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#include <linux/wait.h>
  63
  64/*
  65 * Tracks how many cpusets are currently defined in system.
  66 * When there is only one cpuset (the root cpuset) we can
  67 * short circuit some hooks.
  68 */
  69int number_of_cpusets __read_mostly;
  70
  71/* See "Frequency meter" comments, below. */
  72
  73struct fmeter {
  74        int cnt;                /* unprocessed events count */
  75        int val;                /* most recent output value */
  76        time_t time;            /* clock (secs) when val computed */
  77        spinlock_t lock;        /* guards read or write of above */
  78};
  79
  80struct cpuset {
  81        struct cgroup_subsys_state css;
  82
  83        unsigned long flags;            /* "unsigned long" so bitops work */
  84        cpumask_var_t cpus_allowed;     /* CPUs allowed to tasks in cpuset */
  85        nodemask_t mems_allowed;        /* Memory Nodes allowed to tasks */
  86
  87        /*
  88         * This is old Memory Nodes tasks took on.
  89         *
  90         * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
  91         * - A new cpuset's old_mems_allowed is initialized when some
  92         *   task is moved into it.
  93         * - old_mems_allowed is used in cpuset_migrate_mm() when we change
  94         *   cpuset.mems_allowed and have tasks' nodemask updated, and
  95         *   then old_mems_allowed is updated to mems_allowed.
  96         */
  97        nodemask_t old_mems_allowed;
  98
  99        struct fmeter fmeter;           /* memory_pressure filter */
 100
 101        /*
 102         * Tasks are being attached to this cpuset.  Used to prevent
 103         * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
 104         */
 105        int attach_in_progress;
 106
 107        /* partition number for rebuild_sched_domains() */
 108        int pn;
 109
 110        /* for custom sched domain */
 111        int relax_domain_level;
 112};
 113
 114static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
 115{
 116        return css ? container_of(css, struct cpuset, css) : NULL;
 117}
 118
 119/* Retrieve the cpuset for a task */
 120static inline struct cpuset *task_cs(struct task_struct *task)
 121{
 122        return css_cs(task_css(task, cpuset_subsys_id));
 123}
 124
 125static inline struct cpuset *parent_cs(struct cpuset *cs)
 126{
 127        return css_cs(css_parent(&cs->css));
 128}
 129
 130#ifdef CONFIG_NUMA
 131static inline bool task_has_mempolicy(struct task_struct *task)
 132{
 133        return task->mempolicy;
 134}
 135#else
 136static inline bool task_has_mempolicy(struct task_struct *task)
 137{
 138        return false;
 139}
 140#endif
 141
 142
 143/* bits in struct cpuset flags field */
 144typedef enum {
 145        CS_ONLINE,
 146        CS_CPU_EXCLUSIVE,
 147        CS_MEM_EXCLUSIVE,
 148        CS_MEM_HARDWALL,
 149        CS_MEMORY_MIGRATE,
 150        CS_SCHED_LOAD_BALANCE,
 151        CS_SPREAD_PAGE,
 152        CS_SPREAD_SLAB,
 153} cpuset_flagbits_t;
 154
 155/* convenient tests for these bits */
 156static inline bool is_cpuset_online(const struct cpuset *cs)
 157{
 158        return test_bit(CS_ONLINE, &cs->flags);
 159}
 160
 161static inline int is_cpu_exclusive(const struct cpuset *cs)
 162{
 163        return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
 164}
 165
 166static inline int is_mem_exclusive(const struct cpuset *cs)
 167{
 168        return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
 169}
 170
 171static inline int is_mem_hardwall(const struct cpuset *cs)
 172{
 173        return test_bit(CS_MEM_HARDWALL, &cs->flags);
 174}
 175
 176static inline int is_sched_load_balance(const struct cpuset *cs)
 177{
 178        return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
 179}
 180
 181static inline int is_memory_migrate(const struct cpuset *cs)
 182{
 183        return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
 184}
 185
 186static inline int is_spread_page(const struct cpuset *cs)
 187{
 188        return test_bit(CS_SPREAD_PAGE, &cs->flags);
 189}
 190
 191static inline int is_spread_slab(const struct cpuset *cs)
 192{
 193        return test_bit(CS_SPREAD_SLAB, &cs->flags);
 194}
 195
 196static struct cpuset top_cpuset = {
 197        .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
 198                  (1 << CS_MEM_EXCLUSIVE)),
 199};
 200
 201/**
 202 * cpuset_for_each_child - traverse online children of a cpuset
 203 * @child_cs: loop cursor pointing to the current child
 204 * @pos_css: used for iteration
 205 * @parent_cs: target cpuset to walk children of
 206 *
 207 * Walk @child_cs through the online children of @parent_cs.  Must be used
 208 * with RCU read locked.
 209 */
 210#define cpuset_for_each_child(child_cs, pos_css, parent_cs)             \
 211        css_for_each_child((pos_css), &(parent_cs)->css)                \
 212                if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
 213
 214/**
 215 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
 216 * @des_cs: loop cursor pointing to the current descendant
 217 * @pos_css: used for iteration
 218 * @root_cs: target cpuset to walk ancestor of
 219 *
 220 * Walk @des_cs through the online descendants of @root_cs.  Must be used
 221 * with RCU read locked.  The caller may modify @pos_css by calling
 222 * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
 223 * iteration and the first node to be visited.
 224 */
 225#define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)        \
 226        css_for_each_descendant_pre((pos_css), &(root_cs)->css)         \
 227                if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
 228
 229/*
 230 * There are two global mutexes guarding cpuset structures - cpuset_mutex
 231 * and callback_mutex.  The latter may nest inside the former.  We also
 232 * require taking task_lock() when dereferencing a task's cpuset pointer.
 233 * See "The task_lock() exception", at the end of this comment.
 234 *
 235 * A task must hold both mutexes to modify cpusets.  If a task holds
 236 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
 237 * is the only task able to also acquire callback_mutex and be able to
 238 * modify cpusets.  It can perform various checks on the cpuset structure
 239 * first, knowing nothing will change.  It can also allocate memory while
 240 * just holding cpuset_mutex.  While it is performing these checks, various
 241 * callback routines can briefly acquire callback_mutex to query cpusets.
 242 * Once it is ready to make the changes, it takes callback_mutex, blocking
 243 * everyone else.
 244 *
 245 * Calls to the kernel memory allocator can not be made while holding
 246 * callback_mutex, as that would risk double tripping on callback_mutex
 247 * from one of the callbacks into the cpuset code from within
 248 * __alloc_pages().
 249 *
 250 * If a task is only holding callback_mutex, then it has read-only
 251 * access to cpusets.
 252 *
 253 * Now, the task_struct fields mems_allowed and mempolicy may be changed
 254 * by other task, we use alloc_lock in the task_struct fields to protect
 255 * them.
 256 *
 257 * The cpuset_common_file_read() handlers only hold callback_mutex across
 258 * small pieces of code, such as when reading out possibly multi-word
 259 * cpumasks and nodemasks.
 260 *
 261 * Accessing a task's cpuset should be done in accordance with the
 262 * guidelines for accessing subsystem state in kernel/cgroup.c
 263 */
 264
 265static DEFINE_MUTEX(cpuset_mutex);
 266static DEFINE_MUTEX(callback_mutex);
 267
 268/*
 269 * CPU / memory hotplug is handled asynchronously.
 270 */
 271static void cpuset_hotplug_workfn(struct work_struct *work);
 272static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
 273
 274static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
 275
 276/*
 277 * This is ugly, but preserves the userspace API for existing cpuset
 278 * users. If someone tries to mount the "cpuset" filesystem, we
 279 * silently switch it to mount "cgroup" instead
 280 */
 281static struct dentry *cpuset_mount(struct file_system_type *fs_type,
 282                         int flags, const char *unused_dev_name, void *data)
 283{
 284        struct file_system_type *cgroup_fs = get_fs_type("cgroup");
 285        struct dentry *ret = ERR_PTR(-ENODEV);
 286        if (cgroup_fs) {
 287                char mountopts[] =
 288                        "cpuset,noprefix,"
 289                        "release_agent=/sbin/cpuset_release_agent";
 290                ret = cgroup_fs->mount(cgroup_fs, flags,
 291                                           unused_dev_name, mountopts);
 292                put_filesystem(cgroup_fs);
 293        }
 294        return ret;
 295}
 296
 297static struct file_system_type cpuset_fs_type = {
 298        .name = "cpuset",
 299        .mount = cpuset_mount,
 300};
 301
 302/*
 303 * Return in pmask the portion of a cpusets's cpus_allowed that
 304 * are online.  If none are online, walk up the cpuset hierarchy
 305 * until we find one that does have some online cpus.  The top
 306 * cpuset always has some cpus online.
 307 *
 308 * One way or another, we guarantee to return some non-empty subset
 309 * of cpu_online_mask.
 310 *
 311 * Call with callback_mutex held.
 312 */
 313static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
 314{
 315        while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
 316                cs = parent_cs(cs);
 317        cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
 318}
 319
 320/*
 321 * Return in *pmask the portion of a cpusets's mems_allowed that
 322 * are online, with memory.  If none are online with memory, walk
 323 * up the cpuset hierarchy until we find one that does have some
 324 * online mems.  The top cpuset always has some mems online.
 325 *
 326 * One way or another, we guarantee to return some non-empty subset
 327 * of node_states[N_MEMORY].
 328 *
 329 * Call with callback_mutex held.
 330 */
 331static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
 332{
 333        while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
 334                cs = parent_cs(cs);
 335        nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
 336}
 337
 338/*
 339 * update task's spread flag if cpuset's page/slab spread flag is set
 340 *
 341 * Called with callback_mutex/cpuset_mutex held
 342 */
 343static void cpuset_update_task_spread_flag(struct cpuset *cs,
 344                                        struct task_struct *tsk)
 345{
 346        if (is_spread_page(cs))
 347                tsk->flags |= PF_SPREAD_PAGE;
 348        else
 349                tsk->flags &= ~PF_SPREAD_PAGE;
 350        if (is_spread_slab(cs))
 351                tsk->flags |= PF_SPREAD_SLAB;
 352        else
 353                tsk->flags &= ~PF_SPREAD_SLAB;
 354}
 355
 356/*
 357 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
 358 *
 359 * One cpuset is a subset of another if all its allowed CPUs and
 360 * Memory Nodes are a subset of the other, and its exclusive flags
 361 * are only set if the other's are set.  Call holding cpuset_mutex.
 362 */
 363
 364static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
 365{
 366        return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
 367                nodes_subset(p->mems_allowed, q->mems_allowed) &&
 368                is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
 369                is_mem_exclusive(p) <= is_mem_exclusive(q);
 370}
 371
 372/**
 373 * alloc_trial_cpuset - allocate a trial cpuset
 374 * @cs: the cpuset that the trial cpuset duplicates
 375 */
 376static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
 377{
 378        struct cpuset *trial;
 379
 380        trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
 381        if (!trial)
 382                return NULL;
 383
 384        if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
 385                kfree(trial);
 386                return NULL;
 387        }
 388        cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
 389
 390        return trial;
 391}
 392
 393/**
 394 * free_trial_cpuset - free the trial cpuset
 395 * @trial: the trial cpuset to be freed
 396 */
 397static void free_trial_cpuset(struct cpuset *trial)
 398{
 399        free_cpumask_var(trial->cpus_allowed);
 400        kfree(trial);
 401}
 402
 403/*
 404 * validate_change() - Used to validate that any proposed cpuset change
 405 *                     follows the structural rules for cpusets.
 406 *
 407 * If we replaced the flag and mask values of the current cpuset
 408 * (cur) with those values in the trial cpuset (trial), would
 409 * our various subset and exclusive rules still be valid?  Presumes
 410 * cpuset_mutex held.
 411 *
 412 * 'cur' is the address of an actual, in-use cpuset.  Operations
 413 * such as list traversal that depend on the actual address of the
 414 * cpuset in the list must use cur below, not trial.
 415 *
 416 * 'trial' is the address of bulk structure copy of cur, with
 417 * perhaps one or more of the fields cpus_allowed, mems_allowed,
 418 * or flags changed to new, trial values.
 419 *
 420 * Return 0 if valid, -errno if not.
 421 */
 422
 423static int validate_change(struct cpuset *cur, struct cpuset *trial)
 424{
 425        struct cgroup_subsys_state *css;
 426        struct cpuset *c, *par;
 427        int ret;
 428
 429        rcu_read_lock();
 430
 431        /* Each of our child cpusets must be a subset of us */
 432        ret = -EBUSY;
 433        cpuset_for_each_child(c, css, cur)
 434                if (!is_cpuset_subset(c, trial))
 435                        goto out;
 436
 437        /* Remaining checks don't apply to root cpuset */
 438        ret = 0;
 439        if (cur == &top_cpuset)
 440                goto out;
 441
 442        par = parent_cs(cur);
 443
 444        /* We must be a subset of our parent cpuset */
 445        ret = -EACCES;
 446        if (!is_cpuset_subset(trial, par))
 447                goto out;
 448
 449        /*
 450         * If either I or some sibling (!= me) is exclusive, we can't
 451         * overlap
 452         */
 453        ret = -EINVAL;
 454        cpuset_for_each_child(c, css, par) {
 455                if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
 456                    c != cur &&
 457                    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
 458                        goto out;
 459                if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
 460                    c != cur &&
 461                    nodes_intersects(trial->mems_allowed, c->mems_allowed))
 462                        goto out;
 463        }
 464
 465        /*
 466         * Cpusets with tasks - existing or newly being attached - can't
 467         * be changed to have empty cpus_allowed or mems_allowed.
 468         */
 469        ret = -ENOSPC;
 470        if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress)) {
 471                if (!cpumask_empty(cur->cpus_allowed) &&
 472                    cpumask_empty(trial->cpus_allowed))
 473                        goto out;
 474                if (!nodes_empty(cur->mems_allowed) &&
 475                    nodes_empty(trial->mems_allowed))
 476                        goto out;
 477        }
 478
 479        ret = 0;
 480out:
 481        rcu_read_unlock();
 482        return ret;
 483}
 484
 485#ifdef CONFIG_SMP
 486/*
 487 * Helper routine for generate_sched_domains().
 488 * Do cpusets a, b have overlapping cpus_allowed masks?
 489 */
 490static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
 491{
 492        return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
 493}
 494
 495static void
 496update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
 497{
 498        if (dattr->relax_domain_level < c->relax_domain_level)
 499                dattr->relax_domain_level = c->relax_domain_level;
 500        return;
 501}
 502
 503static void update_domain_attr_tree(struct sched_domain_attr *dattr,
 504                                    struct cpuset *root_cs)
 505{
 506        struct cpuset *cp;
 507        struct cgroup_subsys_state *pos_css;
 508
 509        rcu_read_lock();
 510        cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
 511                if (cp == root_cs)
 512                        continue;
 513
 514                /* skip the whole subtree if @cp doesn't have any CPU */
 515                if (cpumask_empty(cp->cpus_allowed)) {
 516                        pos_css = css_rightmost_descendant(pos_css);
 517                        continue;
 518                }
 519
 520                if (is_sched_load_balance(cp))
 521                        update_domain_attr(dattr, cp);
 522        }
 523        rcu_read_unlock();
 524}
 525
 526/*
 527 * generate_sched_domains()
 528 *
 529 * This function builds a partial partition of the systems CPUs
 530 * A 'partial partition' is a set of non-overlapping subsets whose
 531 * union is a subset of that set.
 532 * The output of this function needs to be passed to kernel/sched/core.c
 533 * partition_sched_domains() routine, which will rebuild the scheduler's
 534 * load balancing domains (sched domains) as specified by that partial
 535 * partition.
 536 *
 537 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
 538 * for a background explanation of this.
 539 *
 540 * Does not return errors, on the theory that the callers of this
 541 * routine would rather not worry about failures to rebuild sched
 542 * domains when operating in the severe memory shortage situations
 543 * that could cause allocation failures below.
 544 *
 545 * Must be called with cpuset_mutex held.
 546 *
 547 * The three key local variables below are:
 548 *    q  - a linked-list queue of cpuset pointers, used to implement a
 549 *         top-down scan of all cpusets.  This scan loads a pointer
 550 *         to each cpuset marked is_sched_load_balance into the
 551 *         array 'csa'.  For our purposes, rebuilding the schedulers
 552 *         sched domains, we can ignore !is_sched_load_balance cpusets.
 553 *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
 554 *         that need to be load balanced, for convenient iterative
 555 *         access by the subsequent code that finds the best partition,
 556 *         i.e the set of domains (subsets) of CPUs such that the
 557 *         cpus_allowed of every cpuset marked is_sched_load_balance
 558 *         is a subset of one of these domains, while there are as
 559 *         many such domains as possible, each as small as possible.
 560 * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
 561 *         the kernel/sched/core.c routine partition_sched_domains() in a
 562 *         convenient format, that can be easily compared to the prior
 563 *         value to determine what partition elements (sched domains)
 564 *         were changed (added or removed.)
 565 *
 566 * Finding the best partition (set of domains):
 567 *      The triple nested loops below over i, j, k scan over the
 568 *      load balanced cpusets (using the array of cpuset pointers in
 569 *      csa[]) looking for pairs of cpusets that have overlapping
 570 *      cpus_allowed, but which don't have the same 'pn' partition
 571 *      number and gives them in the same partition number.  It keeps
 572 *      looping on the 'restart' label until it can no longer find
 573 *      any such pairs.
 574 *
 575 *      The union of the cpus_allowed masks from the set of
 576 *      all cpusets having the same 'pn' value then form the one
 577 *      element of the partition (one sched domain) to be passed to
 578 *      partition_sched_domains().
 579 */
 580static int generate_sched_domains(cpumask_var_t **domains,
 581                        struct sched_domain_attr **attributes)
 582{
 583        struct cpuset *cp;      /* scans q */
 584        struct cpuset **csa;    /* array of all cpuset ptrs */
 585        int csn;                /* how many cpuset ptrs in csa so far */
 586        int i, j, k;            /* indices for partition finding loops */
 587        cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
 588        struct sched_domain_attr *dattr;  /* attributes for custom domains */
 589        int ndoms = 0;          /* number of sched domains in result */
 590        int nslot;              /* next empty doms[] struct cpumask slot */
 591        struct cgroup_subsys_state *pos_css;
 592
 593        doms = NULL;
 594        dattr = NULL;
 595        csa = NULL;
 596
 597        /* Special case for the 99% of systems with one, full, sched domain */
 598        if (is_sched_load_balance(&top_cpuset)) {
 599                ndoms = 1;
 600                doms = alloc_sched_domains(ndoms);
 601                if (!doms)
 602                        goto done;
 603
 604                dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
 605                if (dattr) {
 606                        *dattr = SD_ATTR_INIT;
 607                        update_domain_attr_tree(dattr, &top_cpuset);
 608                }
 609                cpumask_copy(doms[0], top_cpuset.cpus_allowed);
 610
 611                goto done;
 612        }
 613
 614        csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
 615        if (!csa)
 616                goto done;
 617        csn = 0;
 618
 619        rcu_read_lock();
 620        cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
 621                if (cp == &top_cpuset)
 622                        continue;
 623                /*
 624                 * Continue traversing beyond @cp iff @cp has some CPUs and
 625                 * isn't load balancing.  The former is obvious.  The
 626                 * latter: All child cpusets contain a subset of the
 627                 * parent's cpus, so just skip them, and then we call
 628                 * update_domain_attr_tree() to calc relax_domain_level of
 629                 * the corresponding sched domain.
 630                 */
 631                if (!cpumask_empty(cp->cpus_allowed) &&
 632                    !is_sched_load_balance(cp))
 633                        continue;
 634
 635                if (is_sched_load_balance(cp))
 636                        csa[csn++] = cp;
 637
 638                /* skip @cp's subtree */
 639                pos_css = css_rightmost_descendant(pos_css);
 640        }
 641        rcu_read_unlock();
 642
 643        for (i = 0; i < csn; i++)
 644                csa[i]->pn = i;
 645        ndoms = csn;
 646
 647restart:
 648        /* Find the best partition (set of sched domains) */
 649        for (i = 0; i < csn; i++) {
 650                struct cpuset *a = csa[i];
 651                int apn = a->pn;
 652
 653                for (j = 0; j < csn; j++) {
 654                        struct cpuset *b = csa[j];
 655                        int bpn = b->pn;
 656
 657                        if (apn != bpn && cpusets_overlap(a, b)) {
 658                                for (k = 0; k < csn; k++) {
 659                                        struct cpuset *c = csa[k];
 660
 661                                        if (c->pn == bpn)
 662                                                c->pn = apn;
 663                                }
 664                                ndoms--;        /* one less element */
 665                                goto restart;
 666                        }
 667                }
 668        }
 669
 670        /*
 671         * Now we know how many domains to create.
 672         * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
 673         */
 674        doms = alloc_sched_domains(ndoms);
 675        if (!doms)
 676                goto done;
 677
 678        /*
 679         * The rest of the code, including the scheduler, can deal with
 680         * dattr==NULL case. No need to abort if alloc fails.
 681         */
 682        dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
 683
 684        for (nslot = 0, i = 0; i < csn; i++) {
 685                struct cpuset *a = csa[i];
 686                struct cpumask *dp;
 687                int apn = a->pn;
 688
 689                if (apn < 0) {
 690                        /* Skip completed partitions */
 691                        continue;
 692                }
 693
 694                dp = doms[nslot];
 695
 696                if (nslot == ndoms) {
 697                        static int warnings = 10;
 698                        if (warnings) {
 699                                printk(KERN_WARNING
 700                                 "rebuild_sched_domains confused:"
 701                                  " nslot %d, ndoms %d, csn %d, i %d,"
 702                                  " apn %d\n",
 703                                  nslot, ndoms, csn, i, apn);
 704                                warnings--;
 705                        }
 706                        continue;
 707                }
 708
 709                cpumask_clear(dp);
 710                if (dattr)
 711                        *(dattr + nslot) = SD_ATTR_INIT;
 712                for (j = i; j < csn; j++) {
 713                        struct cpuset *b = csa[j];
 714
 715                        if (apn == b->pn) {
 716                                cpumask_or(dp, dp, b->cpus_allowed);
 717                                if (dattr)
 718                                        update_domain_attr_tree(dattr + nslot, b);
 719
 720                                /* Done with this partition */
 721                                b->pn = -1;
 722                        }
 723                }
 724                nslot++;
 725        }
 726        BUG_ON(nslot != ndoms);
 727
 728done:
 729        kfree(csa);
 730
 731        /*
 732         * Fallback to the default domain if kmalloc() failed.
 733         * See comments in partition_sched_domains().
 734         */
 735        if (doms == NULL)
 736                ndoms = 1;
 737
 738        *domains    = doms;
 739        *attributes = dattr;
 740        return ndoms;
 741}
 742
 743/*
 744 * Rebuild scheduler domains.
 745 *
 746 * If the flag 'sched_load_balance' of any cpuset with non-empty
 747 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
 748 * which has that flag enabled, or if any cpuset with a non-empty
 749 * 'cpus' is removed, then call this routine to rebuild the
 750 * scheduler's dynamic sched domains.
 751 *
 752 * Call with cpuset_mutex held.  Takes get_online_cpus().
 753 */
 754static void rebuild_sched_domains_locked(void)
 755{
 756        struct sched_domain_attr *attr;
 757        cpumask_var_t *doms;
 758        int ndoms;
 759
 760        lockdep_assert_held(&cpuset_mutex);
 761        get_online_cpus();
 762
 763        /*
 764         * We have raced with CPU hotplug. Don't do anything to avoid
 765         * passing doms with offlined cpu to partition_sched_domains().
 766         * Anyways, hotplug work item will rebuild sched domains.
 767         */
 768        if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
 769                goto out;
 770
 771        /* Generate domain masks and attrs */
 772        ndoms = generate_sched_domains(&doms, &attr);
 773
 774        /* Have scheduler rebuild the domains */
 775        partition_sched_domains(ndoms, doms, attr);
 776out:
 777        put_online_cpus();
 778}
 779#else /* !CONFIG_SMP */
 780static void rebuild_sched_domains_locked(void)
 781{
 782}
 783#endif /* CONFIG_SMP */
 784
 785void rebuild_sched_domains(void)
 786{
 787        mutex_lock(&cpuset_mutex);
 788        rebuild_sched_domains_locked();
 789        mutex_unlock(&cpuset_mutex);
 790}
 791
 792/*
 793 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
 794 * @cs: the cpuset in interest
 795 *
 796 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
 797 * with non-empty cpus. We use effective cpumask whenever:
 798 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
 799 *   if the cpuset they reside in has no cpus)
 800 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
 801 *
 802 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
 803 * exception. See comments there.
 804 */
 805static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
 806{
 807        while (cpumask_empty(cs->cpus_allowed))
 808                cs = parent_cs(cs);
 809        return cs;
 810}
 811
 812/*
 813 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
 814 * @cs: the cpuset in interest
 815 *
 816 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
 817 * with non-empty memss. We use effective nodemask whenever:
 818 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
 819 *   if the cpuset they reside in has no mems)
 820 * - we want to retrieve task_cs(tsk)'s mems_allowed.
 821 *
 822 * Called with cpuset_mutex held.
 823 */
 824static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
 825{
 826        while (nodes_empty(cs->mems_allowed))
 827                cs = parent_cs(cs);
 828        return cs;
 829}
 830
 831/**
 832 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
 833 * @tsk: task to test
 834 * @data: cpuset to @tsk belongs to
 835 *
 836 * Called by css_scan_tasks() for each task in a cgroup whose cpus_allowed
 837 * mask needs to be changed.
 838 *
 839 * We don't need to re-check for the cgroup/cpuset membership, since we're
 840 * holding cpuset_mutex at this point.
 841 */
 842static void cpuset_change_cpumask(struct task_struct *tsk, void *data)
 843{
 844        struct cpuset *cs = data;
 845        struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
 846
 847        set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
 848}
 849
 850/**
 851 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
 852 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
 853 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
 854 *
 855 * Called with cpuset_mutex held
 856 *
 857 * The css_scan_tasks() function will scan all the tasks in a cgroup,
 858 * calling callback functions for each.
 859 *
 860 * No return value. It's guaranteed that css_scan_tasks() always returns 0
 861 * if @heap != NULL.
 862 */
 863static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 864{
 865        css_scan_tasks(&cs->css, NULL, cpuset_change_cpumask, cs, heap);
 866}
 867
 868/*
 869 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
 870 * @root_cs: the root cpuset of the hierarchy
 871 * @update_root: update root cpuset or not?
 872 * @heap: the heap used by css_scan_tasks()
 873 *
 874 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
 875 * which take on cpumask of @root_cs.
 876 *
 877 * Called with cpuset_mutex held
 878 */
 879static void update_tasks_cpumask_hier(struct cpuset *root_cs,
 880                                      bool update_root, struct ptr_heap *heap)
 881{
 882        struct cpuset *cp;
 883        struct cgroup_subsys_state *pos_css;
 884
 885        rcu_read_lock();
 886        cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
 887                if (cp == root_cs) {
 888                        if (!update_root)
 889                                continue;
 890                } else {
 891                        /* skip the whole subtree if @cp have some CPU */
 892                        if (!cpumask_empty(cp->cpus_allowed)) {
 893                                pos_css = css_rightmost_descendant(pos_css);
 894                                continue;
 895                        }
 896                }
 897                if (!css_tryget(&cp->css))
 898                        continue;
 899                rcu_read_unlock();
 900
 901                update_tasks_cpumask(cp, heap);
 902
 903                rcu_read_lock();
 904                css_put(&cp->css);
 905        }
 906        rcu_read_unlock();
 907}
 908
 909/**
 910 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
 911 * @cs: the cpuset to consider
 912 * @buf: buffer of cpu numbers written to this cpuset
 913 */
 914static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 915                          const char *buf)
 916{
 917        struct ptr_heap heap;
 918        int retval;
 919        int is_load_balanced;
 920
 921        /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
 922        if (cs == &top_cpuset)
 923                return -EACCES;
 924
 925        /*
 926         * An empty cpus_allowed is ok only if the cpuset has no tasks.
 927         * Since cpulist_parse() fails on an empty mask, we special case
 928         * that parsing.  The validate_change() call ensures that cpusets
 929         * with tasks have cpus.
 930         */
 931        if (!*buf) {
 932                cpumask_clear(trialcs->cpus_allowed);
 933        } else {
 934                retval = cpulist_parse(buf, trialcs->cpus_allowed);
 935                if (retval < 0)
 936                        return retval;
 937
 938                if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
 939                        return -EINVAL;
 940        }
 941
 942        /* Nothing to do if the cpus didn't change */
 943        if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
 944                return 0;
 945
 946        retval = validate_change(cs, trialcs);
 947        if (retval < 0)
 948                return retval;
 949
 950        retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
 951        if (retval)
 952                return retval;
 953
 954        is_load_balanced = is_sched_load_balance(trialcs);
 955
 956        mutex_lock(&callback_mutex);
 957        cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
 958        mutex_unlock(&callback_mutex);
 959
 960        update_tasks_cpumask_hier(cs, true, &heap);
 961
 962        heap_free(&heap);
 963
 964        if (is_load_balanced)
 965                rebuild_sched_domains_locked();
 966        return 0;
 967}
 968
 969/*
 970 * cpuset_migrate_mm
 971 *
 972 *    Migrate memory region from one set of nodes to another.
 973 *
 974 *    Temporarilly set tasks mems_allowed to target nodes of migration,
 975 *    so that the migration code can allocate pages on these nodes.
 976 *
 977 *    Call holding cpuset_mutex, so current's cpuset won't change
 978 *    during this call, as manage_mutex holds off any cpuset_attach()
 979 *    calls.  Therefore we don't need to take task_lock around the
 980 *    call to guarantee_online_mems(), as we know no one is changing
 981 *    our task's cpuset.
 982 *
 983 *    While the mm_struct we are migrating is typically from some
 984 *    other task, the task_struct mems_allowed that we are hacking
 985 *    is for our current task, which must allocate new pages for that
 986 *    migrating memory region.
 987 */
 988
 989static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
 990                                                        const nodemask_t *to)
 991{
 992        struct task_struct *tsk = current;
 993        struct cpuset *mems_cs;
 994
 995        tsk->mems_allowed = *to;
 996
 997        do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
 998
 999        mems_cs = effective_nodemask_cpuset(task_cs(tsk));
1000        guarantee_online_mems(mems_cs, &tsk->mems_allowed);
1001}
1002
1003/*
1004 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1005 * @tsk: the task to change
1006 * @newmems: new nodes that the task will be set
1007 *
1008 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1009 * we structure updates as setting all new allowed nodes, then clearing newly
1010 * disallowed ones.
1011 */
1012static void cpuset_change_task_nodemask(struct task_struct *tsk,
1013                                        nodemask_t *newmems)
1014{
1015        bool need_loop;
1016
1017        /*
1018         * Allow tasks that have access to memory reserves because they have
1019         * been OOM killed to get memory anywhere.
1020         */
1021        if (unlikely(test_thread_flag(TIF_MEMDIE)))
1022                return;
1023        if (current->flags & PF_EXITING) /* Let dying task have memory */
1024                return;
1025
1026        task_lock(tsk);
1027        /*
1028         * Determine if a loop is necessary if another thread is doing
1029         * get_mems_allowed().  If at least one node remains unchanged and
1030         * tsk does not have a mempolicy, then an empty nodemask will not be
1031         * possible when mems_allowed is larger than a word.
1032         */
1033        need_loop = task_has_mempolicy(tsk) ||
1034                        !nodes_intersects(*newmems, tsk->mems_allowed);
1035
1036        if (need_loop)
1037                write_seqcount_begin(&tsk->mems_allowed_seq);
1038
1039        nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1040        mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1041
1042        mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1043        tsk->mems_allowed = *newmems;
1044
1045        if (need_loop)
1046                write_seqcount_end(&tsk->mems_allowed_seq);
1047
1048        task_unlock(tsk);
1049}
1050
1051struct cpuset_change_nodemask_arg {
1052        struct cpuset           *cs;
1053        nodemask_t              *newmems;
1054};
1055
1056/*
1057 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1058 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1059 * memory_migrate flag is set. Called with cpuset_mutex held.
1060 */
1061static void cpuset_change_nodemask(struct task_struct *p, void *data)
1062{
1063        struct cpuset_change_nodemask_arg *arg = data;
1064        struct cpuset *cs = arg->cs;
1065        struct mm_struct *mm;
1066        int migrate;
1067
1068        cpuset_change_task_nodemask(p, arg->newmems);
1069
1070        mm = get_task_mm(p);
1071        if (!mm)
1072                return;
1073
1074        migrate = is_memory_migrate(cs);
1075
1076        mpol_rebind_mm(mm, &cs->mems_allowed);
1077        if (migrate)
1078                cpuset_migrate_mm(mm, &cs->old_mems_allowed, arg->newmems);
1079        mmput(mm);
1080}
1081
1082static void *cpuset_being_rebound;
1083
1084/**
1085 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1086 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1087 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
1088 *
1089 * Called with cpuset_mutex held.  No return value. It's guaranteed that
1090 * css_scan_tasks() always returns 0 if @heap != NULL.
1091 */
1092static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
1093{
1094        static nodemask_t newmems;      /* protected by cpuset_mutex */
1095        struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1096        struct cpuset_change_nodemask_arg arg = { .cs = cs,
1097                                                  .newmems = &newmems };
1098
1099        cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1100
1101        guarantee_online_mems(mems_cs, &newmems);
1102
1103        /*
1104         * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1105         * take while holding tasklist_lock.  Forks can happen - the
1106         * mpol_dup() cpuset_being_rebound check will catch such forks,
1107         * and rebind their vma mempolicies too.  Because we still hold
1108         * the global cpuset_mutex, we know that no other rebind effort
1109         * will be contending for the global variable cpuset_being_rebound.
1110         * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1111         * is idempotent.  Also migrate pages in each mm to new nodes.
1112         */
1113        css_scan_tasks(&cs->css, NULL, cpuset_change_nodemask, &arg, heap);
1114
1115        /*
1116         * All the tasks' nodemasks have been updated, update
1117         * cs->old_mems_allowed.
1118         */
1119        cs->old_mems_allowed = newmems;
1120
1121        /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1122        cpuset_being_rebound = NULL;
1123}
1124
1125/*
1126 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1127 * @cs: the root cpuset of the hierarchy
1128 * @update_root: update the root cpuset or not?
1129 * @heap: the heap used by css_scan_tasks()
1130 *
1131 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1132 * which take on nodemask of @root_cs.
1133 *
1134 * Called with cpuset_mutex held
1135 */
1136static void update_tasks_nodemask_hier(struct cpuset *root_cs,
1137                                       bool update_root, struct ptr_heap *heap)
1138{
1139        struct cpuset *cp;
1140        struct cgroup_subsys_state *pos_css;
1141
1142        rcu_read_lock();
1143        cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
1144                if (cp == root_cs) {
1145                        if (!update_root)
1146                                continue;
1147                } else {
1148                        /* skip the whole subtree if @cp have some CPU */
1149                        if (!nodes_empty(cp->mems_allowed)) {
1150                                pos_css = css_rightmost_descendant(pos_css);
1151                                continue;
1152                        }
1153                }
1154                if (!css_tryget(&cp->css))
1155                        continue;
1156                rcu_read_unlock();
1157
1158                update_tasks_nodemask(cp, heap);
1159
1160                rcu_read_lock();
1161                css_put(&cp->css);
1162        }
1163        rcu_read_unlock();
1164}
1165
1166/*
1167 * Handle user request to change the 'mems' memory placement
1168 * of a cpuset.  Needs to validate the request, update the
1169 * cpusets mems_allowed, and for each task in the cpuset,
1170 * update mems_allowed and rebind task's mempolicy and any vma
1171 * mempolicies and if the cpuset is marked 'memory_migrate',
1172 * migrate the tasks pages to the new memory.
1173 *
1174 * Call with cpuset_mutex held.  May take callback_mutex during call.
1175 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1176 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1177 * their mempolicies to the cpusets new mems_allowed.
1178 */
1179static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1180                           const char *buf)
1181{
1182        int retval;
1183        struct ptr_heap heap;
1184
1185        /*
1186         * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1187         * it's read-only
1188         */
1189        if (cs == &top_cpuset) {
1190                retval = -EACCES;
1191                goto done;
1192        }
1193
1194        /*
1195         * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1196         * Since nodelist_parse() fails on an empty mask, we special case
1197         * that parsing.  The validate_change() call ensures that cpusets
1198         * with tasks have memory.
1199         */
1200        if (!*buf) {
1201                nodes_clear(trialcs->mems_allowed);
1202        } else {
1203                retval = nodelist_parse(buf, trialcs->mems_allowed);
1204                if (retval < 0)
1205                        goto done;
1206
1207                if (!nodes_subset(trialcs->mems_allowed,
1208                                node_states[N_MEMORY])) {
1209                        retval =  -EINVAL;
1210                        goto done;
1211                }
1212        }
1213
1214        if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1215                retval = 0;             /* Too easy - nothing to do */
1216                goto done;
1217        }
1218        retval = validate_change(cs, trialcs);
1219        if (retval < 0)
1220                goto done;
1221
1222        retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1223        if (retval < 0)
1224                goto done;
1225
1226        mutex_lock(&callback_mutex);
1227        cs->mems_allowed = trialcs->mems_allowed;
1228        mutex_unlock(&callback_mutex);
1229
1230        update_tasks_nodemask_hier(cs, true, &heap);
1231
1232        heap_free(&heap);
1233done:
1234        return retval;
1235}
1236
1237int current_cpuset_is_being_rebound(void)
1238{
1239        return task_cs(current) == cpuset_being_rebound;
1240}
1241
1242static int update_relax_domain_level(struct cpuset *cs, s64 val)
1243{
1244#ifdef CONFIG_SMP
1245        if (val < -1 || val >= sched_domain_level_max)
1246                return -EINVAL;
1247#endif
1248
1249        if (val != cs->relax_domain_level) {
1250                cs->relax_domain_level = val;
1251                if (!cpumask_empty(cs->cpus_allowed) &&
1252                    is_sched_load_balance(cs))
1253                        rebuild_sched_domains_locked();
1254        }
1255
1256        return 0;
1257}
1258
1259/**
1260 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1261 * @tsk: task to be updated
1262 * @data: cpuset to @tsk belongs to
1263 *
1264 * Called by css_scan_tasks() for each task in a cgroup.
1265 *
1266 * We don't need to re-check for the cgroup/cpuset membership, since we're
1267 * holding cpuset_mutex at this point.
1268 */
1269static void cpuset_change_flag(struct task_struct *tsk, void *data)
1270{
1271        struct cpuset *cs = data;
1272
1273        cpuset_update_task_spread_flag(cs, tsk);
1274}
1275
1276/**
1277 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1278 * @cs: the cpuset in which each task's spread flags needs to be changed
1279 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
1280 *
1281 * Called with cpuset_mutex held
1282 *
1283 * The css_scan_tasks() function will scan all the tasks in a cgroup,
1284 * calling callback functions for each.
1285 *
1286 * No return value. It's guaranteed that css_scan_tasks() always returns 0
1287 * if @heap != NULL.
1288 */
1289static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1290{
1291        css_scan_tasks(&cs->css, NULL, cpuset_change_flag, cs, heap);
1292}
1293
1294/*
1295 * update_flag - read a 0 or a 1 in a file and update associated flag
1296 * bit:         the bit to update (see cpuset_flagbits_t)
1297 * cs:          the cpuset to update
1298 * turning_on:  whether the flag is being set or cleared
1299 *
1300 * Call with cpuset_mutex held.
1301 */
1302
1303static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1304                       int turning_on)
1305{
1306        struct cpuset *trialcs;
1307        int balance_flag_changed;
1308        int spread_flag_changed;
1309        struct ptr_heap heap;
1310        int err;
1311
1312        trialcs = alloc_trial_cpuset(cs);
1313        if (!trialcs)
1314                return -ENOMEM;
1315
1316        if (turning_on)
1317                set_bit(bit, &trialcs->flags);
1318        else
1319                clear_bit(bit, &trialcs->flags);
1320
1321        err = validate_change(cs, trialcs);
1322        if (err < 0)
1323                goto out;
1324
1325        err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1326        if (err < 0)
1327                goto out;
1328
1329        balance_flag_changed = (is_sched_load_balance(cs) !=
1330                                is_sched_load_balance(trialcs));
1331
1332        spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1333                        || (is_spread_page(cs) != is_spread_page(trialcs)));
1334
1335        mutex_lock(&callback_mutex);
1336        cs->flags = trialcs->flags;
1337        mutex_unlock(&callback_mutex);
1338
1339        if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1340                rebuild_sched_domains_locked();
1341
1342        if (spread_flag_changed)
1343                update_tasks_flags(cs, &heap);
1344        heap_free(&heap);
1345out:
1346        free_trial_cpuset(trialcs);
1347        return err;
1348}
1349
1350/*
1351 * Frequency meter - How fast is some event occurring?
1352 *
1353 * These routines manage a digitally filtered, constant time based,
1354 * event frequency meter.  There are four routines:
1355 *   fmeter_init() - initialize a frequency meter.
1356 *   fmeter_markevent() - called each time the event happens.
1357 *   fmeter_getrate() - returns the recent rate of such events.
1358 *   fmeter_update() - internal routine used to update fmeter.
1359 *
1360 * A common data structure is passed to each of these routines,
1361 * which is used to keep track of the state required to manage the
1362 * frequency meter and its digital filter.
1363 *
1364 * The filter works on the number of events marked per unit time.
1365 * The filter is single-pole low-pass recursive (IIR).  The time unit
1366 * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1367 * simulate 3 decimal digits of precision (multiplied by 1000).
1368 *
1369 * With an FM_COEF of 933, and a time base of 1 second, the filter
1370 * has a half-life of 10 seconds, meaning that if the events quit
1371 * happening, then the rate returned from the fmeter_getrate()
1372 * will be cut in half each 10 seconds, until it converges to zero.
1373 *
1374 * It is not worth doing a real infinitely recursive filter.  If more
1375 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1376 * just compute FM_MAXTICKS ticks worth, by which point the level
1377 * will be stable.
1378 *
1379 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1380 * arithmetic overflow in the fmeter_update() routine.
1381 *
1382 * Given the simple 32 bit integer arithmetic used, this meter works
1383 * best for reporting rates between one per millisecond (msec) and
1384 * one per 32 (approx) seconds.  At constant rates faster than one
1385 * per msec it maxes out at values just under 1,000,000.  At constant
1386 * rates between one per msec, and one per second it will stabilize
1387 * to a value N*1000, where N is the rate of events per second.
1388 * At constant rates between one per second and one per 32 seconds,
1389 * it will be choppy, moving up on the seconds that have an event,
1390 * and then decaying until the next event.  At rates slower than
1391 * about one in 32 seconds, it decays all the way back to zero between
1392 * each event.
1393 */
1394
1395#define FM_COEF 933             /* coefficient for half-life of 10 secs */
1396#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1397#define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1398#define FM_SCALE 1000           /* faux fixed point scale */
1399
1400/* Initialize a frequency meter */
1401static void fmeter_init(struct fmeter *fmp)
1402{
1403        fmp->cnt = 0;
1404        fmp->val = 0;
1405        fmp->time = 0;
1406        spin_lock_init(&fmp->lock);
1407}
1408
1409/* Internal meter update - process cnt events and update value */
1410static void fmeter_update(struct fmeter *fmp)
1411{
1412        time_t now = get_seconds();
1413        time_t ticks = now - fmp->time;
1414
1415        if (ticks == 0)
1416                return;
1417
1418        ticks = min(FM_MAXTICKS, ticks);
1419        while (ticks-- > 0)
1420                fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1421        fmp->time = now;
1422
1423        fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1424        fmp->cnt = 0;
1425}
1426
1427/* Process any previous ticks, then bump cnt by one (times scale). */
1428static void fmeter_markevent(struct fmeter *fmp)
1429{
1430        spin_lock(&fmp->lock);
1431        fmeter_update(fmp);
1432        fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1433        spin_unlock(&fmp->lock);
1434}
1435
1436/* Process any previous ticks, then return current value. */
1437static int fmeter_getrate(struct fmeter *fmp)
1438{
1439        int val;
1440
1441        spin_lock(&fmp->lock);
1442        fmeter_update(fmp);
1443        val = fmp->val;
1444        spin_unlock(&fmp->lock);
1445        return val;
1446}
1447
1448/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1449static int cpuset_can_attach(struct cgroup_subsys_state *css,
1450                             struct cgroup_taskset *tset)
1451{
1452        struct cpuset *cs = css_cs(css);
1453        struct task_struct *task;
1454        int ret;
1455
1456        mutex_lock(&cpuset_mutex);
1457
1458        /*
1459         * We allow to move tasks into an empty cpuset if sane_behavior
1460         * flag is set.
1461         */
1462        ret = -ENOSPC;
1463        if (!cgroup_sane_behavior(css->cgroup) &&
1464            (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1465                goto out_unlock;
1466
1467        cgroup_taskset_for_each(task, css, tset) {
1468                /*
1469                 * Kthreads which disallow setaffinity shouldn't be moved
1470                 * to a new cpuset; we don't want to change their cpu
1471                 * affinity and isolating such threads by their set of
1472                 * allowed nodes is unnecessary.  Thus, cpusets are not
1473                 * applicable for such threads.  This prevents checking for
1474                 * success of set_cpus_allowed_ptr() on all attached tasks
1475                 * before cpus_allowed may be changed.
1476                 */
1477                ret = -EINVAL;
1478                if (task->flags & PF_NO_SETAFFINITY)
1479                        goto out_unlock;
1480                ret = security_task_setscheduler(task);
1481                if (ret)
1482                        goto out_unlock;
1483        }
1484
1485        /*
1486         * Mark attach is in progress.  This makes validate_change() fail
1487         * changes which zero cpus/mems_allowed.
1488         */
1489        cs->attach_in_progress++;
1490        ret = 0;
1491out_unlock:
1492        mutex_unlock(&cpuset_mutex);
1493        return ret;
1494}
1495
1496static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1497                                 struct cgroup_taskset *tset)
1498{
1499        mutex_lock(&cpuset_mutex);
1500        css_cs(css)->attach_in_progress--;
1501        mutex_unlock(&cpuset_mutex);
1502}
1503
1504/*
1505 * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1506 * but we can't allocate it dynamically there.  Define it global and
1507 * allocate from cpuset_init().
1508 */
1509static cpumask_var_t cpus_attach;
1510
1511static void cpuset_attach(struct cgroup_subsys_state *css,
1512                          struct cgroup_taskset *tset)
1513{
1514        /* static buf protected by cpuset_mutex */
1515        static nodemask_t cpuset_attach_nodemask_to;
1516        struct mm_struct *mm;
1517        struct task_struct *task;
1518        struct task_struct *leader = cgroup_taskset_first(tset);
1519        struct cgroup_subsys_state *oldcss = cgroup_taskset_cur_css(tset,
1520                                                        cpuset_subsys_id);
1521        struct cpuset *cs = css_cs(css);
1522        struct cpuset *oldcs = css_cs(oldcss);
1523        struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1524        struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1525
1526        mutex_lock(&cpuset_mutex);
1527
1528        /* prepare for attach */
1529        if (cs == &top_cpuset)
1530                cpumask_copy(cpus_attach, cpu_possible_mask);
1531        else
1532                guarantee_online_cpus(cpus_cs, cpus_attach);
1533
1534        guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1535
1536        cgroup_taskset_for_each(task, css, tset) {
1537                /*
1538                 * can_attach beforehand should guarantee that this doesn't
1539                 * fail.  TODO: have a better way to handle failure here
1540                 */
1541                WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1542
1543                cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1544                cpuset_update_task_spread_flag(cs, task);
1545        }
1546
1547        /*
1548         * Change mm, possibly for multiple threads in a threadgroup. This is
1549         * expensive and may sleep.
1550         */
1551        cpuset_attach_nodemask_to = cs->mems_allowed;
1552        mm = get_task_mm(leader);
1553        if (mm) {
1554                struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1555
1556                mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1557
1558                /*
1559                 * old_mems_allowed is the same with mems_allowed here, except
1560                 * if this task is being moved automatically due to hotplug.
1561                 * In that case @mems_allowed has been updated and is empty,
1562                 * so @old_mems_allowed is the right nodesets that we migrate
1563                 * mm from.
1564                 */
1565                if (is_memory_migrate(cs)) {
1566                        cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
1567                                          &cpuset_attach_nodemask_to);
1568                }
1569                mmput(mm);
1570        }
1571
1572        cs->old_mems_allowed = cpuset_attach_nodemask_to;
1573
1574        cs->attach_in_progress--;
1575        if (!cs->attach_in_progress)
1576                wake_up(&cpuset_attach_wq);
1577
1578        mutex_unlock(&cpuset_mutex);
1579}
1580
1581/* The various types of files and directories in a cpuset file system */
1582
1583typedef enum {
1584        FILE_MEMORY_MIGRATE,
1585        FILE_CPULIST,
1586        FILE_MEMLIST,
1587        FILE_CPU_EXCLUSIVE,
1588        FILE_MEM_EXCLUSIVE,
1589        FILE_MEM_HARDWALL,
1590        FILE_SCHED_LOAD_BALANCE,
1591        FILE_SCHED_RELAX_DOMAIN_LEVEL,
1592        FILE_MEMORY_PRESSURE_ENABLED,
1593        FILE_MEMORY_PRESSURE,
1594        FILE_SPREAD_PAGE,
1595        FILE_SPREAD_SLAB,
1596} cpuset_filetype_t;
1597
1598static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1599                            u64 val)
1600{
1601        struct cpuset *cs = css_cs(css);
1602        cpuset_filetype_t type = cft->private;
1603        int retval = 0;
1604
1605        mutex_lock(&cpuset_mutex);
1606        if (!is_cpuset_online(cs)) {
1607                retval = -ENODEV;
1608                goto out_unlock;
1609        }
1610
1611        switch (type) {
1612        case FILE_CPU_EXCLUSIVE:
1613                retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1614                break;
1615        case FILE_MEM_EXCLUSIVE:
1616                retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1617                break;
1618        case FILE_MEM_HARDWALL:
1619                retval = update_flag(CS_MEM_HARDWALL, cs, val);
1620                break;
1621        case FILE_SCHED_LOAD_BALANCE:
1622                retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1623                break;
1624        case FILE_MEMORY_MIGRATE:
1625                retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1626                break;
1627        case FILE_MEMORY_PRESSURE_ENABLED:
1628                cpuset_memory_pressure_enabled = !!val;
1629                break;
1630        case FILE_MEMORY_PRESSURE:
1631                retval = -EACCES;
1632                break;
1633        case FILE_SPREAD_PAGE:
1634                retval = update_flag(CS_SPREAD_PAGE, cs, val);
1635                break;
1636        case FILE_SPREAD_SLAB:
1637                retval = update_flag(CS_SPREAD_SLAB, cs, val);
1638                break;
1639        default:
1640                retval = -EINVAL;
1641                break;
1642        }
1643out_unlock:
1644        mutex_unlock(&cpuset_mutex);
1645        return retval;
1646}
1647
1648static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1649                            s64 val)
1650{
1651        struct cpuset *cs = css_cs(css);
1652        cpuset_filetype_t type = cft->private;
1653        int retval = -ENODEV;
1654
1655        mutex_lock(&cpuset_mutex);
1656        if (!is_cpuset_online(cs))
1657                goto out_unlock;
1658
1659        switch (type) {
1660        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1661                retval = update_relax_domain_level(cs, val);
1662                break;
1663        default:
1664                retval = -EINVAL;
1665                break;
1666        }
1667out_unlock:
1668        mutex_unlock(&cpuset_mutex);
1669        return retval;
1670}
1671
1672/*
1673 * Common handling for a write to a "cpus" or "mems" file.
1674 */
1675static int cpuset_write_resmask(struct cgroup_subsys_state *css,
1676                                struct cftype *cft, const char *buf)
1677{
1678        struct cpuset *cs = css_cs(css);
1679        struct cpuset *trialcs;
1680        int retval = -ENODEV;
1681
1682        /*
1683         * CPU or memory hotunplug may leave @cs w/o any execution
1684         * resources, in which case the hotplug code asynchronously updates
1685         * configuration and transfers all tasks to the nearest ancestor
1686         * which can execute.
1687         *
1688         * As writes to "cpus" or "mems" may restore @cs's execution
1689         * resources, wait for the previously scheduled operations before
1690         * proceeding, so that we don't end up keep removing tasks added
1691         * after execution capability is restored.
1692         */
1693        flush_work(&cpuset_hotplug_work);
1694
1695        mutex_lock(&cpuset_mutex);
1696        if (!is_cpuset_online(cs))
1697                goto out_unlock;
1698
1699        trialcs = alloc_trial_cpuset(cs);
1700        if (!trialcs) {
1701                retval = -ENOMEM;
1702                goto out_unlock;
1703        }
1704
1705        switch (cft->private) {
1706        case FILE_CPULIST:
1707                retval = update_cpumask(cs, trialcs, buf);
1708                break;
1709        case FILE_MEMLIST:
1710                retval = update_nodemask(cs, trialcs, buf);
1711                break;
1712        default:
1713                retval = -EINVAL;
1714                break;
1715        }
1716
1717        free_trial_cpuset(trialcs);
1718out_unlock:
1719        mutex_unlock(&cpuset_mutex);
1720        return retval;
1721}
1722
1723/*
1724 * These ascii lists should be read in a single call, by using a user
1725 * buffer large enough to hold the entire map.  If read in smaller
1726 * chunks, there is no guarantee of atomicity.  Since the display format
1727 * used, list of ranges of sequential numbers, is variable length,
1728 * and since these maps can change value dynamically, one could read
1729 * gibberish by doing partial reads while a list was changing.
1730 * A single large read to a buffer that crosses a page boundary is
1731 * ok, because the result being copied to user land is not recomputed
1732 * across a page fault.
1733 */
1734
1735static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1736{
1737        size_t count;
1738
1739        mutex_lock(&callback_mutex);
1740        count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1741        mutex_unlock(&callback_mutex);
1742
1743        return count;
1744}
1745
1746static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1747{
1748        size_t count;
1749
1750        mutex_lock(&callback_mutex);
1751        count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1752        mutex_unlock(&callback_mutex);
1753
1754        return count;
1755}
1756
1757static ssize_t cpuset_common_file_read(struct cgroup_subsys_state *css,
1758                                       struct cftype *cft, struct file *file,
1759                                       char __user *buf, size_t nbytes,
1760                                       loff_t *ppos)
1761{
1762        struct cpuset *cs = css_cs(css);
1763        cpuset_filetype_t type = cft->private;
1764        char *page;
1765        ssize_t retval = 0;
1766        char *s;
1767
1768        if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1769                return -ENOMEM;
1770
1771        s = page;
1772
1773        switch (type) {
1774        case FILE_CPULIST:
1775                s += cpuset_sprintf_cpulist(s, cs);
1776                break;
1777        case FILE_MEMLIST:
1778                s += cpuset_sprintf_memlist(s, cs);
1779                break;
1780        default:
1781                retval = -EINVAL;
1782                goto out;
1783        }
1784        *s++ = '\n';
1785
1786        retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1787out:
1788        free_page((unsigned long)page);
1789        return retval;
1790}
1791
1792static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1793{
1794        struct cpuset *cs = css_cs(css);
1795        cpuset_filetype_t type = cft->private;
1796        switch (type) {
1797        case FILE_CPU_EXCLUSIVE:
1798                return is_cpu_exclusive(cs);
1799        case FILE_MEM_EXCLUSIVE:
1800                return is_mem_exclusive(cs);
1801        case FILE_MEM_HARDWALL:
1802                return is_mem_hardwall(cs);
1803        case FILE_SCHED_LOAD_BALANCE:
1804                return is_sched_load_balance(cs);
1805        case FILE_MEMORY_MIGRATE:
1806                return is_memory_migrate(cs);
1807        case FILE_MEMORY_PRESSURE_ENABLED:
1808                return cpuset_memory_pressure_enabled;
1809        case FILE_MEMORY_PRESSURE:
1810                return fmeter_getrate(&cs->fmeter);
1811        case FILE_SPREAD_PAGE:
1812                return is_spread_page(cs);
1813        case FILE_SPREAD_SLAB:
1814                return is_spread_slab(cs);
1815        default:
1816                BUG();
1817        }
1818
1819        /* Unreachable but makes gcc happy */
1820        return 0;
1821}
1822
1823static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1824{
1825        struct cpuset *cs = css_cs(css);
1826        cpuset_filetype_t type = cft->private;
1827        switch (type) {
1828        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1829                return cs->relax_domain_level;
1830        default:
1831                BUG();
1832        }
1833
1834        /* Unrechable but makes gcc happy */
1835        return 0;
1836}
1837
1838
1839/*
1840 * for the common functions, 'private' gives the type of file
1841 */
1842
1843static struct cftype files[] = {
1844        {
1845                .name = "cpus",
1846                .read = cpuset_common_file_read,
1847                .write_string = cpuset_write_resmask,
1848                .max_write_len = (100U + 6 * NR_CPUS),
1849                .private = FILE_CPULIST,
1850        },
1851
1852        {
1853                .name = "mems",
1854                .read = cpuset_common_file_read,
1855                .write_string = cpuset_write_resmask,
1856                .max_write_len = (100U + 6 * MAX_NUMNODES),
1857                .private = FILE_MEMLIST,
1858        },
1859
1860        {
1861                .name = "cpu_exclusive",
1862                .read_u64 = cpuset_read_u64,
1863                .write_u64 = cpuset_write_u64,
1864                .private = FILE_CPU_EXCLUSIVE,
1865        },
1866
1867        {
1868                .name = "mem_exclusive",
1869                .read_u64 = cpuset_read_u64,
1870                .write_u64 = cpuset_write_u64,
1871                .private = FILE_MEM_EXCLUSIVE,
1872        },
1873
1874        {
1875                .name = "mem_hardwall",
1876                .read_u64 = cpuset_read_u64,
1877                .write_u64 = cpuset_write_u64,
1878                .private = FILE_MEM_HARDWALL,
1879        },
1880
1881        {
1882                .name = "sched_load_balance",
1883                .read_u64 = cpuset_read_u64,
1884                .write_u64 = cpuset_write_u64,
1885                .private = FILE_SCHED_LOAD_BALANCE,
1886        },
1887
1888        {
1889                .name = "sched_relax_domain_level",
1890                .read_s64 = cpuset_read_s64,
1891                .write_s64 = cpuset_write_s64,
1892                .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1893        },
1894
1895        {
1896                .name = "memory_migrate",
1897                .read_u64 = cpuset_read_u64,
1898                .write_u64 = cpuset_write_u64,
1899                .private = FILE_MEMORY_MIGRATE,
1900        },
1901
1902        {
1903                .name = "memory_pressure",
1904                .read_u64 = cpuset_read_u64,
1905                .write_u64 = cpuset_write_u64,
1906                .private = FILE_MEMORY_PRESSURE,
1907                .mode = S_IRUGO,
1908        },
1909
1910        {
1911                .name = "memory_spread_page",
1912                .read_u64 = cpuset_read_u64,
1913                .write_u64 = cpuset_write_u64,
1914                .private = FILE_SPREAD_PAGE,
1915        },
1916
1917        {
1918                .name = "memory_spread_slab",
1919                .read_u64 = cpuset_read_u64,
1920                .write_u64 = cpuset_write_u64,
1921                .private = FILE_SPREAD_SLAB,
1922        },
1923
1924        {
1925                .name = "memory_pressure_enabled",
1926                .flags = CFTYPE_ONLY_ON_ROOT,
1927                .read_u64 = cpuset_read_u64,
1928                .write_u64 = cpuset_write_u64,
1929                .private = FILE_MEMORY_PRESSURE_ENABLED,
1930        },
1931
1932        { }     /* terminate */
1933};
1934
1935/*
1936 *      cpuset_css_alloc - allocate a cpuset css
1937 *      cgrp:   control group that the new cpuset will be part of
1938 */
1939
1940static struct cgroup_subsys_state *
1941cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1942{
1943        struct cpuset *cs;
1944
1945        if (!parent_css)
1946                return &top_cpuset.css;
1947
1948        cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1949        if (!cs)
1950                return ERR_PTR(-ENOMEM);
1951        if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1952                kfree(cs);
1953                return ERR_PTR(-ENOMEM);
1954        }
1955
1956        set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1957        cpumask_clear(cs->cpus_allowed);
1958        nodes_clear(cs->mems_allowed);
1959        fmeter_init(&cs->fmeter);
1960        cs->relax_domain_level = -1;
1961
1962        return &cs->css;
1963}
1964
1965static int cpuset_css_online(struct cgroup_subsys_state *css)
1966{
1967        struct cpuset *cs = css_cs(css);
1968        struct cpuset *parent = parent_cs(cs);
1969        struct cpuset *tmp_cs;
1970        struct cgroup_subsys_state *pos_css;
1971
1972        if (!parent)
1973                return 0;
1974
1975        mutex_lock(&cpuset_mutex);
1976
1977        set_bit(CS_ONLINE, &cs->flags);
1978        if (is_spread_page(parent))
1979                set_bit(CS_SPREAD_PAGE, &cs->flags);
1980        if (is_spread_slab(parent))
1981                set_bit(CS_SPREAD_SLAB, &cs->flags);
1982
1983        number_of_cpusets++;
1984
1985        if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1986                goto out_unlock;
1987
1988        /*
1989         * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1990         * set.  This flag handling is implemented in cgroup core for
1991         * histrical reasons - the flag may be specified during mount.
1992         *
1993         * Currently, if any sibling cpusets have exclusive cpus or mem, we
1994         * refuse to clone the configuration - thereby refusing the task to
1995         * be entered, and as a result refusing the sys_unshare() or
1996         * clone() which initiated it.  If this becomes a problem for some
1997         * users who wish to allow that scenario, then this could be
1998         * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1999         * (and likewise for mems) to the new cgroup.
2000         */
2001        rcu_read_lock();
2002        cpuset_for_each_child(tmp_cs, pos_css, parent) {
2003                if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2004                        rcu_read_unlock();
2005                        goto out_unlock;
2006                }
2007        }
2008        rcu_read_unlock();
2009
2010        mutex_lock(&callback_mutex);
2011        cs->mems_allowed = parent->mems_allowed;
2012        cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2013        mutex_unlock(&callback_mutex);
2014out_unlock:
2015        mutex_unlock(&cpuset_mutex);
2016        return 0;
2017}
2018
2019/*
2020 * If the cpuset being removed has its flag 'sched_load_balance'
2021 * enabled, then simulate turning sched_load_balance off, which
2022 * will call rebuild_sched_domains_locked().
2023 */
2024
2025static void cpuset_css_offline(struct cgroup_subsys_state *css)
2026{
2027        struct cpuset *cs = css_cs(css);
2028
2029        mutex_lock(&cpuset_mutex);
2030
2031        if (is_sched_load_balance(cs))
2032                update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2033
2034        number_of_cpusets--;
2035        clear_bit(CS_ONLINE, &cs->flags);
2036
2037        mutex_unlock(&cpuset_mutex);
2038}
2039
2040static void cpuset_css_free(struct cgroup_subsys_state *css)
2041{
2042        struct cpuset *cs = css_cs(css);
2043
2044        free_cpumask_var(cs->cpus_allowed);
2045        kfree(cs);
2046}
2047
2048struct cgroup_subsys cpuset_subsys = {
2049        .name = "cpuset",
2050        .css_alloc = cpuset_css_alloc,
2051        .css_online = cpuset_css_online,
2052        .css_offline = cpuset_css_offline,
2053        .css_free = cpuset_css_free,
2054        .can_attach = cpuset_can_attach,
2055        .cancel_attach = cpuset_cancel_attach,
2056        .attach = cpuset_attach,
2057        .subsys_id = cpuset_subsys_id,
2058        .base_cftypes = files,
2059        .early_init = 1,
2060};
2061
2062/**
2063 * cpuset_init - initialize cpusets at system boot
2064 *
2065 * Description: Initialize top_cpuset and the cpuset internal file system,
2066 **/
2067
2068int __init cpuset_init(void)
2069{
2070        int err = 0;
2071
2072        if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2073                BUG();
2074
2075        cpumask_setall(top_cpuset.cpus_allowed);
2076        nodes_setall(top_cpuset.mems_allowed);
2077
2078        fmeter_init(&top_cpuset.fmeter);
2079        set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2080        top_cpuset.relax_domain_level = -1;
2081
2082        err = register_filesystem(&cpuset_fs_type);
2083        if (err < 0)
2084                return err;
2085
2086        if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2087                BUG();
2088
2089        number_of_cpusets = 1;
2090        return 0;
2091}
2092
2093/*
2094 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2095 * or memory nodes, we need to walk over the cpuset hierarchy,
2096 * removing that CPU or node from all cpusets.  If this removes the
2097 * last CPU or node from a cpuset, then move the tasks in the empty
2098 * cpuset to its next-highest non-empty parent.
2099 */
2100static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2101{
2102        struct cpuset *parent;
2103
2104        /*
2105         * Find its next-highest non-empty parent, (top cpuset
2106         * has online cpus, so can't be empty).
2107         */
2108        parent = parent_cs(cs);
2109        while (cpumask_empty(parent->cpus_allowed) ||
2110                        nodes_empty(parent->mems_allowed))
2111                parent = parent_cs(parent);
2112
2113        if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2114                rcu_read_lock();
2115                printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2116                       cgroup_name(cs->css.cgroup));
2117                rcu_read_unlock();
2118        }
2119}
2120
2121/**
2122 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2123 * @cs: cpuset in interest
2124 *
2125 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2126 * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2127 * all its tasks are moved to the nearest ancestor with both resources.
2128 */
2129static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2130{
2131        static cpumask_t off_cpus;
2132        static nodemask_t off_mems;
2133        bool is_empty;
2134        bool sane = cgroup_sane_behavior(cs->css.cgroup);
2135
2136retry:
2137        wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2138
2139        mutex_lock(&cpuset_mutex);
2140
2141        /*
2142         * We have raced with task attaching. We wait until attaching
2143         * is finished, so we won't attach a task to an empty cpuset.
2144         */
2145        if (cs->attach_in_progress) {
2146                mutex_unlock(&cpuset_mutex);
2147                goto retry;
2148        }
2149
2150        cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2151        nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2152
2153        mutex_lock(&callback_mutex);
2154        cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2155        mutex_unlock(&callback_mutex);
2156
2157        /*
2158         * If sane_behavior flag is set, we need to update tasks' cpumask
2159         * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2160         * call update_tasks_cpumask() if the cpuset becomes empty, as
2161         * the tasks in it will be migrated to an ancestor.
2162         */
2163        if ((sane && cpumask_empty(cs->cpus_allowed)) ||
2164            (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
2165                update_tasks_cpumask(cs, NULL);
2166
2167        mutex_lock(&callback_mutex);
2168        nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2169        mutex_unlock(&callback_mutex);
2170
2171        /*
2172         * If sane_behavior flag is set, we need to update tasks' nodemask
2173         * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2174         * call update_tasks_nodemask() if the cpuset becomes empty, as
2175         * the tasks in it will be migratd to an ancestor.
2176         */
2177        if ((sane && nodes_empty(cs->mems_allowed)) ||
2178            (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
2179                update_tasks_nodemask(cs, NULL);
2180
2181        is_empty = cpumask_empty(cs->cpus_allowed) ||
2182                nodes_empty(cs->mems_allowed);
2183
2184        mutex_unlock(&cpuset_mutex);
2185
2186        /*
2187         * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2188         *
2189         * Otherwise move tasks to the nearest ancestor with execution
2190         * resources.  This is full cgroup operation which will
2191         * also call back into cpuset.  Should be done outside any lock.
2192         */
2193        if (!sane && is_empty)
2194                remove_tasks_in_empty_cpuset(cs);
2195}
2196
2197/**
2198 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2199 *
2200 * This function is called after either CPU or memory configuration has
2201 * changed and updates cpuset accordingly.  The top_cpuset is always
2202 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2203 * order to make cpusets transparent (of no affect) on systems that are
2204 * actively using CPU hotplug but making no active use of cpusets.
2205 *
2206 * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2207 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2208 * all descendants.
2209 *
2210 * Note that CPU offlining during suspend is ignored.  We don't modify
2211 * cpusets across suspend/resume cycles at all.
2212 */
2213static void cpuset_hotplug_workfn(struct work_struct *work)
2214{
2215        static cpumask_t new_cpus;
2216        static nodemask_t new_mems;
2217        bool cpus_updated, mems_updated;
2218
2219        mutex_lock(&cpuset_mutex);
2220
2221        /* fetch the available cpus/mems and find out which changed how */
2222        cpumask_copy(&new_cpus, cpu_active_mask);
2223        new_mems = node_states[N_MEMORY];
2224
2225        cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2226        mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2227
2228        /* synchronize cpus_allowed to cpu_active_mask */
2229        if (cpus_updated) {
2230                mutex_lock(&callback_mutex);
2231                cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2232                mutex_unlock(&callback_mutex);
2233                /* we don't mess with cpumasks of tasks in top_cpuset */
2234        }
2235
2236        /* synchronize mems_allowed to N_MEMORY */
2237        if (mems_updated) {
2238                mutex_lock(&callback_mutex);
2239                top_cpuset.mems_allowed = new_mems;
2240                mutex_unlock(&callback_mutex);
2241                update_tasks_nodemask(&top_cpuset, NULL);
2242        }
2243
2244        mutex_unlock(&cpuset_mutex);
2245
2246        /* if cpus or mems changed, we need to propagate to descendants */
2247        if (cpus_updated || mems_updated) {
2248                struct cpuset *cs;
2249                struct cgroup_subsys_state *pos_css;
2250
2251                rcu_read_lock();
2252                cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2253                        if (cs == &top_cpuset || !css_tryget(&cs->css))
2254                                continue;
2255                        rcu_read_unlock();
2256
2257                        cpuset_hotplug_update_tasks(cs);
2258
2259                        rcu_read_lock();
2260                        css_put(&cs->css);
2261                }
2262                rcu_read_unlock();
2263        }
2264
2265        /* rebuild sched domains if cpus_allowed has changed */
2266        if (cpus_updated)
2267                rebuild_sched_domains();
2268}
2269
2270void cpuset_update_active_cpus(bool cpu_online)
2271{
2272        /*
2273         * We're inside cpu hotplug critical region which usually nests
2274         * inside cgroup synchronization.  Bounce actual hotplug processing
2275         * to a work item to avoid reverse locking order.
2276         *
2277         * We still need to do partition_sched_domains() synchronously;
2278         * otherwise, the scheduler will get confused and put tasks to the
2279         * dead CPU.  Fall back to the default single domain.
2280         * cpuset_hotplug_workfn() will rebuild it as necessary.
2281         */
2282        partition_sched_domains(1, NULL, NULL);
2283        schedule_work(&cpuset_hotplug_work);
2284}
2285
2286/*
2287 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2288 * Call this routine anytime after node_states[N_MEMORY] changes.
2289 * See cpuset_update_active_cpus() for CPU hotplug handling.
2290 */
2291static int cpuset_track_online_nodes(struct notifier_block *self,
2292                                unsigned long action, void *arg)
2293{
2294        schedule_work(&cpuset_hotplug_work);
2295        return NOTIFY_OK;
2296}
2297
2298static struct notifier_block cpuset_track_online_nodes_nb = {
2299        .notifier_call = cpuset_track_online_nodes,
2300        .priority = 10,         /* ??! */
2301};
2302
2303/**
2304 * cpuset_init_smp - initialize cpus_allowed
2305 *
2306 * Description: Finish top cpuset after cpu, node maps are initialized
2307 */
2308void __init cpuset_init_smp(void)
2309{
2310        cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2311        top_cpuset.mems_allowed = node_states[N_MEMORY];
2312        top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2313
2314        register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2315}
2316
2317/**
2318 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2319 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2320 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2321 *
2322 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2323 * attached to the specified @tsk.  Guaranteed to return some non-empty
2324 * subset of cpu_online_mask, even if this means going outside the
2325 * tasks cpuset.
2326 **/
2327
2328void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2329{
2330        struct cpuset *cpus_cs;
2331
2332        mutex_lock(&callback_mutex);
2333        task_lock(tsk);
2334        cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2335        guarantee_online_cpus(cpus_cs, pmask);
2336        task_unlock(tsk);
2337        mutex_unlock(&callback_mutex);
2338}
2339
2340void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2341{
2342        struct cpuset *cpus_cs;
2343
2344        rcu_read_lock();
2345        cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2346        do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2347        rcu_read_unlock();
2348
2349        /*
2350         * We own tsk->cpus_allowed, nobody can change it under us.
2351         *
2352         * But we used cs && cs->cpus_allowed lockless and thus can
2353         * race with cgroup_attach_task() or update_cpumask() and get
2354         * the wrong tsk->cpus_allowed. However, both cases imply the
2355         * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2356         * which takes task_rq_lock().
2357         *
2358         * If we are called after it dropped the lock we must see all
2359         * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2360         * set any mask even if it is not right from task_cs() pov,
2361         * the pending set_cpus_allowed_ptr() will fix things.
2362         *
2363         * select_fallback_rq() will fix things ups and set cpu_possible_mask
2364         * if required.
2365         */
2366}
2367
2368void cpuset_init_current_mems_allowed(void)
2369{
2370        nodes_setall(current->mems_allowed);
2371}
2372
2373/**
2374 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2375 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2376 *
2377 * Description: Returns the nodemask_t mems_allowed of the cpuset
2378 * attached to the specified @tsk.  Guaranteed to return some non-empty
2379 * subset of node_states[N_MEMORY], even if this means going outside the
2380 * tasks cpuset.
2381 **/
2382
2383nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2384{
2385        struct cpuset *mems_cs;
2386        nodemask_t mask;
2387
2388        mutex_lock(&callback_mutex);
2389        task_lock(tsk);
2390        mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2391        guarantee_online_mems(mems_cs, &mask);
2392        task_unlock(tsk);
2393        mutex_unlock(&callback_mutex);
2394
2395        return mask;
2396}
2397
2398/**
2399 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2400 * @nodemask: the nodemask to be checked
2401 *
2402 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2403 */
2404int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2405{
2406        return nodes_intersects(*nodemask, current->mems_allowed);
2407}
2408
2409/*
2410 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2411 * mem_hardwall ancestor to the specified cpuset.  Call holding
2412 * callback_mutex.  If no ancestor is mem_exclusive or mem_hardwall
2413 * (an unusual configuration), then returns the root cpuset.
2414 */
2415static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2416{
2417        while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2418                cs = parent_cs(cs);
2419        return cs;
2420}
2421
2422/**
2423 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2424 * @node: is this an allowed node?
2425 * @gfp_mask: memory allocation flags
2426 *
2427 * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2428 * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2429 * yes.  If it's not a __GFP_HARDWALL request and this node is in the nearest
2430 * hardwalled cpuset ancestor to this task's cpuset, yes.  If the task has been
2431 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2432 * flag, yes.
2433 * Otherwise, no.
2434 *
2435 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2436 * cpuset_node_allowed_hardwall().  Otherwise, cpuset_node_allowed_softwall()
2437 * might sleep, and might allow a node from an enclosing cpuset.
2438 *
2439 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2440 * cpusets, and never sleeps.
2441 *
2442 * The __GFP_THISNODE placement logic is really handled elsewhere,
2443 * by forcibly using a zonelist starting at a specified node, and by
2444 * (in get_page_from_freelist()) refusing to consider the zones for
2445 * any node on the zonelist except the first.  By the time any such
2446 * calls get to this routine, we should just shut up and say 'yes'.
2447 *
2448 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2449 * and do not allow allocations outside the current tasks cpuset
2450 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2451 * GFP_KERNEL allocations are not so marked, so can escape to the
2452 * nearest enclosing hardwalled ancestor cpuset.
2453 *
2454 * Scanning up parent cpusets requires callback_mutex.  The
2455 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2456 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2457 * current tasks mems_allowed came up empty on the first pass over
2458 * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2459 * cpuset are short of memory, might require taking the callback_mutex
2460 * mutex.
2461 *
2462 * The first call here from mm/page_alloc:get_page_from_freelist()
2463 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2464 * so no allocation on a node outside the cpuset is allowed (unless
2465 * in interrupt, of course).
2466 *
2467 * The second pass through get_page_from_freelist() doesn't even call
2468 * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2469 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2470 * in alloc_flags.  That logic and the checks below have the combined
2471 * affect that:
2472 *      in_interrupt - any node ok (current task context irrelevant)
2473 *      GFP_ATOMIC   - any node ok
2474 *      TIF_MEMDIE   - any node ok
2475 *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2476 *      GFP_USER     - only nodes in current tasks mems allowed ok.
2477 *
2478 * Rule:
2479 *    Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2480 *    pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2481 *    the code that might scan up ancestor cpusets and sleep.
2482 */
2483int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2484{
2485        struct cpuset *cs;              /* current cpuset ancestors */
2486        int allowed;                    /* is allocation in zone z allowed? */
2487
2488        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2489                return 1;
2490        might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2491        if (node_isset(node, current->mems_allowed))
2492                return 1;
2493        /*
2494         * Allow tasks that have access to memory reserves because they have
2495         * been OOM killed to get memory anywhere.
2496         */
2497        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2498                return 1;
2499        if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2500                return 0;
2501
2502        if (current->flags & PF_EXITING) /* Let dying task have memory */
2503                return 1;
2504
2505        /* Not hardwall and node outside mems_allowed: scan up cpusets */
2506        mutex_lock(&callback_mutex);
2507
2508        task_lock(current);
2509        cs = nearest_hardwall_ancestor(task_cs(current));
2510        task_unlock(current);
2511
2512        allowed = node_isset(node, cs->mems_allowed);
2513        mutex_unlock(&callback_mutex);
2514        return allowed;
2515}
2516
2517/*
2518 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2519 * @node: is this an allowed node?
2520 * @gfp_mask: memory allocation flags
2521 *
2522 * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2523 * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2524 * yes.  If the task has been OOM killed and has access to memory reserves as
2525 * specified by the TIF_MEMDIE flag, yes.
2526 * Otherwise, no.
2527 *
2528 * The __GFP_THISNODE placement logic is really handled elsewhere,
2529 * by forcibly using a zonelist starting at a specified node, and by
2530 * (in get_page_from_freelist()) refusing to consider the zones for
2531 * any node on the zonelist except the first.  By the time any such
2532 * calls get to this routine, we should just shut up and say 'yes'.
2533 *
2534 * Unlike the cpuset_node_allowed_softwall() variant, above,
2535 * this variant requires that the node be in the current task's
2536 * mems_allowed or that we're in interrupt.  It does not scan up the
2537 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2538 * It never sleeps.
2539 */
2540int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2541{
2542        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2543                return 1;
2544        if (node_isset(node, current->mems_allowed))
2545                return 1;
2546        /*
2547         * Allow tasks that have access to memory reserves because they have
2548         * been OOM killed to get memory anywhere.
2549         */
2550        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2551                return 1;
2552        return 0;
2553}
2554
2555/**
2556 * cpuset_mem_spread_node() - On which node to begin search for a file page
2557 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2558 *
2559 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2560 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2561 * and if the memory allocation used cpuset_mem_spread_node()
2562 * to determine on which node to start looking, as it will for
2563 * certain page cache or slab cache pages such as used for file
2564 * system buffers and inode caches, then instead of starting on the
2565 * local node to look for a free page, rather spread the starting
2566 * node around the tasks mems_allowed nodes.
2567 *
2568 * We don't have to worry about the returned node being offline
2569 * because "it can't happen", and even if it did, it would be ok.
2570 *
2571 * The routines calling guarantee_online_mems() are careful to
2572 * only set nodes in task->mems_allowed that are online.  So it
2573 * should not be possible for the following code to return an
2574 * offline node.  But if it did, that would be ok, as this routine
2575 * is not returning the node where the allocation must be, only
2576 * the node where the search should start.  The zonelist passed to
2577 * __alloc_pages() will include all nodes.  If the slab allocator
2578 * is passed an offline node, it will fall back to the local node.
2579 * See kmem_cache_alloc_node().
2580 */
2581
2582static int cpuset_spread_node(int *rotor)
2583{
2584        int node;
2585
2586        node = next_node(*rotor, current->mems_allowed);
2587        if (node == MAX_NUMNODES)
2588                node = first_node(current->mems_allowed);
2589        *rotor = node;
2590        return node;
2591}
2592
2593int cpuset_mem_spread_node(void)
2594{
2595        if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2596                current->cpuset_mem_spread_rotor =
2597                        node_random(&current->mems_allowed);
2598
2599        return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2600}
2601
2602int cpuset_slab_spread_node(void)
2603{
2604        if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2605                current->cpuset_slab_spread_rotor =
2606                        node_random(&current->mems_allowed);
2607
2608        return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2609}
2610
2611EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2612
2613/**
2614 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2615 * @tsk1: pointer to task_struct of some task.
2616 * @tsk2: pointer to task_struct of some other task.
2617 *
2618 * Description: Return true if @tsk1's mems_allowed intersects the
2619 * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2620 * one of the task's memory usage might impact the memory available
2621 * to the other.
2622 **/
2623
2624int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2625                                   const struct task_struct *tsk2)
2626{
2627        return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2628}
2629
2630#define CPUSET_NODELIST_LEN     (256)
2631
2632/**
2633 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2634 * @task: pointer to task_struct of some task.
2635 *
2636 * Description: Prints @task's name, cpuset name, and cached copy of its
2637 * mems_allowed to the kernel log.  Must hold task_lock(task) to allow
2638 * dereferencing task_cs(task).
2639 */
2640void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2641{
2642         /* Statically allocated to prevent using excess stack. */
2643        static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2644        static DEFINE_SPINLOCK(cpuset_buffer_lock);
2645
2646        struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2647
2648        rcu_read_lock();
2649        spin_lock(&cpuset_buffer_lock);
2650
2651        nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2652                           tsk->mems_allowed);
2653        printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2654               tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2655
2656        spin_unlock(&cpuset_buffer_lock);
2657        rcu_read_unlock();
2658}
2659
2660/*
2661 * Collection of memory_pressure is suppressed unless
2662 * this flag is enabled by writing "1" to the special
2663 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2664 */
2665
2666int cpuset_memory_pressure_enabled __read_mostly;
2667
2668/**
2669 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2670 *
2671 * Keep a running average of the rate of synchronous (direct)
2672 * page reclaim efforts initiated by tasks in each cpuset.
2673 *
2674 * This represents the rate at which some task in the cpuset
2675 * ran low on memory on all nodes it was allowed to use, and
2676 * had to enter the kernels page reclaim code in an effort to
2677 * create more free memory by tossing clean pages or swapping
2678 * or writing dirty pages.
2679 *
2680 * Display to user space in the per-cpuset read-only file
2681 * "memory_pressure".  Value displayed is an integer
2682 * representing the recent rate of entry into the synchronous
2683 * (direct) page reclaim by any task attached to the cpuset.
2684 **/
2685
2686void __cpuset_memory_pressure_bump(void)
2687{
2688        task_lock(current);
2689        fmeter_markevent(&task_cs(current)->fmeter);
2690        task_unlock(current);
2691}
2692
2693#ifdef CONFIG_PROC_PID_CPUSET
2694/*
2695 * proc_cpuset_show()
2696 *  - Print tasks cpuset path into seq_file.
2697 *  - Used for /proc/<pid>/cpuset.
2698 *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2699 *    doesn't really matter if tsk->cpuset changes after we read it,
2700 *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2701 *    anyway.
2702 */
2703int proc_cpuset_show(struct seq_file *m, void *unused_v)
2704{
2705        struct pid *pid;
2706        struct task_struct *tsk;
2707        char *buf;
2708        struct cgroup_subsys_state *css;
2709        int retval;
2710
2711        retval = -ENOMEM;
2712        buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2713        if (!buf)
2714                goto out;
2715
2716        retval = -ESRCH;
2717        pid = m->private;
2718        tsk = get_pid_task(pid, PIDTYPE_PID);
2719        if (!tsk)
2720                goto out_free;
2721
2722        rcu_read_lock();
2723        css = task_css(tsk, cpuset_subsys_id);
2724        retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2725        rcu_read_unlock();
2726        if (retval < 0)
2727                goto out_put_task;
2728        seq_puts(m, buf);
2729        seq_putc(m, '\n');
2730out_put_task:
2731        put_task_struct(tsk);
2732out_free:
2733        kfree(buf);
2734out:
2735        return retval;
2736}
2737#endif /* CONFIG_PROC_PID_CPUSET */
2738
2739/* Display task mems_allowed in /proc/<pid>/status file. */
2740void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2741{
2742        seq_printf(m, "Mems_allowed:\t");
2743        seq_nodemask(m, &task->mems_allowed);
2744        seq_printf(m, "\n");
2745        seq_printf(m, "Mems_allowed_list:\t");
2746        seq_nodemask_list(m, &task->mems_allowed);
2747        seq_printf(m, "\n");
2748}
2749
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.