]> bbs.cooldavid.org Git - net-next-2.6.git/blame - kernel/cpuset.c
xps: Transmit Packet Steering
[net-next-2.6.git] / kernel / cpuset.c
CommitLineData
1da177e4
LT
1/*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
029190c5 7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8793d854 8 * Copyright (C) 2006 Google, Inc
1da177e4
LT
9 *
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
1da177e4 12 *
825a46af 13 * 2003-10-10 Written by Simon Derr.
1da177e4 14 * 2003-10-22 Updates by Stephen Hemminger.
825a46af 15 * 2004 May-July Rework by Paul Jackson.
8793d854 16 * 2006 Rework by Paul Menage to use generic cgroups
cf417141
MK
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
18 * by Max Krasnyansky
1da177e4
LT
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
1da177e4
LT
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>
68860ec1 37#include <linux/mempolicy.h>
1da177e4 38#include <linux/mm.h>
f481891f 39#include <linux/memory.h>
1da177e4
LT
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>
6b9c2603 45#include <linux/rcupdate.h>
1da177e4
LT
46#include <linux/sched.h>
47#include <linux/seq_file.h>
22fb52dd 48#include <linux/security.h>
1da177e4 49#include <linux/slab.h>
1da177e4
LT
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>
3d3f26a7 59#include <linux/mutex.h>
956db3ca
CW
60#include <linux/workqueue.h>
61#include <linux/cgroup.h>
1da177e4 62
f90d4118
MX
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
202f72d5
PJ
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 */
7edc5962 76int number_of_cpusets __read_mostly;
202f72d5 77
2df167a3 78/* Forward declare cgroup structures */
8793d854
PM
79struct cgroup_subsys cpuset_subsys;
80struct cpuset;
81
3e0d98b9
PJ
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
1da177e4 91struct cpuset {
8793d854
PM
92 struct cgroup_subsys_state css;
93
1da177e4 94 unsigned long flags; /* "unsigned long" so bitops work */
300ed6cb 95 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
1da177e4
LT
96 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
97
1da177e4 98 struct cpuset *parent; /* my parent */
1da177e4 99
3e0d98b9 100 struct fmeter fmeter; /* memory_pressure filter */
029190c5
PJ
101
102 /* partition number for rebuild_sched_domains() */
103 int pn;
956db3ca 104
1d3504fc
HS
105 /* for custom sched domain */
106 int relax_domain_level;
107
732bee7a 108 /* used for walking a cpuset hierarchy */
956db3ca 109 struct list_head stack_list;
1da177e4
LT
110};
111
8793d854
PM
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}
8793d854 125
1da177e4
LT
126/* bits in struct cpuset flags field */
127typedef enum {
128 CS_CPU_EXCLUSIVE,
129 CS_MEM_EXCLUSIVE,
78608366 130 CS_MEM_HARDWALL,
45b07ef3 131 CS_MEMORY_MIGRATE,
029190c5 132 CS_SCHED_LOAD_BALANCE,
825a46af
PJ
133 CS_SPREAD_PAGE,
134 CS_SPREAD_SLAB,
1da177e4
LT
135} cpuset_flagbits_t;
136
137/* convenient tests for these bits */
138static inline int is_cpu_exclusive(const struct cpuset *cs)
139{
7b5b9ef0 140 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
1da177e4
LT
141}
142
143static inline int is_mem_exclusive(const struct cpuset *cs)
144{
7b5b9ef0 145 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
1da177e4
LT
146}
147
78608366
PM
148static inline int is_mem_hardwall(const struct cpuset *cs)
149{
150 return test_bit(CS_MEM_HARDWALL, &cs->flags);
151}
152
029190c5
PJ
153static inline int is_sched_load_balance(const struct cpuset *cs)
154{
155 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
156}
157
45b07ef3
PJ
158static inline int is_memory_migrate(const struct cpuset *cs)
159{
7b5b9ef0 160 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
45b07ef3
PJ
161}
162
825a46af
PJ
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
1da177e4
LT
173static struct cpuset top_cpuset = {
174 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
1da177e4
LT
175};
176
1da177e4 177/*
2df167a3
PM
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.
053199ed 185 *
3d3f26a7 186 * A task must hold both mutexes to modify cpusets. If a task
2df167a3 187 * holds cgroup_mutex, then it blocks others wanting that mutex,
3d3f26a7 188 * ensuring that it is the only task able to also acquire callback_mutex
053199ed
PJ
189 * and be able to modify cpusets. It can perform various checks on
190 * the cpuset structure first, knowing nothing will change. It can
2df167a3 191 * also allocate memory while just holding cgroup_mutex. While it is
053199ed 192 * performing these checks, various callback routines can briefly
3d3f26a7
IM
193 * acquire callback_mutex to query cpusets. Once it is ready to make
194 * the changes, it takes callback_mutex, blocking everyone else.
053199ed
PJ
195 *
196 * Calls to the kernel memory allocator can not be made while holding
3d3f26a7 197 * callback_mutex, as that would risk double tripping on callback_mutex
053199ed
PJ
198 * from one of the callbacks into the cpuset code from within
199 * __alloc_pages().
200 *
3d3f26a7 201 * If a task is only holding callback_mutex, then it has read-only
053199ed
PJ
202 * access to cpusets.
203 *
58568d2a
MX
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.
053199ed 207 *
3d3f26a7 208 * The cpuset_common_file_read() handlers only hold callback_mutex across
053199ed
PJ
209 * small pieces of code, such as when reading out possibly multi-word
210 * cpumasks and nodemasks.
211 *
2df167a3
PM
212 * Accessing a task's cpuset should be done in accordance with the
213 * guidelines for accessing subsystem state in kernel/cgroup.c
1da177e4
LT
214 */
215
3d3f26a7 216static DEFINE_MUTEX(callback_mutex);
4247bdc6 217
75aa1994
DR
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
cf417141
MK
229/*
230 * This is ugly, but preserves the userspace API for existing cpuset
8793d854 231 * users. If someone tries to mount the "cpuset" filesystem, we
cf417141
MK
232 * silently switch it to mount "cgroup" instead
233 */
f7e83571
AV
234static struct dentry *cpuset_mount(struct file_system_type *fs_type,
235 int flags, const char *unused_dev_name, void *data)
1da177e4 236{
8793d854 237 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
f7e83571 238 struct dentry *ret = ERR_PTR(-ENODEV);
8793d854
PM
239 if (cgroup_fs) {
240 char mountopts[] =
241 "cpuset,noprefix,"
242 "release_agent=/sbin/cpuset_release_agent";
f7e83571
AV
243 ret = cgroup_fs->mount(cgroup_fs, flags,
244 unused_dev_name, mountopts);
8793d854
PM
245 put_filesystem(cgroup_fs);
246 }
247 return ret;
1da177e4
LT
248}
249
250static struct file_system_type cpuset_fs_type = {
251 .name = "cpuset",
f7e83571 252 .mount = cpuset_mount,
1da177e4
LT
253};
254
1da177e4 255/*
300ed6cb 256 * Return in pmask the portion of a cpusets's cpus_allowed that
1da177e4
LT
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 *
3d3f26a7 266 * Call with callback_mutex held.
1da177e4
LT
267 */
268
6af866af
LZ
269static void guarantee_online_cpus(const struct cpuset *cs,
270 struct cpumask *pmask)
1da177e4 271{
300ed6cb 272 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
1da177e4
LT
273 cs = cs->parent;
274 if (cs)
300ed6cb 275 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
1da177e4 276 else
300ed6cb
LZ
277 cpumask_copy(pmask, cpu_online_mask);
278 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
1da177e4
LT
279}
280
281/*
282 * Return in *pmask the portion of a cpusets's mems_allowed that
0e1e7c7a
CL
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].
1da177e4
LT
287 *
288 * One way or another, we guarantee to return some non-empty subset
0e1e7c7a 289 * of node_states[N_HIGH_MEMORY].
1da177e4 290 *
3d3f26a7 291 * Call with callback_mutex held.
1da177e4
LT
292 */
293
294static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
295{
0e1e7c7a
CL
296 while (cs && !nodes_intersects(cs->mems_allowed,
297 node_states[N_HIGH_MEMORY]))
1da177e4
LT
298 cs = cs->parent;
299 if (cs)
0e1e7c7a
CL
300 nodes_and(*pmask, cs->mems_allowed,
301 node_states[N_HIGH_MEMORY]);
1da177e4 302 else
0e1e7c7a
CL
303 *pmask = node_states[N_HIGH_MEMORY];
304 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
1da177e4
LT
305}
306
f3b39d47
MX
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
1da177e4
LT
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
2df167a3 330 * are only set if the other's are set. Call holding cgroup_mutex.
1da177e4
LT
331 */
332
333static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
334{
300ed6cb 335 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
1da177e4
LT
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
645fcc9d
LZ
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{
300ed6cb
LZ
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;
645fcc9d
LZ
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{
300ed6cb 368 free_cpumask_var(trial->cpus_allowed);
645fcc9d
LZ
369 kfree(trial);
370}
371
1da177e4
LT
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
2df167a3 379 * cgroup_mutex held.
1da177e4
LT
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{
8793d854 394 struct cgroup *cont;
1da177e4
LT
395 struct cpuset *c, *par;
396
397 /* Each of our child cpusets must be a subset of us */
8793d854
PM
398 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
399 if (!is_cpuset_subset(cgroup_cs(cont), trial))
1da177e4
LT
400 return -EBUSY;
401 }
402
403 /* Remaining checks don't apply to root cpuset */
69604067 404 if (cur == &top_cpuset)
1da177e4
LT
405 return 0;
406
69604067
PJ
407 par = cur->parent;
408
1da177e4
LT
409 /* We must be a subset of our parent cpuset */
410 if (!is_cpuset_subset(trial, par))
411 return -EACCES;
412
2df167a3
PM
413 /*
414 * If either I or some sibling (!= me) is exclusive, we can't
415 * overlap
416 */
8793d854
PM
417 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
418 c = cgroup_cs(cont);
1da177e4
LT
419 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
420 c != cur &&
300ed6cb 421 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
1da177e4
LT
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
020958b6
PJ
429 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
430 if (cgroup_task_count(cur->css.cgroup)) {
300ed6cb 431 if (cpumask_empty(trial->cpus_allowed) ||
020958b6
PJ
432 nodes_empty(trial->mems_allowed)) {
433 return -ENOSPC;
434 }
435 }
436
1da177e4
LT
437 return 0;
438}
439
db7f47cf 440#ifdef CONFIG_SMP
029190c5 441/*
cf417141 442 * Helper routine for generate_sched_domains().
029190c5
PJ
443 * Do cpusets a, b have overlapping cpus_allowed masks?
444 */
029190c5
PJ
445static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
446{
300ed6cb 447 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
029190c5
PJ
448}
449
1d3504fc
HS
450static void
451update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
452{
1d3504fc
HS
453 if (dattr->relax_domain_level < c->relax_domain_level)
454 dattr->relax_domain_level = c->relax_domain_level;
455 return;
456}
457
f5393693
LJ
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
300ed6cb 472 if (cpumask_empty(cp->cpus_allowed))
f5393693
LJ
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
029190c5 485/*
cf417141
MK
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.
029190c5 495 *
45ce80fb 496 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
029190c5
PJ
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 *
cf417141 504 * Must be called with cgroup_lock held.
029190c5
PJ
505 *
506 * The three key local variables below are:
aeed6824 507 * q - a linked-list queue of cpuset pointers, used to implement a
029190c5
PJ
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 */
acc3f5d7 539static int generate_sched_domains(cpumask_var_t **domains,
cf417141 540 struct sched_domain_attr **attributes)
029190c5 541{
cf417141 542 LIST_HEAD(q); /* queue of cpusets to be scanned */
029190c5
PJ
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 */
acc3f5d7 547 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
1d3504fc 548 struct sched_domain_attr *dattr; /* attributes for custom domains */
1583715d 549 int ndoms = 0; /* number of sched domains in result */
6af866af 550 int nslot; /* next empty doms[] struct cpumask slot */
029190c5 551
029190c5 552 doms = NULL;
1d3504fc 553 dattr = NULL;
cf417141 554 csa = NULL;
029190c5
PJ
555
556 /* Special case for the 99% of systems with one, full, sched domain */
557 if (is_sched_load_balance(&top_cpuset)) {
acc3f5d7
RR
558 ndoms = 1;
559 doms = alloc_sched_domains(ndoms);
029190c5 560 if (!doms)
cf417141
MK
561 goto done;
562
1d3504fc
HS
563 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
564 if (dattr) {
565 *dattr = SD_ATTR_INIT;
93a65575 566 update_domain_attr_tree(dattr, &top_cpuset);
1d3504fc 567 }
acc3f5d7 568 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
cf417141 569
cf417141 570 goto done;
029190c5
PJ
571 }
572
029190c5
PJ
573 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
574 if (!csa)
575 goto done;
576 csn = 0;
577
aeed6824
LZ
578 list_add(&top_cpuset.stack_list, &q);
579 while (!list_empty(&q)) {
029190c5
PJ
580 struct cgroup *cont;
581 struct cpuset *child; /* scans child cpusets of cp */
489a5393 582
aeed6824
LZ
583 cp = list_first_entry(&q, struct cpuset, stack_list);
584 list_del(q.next);
585
300ed6cb 586 if (cpumask_empty(cp->cpus_allowed))
489a5393
LJ
587 continue;
588
f5393693
LJ
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)) {
029190c5 596 csa[csn++] = cp;
f5393693
LJ
597 continue;
598 }
489a5393 599
029190c5
PJ
600 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
601 child = cgroup_cs(cont);
aeed6824 602 list_add_tail(&child->stack_list, &q);
029190c5
PJ
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
cf417141
MK
633 /*
634 * Now we know how many domains to create.
635 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
636 */
acc3f5d7 637 doms = alloc_sched_domains(ndoms);
700018e0 638 if (!doms)
cf417141 639 goto done;
cf417141
MK
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 */
1d3504fc 645 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
029190c5
PJ
646
647 for (nslot = 0, i = 0; i < csn; i++) {
648 struct cpuset *a = csa[i];
6af866af 649 struct cpumask *dp;
029190c5
PJ
650 int apn = a->pn;
651
cf417141
MK
652 if (apn < 0) {
653 /* Skip completed partitions */
654 continue;
655 }
656
acc3f5d7 657 dp = doms[nslot];
cf417141
MK
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--;
029190c5 668 }
cf417141
MK
669 continue;
670 }
029190c5 671
6af866af 672 cpumask_clear(dp);
cf417141
MK
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) {
300ed6cb 679 cpumask_or(dp, dp, b->cpus_allowed);
cf417141
MK
680 if (dattr)
681 update_domain_attr_tree(dattr + nslot, b);
682
683 /* Done with this partition */
684 b->pn = -1;
029190c5 685 }
029190c5 686 }
cf417141 687 nslot++;
029190c5
PJ
688 }
689 BUG_ON(nslot != ndoms);
690
cf417141
MK
691done:
692 kfree(csa);
693
700018e0
LZ
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
cf417141
MK
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;
acc3f5d7 719 cpumask_var_t *doms;
cf417141
MK
720 int ndoms;
721
86ef5c9a 722 get_online_cpus();
cf417141
MK
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
86ef5c9a 732 put_online_cpus();
cf417141 733}
db7f47cf
PM
734#else /* !CONFIG_SMP */
735static void do_rebuild_sched_domains(struct work_struct *unused)
736{
737}
738
e1b8090b 739static int generate_sched_domains(cpumask_var_t **domains,
db7f47cf
PM
740 struct sched_domain_attr **attributes)
741{
742 *domains = NULL;
743 return 1;
744}
745#endif /* CONFIG_SMP */
029190c5 746
cf417141
MK
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{
f90d4118 770 queue_work(cpuset_wq, &rebuild_sched_domains_work);
cf417141
MK
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);
029190c5
PJ
785}
786
58f4790b
CW
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 *
2df167a3 792 * Call with cgroup_mutex held. May take callback_mutex during call.
58f4790b
CW
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).
053199ed 796 */
9e0c914c
AB
797static int cpuset_test_cpumask(struct task_struct *tsk,
798 struct cgroup_scanner *scan)
58f4790b 799{
300ed6cb 800 return !cpumask_equal(&tsk->cpus_allowed,
58f4790b
CW
801 (cgroup_cs(scan->cg))->cpus_allowed);
802}
053199ed 803
58f4790b
CW
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 */
9e0c914c
AB
815static void cpuset_change_cpumask(struct task_struct *tsk,
816 struct cgroup_scanner *scan)
58f4790b 817{
300ed6cb 818 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
58f4790b
CW
819}
820
0b2f630a
MX
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
4e74339a 824 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
0b2f630a
MX
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 *
4e74339a
LZ
831 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
832 * if @heap != NULL.
0b2f630a 833 */
4e74339a 834static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
0b2f630a
MX
835{
836 struct cgroup_scanner scan;
0b2f630a
MX
837
838 scan.cg = cs->css.cgroup;
839 scan.test_task = cpuset_test_cpumask;
840 scan.process_task = cpuset_change_cpumask;
4e74339a
LZ
841 scan.heap = heap;
842 cgroup_scan_tasks(&scan);
0b2f630a
MX
843}
844
58f4790b
CW
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 */
645fcc9d
LZ
850static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
851 const char *buf)
1da177e4 852{
4e74339a 853 struct ptr_heap heap;
58f4790b
CW
854 int retval;
855 int is_load_balanced;
1da177e4 856
4c4d50f7
PJ
857 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
858 if (cs == &top_cpuset)
859 return -EACCES;
860
6f7f02e7 861 /*
c8d9c90c 862 * An empty cpus_allowed is ok only if the cpuset has no tasks.
020958b6
PJ
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.
6f7f02e7 866 */
020958b6 867 if (!*buf) {
300ed6cb 868 cpumask_clear(trialcs->cpus_allowed);
6f7f02e7 869 } else {
300ed6cb 870 retval = cpulist_parse(buf, trialcs->cpus_allowed);
6f7f02e7
DR
871 if (retval < 0)
872 return retval;
37340746 873
6ad4c188 874 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
37340746 875 return -EINVAL;
6f7f02e7 876 }
645fcc9d 877 retval = validate_change(cs, trialcs);
85d7b949
DG
878 if (retval < 0)
879 return retval;
029190c5 880
8707d8b8 881 /* Nothing to do if the cpus didn't change */
300ed6cb 882 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
8707d8b8 883 return 0;
58f4790b 884
4e74339a
LZ
885 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
886 if (retval)
887 return retval;
888
645fcc9d 889 is_load_balanced = is_sched_load_balance(trialcs);
029190c5 890
3d3f26a7 891 mutex_lock(&callback_mutex);
300ed6cb 892 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
3d3f26a7 893 mutex_unlock(&callback_mutex);
029190c5 894
8707d8b8
PM
895 /*
896 * Scan tasks in the cpuset, and update the cpumasks of any
58f4790b 897 * that need an update.
8707d8b8 898 */
4e74339a
LZ
899 update_tasks_cpumask(cs, &heap);
900
901 heap_free(&heap);
58f4790b 902
8707d8b8 903 if (is_load_balanced)
cf417141 904 async_rebuild_sched_domains();
85d7b949 905 return 0;
1da177e4
LT
906}
907
e4e364e8
PJ
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 *
2df167a3 916 * Call holding cgroup_mutex, so current's cpuset won't change
c8d9c90c 917 * during this call, as manage_mutex holds off any cpuset_attach()
e4e364e8
PJ
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
2df167a3 920 * our task's cpuset.
e4e364e8 921 *
e4e364e8
PJ
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.
e4e364e8
PJ
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
e4e364e8 933 tsk->mems_allowed = *to;
e4e364e8
PJ
934
935 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
936
8793d854 937 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
e4e364e8
PJ
938}
939
3b6766fe 940/*
58568d2a
MX
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.
58568d2a
MX
948 */
949static void cpuset_change_task_nodemask(struct task_struct *tsk,
950 nodemask_t *newmems)
951{
c0ff7453
MX
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);
58568d2a 963 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
c0ff7453
MX
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);
58568d2a 1002 tsk->mems_allowed = *newmems;
c0ff7453 1003 task_unlock(tsk);
58568d2a
MX
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.
3b6766fe
LZ
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;
53feb297
MX
1018 NODEMASK_ALLOC(nodemask_t, newmems, GFP_KERNEL);
1019
1020 if (!newmems)
1021 return;
58568d2a
MX
1022
1023 cs = cgroup_cs(scan->cg);
53feb297 1024 guarantee_online_mems(cs, newmems);
58568d2a 1025
53feb297 1026 cpuset_change_task_nodemask(p, newmems);
3b6766fe 1027
53feb297
MX
1028 NODEMASK_FREE(newmems);
1029
3b6766fe
LZ
1030 mm = get_task_mm(p);
1031 if (!mm)
1032 return;
1033
3b6766fe
LZ
1034 migrate = is_memory_migrate(cs);
1035
1036 mpol_rebind_mm(mm, &cs->mems_allowed);
1037 if (migrate)
1038 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1039 mmput(mm);
1040}
1041
8793d854
PM
1042static void *cpuset_being_rebound;
1043
0b2f630a
MX
1044/**
1045 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1046 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1047 * @oldmem: old mems_allowed of cpuset cs
010cfac4 1048 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
0b2f630a
MX
1049 *
1050 * Called with cgroup_mutex held
010cfac4
LZ
1051 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1052 * if @heap != NULL.
0b2f630a 1053 */
010cfac4
LZ
1054static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1055 struct ptr_heap *heap)
1da177e4 1056{
3b6766fe 1057 struct cgroup_scanner scan;
59dac16f 1058
846a16bf 1059 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
4225399a 1060
3b6766fe
LZ
1061 scan.cg = cs->css.cgroup;
1062 scan.test_task = NULL;
1063 scan.process_task = cpuset_change_nodemask;
010cfac4 1064 scan.heap = heap;
3b6766fe 1065 scan.data = (nodemask_t *)oldmem;
4225399a
PJ
1066
1067 /*
3b6766fe
LZ
1068 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1069 * take while holding tasklist_lock. Forks can happen - the
1070 * mpol_dup() cpuset_being_rebound check will catch such forks,
1071 * and rebind their vma mempolicies too. Because we still hold
1072 * the global cgroup_mutex, we know that no other rebind effort
1073 * will be contending for the global variable cpuset_being_rebound.
4225399a 1074 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
04c19fa6 1075 * is idempotent. Also migrate pages in each mm to new nodes.
4225399a 1076 */
010cfac4 1077 cgroup_scan_tasks(&scan);
4225399a 1078
2df167a3 1079 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
8793d854 1080 cpuset_being_rebound = NULL;
1da177e4
LT
1081}
1082
0b2f630a
MX
1083/*
1084 * Handle user request to change the 'mems' memory placement
1085 * of a cpuset. Needs to validate the request, update the
58568d2a
MX
1086 * cpusets mems_allowed, and for each task in the cpuset,
1087 * update mems_allowed and rebind task's mempolicy and any vma
1088 * mempolicies and if the cpuset is marked 'memory_migrate',
1089 * migrate the tasks pages to the new memory.
0b2f630a
MX
1090 *
1091 * Call with cgroup_mutex held. May take callback_mutex during call.
1092 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1093 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1094 * their mempolicies to the cpusets new mems_allowed.
1095 */
645fcc9d
LZ
1096static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1097 const char *buf)
0b2f630a 1098{
53feb297 1099 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
0b2f630a 1100 int retval;
010cfac4 1101 struct ptr_heap heap;
0b2f630a 1102
53feb297
MX
1103 if (!oldmem)
1104 return -ENOMEM;
1105
0b2f630a
MX
1106 /*
1107 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1108 * it's read-only
1109 */
53feb297
MX
1110 if (cs == &top_cpuset) {
1111 retval = -EACCES;
1112 goto done;
1113 }
0b2f630a 1114
0b2f630a
MX
1115 /*
1116 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1117 * Since nodelist_parse() fails on an empty mask, we special case
1118 * that parsing. The validate_change() call ensures that cpusets
1119 * with tasks have memory.
1120 */
1121 if (!*buf) {
645fcc9d 1122 nodes_clear(trialcs->mems_allowed);
0b2f630a 1123 } else {
645fcc9d 1124 retval = nodelist_parse(buf, trialcs->mems_allowed);
0b2f630a
MX
1125 if (retval < 0)
1126 goto done;
1127
645fcc9d 1128 if (!nodes_subset(trialcs->mems_allowed,
53feb297
MX
1129 node_states[N_HIGH_MEMORY])) {
1130 retval = -EINVAL;
1131 goto done;
1132 }
0b2f630a 1133 }
53feb297
MX
1134 *oldmem = cs->mems_allowed;
1135 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
0b2f630a
MX
1136 retval = 0; /* Too easy - nothing to do */
1137 goto done;
1138 }
645fcc9d 1139 retval = validate_change(cs, trialcs);
0b2f630a
MX
1140 if (retval < 0)
1141 goto done;
1142
010cfac4
LZ
1143 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1144 if (retval < 0)
1145 goto done;
1146
0b2f630a 1147 mutex_lock(&callback_mutex);
645fcc9d 1148 cs->mems_allowed = trialcs->mems_allowed;
0b2f630a
MX
1149 mutex_unlock(&callback_mutex);
1150
53feb297 1151 update_tasks_nodemask(cs, oldmem, &heap);
010cfac4
LZ
1152
1153 heap_free(&heap);
0b2f630a 1154done:
53feb297 1155 NODEMASK_FREE(oldmem);
0b2f630a
MX
1156 return retval;
1157}
1158
8793d854
PM
1159int current_cpuset_is_being_rebound(void)
1160{
1161 return task_cs(current) == cpuset_being_rebound;
1162}
1163
5be7a479 1164static int update_relax_domain_level(struct cpuset *cs, s64 val)
1d3504fc 1165{
db7f47cf 1166#ifdef CONFIG_SMP
30e0e178
LZ
1167 if (val < -1 || val >= SD_LV_MAX)
1168 return -EINVAL;
db7f47cf 1169#endif
1d3504fc
HS
1170
1171 if (val != cs->relax_domain_level) {
1172 cs->relax_domain_level = val;
300ed6cb
LZ
1173 if (!cpumask_empty(cs->cpus_allowed) &&
1174 is_sched_load_balance(cs))
cf417141 1175 async_rebuild_sched_domains();
1d3504fc
HS
1176 }
1177
1178 return 0;
1179}
1180
950592f7
MX
1181/*
1182 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1183 * @tsk: task to be updated
1184 * @scan: struct cgroup_scanner containing the cgroup of the task
1185 *
1186 * Called by cgroup_scan_tasks() for each task in a cgroup.
1187 *
1188 * We don't need to re-check for the cgroup/cpuset membership, since we're
1189 * holding cgroup_lock() at this point.
1190 */
1191static void cpuset_change_flag(struct task_struct *tsk,
1192 struct cgroup_scanner *scan)
1193{
1194 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1195}
1196
1197/*
1198 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1199 * @cs: the cpuset in which each task's spread flags needs to be changed
1200 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1201 *
1202 * Called with cgroup_mutex held
1203 *
1204 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1205 * calling callback functions for each.
1206 *
1207 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1208 * if @heap != NULL.
1209 */
1210static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1211{
1212 struct cgroup_scanner scan;
1213
1214 scan.cg = cs->css.cgroup;
1215 scan.test_task = NULL;
1216 scan.process_task = cpuset_change_flag;
1217 scan.heap = heap;
1218 cgroup_scan_tasks(&scan);
1219}
1220
1da177e4
LT
1221/*
1222 * update_flag - read a 0 or a 1 in a file and update associated flag
78608366
PM
1223 * bit: the bit to update (see cpuset_flagbits_t)
1224 * cs: the cpuset to update
1225 * turning_on: whether the flag is being set or cleared
053199ed 1226 *
2df167a3 1227 * Call with cgroup_mutex held.
1da177e4
LT
1228 */
1229
700fe1ab
PM
1230static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1231 int turning_on)
1da177e4 1232{
645fcc9d 1233 struct cpuset *trialcs;
40b6a762 1234 int balance_flag_changed;
950592f7
MX
1235 int spread_flag_changed;
1236 struct ptr_heap heap;
1237 int err;
1da177e4 1238
645fcc9d
LZ
1239 trialcs = alloc_trial_cpuset(cs);
1240 if (!trialcs)
1241 return -ENOMEM;
1242
1da177e4 1243 if (turning_on)
645fcc9d 1244 set_bit(bit, &trialcs->flags);
1da177e4 1245 else
645fcc9d 1246 clear_bit(bit, &trialcs->flags);
1da177e4 1247
645fcc9d 1248 err = validate_change(cs, trialcs);
85d7b949 1249 if (err < 0)
645fcc9d 1250 goto out;
029190c5 1251
950592f7
MX
1252 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1253 if (err < 0)
1254 goto out;
1255
029190c5 1256 balance_flag_changed = (is_sched_load_balance(cs) !=
645fcc9d 1257 is_sched_load_balance(trialcs));
029190c5 1258
950592f7
MX
1259 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1260 || (is_spread_page(cs) != is_spread_page(trialcs)));
1261
3d3f26a7 1262 mutex_lock(&callback_mutex);
645fcc9d 1263 cs->flags = trialcs->flags;
3d3f26a7 1264 mutex_unlock(&callback_mutex);
85d7b949 1265
300ed6cb 1266 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
cf417141 1267 async_rebuild_sched_domains();
029190c5 1268
950592f7
MX
1269 if (spread_flag_changed)
1270 update_tasks_flags(cs, &heap);
1271 heap_free(&heap);
645fcc9d
LZ
1272out:
1273 free_trial_cpuset(trialcs);
1274 return err;
1da177e4
LT
1275}
1276
3e0d98b9 1277/*
80f7228b 1278 * Frequency meter - How fast is some event occurring?
3e0d98b9
PJ
1279 *
1280 * These routines manage a digitally filtered, constant time based,
1281 * event frequency meter. There are four routines:
1282 * fmeter_init() - initialize a frequency meter.
1283 * fmeter_markevent() - called each time the event happens.
1284 * fmeter_getrate() - returns the recent rate of such events.
1285 * fmeter_update() - internal routine used to update fmeter.
1286 *
1287 * A common data structure is passed to each of these routines,
1288 * which is used to keep track of the state required to manage the
1289 * frequency meter and its digital filter.
1290 *
1291 * The filter works on the number of events marked per unit time.
1292 * The filter is single-pole low-pass recursive (IIR). The time unit
1293 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1294 * simulate 3 decimal digits of precision (multiplied by 1000).
1295 *
1296 * With an FM_COEF of 933, and a time base of 1 second, the filter
1297 * has a half-life of 10 seconds, meaning that if the events quit
1298 * happening, then the rate returned from the fmeter_getrate()
1299 * will be cut in half each 10 seconds, until it converges to zero.
1300 *
1301 * It is not worth doing a real infinitely recursive filter. If more
1302 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1303 * just compute FM_MAXTICKS ticks worth, by which point the level
1304 * will be stable.
1305 *
1306 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1307 * arithmetic overflow in the fmeter_update() routine.
1308 *
1309 * Given the simple 32 bit integer arithmetic used, this meter works
1310 * best for reporting rates between one per millisecond (msec) and
1311 * one per 32 (approx) seconds. At constant rates faster than one
1312 * per msec it maxes out at values just under 1,000,000. At constant
1313 * rates between one per msec, and one per second it will stabilize
1314 * to a value N*1000, where N is the rate of events per second.
1315 * At constant rates between one per second and one per 32 seconds,
1316 * it will be choppy, moving up on the seconds that have an event,
1317 * and then decaying until the next event. At rates slower than
1318 * about one in 32 seconds, it decays all the way back to zero between
1319 * each event.
1320 */
1321
1322#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1323#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1324#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1325#define FM_SCALE 1000 /* faux fixed point scale */
1326
1327/* Initialize a frequency meter */
1328static void fmeter_init(struct fmeter *fmp)
1329{
1330 fmp->cnt = 0;
1331 fmp->val = 0;
1332 fmp->time = 0;
1333 spin_lock_init(&fmp->lock);
1334}
1335
1336/* Internal meter update - process cnt events and update value */
1337static void fmeter_update(struct fmeter *fmp)
1338{
1339 time_t now = get_seconds();
1340 time_t ticks = now - fmp->time;
1341
1342 if (ticks == 0)
1343 return;
1344
1345 ticks = min(FM_MAXTICKS, ticks);
1346 while (ticks-- > 0)
1347 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1348 fmp->time = now;
1349
1350 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1351 fmp->cnt = 0;
1352}
1353
1354/* Process any previous ticks, then bump cnt by one (times scale). */
1355static void fmeter_markevent(struct fmeter *fmp)
1356{
1357 spin_lock(&fmp->lock);
1358 fmeter_update(fmp);
1359 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1360 spin_unlock(&fmp->lock);
1361}
1362
1363/* Process any previous ticks, then return current value. */
1364static int fmeter_getrate(struct fmeter *fmp)
1365{
1366 int val;
1367
1368 spin_lock(&fmp->lock);
1369 fmeter_update(fmp);
1370 val = fmp->val;
1371 spin_unlock(&fmp->lock);
1372 return val;
1373}
1374
2341d1b6
LZ
1375/* Protected by cgroup_lock */
1376static cpumask_var_t cpus_attach;
1377
2df167a3 1378/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
be367d09
BB
1379static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1380 struct task_struct *tsk, bool threadgroup)
1da177e4 1381{
be367d09 1382 int ret;
8793d854 1383 struct cpuset *cs = cgroup_cs(cont);
1da177e4 1384
300ed6cb 1385 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1da177e4 1386 return -ENOSPC;
9985b0ba 1387
6d7b2f5f
DR
1388 /*
1389 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1390 * cannot change their cpu affinity and isolating such threads by their
1391 * set of allowed nodes is unnecessary. Thus, cpusets are not
1392 * applicable for such threads. This prevents checking for success of
1393 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1394 * be changed.
1395 */
1396 if (tsk->flags & PF_THREAD_BOUND)
1397 return -EINVAL;
1da177e4 1398
b0ae1981 1399 ret = security_task_setscheduler(tsk);
be367d09
BB
1400 if (ret)
1401 return ret;
1402 if (threadgroup) {
1403 struct task_struct *c;
1404
1405 rcu_read_lock();
1406 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
b0ae1981 1407 ret = security_task_setscheduler(c);
be367d09
BB
1408 if (ret) {
1409 rcu_read_unlock();
1410 return ret;
1411 }
1412 }
1413 rcu_read_unlock();
1414 }
1415 return 0;
1416}
1417
1418static void cpuset_attach_task(struct task_struct *tsk, nodemask_t *to,
1419 struct cpuset *cs)
1420{
1421 int err;
1422 /*
1423 * can_attach beforehand should guarantee that this doesn't fail.
1424 * TODO: have a better way to handle failure here
1425 */
1426 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1427 WARN_ON_ONCE(err);
1428
be367d09 1429 cpuset_change_task_nodemask(tsk, to);
be367d09
BB
1430 cpuset_update_task_spread_flag(cs, tsk);
1431
8793d854 1432}
1da177e4 1433
be367d09
BB
1434static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1435 struct cgroup *oldcont, struct task_struct *tsk,
1436 bool threadgroup)
8793d854 1437{
8793d854
PM
1438 struct mm_struct *mm;
1439 struct cpuset *cs = cgroup_cs(cont);
1440 struct cpuset *oldcs = cgroup_cs(oldcont);
53feb297
MX
1441 NODEMASK_ALLOC(nodemask_t, from, GFP_KERNEL);
1442 NODEMASK_ALLOC(nodemask_t, to, GFP_KERNEL);
1443
1444 if (from == NULL || to == NULL)
1445 goto alloc_fail;
22fb52dd 1446
f5813d94 1447 if (cs == &top_cpuset) {
2341d1b6 1448 cpumask_copy(cpus_attach, cpu_possible_mask);
f5813d94 1449 } else {
2341d1b6 1450 guarantee_online_cpus(cs, cpus_attach);
f5813d94 1451 }
53feb297 1452 guarantee_online_mems(cs, to);
1da177e4 1453
be367d09 1454 /* do per-task migration stuff possibly for each in the threadgroup */
53feb297 1455 cpuset_attach_task(tsk, to, cs);
be367d09
BB
1456 if (threadgroup) {
1457 struct task_struct *c;
1458 rcu_read_lock();
1459 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
53feb297 1460 cpuset_attach_task(c, to, cs);
be367d09
BB
1461 }
1462 rcu_read_unlock();
1463 }
950592f7 1464
be367d09 1465 /* change mm; only needs to be done once even if threadgroup */
53feb297
MX
1466 *from = oldcs->mems_allowed;
1467 *to = cs->mems_allowed;
4225399a
PJ
1468 mm = get_task_mm(tsk);
1469 if (mm) {
53feb297 1470 mpol_rebind_mm(mm, to);
2741a559 1471 if (is_memory_migrate(cs))
53feb297 1472 cpuset_migrate_mm(mm, from, to);
4225399a
PJ
1473 mmput(mm);
1474 }
53feb297
MX
1475
1476alloc_fail:
1477 NODEMASK_FREE(from);
1478 NODEMASK_FREE(to);
1da177e4
LT
1479}
1480
1481/* The various types of files and directories in a cpuset file system */
1482
1483typedef enum {
45b07ef3 1484 FILE_MEMORY_MIGRATE,
1da177e4
LT
1485 FILE_CPULIST,
1486 FILE_MEMLIST,
1487 FILE_CPU_EXCLUSIVE,
1488 FILE_MEM_EXCLUSIVE,
78608366 1489 FILE_MEM_HARDWALL,
029190c5 1490 FILE_SCHED_LOAD_BALANCE,
1d3504fc 1491 FILE_SCHED_RELAX_DOMAIN_LEVEL,
3e0d98b9
PJ
1492 FILE_MEMORY_PRESSURE_ENABLED,
1493 FILE_MEMORY_PRESSURE,
825a46af
PJ
1494 FILE_SPREAD_PAGE,
1495 FILE_SPREAD_SLAB,
1da177e4
LT
1496} cpuset_filetype_t;
1497
700fe1ab
PM
1498static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1499{
1500 int retval = 0;
1501 struct cpuset *cs = cgroup_cs(cgrp);
1502 cpuset_filetype_t type = cft->private;
1503
e3712395 1504 if (!cgroup_lock_live_group(cgrp))
700fe1ab 1505 return -ENODEV;
700fe1ab
PM
1506
1507 switch (type) {
1da177e4 1508 case FILE_CPU_EXCLUSIVE:
700fe1ab 1509 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1da177e4
LT
1510 break;
1511 case FILE_MEM_EXCLUSIVE:
700fe1ab 1512 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1da177e4 1513 break;
78608366
PM
1514 case FILE_MEM_HARDWALL:
1515 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1516 break;
029190c5 1517 case FILE_SCHED_LOAD_BALANCE:
700fe1ab 1518 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1d3504fc 1519 break;
45b07ef3 1520 case FILE_MEMORY_MIGRATE:
700fe1ab 1521 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
45b07ef3 1522 break;
3e0d98b9 1523 case FILE_MEMORY_PRESSURE_ENABLED:
700fe1ab 1524 cpuset_memory_pressure_enabled = !!val;
3e0d98b9
PJ
1525 break;
1526 case FILE_MEMORY_PRESSURE:
1527 retval = -EACCES;
1528 break;
825a46af 1529 case FILE_SPREAD_PAGE:
700fe1ab 1530 retval = update_flag(CS_SPREAD_PAGE, cs, val);
825a46af
PJ
1531 break;
1532 case FILE_SPREAD_SLAB:
700fe1ab 1533 retval = update_flag(CS_SPREAD_SLAB, cs, val);
825a46af 1534 break;
1da177e4
LT
1535 default:
1536 retval = -EINVAL;
700fe1ab 1537 break;
1da177e4 1538 }
8793d854 1539 cgroup_unlock();
1da177e4
LT
1540 return retval;
1541}
1542
5be7a479
PM
1543static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1544{
1545 int retval = 0;
1546 struct cpuset *cs = cgroup_cs(cgrp);
1547 cpuset_filetype_t type = cft->private;
1548
e3712395 1549 if (!cgroup_lock_live_group(cgrp))
5be7a479 1550 return -ENODEV;
e3712395 1551
5be7a479
PM
1552 switch (type) {
1553 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1554 retval = update_relax_domain_level(cs, val);
1555 break;
1556 default:
1557 retval = -EINVAL;
1558 break;
1559 }
1560 cgroup_unlock();
1561 return retval;
1562}
1563
e3712395
PM
1564/*
1565 * Common handling for a write to a "cpus" or "mems" file.
1566 */
1567static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1568 const char *buf)
1569{
1570 int retval = 0;
645fcc9d
LZ
1571 struct cpuset *cs = cgroup_cs(cgrp);
1572 struct cpuset *trialcs;
e3712395
PM
1573
1574 if (!cgroup_lock_live_group(cgrp))
1575 return -ENODEV;
1576
645fcc9d
LZ
1577 trialcs = alloc_trial_cpuset(cs);
1578 if (!trialcs)
1579 return -ENOMEM;
1580
e3712395
PM
1581 switch (cft->private) {
1582 case FILE_CPULIST:
645fcc9d 1583 retval = update_cpumask(cs, trialcs, buf);
e3712395
PM
1584 break;
1585 case FILE_MEMLIST:
645fcc9d 1586 retval = update_nodemask(cs, trialcs, buf);
e3712395
PM
1587 break;
1588 default:
1589 retval = -EINVAL;
1590 break;
1591 }
645fcc9d
LZ
1592
1593 free_trial_cpuset(trialcs);
e3712395
PM
1594 cgroup_unlock();
1595 return retval;
1596}
1597
1da177e4
LT
1598/*
1599 * These ascii lists should be read in a single call, by using a user
1600 * buffer large enough to hold the entire map. If read in smaller
1601 * chunks, there is no guarantee of atomicity. Since the display format
1602 * used, list of ranges of sequential numbers, is variable length,
1603 * and since these maps can change value dynamically, one could read
1604 * gibberish by doing partial reads while a list was changing.
1605 * A single large read to a buffer that crosses a page boundary is
1606 * ok, because the result being copied to user land is not recomputed
1607 * across a page fault.
1608 */
1609
1610static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1611{
5a7625df 1612 int ret;
1da177e4 1613
3d3f26a7 1614 mutex_lock(&callback_mutex);
300ed6cb 1615 ret = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
3d3f26a7 1616 mutex_unlock(&callback_mutex);
1da177e4 1617
5a7625df 1618 return ret;
1da177e4
LT
1619}
1620
1621static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1622{
53feb297
MX
1623 NODEMASK_ALLOC(nodemask_t, mask, GFP_KERNEL);
1624 int retval;
1625
1626 if (mask == NULL)
1627 return -ENOMEM;
1da177e4 1628
3d3f26a7 1629 mutex_lock(&callback_mutex);
53feb297 1630 *mask = cs->mems_allowed;
3d3f26a7 1631 mutex_unlock(&callback_mutex);
1da177e4 1632
53feb297
MX
1633 retval = nodelist_scnprintf(page, PAGE_SIZE, *mask);
1634
1635 NODEMASK_FREE(mask);
1636
1637 return retval;
1da177e4
LT
1638}
1639
8793d854
PM
1640static ssize_t cpuset_common_file_read(struct cgroup *cont,
1641 struct cftype *cft,
1642 struct file *file,
1643 char __user *buf,
1644 size_t nbytes, loff_t *ppos)
1da177e4 1645{
8793d854 1646 struct cpuset *cs = cgroup_cs(cont);
1da177e4
LT
1647 cpuset_filetype_t type = cft->private;
1648 char *page;
1649 ssize_t retval = 0;
1650 char *s;
1da177e4 1651
e12ba74d 1652 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1da177e4
LT
1653 return -ENOMEM;
1654
1655 s = page;
1656
1657 switch (type) {
1658 case FILE_CPULIST:
1659 s += cpuset_sprintf_cpulist(s, cs);
1660 break;
1661 case FILE_MEMLIST:
1662 s += cpuset_sprintf_memlist(s, cs);
1663 break;
1da177e4
LT
1664 default:
1665 retval = -EINVAL;
1666 goto out;
1667 }
1668 *s++ = '\n';
1da177e4 1669
eacaa1f5 1670 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1da177e4
LT
1671out:
1672 free_page((unsigned long)page);
1673 return retval;
1674}
1675
700fe1ab
PM
1676static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1677{
1678 struct cpuset *cs = cgroup_cs(cont);
1679 cpuset_filetype_t type = cft->private;
1680 switch (type) {
1681 case FILE_CPU_EXCLUSIVE:
1682 return is_cpu_exclusive(cs);
1683 case FILE_MEM_EXCLUSIVE:
1684 return is_mem_exclusive(cs);
78608366
PM
1685 case FILE_MEM_HARDWALL:
1686 return is_mem_hardwall(cs);
700fe1ab
PM
1687 case FILE_SCHED_LOAD_BALANCE:
1688 return is_sched_load_balance(cs);
1689 case FILE_MEMORY_MIGRATE:
1690 return is_memory_migrate(cs);
1691 case FILE_MEMORY_PRESSURE_ENABLED:
1692 return cpuset_memory_pressure_enabled;
1693 case FILE_MEMORY_PRESSURE:
1694 return fmeter_getrate(&cs->fmeter);
1695 case FILE_SPREAD_PAGE:
1696 return is_spread_page(cs);
1697 case FILE_SPREAD_SLAB:
1698 return is_spread_slab(cs);
1699 default:
1700 BUG();
1701 }
cf417141
MK
1702
1703 /* Unreachable but makes gcc happy */
1704 return 0;
700fe1ab 1705}
1da177e4 1706
5be7a479
PM
1707static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1708{
1709 struct cpuset *cs = cgroup_cs(cont);
1710 cpuset_filetype_t type = cft->private;
1711 switch (type) {
1712 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1713 return cs->relax_domain_level;
1714 default:
1715 BUG();
1716 }
cf417141
MK
1717
1718 /* Unrechable but makes gcc happy */
1719 return 0;
5be7a479
PM
1720}
1721
1da177e4
LT
1722
1723/*
1724 * for the common functions, 'private' gives the type of file
1725 */
1726
addf2c73
PM
1727static struct cftype files[] = {
1728 {
1729 .name = "cpus",
1730 .read = cpuset_common_file_read,
e3712395
PM
1731 .write_string = cpuset_write_resmask,
1732 .max_write_len = (100U + 6 * NR_CPUS),
addf2c73
PM
1733 .private = FILE_CPULIST,
1734 },
1735
1736 {
1737 .name = "mems",
1738 .read = cpuset_common_file_read,
e3712395
PM
1739 .write_string = cpuset_write_resmask,
1740 .max_write_len = (100U + 6 * MAX_NUMNODES),
addf2c73
PM
1741 .private = FILE_MEMLIST,
1742 },
1743
1744 {
1745 .name = "cpu_exclusive",
1746 .read_u64 = cpuset_read_u64,
1747 .write_u64 = cpuset_write_u64,
1748 .private = FILE_CPU_EXCLUSIVE,
1749 },
1750
1751 {
1752 .name = "mem_exclusive",
1753 .read_u64 = cpuset_read_u64,
1754 .write_u64 = cpuset_write_u64,
1755 .private = FILE_MEM_EXCLUSIVE,
1756 },
1757
78608366
PM
1758 {
1759 .name = "mem_hardwall",
1760 .read_u64 = cpuset_read_u64,
1761 .write_u64 = cpuset_write_u64,
1762 .private = FILE_MEM_HARDWALL,
1763 },
1764
addf2c73
PM
1765 {
1766 .name = "sched_load_balance",
1767 .read_u64 = cpuset_read_u64,
1768 .write_u64 = cpuset_write_u64,
1769 .private = FILE_SCHED_LOAD_BALANCE,
1770 },
1771
1772 {
1773 .name = "sched_relax_domain_level",
5be7a479
PM
1774 .read_s64 = cpuset_read_s64,
1775 .write_s64 = cpuset_write_s64,
addf2c73
PM
1776 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1777 },
1778
1779 {
1780 .name = "memory_migrate",
1781 .read_u64 = cpuset_read_u64,
1782 .write_u64 = cpuset_write_u64,
1783 .private = FILE_MEMORY_MIGRATE,
1784 },
1785
1786 {
1787 .name = "memory_pressure",
1788 .read_u64 = cpuset_read_u64,
1789 .write_u64 = cpuset_write_u64,
1790 .private = FILE_MEMORY_PRESSURE,
099fca32 1791 .mode = S_IRUGO,
addf2c73
PM
1792 },
1793
1794 {
1795 .name = "memory_spread_page",
1796 .read_u64 = cpuset_read_u64,
1797 .write_u64 = cpuset_write_u64,
1798 .private = FILE_SPREAD_PAGE,
1799 },
1800
1801 {
1802 .name = "memory_spread_slab",
1803 .read_u64 = cpuset_read_u64,
1804 .write_u64 = cpuset_write_u64,
1805 .private = FILE_SPREAD_SLAB,
1806 },
45b07ef3
PJ
1807};
1808
3e0d98b9
PJ
1809static struct cftype cft_memory_pressure_enabled = {
1810 .name = "memory_pressure_enabled",
700fe1ab
PM
1811 .read_u64 = cpuset_read_u64,
1812 .write_u64 = cpuset_write_u64,
3e0d98b9
PJ
1813 .private = FILE_MEMORY_PRESSURE_ENABLED,
1814};
1815
8793d854 1816static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1da177e4
LT
1817{
1818 int err;
1819
addf2c73
PM
1820 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1821 if (err)
1da177e4 1822 return err;
8793d854 1823 /* memory_pressure_enabled is in root cpuset only */
addf2c73 1824 if (!cont->parent)
8793d854 1825 err = cgroup_add_file(cont, ss,
addf2c73
PM
1826 &cft_memory_pressure_enabled);
1827 return err;
1da177e4
LT
1828}
1829
8793d854
PM
1830/*
1831 * post_clone() is called at the end of cgroup_clone().
1832 * 'cgroup' was just created automatically as a result of
1833 * a cgroup_clone(), and the current task is about to
1834 * be moved into 'cgroup'.
1835 *
1836 * Currently we refuse to set up the cgroup - thereby
1837 * refusing the task to be entered, and as a result refusing
1838 * the sys_unshare() or clone() which initiated it - if any
1839 * sibling cpusets have exclusive cpus or mem.
1840 *
1841 * If this becomes a problem for some users who wish to
1842 * allow that scenario, then cpuset_post_clone() could be
1843 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2df167a3
PM
1844 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1845 * held.
8793d854
PM
1846 */
1847static void cpuset_post_clone(struct cgroup_subsys *ss,
1848 struct cgroup *cgroup)
1849{
1850 struct cgroup *parent, *child;
1851 struct cpuset *cs, *parent_cs;
1852
1853 parent = cgroup->parent;
1854 list_for_each_entry(child, &parent->children, sibling) {
1855 cs = cgroup_cs(child);
1856 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1857 return;
1858 }
1859 cs = cgroup_cs(cgroup);
1860 parent_cs = cgroup_cs(parent);
1861
1862 cs->mems_allowed = parent_cs->mems_allowed;
300ed6cb 1863 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
8793d854
PM
1864 return;
1865}
1866
1da177e4
LT
1867/*
1868 * cpuset_create - create a cpuset
2df167a3
PM
1869 * ss: cpuset cgroup subsystem
1870 * cont: control group that the new cpuset will be part of
1da177e4
LT
1871 */
1872
8793d854
PM
1873static struct cgroup_subsys_state *cpuset_create(
1874 struct cgroup_subsys *ss,
1875 struct cgroup *cont)
1da177e4
LT
1876{
1877 struct cpuset *cs;
8793d854 1878 struct cpuset *parent;
1da177e4 1879
8793d854 1880 if (!cont->parent) {
8793d854
PM
1881 return &top_cpuset.css;
1882 }
1883 parent = cgroup_cs(cont->parent);
1da177e4
LT
1884 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1885 if (!cs)
8793d854 1886 return ERR_PTR(-ENOMEM);
300ed6cb
LZ
1887 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1888 kfree(cs);
1889 return ERR_PTR(-ENOMEM);
1890 }
1da177e4 1891
1da177e4 1892 cs->flags = 0;
825a46af
PJ
1893 if (is_spread_page(parent))
1894 set_bit(CS_SPREAD_PAGE, &cs->flags);
1895 if (is_spread_slab(parent))
1896 set_bit(CS_SPREAD_SLAB, &cs->flags);
029190c5 1897 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
300ed6cb 1898 cpumask_clear(cs->cpus_allowed);
f9a86fcb 1899 nodes_clear(cs->mems_allowed);
3e0d98b9 1900 fmeter_init(&cs->fmeter);
1d3504fc 1901 cs->relax_domain_level = -1;
1da177e4
LT
1902
1903 cs->parent = parent;
202f72d5 1904 number_of_cpusets++;
8793d854 1905 return &cs->css ;
1da177e4
LT
1906}
1907
029190c5 1908/*
029190c5
PJ
1909 * If the cpuset being removed has its flag 'sched_load_balance'
1910 * enabled, then simulate turning sched_load_balance off, which
cf417141 1911 * will call async_rebuild_sched_domains().
029190c5
PJ
1912 */
1913
8793d854 1914static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1da177e4 1915{
8793d854 1916 struct cpuset *cs = cgroup_cs(cont);
1da177e4 1917
029190c5 1918 if (is_sched_load_balance(cs))
700fe1ab 1919 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
029190c5 1920
202f72d5 1921 number_of_cpusets--;
300ed6cb 1922 free_cpumask_var(cs->cpus_allowed);
8793d854 1923 kfree(cs);
1da177e4
LT
1924}
1925
8793d854
PM
1926struct cgroup_subsys cpuset_subsys = {
1927 .name = "cpuset",
1928 .create = cpuset_create,
cf417141 1929 .destroy = cpuset_destroy,
8793d854
PM
1930 .can_attach = cpuset_can_attach,
1931 .attach = cpuset_attach,
1932 .populate = cpuset_populate,
1933 .post_clone = cpuset_post_clone,
1934 .subsys_id = cpuset_subsys_id,
1935 .early_init = 1,
1936};
1937
1da177e4
LT
1938/**
1939 * cpuset_init - initialize cpusets at system boot
1940 *
1941 * Description: Initialize top_cpuset and the cpuset internal file system,
1942 **/
1943
1944int __init cpuset_init(void)
1945{
8793d854 1946 int err = 0;
1da177e4 1947
58568d2a
MX
1948 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1949 BUG();
1950
300ed6cb 1951 cpumask_setall(top_cpuset.cpus_allowed);
f9a86fcb 1952 nodes_setall(top_cpuset.mems_allowed);
1da177e4 1953
3e0d98b9 1954 fmeter_init(&top_cpuset.fmeter);
029190c5 1955 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1d3504fc 1956 top_cpuset.relax_domain_level = -1;
1da177e4 1957
1da177e4
LT
1958 err = register_filesystem(&cpuset_fs_type);
1959 if (err < 0)
8793d854
PM
1960 return err;
1961
2341d1b6
LZ
1962 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1963 BUG();
1964
202f72d5 1965 number_of_cpusets = 1;
8793d854 1966 return 0;
1da177e4
LT
1967}
1968
956db3ca
CW
1969/**
1970 * cpuset_do_move_task - move a given task to another cpuset
1971 * @tsk: pointer to task_struct the task to move
1972 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1973 *
1974 * Called by cgroup_scan_tasks() for each task in a cgroup.
1975 * Return nonzero to stop the walk through the tasks.
1976 */
9e0c914c
AB
1977static void cpuset_do_move_task(struct task_struct *tsk,
1978 struct cgroup_scanner *scan)
956db3ca 1979{
7f81b1ae 1980 struct cgroup *new_cgroup = scan->data;
956db3ca 1981
7f81b1ae 1982 cgroup_attach_task(new_cgroup, tsk);
956db3ca
CW
1983}
1984
1985/**
1986 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1987 * @from: cpuset in which the tasks currently reside
1988 * @to: cpuset to which the tasks will be moved
1989 *
c8d9c90c
PJ
1990 * Called with cgroup_mutex held
1991 * callback_mutex must not be held, as cpuset_attach() will take it.
956db3ca
CW
1992 *
1993 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1994 * calling callback functions for each.
1995 */
1996static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1997{
7f81b1ae 1998 struct cgroup_scanner scan;
956db3ca 1999
7f81b1ae
LZ
2000 scan.cg = from->css.cgroup;
2001 scan.test_task = NULL; /* select all tasks in cgroup */
2002 scan.process_task = cpuset_do_move_task;
2003 scan.heap = NULL;
2004 scan.data = to->css.cgroup;
956db3ca 2005
7f81b1ae 2006 if (cgroup_scan_tasks(&scan))
956db3ca
CW
2007 printk(KERN_ERR "move_member_tasks_to_cpuset: "
2008 "cgroup_scan_tasks failed\n");
2009}
2010
b1aac8bb 2011/*
cf417141 2012 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
b1aac8bb
PJ
2013 * or memory nodes, we need to walk over the cpuset hierarchy,
2014 * removing that CPU or node from all cpusets. If this removes the
956db3ca
CW
2015 * last CPU or node from a cpuset, then move the tasks in the empty
2016 * cpuset to its next-highest non-empty parent.
b1aac8bb 2017 *
c8d9c90c
PJ
2018 * Called with cgroup_mutex held
2019 * callback_mutex must not be held, as cpuset_attach() will take it.
b1aac8bb 2020 */
956db3ca
CW
2021static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2022{
2023 struct cpuset *parent;
2024
c8d9c90c
PJ
2025 /*
2026 * The cgroup's css_sets list is in use if there are tasks
2027 * in the cpuset; the list is empty if there are none;
2028 * the cs->css.refcnt seems always 0.
2029 */
956db3ca
CW
2030 if (list_empty(&cs->css.cgroup->css_sets))
2031 return;
b1aac8bb 2032
956db3ca
CW
2033 /*
2034 * Find its next-highest non-empty parent, (top cpuset
2035 * has online cpus, so can't be empty).
2036 */
2037 parent = cs->parent;
300ed6cb 2038 while (cpumask_empty(parent->cpus_allowed) ||
b4501295 2039 nodes_empty(parent->mems_allowed))
956db3ca 2040 parent = parent->parent;
956db3ca
CW
2041
2042 move_member_tasks_to_cpuset(cs, parent);
2043}
2044
2045/*
2046 * Walk the specified cpuset subtree and look for empty cpusets.
2047 * The tasks of such cpuset must be moved to a parent cpuset.
2048 *
2df167a3 2049 * Called with cgroup_mutex held. We take callback_mutex to modify
956db3ca
CW
2050 * cpus_allowed and mems_allowed.
2051 *
2052 * This walk processes the tree from top to bottom, completing one layer
2053 * before dropping down to the next. It always processes a node before
2054 * any of its children.
2055 *
2056 * For now, since we lack memory hot unplug, we'll never see a cpuset
2057 * that has tasks along with an empty 'mems'. But if we did see such
2058 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2059 */
d294eb83 2060static void scan_for_empty_cpusets(struct cpuset *root)
b1aac8bb 2061{
8d1e6266 2062 LIST_HEAD(queue);
956db3ca
CW
2063 struct cpuset *cp; /* scans cpusets being updated */
2064 struct cpuset *child; /* scans child cpusets of cp */
8793d854 2065 struct cgroup *cont;
53feb297
MX
2066 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2067
2068 if (oldmems == NULL)
2069 return;
b1aac8bb 2070
956db3ca
CW
2071 list_add_tail((struct list_head *)&root->stack_list, &queue);
2072
956db3ca 2073 while (!list_empty(&queue)) {
8d1e6266 2074 cp = list_first_entry(&queue, struct cpuset, stack_list);
956db3ca
CW
2075 list_del(queue.next);
2076 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2077 child = cgroup_cs(cont);
2078 list_add_tail(&child->stack_list, &queue);
2079 }
b4501295
PJ
2080
2081 /* Continue past cpusets with all cpus, mems online */
6ad4c188 2082 if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
b4501295
PJ
2083 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2084 continue;
2085
53feb297 2086 *oldmems = cp->mems_allowed;
f9b4fb8d 2087
956db3ca 2088 /* Remove offline cpus and mems from this cpuset. */
b4501295 2089 mutex_lock(&callback_mutex);
300ed6cb 2090 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
6ad4c188 2091 cpu_active_mask);
956db3ca
CW
2092 nodes_and(cp->mems_allowed, cp->mems_allowed,
2093 node_states[N_HIGH_MEMORY]);
b4501295
PJ
2094 mutex_unlock(&callback_mutex);
2095
2096 /* Move tasks from the empty cpuset to a parent */
300ed6cb 2097 if (cpumask_empty(cp->cpus_allowed) ||
b4501295 2098 nodes_empty(cp->mems_allowed))
956db3ca 2099 remove_tasks_in_empty_cpuset(cp);
f9b4fb8d 2100 else {
4e74339a 2101 update_tasks_cpumask(cp, NULL);
53feb297 2102 update_tasks_nodemask(cp, oldmems, NULL);
f9b4fb8d 2103 }
b1aac8bb 2104 }
53feb297 2105 NODEMASK_FREE(oldmems);
b1aac8bb
PJ
2106}
2107
4c4d50f7
PJ
2108/*
2109 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2110 * period. This is necessary in order to make cpusets transparent
2111 * (of no affect) on systems that are actively using CPU hotplug
2112 * but making no active use of cpusets.
2113 *
38837fc7 2114 * This routine ensures that top_cpuset.cpus_allowed tracks
3a101d05 2115 * cpu_active_mask on each CPU hotplug (cpuhp) event.
cf417141
MK
2116 *
2117 * Called within get_online_cpus(). Needs to call cgroup_lock()
2118 * before calling generate_sched_domains().
4c4d50f7 2119 */
0b2e918a 2120void cpuset_update_active_cpus(void)
4c4d50f7 2121{
cf417141 2122 struct sched_domain_attr *attr;
acc3f5d7 2123 cpumask_var_t *doms;
cf417141
MK
2124 int ndoms;
2125
cf417141 2126 cgroup_lock();
0b4217b3 2127 mutex_lock(&callback_mutex);
6ad4c188 2128 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
0b4217b3 2129 mutex_unlock(&callback_mutex);
cf417141
MK
2130 scan_for_empty_cpusets(&top_cpuset);
2131 ndoms = generate_sched_domains(&doms, &attr);
2132 cgroup_unlock();
2133
2134 /* Have scheduler rebuild the domains */
2135 partition_sched_domains(ndoms, doms, attr);
4c4d50f7 2136}
4c4d50f7 2137
b1aac8bb 2138#ifdef CONFIG_MEMORY_HOTPLUG
38837fc7 2139/*
0e1e7c7a 2140 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
cf417141
MK
2141 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2142 * See also the previous routine cpuset_track_online_cpus().
38837fc7 2143 */
f481891f
MX
2144static int cpuset_track_online_nodes(struct notifier_block *self,
2145 unsigned long action, void *arg)
38837fc7 2146{
53feb297
MX
2147 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2148
2149 if (oldmems == NULL)
2150 return NOTIFY_DONE;
5ab116c9 2151
cf417141 2152 cgroup_lock();
f481891f
MX
2153 switch (action) {
2154 case MEM_ONLINE:
53feb297 2155 *oldmems = top_cpuset.mems_allowed;
0b4217b3 2156 mutex_lock(&callback_mutex);
f481891f 2157 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
0b4217b3 2158 mutex_unlock(&callback_mutex);
53feb297 2159 update_tasks_nodemask(&top_cpuset, oldmems, NULL);
5ab116c9
MX
2160 break;
2161 case MEM_OFFLINE:
2162 /*
2163 * needn't update top_cpuset.mems_allowed explicitly because
2164 * scan_for_empty_cpusets() will update it.
2165 */
2166 scan_for_empty_cpusets(&top_cpuset);
f481891f
MX
2167 break;
2168 default:
2169 break;
2170 }
cf417141 2171 cgroup_unlock();
53feb297
MX
2172
2173 NODEMASK_FREE(oldmems);
f481891f 2174 return NOTIFY_OK;
38837fc7
PJ
2175}
2176#endif
2177
1da177e4
LT
2178/**
2179 * cpuset_init_smp - initialize cpus_allowed
2180 *
2181 * Description: Finish top cpuset after cpu, node maps are initialized
2182 **/
2183
2184void __init cpuset_init_smp(void)
2185{
6ad4c188 2186 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
0e1e7c7a 2187 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
4c4d50f7 2188
f481891f 2189 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
f90d4118
MX
2190
2191 cpuset_wq = create_singlethread_workqueue("cpuset");
2192 BUG_ON(!cpuset_wq);
1da177e4
LT
2193}
2194
2195/**
1da177e4
LT
2196 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2197 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
6af866af 2198 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
1da177e4 2199 *
300ed6cb 2200 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
1da177e4
LT
2201 * attached to the specified @tsk. Guaranteed to return some non-empty
2202 * subset of cpu_online_map, even if this means going outside the
2203 * tasks cpuset.
2204 **/
2205
6af866af 2206void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
1da177e4 2207{
3d3f26a7 2208 mutex_lock(&callback_mutex);
909d75a3 2209 task_lock(tsk);
f9a86fcb 2210 guarantee_online_cpus(task_cs(tsk), pmask);
909d75a3 2211 task_unlock(tsk);
897f0b3c 2212 mutex_unlock(&callback_mutex);
1da177e4
LT
2213}
2214
9084bb82
ON
2215int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2216{
2217 const struct cpuset *cs;
2218 int cpu;
2219
2220 rcu_read_lock();
2221 cs = task_cs(tsk);
2222 if (cs)
2223 cpumask_copy(&tsk->cpus_allowed, cs->cpus_allowed);
2224 rcu_read_unlock();
2225
2226 /*
2227 * We own tsk->cpus_allowed, nobody can change it under us.
2228 *
2229 * But we used cs && cs->cpus_allowed lockless and thus can
2230 * race with cgroup_attach_task() or update_cpumask() and get
2231 * the wrong tsk->cpus_allowed. However, both cases imply the
2232 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2233 * which takes task_rq_lock().
2234 *
2235 * If we are called after it dropped the lock we must see all
2236 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2237 * set any mask even if it is not right from task_cs() pov,
2238 * the pending set_cpus_allowed_ptr() will fix things.
2239 */
2240
2241 cpu = cpumask_any_and(&tsk->cpus_allowed, cpu_active_mask);
2242 if (cpu >= nr_cpu_ids) {
2243 /*
2244 * Either tsk->cpus_allowed is wrong (see above) or it
2245 * is actually empty. The latter case is only possible
2246 * if we are racing with remove_tasks_in_empty_cpuset().
2247 * Like above we can temporary set any mask and rely on
2248 * set_cpus_allowed_ptr() as synchronization point.
2249 */
2250 cpumask_copy(&tsk->cpus_allowed, cpu_possible_mask);
2251 cpu = cpumask_any(cpu_active_mask);
2252 }
2253
2254 return cpu;
2255}
2256
1da177e4
LT
2257void cpuset_init_current_mems_allowed(void)
2258{
f9a86fcb 2259 nodes_setall(current->mems_allowed);
1da177e4
LT
2260}
2261
909d75a3
PJ
2262/**
2263 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2264 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2265 *
2266 * Description: Returns the nodemask_t mems_allowed of the cpuset
2267 * attached to the specified @tsk. Guaranteed to return some non-empty
0e1e7c7a 2268 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
909d75a3
PJ
2269 * tasks cpuset.
2270 **/
2271
2272nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2273{
2274 nodemask_t mask;
2275
3d3f26a7 2276 mutex_lock(&callback_mutex);
909d75a3 2277 task_lock(tsk);
8793d854 2278 guarantee_online_mems(task_cs(tsk), &mask);
909d75a3 2279 task_unlock(tsk);
3d3f26a7 2280 mutex_unlock(&callback_mutex);
909d75a3
PJ
2281
2282 return mask;
2283}
2284
d9fd8a6d 2285/**
19770b32
MG
2286 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2287 * @nodemask: the nodemask to be checked
d9fd8a6d 2288 *
19770b32 2289 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
1da177e4 2290 */
19770b32 2291int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
1da177e4 2292{
19770b32 2293 return nodes_intersects(*nodemask, current->mems_allowed);
1da177e4
LT
2294}
2295
9bf2229f 2296/*
78608366
PM
2297 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2298 * mem_hardwall ancestor to the specified cpuset. Call holding
2299 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2300 * (an unusual configuration), then returns the root cpuset.
9bf2229f 2301 */
78608366 2302static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
9bf2229f 2303{
78608366 2304 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
9bf2229f
PJ
2305 cs = cs->parent;
2306 return cs;
2307}
2308
d9fd8a6d 2309/**
a1bc5a4e
DR
2310 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2311 * @node: is this an allowed node?
02a0e53d 2312 * @gfp_mask: memory allocation flags
d9fd8a6d 2313 *
a1bc5a4e
DR
2314 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2315 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2316 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2317 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2318 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2319 * flag, yes.
9bf2229f
PJ
2320 * Otherwise, no.
2321 *
a1bc5a4e
DR
2322 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2323 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2324 * might sleep, and might allow a node from an enclosing cpuset.
02a0e53d 2325 *
a1bc5a4e
DR
2326 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2327 * cpusets, and never sleeps.
02a0e53d
PJ
2328 *
2329 * The __GFP_THISNODE placement logic is really handled elsewhere,
2330 * by forcibly using a zonelist starting at a specified node, and by
2331 * (in get_page_from_freelist()) refusing to consider the zones for
2332 * any node on the zonelist except the first. By the time any such
2333 * calls get to this routine, we should just shut up and say 'yes'.
2334 *
9bf2229f 2335 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
c596d9f3
DR
2336 * and do not allow allocations outside the current tasks cpuset
2337 * unless the task has been OOM killed as is marked TIF_MEMDIE.
9bf2229f 2338 * GFP_KERNEL allocations are not so marked, so can escape to the
78608366 2339 * nearest enclosing hardwalled ancestor cpuset.
9bf2229f 2340 *
02a0e53d
PJ
2341 * Scanning up parent cpusets requires callback_mutex. The
2342 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2343 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2344 * current tasks mems_allowed came up empty on the first pass over
2345 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2346 * cpuset are short of memory, might require taking the callback_mutex
2347 * mutex.
9bf2229f 2348 *
36be57ff 2349 * The first call here from mm/page_alloc:get_page_from_freelist()
02a0e53d
PJ
2350 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2351 * so no allocation on a node outside the cpuset is allowed (unless
2352 * in interrupt, of course).
36be57ff
PJ
2353 *
2354 * The second pass through get_page_from_freelist() doesn't even call
2355 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2356 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2357 * in alloc_flags. That logic and the checks below have the combined
2358 * affect that:
9bf2229f
PJ
2359 * in_interrupt - any node ok (current task context irrelevant)
2360 * GFP_ATOMIC - any node ok
c596d9f3 2361 * TIF_MEMDIE - any node ok
78608366 2362 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
9bf2229f 2363 * GFP_USER - only nodes in current tasks mems allowed ok.
36be57ff
PJ
2364 *
2365 * Rule:
a1bc5a4e 2366 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
36be57ff
PJ
2367 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2368 * the code that might scan up ancestor cpusets and sleep.
02a0e53d 2369 */
a1bc5a4e 2370int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
1da177e4 2371{
9bf2229f 2372 const struct cpuset *cs; /* current cpuset ancestors */
29afd49b 2373 int allowed; /* is allocation in zone z allowed? */
9bf2229f 2374
9b819d20 2375 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
9bf2229f 2376 return 1;
92d1dbd2 2377 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
9bf2229f
PJ
2378 if (node_isset(node, current->mems_allowed))
2379 return 1;
c596d9f3
DR
2380 /*
2381 * Allow tasks that have access to memory reserves because they have
2382 * been OOM killed to get memory anywhere.
2383 */
2384 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2385 return 1;
9bf2229f
PJ
2386 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2387 return 0;
2388
5563e770
BP
2389 if (current->flags & PF_EXITING) /* Let dying task have memory */
2390 return 1;
2391
9bf2229f 2392 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3d3f26a7 2393 mutex_lock(&callback_mutex);
053199ed 2394
053199ed 2395 task_lock(current);
78608366 2396 cs = nearest_hardwall_ancestor(task_cs(current));
053199ed
PJ
2397 task_unlock(current);
2398
9bf2229f 2399 allowed = node_isset(node, cs->mems_allowed);
3d3f26a7 2400 mutex_unlock(&callback_mutex);
9bf2229f 2401 return allowed;
1da177e4
LT
2402}
2403
02a0e53d 2404/*
a1bc5a4e
DR
2405 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2406 * @node: is this an allowed node?
02a0e53d
PJ
2407 * @gfp_mask: memory allocation flags
2408 *
a1bc5a4e
DR
2409 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2410 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2411 * yes. If the task has been OOM killed and has access to memory reserves as
2412 * specified by the TIF_MEMDIE flag, yes.
2413 * Otherwise, no.
02a0e53d
PJ
2414 *
2415 * The __GFP_THISNODE placement logic is really handled elsewhere,
2416 * by forcibly using a zonelist starting at a specified node, and by
2417 * (in get_page_from_freelist()) refusing to consider the zones for
2418 * any node on the zonelist except the first. By the time any such
2419 * calls get to this routine, we should just shut up and say 'yes'.
2420 *
a1bc5a4e
DR
2421 * Unlike the cpuset_node_allowed_softwall() variant, above,
2422 * this variant requires that the node be in the current task's
02a0e53d
PJ
2423 * mems_allowed or that we're in interrupt. It does not scan up the
2424 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2425 * It never sleeps.
2426 */
a1bc5a4e 2427int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
02a0e53d 2428{
02a0e53d
PJ
2429 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2430 return 1;
02a0e53d
PJ
2431 if (node_isset(node, current->mems_allowed))
2432 return 1;
dedf8b79
DW
2433 /*
2434 * Allow tasks that have access to memory reserves because they have
2435 * been OOM killed to get memory anywhere.
2436 */
2437 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2438 return 1;
02a0e53d
PJ
2439 return 0;
2440}
2441
505970b9
PJ
2442/**
2443 * cpuset_unlock - release lock on cpuset changes
2444 *
2445 * Undo the lock taken in a previous cpuset_lock() call.
2446 */
2447
2448void cpuset_unlock(void)
2449{
3d3f26a7 2450 mutex_unlock(&callback_mutex);
505970b9
PJ
2451}
2452
825a46af 2453/**
6adef3eb
JS
2454 * cpuset_mem_spread_node() - On which node to begin search for a file page
2455 * cpuset_slab_spread_node() - On which node to begin search for a slab page
825a46af
PJ
2456 *
2457 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2458 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2459 * and if the memory allocation used cpuset_mem_spread_node()
2460 * to determine on which node to start looking, as it will for
2461 * certain page cache or slab cache pages such as used for file
2462 * system buffers and inode caches, then instead of starting on the
2463 * local node to look for a free page, rather spread the starting
2464 * node around the tasks mems_allowed nodes.
2465 *
2466 * We don't have to worry about the returned node being offline
2467 * because "it can't happen", and even if it did, it would be ok.
2468 *
2469 * The routines calling guarantee_online_mems() are careful to
2470 * only set nodes in task->mems_allowed that are online. So it
2471 * should not be possible for the following code to return an
2472 * offline node. But if it did, that would be ok, as this routine
2473 * is not returning the node where the allocation must be, only
2474 * the node where the search should start. The zonelist passed to
2475 * __alloc_pages() will include all nodes. If the slab allocator
2476 * is passed an offline node, it will fall back to the local node.
2477 * See kmem_cache_alloc_node().
2478 */
2479
6adef3eb 2480static int cpuset_spread_node(int *rotor)
825a46af
PJ
2481{
2482 int node;
2483
6adef3eb 2484 node = next_node(*rotor, current->mems_allowed);
825a46af
PJ
2485 if (node == MAX_NUMNODES)
2486 node = first_node(current->mems_allowed);
6adef3eb 2487 *rotor = node;
825a46af
PJ
2488 return node;
2489}
6adef3eb
JS
2490
2491int cpuset_mem_spread_node(void)
2492{
2493 return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2494}
2495
2496int cpuset_slab_spread_node(void)
2497{
2498 return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2499}
2500
825a46af
PJ
2501EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2502
ef08e3b4 2503/**
bbe373f2
DR
2504 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2505 * @tsk1: pointer to task_struct of some task.
2506 * @tsk2: pointer to task_struct of some other task.
2507 *
2508 * Description: Return true if @tsk1's mems_allowed intersects the
2509 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2510 * one of the task's memory usage might impact the memory available
2511 * to the other.
ef08e3b4
PJ
2512 **/
2513
bbe373f2
DR
2514int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2515 const struct task_struct *tsk2)
ef08e3b4 2516{
bbe373f2 2517 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
ef08e3b4
PJ
2518}
2519
75aa1994
DR
2520/**
2521 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2522 * @task: pointer to task_struct of some task.
2523 *
2524 * Description: Prints @task's name, cpuset name, and cached copy of its
2525 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2526 * dereferencing task_cs(task).
2527 */
2528void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2529{
2530 struct dentry *dentry;
2531
2532 dentry = task_cs(tsk)->css.cgroup->dentry;
2533 spin_lock(&cpuset_buffer_lock);
2534 snprintf(cpuset_name, CPUSET_NAME_LEN,
2535 dentry ? (const char *)dentry->d_name.name : "/");
2536 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2537 tsk->mems_allowed);
2538 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2539 tsk->comm, cpuset_name, cpuset_nodelist);
2540 spin_unlock(&cpuset_buffer_lock);
2541}
2542
3e0d98b9
PJ
2543/*
2544 * Collection of memory_pressure is suppressed unless
2545 * this flag is enabled by writing "1" to the special
2546 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2547 */
2548
c5b2aff8 2549int cpuset_memory_pressure_enabled __read_mostly;
3e0d98b9
PJ
2550
2551/**
2552 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2553 *
2554 * Keep a running average of the rate of synchronous (direct)
2555 * page reclaim efforts initiated by tasks in each cpuset.
2556 *
2557 * This represents the rate at which some task in the cpuset
2558 * ran low on memory on all nodes it was allowed to use, and
2559 * had to enter the kernels page reclaim code in an effort to
2560 * create more free memory by tossing clean pages or swapping
2561 * or writing dirty pages.
2562 *
2563 * Display to user space in the per-cpuset read-only file
2564 * "memory_pressure". Value displayed is an integer
2565 * representing the recent rate of entry into the synchronous
2566 * (direct) page reclaim by any task attached to the cpuset.
2567 **/
2568
2569void __cpuset_memory_pressure_bump(void)
2570{
3e0d98b9 2571 task_lock(current);
8793d854 2572 fmeter_markevent(&task_cs(current)->fmeter);
3e0d98b9
PJ
2573 task_unlock(current);
2574}
2575
8793d854 2576#ifdef CONFIG_PROC_PID_CPUSET
1da177e4
LT
2577/*
2578 * proc_cpuset_show()
2579 * - Print tasks cpuset path into seq_file.
2580 * - Used for /proc/<pid>/cpuset.
053199ed
PJ
2581 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2582 * doesn't really matter if tsk->cpuset changes after we read it,
c8d9c90c 2583 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2df167a3 2584 * anyway.
1da177e4 2585 */
029190c5 2586static int proc_cpuset_show(struct seq_file *m, void *unused_v)
1da177e4 2587{
13b41b09 2588 struct pid *pid;
1da177e4
LT
2589 struct task_struct *tsk;
2590 char *buf;
8793d854 2591 struct cgroup_subsys_state *css;
99f89551 2592 int retval;
1da177e4 2593
99f89551 2594 retval = -ENOMEM;
1da177e4
LT
2595 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2596 if (!buf)
99f89551
EB
2597 goto out;
2598
2599 retval = -ESRCH;
13b41b09
EB
2600 pid = m->private;
2601 tsk = get_pid_task(pid, PIDTYPE_PID);
99f89551
EB
2602 if (!tsk)
2603 goto out_free;
1da177e4 2604
99f89551 2605 retval = -EINVAL;
8793d854
PM
2606 cgroup_lock();
2607 css = task_subsys_state(tsk, cpuset_subsys_id);
2608 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
1da177e4 2609 if (retval < 0)
99f89551 2610 goto out_unlock;
1da177e4
LT
2611 seq_puts(m, buf);
2612 seq_putc(m, '\n');
99f89551 2613out_unlock:
8793d854 2614 cgroup_unlock();
99f89551
EB
2615 put_task_struct(tsk);
2616out_free:
1da177e4 2617 kfree(buf);
99f89551 2618out:
1da177e4
LT
2619 return retval;
2620}
2621
2622static int cpuset_open(struct inode *inode, struct file *file)
2623{
13b41b09
EB
2624 struct pid *pid = PROC_I(inode)->pid;
2625 return single_open(file, proc_cpuset_show, pid);
1da177e4
LT
2626}
2627
9a32144e 2628const struct file_operations proc_cpuset_operations = {
1da177e4
LT
2629 .open = cpuset_open,
2630 .read = seq_read,
2631 .llseek = seq_lseek,
2632 .release = single_release,
2633};
8793d854 2634#endif /* CONFIG_PROC_PID_CPUSET */
1da177e4 2635
d01d4827 2636/* Display task mems_allowed in /proc/<pid>/status file. */
df5f8314
EB
2637void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2638{
df5f8314 2639 seq_printf(m, "Mems_allowed:\t");
30e8e136 2640 seq_nodemask(m, &task->mems_allowed);
df5f8314 2641 seq_printf(m, "\n");
39106dcf 2642 seq_printf(m, "Mems_allowed_list:\t");
30e8e136 2643 seq_nodemask_list(m, &task->mems_allowed);
39106dcf 2644 seq_printf(m, "\n");
1da177e4 2645}