Root/kernel/cpuset.c

1/*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
9 *
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
12 *
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
18 * by Max Krasnyansky
19 *
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
23 */
24
25#include <linux/cpu.h>
26#include <linux/cpumask.h>
27#include <linux/cpuset.h>
28#include <linux/err.h>
29#include <linux/errno.h>
30#include <linux/file.h>
31#include <linux/fs.h>
32#include <linux/init.h>
33#include <linux/interrupt.h>
34#include <linux/kernel.h>
35#include <linux/kmod.h>
36#include <linux/list.h>
37#include <linux/mempolicy.h>
38#include <linux/mm.h>
39#include <linux/memory.h>
40#include <linux/module.h>
41#include <linux/mount.h>
42#include <linux/namei.h>
43#include <linux/pagemap.h>
44#include <linux/proc_fs.h>
45#include <linux/rcupdate.h>
46#include <linux/sched.h>
47#include <linux/seq_file.h>
48#include <linux/security.h>
49#include <linux/slab.h>
50#include <linux/spinlock.h>
51#include <linux/stat.h>
52#include <linux/string.h>
53#include <linux/time.h>
54#include <linux/backing-dev.h>
55#include <linux/sort.h>
56
57#include <asm/uaccess.h>
58#include <asm/atomic.h>
59#include <linux/mutex.h>
60#include <linux/workqueue.h>
61#include <linux/cgroup.h>
62
63/*
64 * Workqueue for cpuset related tasks.
65 *
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
68 */
69static struct workqueue_struct *cpuset_wq;
70
71/*
72 * Tracks how many cpusets are currently defined in system.
73 * When there is only one cpuset (the root cpuset) we can
74 * short circuit some hooks.
75 */
76int number_of_cpusets __read_mostly;
77
78/* Forward declare cgroup structures */
79struct cgroup_subsys cpuset_subsys;
80struct cpuset;
81
82/* See "Frequency meter" comments, below. */
83
84struct fmeter {
85    int cnt; /* unprocessed events count */
86    int val; /* most recent output value */
87    time_t time; /* clock (secs) when val computed */
88    spinlock_t lock; /* guards read or write of above */
89};
90
91struct cpuset {
92    struct cgroup_subsys_state css;
93
94    unsigned long flags; /* "unsigned long" so bitops work */
95    cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
96    nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
97
98    struct cpuset *parent; /* my parent */
99
100    struct fmeter fmeter; /* memory_pressure filter */
101
102    /* partition number for rebuild_sched_domains() */
103    int pn;
104
105    /* for custom sched domain */
106    int relax_domain_level;
107
108    /* used for walking a cpuset hierarchy */
109    struct list_head stack_list;
110};
111
112/* Retrieve the cpuset for a cgroup */
113static inline struct cpuset *cgroup_cs(struct cgroup *cont)
114{
115    return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
116                struct cpuset, css);
117}
118
119/* Retrieve the cpuset for a task */
120static inline struct cpuset *task_cs(struct task_struct *task)
121{
122    return container_of(task_subsys_state(task, cpuset_subsys_id),
123                struct cpuset, css);
124}
125
126/* bits in struct cpuset flags field */
127typedef enum {
128    CS_CPU_EXCLUSIVE,
129    CS_MEM_EXCLUSIVE,
130    CS_MEM_HARDWALL,
131    CS_MEMORY_MIGRATE,
132    CS_SCHED_LOAD_BALANCE,
133    CS_SPREAD_PAGE,
134    CS_SPREAD_SLAB,
135} cpuset_flagbits_t;
136
137/* convenient tests for these bits */
138static inline int is_cpu_exclusive(const struct cpuset *cs)
139{
140    return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
141}
142
143static inline int is_mem_exclusive(const struct cpuset *cs)
144{
145    return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
146}
147
148static inline int is_mem_hardwall(const struct cpuset *cs)
149{
150    return test_bit(CS_MEM_HARDWALL, &cs->flags);
151}
152
153static inline int is_sched_load_balance(const struct cpuset *cs)
154{
155    return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
156}
157
158static inline int is_memory_migrate(const struct cpuset *cs)
159{
160    return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
161}
162
163static inline int is_spread_page(const struct cpuset *cs)
164{
165    return test_bit(CS_SPREAD_PAGE, &cs->flags);
166}
167
168static inline int is_spread_slab(const struct cpuset *cs)
169{
170    return test_bit(CS_SPREAD_SLAB, &cs->flags);
171}
172
173static struct cpuset top_cpuset = {
174    .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
175};
176
177/*
178 * There are two global mutexes guarding cpuset structures. The first
179 * is the main control groups cgroup_mutex, accessed via
180 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
181 * callback_mutex, below. They can nest. It is ok to first take
182 * cgroup_mutex, then nest callback_mutex. We also require taking
183 * task_lock() when dereferencing a task's cpuset pointer. See "The
184 * task_lock() exception", at the end of this comment.
185 *
186 * A task must hold both mutexes to modify cpusets. If a task
187 * holds cgroup_mutex, then it blocks others wanting that mutex,
188 * ensuring that it is the only task able to also acquire callback_mutex
189 * and be able to modify cpusets. It can perform various checks on
190 * the cpuset structure first, knowing nothing will change. It can
191 * also allocate memory while just holding cgroup_mutex. While it is
192 * performing these checks, various callback routines can briefly
193 * acquire callback_mutex to query cpusets. Once it is ready to make
194 * the changes, it takes callback_mutex, blocking everyone else.
195 *
196 * Calls to the kernel memory allocator can not be made while holding
197 * callback_mutex, as that would risk double tripping on callback_mutex
198 * from one of the callbacks into the cpuset code from within
199 * __alloc_pages().
200 *
201 * If a task is only holding callback_mutex, then it has read-only
202 * access to cpusets.
203 *
204 * Now, the task_struct fields mems_allowed and mempolicy may be changed
205 * by other task, we use alloc_lock in the task_struct fields to protect
206 * them.
207 *
208 * The cpuset_common_file_read() handlers only hold callback_mutex across
209 * small pieces of code, such as when reading out possibly multi-word
210 * cpumasks and nodemasks.
211 *
212 * Accessing a task's cpuset should be done in accordance with the
213 * guidelines for accessing subsystem state in kernel/cgroup.c
214 */
215
216static DEFINE_MUTEX(callback_mutex);
217
218/*
219 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
220 * buffers. They are statically allocated to prevent using excess stack
221 * when calling cpuset_print_task_mems_allowed().
222 */
223#define CPUSET_NAME_LEN (128)
224#define CPUSET_NODELIST_LEN (256)
225static char cpuset_name[CPUSET_NAME_LEN];
226static char cpuset_nodelist[CPUSET_NODELIST_LEN];
227static DEFINE_SPINLOCK(cpuset_buffer_lock);
228
229/*
230 * This is ugly, but preserves the userspace API for existing cpuset
231 * users. If someone tries to mount the "cpuset" filesystem, we
232 * silently switch it to mount "cgroup" instead
233 */
234static struct dentry *cpuset_mount(struct file_system_type *fs_type,
235             int flags, const char *unused_dev_name, void *data)
236{
237    struct file_system_type *cgroup_fs = get_fs_type("cgroup");
238    struct dentry *ret = ERR_PTR(-ENODEV);
239    if (cgroup_fs) {
240        char mountopts[] =
241            "cpuset,noprefix,"
242            "release_agent=/sbin/cpuset_release_agent";
243        ret = cgroup_fs->mount(cgroup_fs, flags,
244                       unused_dev_name, mountopts);
245        put_filesystem(cgroup_fs);
246    }
247    return ret;
248}
249
250static struct file_system_type cpuset_fs_type = {
251    .name = "cpuset",
252    .mount = cpuset_mount,
253};
254
255/*
256 * Return in pmask the portion of a cpusets's cpus_allowed that
257 * are online. If none are online, walk up the cpuset hierarchy
258 * until we find one that does have some online cpus. If we get
259 * all the way to the top and still haven't found any online cpus,
260 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
261 * task, return cpu_online_map.
262 *
263 * One way or another, we guarantee to return some non-empty subset
264 * of cpu_online_map.
265 *
266 * Call with callback_mutex held.
267 */
268
269static void guarantee_online_cpus(const struct cpuset *cs,
270                  struct cpumask *pmask)
271{
272    while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
273        cs = cs->parent;
274    if (cs)
275        cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
276    else
277        cpumask_copy(pmask, cpu_online_mask);
278    BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
279}
280
281/*
282 * Return in *pmask the portion of a cpusets's mems_allowed that
283 * are online, with memory. If none are online with memory, walk
284 * up the cpuset hierarchy until we find one that does have some
285 * online mems. If we get all the way to the top and still haven't
286 * found any online mems, return node_states[N_HIGH_MEMORY].
287 *
288 * One way or another, we guarantee to return some non-empty subset
289 * of node_states[N_HIGH_MEMORY].
290 *
291 * Call with callback_mutex held.
292 */
293
294static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
295{
296    while (cs && !nodes_intersects(cs->mems_allowed,
297                    node_states[N_HIGH_MEMORY]))
298        cs = cs->parent;
299    if (cs)
300        nodes_and(*pmask, cs->mems_allowed,
301                    node_states[N_HIGH_MEMORY]);
302    else
303        *pmask = node_states[N_HIGH_MEMORY];
304    BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
305}
306
307/*
308 * update task's spread flag if cpuset's page/slab spread flag is set
309 *
310 * Called with callback_mutex/cgroup_mutex held
311 */
312static void cpuset_update_task_spread_flag(struct cpuset *cs,
313                    struct task_struct *tsk)
314{
315    if (is_spread_page(cs))
316        tsk->flags |= PF_SPREAD_PAGE;
317    else
318        tsk->flags &= ~PF_SPREAD_PAGE;
319    if (is_spread_slab(cs))
320        tsk->flags |= PF_SPREAD_SLAB;
321    else
322        tsk->flags &= ~PF_SPREAD_SLAB;
323}
324
325/*
326 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
327 *
328 * One cpuset is a subset of another if all its allowed CPUs and
329 * Memory Nodes are a subset of the other, and its exclusive flags
330 * are only set if the other's are set. Call holding cgroup_mutex.
331 */
332
333static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
334{
335    return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
336        nodes_subset(p->mems_allowed, q->mems_allowed) &&
337        is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
338        is_mem_exclusive(p) <= is_mem_exclusive(q);
339}
340
341/**
342 * alloc_trial_cpuset - allocate a trial cpuset
343 * @cs: the cpuset that the trial cpuset duplicates
344 */
345static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
346{
347    struct cpuset *trial;
348
349    trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
350    if (!trial)
351        return NULL;
352
353    if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
354        kfree(trial);
355        return NULL;
356    }
357    cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
358
359    return trial;
360}
361
362/**
363 * free_trial_cpuset - free the trial cpuset
364 * @trial: the trial cpuset to be freed
365 */
366static void free_trial_cpuset(struct cpuset *trial)
367{
368    free_cpumask_var(trial->cpus_allowed);
369    kfree(trial);
370}
371
372/*
373 * validate_change() - Used to validate that any proposed cpuset change
374 * follows the structural rules for cpusets.
375 *
376 * If we replaced the flag and mask values of the current cpuset
377 * (cur) with those values in the trial cpuset (trial), would
378 * our various subset and exclusive rules still be valid? Presumes
379 * cgroup_mutex held.
380 *
381 * 'cur' is the address of an actual, in-use cpuset. Operations
382 * such as list traversal that depend on the actual address of the
383 * cpuset in the list must use cur below, not trial.
384 *
385 * 'trial' is the address of bulk structure copy of cur, with
386 * perhaps one or more of the fields cpus_allowed, mems_allowed,
387 * or flags changed to new, trial values.
388 *
389 * Return 0 if valid, -errno if not.
390 */
391
392static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
393{
394    struct cgroup *cont;
395    struct cpuset *c, *par;
396
397    /* Each of our child cpusets must be a subset of us */
398    list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
399        if (!is_cpuset_subset(cgroup_cs(cont), trial))
400            return -EBUSY;
401    }
402
403    /* Remaining checks don't apply to root cpuset */
404    if (cur == &top_cpuset)
405        return 0;
406
407    par = cur->parent;
408
409    /* We must be a subset of our parent cpuset */
410    if (!is_cpuset_subset(trial, par))
411        return -EACCES;
412
413    /*
414     * If either I or some sibling (!= me) is exclusive, we can't
415     * overlap
416     */
417    list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
418        c = cgroup_cs(cont);
419        if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
420            c != cur &&
421            cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
422            return -EINVAL;
423        if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
424            c != cur &&
425            nodes_intersects(trial->mems_allowed, c->mems_allowed))
426            return -EINVAL;
427    }
428
429    /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
430    if (cgroup_task_count(cur->css.cgroup)) {
431        if (cpumask_empty(trial->cpus_allowed) ||
432            nodes_empty(trial->mems_allowed)) {
433            return -ENOSPC;
434        }
435    }
436
437    return 0;
438}
439
440#ifdef CONFIG_SMP
441/*
442 * Helper routine for generate_sched_domains().
443 * Do cpusets a, b have overlapping cpus_allowed masks?
444 */
445static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
446{
447    return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
448}
449
450static void
451update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
452{
453    if (dattr->relax_domain_level < c->relax_domain_level)
454        dattr->relax_domain_level = c->relax_domain_level;
455    return;
456}
457
458static void
459update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
460{
461    LIST_HEAD(q);
462
463    list_add(&c->stack_list, &q);
464    while (!list_empty(&q)) {
465        struct cpuset *cp;
466        struct cgroup *cont;
467        struct cpuset *child;
468
469        cp = list_first_entry(&q, struct cpuset, stack_list);
470        list_del(q.next);
471
472        if (cpumask_empty(cp->cpus_allowed))
473            continue;
474
475        if (is_sched_load_balance(cp))
476            update_domain_attr(dattr, cp);
477
478        list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
479            child = cgroup_cs(cont);
480            list_add_tail(&child->stack_list, &q);
481        }
482    }
483}
484
485/*
486 * generate_sched_domains()
487 *
488 * This function builds a partial partition of the systems CPUs
489 * A 'partial partition' is a set of non-overlapping subsets whose
490 * union is a subset of that set.
491 * The output of this function needs to be passed to kernel/sched.c
492 * partition_sched_domains() routine, which will rebuild the scheduler's
493 * load balancing domains (sched domains) as specified by that partial
494 * partition.
495 *
496 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
497 * for a background explanation of this.
498 *
499 * Does not return errors, on the theory that the callers of this
500 * routine would rather not worry about failures to rebuild sched
501 * domains when operating in the severe memory shortage situations
502 * that could cause allocation failures below.
503 *
504 * Must be called with cgroup_lock held.
505 *
506 * The three key local variables below are:
507 * q - a linked-list queue of cpuset pointers, used to implement a
508 * top-down scan of all cpusets. This scan loads a pointer
509 * to each cpuset marked is_sched_load_balance into the
510 * array 'csa'. For our purposes, rebuilding the schedulers
511 * sched domains, we can ignore !is_sched_load_balance cpusets.
512 * csa - (for CpuSet Array) Array of pointers to all the cpusets
513 * that need to be load balanced, for convenient iterative
514 * access by the subsequent code that finds the best partition,
515 * i.e the set of domains (subsets) of CPUs such that the
516 * cpus_allowed of every cpuset marked is_sched_load_balance
517 * is a subset of one of these domains, while there are as
518 * many such domains as possible, each as small as possible.
519 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
520 * the kernel/sched.c routine partition_sched_domains() in a
521 * convenient format, that can be easily compared to the prior
522 * value to determine what partition elements (sched domains)
523 * were changed (added or removed.)
524 *
525 * Finding the best partition (set of domains):
526 * The triple nested loops below over i, j, k scan over the
527 * load balanced cpusets (using the array of cpuset pointers in
528 * csa[]) looking for pairs of cpusets that have overlapping
529 * cpus_allowed, but which don't have the same 'pn' partition
530 * number and gives them in the same partition number. It keeps
531 * looping on the 'restart' label until it can no longer find
532 * any such pairs.
533 *
534 * The union of the cpus_allowed masks from the set of
535 * all cpusets having the same 'pn' value then form the one
536 * element of the partition (one sched domain) to be passed to
537 * partition_sched_domains().
538 */
539static int generate_sched_domains(cpumask_var_t **domains,
540            struct sched_domain_attr **attributes)
541{
542    LIST_HEAD(q); /* queue of cpusets to be scanned */
543    struct cpuset *cp; /* scans q */
544    struct cpuset **csa; /* array of all cpuset ptrs */
545    int csn; /* how many cpuset ptrs in csa so far */
546    int i, j, k; /* indices for partition finding loops */
547    cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
548    struct sched_domain_attr *dattr; /* attributes for custom domains */
549    int ndoms = 0; /* number of sched domains in result */
550    int nslot; /* next empty doms[] struct cpumask slot */
551
552    doms = NULL;
553    dattr = NULL;
554    csa = NULL;
555
556    /* Special case for the 99% of systems with one, full, sched domain */
557    if (is_sched_load_balance(&top_cpuset)) {
558        ndoms = 1;
559        doms = alloc_sched_domains(ndoms);
560        if (!doms)
561            goto done;
562
563        dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
564        if (dattr) {
565            *dattr = SD_ATTR_INIT;
566            update_domain_attr_tree(dattr, &top_cpuset);
567        }
568        cpumask_copy(doms[0], top_cpuset.cpus_allowed);
569
570        goto done;
571    }
572
573    csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
574    if (!csa)
575        goto done;
576    csn = 0;
577
578    list_add(&top_cpuset.stack_list, &q);
579    while (!list_empty(&q)) {
580        struct cgroup *cont;
581        struct cpuset *child; /* scans child cpusets of cp */
582
583        cp = list_first_entry(&q, struct cpuset, stack_list);
584        list_del(q.next);
585
586        if (cpumask_empty(cp->cpus_allowed))
587            continue;
588
589        /*
590         * All child cpusets contain a subset of the parent's cpus, so
591         * just skip them, and then we call update_domain_attr_tree()
592         * to calc relax_domain_level of the corresponding sched
593         * domain.
594         */
595        if (is_sched_load_balance(cp)) {
596            csa[csn++] = cp;
597            continue;
598        }
599
600        list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
601            child = cgroup_cs(cont);
602            list_add_tail(&child->stack_list, &q);
603        }
604      }
605
606    for (i = 0; i < csn; i++)
607        csa[i]->pn = i;
608    ndoms = csn;
609
610restart:
611    /* Find the best partition (set of sched domains) */
612    for (i = 0; i < csn; i++) {
613        struct cpuset *a = csa[i];
614        int apn = a->pn;
615
616        for (j = 0; j < csn; j++) {
617            struct cpuset *b = csa[j];
618            int bpn = b->pn;
619
620            if (apn != bpn && cpusets_overlap(a, b)) {
621                for (k = 0; k < csn; k++) {
622                    struct cpuset *c = csa[k];
623
624                    if (c->pn == bpn)
625                        c->pn = apn;
626                }
627                ndoms--; /* one less element */
628                goto restart;
629            }
630        }
631    }
632
633    /*
634     * Now we know how many domains to create.
635     * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
636     */
637    doms = alloc_sched_domains(ndoms);
638    if (!doms)
639        goto done;
640
641    /*
642     * The rest of the code, including the scheduler, can deal with
643     * dattr==NULL case. No need to abort if alloc fails.
644     */
645    dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
646
647    for (nslot = 0, i = 0; i < csn; i++) {
648        struct cpuset *a = csa[i];
649        struct cpumask *dp;
650        int apn = a->pn;
651
652        if (apn < 0) {
653            /* Skip completed partitions */
654            continue;
655        }
656
657        dp = doms[nslot];
658
659        if (nslot == ndoms) {
660            static int warnings = 10;
661            if (warnings) {
662                printk(KERN_WARNING
663                 "rebuild_sched_domains confused:"
664                  " nslot %d, ndoms %d, csn %d, i %d,"
665                  " apn %d\n",
666                  nslot, ndoms, csn, i, apn);
667                warnings--;
668            }
669            continue;
670        }
671
672        cpumask_clear(dp);
673        if (dattr)
674            *(dattr + nslot) = SD_ATTR_INIT;
675        for (j = i; j < csn; j++) {
676            struct cpuset *b = csa[j];
677
678            if (apn == b->pn) {
679                cpumask_or(dp, dp, b->cpus_allowed);
680                if (dattr)
681                    update_domain_attr_tree(dattr + nslot, b);
682
683                /* Done with this partition */
684                b->pn = -1;
685            }
686        }
687        nslot++;
688    }
689    BUG_ON(nslot != ndoms);
690
691done:
692    kfree(csa);
693
694    /*
695     * Fallback to the default domain if kmalloc() failed.
696     * See comments in partition_sched_domains().
697     */
698    if (doms == NULL)
699        ndoms = 1;
700
701    *domains = doms;
702    *attributes = dattr;
703    return ndoms;
704}
705
706/*
707 * Rebuild scheduler domains.
708 *
709 * Call with neither cgroup_mutex held nor within get_online_cpus().
710 * Takes both cgroup_mutex and get_online_cpus().
711 *
712 * Cannot be directly called from cpuset code handling changes
713 * to the cpuset pseudo-filesystem, because it cannot be called
714 * from code that already holds cgroup_mutex.
715 */
716static void do_rebuild_sched_domains(struct work_struct *unused)
717{
718    struct sched_domain_attr *attr;
719    cpumask_var_t *doms;
720    int ndoms;
721
722    get_online_cpus();
723
724    /* Generate domain masks and attrs */
725    cgroup_lock();
726    ndoms = generate_sched_domains(&doms, &attr);
727    cgroup_unlock();
728
729    /* Have scheduler rebuild the domains */
730    partition_sched_domains(ndoms, doms, attr);
731
732    put_online_cpus();
733}
734#else /* !CONFIG_SMP */
735static void do_rebuild_sched_domains(struct work_struct *unused)
736{
737}
738
739static int generate_sched_domains(cpumask_var_t **domains,
740            struct sched_domain_attr **attributes)
741{
742    *domains = NULL;
743    return 1;
744}
745#endif /* CONFIG_SMP */
746
747static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
748
749/*
750 * Rebuild scheduler domains, asynchronously via workqueue.
751 *
752 * If the flag 'sched_load_balance' of any cpuset with non-empty
753 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
754 * which has that flag enabled, or if any cpuset with a non-empty
755 * 'cpus' is removed, then call this routine to rebuild the
756 * scheduler's dynamic sched domains.
757 *
758 * The rebuild_sched_domains() and partition_sched_domains()
759 * routines must nest cgroup_lock() inside get_online_cpus(),
760 * but such cpuset changes as these must nest that locking the
761 * other way, holding cgroup_lock() for much of the code.
762 *
763 * So in order to avoid an ABBA deadlock, the cpuset code handling
764 * these user changes delegates the actual sched domain rebuilding
765 * to a separate workqueue thread, which ends up processing the
766 * above do_rebuild_sched_domains() function.
767 */
768static void async_rebuild_sched_domains(void)
769{
770    queue_work(cpuset_wq, &rebuild_sched_domains_work);
771}
772
773/*
774 * Accomplishes the same scheduler domain rebuild as the above
775 * async_rebuild_sched_domains(), however it directly calls the
776 * rebuild routine synchronously rather than calling it via an
777 * asynchronous work thread.
778 *
779 * This can only be called from code that is not holding
780 * cgroup_mutex (not nested in a cgroup_lock() call.)
781 */
782void rebuild_sched_domains(void)
783{
784    do_rebuild_sched_domains(NULL);
785}
786
787/**
788 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
789 * @tsk: task to test
790 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
791 *
792 * Call with cgroup_mutex held. May take callback_mutex during call.
793 * Called for each task in a cgroup by cgroup_scan_tasks().
794 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
795 * words, if its mask is not equal to its cpuset's mask).
796 */
797static int cpuset_test_cpumask(struct task_struct *tsk,
798                   struct cgroup_scanner *scan)
799{
800    return !cpumask_equal(&tsk->cpus_allowed,
801            (cgroup_cs(scan->cg))->cpus_allowed);
802}
803
804/**
805 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
806 * @tsk: task to test
807 * @scan: struct cgroup_scanner containing the cgroup of the task
808 *
809 * Called by cgroup_scan_tasks() for each task in a cgroup whose
810 * cpus_allowed mask needs to be changed.
811 *
812 * We don't need to re-check for the cgroup/cpuset membership, since we're
813 * holding cgroup_lock() at this point.
814 */
815static void cpuset_change_cpumask(struct task_struct *tsk,
816                  struct cgroup_scanner *scan)
817{
818    set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
819}
820
821/**
822 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
823 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
824 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
825 *
826 * Called with cgroup_mutex held
827 *
828 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
829 * calling callback functions for each.
830 *
831 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
832 * if @heap != NULL.
833 */
834static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
835{
836    struct cgroup_scanner scan;
837
838    scan.cg = cs->css.cgroup;
839    scan.test_task = cpuset_test_cpumask;
840    scan.process_task = cpuset_change_cpumask;
841    scan.heap = heap;
842    cgroup_scan_tasks(&scan);
843}
844
845/**
846 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
847 * @cs: the cpuset to consider
848 * @buf: buffer of cpu numbers written to this cpuset
849 */
850static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
851              const char *buf)
852{
853    struct ptr_heap heap;
854    int retval;
855    int is_load_balanced;
856
857    /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
858    if (cs == &top_cpuset)
859        return -EACCES;
860
861    /*
862     * An empty cpus_allowed is ok only if the cpuset has no tasks.
863     * Since cpulist_parse() fails on an empty mask, we special case
864     * that parsing. The validate_change() call ensures that cpusets
865     * with tasks have cpus.
866     */
867    if (!*buf) {
868        cpumask_clear(trialcs->cpus_allowed);
869    } else {
870        retval = cpulist_parse(buf, trialcs->cpus_allowed);
871        if (retval < 0)
872            return retval;
873
874        if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
875            return -EINVAL;
876    }
877    retval = validate_change(cs, trialcs);
878    if (retval < 0)
879        return retval;
880
881    /* Nothing to do if the cpus didn't change */
882    if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
883        return 0;
884
885    retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
886    if (retval)
887        return retval;
888
889    is_load_balanced = is_sched_load_balance(trialcs);
890
891    mutex_lock(&callback_mutex);
892    cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
893    mutex_unlock(&callback_mutex);
894
895    /*
896     * Scan tasks in the cpuset, and update the cpumasks of any
897     * that need an update.
898     */
899    update_tasks_cpumask(cs, &heap);
900
901    heap_free(&heap);
902
903    if (is_load_balanced)
904        async_rebuild_sched_domains();
905    return 0;
906}
907
908/*
909 * cpuset_migrate_mm
910 *
911 * Migrate memory region from one set of nodes to another.
912 *
913 * Temporarilly set tasks mems_allowed to target nodes of migration,
914 * so that the migration code can allocate pages on these nodes.
915 *
916 * Call holding cgroup_mutex, so current's cpuset won't change
917 * during this call, as manage_mutex holds off any cpuset_attach()
918 * calls. Therefore we don't need to take task_lock around the
919 * call to guarantee_online_mems(), as we know no one is changing
920 * our task's cpuset.
921 *
922 * While the mm_struct we are migrating is typically from some
923 * other task, the task_struct mems_allowed that we are hacking
924 * is for our current task, which must allocate new pages for that
925 * migrating memory region.
926 */
927
928static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
929                            const nodemask_t *to)
930{
931    struct task_struct *tsk = current;
932
933    tsk->mems_allowed = *to;
934
935    do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
936
937    guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
938}
939
940/*
941 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
942 * @tsk: the task to change
943 * @newmems: new nodes that the task will be set
944 *
945 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
946 * we structure updates as setting all new allowed nodes, then clearing newly
947 * disallowed ones.
948 */
949static void cpuset_change_task_nodemask(struct task_struct *tsk,
950                    nodemask_t *newmems)
951{
952repeat:
953    /*
954     * Allow tasks that have access to memory reserves because they have
955     * been OOM killed to get memory anywhere.
956     */
957    if (unlikely(test_thread_flag(TIF_MEMDIE)))
958        return;
959    if (current->flags & PF_EXITING) /* Let dying task have memory */
960        return;
961
962    task_lock(tsk);
963    nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
964    mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
965
966
967    /*
968     * ensure checking ->mems_allowed_change_disable after setting all new
969     * allowed nodes.
970     *
971     * the read-side task can see an nodemask with new allowed nodes and
972     * old allowed nodes. and if it allocates page when cpuset clears newly
973     * disallowed ones continuous, it can see the new allowed bits.
974     *
975     * And if setting all new allowed nodes is after the checking, setting
976     * all new allowed nodes and clearing newly disallowed ones will be done
977     * continuous, and the read-side task may find no node to alloc page.
978     */
979    smp_mb();
980
981    /*
982     * Allocation of memory is very fast, we needn't sleep when waiting
983     * for the read-side.
984     */
985    while (ACCESS_ONCE(tsk->mems_allowed_change_disable)) {
986        task_unlock(tsk);
987        if (!task_curr(tsk))
988            yield();
989        goto repeat;
990    }
991
992    /*
993     * ensure checking ->mems_allowed_change_disable before clearing all new
994     * disallowed nodes.
995     *
996     * if clearing newly disallowed bits before the checking, the read-side
997     * task may find no node to alloc page.
998     */
999    smp_mb();
1000
1001    mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1002    tsk->mems_allowed = *newmems;
1003    task_unlock(tsk);
1004}
1005
1006/*
1007 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1008 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1009 * memory_migrate flag is set. Called with cgroup_mutex held.
1010 */
1011static void cpuset_change_nodemask(struct task_struct *p,
1012                   struct cgroup_scanner *scan)
1013{
1014    struct mm_struct *mm;
1015    struct cpuset *cs;
1016    int migrate;
1017    const nodemask_t *oldmem = scan->data;
1018    static nodemask_t newmems; /* protected by cgroup_mutex */
1019
1020    cs = cgroup_cs(scan->cg);
1021    guarantee_online_mems(cs, &newmems);
1022
1023    cpuset_change_task_nodemask(p, &newmems);
1024
1025    mm = get_task_mm(p);
1026    if (!mm)
1027        return;
1028
1029    migrate = is_memory_migrate(cs);
1030
1031    mpol_rebind_mm(mm, &cs->mems_allowed);
1032    if (migrate)
1033        cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1034    mmput(mm);
1035}
1036
1037static void *cpuset_being_rebound;
1038
1039/**
1040 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1041 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1042 * @oldmem: old mems_allowed of cpuset cs
1043 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1044 *
1045 * Called with cgroup_mutex held
1046 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1047 * if @heap != NULL.
1048 */
1049static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1050                 struct ptr_heap *heap)
1051{
1052    struct cgroup_scanner scan;
1053
1054    cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1055
1056    scan.cg = cs->css.cgroup;
1057    scan.test_task = NULL;
1058    scan.process_task = cpuset_change_nodemask;
1059    scan.heap = heap;
1060    scan.data = (nodemask_t *)oldmem;
1061
1062    /*
1063     * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1064     * take while holding tasklist_lock. Forks can happen - the
1065     * mpol_dup() cpuset_being_rebound check will catch such forks,
1066     * and rebind their vma mempolicies too. Because we still hold
1067     * the global cgroup_mutex, we know that no other rebind effort
1068     * will be contending for the global variable cpuset_being_rebound.
1069     * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1070     * is idempotent. Also migrate pages in each mm to new nodes.
1071     */
1072    cgroup_scan_tasks(&scan);
1073
1074    /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1075    cpuset_being_rebound = NULL;
1076}
1077
1078/*
1079 * Handle user request to change the 'mems' memory placement
1080 * of a cpuset. Needs to validate the request, update the
1081 * cpusets mems_allowed, and for each task in the cpuset,
1082 * update mems_allowed and rebind task's mempolicy and any vma
1083 * mempolicies and if the cpuset is marked 'memory_migrate',
1084 * migrate the tasks pages to the new memory.
1085 *
1086 * Call with cgroup_mutex held. May take callback_mutex during call.
1087 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1088 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1089 * their mempolicies to the cpusets new mems_allowed.
1090 */
1091static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1092               const char *buf)
1093{
1094    NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1095    int retval;
1096    struct ptr_heap heap;
1097
1098    if (!oldmem)
1099        return -ENOMEM;
1100
1101    /*
1102     * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1103     * it's read-only
1104     */
1105    if (cs == &top_cpuset) {
1106        retval = -EACCES;
1107        goto done;
1108    }
1109
1110    /*
1111     * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1112     * Since nodelist_parse() fails on an empty mask, we special case
1113     * that parsing. The validate_change() call ensures that cpusets
1114     * with tasks have memory.
1115     */
1116    if (!*buf) {
1117        nodes_clear(trialcs->mems_allowed);
1118    } else {
1119        retval = nodelist_parse(buf, trialcs->mems_allowed);
1120        if (retval < 0)
1121            goto done;
1122
1123        if (!nodes_subset(trialcs->mems_allowed,
1124                node_states[N_HIGH_MEMORY])) {
1125            retval = -EINVAL;
1126            goto done;
1127        }
1128    }
1129    *oldmem = cs->mems_allowed;
1130    if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1131        retval = 0; /* Too easy - nothing to do */
1132        goto done;
1133    }
1134    retval = validate_change(cs, trialcs);
1135    if (retval < 0)
1136        goto done;
1137
1138    retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1139    if (retval < 0)
1140        goto done;
1141
1142    mutex_lock(&callback_mutex);
1143    cs->mems_allowed = trialcs->mems_allowed;
1144    mutex_unlock(&callback_mutex);
1145
1146    update_tasks_nodemask(cs, oldmem, &heap);
1147
1148    heap_free(&heap);
1149done:
1150    NODEMASK_FREE(oldmem);
1151    return retval;
1152}
1153
1154int current_cpuset_is_being_rebound(void)
1155{
1156    return task_cs(current) == cpuset_being_rebound;
1157}
1158
1159static int update_relax_domain_level(struct cpuset *cs, s64 val)
1160{
1161#ifdef CONFIG_SMP
1162    if (val < -1 || val >= sched_domain_level_max)
1163        return -EINVAL;
1164#endif
1165
1166    if (val != cs->relax_domain_level) {
1167        cs->relax_domain_level = val;
1168        if (!cpumask_empty(cs->cpus_allowed) &&
1169            is_sched_load_balance(cs))
1170            async_rebuild_sched_domains();
1171    }
1172
1173    return 0;
1174}
1175
1176/*
1177 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1178 * @tsk: task to be updated
1179 * @scan: struct cgroup_scanner containing the cgroup of the task
1180 *
1181 * Called by cgroup_scan_tasks() for each task in a cgroup.
1182 *
1183 * We don't need to re-check for the cgroup/cpuset membership, since we're
1184 * holding cgroup_lock() at this point.
1185 */
1186static void cpuset_change_flag(struct task_struct *tsk,
1187                struct cgroup_scanner *scan)
1188{
1189    cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1190}
1191
1192/*
1193 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1194 * @cs: the cpuset in which each task's spread flags needs to be changed
1195 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1196 *
1197 * Called with cgroup_mutex held
1198 *
1199 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1200 * calling callback functions for each.
1201 *
1202 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1203 * if @heap != NULL.
1204 */
1205static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1206{
1207    struct cgroup_scanner scan;
1208
1209    scan.cg = cs->css.cgroup;
1210    scan.test_task = NULL;
1211    scan.process_task = cpuset_change_flag;
1212    scan.heap = heap;
1213    cgroup_scan_tasks(&scan);
1214}
1215
1216/*
1217 * update_flag - read a 0 or a 1 in a file and update associated flag
1218 * bit: the bit to update (see cpuset_flagbits_t)
1219 * cs: the cpuset to update
1220 * turning_on: whether the flag is being set or cleared
1221 *
1222 * Call with cgroup_mutex held.
1223 */
1224
1225static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1226               int turning_on)
1227{
1228    struct cpuset *trialcs;
1229    int balance_flag_changed;
1230    int spread_flag_changed;
1231    struct ptr_heap heap;
1232    int err;
1233
1234    trialcs = alloc_trial_cpuset(cs);
1235    if (!trialcs)
1236        return -ENOMEM;
1237
1238    if (turning_on)
1239        set_bit(bit, &trialcs->flags);
1240    else
1241        clear_bit(bit, &trialcs->flags);
1242
1243    err = validate_change(cs, trialcs);
1244    if (err < 0)
1245        goto out;
1246
1247    err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1248    if (err < 0)
1249        goto out;
1250
1251    balance_flag_changed = (is_sched_load_balance(cs) !=
1252                is_sched_load_balance(trialcs));
1253
1254    spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1255            || (is_spread_page(cs) != is_spread_page(trialcs)));
1256
1257    mutex_lock(&callback_mutex);
1258    cs->flags = trialcs->flags;
1259    mutex_unlock(&callback_mutex);
1260
1261    if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1262        async_rebuild_sched_domains();
1263
1264    if (spread_flag_changed)
1265        update_tasks_flags(cs, &heap);
1266    heap_free(&heap);
1267out:
1268    free_trial_cpuset(trialcs);
1269    return err;
1270}
1271
1272/*
1273 * Frequency meter - How fast is some event occurring?
1274 *
1275 * These routines manage a digitally filtered, constant time based,
1276 * event frequency meter. There are four routines:
1277 * fmeter_init() - initialize a frequency meter.
1278 * fmeter_markevent() - called each time the event happens.
1279 * fmeter_getrate() - returns the recent rate of such events.
1280 * fmeter_update() - internal routine used to update fmeter.
1281 *
1282 * A common data structure is passed to each of these routines,
1283 * which is used to keep track of the state required to manage the
1284 * frequency meter and its digital filter.
1285 *
1286 * The filter works on the number of events marked per unit time.
1287 * The filter is single-pole low-pass recursive (IIR). The time unit
1288 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1289 * simulate 3 decimal digits of precision (multiplied by 1000).
1290 *
1291 * With an FM_COEF of 933, and a time base of 1 second, the filter
1292 * has a half-life of 10 seconds, meaning that if the events quit
1293 * happening, then the rate returned from the fmeter_getrate()
1294 * will be cut in half each 10 seconds, until it converges to zero.
1295 *
1296 * It is not worth doing a real infinitely recursive filter. If more
1297 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1298 * just compute FM_MAXTICKS ticks worth, by which point the level
1299 * will be stable.
1300 *
1301 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1302 * arithmetic overflow in the fmeter_update() routine.
1303 *
1304 * Given the simple 32 bit integer arithmetic used, this meter works
1305 * best for reporting rates between one per millisecond (msec) and
1306 * one per 32 (approx) seconds. At constant rates faster than one
1307 * per msec it maxes out at values just under 1,000,000. At constant
1308 * rates between one per msec, and one per second it will stabilize
1309 * to a value N*1000, where N is the rate of events per second.
1310 * At constant rates between one per second and one per 32 seconds,
1311 * it will be choppy, moving up on the seconds that have an event,
1312 * and then decaying until the next event. At rates slower than
1313 * about one in 32 seconds, it decays all the way back to zero between
1314 * each event.
1315 */
1316
1317#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1318#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1319#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1320#define FM_SCALE 1000 /* faux fixed point scale */
1321
1322/* Initialize a frequency meter */
1323static void fmeter_init(struct fmeter *fmp)
1324{
1325    fmp->cnt = 0;
1326    fmp->val = 0;
1327    fmp->time = 0;
1328    spin_lock_init(&fmp->lock);
1329}
1330
1331/* Internal meter update - process cnt events and update value */
1332static void fmeter_update(struct fmeter *fmp)
1333{
1334    time_t now = get_seconds();
1335    time_t ticks = now - fmp->time;
1336
1337    if (ticks == 0)
1338        return;
1339
1340    ticks = min(FM_MAXTICKS, ticks);
1341    while (ticks-- > 0)
1342        fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1343    fmp->time = now;
1344
1345    fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1346    fmp->cnt = 0;
1347}
1348
1349/* Process any previous ticks, then bump cnt by one (times scale). */
1350static void fmeter_markevent(struct fmeter *fmp)
1351{
1352    spin_lock(&fmp->lock);
1353    fmeter_update(fmp);
1354    fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1355    spin_unlock(&fmp->lock);
1356}
1357
1358/* Process any previous ticks, then return current value. */
1359static int fmeter_getrate(struct fmeter *fmp)
1360{
1361    int val;
1362
1363    spin_lock(&fmp->lock);
1364    fmeter_update(fmp);
1365    val = fmp->val;
1366    spin_unlock(&fmp->lock);
1367    return val;
1368}
1369
1370/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1371static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1372                 struct task_struct *tsk)
1373{
1374    struct cpuset *cs = cgroup_cs(cont);
1375
1376    if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1377        return -ENOSPC;
1378
1379    /*
1380     * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1381     * cannot change their cpu affinity and isolating such threads by their
1382     * set of allowed nodes is unnecessary. Thus, cpusets are not
1383     * applicable for such threads. This prevents checking for success of
1384     * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1385     * be changed.
1386     */
1387    if (tsk->flags & PF_THREAD_BOUND)
1388        return -EINVAL;
1389
1390    return 0;
1391}
1392
1393static int cpuset_can_attach_task(struct cgroup *cgrp, struct task_struct *task)
1394{
1395    return security_task_setscheduler(task);
1396}
1397
1398/*
1399 * Protected by cgroup_lock. The nodemasks must be stored globally because
1400 * dynamically allocating them is not allowed in pre_attach, and they must
1401 * persist among pre_attach, attach_task, and attach.
1402 */
1403static cpumask_var_t cpus_attach;
1404static nodemask_t cpuset_attach_nodemask_from;
1405static nodemask_t cpuset_attach_nodemask_to;
1406
1407/* Set-up work for before attaching each task. */
1408static void cpuset_pre_attach(struct cgroup *cont)
1409{
1410    struct cpuset *cs = cgroup_cs(cont);
1411
1412    if (cs == &top_cpuset)
1413        cpumask_copy(cpus_attach, cpu_possible_mask);
1414    else
1415        guarantee_online_cpus(cs, cpus_attach);
1416
1417    guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1418}
1419
1420/* Per-thread attachment work. */
1421static void cpuset_attach_task(struct cgroup *cont, struct task_struct *tsk)
1422{
1423    int err;
1424    struct cpuset *cs = cgroup_cs(cont);
1425
1426    /*
1427     * can_attach beforehand should guarantee that this doesn't fail.
1428     * TODO: have a better way to handle failure here
1429     */
1430    err = set_cpus_allowed_ptr(tsk, cpus_attach);
1431    WARN_ON_ONCE(err);
1432
1433    cpuset_change_task_nodemask(tsk, &cpuset_attach_nodemask_to);
1434    cpuset_update_task_spread_flag(cs, tsk);
1435}
1436
1437static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1438              struct cgroup *oldcont, struct task_struct *tsk)
1439{
1440    struct mm_struct *mm;
1441    struct cpuset *cs = cgroup_cs(cont);
1442    struct cpuset *oldcs = cgroup_cs(oldcont);
1443
1444    /*
1445     * Change mm, possibly for multiple threads in a threadgroup. This is
1446     * expensive and may sleep.
1447     */
1448    cpuset_attach_nodemask_from = oldcs->mems_allowed;
1449    cpuset_attach_nodemask_to = cs->mems_allowed;
1450    mm = get_task_mm(tsk);
1451    if (mm) {
1452        mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1453        if (is_memory_migrate(cs))
1454            cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1455                      &cpuset_attach_nodemask_to);
1456        mmput(mm);
1457    }
1458}
1459
1460/* The various types of files and directories in a cpuset file system */
1461
1462typedef enum {
1463    FILE_MEMORY_MIGRATE,
1464    FILE_CPULIST,
1465    FILE_MEMLIST,
1466    FILE_CPU_EXCLUSIVE,
1467    FILE_MEM_EXCLUSIVE,
1468    FILE_MEM_HARDWALL,
1469    FILE_SCHED_LOAD_BALANCE,
1470    FILE_SCHED_RELAX_DOMAIN_LEVEL,
1471    FILE_MEMORY_PRESSURE_ENABLED,
1472    FILE_MEMORY_PRESSURE,
1473    FILE_SPREAD_PAGE,
1474    FILE_SPREAD_SLAB,
1475} cpuset_filetype_t;
1476
1477static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1478{
1479    int retval = 0;
1480    struct cpuset *cs = cgroup_cs(cgrp);
1481    cpuset_filetype_t type = cft->private;
1482
1483    if (!cgroup_lock_live_group(cgrp))
1484        return -ENODEV;
1485
1486    switch (type) {
1487    case FILE_CPU_EXCLUSIVE:
1488        retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1489        break;
1490    case FILE_MEM_EXCLUSIVE:
1491        retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1492        break;
1493    case FILE_MEM_HARDWALL:
1494        retval = update_flag(CS_MEM_HARDWALL, cs, val);
1495        break;
1496    case FILE_SCHED_LOAD_BALANCE:
1497        retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1498        break;
1499    case FILE_MEMORY_MIGRATE:
1500        retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1501        break;
1502    case FILE_MEMORY_PRESSURE_ENABLED:
1503        cpuset_memory_pressure_enabled = !!val;
1504        break;
1505    case FILE_MEMORY_PRESSURE:
1506        retval = -EACCES;
1507        break;
1508    case FILE_SPREAD_PAGE:
1509        retval = update_flag(CS_SPREAD_PAGE, cs, val);
1510        break;
1511    case FILE_SPREAD_SLAB:
1512        retval = update_flag(CS_SPREAD_SLAB, cs, val);
1513        break;
1514    default:
1515        retval = -EINVAL;
1516        break;
1517    }
1518    cgroup_unlock();
1519    return retval;
1520}
1521
1522static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1523{
1524    int retval = 0;
1525    struct cpuset *cs = cgroup_cs(cgrp);
1526    cpuset_filetype_t type = cft->private;
1527
1528    if (!cgroup_lock_live_group(cgrp))
1529        return -ENODEV;
1530
1531    switch (type) {
1532    case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1533        retval = update_relax_domain_level(cs, val);
1534        break;
1535    default:
1536        retval = -EINVAL;
1537        break;
1538    }
1539    cgroup_unlock();
1540    return retval;
1541}
1542
1543/*
1544 * Common handling for a write to a "cpus" or "mems" file.
1545 */
1546static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1547                const char *buf)
1548{
1549    int retval = 0;
1550    struct cpuset *cs = cgroup_cs(cgrp);
1551    struct cpuset *trialcs;
1552
1553    if (!cgroup_lock_live_group(cgrp))
1554        return -ENODEV;
1555
1556    trialcs = alloc_trial_cpuset(cs);
1557    if (!trialcs) {
1558        retval = -ENOMEM;
1559        goto out;
1560    }
1561
1562    switch (cft->private) {
1563    case FILE_CPULIST:
1564        retval = update_cpumask(cs, trialcs, buf);
1565        break;
1566    case FILE_MEMLIST:
1567        retval = update_nodemask(cs, trialcs, buf);
1568        break;
1569    default:
1570        retval = -EINVAL;
1571        break;
1572    }
1573
1574    free_trial_cpuset(trialcs);
1575out:
1576    cgroup_unlock();
1577    return retval;
1578}
1579
1580/*
1581 * These ascii lists should be read in a single call, by using a user
1582 * buffer large enough to hold the entire map. If read in smaller
1583 * chunks, there is no guarantee of atomicity. Since the display format
1584 * used, list of ranges of sequential numbers, is variable length,
1585 * and since these maps can change value dynamically, one could read
1586 * gibberish by doing partial reads while a list was changing.
1587 * A single large read to a buffer that crosses a page boundary is
1588 * ok, because the result being copied to user land is not recomputed
1589 * across a page fault.
1590 */
1591
1592static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1593{
1594    size_t count;
1595
1596    mutex_lock(&callback_mutex);
1597    count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1598    mutex_unlock(&callback_mutex);
1599
1600    return count;
1601}
1602
1603static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1604{
1605    size_t count;
1606
1607    mutex_lock(&callback_mutex);
1608    count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1609    mutex_unlock(&callback_mutex);
1610
1611    return count;
1612}
1613
1614static ssize_t cpuset_common_file_read(struct cgroup *cont,
1615                       struct cftype *cft,
1616                       struct file *file,
1617                       char __user *buf,
1618                       size_t nbytes, loff_t *ppos)
1619{
1620    struct cpuset *cs = cgroup_cs(cont);
1621    cpuset_filetype_t type = cft->private;
1622    char *page;
1623    ssize_t retval = 0;
1624    char *s;
1625
1626    if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1627        return -ENOMEM;
1628
1629    s = page;
1630
1631    switch (type) {
1632    case FILE_CPULIST:
1633        s += cpuset_sprintf_cpulist(s, cs);
1634        break;
1635    case FILE_MEMLIST:
1636        s += cpuset_sprintf_memlist(s, cs);
1637        break;
1638    default:
1639        retval = -EINVAL;
1640        goto out;
1641    }
1642    *s++ = '\n';
1643
1644    retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1645out:
1646    free_page((unsigned long)page);
1647    return retval;
1648}
1649
1650static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1651{
1652    struct cpuset *cs = cgroup_cs(cont);
1653    cpuset_filetype_t type = cft->private;
1654    switch (type) {
1655    case FILE_CPU_EXCLUSIVE:
1656        return is_cpu_exclusive(cs);
1657    case FILE_MEM_EXCLUSIVE:
1658        return is_mem_exclusive(cs);
1659    case FILE_MEM_HARDWALL:
1660        return is_mem_hardwall(cs);
1661    case FILE_SCHED_LOAD_BALANCE:
1662        return is_sched_load_balance(cs);
1663    case FILE_MEMORY_MIGRATE:
1664        return is_memory_migrate(cs);
1665    case FILE_MEMORY_PRESSURE_ENABLED:
1666        return cpuset_memory_pressure_enabled;
1667    case FILE_MEMORY_PRESSURE:
1668        return fmeter_getrate(&cs->fmeter);
1669    case FILE_SPREAD_PAGE:
1670        return is_spread_page(cs);
1671    case FILE_SPREAD_SLAB:
1672        return is_spread_slab(cs);
1673    default:
1674        BUG();
1675    }
1676
1677    /* Unreachable but makes gcc happy */
1678    return 0;
1679}
1680
1681static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1682{
1683    struct cpuset *cs = cgroup_cs(cont);
1684    cpuset_filetype_t type = cft->private;
1685    switch (type) {
1686    case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1687        return cs->relax_domain_level;
1688    default:
1689        BUG();
1690    }
1691
1692    /* Unrechable but makes gcc happy */
1693    return 0;
1694}
1695
1696
1697/*
1698 * for the common functions, 'private' gives the type of file
1699 */
1700
1701static struct cftype files[] = {
1702    {
1703        .name = "cpus",
1704        .read = cpuset_common_file_read,
1705        .write_string = cpuset_write_resmask,
1706        .max_write_len = (100U + 6 * NR_CPUS),
1707        .private = FILE_CPULIST,
1708    },
1709
1710    {
1711        .name = "mems",
1712        .read = cpuset_common_file_read,
1713        .write_string = cpuset_write_resmask,
1714        .max_write_len = (100U + 6 * MAX_NUMNODES),
1715        .private = FILE_MEMLIST,
1716    },
1717
1718    {
1719        .name = "cpu_exclusive",
1720        .read_u64 = cpuset_read_u64,
1721        .write_u64 = cpuset_write_u64,
1722        .private = FILE_CPU_EXCLUSIVE,
1723    },
1724
1725    {
1726        .name = "mem_exclusive",
1727        .read_u64 = cpuset_read_u64,
1728        .write_u64 = cpuset_write_u64,
1729        .private = FILE_MEM_EXCLUSIVE,
1730    },
1731
1732    {
1733        .name = "mem_hardwall",
1734        .read_u64 = cpuset_read_u64,
1735        .write_u64 = cpuset_write_u64,
1736        .private = FILE_MEM_HARDWALL,
1737    },
1738
1739    {
1740        .name = "sched_load_balance",
1741        .read_u64 = cpuset_read_u64,
1742        .write_u64 = cpuset_write_u64,
1743        .private = FILE_SCHED_LOAD_BALANCE,
1744    },
1745
1746    {
1747        .name = "sched_relax_domain_level",
1748        .read_s64 = cpuset_read_s64,
1749        .write_s64 = cpuset_write_s64,
1750        .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1751    },
1752
1753    {
1754        .name = "memory_migrate",
1755        .read_u64 = cpuset_read_u64,
1756        .write_u64 = cpuset_write_u64,
1757        .private = FILE_MEMORY_MIGRATE,
1758    },
1759
1760    {
1761        .name = "memory_pressure",
1762        .read_u64 = cpuset_read_u64,
1763        .write_u64 = cpuset_write_u64,
1764        .private = FILE_MEMORY_PRESSURE,
1765        .mode = S_IRUGO,
1766    },
1767
1768    {
1769        .name = "memory_spread_page",
1770        .read_u64 = cpuset_read_u64,
1771        .write_u64 = cpuset_write_u64,
1772        .private = FILE_SPREAD_PAGE,
1773    },
1774
1775    {
1776        .name = "memory_spread_slab",
1777        .read_u64 = cpuset_read_u64,
1778        .write_u64 = cpuset_write_u64,
1779        .private = FILE_SPREAD_SLAB,
1780    },
1781};
1782
1783static struct cftype cft_memory_pressure_enabled = {
1784    .name = "memory_pressure_enabled",
1785    .read_u64 = cpuset_read_u64,
1786    .write_u64 = cpuset_write_u64,
1787    .private = FILE_MEMORY_PRESSURE_ENABLED,
1788};
1789
1790static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1791{
1792    int err;
1793
1794    err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1795    if (err)
1796        return err;
1797    /* memory_pressure_enabled is in root cpuset only */
1798    if (!cont->parent)
1799        err = cgroup_add_file(cont, ss,
1800                      &cft_memory_pressure_enabled);
1801    return err;
1802}
1803
1804/*
1805 * post_clone() is called during cgroup_create() when the
1806 * clone_children mount argument was specified. The cgroup
1807 * can not yet have any tasks.
1808 *
1809 * Currently we refuse to set up the cgroup - thereby
1810 * refusing the task to be entered, and as a result refusing
1811 * the sys_unshare() or clone() which initiated it - if any
1812 * sibling cpusets have exclusive cpus or mem.
1813 *
1814 * If this becomes a problem for some users who wish to
1815 * allow that scenario, then cpuset_post_clone() could be
1816 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1817 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1818 * held.
1819 */
1820static void cpuset_post_clone(struct cgroup_subsys *ss,
1821                  struct cgroup *cgroup)
1822{
1823    struct cgroup *parent, *child;
1824    struct cpuset *cs, *parent_cs;
1825
1826    parent = cgroup->parent;
1827    list_for_each_entry(child, &parent->children, sibling) {
1828        cs = cgroup_cs(child);
1829        if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1830            return;
1831    }
1832    cs = cgroup_cs(cgroup);
1833    parent_cs = cgroup_cs(parent);
1834
1835    mutex_lock(&callback_mutex);
1836    cs->mems_allowed = parent_cs->mems_allowed;
1837    cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1838    mutex_unlock(&callback_mutex);
1839    return;
1840}
1841
1842/*
1843 * cpuset_create - create a cpuset
1844 * ss: cpuset cgroup subsystem
1845 * cont: control group that the new cpuset will be part of
1846 */
1847
1848static struct cgroup_subsys_state *cpuset_create(
1849    struct cgroup_subsys *ss,
1850    struct cgroup *cont)
1851{
1852    struct cpuset *cs;
1853    struct cpuset *parent;
1854
1855    if (!cont->parent) {
1856        return &top_cpuset.css;
1857    }
1858    parent = cgroup_cs(cont->parent);
1859    cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1860    if (!cs)
1861        return ERR_PTR(-ENOMEM);
1862    if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1863        kfree(cs);
1864        return ERR_PTR(-ENOMEM);
1865    }
1866
1867    cs->flags = 0;
1868    if (is_spread_page(parent))
1869        set_bit(CS_SPREAD_PAGE, &cs->flags);
1870    if (is_spread_slab(parent))
1871        set_bit(CS_SPREAD_SLAB, &cs->flags);
1872    set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1873    cpumask_clear(cs->cpus_allowed);
1874    nodes_clear(cs->mems_allowed);
1875    fmeter_init(&cs->fmeter);
1876    cs->relax_domain_level = -1;
1877
1878    cs->parent = parent;
1879    number_of_cpusets++;
1880    return &cs->css ;
1881}
1882
1883/*
1884 * If the cpuset being removed has its flag 'sched_load_balance'
1885 * enabled, then simulate turning sched_load_balance off, which
1886 * will call async_rebuild_sched_domains().
1887 */
1888
1889static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1890{
1891    struct cpuset *cs = cgroup_cs(cont);
1892
1893    if (is_sched_load_balance(cs))
1894        update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1895
1896    number_of_cpusets--;
1897    free_cpumask_var(cs->cpus_allowed);
1898    kfree(cs);
1899}
1900
1901struct cgroup_subsys cpuset_subsys = {
1902    .name = "cpuset",
1903    .create = cpuset_create,
1904    .destroy = cpuset_destroy,
1905    .can_attach = cpuset_can_attach,
1906    .can_attach_task = cpuset_can_attach_task,
1907    .pre_attach = cpuset_pre_attach,
1908    .attach_task = cpuset_attach_task,
1909    .attach = cpuset_attach,
1910    .populate = cpuset_populate,
1911    .post_clone = cpuset_post_clone,
1912    .subsys_id = cpuset_subsys_id,
1913    .early_init = 1,
1914};
1915
1916/**
1917 * cpuset_init - initialize cpusets at system boot
1918 *
1919 * Description: Initialize top_cpuset and the cpuset internal file system,
1920 **/
1921
1922int __init cpuset_init(void)
1923{
1924    int err = 0;
1925
1926    if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1927        BUG();
1928
1929    cpumask_setall(top_cpuset.cpus_allowed);
1930    nodes_setall(top_cpuset.mems_allowed);
1931
1932    fmeter_init(&top_cpuset.fmeter);
1933    set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1934    top_cpuset.relax_domain_level = -1;
1935
1936    err = register_filesystem(&cpuset_fs_type);
1937    if (err < 0)
1938        return err;
1939
1940    if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1941        BUG();
1942
1943    number_of_cpusets = 1;
1944    return 0;
1945}
1946
1947/**
1948 * cpuset_do_move_task - move a given task to another cpuset
1949 * @tsk: pointer to task_struct the task to move
1950 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1951 *
1952 * Called by cgroup_scan_tasks() for each task in a cgroup.
1953 * Return nonzero to stop the walk through the tasks.
1954 */
1955static void cpuset_do_move_task(struct task_struct *tsk,
1956                struct cgroup_scanner *scan)
1957{
1958    struct cgroup *new_cgroup = scan->data;
1959
1960    cgroup_attach_task(new_cgroup, tsk);
1961}
1962
1963/**
1964 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1965 * @from: cpuset in which the tasks currently reside
1966 * @to: cpuset to which the tasks will be moved
1967 *
1968 * Called with cgroup_mutex held
1969 * callback_mutex must not be held, as cpuset_attach() will take it.
1970 *
1971 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1972 * calling callback functions for each.
1973 */
1974static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1975{
1976    struct cgroup_scanner scan;
1977
1978    scan.cg = from->css.cgroup;
1979    scan.test_task = NULL; /* select all tasks in cgroup */
1980    scan.process_task = cpuset_do_move_task;
1981    scan.heap = NULL;
1982    scan.data = to->css.cgroup;
1983
1984    if (cgroup_scan_tasks(&scan))
1985        printk(KERN_ERR "move_member_tasks_to_cpuset: "
1986                "cgroup_scan_tasks failed\n");
1987}
1988
1989/*
1990 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1991 * or memory nodes, we need to walk over the cpuset hierarchy,
1992 * removing that CPU or node from all cpusets. If this removes the
1993 * last CPU or node from a cpuset, then move the tasks in the empty
1994 * cpuset to its next-highest non-empty parent.
1995 *
1996 * Called with cgroup_mutex held
1997 * callback_mutex must not be held, as cpuset_attach() will take it.
1998 */
1999static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2000{
2001    struct cpuset *parent;
2002
2003    /*
2004     * The cgroup's css_sets list is in use if there are tasks
2005     * in the cpuset; the list is empty if there are none;
2006     * the cs->css.refcnt seems always 0.
2007     */
2008    if (list_empty(&cs->css.cgroup->css_sets))
2009        return;
2010
2011    /*
2012     * Find its next-highest non-empty parent, (top cpuset
2013     * has online cpus, so can't be empty).
2014     */
2015    parent = cs->parent;
2016    while (cpumask_empty(parent->cpus_allowed) ||
2017            nodes_empty(parent->mems_allowed))
2018        parent = parent->parent;
2019
2020    move_member_tasks_to_cpuset(cs, parent);
2021}
2022
2023/*
2024 * Walk the specified cpuset subtree and look for empty cpusets.
2025 * The tasks of such cpuset must be moved to a parent cpuset.
2026 *
2027 * Called with cgroup_mutex held. We take callback_mutex to modify
2028 * cpus_allowed and mems_allowed.
2029 *
2030 * This walk processes the tree from top to bottom, completing one layer
2031 * before dropping down to the next. It always processes a node before
2032 * any of its children.
2033 *
2034 * For now, since we lack memory hot unplug, we'll never see a cpuset
2035 * that has tasks along with an empty 'mems'. But if we did see such
2036 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2037 */
2038static void scan_for_empty_cpusets(struct cpuset *root)
2039{
2040    LIST_HEAD(queue);
2041    struct cpuset *cp; /* scans cpusets being updated */
2042    struct cpuset *child; /* scans child cpusets of cp */
2043    struct cgroup *cont;
2044    static nodemask_t oldmems; /* protected by cgroup_mutex */
2045
2046    list_add_tail((struct list_head *)&root->stack_list, &queue);
2047
2048    while (!list_empty(&queue)) {
2049        cp = list_first_entry(&queue, struct cpuset, stack_list);
2050        list_del(queue.next);
2051        list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2052            child = cgroup_cs(cont);
2053            list_add_tail(&child->stack_list, &queue);
2054        }
2055
2056        /* Continue past cpusets with all cpus, mems online */
2057        if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
2058            nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2059            continue;
2060
2061        oldmems = cp->mems_allowed;
2062
2063        /* Remove offline cpus and mems from this cpuset. */
2064        mutex_lock(&callback_mutex);
2065        cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2066                cpu_active_mask);
2067        nodes_and(cp->mems_allowed, cp->mems_allowed,
2068                        node_states[N_HIGH_MEMORY]);
2069        mutex_unlock(&callback_mutex);
2070
2071        /* Move tasks from the empty cpuset to a parent */
2072        if (cpumask_empty(cp->cpus_allowed) ||
2073             nodes_empty(cp->mems_allowed))
2074            remove_tasks_in_empty_cpuset(cp);
2075        else {
2076            update_tasks_cpumask(cp, NULL);
2077            update_tasks_nodemask(cp, &oldmems, NULL);
2078        }
2079    }
2080}
2081
2082/*
2083 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2084 * period. This is necessary in order to make cpusets transparent
2085 * (of no affect) on systems that are actively using CPU hotplug
2086 * but making no active use of cpusets.
2087 *
2088 * This routine ensures that top_cpuset.cpus_allowed tracks
2089 * cpu_active_mask on each CPU hotplug (cpuhp) event.
2090 *
2091 * Called within get_online_cpus(). Needs to call cgroup_lock()
2092 * before calling generate_sched_domains().
2093 */
2094void cpuset_update_active_cpus(void)
2095{
2096    struct sched_domain_attr *attr;
2097    cpumask_var_t *doms;
2098    int ndoms;
2099
2100    cgroup_lock();
2101    mutex_lock(&callback_mutex);
2102    cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2103    mutex_unlock(&callback_mutex);
2104    scan_for_empty_cpusets(&top_cpuset);
2105    ndoms = generate_sched_domains(&doms, &attr);
2106    cgroup_unlock();
2107
2108    /* Have scheduler rebuild the domains */
2109    partition_sched_domains(ndoms, doms, attr);
2110}
2111
2112#ifdef CONFIG_MEMORY_HOTPLUG
2113/*
2114 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2115 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2116 * See also the previous routine cpuset_track_online_cpus().
2117 */
2118static int cpuset_track_online_nodes(struct notifier_block *self,
2119                unsigned long action, void *arg)
2120{
2121    static nodemask_t oldmems; /* protected by cgroup_mutex */
2122
2123    cgroup_lock();
2124    switch (action) {
2125    case MEM_ONLINE:
2126        oldmems = top_cpuset.mems_allowed;
2127        mutex_lock(&callback_mutex);
2128        top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2129        mutex_unlock(&callback_mutex);
2130        update_tasks_nodemask(&top_cpuset, &oldmems, NULL);
2131        break;
2132    case MEM_OFFLINE:
2133        /*
2134         * needn't update top_cpuset.mems_allowed explicitly because
2135         * scan_for_empty_cpusets() will update it.
2136         */
2137        scan_for_empty_cpusets(&top_cpuset);
2138        break;
2139    default:
2140        break;
2141    }
2142    cgroup_unlock();
2143
2144    return NOTIFY_OK;
2145}
2146#endif
2147
2148/**
2149 * cpuset_init_smp - initialize cpus_allowed
2150 *
2151 * Description: Finish top cpuset after cpu, node maps are initialized
2152 **/
2153
2154void __init cpuset_init_smp(void)
2155{
2156    cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2157    top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2158
2159    hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2160
2161    cpuset_wq = create_singlethread_workqueue("cpuset");
2162    BUG_ON(!cpuset_wq);
2163}
2164
2165/**
2166 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2167 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2168 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2169 *
2170 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2171 * attached to the specified @tsk. Guaranteed to return some non-empty
2172 * subset of cpu_online_map, even if this means going outside the
2173 * tasks cpuset.
2174 **/
2175
2176void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2177{
2178    mutex_lock(&callback_mutex);
2179    task_lock(tsk);
2180    guarantee_online_cpus(task_cs(tsk), pmask);
2181    task_unlock(tsk);
2182    mutex_unlock(&callback_mutex);
2183}
2184
2185int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2186{
2187    const struct cpuset *cs;
2188    int cpu;
2189
2190    rcu_read_lock();
2191    cs = task_cs(tsk);
2192    if (cs)
2193        do_set_cpus_allowed(tsk, cs->cpus_allowed);
2194    rcu_read_unlock();
2195
2196    /*
2197     * We own tsk->cpus_allowed, nobody can change it under us.
2198     *
2199     * But we used cs && cs->cpus_allowed lockless and thus can
2200     * race with cgroup_attach_task() or update_cpumask() and get
2201     * the wrong tsk->cpus_allowed. However, both cases imply the
2202     * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2203     * which takes task_rq_lock().
2204     *
2205     * If we are called after it dropped the lock we must see all
2206     * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2207     * set any mask even if it is not right from task_cs() pov,
2208     * the pending set_cpus_allowed_ptr() will fix things.
2209     */
2210
2211    cpu = cpumask_any_and(&tsk->cpus_allowed, cpu_active_mask);
2212    if (cpu >= nr_cpu_ids) {
2213        /*
2214         * Either tsk->cpus_allowed is wrong (see above) or it
2215         * is actually empty. The latter case is only possible
2216         * if we are racing with remove_tasks_in_empty_cpuset().
2217         * Like above we can temporary set any mask and rely on
2218         * set_cpus_allowed_ptr() as synchronization point.
2219         */
2220        do_set_cpus_allowed(tsk, cpu_possible_mask);
2221        cpu = cpumask_any(cpu_active_mask);
2222    }
2223
2224    return cpu;
2225}
2226
2227void cpuset_init_current_mems_allowed(void)
2228{
2229    nodes_setall(current->mems_allowed);
2230}
2231
2232/**
2233 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2234 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2235 *
2236 * Description: Returns the nodemask_t mems_allowed of the cpuset
2237 * attached to the specified @tsk. Guaranteed to return some non-empty
2238 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2239 * tasks cpuset.
2240 **/
2241
2242nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2243{
2244    nodemask_t mask;
2245
2246    mutex_lock(&callback_mutex);
2247    task_lock(tsk);
2248    guarantee_online_mems(task_cs(tsk), &mask);
2249    task_unlock(tsk);
2250    mutex_unlock(&callback_mutex);
2251
2252    return mask;
2253}
2254
2255/**
2256 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2257 * @nodemask: the nodemask to be checked
2258 *
2259 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2260 */
2261int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2262{
2263    return nodes_intersects(*nodemask, current->mems_allowed);
2264}
2265
2266/*
2267 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2268 * mem_hardwall ancestor to the specified cpuset. Call holding
2269 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2270 * (an unusual configuration), then returns the root cpuset.
2271 */
2272static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2273{
2274    while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2275        cs = cs->parent;
2276    return cs;
2277}
2278
2279/**
2280 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2281 * @node: is this an allowed node?
2282 * @gfp_mask: memory allocation flags
2283 *
2284 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2285 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2286 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2287 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2288 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2289 * flag, yes.
2290 * Otherwise, no.
2291 *
2292 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2293 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2294 * might sleep, and might allow a node from an enclosing cpuset.
2295 *
2296 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2297 * cpusets, and never sleeps.
2298 *
2299 * The __GFP_THISNODE placement logic is really handled elsewhere,
2300 * by forcibly using a zonelist starting at a specified node, and by
2301 * (in get_page_from_freelist()) refusing to consider the zones for
2302 * any node on the zonelist except the first. By the time any such
2303 * calls get to this routine, we should just shut up and say 'yes'.
2304 *
2305 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2306 * and do not allow allocations outside the current tasks cpuset
2307 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2308 * GFP_KERNEL allocations are not so marked, so can escape to the
2309 * nearest enclosing hardwalled ancestor cpuset.
2310 *
2311 * Scanning up parent cpusets requires callback_mutex. The
2312 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2313 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2314 * current tasks mems_allowed came up empty on the first pass over
2315 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2316 * cpuset are short of memory, might require taking the callback_mutex
2317 * mutex.
2318 *
2319 * The first call here from mm/page_alloc:get_page_from_freelist()
2320 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2321 * so no allocation on a node outside the cpuset is allowed (unless
2322 * in interrupt, of course).
2323 *
2324 * The second pass through get_page_from_freelist() doesn't even call
2325 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2326 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2327 * in alloc_flags. That logic and the checks below have the combined
2328 * affect that:
2329 * in_interrupt - any node ok (current task context irrelevant)
2330 * GFP_ATOMIC - any node ok
2331 * TIF_MEMDIE - any node ok
2332 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2333 * GFP_USER - only nodes in current tasks mems allowed ok.
2334 *
2335 * Rule:
2336 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2337 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2338 * the code that might scan up ancestor cpusets and sleep.
2339 */
2340int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2341{
2342    const struct cpuset *cs; /* current cpuset ancestors */
2343    int allowed; /* is allocation in zone z allowed? */
2344
2345    if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2346        return 1;
2347    might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2348    if (node_isset(node, current->mems_allowed))
2349        return 1;
2350    /*
2351     * Allow tasks that have access to memory reserves because they have
2352     * been OOM killed to get memory anywhere.
2353     */
2354    if (unlikely(test_thread_flag(TIF_MEMDIE)))
2355        return 1;
2356    if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2357        return 0;
2358
2359    if (current->flags & PF_EXITING) /* Let dying task have memory */
2360        return 1;
2361
2362    /* Not hardwall and node outside mems_allowed: scan up cpusets */
2363    mutex_lock(&callback_mutex);
2364
2365    task_lock(current);
2366    cs = nearest_hardwall_ancestor(task_cs(current));
2367    task_unlock(current);
2368
2369    allowed = node_isset(node, cs->mems_allowed);
2370    mutex_unlock(&callback_mutex);
2371    return allowed;
2372}
2373
2374/*
2375 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2376 * @node: is this an allowed node?
2377 * @gfp_mask: memory allocation flags
2378 *
2379 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2380 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2381 * yes. If the task has been OOM killed and has access to memory reserves as
2382 * specified by the TIF_MEMDIE flag, yes.
2383 * Otherwise, no.
2384 *
2385 * The __GFP_THISNODE placement logic is really handled elsewhere,
2386 * by forcibly using a zonelist starting at a specified node, and by
2387 * (in get_page_from_freelist()) refusing to consider the zones for
2388 * any node on the zonelist except the first. By the time any such
2389 * calls get to this routine, we should just shut up and say 'yes'.
2390 *
2391 * Unlike the cpuset_node_allowed_softwall() variant, above,
2392 * this variant requires that the node be in the current task's
2393 * mems_allowed or that we're in interrupt. It does not scan up the
2394 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2395 * It never sleeps.
2396 */
2397int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2398{
2399    if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2400        return 1;
2401    if (node_isset(node, current->mems_allowed))
2402        return 1;
2403    /*
2404     * Allow tasks that have access to memory reserves because they have
2405     * been OOM killed to get memory anywhere.
2406     */
2407    if (unlikely(test_thread_flag(TIF_MEMDIE)))
2408        return 1;
2409    return 0;
2410}
2411
2412/**
2413 * cpuset_unlock - release lock on cpuset changes
2414 *
2415 * Undo the lock taken in a previous cpuset_lock() call.
2416 */
2417
2418void cpuset_unlock(void)
2419{
2420    mutex_unlock(&callback_mutex);
2421}
2422
2423/**
2424 * cpuset_mem_spread_node() - On which node to begin search for a file page
2425 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2426 *
2427 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2428 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2429 * and if the memory allocation used cpuset_mem_spread_node()
2430 * to determine on which node to start looking, as it will for
2431 * certain page cache or slab cache pages such as used for file
2432 * system buffers and inode caches, then instead of starting on the
2433 * local node to look for a free page, rather spread the starting
2434 * node around the tasks mems_allowed nodes.
2435 *
2436 * We don't have to worry about the returned node being offline
2437 * because "it can't happen", and even if it did, it would be ok.
2438 *
2439 * The routines calling guarantee_online_mems() are careful to
2440 * only set nodes in task->mems_allowed that are online. So it
2441 * should not be possible for the following code to return an
2442 * offline node. But if it did, that would be ok, as this routine
2443 * is not returning the node where the allocation must be, only
2444 * the node where the search should start. The zonelist passed to
2445 * __alloc_pages() will include all nodes. If the slab allocator
2446 * is passed an offline node, it will fall back to the local node.
2447 * See kmem_cache_alloc_node().
2448 */
2449
2450static int cpuset_spread_node(int *rotor)
2451{
2452    int node;
2453
2454    node = next_node(*rotor, current->mems_allowed);
2455    if (node == MAX_NUMNODES)
2456        node = first_node(current->mems_allowed);
2457    *rotor = node;
2458    return node;
2459}
2460
2461int cpuset_mem_spread_node(void)
2462{
2463    return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2464}
2465
2466int cpuset_slab_spread_node(void)
2467{
2468    return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2469}
2470
2471EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2472
2473/**
2474 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2475 * @tsk1: pointer to task_struct of some task.
2476 * @tsk2: pointer to task_struct of some other task.
2477 *
2478 * Description: Return true if @tsk1's mems_allowed intersects the
2479 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2480 * one of the task's memory usage might impact the memory available
2481 * to the other.
2482 **/
2483
2484int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2485                   const struct task_struct *tsk2)
2486{
2487    return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2488}
2489
2490/**
2491 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2492 * @task: pointer to task_struct of some task.
2493 *
2494 * Description: Prints @task's name, cpuset name, and cached copy of its
2495 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2496 * dereferencing task_cs(task).
2497 */
2498void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2499{
2500    struct dentry *dentry;
2501
2502    dentry = task_cs(tsk)->css.cgroup->dentry;
2503    spin_lock(&cpuset_buffer_lock);
2504    snprintf(cpuset_name, CPUSET_NAME_LEN,
2505         dentry ? (const char *)dentry->d_name.name : "/");
2506    nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2507               tsk->mems_allowed);
2508    printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2509           tsk->comm, cpuset_name, cpuset_nodelist);
2510    spin_unlock(&cpuset_buffer_lock);
2511}
2512
2513/*
2514 * Collection of memory_pressure is suppressed unless
2515 * this flag is enabled by writing "1" to the special
2516 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2517 */
2518
2519int cpuset_memory_pressure_enabled __read_mostly;
2520
2521/**
2522 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2523 *
2524 * Keep a running average of the rate of synchronous (direct)
2525 * page reclaim efforts initiated by tasks in each cpuset.
2526 *
2527 * This represents the rate at which some task in the cpuset
2528 * ran low on memory on all nodes it was allowed to use, and
2529 * had to enter the kernels page reclaim code in an effort to
2530 * create more free memory by tossing clean pages or swapping
2531 * or writing dirty pages.
2532 *
2533 * Display to user space in the per-cpuset read-only file
2534 * "memory_pressure". Value displayed is an integer
2535 * representing the recent rate of entry into the synchronous
2536 * (direct) page reclaim by any task attached to the cpuset.
2537 **/
2538
2539void __cpuset_memory_pressure_bump(void)
2540{
2541    task_lock(current);
2542    fmeter_markevent(&task_cs(current)->fmeter);
2543    task_unlock(current);
2544}
2545
2546#ifdef CONFIG_PROC_PID_CPUSET
2547/*
2548 * proc_cpuset_show()
2549 * - Print tasks cpuset path into seq_file.
2550 * - Used for /proc/<pid>/cpuset.
2551 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2552 * doesn't really matter if tsk->cpuset changes after we read it,
2553 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2554 * anyway.
2555 */
2556static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2557{
2558    struct pid *pid;
2559    struct task_struct *tsk;
2560    char *buf;
2561    struct cgroup_subsys_state *css;
2562    int retval;
2563
2564    retval = -ENOMEM;
2565    buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2566    if (!buf)
2567        goto out;
2568
2569    retval = -ESRCH;
2570    pid = m->private;
2571    tsk = get_pid_task(pid, PIDTYPE_PID);
2572    if (!tsk)
2573        goto out_free;
2574
2575    retval = -EINVAL;
2576    cgroup_lock();
2577    css = task_subsys_state(tsk, cpuset_subsys_id);
2578    retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2579    if (retval < 0)
2580        goto out_unlock;
2581    seq_puts(m, buf);
2582    seq_putc(m, '\n');
2583out_unlock:
2584    cgroup_unlock();
2585    put_task_struct(tsk);
2586out_free:
2587    kfree(buf);
2588out:
2589    return retval;
2590}
2591
2592static int cpuset_open(struct inode *inode, struct file *file)
2593{
2594    struct pid *pid = PROC_I(inode)->pid;
2595    return single_open(file, proc_cpuset_show, pid);
2596}
2597
2598const struct file_operations proc_cpuset_operations = {
2599    .open = cpuset_open,
2600    .read = seq_read,
2601    .llseek = seq_lseek,
2602    .release = single_release,
2603};
2604#endif /* CONFIG_PROC_PID_CPUSET */
2605
2606/* Display task mems_allowed in /proc/<pid>/status file. */
2607void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2608{
2609    seq_printf(m, "Mems_allowed:\t");
2610    seq_nodemask(m, &task->mems_allowed);
2611    seq_printf(m, "\n");
2612    seq_printf(m, "Mems_allowed_list:\t");
2613    seq_nodemask_list(m, &task->mems_allowed);
2614    seq_printf(m, "\n");
2615}
2616

Archive Download this file



interactive