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/export.h>
  41#include <linux/mount.h>
  42#include <linux/namei.h>
  43#include <linux/pagemap.h>
  44#include <linux/proc_fs.h>
  45#include <linux/rcupdate.h>
  46#include <linux/sched.h>
  47#include <linux/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 <linux/atomic.h>
  59#include <linux/mutex.h>
  60#include <linux/workqueue.h>
  61#include <linux/cgroup.h>
  62
  63/*
  64 * Workqueue for cpuset related tasks.
  65 *
  66 * Using kevent workqueue may cause deadlock when memory_migrate
  67 * is set. So we create a separate workqueue thread for cpuset.
  68 */
  69static struct workqueue_struct *cpuset_wq;
  70
  71/*
  72 * Tracks how many cpusets are currently defined in system.
  73 * When there is only one cpuset (the root cpuset) we can
  74 * short circuit some hooks.
  75 */
  76int number_of_cpusets __read_mostly;
  77
  78/* Forward declare cgroup structures */
  79struct cgroup_subsys cpuset_subsys;
  80struct cpuset;
  81
  82/* See "Frequency meter" comments, below. */
  83
  84struct fmeter {
  85        int cnt;                /* unprocessed events count */
  86        int val;                /* most recent output value */
  87        time_t time;            /* clock (secs) when val computed */
  88        spinlock_t lock;        /* guards read or write of above */
  89};
  90
  91struct cpuset {
  92        struct cgroup_subsys_state css;
  93
  94        unsigned long flags;            /* "unsigned long" so bitops work */
  95        cpumask_var_t cpus_allowed;     /* CPUs allowed to tasks in cpuset */
  96        nodemask_t mems_allowed;        /* Memory Nodes allowed to tasks */
  97
  98        struct cpuset *parent;          /* my parent */
  99
 100        struct fmeter fmeter;           /* memory_pressure filter */
 101
 102        /* partition number for rebuild_sched_domains() */
 103        int pn;
 104
 105        /* for custom sched domain */
 106        int relax_domain_level;
 107
 108        /* used for walking a cpuset hierarchy */
 109        struct list_head stack_list;
 110};
 111
 112/* Retrieve the cpuset for a cgroup */
 113static inline struct cpuset *cgroup_cs(struct cgroup *cont)
 114{
 115        return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
 116                            struct cpuset, css);
 117}
 118
 119/* Retrieve the cpuset for a task */
 120static inline struct cpuset *task_cs(struct task_struct *task)
 121{
 122        return container_of(task_subsys_state(task, cpuset_subsys_id),
 123                            struct cpuset, css);
 124}
 125
 126#ifdef CONFIG_NUMA
 127static inline bool task_has_mempolicy(struct task_struct *task)
 128{
 129        return task->mempolicy;
 130}
 131#else
 132static inline bool task_has_mempolicy(struct task_struct *task)
 133{
 134        return false;
 135}
 136#endif
 137
 138
 139/* bits in struct cpuset flags field */
 140typedef enum {
 141        CS_CPU_EXCLUSIVE,
 142        CS_MEM_EXCLUSIVE,
 143        CS_MEM_HARDWALL,
 144        CS_MEMORY_MIGRATE,
 145        CS_SCHED_LOAD_BALANCE,
 146        CS_SPREAD_PAGE,
 147        CS_SPREAD_SLAB,
 148} cpuset_flagbits_t;
 149
 150/* convenient tests for these bits */
 151static inline int is_cpu_exclusive(const struct cpuset *cs)
 152{
 153        return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
 154}
 155
 156static inline int is_mem_exclusive(const struct cpuset *cs)
 157{
 158        return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
 159}
 160
 161static inline int is_mem_hardwall(const struct cpuset *cs)
 162{
 163        return test_bit(CS_MEM_HARDWALL, &cs->flags);
 164}
 165
 166static inline int is_sched_load_balance(const struct cpuset *cs)
 167{
 168        return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
 169}
 170
 171static inline int is_memory_migrate(const struct cpuset *cs)
 172{
 173        return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
 174}
 175
 176static inline int is_spread_page(const struct cpuset *cs)
 177{
 178        return test_bit(CS_SPREAD_PAGE, &cs->flags);
 179}
 180
 181static inline int is_spread_slab(const struct cpuset *cs)
 182{
 183        return test_bit(CS_SPREAD_SLAB, &cs->flags);
 184}
 185
 186static struct cpuset top_cpuset = {
 187        .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
 188};
 189
 190/*
 191 * There are two global mutexes guarding cpuset structures.  The first
 192 * is the main control groups cgroup_mutex, accessed via
 193 * cgroup_lock()/cgroup_unlock().  The second is the cpuset-specific
 194 * callback_mutex, below. They can nest.  It is ok to first take
 195 * cgroup_mutex, then nest callback_mutex.  We also require taking
 196 * task_lock() when dereferencing a task's cpuset pointer.  See "The
 197 * task_lock() exception", at the end of this comment.
 198 *
 199 * A task must hold both mutexes to modify cpusets.  If a task
 200 * holds cgroup_mutex, then it blocks others wanting that mutex,
 201 * ensuring that it is the only task able to also acquire callback_mutex
 202 * and be able to modify cpusets.  It can perform various checks on
 203 * the cpuset structure first, knowing nothing will change.  It can
 204 * also allocate memory while just holding cgroup_mutex.  While it is
 205 * performing these checks, various callback routines can briefly
 206 * acquire callback_mutex to query cpusets.  Once it is ready to make
 207 * the changes, it takes callback_mutex, blocking everyone else.
 208 *
 209 * Calls to the kernel memory allocator can not be made while holding
 210 * callback_mutex, as that would risk double tripping on callback_mutex
 211 * from one of the callbacks into the cpuset code from within
 212 * __alloc_pages().
 213 *
 214 * If a task is only holding callback_mutex, then it has read-only
 215 * access to cpusets.
 216 *
 217 * Now, the task_struct fields mems_allowed and mempolicy may be changed
 218 * by other task, we use alloc_lock in the task_struct fields to protect
 219 * them.
 220 *
 221 * The cpuset_common_file_read() handlers only hold callback_mutex across
 222 * small pieces of code, such as when reading out possibly multi-word
 223 * cpumasks and nodemasks.
 224 *
 225 * Accessing a task's cpuset should be done in accordance with the
 226 * guidelines for accessing subsystem state in kernel/cgroup.c
 227 */
 228
 229static DEFINE_MUTEX(callback_mutex);
 230
 231/*
 232 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
 233 * buffers.  They are statically allocated to prevent using excess stack
 234 * when calling cpuset_print_task_mems_allowed().
 235 */
 236#define CPUSET_NAME_LEN         (128)
 237#define CPUSET_NODELIST_LEN     (256)
 238static char cpuset_name[CPUSET_NAME_LEN];
 239static char cpuset_nodelist[CPUSET_NODELIST_LEN];
 240static DEFINE_SPINLOCK(cpuset_buffer_lock);
 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 struct dentry *cpuset_mount(struct file_system_type *fs_type,
 248                         int flags, const char *unused_dev_name, void *data)
 249{
 250        struct file_system_type *cgroup_fs = get_fs_type("cgroup");
 251        struct dentry *ret = ERR_PTR(-ENODEV);
 252        if (cgroup_fs) {
 253                char mountopts[] =
 254                        "cpuset,noprefix,"
 255                        "release_agent=/sbin/cpuset_release_agent";
 256                ret = cgroup_fs->mount(cgroup_fs, flags,
 257                                           unused_dev_name, mountopts);
 258                put_filesystem(cgroup_fs);
 259        }
 260        return ret;
 261}
 262
 263static struct file_system_type cpuset_fs_type = {
 264        .name = "cpuset",
 265        .mount = cpuset_mount,
 266};
 267
 268/*
 269 * Return in pmask the portion of a cpusets's cpus_allowed that
 270 * are online.  If none are online, walk up the cpuset hierarchy
 271 * until we find one that does have some online cpus.  If we get
 272 * all the way to the top and still haven't found any online cpus,
 273 * return cpu_online_map.  Or if passed a NULL cs from an exit'ing
 274 * task, return cpu_online_map.
 275 *
 276 * One way or another, we guarantee to return some non-empty subset
 277 * of cpu_online_map.
 278 *
 279 * Call with callback_mutex held.
 280 */
 281
 282static void guarantee_online_cpus(const struct cpuset *cs,
 283                                  struct cpumask *pmask)
 284{
 285        while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
 286                cs = cs->parent;
 287        if (cs)
 288                cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
 289        else
 290                cpumask_copy(pmask, cpu_online_mask);
 291        BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
 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 * update task's spread flag if cpuset's page/slab spread flag is set
 322 *
 323 * Called with callback_mutex/cgroup_mutex held
 324 */
 325static void cpuset_update_task_spread_flag(struct cpuset *cs,
 326                                        struct task_struct *tsk)
 327{
 328        if (is_spread_page(cs))
 329                tsk->flags |= PF_SPREAD_PAGE;
 330        else
 331                tsk->flags &= ~PF_SPREAD_PAGE;
 332        if (is_spread_slab(cs))
 333                tsk->flags |= PF_SPREAD_SLAB;
 334        else
 335                tsk->flags &= ~PF_SPREAD_SLAB;
 336}
 337
 338/*
 339 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
 340 *
 341 * One cpuset is a subset of another if all its allowed CPUs and
 342 * Memory Nodes are a subset of the other, and its exclusive flags
 343 * are only set if the other's are set.  Call holding cgroup_mutex.
 344 */
 345
 346static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
 347{
 348        return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
 349                nodes_subset(p->mems_allowed, q->mems_allowed) &&
 350                is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
 351                is_mem_exclusive(p) <= is_mem_exclusive(q);
 352}
 353
 354/**
 355 * alloc_trial_cpuset - allocate a trial cpuset
 356 * @cs: the cpuset that the trial cpuset duplicates
 357 */
 358static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
 359{
 360        struct cpuset *trial;
 361
 362        trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
 363        if (!trial)
 364                return NULL;
 365
 366        if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
 367                kfree(trial);
 368                return NULL;
 369        }
 370        cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
 371
 372        return trial;
 373}
 374
 375/**
 376 * free_trial_cpuset - free the trial cpuset
 377 * @trial: the trial cpuset to be freed
 378 */
 379static void free_trial_cpuset(struct cpuset *trial)
 380{
 381        free_cpumask_var(trial->cpus_allowed);
 382        kfree(trial);
 383}
 384
 385/*
 386 * validate_change() - Used to validate that any proposed cpuset change
 387 *                     follows the structural rules for cpusets.
 388 *
 389 * If we replaced the flag and mask values of the current cpuset
 390 * (cur) with those values in the trial cpuset (trial), would
 391 * our various subset and exclusive rules still be valid?  Presumes
 392 * cgroup_mutex held.
 393 *
 394 * 'cur' is the address of an actual, in-use cpuset.  Operations
 395 * such as list traversal that depend on the actual address of the
 396 * cpuset in the list must use cur below, not trial.
 397 *
 398 * 'trial' is the address of bulk structure copy of cur, with
 399 * perhaps one or more of the fields cpus_allowed, mems_allowed,
 400 * or flags changed to new, trial values.
 401 *
 402 * Return 0 if valid, -errno if not.
 403 */
 404
 405static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 406{
 407        struct cgroup *cont;
 408        struct cpuset *c, *par;
 409
 410        /* Each of our child cpusets must be a subset of us */
 411        list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
 412                if (!is_cpuset_subset(cgroup_cs(cont), trial))
 413                        return -EBUSY;
 414        }
 415
 416        /* Remaining checks don't apply to root cpuset */
 417        if (cur == &top_cpuset)
 418                return 0;
 419
 420        par = cur->parent;
 421
 422        /* We must be a subset of our parent cpuset */
 423        if (!is_cpuset_subset(trial, par))
 424                return -EACCES;
 425
 426        /*
 427         * If either I or some sibling (!= me) is exclusive, we can't
 428         * overlap
 429         */
 430        list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
 431                c = cgroup_cs(cont);
 432                if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
 433                    c != cur &&
 434                    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
 435                        return -EINVAL;
 436                if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
 437                    c != cur &&
 438                    nodes_intersects(trial->mems_allowed, c->mems_allowed))
 439                        return -EINVAL;
 440        }
 441
 442        /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
 443        if (cgroup_task_count(cur->css.cgroup)) {
 444                if (cpumask_empty(trial->cpus_allowed) ||
 445                    nodes_empty(trial->mems_allowed)) {
 446                        return -ENOSPC;
 447                }
 448        }
 449
 450        return 0;
 451}
 452
 453#ifdef CONFIG_SMP
 454/*
 455 * Helper routine for generate_sched_domains().
 456 * Do cpusets a, b have overlapping cpus_allowed masks?
 457 */
 458static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
 459{
 460        return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
 461}
 462
 463static void
 464update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
 465{
 466        if (dattr->relax_domain_level < c->relax_domain_level)
 467                dattr->relax_domain_level = c->relax_domain_level;
 468        return;
 469}
 470
 471static void
 472update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
 473{
 474        LIST_HEAD(q);
 475
 476        list_add(&c->stack_list, &q);
 477        while (!list_empty(&q)) {
 478                struct cpuset *cp;
 479                struct cgroup *cont;
 480                struct cpuset *child;
 481
 482                cp = list_first_entry(&q, struct cpuset, stack_list);
 483                list_del(q.next);
 484
 485                if (cpumask_empty(cp->cpus_allowed))
 486                        continue;
 487
 488                if (is_sched_load_balance(cp))
 489                        update_domain_attr(dattr, cp);
 490
 491                list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
 492                        child = cgroup_cs(cont);
 493                        list_add_tail(&child->stack_list, &q);
 494                }
 495        }
 496}
 497
 498/*
 499 * generate_sched_domains()
 500 *
 501 * This function builds a partial partition of the systems CPUs
 502 * A 'partial partition' is a set of non-overlapping subsets whose
 503 * union is a subset of that set.
 504 * The output of this function needs to be passed to kernel/sched.c
 505 * partition_sched_domains() routine, which will rebuild the scheduler's
 506 * load balancing domains (sched domains) as specified by that partial
 507 * partition.
 508 *
 509 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
 510 * for a background explanation of this.
 511 *
 512 * Does not return errors, on the theory that the callers of this
 513 * routine would rather not worry about failures to rebuild sched
 514 * domains when operating in the severe memory shortage situations
 515 * that could cause allocation failures below.
 516 *
 517 * Must be called with cgroup_lock held.
 518 *
 519 * The three key local variables below are:
 520 *    q  - a linked-list queue of cpuset pointers, used to implement a
 521 *         top-down scan of all cpusets.  This scan loads a pointer
 522 *         to each cpuset marked is_sched_load_balance into the
 523 *         array 'csa'.  For our purposes, rebuilding the schedulers
 524 *         sched domains, we can ignore !is_sched_load_balance cpusets.
 525 *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
 526 *         that need to be load balanced, for convenient iterative
 527 *         access by the subsequent code that finds the best partition,
 528 *         i.e the set of domains (subsets) of CPUs such that the
 529 *         cpus_allowed of every cpuset marked is_sched_load_balance
 530 *         is a subset of one of these domains, while there are as
 531 *         many such domains as possible, each as small as possible.
 532 * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
 533 *         the kernel/sched.c routine partition_sched_domains() in a
 534 *         convenient format, that can be easily compared to the prior
 535 *         value to determine what partition elements (sched domains)
 536 *         were changed (added or removed.)
 537 *
 538 * Finding the best partition (set of domains):
 539 *      The triple nested loops below over i, j, k scan over the
 540 *      load balanced cpusets (using the array of cpuset pointers in
 541 *      csa[]) looking for pairs of cpusets that have overlapping
 542 *      cpus_allowed, but which don't have the same 'pn' partition
 543 *      number and gives them in the same partition number.  It keeps
 544 *      looping on the 'restart' label until it can no longer find
 545 *      any such pairs.
 546 *
 547 *      The union of the cpus_allowed masks from the set of
 548 *      all cpusets having the same 'pn' value then form the one
 549 *      element of the partition (one sched domain) to be passed to
 550 *      partition_sched_domains().
 551 */
 552static int generate_sched_domains(cpumask_var_t **domains,
 553                        struct sched_domain_attr **attributes)
 554{
 555        LIST_HEAD(q);           /* queue of cpusets to be scanned */
 556        struct cpuset *cp;      /* scans q */
 557        struct cpuset **csa;    /* array of all cpuset ptrs */
 558        int csn;                /* how many cpuset ptrs in csa so far */
 559        int i, j, k;            /* indices for partition finding loops */
 560        cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
 561        struct sched_domain_attr *dattr;  /* attributes for custom domains */
 562        int ndoms = 0;          /* number of sched domains in result */
 563        int nslot;              /* next empty doms[] struct cpumask slot */
 564
 565        doms = NULL;
 566        dattr = NULL;
 567        csa = NULL;
 568
 569        /* Special case for the 99% of systems with one, full, sched domain */
 570        if (is_sched_load_balance(&top_cpuset)) {
 571                ndoms = 1;
 572                doms = alloc_sched_domains(ndoms);
 573                if (!doms)
 574                        goto done;
 575
 576                dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
 577                if (dattr) {
 578                        *dattr = SD_ATTR_INIT;
 579                        update_domain_attr_tree(dattr, &top_cpuset);
 580                }
 581                cpumask_copy(doms[0], top_cpuset.cpus_allowed);
 582
 583                goto done;
 584        }
 585
 586        csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
 587        if (!csa)
 588                goto done;
 589        csn = 0;
 590
 591        list_add(&top_cpuset.stack_list, &q);
 592        while (!list_empty(&q)) {
 593                struct cgroup *cont;
 594                struct cpuset *child;   /* scans child cpusets of cp */
 595
 596                cp = list_first_entry(&q, struct cpuset, stack_list);
 597                list_del(q.next);
 598
 599                if (cpumask_empty(cp->cpus_allowed))
 600                        continue;
 601
 602                /*
 603                 * All child cpusets contain a subset of the parent's cpus, so
 604                 * just skip them, and then we call update_domain_attr_tree()
 605                 * to calc relax_domain_level of the corresponding sched
 606                 * domain.
 607                 */
 608                if (is_sched_load_balance(cp)) {
 609                        csa[csn++] = cp;
 610                        continue;
 611                }
 612
 613                list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
 614                        child = cgroup_cs(cont);
 615                        list_add_tail(&child->stack_list, &q);
 616                }
 617        }
 618
 619        for (i = 0; i < csn; i++)
 620                csa[i]->pn = i;
 621        ndoms = csn;
 622
 623restart:
 624        /* Find the best partition (set of sched domains) */
 625        for (i = 0; i < csn; i++) {
 626                struct cpuset *a = csa[i];
 627                int apn = a->pn;
 628
 629                for (j = 0; j < csn; j++) {
 630                        struct cpuset *b = csa[j];
 631                        int bpn = b->pn;
 632
 633                        if (apn != bpn && cpusets_overlap(a, b)) {
 634                                for (k = 0; k < csn; k++) {
 635                                        struct cpuset *c = csa[k];
 636
 637                                        if (c->pn == bpn)
 638                                                c->pn = apn;
 639                                }
 640                                ndoms--;        /* one less element */
 641                                goto restart;
 642                        }
 643                }
 644        }
 645
 646        /*
 647         * Now we know how many domains to create.
 648         * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
 649         */
 650        doms = alloc_sched_domains(ndoms);
 651        if (!doms)
 652                goto done;
 653
 654        /*
 655         * The rest of the code, including the scheduler, can deal with
 656         * dattr==NULL case. No need to abort if alloc fails.
 657         */
 658        dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
 659
 660        for (nslot = 0, i = 0; i < csn; i++) {
 661                struct cpuset *a = csa[i];
 662                struct cpumask *dp;
 663                int apn = a->pn;
 664
 665                if (apn < 0) {
 666                        /* Skip completed partitions */
 667                        continue;
 668                }
 669
 670                dp = doms[nslot];
 671
 672                if (nslot == ndoms) {
 673                        static int warnings = 10;
 674                        if (warnings) {
 675                                printk(KERN_WARNING
 676                                 "rebuild_sched_domains confused:"
 677                                  " nslot %d, ndoms %d, csn %d, i %d,"
 678                                  " apn %d\n",
 679                                  nslot, ndoms, csn, i, apn);
 680                                warnings--;
 681                        }
 682                        continue;
 683                }
 684
 685                cpumask_clear(dp);
 686                if (dattr)
 687                        *(dattr + nslot) = SD_ATTR_INIT;
 688                for (j = i; j < csn; j++) {
 689                        struct cpuset *b = csa[j];
 690
 691                        if (apn == b->pn) {
 692                                cpumask_or(dp, dp, b->cpus_allowed);
 693                                if (dattr)
 694                                        update_domain_attr_tree(dattr + nslot, b);
 695
 696                                /* Done with this partition */
 697                                b->pn = -1;
 698                        }
 699                }
 700                nslot++;
 701        }
 702        BUG_ON(nslot != ndoms);
 703
 704done:
 705        kfree(csa);
 706
 707        /*
 708         * Fallback to the default domain if kmalloc() failed.
 709         * See comments in partition_sched_domains().
 710         */
 711        if (doms == NULL)
 712                ndoms = 1;
 713
 714        *domains    = doms;
 715        *attributes = dattr;
 716        return ndoms;
 717}
 718
 719/*
 720 * Rebuild scheduler domains.
 721 *
 722 * Call with neither cgroup_mutex held nor within get_online_cpus().
 723 * Takes both cgroup_mutex and get_online_cpus().
 724 *
 725 * Cannot be directly called from cpuset code handling changes
 726 * to the cpuset pseudo-filesystem, because it cannot be called
 727 * from code that already holds cgroup_mutex.
 728 */
 729static void do_rebuild_sched_domains(struct work_struct *unused)
 730{
 731        struct sched_domain_attr *attr;
 732        cpumask_var_t *doms;
 733        int ndoms;
 734
 735        get_online_cpus();
 736
 737        /* Generate domain masks and attrs */
 738        cgroup_lock();
 739        ndoms = generate_sched_domains(&doms, &attr);
 740        cgroup_unlock();
 741
 742        /* Have scheduler rebuild the domains */
 743        partition_sched_domains(ndoms, doms, attr);
 744
 745        put_online_cpus();
 746}
 747#else /* !CONFIG_SMP */
 748static void do_rebuild_sched_domains(struct work_struct *unused)
 749{
 750}
 751
 752static int generate_sched_domains(cpumask_var_t **domains,
 753                        struct sched_domain_attr **attributes)
 754{
 755        *domains = NULL;
 756        return 1;
 757}
 758#endif /* CONFIG_SMP */
 759
 760static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
 761
 762/*
 763 * Rebuild scheduler domains, asynchronously via workqueue.
 764 *
 765 * If the flag 'sched_load_balance' of any cpuset with non-empty
 766 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
 767 * which has that flag enabled, or if any cpuset with a non-empty
 768 * 'cpus' is removed, then call this routine to rebuild the
 769 * scheduler's dynamic sched domains.
 770 *
 771 * The rebuild_sched_domains() and partition_sched_domains()
 772 * routines must nest cgroup_lock() inside get_online_cpus(),
 773 * but such cpuset changes as these must nest that locking the
 774 * other way, holding cgroup_lock() for much of the code.
 775 *
 776 * So in order to avoid an ABBA deadlock, the cpuset code handling
 777 * these user changes delegates the actual sched domain rebuilding
 778 * to a separate workqueue thread, which ends up processing the
 779 * above do_rebuild_sched_domains() function.
 780 */
 781static void async_rebuild_sched_domains(void)
 782{
 783        queue_work(cpuset_wq, &rebuild_sched_domains_work);
 784}
 785
 786/*
 787 * Accomplishes the same scheduler domain rebuild as the above
 788 * async_rebuild_sched_domains(), however it directly calls the
 789 * rebuild routine synchronously rather than calling it via an
 790 * asynchronous work thread.
 791 *
 792 * This can only be called from code that is not holding
 793 * cgroup_mutex (not nested in a cgroup_lock() call.)
 794 */
 795void rebuild_sched_domains(void)
 796{
 797        do_rebuild_sched_domains(NULL);
 798}
 799
 800/**
 801 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
 802 * @tsk: task to test
 803 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
 804 *
 805 * Call with cgroup_mutex held.  May take callback_mutex during call.
 806 * Called for each task in a cgroup by cgroup_scan_tasks().
 807 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
 808 * words, if its mask is not equal to its cpuset's mask).
 809 */
 810static int cpuset_test_cpumask(struct task_struct *tsk,
 811                               struct cgroup_scanner *scan)
 812{
 813        return !cpumask_equal(&tsk->cpus_allowed,
 814                        (cgroup_cs(scan->cg))->cpus_allowed);
 815}
 816
 817/**
 818 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
 819 * @tsk: task to test
 820 * @scan: struct cgroup_scanner containing the cgroup of the task
 821 *
 822 * Called by cgroup_scan_tasks() for each task in a cgroup whose
 823 * cpus_allowed mask needs to be changed.
 824 *
 825 * We don't need to re-check for the cgroup/cpuset membership, since we're
 826 * holding cgroup_lock() at this point.
 827 */
 828static void cpuset_change_cpumask(struct task_struct *tsk,
 829                                  struct cgroup_scanner *scan)
 830{
 831        set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
 832}
 833
 834/**
 835 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
 836 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
 837 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
 838 *
 839 * Called with cgroup_mutex held
 840 *
 841 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
 842 * calling callback functions for each.
 843 *
 844 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
 845 * if @heap != NULL.
 846 */
 847static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 848{
 849        struct cgroup_scanner scan;
 850
 851        scan.cg = cs->css.cgroup;
 852        scan.test_task = cpuset_test_cpumask;
 853        scan.process_task = cpuset_change_cpumask;
 854        scan.heap = heap;
 855        cgroup_scan_tasks(&scan);
 856}
 857
 858/**
 859 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
 860 * @cs: the cpuset to consider
 861 * @buf: buffer of cpu numbers written to this cpuset
 862 */
 863static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 864                          const char *buf)
 865{
 866        struct ptr_heap heap;
 867        int retval;
 868        int is_load_balanced;
 869
 870        /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
 871        if (cs == &top_cpuset)
 872                return -EACCES;
 873
 874        /*
 875         * An empty cpus_allowed is ok only if the cpuset has no tasks.
 876         * Since cpulist_parse() fails on an empty mask, we special case
 877         * that parsing.  The validate_change() call ensures that cpusets
 878         * with tasks have cpus.
 879         */
 880        if (!*buf) {
 881                cpumask_clear(trialcs->cpus_allowed);
 882        } else {
 883                retval = cpulist_parse(buf, trialcs->cpus_allowed);
 884                if (retval < 0)
 885                        return retval;
 886
 887                if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
 888                        return -EINVAL;
 889        }
 890        retval = validate_change(cs, trialcs);
 891        if (retval < 0)
 892                return retval;
 893
 894        /* Nothing to do if the cpus didn't change */
 895        if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
 896                return 0;
 897
 898        retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
 899        if (retval)
 900                return retval;
 901
 902        is_load_balanced = is_sched_load_balance(trialcs);
 903
 904        mutex_lock(&callback_mutex);
 905        cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
 906        mutex_unlock(&callback_mutex);
 907
 908        /*
 909         * Scan tasks in the cpuset, and update the cpumasks of any
 910         * that need an update.
 911         */
 912        update_tasks_cpumask(cs, &heap);
 913
 914        heap_free(&heap);
 915
 916        if (is_load_balanced)
 917                async_rebuild_sched_domains();
 918        return 0;
 919}
 920
 921/*
 922 * cpuset_migrate_mm
 923 *
 924 *    Migrate memory region from one set of nodes to another.
 925 *
 926 *    Temporarilly set tasks mems_allowed to target nodes of migration,
 927 *    so that the migration code can allocate pages on these nodes.
 928 *
 929 *    Call holding cgroup_mutex, so current's cpuset won't change
 930 *    during this call, as manage_mutex holds off any cpuset_attach()
 931 *    calls.  Therefore we don't need to take task_lock around the
 932 *    call to guarantee_online_mems(), as we know no one is changing
 933 *    our task's cpuset.
 934 *
 935 *    While the mm_struct we are migrating is typically from some
 936 *    other task, the task_struct mems_allowed that we are hacking
 937 *    is for our current task, which must allocate new pages for that
 938 *    migrating memory region.
 939 */
 940
 941static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
 942                                                        const nodemask_t *to)
 943{
 944        struct task_struct *tsk = current;
 945
 946        tsk->mems_allowed = *to;
 947
 948        do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
 949
 950        guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
 951}
 952
 953/*
 954 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
 955 * @tsk: the task to change
 956 * @newmems: new nodes that the task will be set
 957 *
 958 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
 959 * we structure updates as setting all new allowed nodes, then clearing newly
 960 * disallowed ones.
 961 */
 962static void cpuset_change_task_nodemask(struct task_struct *tsk,
 963                                        nodemask_t *newmems)
 964{
 965        bool need_loop;
 966
 967repeat:
 968        /*
 969         * Allow tasks that have access to memory reserves because they have
 970         * been OOM killed to get memory anywhere.
 971         */
 972        if (unlikely(test_thread_flag(TIF_MEMDIE)))
 973                return;
 974        if (current->flags & PF_EXITING) /* Let dying task have memory */
 975                return;
 976
 977        task_lock(tsk);
 978        /*
 979         * Determine if a loop is necessary if another thread is doing
 980         * get_mems_allowed().  If at least one node remains unchanged and
 981         * tsk does not have a mempolicy, then an empty nodemask will not be
 982         * possible when mems_allowed is larger than a word.
 983         */
 984        need_loop = task_has_mempolicy(tsk) ||
 985                        !nodes_intersects(*newmems, tsk->mems_allowed);
 986        nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
 987        mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
 988
 989        /*
 990         * ensure checking ->mems_allowed_change_disable after setting all new
 991         * allowed nodes.
 992         *
 993         * the read-side task can see an nodemask with new allowed nodes and
 994         * old allowed nodes. and if it allocates page when cpuset clears newly
 995         * disallowed ones continuous, it can see the new allowed bits.
 996         *
 997         * And if setting all new allowed nodes is after the checking, setting
 998         * all new allowed nodes and clearing newly disallowed ones will be done
 999         * continuous, and the read-side task may find no node to alloc page.
1000         */
1001        smp_mb();
1002
1003        /*
1004         * Allocation of memory is very fast, we needn't sleep when waiting
1005         * for the read-side.
1006         */
1007        while (need_loop && ACCESS_ONCE(tsk->mems_allowed_change_disable)) {
1008                task_unlock(tsk);
1009                if (!task_curr(tsk))
1010                        yield();
1011                goto repeat;
1012        }
1013
1014        /*
1015         * ensure checking ->mems_allowed_change_disable before clearing all new
1016         * disallowed nodes.
1017         *
1018         * if clearing newly disallowed bits before the checking, the read-side
1019         * task may find no node to alloc page.
1020         */
1021        smp_mb();
1022
1023        mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1024        tsk->mems_allowed = *newmems;
1025        task_unlock(tsk);
1026}
1027
1028/*
1029 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1030 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1031 * memory_migrate flag is set. Called with cgroup_mutex held.
1032 */
1033static void cpuset_change_nodemask(struct task_struct *p,
1034                                   struct cgroup_scanner *scan)
1035{
1036        struct mm_struct *mm;
1037        struct cpuset *cs;
1038        int migrate;
1039        const nodemask_t *oldmem = scan->data;
1040        static nodemask_t newmems;      /* protected by cgroup_mutex */
1041
1042        cs = cgroup_cs(scan->cg);
1043        guarantee_online_mems(cs, &newmems);
1044
1045        cpuset_change_task_nodemask(p, &newmems);
1046
1047        mm = get_task_mm(p);
1048        if (!mm)
1049                return;
1050
1051        migrate = is_memory_migrate(cs);
1052
1053        mpol_rebind_mm(mm, &cs->mems_allowed);
1054        if (migrate)
1055                cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1056        mmput(mm);
1057}
1058
1059static void *cpuset_being_rebound;
1060
1061/**
1062 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1063 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1064 * @oldmem: old mems_allowed of cpuset cs
1065 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1066 *
1067 * Called with cgroup_mutex held
1068 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1069 * if @heap != NULL.
1070 */
1071static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1072                                 struct ptr_heap *heap)
1073{
1074        struct cgroup_scanner scan;
1075
1076        cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1077
1078        scan.cg = cs->css.cgroup;
1079        scan.test_task = NULL;
1080        scan.process_task = cpuset_change_nodemask;
1081        scan.heap = heap;
1082        scan.data = (nodemask_t *)oldmem;
1083
1084        /*
1085         * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1086         * take while holding tasklist_lock.  Forks can happen - the
1087         * mpol_dup() cpuset_being_rebound check will catch such forks,
1088         * and rebind their vma mempolicies too.  Because we still hold
1089         * the global cgroup_mutex, we know that no other rebind effort
1090         * will be contending for the global variable cpuset_being_rebound.
1091         * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1092         * is idempotent.  Also migrate pages in each mm to new nodes.
1093         */
1094        cgroup_scan_tasks(&scan);
1095
1096        /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1097        cpuset_being_rebound = NULL;
1098}
1099
1100/*
1101 * Handle user request to change the 'mems' memory placement
1102 * of a cpuset.  Needs to validate the request, update the
1103 * cpusets mems_allowed, and for each task in the cpuset,
1104 * update mems_allowed and rebind task's mempolicy and any vma
1105 * mempolicies and if the cpuset is marked 'memory_migrate',
1106 * migrate the tasks pages to the new memory.
1107 *
1108 * Call with cgroup_mutex held.  May take callback_mutex during call.
1109 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1110 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1111 * their mempolicies to the cpusets new mems_allowed.
1112 */
1113static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1114                           const char *buf)
1115{
1116        NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1117        int retval;
1118        struct ptr_heap heap;
1119
1120        if (!oldmem)
1121                return -ENOMEM;
1122
1123        /*
1124         * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1125         * it's read-only
1126         */
1127        if (cs == &top_cpuset) {
1128                retval = -EACCES;
1129                goto done;
1130        }
1131
1132        /*
1133         * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1134         * Since nodelist_parse() fails on an empty mask, we special case
1135         * that parsing.  The validate_change() call ensures that cpusets
1136         * with tasks have memory.
1137         */
1138        if (!*buf) {
1139                nodes_clear(trialcs->mems_allowed);
1140        } else {
1141                retval = nodelist_parse(buf, trialcs->mems_allowed);
1142                if (retval < 0)
1143                        goto done;
1144
1145                if (!nodes_subset(trialcs->mems_allowed,
1146                                node_states[N_HIGH_MEMORY])) {
1147                        retval =  -EINVAL;
1148                        goto done;
1149                }
1150        }
1151        *oldmem = cs->mems_allowed;
1152        if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1153                retval = 0;             /* Too easy - nothing to do */
1154                goto done;
1155        }
1156        retval = validate_change(cs, trialcs);
1157        if (retval < 0)
1158                goto done;
1159
1160        retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1161        if (retval < 0)
1162                goto done;
1163
1164        mutex_lock(&callback_mutex);
1165        cs->mems_allowed = trialcs->mems_allowed;
1166        mutex_unlock(&callback_mutex);
1167
1168        update_tasks_nodemask(cs, oldmem, &heap);
1169
1170        heap_free(&heap);
1171done:
1172        NODEMASK_FREE(oldmem);
1173        return retval;
1174}
1175
1176int current_cpuset_is_being_rebound(void)
1177{
1178        return task_cs(current) == cpuset_being_rebound;
1179}
1180
1181static int update_relax_domain_level(struct cpuset *cs, s64 val)
1182{
1183#ifdef CONFIG_SMP
1184        if (val < -1 || val >= sched_domain_level_max)
1185                return -EINVAL;
1186#endif
1187
1188        if (val != cs->relax_domain_level) {
1189                cs->relax_domain_level = val;
1190                if (!cpumask_empty(cs->cpus_allowed) &&
1191                    is_sched_load_balance(cs))
1192                        async_rebuild_sched_domains();
1193        }
1194
1195        return 0;
1196}
1197
1198/*
1199 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1200 * @tsk: task to be updated
1201 * @scan: struct cgroup_scanner containing the cgroup of the task
1202 *
1203 * Called by cgroup_scan_tasks() for each task in a cgroup.
1204 *
1205 * We don't need to re-check for the cgroup/cpuset membership, since we're
1206 * holding cgroup_lock() at this point.
1207 */
1208static void cpuset_change_flag(struct task_struct *tsk,
1209                                struct cgroup_scanner *scan)
1210{
1211        cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1212}
1213
1214/*
1215 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1216 * @cs: the cpuset in which each task's spread flags needs to be changed
1217 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1218 *
1219 * Called with cgroup_mutex held
1220 *
1221 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1222 * calling callback functions for each.
1223 *
1224 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1225 * if @heap != NULL.
1226 */
1227static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1228{
1229        struct cgroup_scanner scan;
1230
1231        scan.cg = cs->css.cgroup;
1232        scan.test_task = NULL;
1233        scan.process_task = cpuset_change_flag;
1234        scan.heap = heap;
1235        cgroup_scan_tasks(&scan);
1236}
1237
1238/*
1239 * update_flag - read a 0 or a 1 in a file and update associated flag
1240 * bit:         the bit to update (see cpuset_flagbits_t)
1241 * cs:          the cpuset to update
1242 * turning_on:  whether the flag is being set or cleared
1243 *
1244 * Call with cgroup_mutex held.
1245 */
1246
1247static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1248                       int turning_on)
1249{
1250        struct cpuset *trialcs;
1251        int balance_flag_changed;
1252        int spread_flag_changed;
1253        struct ptr_heap heap;
1254        int err;
1255
1256        trialcs = alloc_trial_cpuset(cs);
1257        if (!trialcs)
1258                return -ENOMEM;
1259
1260        if (turning_on)
1261                set_bit(bit, &trialcs->flags);
1262        else
1263                clear_bit(bit, &trialcs->flags);
1264
1265        err = validate_change(cs, trialcs);
1266        if (err < 0)
1267                goto out;
1268
1269        err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1270        if (err < 0)
1271                goto out;
1272
1273        balance_flag_changed = (is_sched_load_balance(cs) !=
1274                                is_sched_load_balance(trialcs));
1275
1276        spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1277                        || (is_spread_page(cs) != is_spread_page(trialcs)));
1278
1279        mutex_lock(&callback_mutex);
1280        cs->flags = trialcs->flags;
1281        mutex_unlock(&callback_mutex);
1282
1283        if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1284                async_rebuild_sched_domains();
1285
1286        if (spread_flag_changed)
1287                update_tasks_flags(cs, &heap);
1288        heap_free(&heap);
1289out:
1290        free_trial_cpuset(trialcs);
1291        return err;
1292}
1293
1294/*
1295 * Frequency meter - How fast is some event occurring?
1296 *
1297 * These routines manage a digitally filtered, constant time based,
1298 * event frequency meter.  There are four routines:
1299 *   fmeter_init() - initialize a frequency meter.
1300 *   fmeter_markevent() - called each time the event happens.
1301 *   fmeter_getrate() - returns the recent rate of such events.
1302 *   fmeter_update() - internal routine used to update fmeter.
1303 *
1304 * A common data structure is passed to each of these routines,
1305 * which is used to keep track of the state required to manage the
1306 * frequency meter and its digital filter.
1307 *
1308 * The filter works on the number of events marked per unit time.
1309 * The filter is single-pole low-pass recursive (IIR).  The time unit
1310 * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1311 * simulate 3 decimal digits of precision (multiplied by 1000).
1312 *
1313 * With an FM_COEF of 933, and a time base of 1 second, the filter
1314 * has a half-life of 10 seconds, meaning that if the events quit
1315 * happening, then the rate returned from the fmeter_getrate()
1316 * will be cut in half each 10 seconds, until it converges to zero.
1317 *
1318 * It is not worth doing a real infinitely recursive filter.  If more
1319 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1320 * just compute FM_MAXTICKS ticks worth, by which point the level
1321 * will be stable.
1322 *
1323 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1324 * arithmetic overflow in the fmeter_update() routine.
1325 *
1326 * Given the simple 32 bit integer arithmetic used, this meter works
1327 * best for reporting rates between one per millisecond (msec) and
1328 * one per 32 (approx) seconds.  At constant rates faster than one
1329 * per msec it maxes out at values just under 1,000,000.  At constant
1330 * rates between one per msec, and one per second it will stabilize
1331 * to a value N*1000, where N is the rate of events per second.
1332 * At constant rates between one per second and one per 32 seconds,
1333 * it will be choppy, moving up on the seconds that have an event,
1334 * and then decaying until the next event.  At rates slower than
1335 * about one in 32 seconds, it decays all the way back to zero between
1336 * each event.
1337 */
1338
1339#define FM_COEF 933             /* coefficient for half-life of 10 secs */
1340#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1341#define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1342#define FM_SCALE 1000           /* faux fixed point scale */
1343
1344/* Initialize a frequency meter */
1345static void fmeter_init(struct fmeter *fmp)
1346{
1347        fmp->cnt = 0;
1348        fmp->val = 0;
1349        fmp->time = 0;
1350        spin_lock_init(&fmp->lock);
1351}
1352
1353/* Internal meter update - process cnt events and update value */
1354static void fmeter_update(struct fmeter *fmp)
1355{
1356        time_t now = get_seconds();
1357        time_t ticks = now - fmp->time;
1358
1359        if (ticks == 0)
1360                return;
1361
1362        ticks = min(FM_MAXTICKS, ticks);
1363        while (ticks-- > 0)
1364                fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1365        fmp->time = now;
1366
1367        fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1368        fmp->cnt = 0;
1369}
1370
1371/* Process any previous ticks, then bump cnt by one (times scale). */
1372static void fmeter_markevent(struct fmeter *fmp)
1373{
1374        spin_lock(&fmp->lock);
1375        fmeter_update(fmp);
1376        fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1377        spin_unlock(&fmp->lock);
1378}
1379
1380/* Process any previous ticks, then return current value. */
1381static int fmeter_getrate(struct fmeter *fmp)
1382{
1383        int val;
1384
1385        spin_lock(&fmp->lock);
1386        fmeter_update(fmp);
1387        val = fmp->val;
1388        spin_unlock(&fmp->lock);
1389        return val;
1390}
1391
1392/*
1393 * Protected by cgroup_lock. The nodemasks must be stored globally because
1394 * dynamically allocating them is not allowed in can_attach, and they must
1395 * persist until attach.
1396 */
1397static cpumask_var_t cpus_attach;
1398static nodemask_t cpuset_attach_nodemask_from;
1399static nodemask_t cpuset_attach_nodemask_to;
1400
1401/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1402static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
1403                             struct cgroup_taskset *tset)
1404{
1405        struct cpuset *cs = cgroup_cs(cgrp);
1406        struct task_struct *task;
1407        int ret;
1408
1409        if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1410                return -ENOSPC;
1411
1412        cgroup_taskset_for_each(task, cgrp, tset) {
1413                /*
1414                 * Kthreads bound to specific cpus cannot be moved to a new
1415                 * cpuset; we cannot change their cpu affinity and
1416                 * isolating such threads by their set of allowed nodes is
1417                 * unnecessary.  Thus, cpusets are not applicable for such
1418                 * threads.  This prevents checking for success of
1419                 * set_cpus_allowed_ptr() on all attached tasks before
1420                 * cpus_allowed may be changed.
1421                 */
1422                if (task->flags & PF_THREAD_BOUND)
1423                        return -EINVAL;
1424                if ((ret = security_task_setscheduler(task)))
1425                        return ret;
1426        }
1427
1428        /* prepare for attach */
1429        if (cs == &top_cpuset)
1430                cpumask_copy(cpus_attach, cpu_possible_mask);
1431        else
1432                guarantee_online_cpus(cs, cpus_attach);
1433
1434        guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1435
1436        return 0;
1437}
1438
1439static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
1440                          struct cgroup_taskset *tset)
1441{
1442        struct mm_struct *mm;
1443        struct task_struct *task;
1444        struct task_struct *leader = cgroup_taskset_first(tset);
1445        struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1446        struct cpuset *cs = cgroup_cs(cgrp);
1447        struct cpuset *oldcs = cgroup_cs(oldcgrp);
1448
1449        cgroup_taskset_for_each(task, cgrp, tset) {
1450                /*
1451                 * can_attach beforehand should guarantee that this doesn't
1452                 * fail.  TODO: have a better way to handle failure here
1453                 */
1454                WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1455
1456                cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1457                cpuset_update_task_spread_flag(cs, task);
1458        }
1459
1460        /*
1461         * Change mm, possibly for multiple threads in a threadgroup. This is
1462         * expensive and may sleep.
1463         */
1464        cpuset_attach_nodemask_from = oldcs->mems_allowed;
1465        cpuset_attach_nodemask_to = cs->mems_allowed;
1466        mm = get_task_mm(leader);
1467        if (mm) {
1468                mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1469                if (is_memory_migrate(cs))
1470                        cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1471                                          &cpuset_attach_nodemask_to);
1472                mmput(mm);
1473        }
1474}
1475
1476/* The various types of files and directories in a cpuset file system */
1477
1478typedef enum {
1479        FILE_MEMORY_MIGRATE,
1480        FILE_CPULIST,
1481        FILE_MEMLIST,
1482        FILE_CPU_EXCLUSIVE,
1483        FILE_MEM_EXCLUSIVE,
1484        FILE_MEM_HARDWALL,
1485        FILE_SCHED_LOAD_BALANCE,
1486        FILE_SCHED_RELAX_DOMAIN_LEVEL,
1487        FILE_MEMORY_PRESSURE_ENABLED,
1488        FILE_MEMORY_PRESSURE,
1489        FILE_SPREAD_PAGE,
1490        FILE_SPREAD_SLAB,
1491} cpuset_filetype_t;
1492
1493static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1494{
1495        int retval = 0;
1496        struct cpuset *cs = cgroup_cs(cgrp);
1497        cpuset_filetype_t type = cft->private;
1498
1499        if (!cgroup_lock_live_group(cgrp))
1500                return -ENODEV;
1501
1502        switch (type) {
1503        case FILE_CPU_EXCLUSIVE:
1504                retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1505                break;
1506        case FILE_MEM_EXCLUSIVE:
1507                retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1508                break;
1509        case FILE_MEM_HARDWALL:
1510                retval = update_flag(CS_MEM_HARDWALL, cs, val);
1511                break;
1512        case FILE_SCHED_LOAD_BALANCE:
1513                retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1514                break;
1515        case FILE_MEMORY_MIGRATE:
1516                retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1517                break;
1518        case FILE_MEMORY_PRESSURE_ENABLED:
1519                cpuset_memory_pressure_enabled = !!val;
1520                break;
1521        case FILE_MEMORY_PRESSURE:
1522                retval = -EACCES;
1523                break;
1524        case FILE_SPREAD_PAGE:
1525                retval = update_flag(CS_SPREAD_PAGE, cs, val);
1526                break;
1527        case FILE_SPREAD_SLAB:
1528                retval = update_flag(CS_SPREAD_SLAB, cs, val);
1529                break;
1530        default:
1531                retval = -EINVAL;
1532                break;
1533        }
1534        cgroup_unlock();
1535        return retval;
1536}
1537
1538static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1539{
1540        int retval = 0;
1541        struct cpuset *cs = cgroup_cs(cgrp);
1542        cpuset_filetype_t type = cft->private;
1543
1544        if (!cgroup_lock_live_group(cgrp))
1545                return -ENODEV;
1546
1547        switch (type) {
1548        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1549                retval = update_relax_domain_level(cs, val);
1550                break;
1551        default:
1552                retval = -EINVAL;
1553                break;
1554        }
1555        cgroup_unlock();
1556        return retval;
1557}
1558
1559/*
1560 * Common handling for a write to a "cpus" or "mems" file.
1561 */
1562static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1563                                const char *buf)
1564{
1565        int retval = 0;
1566        struct cpuset *cs = cgroup_cs(cgrp);
1567        struct cpuset *trialcs;
1568
1569        if (!cgroup_lock_live_group(cgrp))
1570                return -ENODEV;
1571
1572        trialcs = alloc_trial_cpuset(cs);
1573        if (!trialcs) {
1574                retval = -ENOMEM;
1575                goto out;
1576        }
1577
1578        switch (cft->private) {
1579        case FILE_CPULIST:
1580                retval = update_cpumask(cs, trialcs, buf);
1581                break;
1582        case FILE_MEMLIST:
1583                retval = update_nodemask(cs, trialcs, buf);
1584                break;
1585        default:
1586                retval = -EINVAL;
1587                break;
1588        }
1589
1590        free_trial_cpuset(trialcs);
1591out:
1592        cgroup_unlock();
1593        return retval;
1594}
1595
1596/*
1597 * These ascii lists should be read in a single call, by using a user
1598 * buffer large enough to hold the entire map.  If read in smaller
1599 * chunks, there is no guarantee of atomicity.  Since the display format
1600 * used, list of ranges of sequential numbers, is variable length,
1601 * and since these maps can change value dynamically, one could read
1602 * gibberish by doing partial reads while a list was changing.
1603 * A single large read to a buffer that crosses a page boundary is
1604 * ok, because the result being copied to user land is not recomputed
1605 * across a page fault.
1606 */
1607
1608static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1609{
1610        size_t count;
1611
1612        mutex_lock(&callback_mutex);
1613        count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1614        mutex_unlock(&callback_mutex);
1615
1616        return count;
1617}
1618
1619static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1620{
1621        size_t count;
1622
1623        mutex_lock(&callback_mutex);
1624        count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1625        mutex_unlock(&callback_mutex);
1626
1627        return count;
1628}
1629
1630static ssize_t cpuset_common_file_read(struct cgroup *cont,
1631                                       struct cftype *cft,
1632                                       struct file *file,
1633                                       char __user *buf,
1634                                       size_t nbytes, loff_t *ppos)
1635{
1636        struct cpuset *cs = cgroup_cs(cont);
1637        cpuset_filetype_t type = cft->private;
1638        char *page;
1639        ssize_t retval = 0;
1640        char *s;
1641
1642        if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1643                return -ENOMEM;
1644
1645        s = page;
1646
1647        switch (type) {
1648        case FILE_CPULIST:
1649                s += cpuset_sprintf_cpulist(s, cs);
1650                break;
1651        case FILE_MEMLIST:
1652                s += cpuset_sprintf_memlist(s, cs);
1653                break;
1654        default:
1655                retval = -EINVAL;
1656                goto out;
1657        }
1658        *s++ = '\n';
1659
1660        retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1661out:
1662        free_page((unsigned long)page);
1663        return retval;
1664}
1665
1666static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1667{
1668        struct cpuset *cs = cgroup_cs(cont);
1669        cpuset_filetype_t type = cft->private;
1670        switch (type) {
1671        case FILE_CPU_EXCLUSIVE:
1672                return is_cpu_exclusive(cs);
1673        case FILE_MEM_EXCLUSIVE:
1674                return is_mem_exclusive(cs);
1675        case FILE_MEM_HARDWALL:
1676                return is_mem_hardwall(cs);
1677        case FILE_SCHED_LOAD_BALANCE:
1678                return is_sched_load_balance(cs);
1679        case FILE_MEMORY_MIGRATE:
1680                return is_memory_migrate(cs);
1681        case FILE_MEMORY_PRESSURE_ENABLED:
1682                return cpuset_memory_pressure_enabled;
1683        case FILE_MEMORY_PRESSURE:
1684                return fmeter_getrate(&cs->fmeter);
1685        case FILE_SPREAD_PAGE:
1686                return is_spread_page(cs);
1687        case FILE_SPREAD_SLAB:
1688                return is_spread_slab(cs);
1689        default:
1690                BUG();
1691        }
1692
1693        /* Unreachable but makes gcc happy */
1694        return 0;
1695}
1696
1697static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1698{
1699        struct cpuset *cs = cgroup_cs(cont);
1700        cpuset_filetype_t type = cft->private;
1701        switch (type) {
1702        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1703                return cs->relax_domain_level;
1704        default:
1705                BUG();
1706        }
1707
1708        /* Unrechable but makes gcc happy */
1709        return 0;
1710}
1711
1712
1713/*
1714 * for the common functions, 'private' gives the type of file
1715 */
1716
1717static struct cftype files[] = {
1718        {
1719                .name = "cpus",
1720                .read = cpuset_common_file_read,
1721                .write_string = cpuset_write_resmask,
1722                .max_write_len = (100U + 6 * NR_CPUS),
1723                .private = FILE_CPULIST,
1724        },
1725
1726        {
1727                .name = "mems",
1728                .read = cpuset_common_file_read,
1729                .write_string = cpuset_write_resmask,
1730                .max_write_len = (100U + 6 * MAX_NUMNODES),
1731                .private = FILE_MEMLIST,
1732        },
1733
1734        {
1735                .name = "cpu_exclusive",
1736                .read_u64 = cpuset_read_u64,
1737                .write_u64 = cpuset_write_u64,
1738                .private = FILE_CPU_EXCLUSIVE,
1739        },
1740
1741        {
1742                .name = "mem_exclusive",
1743                .read_u64 = cpuset_read_u64,
1744                .write_u64 = cpuset_write_u64,
1745                .private = FILE_MEM_EXCLUSIVE,
1746        },
1747
1748        {
1749                .name = "mem_hardwall",
1750                .read_u64 = cpuset_read_u64,
1751                .write_u64 = cpuset_write_u64,
1752                .private = FILE_MEM_HARDWALL,
1753        },
1754
1755        {
1756                .name = "sched_load_balance",
1757                .read_u64 = cpuset_read_u64,
1758                .write_u64 = cpuset_write_u64,
1759                .private = FILE_SCHED_LOAD_BALANCE,
1760        },
1761
1762        {
1763                .name = "sched_relax_domain_level",
1764                .read_s64 = cpuset_read_s64,
1765                .write_s64 = cpuset_write_s64,
1766                .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1767        },
1768
1769        {
1770                .name = "memory_migrate",
1771                .read_u64 = cpuset_read_u64,
1772                .write_u64 = cpuset_write_u64,
1773                .private = FILE_MEMORY_MIGRATE,
1774        },
1775
1776        {
1777                .name = "memory_pressure",
1778                .read_u64 = cpuset_read_u64,
1779                .write_u64 = cpuset_write_u64,
1780                .private = FILE_MEMORY_PRESSURE,
1781                .mode = S_IRUGO,
1782        },
1783
1784        {
1785                .name = "memory_spread_page",
1786                .read_u64 = cpuset_read_u64,
1787                .write_u64 = cpuset_write_u64,
1788                .private = FILE_SPREAD_PAGE,
1789        },
1790
1791        {
1792                .name = "memory_spread_slab",
1793                .read_u64 = cpuset_read_u64,
1794                .write_u64 = cpuset_write_u64,
1795                .private = FILE_SPREAD_SLAB,
1796        },
1797};
1798
1799static struct cftype cft_memory_pressure_enabled = {
1800        .name = "memory_pressure_enabled",
1801        .read_u64 = cpuset_read_u64,
1802        .write_u64 = cpuset_write_u64,
1803        .private = FILE_MEMORY_PRESSURE_ENABLED,
1804};
1805
1806static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1807{
1808        int err;
1809
1810        err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1811        if (err)
1812                return err;
1813        /* memory_pressure_enabled is in root cpuset only */
1814        if (!cont->parent)
1815                err = cgroup_add_file(cont, ss,
1816                                      &cft_memory_pressure_enabled);
1817        return err;
1818}
1819
1820/*
1821 * post_clone() is called during cgroup_create() when the
1822 * clone_children mount argument was specified.  The cgroup
1823 * can not yet have any tasks.
1824 *
1825 * Currently we refuse to set up the cgroup - thereby
1826 * refusing the task to be entered, and as a result refusing
1827 * the sys_unshare() or clone() which initiated it - if any
1828 * sibling cpusets have exclusive cpus or mem.
1829 *
1830 * If this becomes a problem for some users who wish to
1831 * allow that scenario, then cpuset_post_clone() could be
1832 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1833 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1834 * held.
1835 */
1836static void cpuset_post_clone(struct cgroup_subsys *ss,
1837                              struct cgroup *cgroup)
1838{
1839        struct cgroup *parent, *child;
1840        struct cpuset *cs, *parent_cs;
1841
1842        parent = cgroup->parent;
1843        list_for_each_entry(child, &parent->children, sibling) {
1844                cs = cgroup_cs(child);
1845                if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1846                        return;
1847        }
1848        cs = cgroup_cs(cgroup);
1849        parent_cs = cgroup_cs(parent);
1850
1851        mutex_lock(&callback_mutex);
1852        cs->mems_allowed = parent_cs->mems_allowed;
1853        cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1854        mutex_unlock(&callback_mutex);
1855        return;
1856}
1857
1858/*
1859 *      cpuset_create - create a cpuset
1860 *      ss:     cpuset cgroup subsystem
1861 *      cont:   control group that the new cpuset will be part of
1862 */
1863
1864static struct cgroup_subsys_state *cpuset_create(
1865        struct cgroup_subsys *ss,
1866        struct cgroup *cont)
1867{
1868        struct cpuset *cs;
1869        struct cpuset *parent;
1870
1871        if (!cont->parent) {
1872                return &top_cpuset.css;
1873        }
1874        parent = cgroup_cs(cont->parent);
1875        cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1876        if (!cs)
1877                return ERR_PTR(-ENOMEM);
1878        if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1879                kfree(cs);
1880                return ERR_PTR(-ENOMEM);
1881        }
1882
1883        cs->flags = 0;
1884        if (is_spread_page(parent))
1885                set_bit(CS_SPREAD_PAGE, &cs->flags);
1886        if (is_spread_slab(parent))
1887                set_bit(CS_SPREAD_SLAB, &cs->flags);
1888        set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1889        cpumask_clear(cs->cpus_allowed);
1890        nodes_clear(cs->mems_allowed);
1891        fmeter_init(&cs->fmeter);
1892        cs->relax_domain_level = -1;
1893
1894        cs->parent = parent;
1895        number_of_cpusets++;
1896        return &cs->css ;
1897}
1898
1899/*
1900 * If the cpuset being removed has its flag 'sched_load_balance'
1901 * enabled, then simulate turning sched_load_balance off, which
1902 * will call async_rebuild_sched_domains().
1903 */
1904
1905static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1906{
1907        struct cpuset *cs = cgroup_cs(cont);
1908
1909        if (is_sched_load_balance(cs))
1910                update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1911
1912        number_of_cpusets--;
1913        free_cpumask_var(cs->cpus_allowed);
1914        kfree(cs);
1915}
1916
1917struct cgroup_subsys cpuset_subsys = {
1918        .name = "cpuset",
1919        .create = cpuset_create,
1920        .destroy = cpuset_destroy,
1921        .can_attach = cpuset_can_attach,
1922        .attach = cpuset_attach,
1923        .populate = cpuset_populate,
1924        .post_clone = cpuset_post_clone,
1925        .subsys_id = cpuset_subsys_id,
1926        .early_init = 1,
1927};
1928
1929/**
1930 * cpuset_init - initialize cpusets at system boot
1931 *
1932 * Description: Initialize top_cpuset and the cpuset internal file system,
1933 **/
1934
1935int __init cpuset_init(void)
1936{
1937        int err = 0;
1938
1939        if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1940                BUG();
1941
1942        cpumask_setall(top_cpuset.cpus_allowed);
1943        nodes_setall(top_cpuset.mems_allowed);
1944
1945        fmeter_init(&top_cpuset.fmeter);
1946        set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1947        top_cpuset.relax_domain_level = -1;
1948
1949        err = register_filesystem(&cpuset_fs_type);
1950        if (err < 0)
1951                return err;
1952
1953        if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1954                BUG();
1955
1956        number_of_cpusets = 1;
1957        return 0;
1958}
1959
1960/**
1961 * cpuset_do_move_task - move a given task to another cpuset
1962 * @tsk: pointer to task_struct the task to move
1963 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1964 *
1965 * Called by cgroup_scan_tasks() for each task in a cgroup.
1966 * Return nonzero to stop the walk through the tasks.
1967 */
1968static void cpuset_do_move_task(struct task_struct *tsk,
1969                                struct cgroup_scanner *scan)
1970{
1971        struct cgroup *new_cgroup = scan->data;
1972
1973        cgroup_attach_task(new_cgroup, tsk);
1974}
1975
1976/**
1977 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1978 * @from: cpuset in which the tasks currently reside
1979 * @to: cpuset to which the tasks will be moved
1980 *
1981 * Called with cgroup_mutex held
1982 * callback_mutex must not be held, as cpuset_attach() will take it.
1983 *
1984 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1985 * calling callback functions for each.
1986 */
1987static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1988{
1989        struct cgroup_scanner scan;
1990
1991        scan.cg = from->css.cgroup;
1992        scan.test_task = NULL; /* select all tasks in cgroup */
1993        scan.process_task = cpuset_do_move_task;
1994        scan.heap = NULL;
1995        scan.data = to->css.cgroup;
1996
1997        if (cgroup_scan_tasks(&scan))
1998                printk(KERN_ERR "move_member_tasks_to_cpuset: "
1999                                "cgroup_scan_tasks failed\n");
2000}
2001
2002/*
2003 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2004 * or memory nodes, we need to walk over the cpuset hierarchy,
2005 * removing that CPU or node from all cpusets.  If this removes the
2006 * last CPU or node from a cpuset, then move the tasks in the empty
2007 * cpuset to its next-highest non-empty parent.
2008 *
2009 * Called with cgroup_mutex held
2010 * callback_mutex must not be held, as cpuset_attach() will take it.
2011 */
2012static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2013{
2014        struct cpuset *parent;
2015
2016        /*
2017         * The cgroup's css_sets list is in use if there are tasks
2018         * in the cpuset; the list is empty if there are none;
2019         * the cs->css.refcnt seems always 0.
2020         */
2021        if (list_empty(&cs->css.cgroup->css_sets))
2022                return;
2023
2024        /*
2025         * Find its next-highest non-empty parent, (top cpuset
2026         * has online cpus, so can't be empty).
2027         */
2028        parent = cs->parent;
2029        while (cpumask_empty(parent->cpus_allowed) ||
2030                        nodes_empty(parent->mems_allowed))
2031                parent = parent->parent;
2032
2033        move_member_tasks_to_cpuset(cs, parent);
2034}
2035
2036/*
2037 * Walk the specified cpuset subtree and look for empty cpusets.
2038 * The tasks of such cpuset must be moved to a parent cpuset.
2039 *
2040 * Called with cgroup_mutex held.  We take callback_mutex to modify
2041 * cpus_allowed and mems_allowed.
2042 *
2043 * This walk processes the tree from top to bottom, completing one layer
2044 * before dropping down to the next.  It always processes a node before
2045 * any of its children.
2046 *
2047 * For now, since we lack memory hot unplug, we'll never see a cpuset
2048 * that has tasks along with an empty 'mems'.  But if we did see such
2049 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2050 */
2051static void scan_for_empty_cpusets(struct cpuset *root)
2052{
2053        LIST_HEAD(queue);
2054        struct cpuset *cp;      /* scans cpusets being updated */
2055        struct cpuset *child;   /* scans child cpusets of cp */
2056        struct cgroup *cont;
2057        static nodemask_t oldmems;      /* protected by cgroup_mutex */
2058
2059        list_add_tail((struct list_head *)&root->stack_list, &queue);
2060
2061        while (!list_empty(&queue)) {
2062                cp = list_first_entry(&queue, struct cpuset, stack_list);
2063                list_del(queue.next);
2064                list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2065                        child = cgroup_cs(cont);
2066                        list_add_tail(&child->stack_list, &queue);
2067                }
2068
2069                /* Continue past cpusets with all cpus, mems online */
2070                if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
2071                    nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2072                        continue;
2073
2074                oldmems = cp->mems_allowed;
2075
2076                /* Remove offline cpus and mems from this cpuset. */
2077                mutex_lock(&callback_mutex);
2078                cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2079                            cpu_active_mask);
2080                nodes_and(cp->mems_allowed, cp->mems_allowed,
2081                                                node_states[N_HIGH_MEMORY]);
2082                mutex_unlock(&callback_mutex);
2083
2084                /* Move tasks from the empty cpuset to a parent */
2085                if (cpumask_empty(cp->cpus_allowed) ||
2086                     nodes_empty(cp->mems_allowed))
2087                        remove_tasks_in_empty_cpuset(cp);
2088                else {
2089                        update_tasks_cpumask(cp, NULL);
2090                        update_tasks_nodemask(cp, &oldmems, NULL);
2091                }
2092        }
2093}
2094
2095/*
2096 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2097 * period.  This is necessary in order to make cpusets transparent
2098 * (of no affect) on systems that are actively using CPU hotplug
2099 * but making no active use of cpusets.
2100 *
2101 * This routine ensures that top_cpuset.cpus_allowed tracks
2102 * cpu_active_mask on each CPU hotplug (cpuhp) event.
2103 *
2104 * Called within get_online_cpus().  Needs to call cgroup_lock()
2105 * before calling generate_sched_domains().
2106 */
2107void cpuset_update_active_cpus(void)
2108{
2109        struct sched_domain_attr *attr;
2110        cpumask_var_t *doms;
2111        int ndoms;
2112
2113        cgroup_lock();
2114        mutex_lock(&callback_mutex);
2115        cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2116        mutex_unlock(&callback_mutex);
2117        scan_for_empty_cpusets(&top_cpuset);
2118        ndoms = generate_sched_domains(&doms, &attr);
2119        cgroup_unlock();
2120
2121        /* Have scheduler rebuild the domains */
2122        partition_sched_domains(ndoms, doms, attr);
2123}
2124
2125#ifdef CONFIG_MEMORY_HOTPLUG
2126/*
2127 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2128 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2129 * See also the previous routine cpuset_track_online_cpus().
2130 */
2131static int cpuset_track_online_nodes(struct notifier_block *self,
2132                                unsigned long action, void *arg)
2133{
2134        static nodemask_t oldmems;      /* protected by cgroup_mutex */
2135
2136        cgroup_lock();
2137        switch (action) {
2138        case MEM_ONLINE:
2139                oldmems = top_cpuset.mems_allowed;
2140                mutex_lock(&callback_mutex);
2141                top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2142                mutex_unlock(&callback_mutex);
2143                update_tasks_nodemask(&top_cpuset, &oldmems, NULL);
2144                break;
2145        case MEM_OFFLINE:
2146                /*
2147                 * needn't update top_cpuset.mems_allowed explicitly because
2148                 * scan_for_empty_cpusets() will update it.
2149                 */
2150                scan_for_empty_cpusets(&top_cpuset);
2151                break;
2152        default:
2153                break;
2154        }
2155        cgroup_unlock();
2156
2157        return NOTIFY_OK;
2158}
2159#endif
2160
2161/**
2162 * cpuset_init_smp - initialize cpus_allowed
2163 *
2164 * Description: Finish top cpuset after cpu, node maps are initialized
2165 **/
2166
2167void __init cpuset_init_smp(void)
2168{
2169        cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2170        top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2171
2172        hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2173
2174        cpuset_wq = create_singlethread_workqueue("cpuset");
2175        BUG_ON(!cpuset_wq);
2176}
2177
2178/**
2179 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2180 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2181 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2182 *
2183 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2184 * attached to the specified @tsk.  Guaranteed to return some non-empty
2185 * subset of cpu_online_map, even if this means going outside the
2186 * tasks cpuset.
2187 **/
2188
2189void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2190{
2191        mutex_lock(&callback_mutex);
2192        task_lock(tsk);
2193        guarantee_online_cpus(task_cs(tsk), pmask);
2194        task_unlock(tsk);
2195        mutex_unlock(&callback_mutex);
2196}
2197
2198int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2199{
2200        const struct cpuset *cs;
2201        int cpu;
2202
2203        rcu_read_lock();
2204        cs = task_cs(tsk);
2205        if (cs)
2206                do_set_cpus_allowed(tsk, cs->cpus_allowed);
2207        rcu_read_unlock();
2208
2209        /*
2210         * We own tsk->cpus_allowed, nobody can change it under us.
2211         *
2212         * But we used cs && cs->cpus_allowed lockless and thus can
2213         * race with cgroup_attach_task() or update_cpumask() and get
2214         * the wrong tsk->cpus_allowed. However, both cases imply the
2215         * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2216         * which takes task_rq_lock().
2217         *
2218         * If we are called after it dropped the lock we must see all
2219         * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2220         * set any mask even if it is not right from task_cs() pov,
2221         * the pending set_cpus_allowed_ptr() will fix things.
2222         */
2223
2224        cpu = cpumask_any_and(&tsk->cpus_allowed, cpu_active_mask);
2225        if (cpu >= nr_cpu_ids) {
2226                /*
2227                 * Either tsk->cpus_allowed is wrong (see above) or it
2228                 * is actually empty. The latter case is only possible
2229                 * if we are racing with remove_tasks_in_empty_cpuset().
2230                 * Like above we can temporary set any mask and rely on
2231                 * set_cpus_allowed_ptr() as synchronization point.
2232                 */
2233                do_set_cpus_allowed(tsk, cpu_possible_mask);
2234                cpu = cpumask_any(cpu_active_mask);
2235        }
2236
2237        return cpu;
2238}
2239
2240void cpuset_init_current_mems_allowed(void)
2241{
2242        nodes_setall(current->mems_allowed);
2243}
2244
2245/**
2246 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2247 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2248 *
2249 * Description: Returns the nodemask_t mems_allowed of the cpuset
2250 * attached to the specified @tsk.  Guaranteed to return some non-empty
2251 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2252 * tasks cpuset.
2253 **/
2254
2255nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2256{
2257        nodemask_t mask;
2258
2259        mutex_lock(&callback_mutex);
2260        task_lock(tsk);
2261        guarantee_online_mems(task_cs(tsk), &mask);
2262        task_unlock(tsk);
2263        mutex_unlock(&callback_mutex);
2264
2265        return mask;
2266}
2267
2268/**
2269 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2270 * @nodemask: the nodemask to be checked
2271 *
2272 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2273 */
2274int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2275{
2276        return nodes_intersects(*nodemask, current->mems_allowed);
2277}
2278
2279/*
2280 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2281 * mem_hardwall ancestor to the specified cpuset.  Call holding
2282 * callback_mutex.  If no ancestor is mem_exclusive or mem_hardwall
2283 * (an unusual configuration), then returns the root cpuset.
2284 */
2285static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2286{
2287        while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2288                cs = cs->parent;
2289        return cs;
2290}
2291
2292/**
2293 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2294 * @node: is this an allowed node?
2295 * @gfp_mask: memory allocation flags
2296 *
2297 * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2298 * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2299 * yes.  If it's not a __GFP_HARDWALL request and this node is in the nearest
2300 * hardwalled cpuset ancestor to this task's cpuset, yes.  If the task has been
2301 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2302 * flag, yes.
2303 * Otherwise, no.
2304 *
2305 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2306 * cpuset_node_allowed_hardwall().  Otherwise, cpuset_node_allowed_softwall()
2307 * might sleep, and might allow a node from an enclosing cpuset.
2308 *
2309 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2310 * cpusets, and never sleeps.
2311 *
2312 * The __GFP_THISNODE placement logic is really handled elsewhere,
2313 * by forcibly using a zonelist starting at a specified node, and by
2314 * (in get_page_from_freelist()) refusing to consider the zones for
2315 * any node on the zonelist except the first.  By the time any such
2316 * calls get to this routine, we should just shut up and say 'yes'.
2317 *
2318 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2319 * and do not allow allocations outside the current tasks cpuset
2320 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2321 * GFP_KERNEL allocations are not so marked, so can escape to the
2322 * nearest enclosing hardwalled ancestor cpuset.
2323 *
2324 * Scanning up parent cpusets requires callback_mutex.  The
2325 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2326 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2327 * current tasks mems_allowed came up empty on the first pass over
2328 * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2329 * cpuset are short of memory, might require taking the callback_mutex
2330 * mutex.
2331 *
2332 * The first call here from mm/page_alloc:get_page_from_freelist()
2333 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2334 * so no allocation on a node outside the cpuset is allowed (unless
2335 * in interrupt, of course).
2336 *
2337 * The second pass through get_page_from_freelist() doesn't even call
2338 * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2339 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2340 * in alloc_flags.  That logic and the checks below have the combined
2341 * affect that:
2342 *      in_interrupt - any node ok (current task context irrelevant)
2343 *      GFP_ATOMIC   - any node ok
2344 *      TIF_MEMDIE   - any node ok
2345 *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2346 *      GFP_USER     - only nodes in current tasks mems allowed ok.
2347 *
2348 * Rule:
2349 *    Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2350 *    pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2351 *    the code that might scan up ancestor cpusets and sleep.
2352 */
2353int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2354{
2355        const struct cpuset *cs;        /* current cpuset ancestors */
2356        int allowed;                    /* is allocation in zone z allowed? */
2357
2358        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2359                return 1;
2360        might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2361        if (node_isset(node, current->mems_allowed))
2362                return 1;
2363        /*
2364         * Allow tasks that have access to memory reserves because they have
2365         * been OOM killed to get memory anywhere.
2366         */
2367        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2368                return 1;
2369        if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2370                return 0;
2371
2372        if (current->flags & PF_EXITING) /* Let dying task have memory */
2373                return 1;
2374
2375        /* Not hardwall and node outside mems_allowed: scan up cpusets */
2376        mutex_lock(&callback_mutex);
2377
2378        task_lock(current);
2379        cs = nearest_hardwall_ancestor(task_cs(current));
2380        task_unlock(current);
2381
2382        allowed = node_isset(node, cs->mems_allowed);
2383        mutex_unlock(&callback_mutex);
2384        return allowed;
2385}
2386
2387/*
2388 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2389 * @node: is this an allowed node?
2390 * @gfp_mask: memory allocation flags
2391 *
2392 * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2393 * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2394 * yes.  If the task has been OOM killed and has access to memory reserves as
2395 * specified by the TIF_MEMDIE flag, yes.
2396 * Otherwise, no.
2397 *
2398 * The __GFP_THISNODE placement logic is really handled elsewhere,
2399 * by forcibly using a zonelist starting at a specified node, and by
2400 * (in get_page_from_freelist()) refusing to consider the zones for
2401 * any node on the zonelist except the first.  By the time any such
2402 * calls get to this routine, we should just shut up and say 'yes'.
2403 *
2404 * Unlike the cpuset_node_allowed_softwall() variant, above,
2405 * this variant requires that the node be in the current task's
2406 * mems_allowed or that we're in interrupt.  It does not scan up the
2407 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2408 * It never sleeps.
2409 */
2410int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2411{
2412        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2413                return 1;
2414        if (node_isset(node, current->mems_allowed))
2415                return 1;
2416        /*
2417         * Allow tasks that have access to memory reserves because they have
2418         * been OOM killed to get memory anywhere.
2419         */
2420        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2421                return 1;
2422        return 0;
2423}
2424
2425/**
2426 * cpuset_unlock - release lock on cpuset changes
2427 *
2428 * Undo the lock taken in a previous cpuset_lock() call.
2429 */
2430
2431void cpuset_unlock(void)
2432{
2433        mutex_unlock(&callback_mutex);
2434}
2435
2436/**
2437 * cpuset_mem_spread_node() - On which node to begin search for a file page
2438 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2439 *
2440 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2441 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2442 * and if the memory allocation used cpuset_mem_spread_node()
2443 * to determine on which node to start looking, as it will for
2444 * certain page cache or slab cache pages such as used for file
2445 * system buffers and inode caches, then instead of starting on the
2446 * local node to look for a free page, rather spread the starting
2447 * node around the tasks mems_allowed nodes.
2448 *
2449 * We don't have to worry about the returned node being offline
2450 * because "it can't happen", and even if it did, it would be ok.
2451 *
2452 * The routines calling guarantee_online_mems() are careful to
2453 * only set nodes in task->mems_allowed that are online.  So it
2454 * should not be possible for the following code to return an
2455 * offline node.  But if it did, that would be ok, as this routine
2456 * is not returning the node where the allocation must be, only
2457 * the node where the search should start.  The zonelist passed to
2458 * __alloc_pages() will include all nodes.  If the slab allocator
2459 * is passed an offline node, it will fall back to the local node.
2460 * See kmem_cache_alloc_node().
2461 */
2462
2463static int cpuset_spread_node(int *rotor)
2464{
2465        int node;
2466
2467        node = next_node(*rotor, current->mems_allowed);
2468        if (node == MAX_NUMNODES)
2469                node = first_node(current->mems_allowed);
2470        *rotor = node;
2471        return node;
2472}
2473
2474int cpuset_mem_spread_node(void)
2475{
2476        if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2477                current->cpuset_mem_spread_rotor =
2478                        node_random(&current->mems_allowed);
2479
2480        return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2481}
2482
2483int cpuset_slab_spread_node(void)
2484{
2485        if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2486                current->cpuset_slab_spread_rotor =
2487                        node_random(&current->mems_allowed);
2488
2489        return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2490}
2491
2492EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2493
2494/**
2495 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2496 * @tsk1: pointer to task_struct of some task.
2497 * @tsk2: pointer to task_struct of some other task.
2498 *
2499 * Description: Return true if @tsk1's mems_allowed intersects the
2500 * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2501 * one of the task's memory usage might impact the memory available
2502 * to the other.
2503 **/
2504
2505int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2506                                   const struct task_struct *tsk2)
2507{
2508        return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2509}
2510
2511/**
2512 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2513 * @task: pointer to task_struct of some task.
2514 *
2515 * Description: Prints @task's name, cpuset name, and cached copy of its
2516 * mems_allowed to the kernel log.  Must hold task_lock(task) to allow
2517 * dereferencing task_cs(task).
2518 */
2519void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2520{
2521        struct dentry *dentry;
2522
2523        dentry = task_cs(tsk)->css.cgroup->dentry;
2524        spin_lock(&cpuset_buffer_lock);
2525        snprintf(cpuset_name, CPUSET_NAME_LEN,
2526                 dentry ? (const char *)dentry->d_name.name : "/");
2527        nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2528                           tsk->mems_allowed);
2529        printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2530               tsk->comm, cpuset_name, cpuset_nodelist);
2531        spin_unlock(&cpuset_buffer_lock);
2532}
2533
2534/*
2535 * Collection of memory_pressure is suppressed unless
2536 * this flag is enabled by writing "1" to the special
2537 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2538 */
2539
2540int cpuset_memory_pressure_enabled __read_mostly;
2541
2542/**
2543 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2544 *
2545 * Keep a running average of the rate of synchronous (direct)
2546 * page reclaim efforts initiated by tasks in each cpuset.
2547 *
2548 * This represents the rate at which some task in the cpuset
2549 * ran low on memory on all nodes it was allowed to use, and
2550 * had to enter the kernels page reclaim code in an effort to
2551 * create more free memory by tossing clean pages or swapping
2552 * or writing dirty pages.
2553 *
2554 * Display to user space in the per-cpuset read-only file
2555 * "memory_pressure".  Value displayed is an integer
2556 * representing the recent rate of entry into the synchronous
2557 * (direct) page reclaim by any task attached to the cpuset.
2558 **/
2559
2560void __cpuset_memory_pressure_bump(void)
2561{
2562        task_lock(current);
2563        fmeter_markevent(&task_cs(current)->fmeter);
2564        task_unlock(current);
2565}
2566
2567#ifdef CONFIG_PROC_PID_CPUSET
2568/*
2569 * proc_cpuset_show()
2570 *  - Print tasks cpuset path into seq_file.
2571 *  - Used for /proc/<pid>/cpuset.
2572 *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2573 *    doesn't really matter if tsk->cpuset changes after we read it,
2574 *    and we take cgroup_mutex, keeping cpuset_attach() from changing it
2575 *    anyway.
2576 */
2577static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2578{
2579        struct pid *pid;
2580        struct task_struct *tsk;
2581        char *buf;
2582        struct cgroup_subsys_state *css;
2583        int retval;
2584
2585        retval = -ENOMEM;
2586        buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2587        if (!buf)
2588                goto out;
2589
2590        retval = -ESRCH;
2591        pid = m->private;
2592        tsk = get_pid_task(pid, PIDTYPE_PID);
2593        if (!tsk)
2594                goto out_free;
2595
2596        retval = -EINVAL;
2597        cgroup_lock();
2598        css = task_subsys_state(tsk, cpuset_subsys_id);
2599        retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2600        if (retval < 0)
2601                goto out_unlock;
2602        seq_puts(m, buf);
2603        seq_putc(m, '\n');
2604out_unlock:
2605        cgroup_unlock();
2606        put_task_struct(tsk);
2607out_free:
2608        kfree(buf);
2609out:
2610        return retval;
2611}
2612
2613static int cpuset_open(struct inode *inode, struct file *file)
2614{
2615        struct pid *pid = PROC_I(inode)->pid;
2616        return single_open(file, proc_cpuset_show, pid);
2617}
2618
2619const struct file_operations proc_cpuset_operations = {
2620        .open           = cpuset_open,
2621        .read           = seq_read,
2622        .llseek         = seq_lseek,
2623        .release        = single_release,
2624};
2625#endif /* CONFIG_PROC_PID_CPUSET */
2626
2627/* Display task mems_allowed in /proc/<pid>/status file. */
2628void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2629{
2630        seq_printf(m, "Mems_allowed:\t");
2631        seq_nodemask(m, &task->mems_allowed);
2632        seq_printf(m, "\n");
2633        seq_printf(m, "Mems_allowed_list:\t");
2634        seq_nodemask_list(m, &task->mems_allowed);
2635        seq_printf(m, "\n");
2636}
2637