linux/kernel/cpuset.c
<<
>>
Prefs
   1/*
   2 *  kernel/cpuset.c
   3 *
   4 *  Processor and Memory placement constraints for sets of tasks.
   5 *
   6 *  Copyright (C) 2003 BULL SA.
   7 *  Copyright (C) 2004-2007 Silicon Graphics, Inc.
   8 *  Copyright (C) 2006 Google, Inc
   9 *
  10 *  Portions derived from Patrick Mochel's sysfs code.
  11 *  sysfs is Copyright (c) 2001-3 Patrick Mochel
  12 *
  13 *  2003-10-10 Written by Simon Derr.
  14 *  2003-10-22 Updates by Stephen Hemminger.
  15 *  2004 May-July Rework by Paul Jackson.
  16 *  2006 Rework by Paul Menage to use generic cgroups
  17 *  2008 Rework of the scheduler domains and CPU hotplug handling
  18 *       by Max Krasnyansky
  19 *
  20 *  This file is subject to the terms and conditions of the GNU General Public
  21 *  License.  See the file COPYING in the main directory of the Linux
  22 *  distribution for more details.
  23 */
  24
  25#include <linux/cpu.h>
  26#include <linux/cpumask.h>
  27#include <linux/cpuset.h>
  28#include <linux/err.h>
  29#include <linux/errno.h>
  30#include <linux/file.h>
  31#include <linux/fs.h>
  32#include <linux/init.h>
  33#include <linux/interrupt.h>
  34#include <linux/kernel.h>
  35#include <linux/kmod.h>
  36#include <linux/list.h>
  37#include <linux/mempolicy.h>
  38#include <linux/mm.h>
  39#include <linux/memory.h>
  40#include <linux/export.h>
  41#include <linux/mount.h>
  42#include <linux/namei.h>
  43#include <linux/pagemap.h>
  44#include <linux/proc_fs.h>
  45#include <linux/rcupdate.h>
  46#include <linux/sched.h>
  47#include <linux/seq_file.h>
  48#include <linux/security.h>
  49#include <linux/slab.h>
  50#include <linux/spinlock.h>
  51#include <linux/stat.h>
  52#include <linux/string.h>
  53#include <linux/time.h>
  54#include <linux/backing-dev.h>
  55#include <linux/sort.h>
  56
  57#include <asm/uaccess.h>
  58#include <linux/atomic.h>
  59#include <linux/mutex.h>
  60#include <linux/workqueue.h>
  61#include <linux/cgroup.h>
  62
  63/*
  64 * Tracks how many cpusets are currently defined in system.
  65 * When there is only one cpuset (the root cpuset) we can
  66 * short circuit some hooks.
  67 */
  68int number_of_cpusets __read_mostly;
  69
  70/* Forward declare cgroup structures */
  71struct cgroup_subsys cpuset_subsys;
  72struct cpuset;
  73
  74/* See "Frequency meter" comments, below. */
  75
  76struct fmeter {
  77        int cnt;                /* unprocessed events count */
  78        int val;                /* most recent output value */
  79        time_t time;            /* clock (secs) when val computed */
  80        spinlock_t lock;        /* guards read or write of above */
  81};
  82
  83struct cpuset {
  84        struct cgroup_subsys_state css;
  85
  86        unsigned long flags;            /* "unsigned long" so bitops work */
  87        cpumask_var_t cpus_allowed;     /* CPUs allowed to tasks in cpuset */
  88        nodemask_t mems_allowed;        /* Memory Nodes allowed to tasks */
  89
  90        struct fmeter fmeter;           /* memory_pressure filter */
  91
  92        /*
  93         * Tasks are being attached to this cpuset.  Used to prevent
  94         * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
  95         */
  96        int attach_in_progress;
  97
  98        /* partition number for rebuild_sched_domains() */
  99        int pn;
 100
 101        /* for custom sched domain */
 102        int relax_domain_level;
 103
 104        struct work_struct hotplug_work;
 105};
 106
 107/* Retrieve the cpuset for a cgroup */
 108static inline struct cpuset *cgroup_cs(struct cgroup *cont)
 109{
 110        return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
 111                            struct cpuset, css);
 112}
 113
 114/* Retrieve the cpuset for a task */
 115static inline struct cpuset *task_cs(struct task_struct *task)
 116{
 117        return container_of(task_subsys_state(task, cpuset_subsys_id),
 118                            struct cpuset, css);
 119}
 120
 121static inline struct cpuset *parent_cs(const struct cpuset *cs)
 122{
 123        struct cgroup *pcgrp = cs->css.cgroup->parent;
 124
 125        if (pcgrp)
 126                return cgroup_cs(pcgrp);
 127        return NULL;
 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_cgrp: 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_cgrp, parent_cs)            \
 211        cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup)      \
 212                if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
 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_cgrp: 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_cgrp by calling
 222 * cgroup_rightmost_descendant() to skip subtree.
 223 */
 224#define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs)       \
 225        cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
 226                if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
 227
 228/*
 229 * There are two global mutexes guarding cpuset structures - cpuset_mutex
 230 * and callback_mutex.  The latter may nest inside the former.  We also
 231 * require taking task_lock() when dereferencing a task's cpuset pointer.
 232 * See "The task_lock() exception", at the end of this comment.
 233 *
 234 * A task must hold both mutexes to modify cpusets.  If a task holds
 235 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
 236 * is the only task able to also acquire callback_mutex and be able to
 237 * modify cpusets.  It can perform various checks on the cpuset structure
 238 * first, knowing nothing will change.  It can also allocate memory while
 239 * just holding cpuset_mutex.  While it is performing these checks, various
 240 * callback routines can briefly acquire callback_mutex to query cpusets.
 241 * Once it is ready to make the changes, it takes callback_mutex, blocking
 242 * everyone else.
 243 *
 244 * Calls to the kernel memory allocator can not be made while holding
 245 * callback_mutex, as that would risk double tripping on callback_mutex
 246 * from one of the callbacks into the cpuset code from within
 247 * __alloc_pages().
 248 *
 249 * If a task is only holding callback_mutex, then it has read-only
 250 * access to cpusets.
 251 *
 252 * Now, the task_struct fields mems_allowed and mempolicy may be changed
 253 * by other task, we use alloc_lock in the task_struct fields to protect
 254 * them.
 255 *
 256 * The cpuset_common_file_read() handlers only hold callback_mutex across
 257 * small pieces of code, such as when reading out possibly multi-word
 258 * cpumasks and nodemasks.
 259 *
 260 * Accessing a task's cpuset should be done in accordance with the
 261 * guidelines for accessing subsystem state in kernel/cgroup.c
 262 */
 263
 264static DEFINE_MUTEX(cpuset_mutex);
 265static DEFINE_MUTEX(callback_mutex);
 266
 267/*
 268 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
 269 * buffers.  They are statically allocated to prevent using excess stack
 270 * when calling cpuset_print_task_mems_allowed().
 271 */
 272#define CPUSET_NAME_LEN         (128)
 273#define CPUSET_NODELIST_LEN     (256)
 274static char cpuset_name[CPUSET_NAME_LEN];
 275static char cpuset_nodelist[CPUSET_NODELIST_LEN];
 276static DEFINE_SPINLOCK(cpuset_buffer_lock);
 277
 278/*
 279 * CPU / memory hotplug is handled asynchronously.
 280 */
 281static struct workqueue_struct *cpuset_propagate_hotplug_wq;
 282
 283static void cpuset_hotplug_workfn(struct work_struct *work);
 284static void cpuset_propagate_hotplug_workfn(struct work_struct *work);
 285static void schedule_cpuset_propagate_hotplug(struct cpuset *cs);
 286
 287static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
 288
 289/*
 290 * This is ugly, but preserves the userspace API for existing cpuset
 291 * users. If someone tries to mount the "cpuset" filesystem, we
 292 * silently switch it to mount "cgroup" instead
 293 */
 294static struct dentry *cpuset_mount(struct file_system_type *fs_type,
 295                         int flags, const char *unused_dev_name, void *data)
 296{
 297        struct file_system_type *cgroup_fs = get_fs_type("cgroup");
 298        struct dentry *ret = ERR_PTR(-ENODEV);
 299        if (cgroup_fs) {
 300                char mountopts[] =
 301                        "cpuset,noprefix,"
 302                        "release_agent=/sbin/cpuset_release_agent";
 303                ret = cgroup_fs->mount(cgroup_fs, flags,
 304                                           unused_dev_name, mountopts);
 305                put_filesystem(cgroup_fs);
 306        }
 307        return ret;
 308}
 309
 310static struct file_system_type cpuset_fs_type = {
 311        .name = "cpuset",
 312        .mount = cpuset_mount,
 313};
 314
 315/*
 316 * Return in pmask the portion of a cpusets's cpus_allowed that
 317 * are online.  If none are online, walk up the cpuset hierarchy
 318 * until we find one that does have some online cpus.  If we get
 319 * all the way to the top and still haven't found any online cpus,
 320 * return cpu_online_mask.  Or if passed a NULL cs from an exit'ing
 321 * task, return cpu_online_mask.
 322 *
 323 * One way or another, we guarantee to return some non-empty subset
 324 * of cpu_online_mask.
 325 *
 326 * Call with callback_mutex held.
 327 */
 328
 329static void guarantee_online_cpus(const struct cpuset *cs,
 330                                  struct cpumask *pmask)
 331{
 332        while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
 333                cs = parent_cs(cs);
 334        if (cs)
 335                cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
 336        else
 337                cpumask_copy(pmask, cpu_online_mask);
 338        BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
 339}
 340
 341/*
 342 * Return in *pmask the portion of a cpusets's mems_allowed that
 343 * are online, with memory.  If none are online with memory, walk
 344 * up the cpuset hierarchy until we find one that does have some
 345 * online mems.  If we get all the way to the top and still haven't
 346 * found any online mems, return node_states[N_MEMORY].
 347 *
 348 * One way or another, we guarantee to return some non-empty subset
 349 * of node_states[N_MEMORY].
 350 *
 351 * Call with callback_mutex held.
 352 */
 353
 354static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
 355{
 356        while (cs && !nodes_intersects(cs->mems_allowed,
 357                                        node_states[N_MEMORY]))
 358                cs = parent_cs(cs);
 359        if (cs)
 360                nodes_and(*pmask, cs->mems_allowed,
 361                                        node_states[N_MEMORY]);
 362        else
 363                *pmask = node_states[N_MEMORY];
 364        BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
 365}
 366
 367/*
 368 * update task's spread flag if cpuset's page/slab spread flag is set
 369 *
 370 * Called with callback_mutex/cpuset_mutex held
 371 */
 372static void cpuset_update_task_spread_flag(struct cpuset *cs,
 373                                        struct task_struct *tsk)
 374{
 375        if (is_spread_page(cs))
 376                tsk->flags |= PF_SPREAD_PAGE;
 377        else
 378                tsk->flags &= ~PF_SPREAD_PAGE;
 379        if (is_spread_slab(cs))
 380                tsk->flags |= PF_SPREAD_SLAB;
 381        else
 382                tsk->flags &= ~PF_SPREAD_SLAB;
 383}
 384
 385/*
 386 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
 387 *
 388 * One cpuset is a subset of another if all its allowed CPUs and
 389 * Memory Nodes are a subset of the other, and its exclusive flags
 390 * are only set if the other's are set.  Call holding cpuset_mutex.
 391 */
 392
 393static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
 394{
 395        return  cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
 396                nodes_subset(p->mems_allowed, q->mems_allowed) &&
 397                is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
 398                is_mem_exclusive(p) <= is_mem_exclusive(q);
 399}
 400
 401/**
 402 * alloc_trial_cpuset - allocate a trial cpuset
 403 * @cs: the cpuset that the trial cpuset duplicates
 404 */
 405static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
 406{
 407        struct cpuset *trial;
 408
 409        trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
 410        if (!trial)
 411                return NULL;
 412
 413        if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
 414                kfree(trial);
 415                return NULL;
 416        }
 417        cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
 418
 419        return trial;
 420}
 421
 422/**
 423 * free_trial_cpuset - free the trial cpuset
 424 * @trial: the trial cpuset to be freed
 425 */
 426static void free_trial_cpuset(struct cpuset *trial)
 427{
 428        free_cpumask_var(trial->cpus_allowed);
 429        kfree(trial);
 430}
 431
 432/*
 433 * validate_change() - Used to validate that any proposed cpuset change
 434 *                     follows the structural rules for cpusets.
 435 *
 436 * If we replaced the flag and mask values of the current cpuset
 437 * (cur) with those values in the trial cpuset (trial), would
 438 * our various subset and exclusive rules still be valid?  Presumes
 439 * cpuset_mutex held.
 440 *
 441 * 'cur' is the address of an actual, in-use cpuset.  Operations
 442 * such as list traversal that depend on the actual address of the
 443 * cpuset in the list must use cur below, not trial.
 444 *
 445 * 'trial' is the address of bulk structure copy of cur, with
 446 * perhaps one or more of the fields cpus_allowed, mems_allowed,
 447 * or flags changed to new, trial values.
 448 *
 449 * Return 0 if valid, -errno if not.
 450 */
 451
 452static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
 453{
 454        struct cgroup *cont;
 455        struct cpuset *c, *par;
 456        int ret;
 457
 458        rcu_read_lock();
 459
 460        /* Each of our child cpusets must be a subset of us */
 461        ret = -EBUSY;
 462        cpuset_for_each_child(c, cont, cur)
 463                if (!is_cpuset_subset(c, trial))
 464                        goto out;
 465
 466        /* Remaining checks don't apply to root cpuset */
 467        ret = 0;
 468        if (cur == &top_cpuset)
 469                goto out;
 470
 471        par = parent_cs(cur);
 472
 473        /* We must be a subset of our parent cpuset */
 474        ret = -EACCES;
 475        if (!is_cpuset_subset(trial, par))
 476                goto out;
 477
 478        /*
 479         * If either I or some sibling (!= me) is exclusive, we can't
 480         * overlap
 481         */
 482        ret = -EINVAL;
 483        cpuset_for_each_child(c, cont, par) {
 484                if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
 485                    c != cur &&
 486                    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
 487                        goto out;
 488                if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
 489                    c != cur &&
 490                    nodes_intersects(trial->mems_allowed, c->mems_allowed))
 491                        goto out;
 492        }
 493
 494        /*
 495         * Cpusets with tasks - existing or newly being attached - can't
 496         * have empty cpus_allowed or mems_allowed.
 497         */
 498        ret = -ENOSPC;
 499        if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
 500            (cpumask_empty(trial->cpus_allowed) ||
 501             nodes_empty(trial->mems_allowed)))
 502                goto out;
 503
 504        ret = 0;
 505out:
 506        rcu_read_unlock();
 507        return ret;
 508}
 509
 510#ifdef CONFIG_SMP
 511/*
 512 * Helper routine for generate_sched_domains().
 513 * Do cpusets a, b have overlapping cpus_allowed masks?
 514 */
 515static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
 516{
 517        return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
 518}
 519
 520static void
 521update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
 522{
 523        if (dattr->relax_domain_level < c->relax_domain_level)
 524                dattr->relax_domain_level = c->relax_domain_level;
 525        return;
 526}
 527
 528static void update_domain_attr_tree(struct sched_domain_attr *dattr,
 529                                    struct cpuset *root_cs)
 530{
 531        struct cpuset *cp;
 532        struct cgroup *pos_cgrp;
 533
 534        rcu_read_lock();
 535        cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
 536                /* skip the whole subtree if @cp doesn't have any CPU */
 537                if (cpumask_empty(cp->cpus_allowed)) {
 538                        pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
 539                        continue;
 540                }
 541
 542                if (is_sched_load_balance(cp))
 543                        update_domain_attr(dattr, cp);
 544        }
 545        rcu_read_unlock();
 546}
 547
 548/*
 549 * generate_sched_domains()
 550 *
 551 * This function builds a partial partition of the systems CPUs
 552 * A 'partial partition' is a set of non-overlapping subsets whose
 553 * union is a subset of that set.
 554 * The output of this function needs to be passed to kernel/sched.c
 555 * partition_sched_domains() routine, which will rebuild the scheduler's
 556 * load balancing domains (sched domains) as specified by that partial
 557 * partition.
 558 *
 559 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
 560 * for a background explanation of this.
 561 *
 562 * Does not return errors, on the theory that the callers of this
 563 * routine would rather not worry about failures to rebuild sched
 564 * domains when operating in the severe memory shortage situations
 565 * that could cause allocation failures below.
 566 *
 567 * Must be called with cpuset_mutex held.
 568 *
 569 * The three key local variables below are:
 570 *    q  - a linked-list queue of cpuset pointers, used to implement a
 571 *         top-down scan of all cpusets.  This scan loads a pointer
 572 *         to each cpuset marked is_sched_load_balance into the
 573 *         array 'csa'.  For our purposes, rebuilding the schedulers
 574 *         sched domains, we can ignore !is_sched_load_balance cpusets.
 575 *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
 576 *         that need to be load balanced, for convenient iterative
 577 *         access by the subsequent code that finds the best partition,
 578 *         i.e the set of domains (subsets) of CPUs such that the
 579 *         cpus_allowed of every cpuset marked is_sched_load_balance
 580 *         is a subset of one of these domains, while there are as
 581 *         many such domains as possible, each as small as possible.
 582 * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
 583 *         the kernel/sched.c routine partition_sched_domains() in a
 584 *         convenient format, that can be easily compared to the prior
 585 *         value to determine what partition elements (sched domains)
 586 *         were changed (added or removed.)
 587 *
 588 * Finding the best partition (set of domains):
 589 *      The triple nested loops below over i, j, k scan over the
 590 *      load balanced cpusets (using the array of cpuset pointers in
 591 *      csa[]) looking for pairs of cpusets that have overlapping
 592 *      cpus_allowed, but which don't have the same 'pn' partition
 593 *      number and gives them in the same partition number.  It keeps
 594 *      looping on the 'restart' label until it can no longer find
 595 *      any such pairs.
 596 *
 597 *      The union of the cpus_allowed masks from the set of
 598 *      all cpusets having the same 'pn' value then form the one
 599 *      element of the partition (one sched domain) to be passed to
 600 *      partition_sched_domains().
 601 */
 602static int generate_sched_domains(cpumask_var_t **domains,
 603                        struct sched_domain_attr **attributes)
 604{
 605        struct cpuset *cp;      /* scans q */
 606        struct cpuset **csa;    /* array of all cpuset ptrs */
 607        int csn;                /* how many cpuset ptrs in csa so far */
 608        int i, j, k;            /* indices for partition finding loops */
 609        cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
 610        struct sched_domain_attr *dattr;  /* attributes for custom domains */
 611        int ndoms = 0;          /* number of sched domains in result */
 612        int nslot;              /* next empty doms[] struct cpumask slot */
 613        struct cgroup *pos_cgrp;
 614
 615        doms = NULL;
 616        dattr = NULL;
 617        csa = NULL;
 618
 619        /* Special case for the 99% of systems with one, full, sched domain */
 620        if (is_sched_load_balance(&top_cpuset)) {
 621                ndoms = 1;
 622                doms = alloc_sched_domains(ndoms);
 623                if (!doms)
 624                        goto done;
 625
 626                dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
 627                if (dattr) {
 628                        *dattr = SD_ATTR_INIT;
 629                        update_domain_attr_tree(dattr, &top_cpuset);
 630                }
 631                cpumask_copy(doms[0], top_cpuset.cpus_allowed);
 632
 633                goto done;
 634        }
 635
 636        csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
 637        if (!csa)
 638                goto done;
 639        csn = 0;
 640
 641        rcu_read_lock();
 642        cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
 643                /*
 644                 * Continue traversing beyond @cp iff @cp has some CPUs and
 645                 * isn't load balancing.  The former is obvious.  The
 646                 * latter: All child cpusets contain a subset of the
 647                 * parent's cpus, so just skip them, and then we call
 648                 * update_domain_attr_tree() to calc relax_domain_level of
 649                 * the corresponding sched domain.
 650                 */
 651                if (!cpumask_empty(cp->cpus_allowed) &&
 652                    !is_sched_load_balance(cp))
 653                        continue;
 654
 655                if (is_sched_load_balance(cp))
 656                        csa[csn++] = cp;
 657
 658                /* skip @cp's subtree */
 659                pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
 660        }
 661        rcu_read_unlock();
 662
 663        for (i = 0; i < csn; i++)
 664                csa[i]->pn = i;
 665        ndoms = csn;
 666
 667restart:
 668        /* Find the best partition (set of sched domains) */
 669        for (i = 0; i < csn; i++) {
 670                struct cpuset *a = csa[i];
 671                int apn = a->pn;
 672
 673                for (j = 0; j < csn; j++) {
 674                        struct cpuset *b = csa[j];
 675                        int bpn = b->pn;
 676
 677                        if (apn != bpn && cpusets_overlap(a, b)) {
 678                                for (k = 0; k < csn; k++) {
 679                                        struct cpuset *c = csa[k];
 680
 681                                        if (c->pn == bpn)
 682                                                c->pn = apn;
 683                                }
 684                                ndoms--;        /* one less element */
 685                                goto restart;
 686                        }
 687                }
 688        }
 689
 690        /*
 691         * Now we know how many domains to create.
 692         * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
 693         */
 694        doms = alloc_sched_domains(ndoms);
 695        if (!doms)
 696                goto done;
 697
 698        /*
 699         * The rest of the code, including the scheduler, can deal with
 700         * dattr==NULL case. No need to abort if alloc fails.
 701         */
 702        dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
 703
 704        for (nslot = 0, i = 0; i < csn; i++) {
 705                struct cpuset *a = csa[i];
 706                struct cpumask *dp;
 707                int apn = a->pn;
 708
 709                if (apn < 0) {
 710                        /* Skip completed partitions */
 711                        continue;
 712                }
 713
 714                dp = doms[nslot];
 715
 716                if (nslot == ndoms) {
 717                        static int warnings = 10;
 718                        if (warnings) {
 719                                printk(KERN_WARNING
 720                                 "rebuild_sched_domains confused:"
 721                                  " nslot %d, ndoms %d, csn %d, i %d,"
 722                                  " apn %d\n",
 723                                  nslot, ndoms, csn, i, apn);
 724                                warnings--;
 725                        }
 726                        continue;
 727                }
 728
 729                cpumask_clear(dp);
 730                if (dattr)
 731                        *(dattr + nslot) = SD_ATTR_INIT;
 732                for (j = i; j < csn; j++) {
 733                        struct cpuset *b = csa[j];
 734
 735                        if (apn == b->pn) {
 736                                cpumask_or(dp, dp, b->cpus_allowed);
 737                                if (dattr)
 738                                        update_domain_attr_tree(dattr + nslot, b);
 739
 740                                /* Done with this partition */
 741                                b->pn = -1;
 742                        }
 743                }
 744                nslot++;
 745        }
 746        BUG_ON(nslot != ndoms);
 747
 748done:
 749        kfree(csa);
 750
 751        /*
 752         * Fallback to the default domain if kmalloc() failed.
 753         * See comments in partition_sched_domains().
 754         */
 755        if (doms == NULL)
 756                ndoms = 1;
 757
 758        *domains    = doms;
 759        *attributes = dattr;
 760        return ndoms;
 761}
 762
 763/*
 764 * Rebuild scheduler domains.
 765 *
 766 * If the flag 'sched_load_balance' of any cpuset with non-empty
 767 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
 768 * which has that flag enabled, or if any cpuset with a non-empty
 769 * 'cpus' is removed, then call this routine to rebuild the
 770 * scheduler's dynamic sched domains.
 771 *
 772 * Call with cpuset_mutex held.  Takes get_online_cpus().
 773 */
 774static void rebuild_sched_domains_locked(void)
 775{
 776        struct sched_domain_attr *attr;
 777        cpumask_var_t *doms;
 778        int ndoms;
 779
 780        lockdep_assert_held(&cpuset_mutex);
 781        get_online_cpus();
 782
 783        /* Generate domain masks and attrs */
 784        ndoms = generate_sched_domains(&doms, &attr);
 785
 786        /* Have scheduler rebuild the domains */
 787        partition_sched_domains(ndoms, doms, attr);
 788
 789        put_online_cpus();
 790}
 791#else /* !CONFIG_SMP */
 792static void rebuild_sched_domains_locked(void)
 793{
 794}
 795
 796static int generate_sched_domains(cpumask_var_t **domains,
 797                        struct sched_domain_attr **attributes)
 798{
 799        *domains = NULL;
 800        return 1;
 801}
 802#endif /* CONFIG_SMP */
 803
 804void rebuild_sched_domains(void)
 805{
 806        mutex_lock(&cpuset_mutex);
 807        rebuild_sched_domains_locked();
 808        mutex_unlock(&cpuset_mutex);
 809}
 810
 811/**
 812 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
 813 * @tsk: task to test
 814 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
 815 *
 816 * Call with cpuset_mutex held.  May take callback_mutex during call.
 817 * Called for each task in a cgroup by cgroup_scan_tasks().
 818 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
 819 * words, if its mask is not equal to its cpuset's mask).
 820 */
 821static int cpuset_test_cpumask(struct task_struct *tsk,
 822                               struct cgroup_scanner *scan)
 823{
 824        return !cpumask_equal(&tsk->cpus_allowed,
 825                        (cgroup_cs(scan->cg))->cpus_allowed);
 826}
 827
 828/**
 829 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
 830 * @tsk: task to test
 831 * @scan: struct cgroup_scanner containing the cgroup of the task
 832 *
 833 * Called by cgroup_scan_tasks() for each task in a cgroup whose
 834 * cpus_allowed mask needs to be changed.
 835 *
 836 * We don't need to re-check for the cgroup/cpuset membership, since we're
 837 * holding cpuset_mutex at this point.
 838 */
 839static void cpuset_change_cpumask(struct task_struct *tsk,
 840                                  struct cgroup_scanner *scan)
 841{
 842        set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
 843}
 844
 845/**
 846 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
 847 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
 848 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
 849 *
 850 * Called with cpuset_mutex held
 851 *
 852 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
 853 * calling callback functions for each.
 854 *
 855 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
 856 * if @heap != NULL.
 857 */
 858static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
 859{
 860        struct cgroup_scanner scan;
 861
 862        scan.cg = cs->css.cgroup;
 863        scan.test_task = cpuset_test_cpumask;
 864        scan.process_task = cpuset_change_cpumask;
 865        scan.heap = heap;
 866        cgroup_scan_tasks(&scan);
 867}
 868
 869/**
 870 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
 871 * @cs: the cpuset to consider
 872 * @buf: buffer of cpu numbers written to this cpuset
 873 */
 874static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
 875                          const char *buf)
 876{
 877        struct ptr_heap heap;
 878        int retval;
 879        int is_load_balanced;
 880
 881        /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
 882        if (cs == &top_cpuset)
 883                return -EACCES;
 884
 885        /*
 886         * An empty cpus_allowed is ok only if the cpuset has no tasks.
 887         * Since cpulist_parse() fails on an empty mask, we special case
 888         * that parsing.  The validate_change() call ensures that cpusets
 889         * with tasks have cpus.
 890         */
 891        if (!*buf) {
 892                cpumask_clear(trialcs->cpus_allowed);
 893        } else {
 894                retval = cpulist_parse(buf, trialcs->cpus_allowed);
 895                if (retval < 0)
 896                        return retval;
 897
 898                if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
 899                        return -EINVAL;
 900        }
 901        retval = validate_change(cs, trialcs);
 902        if (retval < 0)
 903                return retval;
 904
 905        /* Nothing to do if the cpus didn't change */
 906        if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
 907                return 0;
 908
 909        retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
 910        if (retval)
 911                return retval;
 912
 913        is_load_balanced = is_sched_load_balance(trialcs);
 914
 915        mutex_lock(&callback_mutex);
 916        cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
 917        mutex_unlock(&callback_mutex);
 918
 919        /*
 920         * Scan tasks in the cpuset, and update the cpumasks of any
 921         * that need an update.
 922         */
 923        update_tasks_cpumask(cs, &heap);
 924
 925        heap_free(&heap);
 926
 927        if (is_load_balanced)
 928                rebuild_sched_domains_locked();
 929        return 0;
 930}
 931
 932/*
 933 * cpuset_migrate_mm
 934 *
 935 *    Migrate memory region from one set of nodes to another.
 936 *
 937 *    Temporarilly set tasks mems_allowed to target nodes of migration,
 938 *    so that the migration code can allocate pages on these nodes.
 939 *
 940 *    Call holding cpuset_mutex, so current's cpuset won't change
 941 *    during this call, as manage_mutex holds off any cpuset_attach()
 942 *    calls.  Therefore we don't need to take task_lock around the
 943 *    call to guarantee_online_mems(), as we know no one is changing
 944 *    our task's cpuset.
 945 *
 946 *    While the mm_struct we are migrating is typically from some
 947 *    other task, the task_struct mems_allowed that we are hacking
 948 *    is for our current task, which must allocate new pages for that
 949 *    migrating memory region.
 950 */
 951
 952static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
 953                                                        const nodemask_t *to)
 954{
 955        struct task_struct *tsk = current;
 956
 957        tsk->mems_allowed = *to;
 958
 959        do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
 960
 961        guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
 962}
 963
 964/*
 965 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
 966 * @tsk: the task to change
 967 * @newmems: new nodes that the task will be set
 968 *
 969 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
 970 * we structure updates as setting all new allowed nodes, then clearing newly
 971 * disallowed ones.
 972 */
 973static void cpuset_change_task_nodemask(struct task_struct *tsk,
 974                                        nodemask_t *newmems)
 975{
 976        bool need_loop;
 977
 978        /*
 979         * Allow tasks that have access to memory reserves because they have
 980         * been OOM killed to get memory anywhere.
 981         */
 982        if (unlikely(test_thread_flag(TIF_MEMDIE)))
 983                return;
 984        if (current->flags & PF_EXITING) /* Let dying task have memory */
 985                return;
 986
 987        task_lock(tsk);
 988        /*
 989         * Determine if a loop is necessary if another thread is doing
 990         * get_mems_allowed().  If at least one node remains unchanged and
 991         * tsk does not have a mempolicy, then an empty nodemask will not be
 992         * possible when mems_allowed is larger than a word.
 993         */
 994        need_loop = task_has_mempolicy(tsk) ||
 995                        !nodes_intersects(*newmems, tsk->mems_allowed);
 996
 997        if (need_loop)
 998                write_seqcount_begin(&tsk->mems_allowed_seq);
 999
1000        nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1001        mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1002
1003        mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1004        tsk->mems_allowed = *newmems;
1005
1006        if (need_loop)
1007                write_seqcount_end(&tsk->mems_allowed_seq);
1008
1009        task_unlock(tsk);
1010}
1011
1012/*
1013 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1014 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1015 * memory_migrate flag is set. Called with cpuset_mutex held.
1016 */
1017static void cpuset_change_nodemask(struct task_struct *p,
1018                                   struct cgroup_scanner *scan)
1019{
1020        struct mm_struct *mm;
1021        struct cpuset *cs;
1022        int migrate;
1023        const nodemask_t *oldmem = scan->data;
1024        static nodemask_t newmems;      /* protected by cpuset_mutex */
1025
1026        cs = cgroup_cs(scan->cg);
1027        guarantee_online_mems(cs, &newmems);
1028
1029        cpuset_change_task_nodemask(p, &newmems);
1030
1031        mm = get_task_mm(p);
1032        if (!mm)
1033                return;
1034
1035        migrate = is_memory_migrate(cs);
1036
1037        mpol_rebind_mm(mm, &cs->mems_allowed);
1038        if (migrate)
1039                cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1040        mmput(mm);
1041}
1042
1043static void *cpuset_being_rebound;
1044
1045/**
1046 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1047 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1048 * @oldmem: old mems_allowed of cpuset cs
1049 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1050 *
1051 * Called with cpuset_mutex held
1052 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1053 * if @heap != NULL.
1054 */
1055static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1056                                 struct ptr_heap *heap)
1057{
1058        struct cgroup_scanner scan;
1059
1060        cpuset_being_rebound = cs;              /* causes mpol_dup() rebind */
1061
1062        scan.cg = cs->css.cgroup;
1063        scan.test_task = NULL;
1064        scan.process_task = cpuset_change_nodemask;
1065        scan.heap = heap;
1066        scan.data = (nodemask_t *)oldmem;
1067
1068        /*
1069         * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1070         * take while holding tasklist_lock.  Forks can happen - the
1071         * mpol_dup() cpuset_being_rebound check will catch such forks,
1072         * and rebind their vma mempolicies too.  Because we still hold
1073         * the global cpuset_mutex, we know that no other rebind effort
1074         * will be contending for the global variable cpuset_being_rebound.
1075         * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1076         * is idempotent.  Also migrate pages in each mm to new nodes.
1077         */
1078        cgroup_scan_tasks(&scan);
1079
1080        /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1081        cpuset_being_rebound = NULL;
1082}
1083
1084/*
1085 * Handle user request to change the 'mems' memory placement
1086 * of a cpuset.  Needs to validate the request, update the
1087 * cpusets mems_allowed, and for each task in the cpuset,
1088 * update mems_allowed and rebind task's mempolicy and any vma
1089 * mempolicies and if the cpuset is marked 'memory_migrate',
1090 * migrate the tasks pages to the new memory.
1091 *
1092 * Call with cpuset_mutex held.  May take callback_mutex during call.
1093 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1094 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1095 * their mempolicies to the cpusets new mems_allowed.
1096 */
1097static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1098                           const char *buf)
1099{
1100        NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1101        int retval;
1102        struct ptr_heap heap;
1103
1104        if (!oldmem)
1105                return -ENOMEM;
1106
1107        /*
1108         * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1109         * it's read-only
1110         */
1111        if (cs == &top_cpuset) {
1112                retval = -EACCES;
1113                goto done;
1114        }
1115
1116        /*
1117         * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1118         * Since nodelist_parse() fails on an empty mask, we special case
1119         * that parsing.  The validate_change() call ensures that cpusets
1120         * with tasks have memory.
1121         */
1122        if (!*buf) {
1123                nodes_clear(trialcs->mems_allowed);
1124        } else {
1125                retval = nodelist_parse(buf, trialcs->mems_allowed);
1126                if (retval < 0)
1127                        goto done;
1128
1129                if (!nodes_subset(trialcs->mems_allowed,
1130                                node_states[N_MEMORY])) {
1131                        retval =  -EINVAL;
1132                        goto done;
1133                }
1134        }
1135        *oldmem = cs->mems_allowed;
1136        if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1137                retval = 0;             /* Too easy - nothing to do */
1138                goto done;
1139        }
1140        retval = validate_change(cs, trialcs);
1141        if (retval < 0)
1142                goto done;
1143
1144        retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1145        if (retval < 0)
1146                goto done;
1147
1148        mutex_lock(&callback_mutex);
1149        cs->mems_allowed = trialcs->mems_allowed;
1150        mutex_unlock(&callback_mutex);
1151
1152        update_tasks_nodemask(cs, oldmem, &heap);
1153
1154        heap_free(&heap);
1155done:
1156        NODEMASK_FREE(oldmem);
1157        return retval;
1158}
1159
1160int current_cpuset_is_being_rebound(void)
1161{
1162        return task_cs(current) == cpuset_being_rebound;
1163}
1164
1165static int update_relax_domain_level(struct cpuset *cs, s64 val)
1166{
1167#ifdef CONFIG_SMP
1168        if (val < -1 || val >= sched_domain_level_max)
1169                return -EINVAL;
1170#endif
1171
1172        if (val != cs->relax_domain_level) {
1173                cs->relax_domain_level = val;
1174                if (!cpumask_empty(cs->cpus_allowed) &&
1175                    is_sched_load_balance(cs))
1176                        rebuild_sched_domains_locked();
1177        }
1178
1179        return 0;
1180}
1181
1182/*
1183 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1184 * @tsk: task to be updated
1185 * @scan: struct cgroup_scanner containing the cgroup of the task
1186 *
1187 * Called by cgroup_scan_tasks() for each task in a cgroup.
1188 *
1189 * We don't need to re-check for the cgroup/cpuset membership, since we're
1190 * holding cpuset_mutex at this point.
1191 */
1192static void cpuset_change_flag(struct task_struct *tsk,
1193                                struct cgroup_scanner *scan)
1194{
1195        cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1196}
1197
1198/*
1199 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1200 * @cs: the cpuset in which each task's spread flags needs to be changed
1201 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1202 *
1203 * Called with cpuset_mutex held
1204 *
1205 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1206 * calling callback functions for each.
1207 *
1208 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1209 * if @heap != NULL.
1210 */
1211static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1212{
1213        struct cgroup_scanner scan;
1214
1215        scan.cg = cs->css.cgroup;
1216        scan.test_task = NULL;
1217        scan.process_task = cpuset_change_flag;
1218        scan.heap = heap;
1219        cgroup_scan_tasks(&scan);
1220}
1221
1222/*
1223 * update_flag - read a 0 or a 1 in a file and update associated flag
1224 * bit:         the bit to update (see cpuset_flagbits_t)
1225 * cs:          the cpuset to update
1226 * turning_on:  whether the flag is being set or cleared
1227 *
1228 * Call with cpuset_mutex held.
1229 */
1230
1231static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1232                       int turning_on)
1233{
1234        struct cpuset *trialcs;
1235        int balance_flag_changed;
1236        int spread_flag_changed;
1237        struct ptr_heap heap;
1238        int err;
1239
1240        trialcs = alloc_trial_cpuset(cs);
1241        if (!trialcs)
1242                return -ENOMEM;
1243
1244        if (turning_on)
1245                set_bit(bit, &trialcs->flags);
1246        else
1247                clear_bit(bit, &trialcs->flags);
1248
1249        err = validate_change(cs, trialcs);
1250        if (err < 0)
1251                goto out;
1252
1253        err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1254        if (err < 0)
1255                goto out;
1256
1257        balance_flag_changed = (is_sched_load_balance(cs) !=
1258                                is_sched_load_balance(trialcs));
1259
1260        spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1261                        || (is_spread_page(cs) != is_spread_page(trialcs)));
1262
1263        mutex_lock(&callback_mutex);
1264        cs->flags = trialcs->flags;
1265        mutex_unlock(&callback_mutex);
1266
1267        if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1268                rebuild_sched_domains_locked();
1269
1270        if (spread_flag_changed)
1271                update_tasks_flags(cs, &heap);
1272        heap_free(&heap);
1273out:
1274        free_trial_cpuset(trialcs);
1275        return err;
1276}
1277
1278/*
1279 * Frequency meter - How fast is some event occurring?
1280 *
1281 * These routines manage a digitally filtered, constant time based,
1282 * event frequency meter.  There are four routines:
1283 *   fmeter_init() - initialize a frequency meter.
1284 *   fmeter_markevent() - called each time the event happens.
1285 *   fmeter_getrate() - returns the recent rate of such events.
1286 *   fmeter_update() - internal routine used to update fmeter.
1287 *
1288 * A common data structure is passed to each of these routines,
1289 * which is used to keep track of the state required to manage the
1290 * frequency meter and its digital filter.
1291 *
1292 * The filter works on the number of events marked per unit time.
1293 * The filter is single-pole low-pass recursive (IIR).  The time unit
1294 * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1295 * simulate 3 decimal digits of precision (multiplied by 1000).
1296 *
1297 * With an FM_COEF of 933, and a time base of 1 second, the filter
1298 * has a half-life of 10 seconds, meaning that if the events quit
1299 * happening, then the rate returned from the fmeter_getrate()
1300 * will be cut in half each 10 seconds, until it converges to zero.
1301 *
1302 * It is not worth doing a real infinitely recursive filter.  If more
1303 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1304 * just compute FM_MAXTICKS ticks worth, by which point the level
1305 * will be stable.
1306 *
1307 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1308 * arithmetic overflow in the fmeter_update() routine.
1309 *
1310 * Given the simple 32 bit integer arithmetic used, this meter works
1311 * best for reporting rates between one per millisecond (msec) and
1312 * one per 32 (approx) seconds.  At constant rates faster than one
1313 * per msec it maxes out at values just under 1,000,000.  At constant
1314 * rates between one per msec, and one per second it will stabilize
1315 * to a value N*1000, where N is the rate of events per second.
1316 * At constant rates between one per second and one per 32 seconds,
1317 * it will be choppy, moving up on the seconds that have an event,
1318 * and then decaying until the next event.  At rates slower than
1319 * about one in 32 seconds, it decays all the way back to zero between
1320 * each event.
1321 */
1322
1323#define FM_COEF 933             /* coefficient for half-life of 10 secs */
1324#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1325#define FM_MAXCNT 1000000       /* limit cnt to avoid overflow */
1326#define FM_SCALE 1000           /* faux fixed point scale */
1327
1328/* Initialize a frequency meter */
1329static void fmeter_init(struct fmeter *fmp)
1330{
1331        fmp->cnt = 0;
1332        fmp->val = 0;
1333        fmp->time = 0;
1334        spin_lock_init(&fmp->lock);
1335}
1336
1337/* Internal meter update - process cnt events and update value */
1338static void fmeter_update(struct fmeter *fmp)
1339{
1340        time_t now = get_seconds();
1341        time_t ticks = now - fmp->time;
1342
1343        if (ticks == 0)
1344                return;
1345
1346        ticks = min(FM_MAXTICKS, ticks);
1347        while (ticks-- > 0)
1348                fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1349        fmp->time = now;
1350
1351        fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1352        fmp->cnt = 0;
1353}
1354
1355/* Process any previous ticks, then bump cnt by one (times scale). */
1356static void fmeter_markevent(struct fmeter *fmp)
1357{
1358        spin_lock(&fmp->lock);
1359        fmeter_update(fmp);
1360        fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1361        spin_unlock(&fmp->lock);
1362}
1363
1364/* Process any previous ticks, then return current value. */
1365static int fmeter_getrate(struct fmeter *fmp)
1366{
1367        int val;
1368
1369        spin_lock(&fmp->lock);
1370        fmeter_update(fmp);
1371        val = fmp->val;
1372        spin_unlock(&fmp->lock);
1373        return val;
1374}
1375
1376/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1377static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1378{
1379        struct cpuset *cs = cgroup_cs(cgrp);
1380        struct task_struct *task;
1381        int ret;
1382
1383        mutex_lock(&cpuset_mutex);
1384
1385        ret = -ENOSPC;
1386        if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1387                goto out_unlock;
1388
1389        cgroup_taskset_for_each(task, cgrp, tset) {
1390                /*
1391                 * Kthreads bound to specific cpus cannot be moved to a new
1392                 * cpuset; we cannot change their cpu affinity and
1393                 * isolating such threads by their set of allowed nodes is
1394                 * unnecessary.  Thus, cpusets are not applicable for such
1395                 * threads.  This prevents checking for success of
1396                 * set_cpus_allowed_ptr() on all attached tasks before
1397                 * cpus_allowed may be changed.
1398                 */
1399                ret = -EINVAL;
1400                if (task->flags & PF_THREAD_BOUND)
1401                        goto out_unlock;
1402                ret = security_task_setscheduler(task);
1403                if (ret)
1404                        goto out_unlock;
1405        }
1406
1407        /*
1408         * Mark attach is in progress.  This makes validate_change() fail
1409         * changes which zero cpus/mems_allowed.
1410         */
1411        cs->attach_in_progress++;
1412        ret = 0;
1413out_unlock:
1414        mutex_unlock(&cpuset_mutex);
1415        return ret;
1416}
1417
1418static void cpuset_cancel_attach(struct cgroup *cgrp,
1419                                 struct cgroup_taskset *tset)
1420{
1421        mutex_lock(&cpuset_mutex);
1422        cgroup_cs(cgrp)->attach_in_progress--;
1423        mutex_unlock(&cpuset_mutex);
1424}
1425
1426/*
1427 * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1428 * but we can't allocate it dynamically there.  Define it global and
1429 * allocate from cpuset_init().
1430 */
1431static cpumask_var_t cpus_attach;
1432
1433static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1434{
1435        /* static bufs protected by cpuset_mutex */
1436        static nodemask_t cpuset_attach_nodemask_from;
1437        static nodemask_t cpuset_attach_nodemask_to;
1438        struct mm_struct *mm;
1439        struct task_struct *task;
1440        struct task_struct *leader = cgroup_taskset_first(tset);
1441        struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1442        struct cpuset *cs = cgroup_cs(cgrp);
1443        struct cpuset *oldcs = cgroup_cs(oldcgrp);
1444
1445        mutex_lock(&cpuset_mutex);
1446
1447        /* prepare for attach */
1448        if (cs == &top_cpuset)
1449                cpumask_copy(cpus_attach, cpu_possible_mask);
1450        else
1451                guarantee_online_cpus(cs, cpus_attach);
1452
1453        guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1454
1455        cgroup_taskset_for_each(task, cgrp, tset) {
1456                /*
1457                 * can_attach beforehand should guarantee that this doesn't
1458                 * fail.  TODO: have a better way to handle failure here
1459                 */
1460                WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1461
1462                cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1463                cpuset_update_task_spread_flag(cs, task);
1464        }
1465
1466        /*
1467         * Change mm, possibly for multiple threads in a threadgroup. This is
1468         * expensive and may sleep.
1469         */
1470        cpuset_attach_nodemask_from = oldcs->mems_allowed;
1471        cpuset_attach_nodemask_to = cs->mems_allowed;
1472        mm = get_task_mm(leader);
1473        if (mm) {
1474                mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1475                if (is_memory_migrate(cs))
1476                        cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1477                                          &cpuset_attach_nodemask_to);
1478                mmput(mm);
1479        }
1480
1481        cs->attach_in_progress--;
1482
1483        /*
1484         * We may have raced with CPU/memory hotunplug.  Trigger hotplug
1485         * propagation if @cs doesn't have any CPU or memory.  It will move
1486         * the newly added tasks to the nearest parent which can execute.
1487         */
1488        if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1489                schedule_cpuset_propagate_hotplug(cs);
1490
1491        mutex_unlock(&cpuset_mutex);
1492}
1493
1494/* The various types of files and directories in a cpuset file system */
1495
1496typedef enum {
1497        FILE_MEMORY_MIGRATE,
1498        FILE_CPULIST,
1499        FILE_MEMLIST,
1500        FILE_CPU_EXCLUSIVE,
1501        FILE_MEM_EXCLUSIVE,
1502        FILE_MEM_HARDWALL,
1503        FILE_SCHED_LOAD_BALANCE,
1504        FILE_SCHED_RELAX_DOMAIN_LEVEL,
1505        FILE_MEMORY_PRESSURE_ENABLED,
1506        FILE_MEMORY_PRESSURE,
1507        FILE_SPREAD_PAGE,
1508        FILE_SPREAD_SLAB,
1509} cpuset_filetype_t;
1510
1511static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1512{
1513        struct cpuset *cs = cgroup_cs(cgrp);
1514        cpuset_filetype_t type = cft->private;
1515        int retval = -ENODEV;
1516
1517        mutex_lock(&cpuset_mutex);
1518        if (!is_cpuset_online(cs))
1519                goto out_unlock;
1520
1521        switch (type) {
1522        case FILE_CPU_EXCLUSIVE:
1523                retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1524                break;
1525        case FILE_MEM_EXCLUSIVE:
1526                retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1527                break;
1528        case FILE_MEM_HARDWALL:
1529                retval = update_flag(CS_MEM_HARDWALL, cs, val);
1530                break;
1531        case FILE_SCHED_LOAD_BALANCE:
1532                retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1533                break;
1534        case FILE_MEMORY_MIGRATE:
1535                retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1536                break;
1537        case FILE_MEMORY_PRESSURE_ENABLED:
1538                cpuset_memory_pressure_enabled = !!val;
1539                break;
1540        case FILE_MEMORY_PRESSURE:
1541                retval = -EACCES;
1542                break;
1543        case FILE_SPREAD_PAGE:
1544                retval = update_flag(CS_SPREAD_PAGE, cs, val);
1545                break;
1546        case FILE_SPREAD_SLAB:
1547                retval = update_flag(CS_SPREAD_SLAB, cs, val);
1548                break;
1549        default:
1550                retval = -EINVAL;
1551                break;
1552        }
1553out_unlock:
1554        mutex_unlock(&cpuset_mutex);
1555        return retval;
1556}
1557
1558static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1559{
1560        struct cpuset *cs = cgroup_cs(cgrp);
1561        cpuset_filetype_t type = cft->private;
1562        int retval = -ENODEV;
1563
1564        mutex_lock(&cpuset_mutex);
1565        if (!is_cpuset_online(cs))
1566                goto out_unlock;
1567
1568        switch (type) {
1569        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1570                retval = update_relax_domain_level(cs, val);
1571                break;
1572        default:
1573                retval = -EINVAL;
1574                break;
1575        }
1576out_unlock:
1577        mutex_unlock(&cpuset_mutex);
1578        return retval;
1579}
1580
1581/*
1582 * Common handling for a write to a "cpus" or "mems" file.
1583 */
1584static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1585                                const char *buf)
1586{
1587        struct cpuset *cs = cgroup_cs(cgrp);
1588        struct cpuset *trialcs;
1589        int retval = -ENODEV;
1590
1591        /*
1592         * CPU or memory hotunplug may leave @cs w/o any execution
1593         * resources, in which case the hotplug code asynchronously updates
1594         * configuration and transfers all tasks to the nearest ancestor
1595         * which can execute.
1596         *
1597         * As writes to "cpus" or "mems" may restore @cs's execution
1598         * resources, wait for the previously scheduled operations before
1599         * proceeding, so that we don't end up keep removing tasks added
1600         * after execution capability is restored.
1601         *
1602         * Flushing cpuset_hotplug_work is enough to synchronize against
1603         * hotplug hanlding; however, cpuset_attach() may schedule
1604         * propagation work directly.  Flush the workqueue too.
1605         */
1606        flush_work(&cpuset_hotplug_work);
1607        flush_workqueue(cpuset_propagate_hotplug_wq);
1608
1609        mutex_lock(&cpuset_mutex);
1610        if (!is_cpuset_online(cs))
1611                goto out_unlock;
1612
1613        trialcs = alloc_trial_cpuset(cs);
1614        if (!trialcs) {
1615                retval = -ENOMEM;
1616                goto out_unlock;
1617        }
1618
1619        switch (cft->private) {
1620        case FILE_CPULIST:
1621                retval = update_cpumask(cs, trialcs, buf);
1622                break;
1623        case FILE_MEMLIST:
1624                retval = update_nodemask(cs, trialcs, buf);
1625                break;
1626        default:
1627                retval = -EINVAL;
1628                break;
1629        }
1630
1631        free_trial_cpuset(trialcs);
1632out_unlock:
1633        mutex_unlock(&cpuset_mutex);
1634        return retval;
1635}
1636
1637/*
1638 * These ascii lists should be read in a single call, by using a user
1639 * buffer large enough to hold the entire map.  If read in smaller
1640 * chunks, there is no guarantee of atomicity.  Since the display format
1641 * used, list of ranges of sequential numbers, is variable length,
1642 * and since these maps can change value dynamically, one could read
1643 * gibberish by doing partial reads while a list was changing.
1644 * A single large read to a buffer that crosses a page boundary is
1645 * ok, because the result being copied to user land is not recomputed
1646 * across a page fault.
1647 */
1648
1649static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1650{
1651        size_t count;
1652
1653        mutex_lock(&callback_mutex);
1654        count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1655        mutex_unlock(&callback_mutex);
1656
1657        return count;
1658}
1659
1660static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1661{
1662        size_t count;
1663
1664        mutex_lock(&callback_mutex);
1665        count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1666        mutex_unlock(&callback_mutex);
1667
1668        return count;
1669}
1670
1671static ssize_t cpuset_common_file_read(struct cgroup *cont,
1672                                       struct cftype *cft,
1673                                       struct file *file,
1674                                       char __user *buf,
1675                                       size_t nbytes, loff_t *ppos)
1676{
1677        struct cpuset *cs = cgroup_cs(cont);
1678        cpuset_filetype_t type = cft->private;
1679        char *page;
1680        ssize_t retval = 0;
1681        char *s;
1682
1683        if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1684                return -ENOMEM;
1685
1686        s = page;
1687
1688        switch (type) {
1689        case FILE_CPULIST:
1690                s += cpuset_sprintf_cpulist(s, cs);
1691                break;
1692        case FILE_MEMLIST:
1693                s += cpuset_sprintf_memlist(s, cs);
1694                break;
1695        default:
1696                retval = -EINVAL;
1697                goto out;
1698        }
1699        *s++ = '\n';
1700
1701        retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1702out:
1703        free_page((unsigned long)page);
1704        return retval;
1705}
1706
1707static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1708{
1709        struct cpuset *cs = cgroup_cs(cont);
1710        cpuset_filetype_t type = cft->private;
1711        switch (type) {
1712        case FILE_CPU_EXCLUSIVE:
1713                return is_cpu_exclusive(cs);
1714        case FILE_MEM_EXCLUSIVE:
1715                return is_mem_exclusive(cs);
1716        case FILE_MEM_HARDWALL:
1717                return is_mem_hardwall(cs);
1718        case FILE_SCHED_LOAD_BALANCE:
1719                return is_sched_load_balance(cs);
1720        case FILE_MEMORY_MIGRATE:
1721                return is_memory_migrate(cs);
1722        case FILE_MEMORY_PRESSURE_ENABLED:
1723                return cpuset_memory_pressure_enabled;
1724        case FILE_MEMORY_PRESSURE:
1725                return fmeter_getrate(&cs->fmeter);
1726        case FILE_SPREAD_PAGE:
1727                return is_spread_page(cs);
1728        case FILE_SPREAD_SLAB:
1729                return is_spread_slab(cs);
1730        default:
1731                BUG();
1732        }
1733
1734        /* Unreachable but makes gcc happy */
1735        return 0;
1736}
1737
1738static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1739{
1740        struct cpuset *cs = cgroup_cs(cont);
1741        cpuset_filetype_t type = cft->private;
1742        switch (type) {
1743        case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1744                return cs->relax_domain_level;
1745        default:
1746                BUG();
1747        }
1748
1749        /* Unrechable but makes gcc happy */
1750        return 0;
1751}
1752
1753
1754/*
1755 * for the common functions, 'private' gives the type of file
1756 */
1757
1758static struct cftype files[] = {
1759        {
1760                .name = "cpus",
1761                .read = cpuset_common_file_read,
1762                .write_string = cpuset_write_resmask,
1763                .max_write_len = (100U + 6 * NR_CPUS),
1764                .private = FILE_CPULIST,
1765        },
1766
1767        {
1768                .name = "mems",
1769                .read = cpuset_common_file_read,
1770                .write_string = cpuset_write_resmask,
1771                .max_write_len = (100U + 6 * MAX_NUMNODES),
1772                .private = FILE_MEMLIST,
1773        },
1774
1775        {
1776                .name = "cpu_exclusive",
1777                .read_u64 = cpuset_read_u64,
1778                .write_u64 = cpuset_write_u64,
1779                .private = FILE_CPU_EXCLUSIVE,
1780        },
1781
1782        {
1783                .name = "mem_exclusive",
1784                .read_u64 = cpuset_read_u64,
1785                .write_u64 = cpuset_write_u64,
1786                .private = FILE_MEM_EXCLUSIVE,
1787        },
1788
1789        {
1790                .name = "mem_hardwall",
1791                .read_u64 = cpuset_read_u64,
1792                .write_u64 = cpuset_write_u64,
1793                .private = FILE_MEM_HARDWALL,
1794        },
1795
1796        {
1797                .name = "sched_load_balance",
1798                .read_u64 = cpuset_read_u64,
1799                .write_u64 = cpuset_write_u64,
1800                .private = FILE_SCHED_LOAD_BALANCE,
1801        },
1802
1803        {
1804                .name = "sched_relax_domain_level",
1805                .read_s64 = cpuset_read_s64,
1806                .write_s64 = cpuset_write_s64,
1807                .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1808        },
1809
1810        {
1811                .name = "memory_migrate",
1812                .read_u64 = cpuset_read_u64,
1813                .write_u64 = cpuset_write_u64,
1814                .private = FILE_MEMORY_MIGRATE,
1815        },
1816
1817        {
1818                .name = "memory_pressure",
1819                .read_u64 = cpuset_read_u64,
1820                .write_u64 = cpuset_write_u64,
1821                .private = FILE_MEMORY_PRESSURE,
1822                .mode = S_IRUGO,
1823        },
1824
1825        {
1826                .name = "memory_spread_page",
1827                .read_u64 = cpuset_read_u64,
1828                .write_u64 = cpuset_write_u64,
1829                .private = FILE_SPREAD_PAGE,
1830        },
1831
1832        {
1833                .name = "memory_spread_slab",
1834                .read_u64 = cpuset_read_u64,
1835                .write_u64 = cpuset_write_u64,
1836                .private = FILE_SPREAD_SLAB,
1837        },
1838
1839        {
1840                .name = "memory_pressure_enabled",
1841                .flags = CFTYPE_ONLY_ON_ROOT,
1842                .read_u64 = cpuset_read_u64,
1843                .write_u64 = cpuset_write_u64,
1844                .private = FILE_MEMORY_PRESSURE_ENABLED,
1845        },
1846
1847        { }     /* terminate */
1848};
1849
1850/*
1851 *      cpuset_css_alloc - allocate a cpuset css
1852 *      cont:   control group that the new cpuset will be part of
1853 */
1854
1855static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
1856{
1857        struct cpuset *cs;
1858
1859        if (!cont->parent)
1860                return &top_cpuset.css;
1861
1862        cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1863        if (!cs)
1864                return ERR_PTR(-ENOMEM);
1865        if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1866                kfree(cs);
1867                return ERR_PTR(-ENOMEM);
1868        }
1869
1870        set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1871        cpumask_clear(cs->cpus_allowed);
1872        nodes_clear(cs->mems_allowed);
1873        fmeter_init(&cs->fmeter);
1874        INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
1875        cs->relax_domain_level = -1;
1876
1877        return &cs->css;
1878}
1879
1880static int cpuset_css_online(struct cgroup *cgrp)
1881{
1882        struct cpuset *cs = cgroup_cs(cgrp);
1883        struct cpuset *parent = parent_cs(cs);
1884        struct cpuset *tmp_cs;
1885        struct cgroup *pos_cg;
1886
1887        if (!parent)
1888                return 0;
1889
1890        mutex_lock(&cpuset_mutex);
1891
1892        set_bit(CS_ONLINE, &cs->flags);
1893        if (is_spread_page(parent))
1894                set_bit(CS_SPREAD_PAGE, &cs->flags);
1895        if (is_spread_slab(parent))
1896                set_bit(CS_SPREAD_SLAB, &cs->flags);
1897
1898        number_of_cpusets++;
1899
1900        if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1901                goto out_unlock;
1902
1903        /*
1904         * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1905         * set.  This flag handling is implemented in cgroup core for
1906         * histrical reasons - the flag may be specified during mount.
1907         *
1908         * Currently, if any sibling cpusets have exclusive cpus or mem, we
1909         * refuse to clone the configuration - thereby refusing the task to
1910         * be entered, and as a result refusing the sys_unshare() or
1911         * clone() which initiated it.  If this becomes a problem for some
1912         * users who wish to allow that scenario, then this could be
1913         * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1914         * (and likewise for mems) to the new cgroup.
1915         */
1916        rcu_read_lock();
1917        cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1918                if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1919                        rcu_read_unlock();
1920                        goto out_unlock;
1921                }
1922        }
1923        rcu_read_unlock();
1924
1925        mutex_lock(&callback_mutex);
1926        cs->mems_allowed = parent->mems_allowed;
1927        cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1928        mutex_unlock(&callback_mutex);
1929out_unlock:
1930        mutex_unlock(&cpuset_mutex);
1931        return 0;
1932}
1933
1934static void cpuset_css_offline(struct cgroup *cgrp)
1935{
1936        struct cpuset *cs = cgroup_cs(cgrp);
1937
1938        mutex_lock(&cpuset_mutex);
1939
1940        if (is_sched_load_balance(cs))
1941                update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1942
1943        number_of_cpusets--;
1944        clear_bit(CS_ONLINE, &cs->flags);
1945
1946        mutex_unlock(&cpuset_mutex);
1947}
1948
1949/*
1950 * If the cpuset being removed has its flag 'sched_load_balance'
1951 * enabled, then simulate turning sched_load_balance off, which
1952 * will call rebuild_sched_domains_locked().
1953 */
1954
1955static void cpuset_css_free(struct cgroup *cont)
1956{
1957        struct cpuset *cs = cgroup_cs(cont);
1958
1959        free_cpumask_var(cs->cpus_allowed);
1960        kfree(cs);
1961}
1962
1963struct cgroup_subsys cpuset_subsys = {
1964        .name = "cpuset",
1965        .css_alloc = cpuset_css_alloc,
1966        .css_online = cpuset_css_online,
1967        .css_offline = cpuset_css_offline,
1968        .css_free = cpuset_css_free,
1969        .can_attach = cpuset_can_attach,
1970        .cancel_attach = cpuset_cancel_attach,
1971        .attach = cpuset_attach,
1972        .subsys_id = cpuset_subsys_id,
1973        .base_cftypes = files,
1974        .early_init = 1,
1975};
1976
1977/**
1978 * cpuset_init - initialize cpusets at system boot
1979 *
1980 * Description: Initialize top_cpuset and the cpuset internal file system,
1981 **/
1982
1983int __init cpuset_init(void)
1984{
1985        int err = 0;
1986
1987        if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1988                BUG();
1989
1990        cpumask_setall(top_cpuset.cpus_allowed);
1991        nodes_setall(top_cpuset.mems_allowed);
1992
1993        fmeter_init(&top_cpuset.fmeter);
1994        set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1995        top_cpuset.relax_domain_level = -1;
1996
1997        err = register_filesystem(&cpuset_fs_type);
1998        if (err < 0)
1999                return err;
2000
2001        if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2002                BUG();
2003
2004        number_of_cpusets = 1;
2005        return 0;
2006}
2007
2008/**
2009 * cpuset_do_move_task - move a given task to another cpuset
2010 * @tsk: pointer to task_struct the task to move
2011 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
2012 *
2013 * Called by cgroup_scan_tasks() for each task in a cgroup.
2014 * Return nonzero to stop the walk through the tasks.
2015 */
2016static void cpuset_do_move_task(struct task_struct *tsk,
2017                                struct cgroup_scanner *scan)
2018{
2019        struct cgroup *new_cgroup = scan->data;
2020
2021        cgroup_lock();
2022        cgroup_attach_task(new_cgroup, tsk);
2023        cgroup_unlock();
2024}
2025
2026/**
2027 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
2028 * @from: cpuset in which the tasks currently reside
2029 * @to: cpuset to which the tasks will be moved
2030 *
2031 * Called with cpuset_mutex held
2032 * callback_mutex must not be held, as cpuset_attach() will take it.
2033 *
2034 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
2035 * calling callback functions for each.
2036 */
2037static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
2038{
2039        struct cgroup_scanner scan;
2040
2041        scan.cg = from->css.cgroup;
2042        scan.test_task = NULL; /* select all tasks in cgroup */
2043        scan.process_task = cpuset_do_move_task;
2044        scan.heap = NULL;
2045        scan.data = to->css.cgroup;
2046
2047        if (cgroup_scan_tasks(&scan))
2048                printk(KERN_ERR "move_member_tasks_to_cpuset: "
2049                                "cgroup_scan_tasks failed\n");
2050}
2051
2052/*
2053 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2054 * or memory nodes, we need to walk over the cpuset hierarchy,
2055 * removing that CPU or node from all cpusets.  If this removes the
2056 * last CPU or node from a cpuset, then move the tasks in the empty
2057 * cpuset to its next-highest non-empty parent.
2058 */
2059static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2060{
2061        struct cpuset *parent;
2062
2063        /*
2064         * Find its next-highest non-empty parent, (top cpuset
2065         * has online cpus, so can't be empty).
2066         */
2067        parent = parent_cs(cs);
2068        while (cpumask_empty(parent->cpus_allowed) ||
2069                        nodes_empty(parent->mems_allowed))
2070                parent = parent_cs(parent);
2071
2072        move_member_tasks_to_cpuset(cs, parent);
2073}
2074
2075/**
2076 * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
2077 * @cs: cpuset in interest
2078 *
2079 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2080 * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2081 * all its tasks are moved to the nearest ancestor with both resources.
2082 */
2083static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
2084{
2085        static cpumask_t off_cpus;
2086        static nodemask_t off_mems, tmp_mems;
2087        struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
2088        bool is_empty;
2089
2090        mutex_lock(&cpuset_mutex);
2091
2092        cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2093        nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2094
2095        /* remove offline cpus from @cs */
2096        if (!cpumask_empty(&off_cpus)) {
2097                mutex_lock(&callback_mutex);
2098                cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2099                mutex_unlock(&callback_mutex);
2100                update_tasks_cpumask(cs, NULL);
2101        }
2102
2103        /* remove offline mems from @cs */
2104        if (!nodes_empty(off_mems)) {
2105                tmp_mems = cs->mems_allowed;
2106                mutex_lock(&callback_mutex);
2107                nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2108                mutex_unlock(&callback_mutex);
2109                update_tasks_nodemask(cs, &tmp_mems, NULL);
2110        }
2111
2112        is_empty = cpumask_empty(cs->cpus_allowed) ||
2113                nodes_empty(cs->mems_allowed);
2114
2115        mutex_unlock(&cpuset_mutex);
2116
2117        /*
2118         * If @cs became empty, move tasks to the nearest ancestor with
2119         * execution resources.  This is full cgroup operation which will
2120         * also call back into cpuset.  Should be done outside any lock.
2121         */
2122        if (is_empty)
2123                remove_tasks_in_empty_cpuset(cs);
2124
2125        /* the following may free @cs, should be the last operation */
2126        css_put(&cs->css);
2127}
2128
2129/**
2130 * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
2131 * @cs: cpuset of interest
2132 *
2133 * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
2134 * memory masks according to top_cpuset.
2135 */
2136static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
2137{
2138        /*
2139         * Pin @cs.  The refcnt will be released when the work item
2140         * finishes executing.
2141         */
2142        if (!css_tryget(&cs->css))
2143                return;
2144
2145        /*
2146         * Queue @cs->hotplug_work.  If already pending, lose the css ref.
2147         * cpuset_propagate_hotplug_wq is ordered and propagation will
2148         * happen in the order this function is called.
2149         */
2150        if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
2151                css_put(&cs->css);
2152}
2153
2154/**
2155 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2156 *
2157 * This function is called after either CPU or memory configuration has
2158 * changed and updates cpuset accordingly.  The top_cpuset is always
2159 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2160 * order to make cpusets transparent (of no affect) on systems that are
2161 * actively using CPU hotplug but making no active use of cpusets.
2162 *
2163 * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2164 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2165 * descendants.
2166 *
2167 * Note that CPU offlining during suspend is ignored.  We don't modify
2168 * cpusets across suspend/resume cycles at all.
2169 */
2170static void cpuset_hotplug_workfn(struct work_struct *work)
2171{
2172        static cpumask_t new_cpus, tmp_cpus;
2173        static nodemask_t new_mems, tmp_mems;
2174        bool cpus_updated, mems_updated;
2175        bool cpus_offlined, mems_offlined;
2176
2177        mutex_lock(&cpuset_mutex);
2178
2179        /* fetch the available cpus/mems and find out which changed how */
2180        cpumask_copy(&new_cpus, cpu_active_mask);
2181        new_mems = node_states[N_MEMORY];
2182
2183        cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2184        cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2185                                       &new_cpus);
2186
2187        mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2188        nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2189        mems_offlined = !nodes_empty(tmp_mems);
2190
2191        /* synchronize cpus_allowed to cpu_active_mask */
2192        if (cpus_updated) {
2193                mutex_lock(&callback_mutex);
2194                cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2195                mutex_unlock(&callback_mutex);
2196                /* we don't mess with cpumasks of tasks in top_cpuset */
2197        }
2198
2199        /* synchronize mems_allowed to N_MEMORY */
2200        if (mems_updated) {
2201                tmp_mems = top_cpuset.mems_allowed;
2202                mutex_lock(&callback_mutex);
2203                top_cpuset.mems_allowed = new_mems;
2204                mutex_unlock(&callback_mutex);
2205                update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
2206        }
2207
2208        /* if cpus or mems went down, we need to propagate to descendants */
2209        if (cpus_offlined || mems_offlined) {
2210                struct cpuset *cs;
2211                struct cgroup *pos_cgrp;
2212
2213                rcu_read_lock();
2214                cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset)
2215                        schedule_cpuset_propagate_hotplug(cs);
2216                rcu_read_unlock();
2217        }
2218
2219        mutex_unlock(&cpuset_mutex);
2220
2221        /* wait for propagations to finish */
2222        flush_workqueue(cpuset_propagate_hotplug_wq);
2223
2224        /* rebuild sched domains if cpus_allowed has changed */
2225        if (cpus_updated) {
2226                struct sched_domain_attr *attr;
2227                cpumask_var_t *doms;
2228                int ndoms;
2229
2230                mutex_lock(&cpuset_mutex);
2231                ndoms = generate_sched_domains(&doms, &attr);
2232                mutex_unlock(&cpuset_mutex);
2233
2234                partition_sched_domains(ndoms, doms, attr);
2235        }
2236}
2237
2238void cpuset_update_active_cpus(bool cpu_online)
2239{
2240        /*
2241         * We're inside cpu hotplug critical region which usually nests
2242         * inside cgroup synchronization.  Bounce actual hotplug processing
2243         * to a work item to avoid reverse locking order.
2244         *
2245         * We still need to do partition_sched_domains() synchronously;
2246         * otherwise, the scheduler will get confused and put tasks to the
2247         * dead CPU.  Fall back to the default single domain.
2248         * cpuset_hotplug_workfn() will rebuild it as necessary.
2249         */
2250        partition_sched_domains(1, NULL, NULL);
2251        schedule_work(&cpuset_hotplug_work);
2252}
2253
2254#ifdef CONFIG_MEMORY_HOTPLUG
2255/*
2256 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2257 * Call this routine anytime after node_states[N_MEMORY] changes.
2258 * See cpuset_update_active_cpus() for CPU hotplug handling.
2259 */
2260static int cpuset_track_online_nodes(struct notifier_block *self,
2261                                unsigned long action, void *arg)
2262{
2263        schedule_work(&cpuset_hotplug_work);
2264        return NOTIFY_OK;
2265}
2266#endif
2267
2268/**
2269 * cpuset_init_smp - initialize cpus_allowed
2270 *
2271 * Description: Finish top cpuset after cpu, node maps are initialized
2272 **/
2273
2274void __init cpuset_init_smp(void)
2275{
2276        cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2277        top_cpuset.mems_allowed = node_states[N_MEMORY];
2278
2279        hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2280
2281        cpuset_propagate_hotplug_wq =
2282                alloc_ordered_workqueue("cpuset_hotplug", 0);
2283        BUG_ON(!cpuset_propagate_hotplug_wq);
2284}
2285
2286/**
2287 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2288 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2289 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2290 *
2291 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2292 * attached to the specified @tsk.  Guaranteed to return some non-empty
2293 * subset of cpu_online_mask, even if this means going outside the
2294 * tasks cpuset.
2295 **/
2296
2297void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2298{
2299        mutex_lock(&callback_mutex);
2300        task_lock(tsk);
2301        guarantee_online_cpus(task_cs(tsk), pmask);
2302        task_unlock(tsk);
2303        mutex_unlock(&callback_mutex);
2304}
2305
2306void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2307{
2308        const struct cpuset *cs;
2309
2310        rcu_read_lock();
2311        cs = task_cs(tsk);
2312        if (cs)
2313                do_set_cpus_allowed(tsk, cs->cpus_allowed);
2314        rcu_read_unlock();
2315
2316        /*
2317         * We own tsk->cpus_allowed, nobody can change it under us.
2318         *
2319         * But we used cs && cs->cpus_allowed lockless and thus can
2320         * race with cgroup_attach_task() or update_cpumask() and get
2321         * the wrong tsk->cpus_allowed. However, both cases imply the
2322         * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2323         * which takes task_rq_lock().
2324         *
2325         * If we are called after it dropped the lock we must see all
2326         * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2327         * set any mask even if it is not right from task_cs() pov,
2328         * the pending set_cpus_allowed_ptr() will fix things.
2329         *
2330         * select_fallback_rq() will fix things ups and set cpu_possible_mask
2331         * if required.
2332         */
2333}
2334
2335void cpuset_init_current_mems_allowed(void)
2336{
2337        nodes_setall(current->mems_allowed);
2338}
2339
2340/**
2341 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2342 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2343 *
2344 * Description: Returns the nodemask_t mems_allowed of the cpuset
2345 * attached to the specified @tsk.  Guaranteed to return some non-empty
2346 * subset of node_states[N_MEMORY], even if this means going outside the
2347 * tasks cpuset.
2348 **/
2349
2350nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2351{
2352        nodemask_t mask;
2353
2354        mutex_lock(&callback_mutex);
2355        task_lock(tsk);
2356        guarantee_online_mems(task_cs(tsk), &mask);
2357        task_unlock(tsk);
2358        mutex_unlock(&callback_mutex);
2359
2360        return mask;
2361}
2362
2363/**
2364 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2365 * @nodemask: the nodemask to be checked
2366 *
2367 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2368 */
2369int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2370{
2371        return nodes_intersects(*nodemask, current->mems_allowed);
2372}
2373
2374/*
2375 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2376 * mem_hardwall ancestor to the specified cpuset.  Call holding
2377 * callback_mutex.  If no ancestor is mem_exclusive or mem_hardwall
2378 * (an unusual configuration), then returns the root cpuset.
2379 */
2380static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2381{
2382        while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2383                cs = parent_cs(cs);
2384        return cs;
2385}
2386
2387/**
2388 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2389 * @node: is this an allowed node?
2390 * @gfp_mask: memory allocation flags
2391 *
2392 * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2393 * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2394 * yes.  If it's not a __GFP_HARDWALL request and this node is in the nearest
2395 * hardwalled cpuset ancestor to this task's cpuset, yes.  If the task has been
2396 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2397 * flag, yes.
2398 * Otherwise, no.
2399 *
2400 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2401 * cpuset_node_allowed_hardwall().  Otherwise, cpuset_node_allowed_softwall()
2402 * might sleep, and might allow a node from an enclosing cpuset.
2403 *
2404 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2405 * cpusets, and never sleeps.
2406 *
2407 * The __GFP_THISNODE placement logic is really handled elsewhere,
2408 * by forcibly using a zonelist starting at a specified node, and by
2409 * (in get_page_from_freelist()) refusing to consider the zones for
2410 * any node on the zonelist except the first.  By the time any such
2411 * calls get to this routine, we should just shut up and say 'yes'.
2412 *
2413 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2414 * and do not allow allocations outside the current tasks cpuset
2415 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2416 * GFP_KERNEL allocations are not so marked, so can escape to the
2417 * nearest enclosing hardwalled ancestor cpuset.
2418 *
2419 * Scanning up parent cpusets requires callback_mutex.  The
2420 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2421 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2422 * current tasks mems_allowed came up empty on the first pass over
2423 * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2424 * cpuset are short of memory, might require taking the callback_mutex
2425 * mutex.
2426 *
2427 * The first call here from mm/page_alloc:get_page_from_freelist()
2428 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2429 * so no allocation on a node outside the cpuset is allowed (unless
2430 * in interrupt, of course).
2431 *
2432 * The second pass through get_page_from_freelist() doesn't even call
2433 * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2434 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2435 * in alloc_flags.  That logic and the checks below have the combined
2436 * affect that:
2437 *      in_interrupt - any node ok (current task context irrelevant)
2438 *      GFP_ATOMIC   - any node ok
2439 *      TIF_MEMDIE   - any node ok
2440 *      GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2441 *      GFP_USER     - only nodes in current tasks mems allowed ok.
2442 *
2443 * Rule:
2444 *    Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2445 *    pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2446 *    the code that might scan up ancestor cpusets and sleep.
2447 */
2448int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2449{
2450        const struct cpuset *cs;        /* current cpuset ancestors */
2451        int allowed;                    /* is allocation in zone z allowed? */
2452
2453        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2454                return 1;
2455        might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2456        if (node_isset(node, current->mems_allowed))
2457                return 1;
2458        /*
2459         * Allow tasks that have access to memory reserves because they have
2460         * been OOM killed to get memory anywhere.
2461         */
2462        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2463                return 1;
2464        if (gfp_mask & __GFP_HARDWALL)  /* If hardwall request, stop here */
2465                return 0;
2466
2467        if (current->flags & PF_EXITING) /* Let dying task have memory */
2468                return 1;
2469
2470        /* Not hardwall and node outside mems_allowed: scan up cpusets */
2471        mutex_lock(&callback_mutex);
2472
2473        task_lock(current);
2474        cs = nearest_hardwall_ancestor(task_cs(current));
2475        task_unlock(current);
2476
2477        allowed = node_isset(node, cs->mems_allowed);
2478        mutex_unlock(&callback_mutex);
2479        return allowed;
2480}
2481
2482/*
2483 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2484 * @node: is this an allowed node?
2485 * @gfp_mask: memory allocation flags
2486 *
2487 * If we're in interrupt, yes, we can always allocate.  If __GFP_THISNODE is
2488 * set, yes, we can always allocate.  If node is in our task's mems_allowed,
2489 * yes.  If the task has been OOM killed and has access to memory reserves as
2490 * specified by the TIF_MEMDIE flag, yes.
2491 * Otherwise, no.
2492 *
2493 * The __GFP_THISNODE placement logic is really handled elsewhere,
2494 * by forcibly using a zonelist starting at a specified node, and by
2495 * (in get_page_from_freelist()) refusing to consider the zones for
2496 * any node on the zonelist except the first.  By the time any such
2497 * calls get to this routine, we should just shut up and say 'yes'.
2498 *
2499 * Unlike the cpuset_node_allowed_softwall() variant, above,
2500 * this variant requires that the node be in the current task's
2501 * mems_allowed or that we're in interrupt.  It does not scan up the
2502 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2503 * It never sleeps.
2504 */
2505int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2506{
2507        if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2508                return 1;
2509        if (node_isset(node, current->mems_allowed))
2510                return 1;
2511        /*
2512         * Allow tasks that have access to memory reserves because they have
2513         * been OOM killed to get memory anywhere.
2514         */
2515        if (unlikely(test_thread_flag(TIF_MEMDIE)))
2516                return 1;
2517        return 0;
2518}
2519
2520/**
2521 * cpuset_mem_spread_node() - On which node to begin search for a file page
2522 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2523 *
2524 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2525 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2526 * and if the memory allocation used cpuset_mem_spread_node()
2527 * to determine on which node to start looking, as it will for
2528 * certain page cache or slab cache pages such as used for file
2529 * system buffers and inode caches, then instead of starting on the
2530 * local node to look for a free page, rather spread the starting
2531 * node around the tasks mems_allowed nodes.
2532 *
2533 * We don't have to worry about the returned node being offline
2534 * because "it can't happen", and even if it did, it would be ok.
2535 *
2536 * The routines calling guarantee_online_mems() are careful to
2537 * only set nodes in task->mems_allowed that are online.  So it
2538 * should not be possible for the following code to return an
2539 * offline node.  But if it did, that would be ok, as this routine
2540 * is not returning the node where the allocation must be, only
2541 * the node where the search should start.  The zonelist passed to
2542 * __alloc_pages() will include all nodes.  If the slab allocator
2543 * is passed an offline node, it will fall back to the local node.
2544 * See kmem_cache_alloc_node().
2545 */
2546
2547static int cpuset_spread_node(int *rotor)
2548{
2549        int node;
2550
2551        node = next_node(*rotor, current->mems_allowed);
2552        if (node == MAX_NUMNODES)
2553                node = first_node(current->mems_allowed);
2554        *rotor = node;
2555        return node;
2556}
2557
2558int cpuset_mem_spread_node(void)
2559{
2560        if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2561                current->cpuset_mem_spread_rotor =
2562                        node_random(&current->mems_allowed);
2563
2564        return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2565}
2566
2567int cpuset_slab_spread_node(void)
2568{
2569        if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2570                current->cpuset_slab_spread_rotor =
2571                        node_random(&current->mems_allowed);
2572
2573        return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2574}
2575
2576EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2577
2578/**
2579 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2580 * @tsk1: pointer to task_struct of some task.
2581 * @tsk2: pointer to task_struct of some other task.
2582 *
2583 * Description: Return true if @tsk1's mems_allowed intersects the
2584 * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2585 * one of the task's memory usage might impact the memory available
2586 * to the other.
2587 **/
2588
2589int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2590                                   const struct task_struct *tsk2)
2591{
2592        return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2593}
2594
2595/**
2596 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2597 * @task: pointer to task_struct of some task.
2598 *
2599 * Description: Prints @task's name, cpuset name, and cached copy of its
2600 * mems_allowed to the kernel log.  Must hold task_lock(task) to allow
2601 * dereferencing task_cs(task).
2602 */
2603void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2604{
2605        struct dentry *dentry;
2606
2607        dentry = task_cs(tsk)->css.cgroup->dentry;
2608        spin_lock(&cpuset_buffer_lock);
2609
2610        if (!dentry) {
2611                strcpy(cpuset_name, "/");
2612        } else {
2613                spin_lock(&dentry->d_lock);
2614                strlcpy(cpuset_name, (const char *)dentry->d_name.name,
2615                        CPUSET_NAME_LEN);
2616                spin_unlock(&dentry->d_lock);
2617        }
2618
2619        nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2620                           tsk->mems_allowed);
2621        printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2622               tsk->comm, cpuset_name, cpuset_nodelist);
2623        spin_unlock(&cpuset_buffer_lock);
2624}
2625
2626/*
2627 * Collection of memory_pressure is suppressed unless
2628 * this flag is enabled by writing "1" to the special
2629 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2630 */
2631
2632int cpuset_memory_pressure_enabled __read_mostly;
2633
2634/**
2635 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2636 *
2637 * Keep a running average of the rate of synchronous (direct)
2638 * page reclaim efforts initiated by tasks in each cpuset.
2639 *
2640 * This represents the rate at which some task in the cpuset
2641 * ran low on memory on all nodes it was allowed to use, and
2642 * had to enter the kernels page reclaim code in an effort to
2643 * create more free memory by tossing clean pages or swapping
2644 * or writing dirty pages.
2645 *
2646 * Display to user space in the per-cpuset read-only file
2647 * "memory_pressure".  Value displayed is an integer
2648 * representing the recent rate of entry into the synchronous
2649 * (direct) page reclaim by any task attached to the cpuset.
2650 **/
2651
2652void __cpuset_memory_pressure_bump(void)
2653{
2654        task_lock(current);
2655        fmeter_markevent(&task_cs(current)->fmeter);
2656        task_unlock(current);
2657}
2658
2659#ifdef CONFIG_PROC_PID_CPUSET
2660/*
2661 * proc_cpuset_show()
2662 *  - Print tasks cpuset path into seq_file.
2663 *  - Used for /proc/<pid>/cpuset.
2664 *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2665 *    doesn't really matter if tsk->cpuset changes after we read it,
2666 *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2667 *    anyway.
2668 */
2669static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2670{
2671        struct pid *pid;
2672        struct task_struct *tsk;
2673        char *buf;
2674        struct cgroup_subsys_state *css;
2675        int retval;
2676
2677        retval = -ENOMEM;
2678        buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2679        if (!buf)
2680                goto out;
2681
2682        retval = -ESRCH;
2683        pid = m->private;
2684        tsk = get_pid_task(pid, PIDTYPE_PID);
2685        if (!tsk)
2686                goto out_free;
2687
2688        rcu_read_lock();
2689        css = task_subsys_state(tsk, cpuset_subsys_id);
2690        retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2691        rcu_read_unlock();
2692        if (retval < 0)
2693                goto out_put_task;
2694        seq_puts(m, buf);
2695        seq_putc(m, '\n');
2696out_put_task:
2697        put_task_struct(tsk);
2698out_free:
2699        kfree(buf);
2700out:
2701        return retval;
2702}
2703
2704static int cpuset_open(struct inode *inode, struct file *file)
2705{
2706        struct pid *pid = PROC_I(inode)->pid;
2707        return single_open(file, proc_cpuset_show, pid);
2708}
2709
2710const struct file_operations proc_cpuset_operations = {
2711        .open           = cpuset_open,
2712        .read           = seq_read,
2713        .llseek         = seq_lseek,
2714        .release        = single_release,
2715};
2716#endif /* CONFIG_PROC_PID_CPUSET */
2717
2718/* Display task mems_allowed in /proc/<pid>/status file. */
2719void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2720{
2721        seq_printf(m, "Mems_allowed:\t");
2722        seq_nodemask(m, &task->mems_allowed);
2723        seq_printf(m, "\n");
2724        seq_printf(m, "Mems_allowed_list:\t");
2725        seq_nodemask_list(m, &task->mems_allowed);
2726        seq_printf(m, "\n");
2727}
2728
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.