_ON_ONCE(cp->partition_root_state
1347					     != PRS_ERROR);
1348				cp->partition_root_state = 0;
1349
1350				/*
1351				 * clear_bit() is an atomic operation and
1352				 * readers aren't interested in the state
1353				 * of CS_CPU_EXCLUSIVE anyway. So we can
1354				 * just update the flag without holding
1355				 * the callback_lock.
1356				 */
1357				clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1358				break;
1359
1360			case PRS_ENABLED:
1361				if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1362					update_tasks_cpumask(parent);
1363				break;
1364
1365			case PRS_ERROR:
1366				/*
1367				 * When parent is invalid, it has to be too.
1368				 */
1369				cp->partition_root_state = PRS_ERROR;
1370				if (cp->nr_subparts_cpus) {
1371					cp->nr_subparts_cpus = 0;
1372					cpumask_clear(cp->subparts_cpus);
1373				}
1374				break;
1375			}
1376		}
1377
1378		if (!css_tryget_online(&cp->css))
1379			continue;
1380		rcu_read_unlock();
1381
1382		spin_lock_irq(&callback_lock);
1383
1384		cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1385		if (cp->nr_subparts_cpus &&
1386		   (cp->partition_root_state != PRS_ENABLED)) {
1387			cp->nr_subparts_cpus = 0;
1388			cpumask_clear(cp->subparts_cpus);
1389		} else if (cp->nr_subparts_cpus) {
1390			/*
1391			 * Make sure that effective_cpus & subparts_cpus
1392			 * are mutually exclusive.
1393			 *
1394			 * In the unlikely event that effective_cpus
1395			 * becomes empty. we clear cp->nr_subparts_cpus and
1396			 * let its child partition roots to compete for
1397			 * CPUs again.
1398			 */
1399			cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1400				       cp->subparts_cpus);
1401			if (cpumask_empty(cp->effective_cpus)) {
1402				cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1403				cpumask_clear(cp->subparts_cpus);
1404				cp->nr_subparts_cpus = 0;
1405			} else if (!cpumask_subset(cp->subparts_cpus,
1406						   tmp->new_cpus)) {
1407				cpumask_andnot(cp->subparts_cpus,
1408					cp->subparts_cpus, tmp->new_cpus);
1409				cp->nr_subparts_cpus
1410					= cpumask_weight(cp->subparts_cpus);
1411			}
1412		}
1413		spin_unlock_irq(&callback_lock);
1414
1415		WARN_ON(!is_in_v2_mode() &&
1416			!cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1417
1418		update_tasks_cpumask(cp);
1419
1420		/*
1421		 * On legacy hierarchy, if the effective cpumask of any non-
1422		 * empty cpuset is changed, we need to rebuild sched domains.
1423		 * On default hierarchy, the cpuset needs to be a partition
1424		 * root as well.
1425		 */
1426		if (!cpumask_empty(cp->cpus_allowed) &&
1427		    is_sched_load_balance(cp) &&
1428		   (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1429		    is_partition_root(cp)))
1430			need_rebuild_sched_domains = true;
1431
1432		rcu_read_lock();
1433		css_put(&cp->css);
1434	}
1435	rcu_read_unlock();
1436
1437	if (need_rebuild_sched_domains)
1438		rebuild_sched_domains_locked();
1439}
1440
1441/**
1442 * update_sibling_cpumasks - Update siblings cpumasks
1443 * @parent:  Parent cpuset
1444 * @cs:      Current cpuset
1445 * @tmp:     Temp variables
1446 */
1447static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1448				    struct tmpmasks *tmp)
1449{
1450	struct cpuset *sibling;
1451	struct cgroup_subsys_state *pos_css;
1452
1453	/*
1454	 * Check all its siblings and call update_cpumasks_hier()
1455	 * if their use_parent_ecpus flag is set in order for them
1456	 * to use the right effective_cpus value.
1457	 */
1458	rcu_read_lock();
1459	cpuset_for_each_child(sibling, pos_css, parent) {
1460		if (sibling == cs)
1461			continue;
1462		if (!sibling->use_parent_ecpus)
1463			continue;
1464
1465		update_cpumasks_hier(sibling, tmp);
1466	}
1467	rcu_read_unlock();
1468}
1469
1470/**
1471 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1472 * @cs: the cpuset to consider
1473 * @trialcs: trial cpuset
1474 * @buf: buffer of cpu numbers written to this cpuset
1475 */
1476static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1477			  const char *buf)
1478{
1479	int retval;
1480	struct tmpmasks tmp;
1481
1482	/* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1483	if (cs == &top_cpuset)
1484		return -EACCES;
1485
1486	/*
1487	 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1488	 * Since cpulist_parse() fails on an empty mask, we special case
1489	 * that parsing.  The validate_change() call ensures that cpusets
1490	 * with tasks have cpus.
1491	 */
1492	if (!*buf) {
1493		cpumask_clear(trialcs->cpus_allowed);
1494	} else {
1495		retval = cpulist_parse(buf, trialcs->cpus_allowed);
1496		if (retval < 0)
1497			return retval;
1498
1499		if (!cpumask_subset(trialcs->cpus_allowed,
1500				    top_cpuset.cpus_allowed))
1501			return -EINVAL;
1502	}
1503
1504	/* Nothing to do if the cpus didn't change */
1505	if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1506		return 0;
1507
1508	retval = validate_change(cs, trialcs);
1509	if (retval < 0)
1510		return retval;
1511
1512#ifdef CONFIG_CPUMASK_OFFSTACK
1513	/*
1514	 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1515	 * to allocated cpumasks.
1516	 */
1517	tmp.addmask  = trialcs->subparts_cpus;
1518	tmp.delmask  = trialcs->effective_cpus;
1519	tmp.new_cpus = trialcs->cpus_allowed;
1520#endif
1521
1522	if (cs->partition_root_state) {
1523		/* Cpumask of a partition root cannot be empty */
1524		if (cpumask_empty(trialcs->cpus_allowed))
1525			return -EINVAL;
1526		if (update_parent_subparts_cpumask(cs, partcmd_update,
1527					trialcs->cpus_allowed, &tmp) < 0)
1528			return -EINVAL;
1529	}
1530
1531	spin_lock_irq(&callback_lock);
1532	cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1533
1534	/*
1535	 * Make sure that subparts_cpus is a subset of cpus_allowed.
1536	 */
1537	if (cs->nr_subparts_cpus) {
1538		cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1539			       cs->cpus_allowed);
1540		cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1541	}
1542	spin_unlock_irq(&callback_lock);
1543
1544	update_cpumasks_hier(cs, &tmp);
1545
1546	if (cs->partition_root_state) {
1547		struct cpuset *parent = parent_cs(cs);
1548
1549		/*
1550		 * For partition root, update the cpumasks of sibling
1551		 * cpusets if they use parent's effective_cpus.
1552		 */
1553		if (parent->child_ecpus_count)
1554			update_sibling_cpumasks(parent, cs, &tmp);
1555	}
1556	return 0;
1557}
1558
1559/*
1560 * Migrate memory region from one set of nodes to another.  This is
1561 * performed asynchronously as it can be called from process migration path
1562 * holding locks involved in process management.  All mm migrations are
1563 * performed in the queued order and can be waited for by flushing
1564 * cpuset_migrate_mm_wq.
1565 */
1566
1567struct cpuset_migrate_mm_work {
1568	struct work_struct	work;
1569	struct mm_struct	*mm;
1570	nodemask_t		from;
1571	nodemask_t		to;
1572};
1573
1574static void cpuset_migrate_mm_workfn(struct work_struct *work)
1575{
1576	struct cpuset_migrate_mm_work *mwork =
1577		container_of(work, struct cpuset_migrate_mm_work, work);
1578
1579	/* on a wq worker, no need to worry about %current's mems_allowed */
1580	do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1581	mmput(mwork->mm);
1582	kfree(mwork);
1583}
1584
1585static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1586							const nodemask_t *to)
1587{
1588	struct cpuset_migrate_mm_work *mwork;
1589
1590	mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1591	if (mwork) {
1592		mwork->mm = mm;
1593		mwork->from = *from;
1594		mwork->to = *to;
1595		INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1596		queue_work(cpuset_migrate_mm_wq, &mwork->work);
1597	} else {
1598		mmput(mm);
1599	}
1600}
1601
1602static void cpuset_post_attach(void)
1603{
1604	flush_workqueue(cpuset_migrate_mm_wq);
1605}
1606
1607/*
1608 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1609 * @tsk: the task to change
1610 * @newmems: new nodes that the task will be set
1611 *
1612 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1613 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1614 * parallel, it might temporarily see an empty intersection, which results in
1615 * a seqlock check and retry before OOM or allocation failure.
1616 */
1617static void cpuset_change_task_nodemask(struct task_struct *tsk,
1618					nodemask_t *newmems)
1619{
1620	task_lock(tsk);
1621
1622	local_irq_disable();
1623	write_seqcount_begin(&tsk->mems_allowed_seq);
1624
1625	nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1626	mpol_rebind_task(tsk, newmems);
1627	tsk->mems_allowed = *newmems;
1628
1629	write_seqcount_end(&tsk->mems_allowed_seq);
1630	local_irq_enable();
1631
1632	task_unlock(tsk);
1633}
1634
1635static void *cpuset_being_rebound;
1636
1637/**
1638 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1639 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1640 *
1641 * Iterate through each task of @cs updating its mems_allowed to the
1642 * effective cpuset's.  As this function is called with cpuset_mutex held,
1643 * cpuset membership stays stable.
1644 */
1645static void update_tasks_nodemask(struct cpuset *cs)
1646{
1647	static nodemask_t newmems;	/* protected by cpuset_mutex */
1648	struct css_task_iter it;
1649	struct task_struct *task;
1650
1651	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
1652
1653	guarantee_online_mems(cs, &newmems);
1654
1655	/*
1656	 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1657	 * take while holding tasklist_lock.  Forks can happen - the
1658	 * mpol_dup() cpuset_being_rebound check will catch such forks,
1659	 * and rebind their vma mempolicies too.  Because we still hold
1660	 * the global cpuset_mutex, we know that no other rebind effort
1661	 * will be contending for the global variable cpuset_being_rebound.
1662	 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1663	 * is idempotent.  Also migrate pages in each mm to new nodes.
1664	 */
1665	css_task_iter_start(&cs->css, 0, &it);
1666	while ((task = css_task_iter_next(&it))) {
1667		struct mm_struct *mm;
1668		bool migrate;
1669
1670		cpuset_change_task_nodemask(task, &newmems);
1671
1672		mm = get_task_mm(task);
1673		if (!mm)
1674			continue;
1675
1676		migrate = is_memory_migrate(cs);
1677
1678		mpol_rebind_mm(mm, &cs->mems_allowed);
1679		if (migrate)
1680			cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1681		else
1682			mmput(mm);
1683	}
1684	css_task_iter_end(&it);
1685
1686	/*
1687	 * All the tasks' nodemasks have been updated, update
1688	 * cs->old_mems_allowed.
1689	 */
1690	cs->old_mems_allowed = newmems;
1691
1692	/* We're done rebinding vmas to this cpuset's new mems_allowed. */
1693	cpuset_being_rebound = NULL;
1694}
1695
1696/*
1697 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1698 * @cs: the cpuset to consider
1699 * @new_mems: a temp variable for calculating new effective_mems
1700 *
1701 * When configured nodemask is changed, the effective nodemasks of this cpuset
1702 * and all its descendants need to be updated.
1703 *
1704 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1705 *
1706 * Called with cpuset_mutex held
1707 */
1708static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1709{
1710	struct cpuset *cp;
1711	struct cgroup_subsys_state *pos_css;
1712
1713	rcu_read_lock();
1714	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1715		struct cpuset *parent = parent_cs(cp);
1716
1717		nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1718
1719		/*
1720		 * If it becomes empty, inherit the effective mask of the
1721		 * parent, which is guaranteed to have some MEMs.
1722		 */
1723		if (is_in_v2_mode() && nodes_empty(*new_mems))
1724			*new_mems = parent->effective_mems;
1725
1726		/* Skip the whole subtree if the nodemask remains the same. */
1727		if (nodes_equal(*new_mems, cp->effective_mems)) {
1728			pos_css = css_rightmost_descendant(pos_css);
1729			continue;
1730		}
1731
1732		if (!css_tryget_online(&cp->css))
1733			continue;
1734		rcu_read_unlock();
1735
1736		spin_lock_irq(&callback_lock);
1737		cp->effective_mems = *new_mems;
1738		spin_unlock_irq(&callback_lock);
1739
1740		WARN_ON(!is_in_v2_mode() &&
1741			!nodes_equal(cp->mems_allowed, cp->effective_mems));
1742
1743		update_tasks_nodemask(cp);
1744
1745		rcu_read_lock();
1746		css_put(&cp->css);
1747	}
1748	rcu_read_unlock();
1749}
1750
1751/*
1752 * Handle user request to change the 'mems' memory placement
1753 * of a cpuset.  Needs to validate the request, update the
1754 * cpusets mems_allowed, and for each task in the cpuset,
1755 * update mems_allowed and rebind task's mempolicy and any vma
1756 * mempolicies and if the cpuset is marked 'memory_migrate',
1757 * migrate the tasks pages to the new memory.
1758 *
1759 * Call with cpuset_mutex held. May take callback_lock during call.
1760 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1761 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1762 * their mempolicies to the cpusets new mems_allowed.
1763 */
1764static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1765			   const char *buf)
1766{
1767	int retval;
1768
1769	/*
1770	 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1771	 * it's read-only
1772	 */
1773	if (cs == &top_cpuset) {
1774		retval = -EACCES;
1775		goto done;
1776	}
1777
1778	/*
1779	 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1780	 * Since nodelist_parse() fails on an empty mask, we special case
1781	 * that parsing.  The validate_change() call ensures that cpusets
1782	 * with tasks have memory.
1783	 */
1784	if (!*buf) {
1785		nodes_clear(trialcs->mems_allowed);
1786	} else {
1787		retval = nodelist_parse(buf, trialcs->mems_allowed);
1788		if (retval < 0)
1789			goto done;
1790
1791		if (!nodes_subset(trialcs->mems_allowed,
1792				  top_cpuset.mems_allowed)) {
1793			retval = -EINVAL;
1794			goto done;
1795		}
1796	}
1797
1798	if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1799		retval = 0;		/* Too easy - nothing to do */
1800		goto done;
1801	}
1802	retval = validate_change(cs, trialcs);
1803	if (retval < 0)
1804		goto done;
1805
1806	spin_lock_irq(&callback_lock);
1807	cs->mems_allowed = trialcs->mems_allowed;
1808	spin_unlock_irq(&callback_lock);
1809
1810	/* use trialcs->mems_allowed as a temp variable */
1811	update_nodemasks_hier(cs, &trialcs->mems_allowed);
1812done:
1813	return retval;
1814}
1815
1816bool current_cpuset_is_being_rebound(void)
1817{
1818	bool ret;
1819
1820	rcu_read_lock();
1821	ret = task_cs(current) == cpuset_being_rebound;
1822	rcu_read_unlock();
1823
1824	return ret;
1825}
1826
1827static int update_relax_domain_level(struct cpuset *cs, s64 val)
1828{
1829#ifdef CONFIG_SMP
1830	if (val < -1 || val >= sched_domain_level_max)
1831		return -EINVAL;
1832#endif
1833
1834	if (val != cs->relax_domain_level) {
1835		cs->relax_domain_level = val;
1836		if (!cpumask_empty(cs->cpus_allowed) &&
1837		    is_sched_load_balance(cs))
1838			rebuild_sched_domains_locked();
1839	}
1840
1841	return 0;
1842}
1843
1844/**
1845 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1846 * @cs: the cpuset in which each task's spread flags needs to be changed
1847 *
1848 * Iterate through each task of @cs updating its spread flags.  As this
1849 * function is called with cpuset_mutex held, cpuset membership stays
1850 * stable.
1851 */
1852static void update_tasks_flags(struct cpuset *cs)
1853{
1854	struct css_task_iter it;
1855	struct task_struct *task;
1856
1857	css_task_iter_start(&cs->css, 0, &it);
1858	while ((task = css_task_iter_next(&it)))
1859		cpuset_update_task_spread_flag(cs, task);
1860	css_task_iter_end(&it);
1861}
1862
1863/*
1864 * update_flag - read a 0 or a 1 in a file and update associated flag
1865 * bit:		the bit to update (see cpuset_flagbits_t)
1866 * cs:		the cpuset to update
1867 * turning_on: 	whether the flag is being set or cleared
1868 *
1869 * Call with cpuset_mutex held.
1870 */
1871
1872static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1873		       int turning_on)
1874{
1875	struct cpuset *trialcs;
1876	int balance_flag_changed;
1877	int spread_flag_changed;
1878	int err;
1879
1880	trialcs = alloc_trial_cpuset(cs);
1881	if (!trialcs)
1882		return -ENOMEM;
1883
1884	if (turning_on)
1885		set_bit(bit, &trialcs->flags);
1886	else
1887		clear_bit(bit, &trialcs->flags);
1888
1889	err = validate_change(cs, trialcs);
1890	if (err < 0)
1891		goto out;
1892
1893	balance_flag_changed = (is_sched_load_balance(cs) !=
1894				is_sched_load_balance(trialcs));
1895
1896	spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1897			|| (is_spread_page(cs) != is_spread_page(trialcs)));
1898
1899	spin_lock_irq(&callback_lock);
1900	cs->flags = trialcs->flags;
1901	spin_unlock_irq(&callback_lock);
1902
1903	if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1904		rebuild_sched_domains_locked();
1905
1906	if (spread_flag_changed)
1907		update_tasks_flags(cs);
1908out:
1909	free_cpuset(trialcs);
1910	return err;
1911}
1912
1913/*
1914 * update_prstate - update partititon_root_state
1915 * cs:	the cpuset to update
1916 * val: 0 - disabled, 1 - enabled
1917 *
1918 * Call with cpuset_mutex held.
1919 */
1920static int update_prstate(struct cpuset *cs, int val)
1921{
1922	int err;
1923	struct cpuset *parent = parent_cs(cs);
1924	struct tmpmasks tmp;
1925
1926	if ((val != 0) && (val != 1))
1927		return -EINVAL;
1928	if (val == cs->partition_root_state)
1929		return 0;
1930
1931	/*
1932	 * Cannot force a partial or invalid partition root to a full
1933	 * partition root.
1934	 */
1935	if (val && cs->partition_root_state)
1936		return -EINVAL;
1937
1938	if (alloc_cpumasks(NULL, &tmp))
1939		return -ENOMEM;
1940
1941	err = -EINVAL;
1942	if (!cs->partition_root_state) {
1943		/*
1944		 * Turning on partition root requires setting the
1945		 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1946		 * cannot be NULL.
1947		 */
1948		if (cpumask_empty(cs->cpus_allowed))
1949			goto out;
1950
1951		err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1952		if (err)
1953			goto out;
1954
1955		err = update_parent_subparts_cpumask(cs, partcmd_enable,
1956						     NULL, &tmp);
1957		if (err) {
1958			update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1959			goto out;
1960		}
1961		cs->partition_root_state = PRS_ENABLED;
1962	} else {
1963		/*
1964		 * Turning off partition root will clear the
1965		 * CS_CPU_EXCLUSIVE bit.
1966		 */
1967		if (cs->partition_root_state == PRS_ERROR) {
1968			cs->partition_root_state = 0;
1969			update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1970			err = 0;
1971			goto out;
1972		}
1973
1974		err = update_parent_subparts_cpumask(cs, partcmd_disable,
1975						     NULL, &tmp);
1976		if (err)
1977			goto out;
1978
1979		cs->partition_root_state = 0;
1980
1981		/* Turning off CS_CPU_EXCLUSIVE will not return error */
1982		update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1983	}
1984
1985	/*
1986	 * Update cpumask of parent's tasks except when it is the top
1987	 * cpuset as some system daemons cannot be mapped to other CPUs.
1988	 */
1989	if (parent != &top_cpuset)
1990		update_tasks_cpumask(parent);
1991
1992	if (parent->child_ecpus_count)
1993		update_sibling_cpumasks(parent, cs, &tmp);
1994
1995	rebuild_sched_domains_locked();
1996out:
1997	free_cpumasks(NULL, &tmp);
1998	return err;
1999}
2000
2001/*
2002 * Frequency meter - How fast is some event occurring?
2003 *
2004 * These routines manage a digitally filtered, constant time based,
2005 * event frequency meter.  There are four routines:
2006 *   fmeter_init() - initialize a frequency meter.
2007 *   fmeter_markevent() - called each time the event happens.
2008 *   fmeter_getrate() - returns the recent rate of such events.
2009 *   fmeter_update() - internal routine used to update fmeter.
2010 *
2011 * A common data structure is passed to each of these routines,
2012 * which is used to keep track of the state required to manage the
2013 * frequency meter and its digital filter.
2014 *
2015 * The filter works on the number of events marked per unit time.
2016 * The filter is single-pole low-pass recursive (IIR).  The time unit
2017 * is 1 second.  Arithmetic is done using 32-bit integers scaled to
2018 * simulate 3 decimal digits of precision (multiplied by 1000).
2019 *
2020 * With an FM_COEF of 933, and a time base of 1 second, the filter
2021 * has a half-life of 10 seconds, meaning that if the events quit
2022 * happening, then the rate returned from the fmeter_getrate()
2023 * will be cut in half each 10 seconds, until it converges to zero.
2024 *
2025 * It is not worth doing a real infinitely recursive filter.  If more
2026 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2027 * just compute FM_MAXTICKS ticks worth, by which point the level
2028 * will be stable.
2029 *
2030 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2031 * arithmetic overflow in the fmeter_update() routine.
2032 *
2033 * Given the simple 32 bit integer arithmetic used, this meter works
2034 * best for reporting rates between one per millisecond (msec) and
2035 * one per 32 (approx) seconds.  At constant rates faster than one
2036 * per msec it maxes out at values just under 1,000,000.  At constant
2037 * rates between one per msec, and one per second it will stabilize
2038 * to a value N*1000, where N is the rate of events per second.
2039 * At constant rates between one per second and one per 32 seconds,
2040 * it will be choppy, moving up on the seconds that have an event,
2041 * and then decaying until the next event.  At rates slower than
2042 * about one in 32 seconds, it decays all the way back to zero between
2043 * each event.
2044 */
2045
2046#define FM_COEF 933		/* coefficient for half-life of 10 secs */
2047#define FM_MAXTICKS ((u32)99)   /* useless computing more ticks than this */
2048#define FM_MAXCNT 1000000	/* limit cnt to avoid overflow */
2049#define FM_SCALE 1000		/* faux fixed point scale */
2050
2051/* Initialize a frequency meter */
2052static void fmeter_init(struct fmeter *fmp)
2053{
2054	fmp->cnt = 0;
2055	fmp->val = 0;
2056	fmp->time = 0;
2057	spin_lock_init(&fmp->lock);
2058}
2059
2060/* Internal meter update - process cnt events and update value */
2061static void fmeter_update(struct fmeter *fmp)
2062{
2063	time64_t now;
2064	u32 ticks;
2065
2066	now = ktime_get_seconds();
2067	ticks = now - fmp->time;
2068
2069	if (ticks == 0)
2070		return;
2071
2072	ticks = min(FM_MAXTICKS, ticks);
2073	while (ticks-- > 0)
2074		fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2075	fmp->time = now;
2076
2077	fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2078	fmp->cnt = 0;
2079}
2080
2081/* Process any previous ticks, then bump cnt by one (times scale). */
2082static void fmeter_markevent(struct fmeter *fmp)
2083{
2084	spin_lock(&fmp->lock);
2085	fmeter_update(fmp);
2086	fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2087	spin_unlock(&fmp->lock);
2088}
2089
2090/* Process any previous ticks, then return current value. */
2091static int fmeter_getrate(struct fmeter *fmp)
2092{
2093	int val;
2094
2095	spin_lock(&fmp->lock);
2096	fmeter_update(fmp);
2097	val = fmp->val;
2098	spin_unlock(&fmp->lock);
2099	return val;
2100}
2101
2102static struct cpuset *cpuset_attach_old_cs;
2103
2104/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2105static int cpuset_can_attach(struct cgroup_taskset *tset)
2106{
2107	struct cgroup_subsys_state *css;
2108	struct cpuset *cs;
2109	struct task_struct *task;
2110	int ret;
2111
2112	/* used later by cpuset_attach() */
2113	cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2114	cs = css_cs(css);
2115
2116	percpu_down_write(&cpuset_rwsem);
2117
2118	/* allow moving tasks into an empty cpuset if on default hierarchy */
2119	ret = -ENOSPC;
2120	if (!is_in_v2_mode() &&
2121	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2122		goto out_unlock;
2123
2124	cgroup_taskset_for_each(task, css, tset) {
2125		ret = task_can_attach(task, cs->cpus_allowed);
2126		if (ret)
2127			goto out_unlock;
2128		ret = security_task_setscheduler(task);
2129		if (ret)
2130			goto out_unlock;
2131	}
2132
2133	/*
2134	 * Mark attach is in progress.  This makes validate_change() fail
2135	 * changes which zero cpus/mems_allowed.
2136	 */
2137	cs->attach_in_progress++;
2138	ret = 0;
2139out_unlock:
2140	percpu_up_write(&cpuset_rwsem);
2141	return ret;
2142}
2143
2144static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2145{
2146	struct cgroup_subsys_state *css;
2147
2148	cgroup_taskset_first(tset, &css);
2149
2150	percpu_down_write(&cpuset_rwsem);
2151	css_cs(css)->attach_in_progress--;
2152	percpu_up_write(&cpuset_rwsem);
2153}
2154
2155/*
2156 * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
2157 * but we can't allocate it dynamically there.  Define it global and
2158 * allocate from cpuset_init().
2159 */
2160static cpumask_var_t cpus_attach;
2161
2162static void cpuset_attach(struct cgroup_taskset *tset)
2163{
2164	/* static buf protected by cpuset_mutex */
2165	static nodemask_t cpuset_attach_nodemask_to;
2166	struct task_struct *task;
2167	struct task_struct *leader;
2168	struct cgroup_subsys_state *css;
2169	struct cpuset *cs;
2170	struct cpuset *oldcs = cpuset_attach_old_cs;
2171
2172	cgroup_taskset_first(tset, &css);
2173	cs = css_cs(css);
2174
2175	percpu_down_write(&cpuset_rwsem);
2176
2177	/* prepare for attach */
2178	if (cs == &top_cpuset)
2179		cpumask_copy(cpus_attach, cpu_possible_mask);
2180	else
2181		guarantee_online_cpus(cs, cpus_attach);
2182
2183	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2184
2185	cgroup_taskset_for_each(task, css, tset) {
2186		/*
2187		 * can_attach beforehand should guarantee that this doesn't
2188		 * fail.  TODO: have a better way to handle failure here
2189		 */
2190		WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2191
2192		cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2193		cpuset_update_task_spread_flag(cs, task);
2194	}
2195
2196	/*
2197	 * Change mm for all threadgroup leaders. This is expensive and may
2198	 * sleep and should be moved outside migration path proper.
2199	 */
2200	cpuset_attach_nodemask_to = cs->effective_mems;
2201	cgroup_taskset_for_each_leader(leader, css, tset) {
2202		struct mm_struct *mm = get_task_mm(leader);
2203
2204		if (mm) {
2205			mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2206
2207			/*
2208			 * old_mems_allowed is the same with mems_allowed
2209			 * here, except if this task is being moved
2210			 * automatically due to hotplug.  In that case
2211			 * @mems_allowed has been updated and is empty, so
2212			 * @old_mems_allowed is the right nodesets that we
2213			 * migrate mm from.
2214			 */
2215			if (is_memory_migrate(cs))
2216				cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2217						  &cpuset_attach_nodemask_to);
2218			else
2219				mmput(mm);
2220		}
2221	}
2222
2223	cs->old_mems_allowed = cpuset_attach_nodemask_to;
2224
2225	cs->attach_in_progress--;
2226	if (!cs->attach_in_progress)
2227		wake_up(&cpuset_attach_wq);
2228
2229	percpu_up_write(&cpuset_rwsem);
2230}
2231
2232/* The various types of files and directories in a cpuset file system */
2233
2234typedef enum {
2235	FILE_MEMORY_MIGRATE,
2236	FILE_CPULIST,
2237	FILE_MEMLIST,
2238	FILE_EFFECTIVE_CPULIST,
2239	FILE_EFFECTIVE_MEMLIST,
2240	FILE_SUBPARTS_CPULIST,
2241	FILE_CPU_EXCLUSIVE,
2242	FILE_MEM_EXCLUSIVE,
2243	FILE_MEM_HARDWALL,
2244	FILE_SCHED_LOAD_BALANCE,
2245	FILE_PARTITION_ROOT,
2246	FILE_SCHED_RELAX_DOMAIN_LEVEL,
2247	FILE_MEMORY_PRESSURE_ENABLED,
2248	FILE_MEMORY_PRESSURE,
2249	FILE_SPREAD_PAGE,
2250	FILE_SPREAD_SLAB,
2251} cpuset_filetype_t;
2252
2253static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2254			    u64 val)
2255{
2256	struct cpuset *cs = css_cs(css);
2257	cpuset_filetype_t type = cft->private;
2258	int retval = 0;
2259
2260	get_online_cpus();
2261	percpu_down_write(&cpuset_rwsem);
2262	if (!is_cpuset_online(cs)) {
2263		retval = -ENODEV;
2264		goto out_unlock;
2265	}
2266
2267	switch (type) {
2268	case FILE_CPU_EXCLUSIVE:
2269		retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2270		break;
2271	case FILE_MEM_EXCLUSIVE:
2272		retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2273		break;
2274	case FILE_MEM_HARDWALL:
2275		retval = update_flag(CS_MEM_HARDWALL, cs, val);
2276		break;
2277	case FILE_SCHED_LOAD_BALANCE:
2278		retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2279		break;
2280	case FILE_MEMORY_MIGRATE:
2281		retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2282		break;
2283	case FILE_MEMORY_PRESSURE_ENABLED:
2284		cpuset_memory_pressure_enabled = !!val;
2285		break;
2286	case FILE_SPREAD_PAGE:
2287		retval = update_flag(CS_SPREAD_PAGE, cs, val);
2288		break;
2289	case FILE_SPREAD_SLAB:
2290		retval = update_flag(CS_SPREAD_SLAB, cs, val);
2291		break;
2292	default:
2293		retval = -EINVAL;
2294		break;
2295	}
2296out_unlock:
2297	percpu_up_write(&cpuset_rwsem);
2298	put_online_cpus();
2299	return retval;
2300}
2301
2302static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2303			    s64 val)
2304{
2305	struct cpuset *cs = css_cs(css);
2306	cpuset_filetype_t type = cft->private;
2307	int retval = -ENODEV;
2308
2309	get_online_cpus();
2310	percpu_down_write(&cpuset_rwsem);
2311	if (!is_cpuset_online(cs))
2312		goto out_unlock;
2313
2314	switch (type) {
2315	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2316		retval = update_relax_domain_level(cs, val);
2317		break;
2318	default:
2319		retval = -EINVAL;
2320		break;
2321	}
2322out_unlock:
2323	percpu_up_write(&cpuset_rwsem);
2324	put_online_cpus();
2325	return retval;
2326}
2327
2328/*
2329 * Common handling for a write to a "cpus" or "mems" file.
2330 */
2331static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2332				    char *buf, size_t nbytes, loff_t off)
2333{
2334	struct cpuset *cs = css_cs(of_css(of));
2335	struct cpuset *trialcs;
2336	int retval = -ENODEV;
2337
2338	buf = strstrip(buf);
2339
2340	/*
2341	 * CPU or memory hotunplug may leave @cs w/o any execution
2342	 * resources, in which case the hotplug code asynchronously updates
2343	 * configuration and transfers all tasks to the nearest ancestor
2344	 * which can execute.
2345	 *
2346	 * As writes to "cpus" or "mems" may restore @cs's execution
2347	 * resources, wait for the previously scheduled operations before
2348	 * proceeding, so that we don't end up keep removing tasks added
2349	 * after execution capability is restored.
2350	 *
2351	 * cpuset_hotplug_work calls back into cgroup core via
2352	 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2353	 * operation like this one can lead to a deadlock through kernfs
2354	 * active_ref protection.  Let's break the protection.  Losing the
2355	 * protection is okay as we check whether @cs is online after
2356	 * grabbing cpuset_mutex anyway.  This only happens on the legacy
2357	 * hierarchies.
2358	 */
2359	css_get(&cs->css);
2360	kernfs_break_active_protection(of->kn);
2361	flush_work(&cpuset_hotplug_work);
2362
2363	get_online_cpus();
2364	percpu_down_write(&cpuset_rwsem);
2365	if (!is_cpuset_online(cs))
2366		goto out_unlock;
2367
2368	trialcs = alloc_trial_cpuset(cs);
2369	if (!trialcs) {
2370		retval = -ENOMEM;
2371		goto out_unlock;
2372	}
2373
2374	switch (of_cft(of)->private) {
2375	case FILE_CPULIST:
2376		retval = update_cpumask(cs, trialcs, buf);
2377		break;
2378	case FILE_MEMLIST:
2379		retval = update_nodemask(cs, trialcs, buf);
2380		break;
2381	default:
2382		retval = -EINVAL;
2383		break;
2384	}
2385
2386	free_cpuset(trialcs);
2387out_unlock:
2388	percpu_up_write(&cpuset_rwsem);
2389	put_online_cpus();
2390	kernfs_unbreak_active_protection(of->kn);
2391	css_put(&cs->css);
2392	flush_workqueue(cpuset_migrate_mm_wq);
2393	return retval ?: nbytes;
2394}
2395
2396/*
2397 * These ascii lists should be read in a single call, by using a user
2398 * buffer large enough to hold the entire map.  If read in smaller
2399 * chunks, there is no guarantee of atomicity.  Since the display format
2400 * used, list of ranges of sequential numbers, is variable length,
2401 * and since these maps can change value dynamically, one could read
2402 * gibberish by doing partial reads while a list was changing.
2403 */
2404static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2405{
2406	struct cpuset *cs = css_cs(seq_css(sf));
2407	cpuset_filetype_t type = seq_cft(sf)->private;
2408	int ret = 0;
2409
2410	spin_lock_irq(&callback_lock);
2411
2412	switch (type) {
2413	case FILE_CPULIST:
2414		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2415		break;
2416	case FILE_MEMLIST:
2417		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2418		break;
2419	case FILE_EFFECTIVE_CPULIST:
2420		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2421		break;
2422	case FILE_EFFECTIVE_MEMLIST:
2423		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2424		break;
2425	case FILE_SUBPARTS_CPULIST:
2426		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2427		break;
2428	default:
2429		ret = -EINVAL;
2430	}
2431
2432	spin_unlock_irq(&callback_lock);
2433	return ret;
2434}
2435
2436static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2437{
2438	struct cpuset *cs = css_cs(css);
2439	cpuset_filetype_t type = cft->private;
2440	switch (type) {
2441	case FILE_CPU_EXCLUSIVE:
2442		return is_cpu_exclusive(cs);
2443	case FILE_MEM_EXCLUSIVE:
2444		return is_mem_exclusive(cs);
2445	case FILE_MEM_HARDWALL:
2446		return is_mem_hardwall(cs);
2447	case FILE_SCHED_LOAD_BALANCE:
2448		return is_sched_load_balance(cs);
2449	case FILE_MEMORY_MIGRATE:
2450		return is_memory_migrate(cs);
2451	case FILE_MEMORY_PRESSURE_ENABLED:
2452		return cpuset_memory_pressure_enabled;
2453	case FILE_MEMORY_PRESSURE:
2454		return fmeter_getrate(&cs->fmeter);
2455	case FILE_SPREAD_PAGE:
2456		return is_spread_page(cs);
2457	case FILE_SPREAD_SLAB:
2458		return is_spread_slab(cs);
2459	default:
2460		BUG();
2461	}
2462
2463	/* Unreachable but makes gcc happy */
2464	return 0;
2465}
2466
2467static s64 cpuset_read_s64(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_SCHED_RELAX_DOMAIN_LEVEL:
2473		return cs->relax_domain_level;
2474	default:
2475		BUG();
2476	}
2477
2478	/* Unrechable but makes gcc happy */
2479	return 0;
2480}
2481
2482static int sched_partition_show(struct seq_file *seq, void *v)
2483{
2484	struct cpuset *cs = css_cs(seq_css(seq));
2485
2486	switch (cs->partition_root_state) {
2487	case PRS_ENABLED:
2488		seq_puts(seq, "root\n");
2489		break;
2490	case PRS_DISABLED:
2491		seq_puts(seq, "member\n");
2492		break;
2493	case PRS_ERROR:
2494		seq_puts(seq, "root invalid\n");
2495		break;
2496	}
2497	return 0;
2498}
2499
2500static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2501				     size_t nbytes, loff_t off)
2502{
2503	struct cpuset *cs = css_cs(of_css(of));
2504	int val;
2505	int retval = -ENODEV;
2506
2507	buf = strstrip(buf);
2508
2509	/*
2510	 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2511	 */
2512	if (!strcmp(buf, "root"))
2513		val = PRS_ENABLED;
2514	else if (!strcmp(buf, "member"))
2515		val = PRS_DISABLED;
2516	else
2517		return -EINVAL;
2518
2519	css_get(&cs->css);
2520	get_online_cpus();
2521	percpu_down_write(&cpuset_rwsem);
2522	if (!is_cpuset_online(cs))
2523		goto out_unlock;
2524
2525	retval = update_prstate(cs, val);
2526out_unlock:
2527	percpu_up_write(&cpuset_rwsem);
2528	put_online_cpus();
2529	css_put(&cs->css);
2530	return retval ?: nbytes;
2531}
2532
2533/*
2534 * for the common functions, 'private' gives the type of file
2535 */
2536
2537static struct cftype legacy_files[] = {
2538	{
2539		.name = "cpus",
2540		.seq_show = cpuset_common_seq_show,
2541		.write = cpuset_write_resmask,
2542		.max_write_len = (100U + 6 * NR_CPUS),
2543		.private = FILE_CPULIST,
2544	},
2545
2546	{
2547		.name = "mems",
2548		.seq_show = cpuset_common_seq_show,
2549		.write = cpuset_write_resmask,
2550		.max_write_len = (100U + 6 * MAX_NUMNODES),
2551		.private = FILE_MEMLIST,
2552	},
2553
2554	{
2555		.name = "effective_cpus",
2556		.seq_show = cpuset_common_seq_show,
2557		.private = FILE_EFFECTIVE_CPULIST,
2558	},
2559
2560	{
2561		.name = "effective_mems",
2562		.seq_show = cpuset_common_seq_show,
2563		.private = FILE_EFFECTIVE_MEMLIST,
2564	},
2565
2566	{
2567		.name = "cpu_exclusive",
2568		.read_u64 = cpuset_read_u64,
2569		.write_u64 = cpuset_write_u64,
2570		.private = FILE_CPU_EXCLUSIVE,
2571	},
2572
2573	{
2574		.name = "mem_exclusive",
2575		.read_u64 = cpuset_read_u64,
2576		.write_u64 = cpuset_write_u64,
2577		.private = FILE_MEM_EXCLUSIVE,
2578	},
2579
2580	{
2581		.name = "mem_hardwall",
2582		.read_u64 = cpuset_read_u64,
2583		.write_u64 = cpuset_write_u64,
2584		.private = FILE_MEM_HARDWALL,
2585	},
2586
2587	{
2588		.name = "sched_load_balance",
2589		.read_u64 = cpuset_read_u64,
2590		.write_u64 = cpuset_write_u64,
2591		.private = FILE_SCHED_LOAD_BALANCE,
2592	},
2593
2594	{
2595		.name = "sched_relax_domain_level",
2596		.read_s64 = cpuset_read_s64,
2597		.write_s64 = cpuset_write_s64,
2598		.private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2599	},
2600
2601	{
2602		.name = "memory_migrate",
2603		.read_u64 = cpuset_read_u64,
2604		.write_u64 = cpuset_write_u64,
2605		.private = FILE_MEMORY_MIGRATE,
2606	},
2607
2608	{
2609		.name = "memory_pressure",
2610		.read_u64 = cpuset_read_u64,
2611		.private = FILE_MEMORY_PRESSURE,
2612	},
2613
2614	{
2615		.name = "memory_spread_page",
2616		.read_u64 = cpuset_read_u64,
2617		.write_u64 = cpuset_write_u64,
2618		.private = FILE_SPREAD_PAGE,
2619	},
2620
2621	{
2622		.name = "memory_spread_slab",
2623		.read_u64 = cpuset_read_u64,
2624		.write_u64 = cpuset_write_u64,
2625		.private = FILE_SPREAD_SLAB,
2626	},
2627
2628	{
2629		.name = "memory_pressure_enabled",
2630		.flags = CFTYPE_ONLY_ON_ROOT,
2631		.read_u64 = cpuset_read_u64,
2632		.write_u64 = cpuset_write_u64,
2633		.private = FILE_MEMORY_PRESSURE_ENABLED,
2634	},
2635
2636	{ }	/* terminate */
2637};
2638
2639/*
2640 * This is currently a minimal set for the default hierarchy. It can be
2641 * expanded later on by migrating more features and control files from v1.
2642 */
2643static struct cftype dfl_files[] = {
2644	{
2645		.name = "cpus",
2646		.seq_show = cpuset_common_seq_show,
2647		.write = cpuset_write_resmask,
2648		.max_write_len = (100U + 6 * NR_CPUS),
2649		.private = FILE_CPULIST,
2650		.flags = CFTYPE_NOT_ON_ROOT,
2651	},
2652
2653	{
2654		.name = "mems",
2655		.seq_show = cpuset_common_seq_show,
2656		.write = cpuset_write_resmask,
2657		.max_write_len = (100U + 6 * MAX_NUMNODES),
2658		.private = FILE_MEMLIST,
2659		.flags = CFTYPE_NOT_ON_ROOT,
2660	},
2661
2662	{
2663		.name = "cpus.effective",
2664		.seq_show = cpuset_common_seq_show,
2665		.private = FILE_EFFECTIVE_CPULIST,
2666	},
2667
2668	{
2669		.name = "mems.effective",
2670		.seq_show = cpuset_common_seq_show,
2671		.private = FILE_EFFECTIVE_MEMLIST,
2672	},
2673
2674	{
2675		.name = "cpus.partition",
2676		.seq_show = sched_partition_show,
2677		.write = sched_partition_write,
2678		.private = FILE_PARTITION_ROOT,
2679		.flags = CFTYPE_NOT_ON_ROOT,
2680	},
2681
2682	{
2683		.name = "cpus.subpartitions",
2684		.seq_show = cpuset_common_seq_show,
2685		.private = FILE_SUBPARTS_CPULIST,
2686		.flags = CFTYPE_DEBUG,
2687	},
2688
2689	{ }	/* terminate */
2690};
2691
2692
2693/*
2694 *	cpuset_css_alloc - allocate a cpuset css
2695 *	cgrp:	control group that the new cpuset will be part of
2696 */
2697
2698static struct cgroup_subsys_state *
2699cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2700{
2701	struct cpuset *cs;
2702
2703	if (!parent_css)
2704		return &top_cpuset.css;
2705
2706	cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2707	if (!cs)
2708		return ERR_PTR(-ENOMEM);
2709
2710	if (alloc_cpumasks(cs, NULL)) {
2711		kfree(cs);
2712		return ERR_PTR(-ENOMEM);
2713	}
2714
2715	set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2716	nodes_clear(cs->mems_allowed);
2717	nodes_clear(cs->effective_mems);
2718	fmeter_init(&cs->fmeter);
2719	cs->relax_domain_level = -1;
2720
2721	return &cs->css;
2722}
2723
2724static int cpuset_css_online(struct cgroup_subsys_state *css)
2725{
2726	struct cpuset *cs = css_cs(css);
2727	struct cpuset *parent = parent_cs(cs);
2728	struct cpuset *tmp_cs;
2729	struct cgroup_subsys_state *pos_css;
2730
2731	if (!parent)
2732		return 0;
2733
2734	get_online_cpus();
2735	percpu_down_write(&cpuset_rwsem);
2736
2737	set_bit(CS_ONLINE, &cs->flags);
2738	if (is_spread_page(parent))
2739		set_bit(CS_SPREAD_PAGE, &cs->flags);
2740	if (is_spread_slab(parent))
2741		set_bit(CS_SPREAD_SLAB, &cs->flags);
2742
2743	cpuset_inc();
2744
2745	spin_lock_irq(&callback_lock);
2746	if (is_in_v2_mode()) {
2747		cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2748		cs->effective_mems = parent->effective_mems;
2749		cs->use_parent_ecpus = true;
2750		parent->child_ecpus_count++;
2751	}
2752	spin_unlock_irq(&callback_lock);
2753
2754	if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2755		goto out_unlock;
2756
2757	/*
2758	 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2759	 * set.  This flag handling is implemented in cgroup core for
2760	 * histrical reasons - the flag may be specified during mount.
2761	 *
2762	 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2763	 * refuse to clone the configuration - thereby refusing the task to
2764	 * be entered, and as a result refusing the sys_unshare() or
2765	 * clone() which initiated it.  If this becomes a problem for some
2766	 * users who wish to allow that scenario, then this could be
2767	 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2768	 * (and likewise for mems) to the new cgroup.
2769	 */
2770	rcu_read_lock();
2771	cpuset_for_each_child(tmp_cs, pos_css, parent) {
2772		if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2773			rcu_read_unlock();
2774			goto out_unlock;
2775		}
2776	}
2777	rcu_read_unlock();
2778
2779	spin_lock_irq(&callback_lock);
2780	cs->mems_allowed = parent->mems_allowed;
2781	cs->effective_mems = parent->mems_allowed;
2782	cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2783	cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2784	spin_unlock_irq(&callback_lock);
2785out_unlock:
2786	percpu_up_write(&cpuset_rwsem);
2787	put_online_cpus();
2788	return 0;
2789}
2790
2791/*
2792 * If the cpuset being removed has its flag 'sched_load_balance'
2793 * enabled, then simulate turning sched_load_balance off, which
2794 * will call rebuild_sched_domains_locked(). That is not needed
2795 * in the default hierarchy where only changes in partition
2796 * will cause repartitioning.
2797 *
2798 * If the cpuset has the 'sched.partition' flag enabled, simulate
2799 * turning 'sched.partition" off.
2800 */
2801
2802static void cpuset_css_offline(struct cgroup_subsys_state *css)
2803{
2804	struct cpuset *cs = css_cs(css);
2805
2806	get_online_cpus();
2807	percpu_down_write(&cpuset_rwsem);
2808
2809	if (is_partition_root(cs))
2810		update_prstate(cs, 0);
2811
2812	if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2813	    is_sched_load_balance(cs))
2814		update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2815
2816	if (cs->use_parent_ecpus) {
2817		struct cpuset *parent = parent_cs(cs);
2818
2819		cs->use_parent_ecpus = false;
2820		parent->child_ecpus_count--;
2821	}
2822
2823	cpuset_dec();
2824	clear_bit(CS_ONLINE, &cs->flags);
2825
2826	percpu_up_write(&cpuset_rwsem);
2827	put_online_cpus();
2828}
2829
2830static void cpuset_css_free(struct cgroup_subsys_state *css)
2831{
2832	struct cpuset *cs = css_cs(css);
2833
2834	free_cpuset(cs);
2835}
2836
2837static void cpuset_bind(struct cgroup_subsys_state *root_css)
2838{
2839	percpu_down_write(&cpuset_rwsem);
2840	spin_lock_irq(&callback_lock);
2841
2842	if (is_in_v2_mode()) {
2843		cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2844		top_cpuset.mems_allowed = node_possible_map;
2845	} else {
2846		cpumask_copy(top_cpuset.cpus_allowed,
2847			     top_cpuset.effective_cpus);
2848		top_cpuset.mems_allowed = top_cpuset.effective_mems;
2849	}
2850
2851	spin_unlock_irq(&callback_lock);
2852	percpu_up_write(&cpuset_rwsem);
2853}
2854
2855/*
2856 * Make sure the new task conform to the current state of its parent,
2857 * which could have been changed by cpuset just after it inherits the
2858 * state from the parent and before it sits on the cgroup's task list.
2859 */
2860static void cpuset_fork(struct task_struct *task)
2861{
2862	if (task_css_is_root(task, cpuset_cgrp_id))
2863		return;
2864
2865	set_cpus_allowed_ptr(task, current->cpus_ptr);
2866	task->mems_allowed = current->mems_allowed;
2867}
2868
2869struct cgroup_subsys cpuset_cgrp_subsys = {
2870	.css_alloc	= cpuset_css_alloc,
2871	.css_online	= cpuset_css_online,
2872	.css_offline	= cpuset_css_offline,
2873	.css_free	= cpuset_css_free,
2874	.can_attach	= cpuset_can_attach,
2875	.cancel_attach	= cpuset_cancel_attach,
2876	.attach		= cpuset_attach,
2877	.post_attach	= cpuset_post_attach,
2878	.bind		= cpuset_bind,
2879	.fork		= cpuset_fork,
2880	.legacy_cftypes	= legacy_files,
2881	.dfl_cftypes	= dfl_files,
2882	.early_init	= true,
2883	.threaded	= true,
2884};
2885
2886/**
2887 * cpuset_init - initialize cpusets at system boot
2888 *
2889 * Description: Initialize top_cpuset
2890 **/
2891
2892int __init cpuset_init(void)
2893{
2894	BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2895
2896	BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2897	BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2898	BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2899
2900	cpumask_setall(top_cpuset.cpus_allowed);
2901	nodes_setall(top_cpuset.mems_allowed);
2902	cpumask_setall(top_cpuset.effective_cpus);
2903	nodes_setall(top_cpuset.effective_mems);
2904
2905	fmeter_init(&top_cpuset.fmeter);
2906	set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2907	top_cpuset.relax_domain_level = -1;
2908
2909	BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2910
2911	return 0;
2912}
2913
2914/*
2915 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2916 * or memory nodes, we need to walk over the cpuset hierarchy,
2917 * removing that CPU or node from all cpusets.  If this removes the
2918 * last CPU or node from a cpuset, then move the tasks in the empty
2919 * cpuset to its next-highest non-empty parent.
2920 */
2921static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2922{
2923	struct cpuset *parent;
2924
2925	/*
2926	 * Find its next-highest non-empty parent, (top cpuset
2927	 * has online cpus, so can't be empty).
2928	 */
2929	parent = parent_cs(cs);
2930	while (cpumask_empty(parent->cpus_allowed) ||
2931			nodes_empty(parent->mems_allowed))
2932		parent = parent_cs(parent);
2933
2934	if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2935		pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2936		pr_cont_cgroup_name(cs->css.cgroup);
2937		pr_cont("\n");
2938	}
2939}
2940
2941static void
2942hotplug_update_tasks_legacy(struct cpuset *cs,
2943			    struct cpumask *new_cpus, nodemask_t *new_mems,
2944			    bool cpus_updated, bool mems_updated)
2945{
2946	bool is_empty;
2947
2948	spin_lock_irq(&callback_lock);
2949	cpumask_copy(cs->cpus_allowed, new_cpus);
2950	cpumask_copy(cs->effective_cpus, new_cpus);
2951	cs->mems_allowed = *new_mems;
2952	cs->effective_mems = *new_mems;
2953	spin_unlock_irq(&callback_lock);
2954
2955	/*
2956	 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2957	 * as the tasks will be migratecd to an ancestor.
2958	 */
2959	if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2960		update_tasks_cpumask(cs);
2961	if (mems_updated && !nodes_empty(cs->mems_allowed))
2962		update_tasks_nodemask(cs);
2963
2964	is_empty = cpumask_empty(cs->cpus_allowed) ||
2965		   nodes_empty(cs->mems_allowed);
2966
2967	percpu_up_write(&cpuset_rwsem);
2968
2969	/*
2970	 * Move tasks to the nearest ancestor with execution resources,
2971	 * This is full cgroup operation which will also call back into
2972	 * cpuset. Should be done outside any lock.
2973	 */
2974	if (is_empty)
2975		remove_tasks_in_empty_cpuset(cs);
2976
2977	percpu_down_write(&cpuset_rwsem);
2978}
2979
2980static void
2981hotplug_update_tasks(struct cpuset *cs,
2982		     struct cpumask *new_cpus, nodemask_t *new_mems,
2983		     bool cpus_updated, bool mems_updated)
2984{
2985	if (cpumask_empty(new_cpus))
2986		cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2987	if (nodes_empty(*new_mems))
2988		*new_mems = parent_cs(cs)->effective_mems;
2989
2990	spin_lock_irq(&callback_lock);
2991	cpumask_copy(cs->effective_cpus, new_cpus);
2992	cs->effective_mems = *new_mems;
2993	spin_unlock_irq(&callback_lock);
2994
2995	if (cpus_updated)
2996		update_tasks_cpumask(cs);
2997	if (mems_updated)
2998		update_tasks_nodemask(cs);
2999}
3000
3001static bool force_rebuild;
3002
3003void cpuset_force_rebuild(void)
3004{
3005	force_rebuild = true;
3006}
3007
3008/**
3009 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3010 * @cs: cpuset in interest
3011 * @tmp: the tmpmasks structure pointer
3012 *
3013 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3014 * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
3015 * all its tasks are moved to the nearest ancestor with both resources.
3016 */
3017static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3018{
3019	static cpumask_t new_cpus;
3020	static nodemask_t new_mems;
3021	bool cpus_updated;
3022	bool mems_updated;
3023	struct cpuset *parent;
3024retry:
3025	wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3026
3027	percpu_down_write(&cpuset_rwsem);
3028
3029	/*
3030	 * We have raced with task attaching. We wait until attaching
3031	 * is finished, so we won't attach a task to an empty cpuset.
3032	 */
3033	if (cs->attach_in_progress) {
3034		percpu_up_write(&cpuset_rwsem);
3035		goto retry;
3036	}
3037
3038	parent =  parent_cs(cs);
3039	compute_effective_cpumask(&new_cpus, cs, parent);
3040	nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3041
3042	if (cs->nr_subparts_cpus)
3043		/*
3044		 * Make sure that CPUs allocated to child partitions
3045		 * do not show up in effective_cpus.
3046		 */
3047		cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3048
3049	if (!tmp || !cs->partition_root_state)
3050		goto update_tasks;
3051
3052	/*
3053	 * In the unlikely event that a partition root has empty
3054	 * effective_cpus or its parent becomes erroneous, we have to
3055	 * transition it to the erroneous state.
3056	 */
3057	if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3058	   (parent->partition_root_state == PRS_ERROR))) {
3059		if (cs->nr_subparts_cpus) {
3060			cs->nr_subparts_cpus = 0;
3061			cpumask_clear(cs->subparts_cpus);
3062			compute_effective_cpumask(&new_cpus, cs, parent);
3063		}
3064
3065		/*
3066		 * If the effective_cpus is empty because the child
3067		 * partitions take away all the CPUs, we can keep
3068		 * the current partition and let the child partitions
3069		 * fight for available CPUs.
3070		 */
3071		if ((parent->partition_root_state == PRS_ERROR) ||
3072		     cpumask_empty(&new_cpus)) {
3073			update_parent_subparts_cpumask(cs, partcmd_disable,
3074						       NULL, tmp);
3075			cs->partition_root_state = PRS_ERROR;
3076		}
3077		cpuset_force_rebuild();
3078	}
3079
3080	/*
3081	 * On the other hand, an erroneous partition root may be transitioned
3082	 * back to a regular one or a partition root with no CPU allocated
3083	 * from the parent may change to erroneous.
3084	 */
3085	if (is_partition_root(parent) &&
3086	   ((cs->partition_root_state == PRS_ERROR) ||
3087	    !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3088	     update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3089		cpuset_force_rebuild();
3090
3091update_tasks:
3092	cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3093	mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3094
3095	if (is_in_v2_mode())
3096		hotplug_update_tasks(cs, &new_cpus, &new_mems,
3097				     cpus_updated, mems_updated);
3098	else
3099		hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3100					    cpus_updated, mems_updated);
3101
3102	percpu_up_write(&cpuset_rwsem);
3103}
3104
3105/**
3106 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3107 *
3108 * This function is called after either CPU or memory configuration has
3109 * changed and updates cpuset accordingly.  The top_cpuset is always
3110 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3111 * order to make cpusets transparent (of no affect) on systems that are
3112 * actively using CPU hotplug but making no active use of cpusets.
3113 *
3114 * Non-root cpusets are only affected by offlining.  If any CPUs or memory
3115 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3116 * all descendants.
3117 *
3118 * Note that CPU offlining during suspend is ignored.  We don't modify
3119 * cpusets across suspend/resume cycles at all.
3120 */
3121static void cpuset_hotplug_workfn(struct work_struct *work)
3122{
3123	static cpumask_t new_cpus;
3124	static nodemask_t new_mems;
3125	bool cpus_updated, mems_updated;
3126	bool on_dfl = is_in_v2_mode();
3127	struct tmpmasks tmp, *ptmp = NULL;
3128
3129	if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3130		ptmp = &tmp;
3131
3132	percpu_down_write(&cpuset_rwsem);
3133
3134	/* fetch the available cpus/mems and find out which changed how */
3135	cpumask_copy(&new_cpus, cpu_active_mask);
3136	new_mems = node_states[N_MEMORY];
3137
3138	/*
3139	 * If subparts_cpus is populated, it is likely that the check below
3140	 * will produce a false positive on cpus_updated when the cpu list
3141	 * isn't changed. It is extra work, but it is better to be safe.
3142	 */
3143	cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3144	mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3145
3146	/* synchronize cpus_allowed to cpu_active_mask */
3147	if (cpus_updated) {
3148		spin_lock_irq(&callback_lock);
3149		if (!on_dfl)
3150			cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3151		/*
3152		 * Make sure that CPUs allocated to child partitions
3153		 * do not show up in effective_cpus. If no CPU is left,
3154		 * we clear the subparts_cpus & let the child partitions
3155		 * fight for the CPUs again.
3156		 */
3157		if (top_cpuset.nr_subparts_cpus) {
3158			if (cpumask_subset(&new_cpus,
3159					   top_cpuset.subparts_cpus)) {
3160				top_cpuset.nr_subparts_cpus = 0;
3161				cpumask_clear(top_cpuset.subparts_cpus);
3162			} else {
3163				cpumask_andnot(&new_cpus, &new_cpus,
3164					       top_cpuset.subparts_cpus);
3165			}
3166		}
3167		cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3168		spin_unlock_irq(&callback_lock);
3169		/* we don't mess with cpumasks of tasks in top_cpuset */
3170	}
3171
3172	/* synchronize mems_allowed to N_MEMORY */
3173	if (mems_updated) {
3174		spin_lock_irq(&callback_lock);
3175		if (!on_dfl)
3176			top_cpuset.mems_allowed = new_mems;
3177		top_cpuset.effective_mems = new_mems;
3178		spin_unlock_irq(&callback_lock);
3179		update_tasks_nodemask(&top_cpuset);
3180	}
3181
3182	percpu_up_write(&cpuset_rwsem);
3183
3184	/* if cpus or mems changed, we need to propagate to descendants */
3185	if (cpus_updated || mems_updated) {
3186		struct cpuset *cs;
3187		struct cgroup_subsys_state *pos_css;
3188
3189		rcu_read_lock();
3190		cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3191			if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3192				continue;
3193			rcu_read_unlock();
3194
3195			cpuset_hotplug_update_tasks(cs, ptmp);
3196
3197			rcu_read_lock();
3198			css_put(&cs->css);
3199		}
3200		rcu_read_unlock();
3201	}
3202
3203	/* rebuild sched domains if cpus_allowed has changed */
3204	if (cpus_updated || force_rebuild) {
3205		force_rebuild = false;
3206		rebuild_sched_domains();
3207	}
3208
3209	free_cpumasks(NULL, ptmp);
3210}
3211
3212void cpuset_update_active_cpus(void)
3213{
3214	/*
3215	 * We're inside cpu hotplug critical region which usually nests
3216	 * inside cgroup synchronization.  Bounce actual hotplug processing
3217	 * to a work item to avoid reverse locking order.
3218	 */
3219	schedule_work(&cpuset_hotplug_work);
3220}
3221
3222void cpuset_wait_for_hotplug(void)
3223{
3224	flush_work(&cpuset_hotplug_work);
3225}
3226
3227/*
3228 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3229 * Call this routine anytime after node_states[N_MEMORY] changes.
3230 * See cpuset_update_active_cpus() for CPU hotplug handling.
3231 */
3232static int cpuset_track_online_nodes(struct notifier_block *self,
3233				unsigned long action, void *arg)
3234{
3235	schedule_work(&cpuset_hotplug_work);
3236	return NOTIFY_OK;
3237}
3238
3239static struct notifier_block cpuset_track_online_nodes_nb = {
3240	.notifier_call = cpuset_track_online_nodes,
3241	.priority = 10,		/* ??! */
3242};
3243
3244/**
3245 * cpuset_init_smp - initialize cpus_allowed
3246 *
3247 * Description: Finish top cpuset after cpu, node maps are initialized
3248 */
3249void __init cpuset_init_smp(void)
3250{
3251	cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3252	top_cpuset.mems_allowed = node_states[N_MEMORY];
3253	top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3254
3255	cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3256	top_cpuset.effective_mems = node_states[N_MEMORY];
3257
3258	register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3259
3260	cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3261	BUG_ON(!cpuset_migrate_mm_wq);
3262}
3263
3264/**
3265 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3266 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3267 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3268 *
3269 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3270 * attached to the specified @tsk.  Guaranteed to return some non-empty
3271 * subset of cpu_online_mask, even if this means going outside the
3272 * tasks cpuset.
3273 **/
3274
3275void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3276{
3277	unsigned long flags;
3278
3279	spin_lock_irqsave(&callback_lock, flags);
3280	rcu_read_lock();
3281	guarantee_online_cpus(task_cs(tsk), pmask);
3282	rcu_read_unlock();
3283	spin_unlock_irqrestore(&callback_lock, flags);
3284}
3285
3286/**
3287 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3288 * @tsk: pointer to task_struct with which the scheduler is struggling
3289 *
3290 * Description: In the case that the scheduler cannot find an allowed cpu in
3291 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3292 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3293 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3294 * This is the absolute last resort for the scheduler and it is only used if
3295 * _every_ other avenue has been traveled.
3296 **/
3297
3298void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3299{
3300	rcu_read_lock();
3301	do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3302		task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3303	rcu_read_unlock();
3304
3305	/*
3306	 * We own tsk->cpus_allowed, nobody can change it under us.
3307	 *
3308	 * But we used cs && cs->cpus_allowed lockless and thus can
3309	 * race with cgroup_attach_task() or update_cpumask() and get
3310	 * the wrong tsk->cpus_allowed. However, both cases imply the
3311	 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3312	 * which takes task_rq_lock().
3313	 *
3314	 * If we are called after it dropped the lock we must see all
3315	 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3316	 * set any mask even if it is not right from task_cs() pov,
3317	 * the pending set_cpus_allowed_ptr() will fix things.
3318	 *
3319	 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3320	 * if required.
3321	 */
3322}
3323
3324void __init cpuset_init_current_mems_allowed(void)
3325{
3326	nodes_setall(current->mems_allowed);
3327}
3328
3329/**
3330 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3331 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3332 *
3333 * Description: Returns the nodemask_t mems_allowed of the cpuset
3334 * attached to the specified @tsk.  Guaranteed to return some non-empty
3335 * subset of node_states[N_MEMORY], even if this means going outside the
3336 * tasks cpuset.
3337 **/
3338
3339nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3340{
3341	nodemask_t mask;
3342	unsigned long flags;
3343
3344	spin_lock_irqsave(&callback_lock, flags);
3345	rcu_read_lock();
3346	guarantee_online_mems(task_cs(tsk), &mask);
3347	rcu_read_unlock();
3348	spin_unlock_irqrestore(&callback_lock, flags);
3349
3350	return mask;
3351}
3352
3353/**
3354 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3355 * @nodemask: the nodemask to be checked
3356 *
3357 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3358 */
3359int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3360{
3361	return nodes_intersects(*nodemask, current->mems_allowed);
3362}
3363
3364/*
3365 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3366 * mem_hardwall ancestor to the specified cpuset.  Call holding
3367 * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
3368 * (an unusual configuration), then returns the root cpuset.
3369 */
3370static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3371{
3372	while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3373		cs = parent_cs(cs);
3374	return cs;
3375}
3376
3377/**
3378 * cpuset_node_allowed - Can we allocate on a memory node?
3379 * @node: is this an allowed node?
3380 * @gfp_mask: memory allocation flags
3381 *
3382 * If we're in interrupt, yes, we can always allocate.  If @node is set in
3383 * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
3384 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3385 * yes.  If current has access to memory reserves as an oom victim, yes.
3386 * Otherwise, no.
3387 *
3388 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3389 * and do not allow allocations outside the current tasks cpuset
3390 * unless the task has been OOM killed.
3391 * GFP_KERNEL allocations are not so marked, so can escape to the
3392 * nearest enclosing hardwalled ancestor cpuset.
3393 *
3394 * Scanning up parent cpusets requires callback_lock.  The
3395 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3396 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3397 * current tasks mems_allowed came up empty on the first pass over
3398 * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
3399 * cpuset are short of memory, might require taking the callback_lock.
3400 *
3401 * The first call here from mm/page_alloc:get_page_from_freelist()
3402 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3403 * so no allocation on a node outside the cpuset is allowed (unless
3404 * in interrupt, of course).
3405 *
3406 * The second pass through get_page_from_freelist() doesn't even call
3407 * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
3408 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3409 * in alloc_flags.  That logic and the checks below have the combined
3410 * affect that:
3411 *	in_interrupt - any node ok (current task context irrelevant)
3412 *	GFP_ATOMIC   - any node ok
3413 *	tsk_is_oom_victim   - any node ok
3414 *	GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
3415 *	GFP_USER     - only nodes in current tasks mems allowed ok.
3416 */
3417bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3418{
3419	struct cpuset *cs;		/* current cpuset ancestors */
3420	int allowed;			/* is allocation in zone z allowed? */
3421	unsigned long flags;
3422
3423	if (in_interrupt())
3424		return true;
3425	if (node_isset(node, current->mems_allowed))
3426		return true;
3427	/*
3428	 * Allow tasks that have access to memory reserves because they have
3429	 * been OOM killed to get memory anywhere.
3430	 */
3431	if (unlikely(tsk_is_oom_victim(current)))
3432		return true;
3433	if (gfp_mask & __GFP_HARDWALL)	/* If hardwall request, stop here */
3434		return false;
3435
3436	if (current->flags & PF_EXITING) /* Let dying task have memory */
3437		return true;
3438
3439	/* Not hardwall and node outside mems_allowed: scan up cpusets */
3440	spin_lock_irqsave(&callback_lock, flags);
3441
3442	rcu_read_lock();
3443	cs = nearest_hardwall_ancestor(task_cs(current));
3444	allowed = node_isset(node, cs->mems_allowed);
3445	rcu_read_unlock();
3446
3447	spin_unlock_irqrestore(&callback_lock, flags);
3448	return allowed;
3449}
3450
3451/**
3452 * cpuset_mem_spread_node() - On which node to begin search for a file page
3453 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3454 *
3455 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3456 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3457 * and if the memory allocation used cpuset_mem_spread_node()
3458 * to determine on which node to start looking, as it will for
3459 * certain page cache or slab cache pages such as used for file
3460 * system buffers and inode caches, then instead of starting on the
3461 * local node to look for a free page, rather spread the starting
3462 * node around the tasks mems_allowed nodes.
3463 *
3464 * We don't have to worry about the returned node being offline
3465 * because "it can't happen", and even if it did, it would be ok.
3466 *
3467 * The routines calling guarantee_online_mems() are careful to
3468 * only set nodes in task->mems_allowed that are online.  So it
3469 * should not be possible for the following code to return an
3470 * offline node.  But if it did, that would be ok, as this routine
3471 * is not returning the node where the allocation must be, only
3472 * the node where the search should start.  The zonelist passed to
3473 * __alloc_pages() will include all nodes.  If the slab allocator
3474 * is passed an offline node, it will fall back to the local node.
3475 * See kmem_cache_alloc_node().
3476 */
3477
3478static int cpuset_spread_node(int *rotor)
3479{
3480	return *rotor = next_node_in(*rotor, current->mems_allowed);
3481}
3482
3483int cpuset_mem_spread_node(void)
3484{
3485	if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3486		current->cpuset_mem_spread_rotor =
3487			node_random(&current->mems_allowed);
3488
3489	return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
3490}
3491
3492int cpuset_slab_spread_node(void)
3493{
3494	if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3495		current->cpuset_slab_spread_rotor =
3496			node_random(&current->mems_allowed);
3497
3498	return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
3499}
3500
3501EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3502
3503/**
3504 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3505 * @tsk1: pointer to task_struct of some task.
3506 * @tsk2: pointer to task_struct of some other task.
3507 *
3508 * Description: Return true if @tsk1's mems_allowed intersects the
3509 * mems_allowed of @tsk2.  Used by the OOM killer to determine if
3510 * one of the task's memory usage might impact the memory available
3511 * to the other.
3512 **/
3513
3514int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3515				   const struct task_struct *tsk2)
3516{
3517	return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3518}
3519
3520/**
3521 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3522 *
3523 * Description: Prints current's name, cpuset name, and cached copy of its
3524 * mems_allowed to the kernel log.
3525 */
3526void cpuset_print_current_mems_allowed(void)
3527{
3528	struct cgroup *cgrp;
3529
3530	rcu_read_lock();
3531
3532	cgrp = task_cs(current)->css.cgroup;
3533	pr_cont(",cpuset=");
3534	pr_cont_cgroup_name(cgrp);
3535	pr_cont(",mems_allowed=%*pbl",
3536		nodemask_pr_args(&current->mems_allowed));
3537
3538	rcu_read_unlock();
3539}
3540
3541/*
3542 * Collection of memory_pressure is suppressed unless
3543 * this flag is enabled by writing "1" to the special
3544 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3545 */
3546
3547int cpuset_memory_pressure_enabled __read_mostly;
3548
3549/**
3550 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3551 *
3552 * Keep a running average of the rate of synchronous (direct)
3553 * page reclaim efforts initiated by tasks in each cpuset.
3554 *
3555 * This represents the rate at which some task in the cpuset
3556 * ran low on memory on all nodes it was allowed to use, and
3557 * had to enter the kernels page reclaim code in an effort to
3558 * create more free memory by tossing clean pages or swapping
3559 * or writing dirty pages.
3560 *
3561 * Display to user space in the per-cpuset read-only file
3562 * "memory_pressure".  Value displayed is an integer
3563 * representing the recent rate of entry into the synchronous
3564 * (direct) page reclaim by any task attached to the cpuset.
3565 **/
3566
3567void __cpuset_memory_pressure_bump(void)
3568{
3569	rcu_read_lock();
3570	fmeter_markevent(&task_cs(current)->fmeter);
3571	rcu_read_unlock();
3572}
3573
3574#ifdef CONFIG_PROC_PID_CPUSET
3575/*
3576 * proc_cpuset_show()
3577 *  - Print tasks cpuset path into seq_file.
3578 *  - Used for /proc/<pid>/cpuset.
3579 *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3580 *    doesn't really matter if tsk->cpuset changes after we read it,
3581 *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
3582 *    anyway.
3583 */
3584int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3585		     struct pid *pid, struct task_struct *tsk)
3586{
3587	char *buf;
3588	struct cgroup_subsys_state *css;
3589	int retval;
3590
3591	retval = -ENOMEM;
3592	buf = kmalloc(PATH_MAX, GFP_KERNEL);
3593	if (!buf)
3594		goto out;
3595
3596	css = task_get_css(tsk, cpuset_cgrp_id);
3597	retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3598				current->nsproxy->cgroup_ns);
3599	css_put(css);
3600	if (retval >= PATH_MAX)
3601		retval = -ENAMETOOLONG;
3602	if (retval < 0)
3603		goto out_free;
3604	seq_puts(m, buf);
3605	seq_putc(m, '\n');
3606	retval = 0;
3607out_free:
3608	kfree(buf);
3609out:
3610	return retval;
3611}
3612#endif /* CONFIG_PROC_PID_CPUSET */
3613
3614/* Display task mems_allowed in /proc/<pid>/status file. */
3615void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3616{
3617	seq_printf(m, "Mems_allowed:\t%*pb\n",
3618		   nodemask_pr_args(&task->mems_allowed));
3619	seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3620		   nodemask_pr_args(&task->mems_allowed));
3621}