linux/kernel/cpuset.c
<<
>>
Prefs
   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/module.h>
  41#include <linux/mount.h>
  42#include <linux/namei.h>
  43#include <linux/pagemap.h>
  44#include <linux/proc_fs.h>
  45#include <linux/rcupdate.h>
  46#include <linux/sched.h>
  47#include <linux/seq_file.h>
  48#include <linux/security.h>
  49#include <linux/slab.h>
  50#include <linux/spinlock.h>
  51#include <linux/stat.h>
  52#include <linux/string.h>
  53#include <linux/time.h>
  54#include <linux/backing-dev.h>
  55#include <linux/sort.h>
  56
  57#include <asm/uaccess.h>
  58#include <asm/atomic.h>
  59#include <linux/mutex.h>
  60#include <linux/workqueue.h>
  61#include <linux/cgroup.h>
  62
  63/*
  64 * Tracks how many cpusets are currently defined in system.
  65 * When there is only one cpuset (the root cpuset) we can
  66 * short circuit some hooks.
  67 */
  68int number_of_cpusets __read_mostly;
  69
  70/* Forward declare cgroup structures */
  71struct cgroup_subsys cpuset_subsys;
  72struct cpuset;
  73
  74/* See "Frequency meter" comments, below. */
  75
  76struct fmeter {
  77        int cnt;                /* unprocessed events count */
  78        int val;                /* most recent output value */
  79        time_t time;            /* clock (secs) when val computed */
  80        spinlock_t lock;        /* guards read or write of above */
  81};
  82
  83struct cpuset {
  84        struct cgroup_subsys_state css;
  85
  86        unsigned long flags;            /* "unsigned long" so bitops work */
  87        cpumask_t cpus_allowed;         /* CPUs allowed to tasks in cpuset */
  88        nodemask_t mems_allowed;        /* Memory Nodes allowed to tasks */
  89
  90        struct cpuset *parent;          /* my parent */
  91
  92        /*
  93         * Copy of global cpuset_mems_generation as of the most
  94         * recent time this cpuset changed its mems_allowed.
  95         */
  96        int mems_generation;
  97
  98        struct fmeter fmeter;           /* memory_pressure filter */
  99
 100        /* partition number for rebuild_sched_domains() */
 101        int pn;
 102
 103        /* for custom sched domain */
 104        int relax_domain_level;
 105
 106        /* used for walking a cpuset heirarchy */
 107        struct list_head stack_list;
 108};
 109
 110/* Retrieve the cpuset for a cgroup */
 111static inline struct cpuset *cgroup_cs(struct cgroup *cont)
 112{
 113        return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
 114                            struct cpuset, css);
 115}
 116
 117/* Retrieve the cpuset for a task */
 118static inline struct cpuset *task_cs(struct task_struct *task)
 119{
 120        return container_of(task_subsys_state(task, cpuset_subsys_id),
 121                            struct cpuset, css);
 122}
 123struct cpuset_hotplug_scanner {
 124        struct cgroup_scanner scan;
 125        struct cgroup *to;
 126};
 127
 128/* bits in struct cpuset flags field */
 129typedef enum {
 130        CS_CPU_EXCLUSIVE,
 131        CS_MEM_EXCLUSIVE,
 132        CS_MEM_HARDWALL,
 133        CS_MEMORY_MIGRATE,
 134        CS_SCHED_LOAD_BALANCE,
 135        CS_SPREAD_PAGE,
 136        CS_SPREAD_SLAB,
 137} cpuset_flagbits_t;
 138
 139/* convenient tests for these bits */
 140static inline int is_cpu_exclusive(const struct cpuset *cs)
 141{
 142        return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
 143}
 144
 145static inline int is_mem_exclusive(const struct cpuset *cs)
 146{
 147        return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
 148}
 149
 150static inline int is_mem_hardwall(const struct cpuset *cs)
 151{
 152        return test_bit(CS_MEM_HARDWALL, &cs->flags);
 153}
 154
 155static inline int is_sched_load_balance(const struct cpuset *cs)
 156{
 157        return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
 158}
 159
 160static inline int is_memory_migrate(const struct cpuset *cs)
 161{
 162        return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
 163}
 164
 165static inline int is_spread_page(const struct cpuset *cs)
 166{
 167        return test_bit(CS_SPREAD_PAGE, &cs->flags);
 168}
 169
 170static inline int is_spread_slab(const struct cpuset *cs)
 171{
 172        return test_bit(CS_SPREAD_SLAB, &cs->flags);
 173}
 174
 175/*
 176 * Increment this integer everytime any cpuset changes its
 177 * mems_allowed value.  Users of cpusets can track this generation
 178 * number, and avoid having to lock and reload mems_allowed unless
 179 * the cpuset they're using changes generation.
 180 *
 181 * A single, global generation is needed because cpuset_attach_task() could
 182 * reattach a task to a different cpuset, which must not have its
 183 * generation numbers aliased with those of that tasks previous cpuset.
 184 *
 185 * Generations are needed for mems_allowed because one task cannot
 186 * modify another's memory placement.  So we must enable every task,
 187 * on every visit to __alloc_pages(), to efficiently check whether
 188 * its current->cpuset->mems_allowed has changed, requiring an update
 189 * of its current->mems_allowed.
 190 *
 191 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
 192 * there is no need to mark it atomic.
 193 */
 194static int cpuset_mems_generation;
 195
 196static struct cpuset top_cpuset = {
 197        .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
 198        .cpus_allowed = CPU_MASK_ALL,
 199        .mems_allowed = NODE_MASK_ALL,
 200};
 201
 202/*
 203 * There are two global mutexes guarding cpuset structures.  The first
 204 * is the main control groups cgroup_mutex, accessed via
 205 * cgroup_lock()/cgroup_unlock().  The second is the cpuset-specific
 206 * callback_mutex, below. They can nest.  It is ok to first take
 207 * cgroup_mutex, then nest callback_mutex.  We also require taking
 208 * task_lock() when dereferencing a task's cpuset pointer.  See "The
 209 * task_lock() exception", at the end of this comment.
 210 *
 211 * A task must hold both mutexes to modify cpusets.  If a task
 212 * holds cgroup_mutex, then it blocks others wanting that mutex,
 213 * ensuring that it is the only task able to also acquire callback_mutex
 214 * and be able to modify cpusets.  It can perform various checks on
 215 * the cpuset structure first, knowing nothing will change.  It can
 216 * also allocate memory while just holding cgroup_mutex.  While it is
 217 * performing these checks, various callback routines can briefly
 218 * acquire callback_mutex to query cpusets.  Once it is ready to make
 219 * the changes, it takes callback_mutex, blocking everyone else.
 220 *
 221 * Calls to the kernel memory allocator can not be made while holding
 222 * callback_mutex, as that would risk double tripping on callback_mutex
 223 * from one of the callbacks into the cpuset code from within
 224 * __alloc_pages().
 225 *
 226 * If a task is only holding callback_mutex, then it has read-only
 227 * access to cpusets.
 228 *
 229 * The task_struct fields mems_allowed and mems_generation may only
 230 * be accessed in the context of that task, so require no locks.
 231 *
 232 * The cpuset_common_file_read() handlers only hold callback_mutex across
 233 * small pieces of code, such as when reading out possibly multi-word
 234 * cpumasks and nodemasks.
 235 *
 236 * Accessing a task's cpuset should be done in accordance with the
 237 * guidelines for accessing subsystem state in kernel/cgroup.c
 238 */
 239
 240static DEFINE_MUTEX(callback_mutex);
 241
 242/*
 243 * This is ugly, but preserves the userspace API for existing cpuset
 244 * users. If someone tries to mount the "cpuset" filesystem, we
 245 * silently switch it to mount "cgroup" instead
 246 */
 247static int cpuset_get_sb(struct file_system_type *fs_type,
 248                         int flags, const char *unused_dev_name,
 249                         void *data, struct vfsmount *mnt)
 250{
 251        struct file_system_type *cgroup_fs = get_fs_type("cgroup");
 252        int ret = -ENODEV;
 253        if (cgroup_fs) {
 254                char mountopts[] =
 255                        "cpuset,noprefix,"
 256                        "release_agent=/sbin/cpuset_release_agent";
 257                ret = cgroup_fs->get_sb(cgroup_fs, flags,
 258                                           unused_dev_name, mountopts, mnt);
 259                put_filesystem(cgroup_fs);
 260        }
 261        return ret;
 262}
 263
 264static struct file_system_type cpuset_fs_type = {
 265        .name = "cpuset",
 266        .get_sb = cpuset_get_sb,
 267};
 268
 269/*
 270 * Return in *pmask the portion of a cpusets's cpus_allowed that
 271 * are online.  If none are online, walk up the cpuset hierarchy
 272 * until we find one that does have some online cpus.  If we get
 273 * all the way to the top and still haven't found any online cpus,
 274 * return cpu_online_map.  Or if passed a NULL cs from an exit'ing
 275 * task, return cpu_online_map.
 276 *
 277 * One way or another, we guarantee to return some non-empty subset
 278 * of cpu_online_map.
 279 *
 280 * Call with callback_mutex held.
 281 */
 282
 283static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
 284{
 285        while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
 286                cs = cs->parent;
 287        if (cs)
 288                cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
 289        else
 290                *pmask = cpu_online_map;
 291        BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
 292}
 293
 294/*
 295 * Return in *pmask the portion of a cpusets's mems_allowed that
 296 * are online, with memory.  If none are online with memory, walk
 297 * up the cpuset hierarchy until we find one that does have some
 298 * online mems.  If we get all the way to the top and still haven't
 299 * found any online mems, return node_states[N_HIGH_MEMORY].
 300 *
 301 * One way or another, we guarantee to return some non-empty subset
 302 * of node_states[N_HIGH_MEMORY].
 303 *
 304 * Call with callback_mutex held.
 305 */
 306
 307static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
 308{
 309        while (cs && !nodes_intersects(cs->mems_allowed,
 310                                        node_states[N_HIGH_MEMORY]))
 311                cs = cs->parent;
 312        if (cs)
 313                nodes_and(*pmask, cs->mems_allowed,
 314                                        node_states[N_HIGH_MEMORY]);
 315        else
 316                *pmask = node_states[N_HIGH_MEMORY];
 317        BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
 318}
 319
 320/**
 321 * cpuset_update_task_memory_state - update task memory placement
 322 *
 323 * If the current tasks cpusets mems_allowed changed behind our
 324 * backs, update current->mems_allowed, mems_generation and task NUMA
 325 * mempolicy to the new value.
 326 *
 327 * Task mempolicy is updated by rebinding it relative to the
 328 * current->cpuset if a task has its memory placement changed.
 329 * Do not call this routine if in_interrupt().
 330 *
 331 * Call without callback_mutex or task_lock() held.  May be
 332 * called with or without cgroup_mutex held.  Thanks in part to
 333 * 'the_top_cpuset_hack', the task's cpuset pointer will never
 334 * be NULL.  This routine also might acquire callback_mutex during
 335 * call.
 336 *
 337 * Reading current->cpuset->mems_generation doesn't need task_lock
 338 * to guard the current->cpuset derefence, because it is guarded
 339 * from concurrent freeing of current->cpuset using RCU.
 340 *
 341 * The rcu_dereference() is technically probably not needed,
 342 * as I don't actually mind if I see a new cpuset pointer but
 343 * an old value of mems_generation.  However this really only
 344 * matters on alpha systems using cpusets heavily.  If I dropped
 345 * that rcu_dereference(), it would save them a memory barrier.
 346 * For all other arch's, rcu_dereference is a no-op anyway, and for
 347 * alpha systems not using cpusets, another planned optimization,
 348 * avoiding the rcu critical section for tasks in the root cpuset
 349 * which is statically allocated, so can't vanish, will make this
 350 * irrelevant.  Better to use RCU as intended, than to engage in
 351 * some cute trick to save a memory barrier that is impossible to
 352 * test, for alpha systems using cpusets heavily, which might not
 353 * even exist.
 354 *
 355 * This routine is needed to update the per-task mems_allowed data,
 356 * within the tasks context, when it is trying to allocate memory
 357 * (in various mm/mempolicy.c routines) and notices that some other
 358 * task has been modifying its cpuset.
 359 */
 360
 361void cpuset_update_task_memory_state(void)
 362{
 363        int my_cpusets_mem_gen;
 364        struct task_struct *tsk = current;
 365        struct cpuset *cs;
 366
 367        if (task_cs(tsk) == &top_cpuset) {
 368                /* Don't need rcu for top_cpuset.  It's never freed. */
 369                my_cpusets_mem_gen = top_cpuset.mems_generation;
 370        } else {
 371                rcu_read_lock();
 372                my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
 373                rcu_read_unlock();
 374        }
 375
 376        if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
 377                mutex_lock(&callback_mutex);
 378                task_lock(tsk);
 379                cs = task_cs(tsk); /* Maybe changed when task not locked */
 380                guarantee_online_mems(cs, &tsk->mems_allowed);
 381                tsk->cpuset_mems_generation = cs->mems_generation;
 382                if (is_spread_page(cs))
 383                        tsk->flags |= PF_SPREAD_PAGE;
 384                else
 385                        tsk->flags &= ~PF_SPREAD_PAGE;
 386                if (is_spread_slab(cs))
 387                        tsk->flags |= PF_SPREAD_SLAB;
 388                else
 389                        tsk->flags &= ~PF_SPREAD_SLAB;
 390                task_unlock(tsk);
 391                mutex_unlock(&callback_mutex);
 392                mpol_rebind_task(tsk, &tsk->mems_allowed);
 393        }
 394}
 395
 396/*
 397 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
 398 *
 399 * One cpuset is a subset of another if all its allowed CPUs and
 400 * Memory Nodes are a subset of the other, and its exclusive flags
 401 * are only set if the other's are set.  Call holding cgroup_mutex.
 402 */
 403
 404static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
 405{
 406        return  cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
 407                nodes_subset(p->mems_allowed, q->mems_allowed) &&
 408                is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
 409                is_mem_exclusive(p) <= is_mem_exclusive(q);
 410}
 411
 412/*
 413 * validate_change() - Used to validate that any proposed cpuset change
 414 *                     follows the structural rules for cpusets.
 415 *
 416 * If we replaced the flag and mask values of the current cpuset
 417 * (cur) with those values in the trial cpuset (trial), would
 418 * our various subset and exclusive rules still be valid?  Presumes
 419 * cgroup_mutex held.
 420 *
 421 * 'cur' is the address of an actual, in-use cpuset.  Operations
 422 * such as list traversal that depend on the actual address of the
 423 * cpuset in the list must use cur below, not trial.
 424 *
 425 * 'trial' is the address of bulk structure copy of cur, with
 426 * perhaps one or more of the fields cpus_allowed, mems_allowed,
 427 * or flags changed to new, trial values.
 428 *
 429 * Return 0 if valid, -errno if not.
 430 */
 431
 432static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 433{
 434        struct cgroup *cont;
 435        struct cpuset *c, *par;
 436
 437        /* Each of our child cpusets must be a subset of us */
 438        list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
 439                if (!is_cpuset_subset(cgroup_cs(cont), trial))
 440                        return -EBUSY;
 441        }
 442
 443        /* Remaining checks don't apply to root cpuset */
 444        if (cur == &top_cpuset)
 445                return 0;
 446
 447        par = cur->parent;
 448
 449        /* We must be a subset of our parent cpuset */
 450        if (!is_cpuset_subset(trial, par))
 451                return -EACCES;
 452
 453        /*
 454         * If either I or some sibling (!= me) is exclusive, we can't
 455         * overlap
 456         */
 457        list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
 458                c = cgroup_cs(cont);
 459                if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
 460                    c != cur &&
 461                    cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
 462                        return -EINVAL;
 463                if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
 464                    c != cur &&
 465                    nodes_intersects(trial->mems_allowed, c->mems_allowed))
 466                        return -EINVAL;
 467        }
 468
 469        /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
 470        if (cgroup_task_count(cur->css.cgroup)) {
 471                if (cpus_empty(trial->cpus_allowed) ||
 472                    nodes_empty(trial->mems_allowed)) {
 473                        return -ENOSPC;
 474                }
 475        }
 476
 477        return 0;
 478}
 479
 480/*
 481 * Helper routine for generate_sched_domains().
 482 * Do cpusets a, b have overlapping cpus_allowed masks?
 483 */
 484static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
 485{
 486        return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
 487}
 488
 489static void
 490update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
 491{
 492        if (dattr->relax_domain_level < c->relax_domain_level)
 493                dattr->relax_domain_level = c->relax_domain_level;
 494        return;
 495}
 496
 497static void
 498update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
 499{
 500        LIST_HEAD(q);
 501
 502        list_add(&c->stack_list, &q);
 503        while (!list_empty(&q)) {
 504                struct cpuset *cp;
 505                struct cgroup *cont;
 506                struct cpuset *child;
 507
 508                cp = list_first_entry(&q, struct cpuset, stack_list);
 509                list_del(q.next);
 510
 511                if (cpus_empty(cp->cpus_allowed))
 512                        continue;
 513
 514                if (is_sched_load_balance(cp))
 515                        update_domain_attr(dattr, cp);
 516
 517                list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
 518                        child = cgroup_cs(cont);
 519                        list_add_tail(&child->stack_list, &q);
 520                }
 521        }
 522}
 523
 524/*
 525 * generate_sched_domains()
 526 *
 527 * This function builds a partial partition of the systems CPUs
 528 * A 'partial partition' is a set of non-overlapping subsets whose
 529 * union is a subset of that set.
 530 * The output of this function needs to be passed to kernel/sched.c
 531 * partition_sched_domains() routine, which will rebuild the scheduler's
 532 * load balancing domains (sched domains) as specified by that partial
 533 * partition.
 534 *
 535 * See "What is sched_load_balance" in Documentation/cpusets.txt
 536 * for a background explanation of this.
 537 *
 538 * Does not return errors, on the theory that the callers of this
 539 * routine would rather not worry about failures to rebuild sched
 540 * domains when operating in the severe memory shortage situations
 541 * that could cause allocation failures below.
 542 *
 543 * Must be called with cgroup_lock held.
 544 *
 545 * The three key local variables below are:
 546 *    q  - a linked-list queue of cpuset pointers, used to implement a
 547 *         top-down scan of all cpusets.  This scan loads a pointer
 548 *         to each cpuset marked is_sched_load_balance into the
 549 *         array 'csa'.  For our purposes, rebuilding the schedulers
 550 *         sched domains, we can ignore !is_sched_load_balance cpusets.
 551 *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
 552 *         that need to be load balanced, for convenient iterative
 553 *         access by the subsequent code that finds the best partition,
 554 *         i.e the set of domains (subsets) of CPUs such that the
 555 *         cpus_allowed of every cpuset marked is_sched_load_balance
 556 *         is a subset of one of these domains, while there are as
 557 *         many such domains as possible, each as small as possible.
 558 * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
 559 *         the kernel/sched.c routine partition_sched_domains() in a
 560 *         convenient format, that can be easily compared to the prior
 561 *         value to determine what partition elements (sched domains)
 562 *         were changed (added or removed.)
 563 *
 564 * Finding the best partition (set of domains):
 565 *      The triple nested loops below over i, j, k scan over the
 566 *      load balanced cpusets (using the array of cpuset pointers in
 567 *      csa[]) looking for pairs of cpusets that have overlapping
 568 *      cpus_allowed, but which don't have the same 'pn' partition
 569 *      number and gives them in the same partition number.  It keeps
 570 *      looping on the 'restart' label until it can no longer find
 571 *      any such pairs.
 572 *
 573 *      The union of the cpus_allowed masks from the set of
 574 *      all cpusets having the same 'pn' value then form the one
 575 *      element of the partition (one sched domain) to be passed to
 576 *      partition_sched_domains().
 577 */
 578static int generate_sched_domains(cpumask_t **domains,
 579                        struct sched_domain_attr **attributes)
 580{
 581        LIST_HEAD(q);           /* queue of cpusets to be scanned */
 582        struct cpuset *cp;      /* scans q */
 583        struct cpuset **csa;    /* array of all cpuset ptrs */
 584        int csn;                /* how many cpuset ptrs in csa so far */
 585        int i, j, k;            /* indices for partition finding loops */
 586        cpumask_t *doms;        /* resulting partition; i.e. sched domains */
 587        struct sched_domain_attr *dattr;  /* attributes for custom domains */
 588        int ndoms = 0;          /* number of sched domains in result */
 589        int nslot;              /* next empty doms[] cpumask_t slot */
 590
 591        doms = NULL;
 592        dattr = NULL;
 593        csa = NULL;
 594
 595        /* Special case for the 99% of systems with one, full, sched domain */
 596        if (is_sched_load_balance(&top_cpuset)) {
 597                doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
 598                if (!doms)
 599                        goto done;
 600
 601                dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
 602                if (dattr) {
 603                        *dattr = SD_ATTR_INIT;
 604                        update_domain_attr_tree(dattr, &top_cpuset);
 605                }
 606                *doms = top_cpuset.cpus_allowed;
 607
 608                ndoms = 1;
 609                goto done;
 610        }
 611
 612        csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
 613        if (!csa)
 614                goto done;
 615        csn = 0;
 616
 617        list_add(&top_cpuset.stack_list, &q);
 618        while (!list_empty(&q)) {
 619                struct cgroup *cont;
 620                struct cpuset *child;   /* scans child cpusets of cp */
 621
 622                cp = list_first_entry(&q, struct cpuset, stack_list);
 623                list_del(q.next);
 624
 625                if (cpus_empty(cp->cpus_allowed))
 626                        continue;
 627
 628                /*
 629                 * All child cpusets contain a subset of the parent's cpus, so
 630                 * just skip them, and then we call update_domain_attr_tree()
 631                 * to calc relax_domain_level of the corresponding sched
 632                 * domain.
 633                 */
 634                if (is_sched_load_balance(cp)) {
 635                        csa[csn++] = cp;
 636                        continue;
 637                }
 638
 639                list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
 640                        child = cgroup_cs(cont);
 641                        list_add_tail(&child->stack_list, &q);
 642                }
 643        }
 644
 645        for (i = 0; i < csn; i++)
 646                csa[i]->pn = i;
 647        ndoms = csn;
 648
 649restart:
 650        /* Find the best partition (set of sched domains) */
 651        for (i = 0; i < csn; i++) {
 652                struct cpuset *a = csa[i];
 653                int apn = a->pn;
 654
 655                for (j = 0; j < csn; j++) {
 656                        struct cpuset *b = csa[j];
 657                        int bpn = b->pn;
 658
 659                        if (apn != bpn && cpusets_overlap(a, b)) {
 660                                for (k = 0; k < csn; k++) {
 661                                        struct cpuset *c = csa[k];
 662
 663                                        if (c->pn == bpn)
 664                                                c->pn = apn;
 665                                }
 666                                ndoms--;        /* one less element */
 667                                goto restart;
 668                        }
 669                }
 670        }
 671
 672        /*
 673         * Now we know how many domains to create.
 674         * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
 675         */
 676        doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
 677        if (!doms)
 678                goto done;
 679
 680        /*
 681         * The rest of the code, including the scheduler, can deal with
 682         * dattr==NULL case. No need to abort if alloc fails.
 683         */
 684        dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
 685
 686        for (nslot = 0, i = 0; i < csn; i++) {
 687                struct cpuset *a = csa[i];
 688                cpumask_t *dp;
 689                int apn = a->pn;
 690
 691                if (apn < 0) {
 692                        /* Skip completed partitions */
 693                        continue;
 694                }
 695
 696                dp = doms + nslot;
 697
 698                if (nslot == ndoms) {
 699                        static int warnings = 10;
 700                        if (warnings) {
 701                                printk(KERN_WARNING
 702                                 "rebuild_sched_domains confused:"
 703                                  " nslot %d, ndoms %d, csn %d, i %d,"
 704                                  " apn %d\n",
 705                                  nslot, ndoms, csn, i, apn);
 706                                warnings--;
 707                        }
 708                        continue;
 709                }
 710
 711                cpus_clear(*dp);
 712                if (dattr)
 713                        *(dattr + nslot) = SD_ATTR_INIT;
 714                for (j = i; j < csn; j++) {
 715                        struct cpuset *b = csa[j];
 716
 717                        if (apn == b->pn) {
 718                                cpus_or(*dp, *dp, b->cpus_allowed);
 719                                if (dattr)
 720                                        update_domain_attr_tree(dattr + nslot, b);
 721
 722                                /* Done with this partition */
 723                                b->pn = -1;
 724                        }
 725                }
 726                nslot++;
 727        }
 728        BUG_ON(nslot != ndoms);
 729
 730done:
 731        kfree(csa);
 732
 733        /*
 734         * Fallback to the default domain if kmalloc() failed.
 735         * See comments in partition_sched_domains().
 736         */
 737        if (doms == NULL)
 738                ndoms = 1;
 739
 740        *domains    = doms;
 741        *attributes = dattr;
 742        return ndoms;
 743}
 744
 745/*
 746 * Rebuild scheduler domains.
 747 *
 748 * Call with neither cgroup_mutex held nor within get_online_cpus().
 749 * Takes both cgroup_mutex and get_online_cpus().
 750 *
 751 * Cannot be directly called from cpuset code handling changes
 752 * to the cpuset pseudo-filesystem, because it cannot be called
 753 * from code that already holds cgroup_mutex.
 754 */
 755static void do_rebuild_sched_domains(struct work_struct *unused)
 756{
 757        struct sched_domain_attr *attr;
 758        cpumask_t *doms;
 759        int ndoms;
 760
 761        get_online_cpus();
 762
 763        /* Generate domain masks and attrs */
 764        cgroup_lock();
 765        ndoms = generate_sched_domains(&doms, &attr);
 766        cgroup_unlock();
 767
 768        /* Have scheduler rebuild the domains */
 769        partition_sched_domains(ndoms, doms, attr);
 770
 771        put_online_cpus();
 772}
 773
 774static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
 775
 776/*
 777 * Rebuild scheduler domains, asynchronously via workqueue.
 778 *
 779 * If the flag 'sched_load_balance' of any cpuset with non-empty
 780 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
 781 * which has that flag enabled, or if any cpuset with a non-empty
 782 * 'cpus' is removed, then call this routine to rebuild the
 783 * scheduler's dynamic sched domains.
 784 *
 785 * The rebuild_sched_domains() and partition_sched_domains()
 786 * routines must nest cgroup_lock() inside get_online_cpus(),
 787 * but such cpuset changes as these must nest that locking the
 788 * other way, holding cgroup_lock() for much of the code.
 789 *
 790 * So in order to avoid an ABBA deadlock, the cpuset code handling
 791 * these user changes delegates the actual sched domain rebuilding
 792 * to a separate workqueue thread, which ends up processing the
 793 * above do_rebuild_sched_domains() function.
 794 */
 795static void async_rebuild_sched_domains(void)
 796{
 797        schedule_work(&rebuild_sched_domains_work);
 798}
 799
 800/*
 801 * Accomplishes the same scheduler domain rebuild as the above
 802 * async_rebuild_sched_domains(), however it directly calls the
 803 * rebuild routine synchronously rather than calling it via an
 804 * asynchronous work thread.
 805 *
 806 * This can only be called from code that is not holding
 807 * cgroup_mutex (not nested in a cgroup_lock() call.)
 808 */
 809void rebuild_sched_domains(void)
 810{
 811        do_rebuild_sched_domains(NULL);
 812}
 813
 814/**
 815 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
 816 * @tsk: task to test
 817 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
 818 *
 819 * Call with cgroup_mutex held.  May take callback_mutex during call.
 820 * Called for each task in a cgroup by cgroup_scan_tasks().
 821 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
 822 * words, if its mask is not equal to its cpuset's mask).
 823 */
 824static int cpuset_test_cpumask(struct task_struct *tsk,
 825                               struct cgroup_scanner *scan)
 826{
 827        return !cpus_equal(tsk->cpus_allowed,
 828                        (cgroup_cs(scan->cg))->cpus_allowed);
 829}
 830
 831/**
 832 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
 833 * @tsk: task to test
 834 * @scan: struct cgroup_scanner containing the cgroup of the task
 835 *
 836 * Called by cgroup_scan_tasks() for each task in a cgroup whose
 837 * cpus_allowed mask needs to be changed.
 838 *
 839 * We don't need to re-check for the cgroup/cpuset membership, since we're
 840 * holding cgroup_lock() at this point.
 841 */
 842static void cpuset_change_cpumask(struct task_struct *tsk,
 843                                  struct cgroup_scanner *scan)
 844{
 845        set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed));
 846}
 847
 848/**
 849 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
 850 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
 851 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
 852 *
 853 * Called with cgroup_mutex held
 854 *
 855 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
 856 * calling callback functions for each.
 857 *
 858 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
 859 * if @heap != NULL.
 860 */
 861static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 862{
 863        struct cgroup_scanner scan;
 864
 865        scan.cg = cs->css.cgroup;
 866        scan.test_task = cpuset_test_cpumask;
 867        scan.process_task = cpuset_change_cpumask;
 868        scan.heap = heap;
 869        cgroup_scan_tasks(&scan);
 870}
 871
 872/**
 873 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
 874 * @cs: the cpuset to consider
 875 * @buf: buffer of cpu numbers written to this cpuset
 876 */
 877static int update_cpumask(struct cpuset *cs, const char *buf)
 878{
 879        struct ptr_heap heap;
 880        struct cpuset trialcs;
 881        int retval;
 882        int is_load_balanced;
 883
 884        /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
 885        if (cs == &top_cpuset)
 886                return -EACCES;
 887
 888        trialcs = *cs;
 889
 890        /*
 891         * An empty cpus_allowed is ok only if the cpuset has no tasks.
 892         * Since cpulist_parse() fails on an empty mask, we special case
 893         * that parsing.  The validate_change() call ensures that cpusets
 894         * with tasks have cpus.
 895         */
 896        if (!*buf) {
 897                cpus_clear(trialcs.cpus_allowed);
 898        } else {
 899                retval = cpulist_parse(buf, trialcs.cpus_allowed);
 900                if (retval < 0)
 901                        return retval;
 902
 903                if (!cpus_subset(trialcs.cpus_allowed, cpu_online_map))
 904                        return -EINVAL;
 905        }
 906        retval = validate_change(cs, &trialcs);
 907        if (retval < 0)
 908                return retval;
 909
 910        /* Nothing to do if the cpus didn't change */
 911        if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
 912                return 0;
 913
 914        retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
 915        if (retval)
 916                return retval;
 917
 918        is_load_balanced = is_sched_load_balance(&trialcs);
 919
 920        mutex_lock(&callback_mutex);
 921        cs->cpus_allowed = trialcs.cpus_allowed;
 922        mutex_unlock(&callback_mutex);
 923
 924        /*
 925         * Scan tasks in the cpuset, and update the cpumasks of any
 926         * that need an update.
 927         */
 928        update_tasks_cpumask(cs, &heap);
 929
 930        heap_free(&heap);
 931
 932        if (is_load_balanced)
 933                async_rebuild_sched_domains();
 934        return 0;
 935}
 936
 937/*
 938 * cpuset_migrate_mm
 939 *
 940 *    Migrate memory region from one set of nodes to another.
 941 *
 942 *    Temporarilly set tasks mems_allowed to target nodes of migration,
 943 *    so that the migration code can allocate pages on these nodes.
 944 *
 945 *    Call holding cgroup_mutex, so current's cpuset won't change
 946 *    during this call, as manage_mutex holds off any cpuset_attach()
 947 *    calls.  Therefore we don't need to take task_lock around the
 948 *    call to guarantee_online_mems(), as we know no one is changing
 949 *    our task's cpuset.
 950 *
 951 *    Hold callback_mutex around the two modifications of our tasks
 952 *    mems_allowed to synchronize with cpuset_mems_allowed().
 953 *
 954 *    While the mm_struct we are migrating is typically from some
 955 *    other task, the task_struct mems_allowed that we are hacking
 956 *    is for our current task, which must allocate new pages for that
 957 *    migrating memory region.
 958 *
 959 *    We call cpuset_update_task_memory_state() before hacking
 960 *    our tasks mems_allowed, so that we are assured of being in
 961 *    sync with our tasks cpuset, and in particular, callbacks to
 962 *    cpuset_update_task_memory_state() from nested page allocations
 963 *    won't see any mismatch of our cpuset and task mems_generation
 964 *    values, so won't overwrite our hacked tasks mems_allowed
 965 *    nodemask.
 966 */
 967
 968static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
 969                                                        const nodemask_t *to)
 970{
 971        struct task_struct *tsk = current;
 972
 973        cpuset_update_task_memory_state();
 974
 975        mutex_lock(&callback_mutex);
 976        tsk->mems_allowed = *to;
 977        mutex_unlock(&callback_mutex);
 978
 979        do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
 980
 981        mutex_lock(&callback_mutex);
 982        guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
 983        mutex_unlock(&callback_mutex);
 984}
 985
 986static void *cpuset_being_rebound;
 987
 988/**
 989 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
 990 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
 991 * @oldmem: old mems_allowed of cpuset cs
 992 *
 993 * Called with cgroup_mutex held
 994 * Return 0 if successful, -errno if not.
 995 */
 996static int update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem)
 997{
 998        struct task_struct *p;
 999        struct mm_struct **mmarray;
1000        int i, n, ntasks;
1001        int migrate;
1002        int fudge;
1003        struct cgroup_iter it;
1004        int retval;
1005
1006        cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1007
1008        fudge = 10;                             /* spare mmarray[] slots */
1009        fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
1010        retval = -ENOMEM;
1011
1012        /*
1013         * Allocate mmarray[] to hold mm reference for each task
1014         * in cpuset cs.  Can't kmalloc GFP_KERNEL while holding
1015         * tasklist_lock.  We could use GFP_ATOMIC, but with a
1016         * few more lines of code, we can retry until we get a big
1017         * enough mmarray[] w/o using GFP_ATOMIC.
1018         */
1019        while (1) {
1020                ntasks = cgroup_task_count(cs->css.cgroup);  /* guess */
1021                ntasks += fudge;
1022                mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
1023                if (!mmarray)
1024                        goto done;
1025                read_lock(&tasklist_lock);              /* block fork */
1026                if (cgroup_task_count(cs->css.cgroup) <= ntasks)
1027                        break;                          /* got enough */
1028                read_unlock(&tasklist_lock);            /* try again */
1029                kfree(mmarray);
1030        }
1031
1032        n = 0;
1033
1034        /* Load up mmarray[] with mm reference for each task in cpuset. */
1035        cgroup_iter_start(cs->css.cgroup, &it);
1036        while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
1037                struct mm_struct *mm;
1038
1039                if (n >= ntasks) {
1040                        printk(KERN_WARNING
1041                                "Cpuset mempolicy rebind incomplete.\n");
1042                        break;
1043                }
1044                mm = get_task_mm(p);
1045                if (!mm)
1046                        continue;
1047                mmarray[n++] = mm;
1048        }
1049        cgroup_iter_end(cs->css.cgroup, &it);
1050        read_unlock(&tasklist_lock);
1051
1052        /*
1053         * Now that we've dropped the tasklist spinlock, we can
1054         * rebind the vma mempolicies of each mm in mmarray[] to their
1055         * new cpuset, and release that mm.  The mpol_rebind_mm()
1056         * call takes mmap_sem, which we couldn't take while holding
1057         * tasklist_lock.  Forks can happen again now - the mpol_dup()
1058         * cpuset_being_rebound check will catch such forks, and rebind
1059         * their vma mempolicies too.  Because we still hold the global
1060         * cgroup_mutex, we know that no other rebind effort will
1061         * be contending for the global variable cpuset_being_rebound.
1062         * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1063         * is idempotent.  Also migrate pages in each mm to new nodes.
1064         */
1065        migrate = is_memory_migrate(cs);
1066        for (i = 0; i < n; i++) {
1067                struct mm_struct *mm = mmarray[i];
1068
1069                mpol_rebind_mm(mm, &cs->mems_allowed);
1070                if (migrate)
1071                        cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1072                mmput(mm);
1073        }
1074
1075        /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1076        kfree(mmarray);
1077        cpuset_being_rebound = NULL;
1078        retval = 0;
1079done:
1080        return retval;
1081}
1082
1083/*
1084 * Handle user request to change the 'mems' memory placement
1085 * of a cpuset.  Needs to validate the request, update the
1086 * cpusets mems_allowed and mems_generation, and for each
1087 * task in the cpuset, rebind any vma mempolicies and if
1088 * the cpuset is marked 'memory_migrate', migrate the tasks
1089 * pages to the new memory.
1090 *
1091 * Call with cgroup_mutex held.  May take callback_mutex during call.
1092 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1093 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1094 * their mempolicies to the cpusets new mems_allowed.
1095 */
1096static int update_nodemask(struct cpuset *cs, const char *buf)
1097{
1098        struct cpuset trialcs;
1099        nodemask_t oldmem;
1100        int retval;
1101
1102        /*
1103         * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1104         * it's read-only
1105         */
1106        if (cs == &top_cpuset)
1107                return -EACCES;
1108
1109        trialcs = *cs;
1110
1111        /*
1112         * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1113         * Since nodelist_parse() fails on an empty mask, we special case
1114         * that parsing.  The validate_change() call ensures that cpusets
1115         * with tasks have memory.
1116         */
1117        if (!*buf) {
1118                nodes_clear(trialcs.mems_allowed);
1119        } else {
1120                retval = nodelist_parse(buf, trialcs.mems_allowed);
1121                if (retval < 0)
1122                        goto done;
1123
1124                if (!nodes_subset(trialcs.mems_allowed,
1125                                node_states[N_HIGH_MEMORY]))
1126                        return -EINVAL;
1127        }
1128        oldmem = cs->mems_allowed;
1129        if (nodes_equal(oldmem, trialcs.mems_allowed)) {
1130                retval = 0;             /* Too easy - nothing to do */
1131                goto done;
1132        }
1133        retval = validate_change(cs, &trialcs);
1134        if (retval < 0)
1135                goto done;
1136
1137        mutex_lock(&callback_mutex);
1138        cs->mems_allowed = trialcs.mems_allowed;
1139        cs->mems_generation = cpuset_mems_generation++;
1140        mutex_unlock(&callback_mutex);
1141
1142        retval = update_tasks_nodemask(cs, &oldmem);
1143done:
1144        return retval;
1145}
1146
1147int current_cpuset_is_being_rebound(void)
1148{
1149        return task_cs(current) == cpuset_being_rebound;
1150}
1151
1152static int update_relax_domain_level(struct cpuset *cs, s64 val)
1153{
1154        if (val < -1 || val >= SD_LV_MAX)
1155                return -EINVAL;
1156
1157        if (val != cs->relax_domain_level) {
1158                cs->relax_domain_level = val;
1159                if (!cpus_empty(cs->cpus_allowed) && is_sched_load_balance(cs))
1160                        async_rebuild_sched_domains();
1161        }
1162
1163        return 0;
1164}
1165
1166/*
1167 * update_flag - read a 0 or a 1 in a file and update associated flag
1168 * bit:         the bit to update (see cpuset_flagbits_t)
1169 * cs:          the cpuset to update
1170 * turning_on:  whether the flag is being set or cleared
1171 *
1172 * Call with cgroup_mutex held.
1173 */
1174
1175static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1176                       int turning_on)
1177{
1178        struct cpuset trialcs;
1179        int err;
1180        int balance_flag_changed;
1181
1182        trialcs = *cs;
1183        if (turning_on)
1184                set_bit(bit, &trialcs.flags);
1185        else
1186                clear_bit(bit, &trialcs.flags);
1187
1188        err = validate_change(cs, &trialcs);
1189        if (err < 0)
1190                return err;
1191
1192        balance_flag_changed = (is_sched_load_balance(cs) !=
1193                                        is_sched_load_balance(&trialcs));
1194
1195        mutex_lock(&callback_mutex);
1196        cs->flags = trialcs.flags;
1197        mutex_unlock(&callback_mutex);
1198
1199        if (!cpus_empty(trialcs.cpus_allowed) && balance_flag_changed)
1200                async_rebuild_sched_domains();
1201
1202        return 0;
1203}
1204
1205/*
1206 * Frequency meter - How fast is some event occurring?
1207 *
1208 * These routines manage a digitally filtered, constant time based,
1209 * event frequency meter.  There are four routines:
1210 *   fmeter_init() - initialize a frequency meter.
1211 *   fmeter_markevent() - called each time the event happens.
1212 *   fmeter_getrate() - returns the recent rate of such events.
1213 *   fmeter_update() - internal routine used to update fmeter.
1214 *
1215 * A common data structure is passed to each of these routines,
1216 * which is used to keep track of the state required to manage the
1217 * frequency meter and its digital filter.
1218 *
1219 * The filter works on the number of events marked per unit time.
1220 * The filter is single-pole low-pass recursive (IIR).  The time unit
1221 * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1222 * simulate 3 decimal digits of precision (multiplied by 1000).
1223 *
1224 * With an FM_COEF of 933, and a time base of 1 second, the filter
1225 * has a half-life of 10 seconds, meaning that if the events quit
1226 * happening, then the rate returned from the fmeter_getrate()
1227 * will be cut in half each 10 seconds, until it converges to zero.
1228 *
1229 * It is not worth doing a real infinitely recursive filter.  If more
1230 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1231 * just compute FM_MAXTICKS ticks worth, by which point the level
1232 * will be stable.
1233 *
1234 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1235 * arithmetic overflow in the fmeter_update() routine.
1236 *
1237 * Given the simple 32 bit integer arithmetic used, this meter works
1238 * best for reporting rates between one per millisecond (msec) and
1239 * one per 32 (approx) seconds.  At constant rates faster than one
1240 * per msec it maxes out at values just under 1,000,000.  At constant
1241 * rates between one per msec, and one per second it will stabilize
1242 * to a value N*1000, where N is the rate of events per second.
1243 * At constant rates between one per second and one per 32 seconds,
1244 * it will be choppy, moving up on the seconds that have an event,
1245 * and then decaying until the next event.  At rates slower than
1246 * about one in 32 seconds, it decays all the way back to zero between
1247 * each event.
1248 */
1249
1250#define FM_COEF 933             /* coefficient for half-life of 10 secs */
1251#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1252#define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1253#define FM_SCALE 1000           /* faux fixed point scale */
1254
1255/* Initialize a frequency meter */
1256static void fmeter_init(struct fmeter *fmp)
1257{
1258        fmp->cnt = 0;
1259        fmp->val = 0;
1260        fmp->time = 0;
1261        spin_lock_init(&fmp->lock);
1262}
1263
1264/* Internal meter update - process cnt events and update value */
1265static void fmeter_update(struct fmeter *fmp)
1266{
1267        time_t now = get_seconds();
1268        time_t ticks = now - fmp->time;
1269
1270        if (ticks == 0)
1271                return;
1272
1273        ticks = min(FM_MAXTICKS, ticks);
1274        while (ticks-- > 0)
1275                fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1276        fmp->time = now;
1277
1278        fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1279        fmp->cnt = 0;
1280}
1281
1282/* Process any previous ticks, then bump cnt by one (times scale). */
1283static void fmeter_markevent(struct fmeter *fmp)
1284{
1285        spin_lock(&fmp->lock);
1286        fmeter_update(fmp);
1287        fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1288        spin_unlock(&fmp->lock);
1289}
1290
1291/* Process any previous ticks, then return current value. */
1292static int fmeter_getrate(struct fmeter *fmp)
1293{
1294        int val;
1295
1296        spin_lock(&fmp->lock);
1297        fmeter_update(fmp);
1298        val = fmp->val;
1299        spin_unlock(&fmp->lock);
1300        return val;
1301}
1302
1303/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1304static int cpuset_can_attach(struct cgroup_subsys *ss,
1305                             struct cgroup *cont, struct task_struct *tsk)
1306{
1307        struct cpuset *cs = cgroup_cs(cont);
1308
1309        if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1310                return -ENOSPC;
1311        if (tsk->flags & PF_THREAD_BOUND) {
1312                cpumask_t mask;
1313
1314                mutex_lock(&callback_mutex);
1315                mask = cs->cpus_allowed;
1316                mutex_unlock(&callback_mutex);
1317                if (!cpus_equal(tsk->cpus_allowed, mask))
1318                        return -EINVAL;
1319        }
1320
1321        return security_task_setscheduler(tsk, 0, NULL);
1322}
1323
1324static void cpuset_attach(struct cgroup_subsys *ss,
1325                          struct cgroup *cont, struct cgroup *oldcont,
1326                          struct task_struct *tsk)
1327{
1328        cpumask_t cpus;
1329        nodemask_t from, to;
1330        struct mm_struct *mm;
1331        struct cpuset *cs = cgroup_cs(cont);
1332        struct cpuset *oldcs = cgroup_cs(oldcont);
1333        int err;
1334
1335        mutex_lock(&callback_mutex);
1336        guarantee_online_cpus(cs, &cpus);
1337        err = set_cpus_allowed_ptr(tsk, &cpus);
1338        mutex_unlock(&callback_mutex);
1339        if (err)
1340                return;
1341
1342        from = oldcs->mems_allowed;
1343        to = cs->mems_allowed;
1344        mm = get_task_mm(tsk);
1345        if (mm) {
1346                mpol_rebind_mm(mm, &to);
1347                if (is_memory_migrate(cs))
1348                        cpuset_migrate_mm(mm, &from, &to);
1349                mmput(mm);
1350        }
1351
1352}
1353
1354/* The various types of files and directories in a cpuset file system */
1355
1356typedef enum {
1357        FILE_MEMORY_MIGRATE,
1358        FILE_CPULIST,
1359        FILE_MEMLIST,
1360        FILE_CPU_EXCLUSIVE,
1361        FILE_MEM_EXCLUSIVE,
1362        FILE_MEM_HARDWALL,
1363        FILE_SCHED_LOAD_BALANCE,
1364        FILE_SCHED_RELAX_DOMAIN_LEVEL,
1365        FILE_MEMORY_PRESSURE_ENABLED,
1366        FILE_MEMORY_PRESSURE,
1367        FILE_SPREAD_PAGE,
1368        FILE_SPREAD_SLAB,
1369} cpuset_filetype_t;
1370
1371static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1372{
1373        int retval = 0;
1374        struct cpuset *cs = cgroup_cs(cgrp);
1375        cpuset_filetype_t type = cft->private;
1376
1377        if (!cgroup_lock_live_group(cgrp))
1378                return -ENODEV;
1379
1380        switch (type) {
1381        case FILE_CPU_EXCLUSIVE:
1382                retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1383                break;
1384        case FILE_MEM_EXCLUSIVE:
1385                retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1386                break;
1387        case FILE_MEM_HARDWALL:
1388                retval = update_flag(CS_MEM_HARDWALL, cs, val);
1389                break;
1390        case FILE_SCHED_LOAD_BALANCE:
1391                retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1392                break;
1393        case FILE_MEMORY_MIGRATE:
1394                retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1395                break;
1396        case FILE_MEMORY_PRESSURE_ENABLED:
1397                cpuset_memory_pressure_enabled = !!val;
1398                break;
1399        case FILE_MEMORY_PRESSURE:
1400                retval = -EACCES;
1401                break;
1402        case FILE_SPREAD_PAGE:
1403                retval = update_flag(CS_SPREAD_PAGE, cs, val);
1404                cs->mems_generation = cpuset_mems_generation++;
1405                break;
1406        case FILE_SPREAD_SLAB:
1407                retval = update_flag(CS_SPREAD_SLAB, cs, val);
1408                cs->mems_generation = cpuset_mems_generation++;
1409                break;
1410        default:
1411                retval = -EINVAL;
1412                break;
1413        }
1414        cgroup_unlock();
1415        return retval;
1416}
1417
1418static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1419{
1420        int retval = 0;
1421        struct cpuset *cs = cgroup_cs(cgrp);
1422        cpuset_filetype_t type = cft->private;
1423
1424        if (!cgroup_lock_live_group(cgrp))
1425                return -ENODEV;
1426
1427        switch (type) {
1428        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1429                retval = update_relax_domain_level(cs, val);
1430                break;
1431        default:
1432                retval = -EINVAL;
1433                break;
1434        }
1435        cgroup_unlock();
1436        return retval;
1437}
1438
1439/*
1440 * Common handling for a write to a "cpus" or "mems" file.
1441 */
1442static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1443                                const char *buf)
1444{
1445        int retval = 0;
1446
1447        if (!cgroup_lock_live_group(cgrp))
1448                return -ENODEV;
1449
1450        switch (cft->private) {
1451        case FILE_CPULIST:
1452                retval = update_cpumask(cgroup_cs(cgrp), buf);
1453                break;
1454        case FILE_MEMLIST:
1455                retval = update_nodemask(cgroup_cs(cgrp), buf);
1456                break;
1457        default:
1458                retval = -EINVAL;
1459                break;
1460        }
1461        cgroup_unlock();
1462        return retval;
1463}
1464
1465/*
1466 * These ascii lists should be read in a single call, by using a user
1467 * buffer large enough to hold the entire map.  If read in smaller
1468 * chunks, there is no guarantee of atomicity.  Since the display format
1469 * used, list of ranges of sequential numbers, is variable length,
1470 * and since these maps can change value dynamically, one could read
1471 * gibberish by doing partial reads while a list was changing.
1472 * A single large read to a buffer that crosses a page boundary is
1473 * ok, because the result being copied to user land is not recomputed
1474 * across a page fault.
1475 */
1476
1477static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1478{
1479        cpumask_t mask;
1480
1481        mutex_lock(&callback_mutex);
1482        mask = cs->cpus_allowed;
1483        mutex_unlock(&callback_mutex);
1484
1485        return cpulist_scnprintf(page, PAGE_SIZE, mask);
1486}
1487
1488static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1489{
1490        nodemask_t mask;
1491
1492        mutex_lock(&callback_mutex);
1493        mask = cs->mems_allowed;
1494        mutex_unlock(&callback_mutex);
1495
1496        return nodelist_scnprintf(page, PAGE_SIZE, mask);
1497}
1498
1499static ssize_t cpuset_common_file_read(struct cgroup *cont,
1500                                       struct cftype *cft,
1501                                       struct file *file,
1502                                       char __user *buf,
1503                                       size_t nbytes, loff_t *ppos)
1504{
1505        struct cpuset *cs = cgroup_cs(cont);
1506        cpuset_filetype_t type = cft->private;
1507        char *page;
1508        ssize_t retval = 0;
1509        char *s;
1510
1511        if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1512                return -ENOMEM;
1513
1514        s = page;
1515
1516        switch (type) {
1517        case FILE_CPULIST:
1518                s += cpuset_sprintf_cpulist(s, cs);
1519                break;
1520        case FILE_MEMLIST:
1521                s += cpuset_sprintf_memlist(s, cs);
1522                break;
1523        default:
1524                retval = -EINVAL;
1525                goto out;
1526        }
1527        *s++ = '\n';
1528
1529        retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1530out:
1531        free_page((unsigned long)page);
1532        return retval;
1533}
1534
1535static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1536{
1537        struct cpuset *cs = cgroup_cs(cont);
1538        cpuset_filetype_t type = cft->private;
1539        switch (type) {
1540        case FILE_CPU_EXCLUSIVE:
1541                return is_cpu_exclusive(cs);
1542        case FILE_MEM_EXCLUSIVE:
1543                return is_mem_exclusive(cs);
1544        case FILE_MEM_HARDWALL:
1545                return is_mem_hardwall(cs);
1546        case FILE_SCHED_LOAD_BALANCE:
1547                return is_sched_load_balance(cs);
1548        case FILE_MEMORY_MIGRATE:
1549                return is_memory_migrate(cs);
1550        case FILE_MEMORY_PRESSURE_ENABLED:
1551                return cpuset_memory_pressure_enabled;
1552        case FILE_MEMORY_PRESSURE:
1553                return fmeter_getrate(&cs->fmeter);
1554        case FILE_SPREAD_PAGE:
1555                return is_spread_page(cs);
1556        case FILE_SPREAD_SLAB:
1557                return is_spread_slab(cs);
1558        default:
1559                BUG();
1560        }
1561
1562        /* Unreachable but makes gcc happy */
1563        return 0;
1564}
1565
1566static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1567{
1568        struct cpuset *cs = cgroup_cs(cont);
1569        cpuset_filetype_t type = cft->private;
1570        switch (type) {
1571        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1572                return cs->relax_domain_level;
1573        default:
1574                BUG();
1575        }
1576
1577        /* Unrechable but makes gcc happy */
1578        return 0;
1579}
1580
1581
1582/*
1583 * for the common functions, 'private' gives the type of file
1584 */
1585
1586static struct cftype files[] = {
1587        {
1588                .name = "cpus",
1589                .read = cpuset_common_file_read,
1590                .write_string = cpuset_write_resmask,
1591                .max_write_len = (100U + 6 * NR_CPUS),
1592                .private = FILE_CPULIST,
1593        },
1594
1595        {
1596                .name = "mems",
1597                .read = cpuset_common_file_read,
1598                .write_string = cpuset_write_resmask,
1599                .max_write_len = (100U + 6 * MAX_NUMNODES),
1600                .private = FILE_MEMLIST,
1601        },
1602
1603        {
1604                .name = "cpu_exclusive",
1605                .read_u64 = cpuset_read_u64,
1606                .write_u64 = cpuset_write_u64,
1607                .private = FILE_CPU_EXCLUSIVE,
1608        },
1609
1610        {
1611                .name = "mem_exclusive",
1612                .read_u64 = cpuset_read_u64,
1613                .write_u64 = cpuset_write_u64,
1614                .private = FILE_MEM_EXCLUSIVE,
1615        },
1616
1617        {
1618                .name = "mem_hardwall",
1619                .read_u64 = cpuset_read_u64,
1620                .write_u64 = cpuset_write_u64,
1621                .private = FILE_MEM_HARDWALL,
1622        },
1623
1624        {
1625                .name = "sched_load_balance",
1626                .read_u64 = cpuset_read_u64,
1627                .write_u64 = cpuset_write_u64,
1628                .private = FILE_SCHED_LOAD_BALANCE,
1629        },
1630
1631        {
1632                .name = "sched_relax_domain_level",
1633                .read_s64 = cpuset_read_s64,
1634                .write_s64 = cpuset_write_s64,
1635                .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1636        },
1637
1638        {
1639                .name = "memory_migrate",
1640                .read_u64 = cpuset_read_u64,
1641                .write_u64 = cpuset_write_u64,
1642                .private = FILE_MEMORY_MIGRATE,
1643        },
1644
1645        {
1646                .name = "memory_pressure",
1647                .read_u64 = cpuset_read_u64,
1648                .write_u64 = cpuset_write_u64,
1649                .private = FILE_MEMORY_PRESSURE,
1650        },
1651
1652        {
1653                .name = "memory_spread_page",
1654                .read_u64 = cpuset_read_u64,
1655                .write_u64 = cpuset_write_u64,
1656                .private = FILE_SPREAD_PAGE,
1657        },
1658
1659        {
1660                .name = "memory_spread_slab",
1661                .read_u64 = cpuset_read_u64,
1662                .write_u64 = cpuset_write_u64,
1663                .private = FILE_SPREAD_SLAB,
1664        },
1665};
1666
1667static struct cftype cft_memory_pressure_enabled = {
1668        .name = "memory_pressure_enabled",
1669        .read_u64 = cpuset_read_u64,
1670        .write_u64 = cpuset_write_u64,
1671        .private = FILE_MEMORY_PRESSURE_ENABLED,
1672};
1673
1674static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1675{
1676        int err;
1677
1678        err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1679        if (err)
1680                return err;
1681        /* memory_pressure_enabled is in root cpuset only */
1682        if (!cont->parent)
1683                err = cgroup_add_file(cont, ss,
1684                                      &cft_memory_pressure_enabled);
1685        return err;
1686}
1687
1688/*
1689 * post_clone() is called at the end of cgroup_clone().
1690 * 'cgroup' was just created automatically as a result of
1691 * a cgroup_clone(), and the current task is about to
1692 * be moved into 'cgroup'.
1693 *
1694 * Currently we refuse to set up the cgroup - thereby
1695 * refusing the task to be entered, and as a result refusing
1696 * the sys_unshare() or clone() which initiated it - if any
1697 * sibling cpusets have exclusive cpus or mem.
1698 *
1699 * If this becomes a problem for some users who wish to
1700 * allow that scenario, then cpuset_post_clone() could be
1701 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1702 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1703 * held.
1704 */
1705static void cpuset_post_clone(struct cgroup_subsys *ss,
1706                              struct cgroup *cgroup)
1707{
1708        struct cgroup *parent, *child;
1709        struct cpuset *cs, *parent_cs;
1710
1711        parent = cgroup->parent;
1712        list_for_each_entry(child, &parent->children, sibling) {
1713                cs = cgroup_cs(child);
1714                if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1715                        return;
1716        }
1717        cs = cgroup_cs(cgroup);
1718        parent_cs = cgroup_cs(parent);
1719
1720        cs->mems_allowed = parent_cs->mems_allowed;
1721        cs->cpus_allowed = parent_cs->cpus_allowed;
1722        return;
1723}
1724
1725/*
1726 *      cpuset_create - create a cpuset
1727 *      ss:     cpuset cgroup subsystem
1728 *      cont:   control group that the new cpuset will be part of
1729 */
1730
1731static struct cgroup_subsys_state *cpuset_create(
1732        struct cgroup_subsys *ss,
1733        struct cgroup *cont)
1734{
1735        struct cpuset *cs;
1736        struct cpuset *parent;
1737
1738        if (!cont->parent) {
1739                /* This is early initialization for the top cgroup */
1740                top_cpuset.mems_generation = cpuset_mems_generation++;
1741                return &top_cpuset.css;
1742        }
1743        parent = cgroup_cs(cont->parent);
1744        cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1745        if (!cs)
1746                return ERR_PTR(-ENOMEM);
1747
1748        cpuset_update_task_memory_state();
1749        cs->flags = 0;
1750        if (is_spread_page(parent))
1751                set_bit(CS_SPREAD_PAGE, &cs->flags);
1752        if (is_spread_slab(parent))
1753                set_bit(CS_SPREAD_SLAB, &cs->flags);
1754        set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1755        cpus_clear(cs->cpus_allowed);
1756        nodes_clear(cs->mems_allowed);
1757        cs->mems_generation = cpuset_mems_generation++;
1758        fmeter_init(&cs->fmeter);
1759        cs->relax_domain_level = -1;
1760
1761        cs->parent = parent;
1762        number_of_cpusets++;
1763        return &cs->css ;
1764}
1765
1766/*
1767 * If the cpuset being removed has its flag 'sched_load_balance'
1768 * enabled, then simulate turning sched_load_balance off, which
1769 * will call async_rebuild_sched_domains().
1770 */
1771
1772static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1773{
1774        struct cpuset *cs = cgroup_cs(cont);
1775
1776        cpuset_update_task_memory_state();
1777
1778        if (is_sched_load_balance(cs))
1779                update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1780
1781        number_of_cpusets--;
1782        kfree(cs);
1783}
1784
1785struct cgroup_subsys cpuset_subsys = {
1786        .name = "cpuset",
1787        .create = cpuset_create,
1788        .destroy = cpuset_destroy,
1789        .can_attach = cpuset_can_attach,
1790        .attach = cpuset_attach,
1791        .populate = cpuset_populate,
1792        .post_clone = cpuset_post_clone,
1793        .subsys_id = cpuset_subsys_id,
1794        .early_init = 1,
1795};
1796
1797/*
1798 * cpuset_init_early - just enough so that the calls to
1799 * cpuset_update_task_memory_state() in early init code
1800 * are harmless.
1801 */
1802
1803int __init cpuset_init_early(void)
1804{
1805        top_cpuset.mems_generation = cpuset_mems_generation++;
1806        return 0;
1807}
1808
1809
1810/**
1811 * cpuset_init - initialize cpusets at system boot
1812 *
1813 * Description: Initialize top_cpuset and the cpuset internal file system,
1814 **/
1815
1816int __init cpuset_init(void)
1817{
1818        int err = 0;
1819
1820        cpus_setall(top_cpuset.cpus_allowed);
1821        nodes_setall(top_cpuset.mems_allowed);
1822
1823        fmeter_init(&top_cpuset.fmeter);
1824        top_cpuset.mems_generation = cpuset_mems_generation++;
1825        set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1826        top_cpuset.relax_domain_level = -1;
1827
1828        err = register_filesystem(&cpuset_fs_type);
1829        if (err < 0)
1830                return err;
1831
1832        number_of_cpusets = 1;
1833        return 0;
1834}
1835
1836/**
1837 * cpuset_do_move_task - move a given task to another cpuset
1838 * @tsk: pointer to task_struct the task to move
1839 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1840 *
1841 * Called by cgroup_scan_tasks() for each task in a cgroup.
1842 * Return nonzero to stop the walk through the tasks.
1843 */
1844static void cpuset_do_move_task(struct task_struct *tsk,
1845                                struct cgroup_scanner *scan)
1846{
1847        struct cpuset_hotplug_scanner *chsp;
1848
1849        chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1850        cgroup_attach_task(chsp->to, tsk);
1851}
1852
1853/**
1854 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1855 * @from: cpuset in which the tasks currently reside
1856 * @to: cpuset to which the tasks will be moved
1857 *
1858 * Called with cgroup_mutex held
1859 * callback_mutex must not be held, as cpuset_attach() will take it.
1860 *
1861 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1862 * calling callback functions for each.
1863 */
1864static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1865{
1866        struct cpuset_hotplug_scanner scan;
1867
1868        scan.scan.cg = from->css.cgroup;
1869        scan.scan.test_task = NULL; /* select all tasks in cgroup */
1870        scan.scan.process_task = cpuset_do_move_task;
1871        scan.scan.heap = NULL;
1872        scan.to = to->css.cgroup;
1873
1874        if (cgroup_scan_tasks(&scan.scan))
1875                printk(KERN_ERR "move_member_tasks_to_cpuset: "
1876                                "cgroup_scan_tasks failed\n");
1877}
1878
1879/*
1880 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1881 * or memory nodes, we need to walk over the cpuset hierarchy,
1882 * removing that CPU or node from all cpusets.  If this removes the
1883 * last CPU or node from a cpuset, then move the tasks in the empty
1884 * cpuset to its next-highest non-empty parent.
1885 *
1886 * Called with cgroup_mutex held
1887 * callback_mutex must not be held, as cpuset_attach() will take it.
1888 */
1889static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1890{
1891        struct cpuset *parent;
1892
1893        /*
1894         * The cgroup's css_sets list is in use if there are tasks
1895         * in the cpuset; the list is empty if there are none;
1896         * the cs->css.refcnt seems always 0.
1897         */
1898        if (list_empty(&cs->css.cgroup->css_sets))
1899                return;
1900
1901        /*
1902         * Find its next-highest non-empty parent, (top cpuset
1903         * has online cpus, so can't be empty).
1904         */
1905        parent = cs->parent;
1906        while (cpus_empty(parent->cpus_allowed) ||
1907                        nodes_empty(parent->mems_allowed))
1908                parent = parent->parent;
1909
1910        move_member_tasks_to_cpuset(cs, parent);
1911}
1912
1913/*
1914 * Walk the specified cpuset subtree and look for empty cpusets.
1915 * The tasks of such cpuset must be moved to a parent cpuset.
1916 *
1917 * Called with cgroup_mutex held.  We take callback_mutex to modify
1918 * cpus_allowed and mems_allowed.
1919 *
1920 * This walk processes the tree from top to bottom, completing one layer
1921 * before dropping down to the next.  It always processes a node before
1922 * any of its children.
1923 *
1924 * For now, since we lack memory hot unplug, we'll never see a cpuset
1925 * that has tasks along with an empty 'mems'.  But if we did see such
1926 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1927 */
1928static void scan_for_empty_cpusets(struct cpuset *root)
1929{
1930        LIST_HEAD(queue);
1931        struct cpuset *cp;      /* scans cpusets being updated */
1932        struct cpuset *child;   /* scans child cpusets of cp */
1933        struct cgroup *cont;
1934        nodemask_t oldmems;
1935
1936        list_add_tail((struct list_head *)&root->stack_list, &queue);
1937
1938        while (!list_empty(&queue)) {
1939                cp = list_first_entry(&queue, struct cpuset, stack_list);
1940                list_del(queue.next);
1941                list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
1942                        child = cgroup_cs(cont);
1943                        list_add_tail(&child->stack_list, &queue);
1944                }
1945
1946                /* Continue past cpusets with all cpus, mems online */
1947                if (cpus_subset(cp->cpus_allowed, cpu_online_map) &&
1948                    nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
1949                        continue;
1950
1951                oldmems = cp->mems_allowed;
1952
1953                /* Remove offline cpus and mems from this cpuset. */
1954                mutex_lock(&callback_mutex);
1955                cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map);
1956                nodes_and(cp->mems_allowed, cp->mems_allowed,
1957                                                node_states[N_HIGH_MEMORY]);
1958                mutex_unlock(&callback_mutex);
1959
1960                /* Move tasks from the empty cpuset to a parent */
1961                if (cpus_empty(cp->cpus_allowed) ||
1962                     nodes_empty(cp->mems_allowed))
1963                        remove_tasks_in_empty_cpuset(cp);
1964                else {
1965                        update_tasks_cpumask(cp, NULL);
1966                        update_tasks_nodemask(cp, &oldmems);
1967                }
1968        }
1969}
1970
1971/*
1972 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1973 * period.  This is necessary in order to make cpusets transparent
1974 * (of no affect) on systems that are actively using CPU hotplug
1975 * but making no active use of cpusets.
1976 *
1977 * This routine ensures that top_cpuset.cpus_allowed tracks
1978 * cpu_online_map on each CPU hotplug (cpuhp) event.
1979 *
1980 * Called within get_online_cpus().  Needs to call cgroup_lock()
1981 * before calling generate_sched_domains().
1982 */
1983static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
1984                                unsigned long phase, void *unused_cpu)
1985{
1986        struct sched_domain_attr *attr;
1987        cpumask_t *doms;
1988        int ndoms;
1989
1990        switch (phase) {
1991        case CPU_ONLINE:
1992        case CPU_ONLINE_FROZEN:
1993        case CPU_DEAD:
1994        case CPU_DEAD_FROZEN:
1995                break;
1996
1997        default:
1998                return NOTIFY_DONE;
1999        }
2000
2001        cgroup_lock();
2002        top_cpuset.cpus_allowed = cpu_online_map;
2003        scan_for_empty_cpusets(&top_cpuset);
2004        ndoms = generate_sched_domains(&doms, &attr);
2005        cgroup_unlock();
2006
2007        /* Have scheduler rebuild the domains */
2008        partition_sched_domains(ndoms, doms, attr);
2009
2010        return NOTIFY_OK;
2011}
2012
2013#ifdef CONFIG_MEMORY_HOTPLUG
2014/*
2015 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2016 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2017 * See also the previous routine cpuset_track_online_cpus().
2018 */
2019static int cpuset_track_online_nodes(struct notifier_block *self,
2020                                unsigned long action, void *arg)
2021{
2022        cgroup_lock();
2023        switch (action) {
2024        case MEM_ONLINE:
2025                top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2026                break;
2027        case MEM_OFFLINE:
2028                top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2029                scan_for_empty_cpusets(&top_cpuset);
2030                break;
2031        default:
2032                break;
2033        }
2034        cgroup_unlock();
2035        return NOTIFY_OK;
2036}
2037#endif
2038
2039/**
2040 * cpuset_init_smp - initialize cpus_allowed
2041 *
2042 * Description: Finish top cpuset after cpu, node maps are initialized
2043 **/
2044
2045void __init cpuset_init_smp(void)
2046{
2047        top_cpuset.cpus_allowed = cpu_online_map;
2048        top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2049
2050        hotcpu_notifier(cpuset_track_online_cpus, 0);
2051        hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2052}
2053
2054/**
2055 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2056 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2057 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
2058 *
2059 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2060 * attached to the specified @tsk.  Guaranteed to return some non-empty
2061 * subset of cpu_online_map, even if this means going outside the
2062 * tasks cpuset.
2063 **/
2064
2065void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask)
2066{
2067        mutex_lock(&callback_mutex);
2068        cpuset_cpus_allowed_locked(tsk, pmask);
2069        mutex_unlock(&callback_mutex);
2070}
2071
2072/**
2073 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2074 * Must be called with callback_mutex held.
2075 **/
2076void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask)
2077{
2078        task_lock(tsk);
2079        guarantee_online_cpus(task_cs(tsk), pmask);
2080        task_unlock(tsk);
2081}
2082
2083void cpuset_init_current_mems_allowed(void)
2084{
2085        nodes_setall(current->mems_allowed);
2086}
2087
2088/**
2089 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2090 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2091 *
2092 * Description: Returns the nodemask_t mems_allowed of the cpuset
2093 * attached to the specified @tsk.  Guaranteed to return some non-empty
2094 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2095 * tasks cpuset.
2096 **/
2097
2098nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2099{
2100        nodemask_t mask;
2101
2102        mutex_lock(&callback_mutex);
2103        task_lock(tsk);
2104        guarantee_online_mems(task_cs(tsk), &mask);
2105        task_unlock(tsk);
2106        mutex_unlock(&callback_mutex);
2107
2108        return mask;
2109}
2110
2111/**
2112 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2113 * @nodemask: the nodemask to be checked
2114 *
2115 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2116 */
2117int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2118{
2119        return nodes_intersects(*nodemask, current->mems_allowed);
2120}
2121
2122/*
2123 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2124 * mem_hardwall ancestor to the specified cpuset.  Call holding
2125 * callback_mutex.  If no ancestor is mem_exclusive or mem_hardwall
2126 * (an unusual configuration), then returns the root cpuset.
2127 */
2128static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2129{
2130        while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2131                cs = cs->parent;
2132        return cs;
2133}
2134
2135/**
2136 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2137 * @z: is this zone on an allowed node?
2138 * @gfp_mask: memory allocation flags
2139 *
2140 * If we're in interrupt, yes, we can always allocate.  If
2141 * __GFP_THISNODE is set, yes, we can always allocate.  If zone
2142 * z's node is in our tasks mems_allowed, yes.  If it's not a
2143 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2144 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2145 * If the task has been OOM killed and has access to memory reserves
2146 * as specified by the TIF_MEMDIE flag, yes.
2147 * Otherwise, no.
2148 *
2149 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2150 * reduces to cpuset_zone_allowed_hardwall().  Otherwise,
2151 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2152 * from an enclosing cpuset.
2153 *
2154 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2155 * hardwall cpusets, and never sleeps.
2156 *
2157 * The __GFP_THISNODE placement logic is really handled elsewhere,
2158 * by forcibly using a zonelist starting at a specified node, and by
2159 * (in get_page_from_freelist()) refusing to consider the zones for
2160 * any node on the zonelist except the first.  By the time any such
2161 * calls get to this routine, we should just shut up and say 'yes'.
2162 *
2163 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2164 * and do not allow allocations outside the current tasks cpuset
2165 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2166 * GFP_KERNEL allocations are not so marked, so can escape to the
2167 * nearest enclosing hardwalled ancestor cpuset.
2168 *
2169 * Scanning up parent cpusets requires callback_mutex.  The
2170 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2171 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2172 * current tasks mems_allowed came up empty on the first pass over
2173 * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2174 * cpuset are short of memory, might require taking the callback_mutex
2175 * mutex.
2176 *
2177 * The first call here from mm/page_alloc:get_page_from_freelist()
2178 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2179 * so no allocation on a node outside the cpuset is allowed (unless
2180 * in interrupt, of course).
2181 *
2182 * The second pass through get_page_from_freelist() doesn't even call
2183 * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2184 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2185 * in alloc_flags.  That logic and the checks below have the combined
2186 * affect that:
2187 *      in_interrupt - any node ok (current task context irrelevant)
2188 *      GFP_ATOMIC   - any node ok
2189 *      TIF_MEMDIE   - any node ok
2190 *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2191 *      GFP_USER     - only nodes in current tasks mems allowed ok.
2192 *
2193 * Rule:
2194 *    Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2195 *    pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2196 *    the code that might scan up ancestor cpusets and sleep.
2197 */
2198
2199int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2200{
2201        int node;                       /* node that zone z is on */
2202        const struct cpuset *cs;        /* current cpuset ancestors */
2203        int allowed;                    /* is allocation in zone z allowed? */
2204
2205        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2206                return 1;
2207        node = zone_to_nid(z);
2208        might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2209        if (node_isset(node, current->mems_allowed))
2210                return 1;
2211        /*
2212         * Allow tasks that have access to memory reserves because they have
2213         * been OOM killed to get memory anywhere.
2214         */
2215        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2216                return 1;
2217        if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2218                return 0;
2219
2220        if (current->flags & PF_EXITING) /* Let dying task have memory */
2221                return 1;
2222
2223        /* Not hardwall and node outside mems_allowed: scan up cpusets */
2224        mutex_lock(&callback_mutex);
2225
2226        task_lock(current);
2227        cs = nearest_hardwall_ancestor(task_cs(current));
2228        task_unlock(current);
2229
2230        allowed = node_isset(node, cs->mems_allowed);
2231        mutex_unlock(&callback_mutex);
2232        return allowed;
2233}
2234
2235/*
2236 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2237 * @z: is this zone on an allowed node?
2238 * @gfp_mask: memory allocation flags
2239 *
2240 * If we're in interrupt, yes, we can always allocate.
2241 * If __GFP_THISNODE is set, yes, we can always allocate.  If zone
2242 * z's node is in our tasks mems_allowed, yes.   If the task has been
2243 * OOM killed and has access to memory reserves as specified by the
2244 * TIF_MEMDIE flag, yes.  Otherwise, no.
2245 *
2246 * The __GFP_THISNODE placement logic is really handled elsewhere,
2247 * by forcibly using a zonelist starting at a specified node, and by
2248 * (in get_page_from_freelist()) refusing to consider the zones for
2249 * any node on the zonelist except the first.  By the time any such
2250 * calls get to this routine, we should just shut up and say 'yes'.
2251 *
2252 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2253 * this variant requires that the zone be in the current tasks
2254 * mems_allowed or that we're in interrupt.  It does not scan up the
2255 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2256 * It never sleeps.
2257 */
2258
2259int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2260{
2261        int node;                       /* node that zone z is on */
2262
2263        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2264                return 1;
2265        node = zone_to_nid(z);
2266        if (node_isset(node, current->mems_allowed))
2267                return 1;
2268        /*
2269         * Allow tasks that have access to memory reserves because they have
2270         * been OOM killed to get memory anywhere.
2271         */
2272        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2273                return 1;
2274        return 0;
2275}
2276
2277/**
2278 * cpuset_lock - lock out any changes to cpuset structures
2279 *
2280 * The out of memory (oom) code needs to mutex_lock cpusets
2281 * from being changed while it scans the tasklist looking for a
2282 * task in an overlapping cpuset.  Expose callback_mutex via this
2283 * cpuset_lock() routine, so the oom code can lock it, before
2284 * locking the task list.  The tasklist_lock is a spinlock, so
2285 * must be taken inside callback_mutex.
2286 */
2287
2288void cpuset_lock(void)
2289{
2290        mutex_lock(&callback_mutex);
2291}
2292
2293/**
2294 * cpuset_unlock - release lock on cpuset changes
2295 *
2296 * Undo the lock taken in a previous cpuset_lock() call.
2297 */
2298
2299void cpuset_unlock(void)
2300{
2301        mutex_unlock(&callback_mutex);
2302}
2303
2304/**
2305 * cpuset_mem_spread_node() - On which node to begin search for a page
2306 *
2307 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2308 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2309 * and if the memory allocation used cpuset_mem_spread_node()
2310 * to determine on which node to start looking, as it will for
2311 * certain page cache or slab cache pages such as used for file
2312 * system buffers and inode caches, then instead of starting on the
2313 * local node to look for a free page, rather spread the starting
2314 * node around the tasks mems_allowed nodes.
2315 *
2316 * We don't have to worry about the returned node being offline
2317 * because "it can't happen", and even if it did, it would be ok.
2318 *
2319 * The routines calling guarantee_online_mems() are careful to
2320 * only set nodes in task->mems_allowed that are online.  So it
2321 * should not be possible for the following code to return an
2322 * offline node.  But if it did, that would be ok, as this routine
2323 * is not returning the node where the allocation must be, only
2324 * the node where the search should start.  The zonelist passed to
2325 * __alloc_pages() will include all nodes.  If the slab allocator
2326 * is passed an offline node, it will fall back to the local node.
2327 * See kmem_cache_alloc_node().
2328 */
2329
2330int cpuset_mem_spread_node(void)
2331{
2332        int node;
2333
2334        node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2335        if (node == MAX_NUMNODES)
2336                node = first_node(current->mems_allowed);
2337        current->cpuset_mem_spread_rotor = node;
2338        return node;
2339}
2340EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2341
2342/**
2343 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2344 * @tsk1: pointer to task_struct of some task.
2345 * @tsk2: pointer to task_struct of some other task.
2346 *
2347 * Description: Return true if @tsk1's mems_allowed intersects the
2348 * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2349 * one of the task's memory usage might impact the memory available
2350 * to the other.
2351 **/
2352
2353int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2354                                   const struct task_struct *tsk2)
2355{
2356        return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2357}
2358
2359/*
2360 * Collection of memory_pressure is suppressed unless
2361 * this flag is enabled by writing "1" to the special
2362 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2363 */
2364
2365int cpuset_memory_pressure_enabled __read_mostly;
2366
2367/**
2368 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2369 *
2370 * Keep a running average of the rate of synchronous (direct)
2371 * page reclaim efforts initiated by tasks in each cpuset.
2372 *
2373 * This represents the rate at which some task in the cpuset
2374 * ran low on memory on all nodes it was allowed to use, and
2375 * had to enter the kernels page reclaim code in an effort to
2376 * create more free memory by tossing clean pages or swapping
2377 * or writing dirty pages.
2378 *
2379 * Display to user space in the per-cpuset read-only file
2380 * "memory_pressure".  Value displayed is an integer
2381 * representing the recent rate of entry into the synchronous
2382 * (direct) page reclaim by any task attached to the cpuset.
2383 **/
2384
2385void __cpuset_memory_pressure_bump(void)
2386{
2387        task_lock(current);
2388        fmeter_markevent(&task_cs(current)->fmeter);
2389        task_unlock(current);
2390}
2391
2392#ifdef CONFIG_PROC_PID_CPUSET
2393/*
2394 * proc_cpuset_show()
2395 *  - Print tasks cpuset path into seq_file.
2396 *  - Used for /proc/<pid>/cpuset.
2397 *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2398 *    doesn't really matter if tsk->cpuset changes after we read it,
2399 *    and we take cgroup_mutex, keeping cpuset_attach() from changing it
2400 *    anyway.
2401 */
2402static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2403{
2404        struct pid *pid;
2405        struct task_struct *tsk;
2406        char *buf;
2407        struct cgroup_subsys_state *css;
2408        int retval;
2409
2410        retval = -ENOMEM;
2411        buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2412        if (!buf)
2413                goto out;
2414
2415        retval = -ESRCH;
2416        pid = m->private;
2417        tsk = get_pid_task(pid, PIDTYPE_PID);
2418        if (!tsk)
2419                goto out_free;
2420
2421        retval = -EINVAL;
2422        cgroup_lock();
2423        css = task_subsys_state(tsk, cpuset_subsys_id);
2424        retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2425        if (retval < 0)
2426                goto out_unlock;
2427        seq_puts(m, buf);
2428        seq_putc(m, '\n');
2429out_unlock:
2430        cgroup_unlock();
2431        put_task_struct(tsk);
2432out_free:
2433        kfree(buf);
2434out:
2435        return retval;
2436}
2437
2438static int cpuset_open(struct inode *inode, struct file *file)
2439{
2440        struct pid *pid = PROC_I(inode)->pid;
2441        return single_open(file, proc_cpuset_show, pid);
2442}
2443
2444const struct file_operations proc_cpuset_operations = {
2445        .open           = cpuset_open,
2446        .read           = seq_read,
2447        .llseek         = seq_lseek,
2448        .release        = single_release,
2449};
2450#endif /* CONFIG_PROC_PID_CPUSET */
2451
2452/* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2453void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2454{
2455        seq_printf(m, "Cpus_allowed:\t");
2456        seq_cpumask(m, &task->cpus_allowed);
2457        seq_printf(m, "\n");
2458        seq_printf(m, "Cpus_allowed_list:\t");
2459        seq_cpumask_list(m, &task->cpus_allowed);
2460        seq_printf(m, "\n");
2461        seq_printf(m, "Mems_allowed:\t");
2462        seq_nodemask(m, &task->mems_allowed);
2463        seq_printf(m, "\n");
2464        seq_printf(m, "Mems_allowed_list:\t");
2465        seq_nodemask_list(m, &task->mems_allowed);
2466        seq_printf(m, "\n");
2467}
2468