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

Archive Download this file



interactive