cgroup: introduce cgroup_taskset and use it in subsys->can_attach(), cancel_attach...
[linux-2.6-block.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>
9984de1a 40#include <linux/export.h>
1da177e4
LT
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>
60063497 58#include <linux/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{
89e8a244
DR
952 bool masks_disjoint = !nodes_intersects(*newmems, tsk->mems_allowed);
953
c0ff7453
MX
954repeat:
955 /*
956 * Allow tasks that have access to memory reserves because they have
957 * been OOM killed to get memory anywhere.
958 */
959 if (unlikely(test_thread_flag(TIF_MEMDIE)))
960 return;
961 if (current->flags & PF_EXITING) /* Let dying task have memory */
962 return;
963
964 task_lock(tsk);
58568d2a 965 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
c0ff7453
MX
966 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
967
c0ff7453
MX
968 /*
969 * ensure checking ->mems_allowed_change_disable after setting all new
970 * allowed nodes.
971 *
972 * the read-side task can see an nodemask with new allowed nodes and
973 * old allowed nodes. and if it allocates page when cpuset clears newly
974 * disallowed ones continuous, it can see the new allowed bits.
975 *
976 * And if setting all new allowed nodes is after the checking, setting
977 * all new allowed nodes and clearing newly disallowed ones will be done
978 * continuous, and the read-side task may find no node to alloc page.
979 */
980 smp_mb();
981
982 /*
983 * Allocation of memory is very fast, we needn't sleep when waiting
89e8a244
DR
984 * for the read-side. No wait is necessary, however, if at least one
985 * node remains unchanged.
c0ff7453 986 */
89e8a244
DR
987 while (masks_disjoint &&
988 ACCESS_ONCE(tsk->mems_allowed_change_disable)) {
c0ff7453
MX
989 task_unlock(tsk);
990 if (!task_curr(tsk))
991 yield();
992 goto repeat;
993 }
994
995 /*
996 * ensure checking ->mems_allowed_change_disable before clearing all new
997 * disallowed nodes.
998 *
999 * if clearing newly disallowed bits before the checking, the read-side
1000 * task may find no node to alloc page.
1001 */
1002 smp_mb();
1003
1004 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
58568d2a 1005 tsk->mems_allowed = *newmems;
c0ff7453 1006 task_unlock(tsk);
58568d2a
MX
1007}
1008
1009/*
1010 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1011 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1012 * memory_migrate flag is set. Called with cgroup_mutex held.
3b6766fe
LZ
1013 */
1014static void cpuset_change_nodemask(struct task_struct *p,
1015 struct cgroup_scanner *scan)
1016{
1017 struct mm_struct *mm;
1018 struct cpuset *cs;
1019 int migrate;
1020 const nodemask_t *oldmem = scan->data;
ee24d379 1021 static nodemask_t newmems; /* protected by cgroup_mutex */
58568d2a
MX
1022
1023 cs = cgroup_cs(scan->cg);
ee24d379 1024 guarantee_online_mems(cs, &newmems);
58568d2a 1025
ee24d379 1026 cpuset_change_task_nodemask(p, &newmems);
53feb297 1027
3b6766fe
LZ
1028 mm = get_task_mm(p);
1029 if (!mm)
1030 return;
1031
3b6766fe
LZ
1032 migrate = is_memory_migrate(cs);
1033
1034 mpol_rebind_mm(mm, &cs->mems_allowed);
1035 if (migrate)
1036 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1037 mmput(mm);
1038}
1039
8793d854
PM
1040static void *cpuset_being_rebound;
1041
0b2f630a
MX
1042/**
1043 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1044 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1045 * @oldmem: old mems_allowed of cpuset cs
010cfac4 1046 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
0b2f630a
MX
1047 *
1048 * Called with cgroup_mutex held
010cfac4
LZ
1049 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1050 * if @heap != NULL.
0b2f630a 1051 */
010cfac4
LZ
1052static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1053 struct ptr_heap *heap)
1da177e4 1054{
3b6766fe 1055 struct cgroup_scanner scan;
59dac16f 1056
846a16bf 1057 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
4225399a 1058
3b6766fe
LZ
1059 scan.cg = cs->css.cgroup;
1060 scan.test_task = NULL;
1061 scan.process_task = cpuset_change_nodemask;
010cfac4 1062 scan.heap = heap;
3b6766fe 1063 scan.data = (nodemask_t *)oldmem;
4225399a
PJ
1064
1065 /*
3b6766fe
LZ
1066 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1067 * take while holding tasklist_lock. Forks can happen - the
1068 * mpol_dup() cpuset_being_rebound check will catch such forks,
1069 * and rebind their vma mempolicies too. Because we still hold
1070 * the global cgroup_mutex, we know that no other rebind effort
1071 * will be contending for the global variable cpuset_being_rebound.
4225399a 1072 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
04c19fa6 1073 * is idempotent. Also migrate pages in each mm to new nodes.
4225399a 1074 */
010cfac4 1075 cgroup_scan_tasks(&scan);
4225399a 1076
2df167a3 1077 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
8793d854 1078 cpuset_being_rebound = NULL;
1da177e4
LT
1079}
1080
0b2f630a
MX
1081/*
1082 * Handle user request to change the 'mems' memory placement
1083 * of a cpuset. Needs to validate the request, update the
58568d2a
MX
1084 * cpusets mems_allowed, and for each task in the cpuset,
1085 * update mems_allowed and rebind task's mempolicy and any vma
1086 * mempolicies and if the cpuset is marked 'memory_migrate',
1087 * migrate the tasks pages to the new memory.
0b2f630a
MX
1088 *
1089 * Call with cgroup_mutex held. May take callback_mutex during call.
1090 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1091 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1092 * their mempolicies to the cpusets new mems_allowed.
1093 */
645fcc9d
LZ
1094static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1095 const char *buf)
0b2f630a 1096{
53feb297 1097 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
0b2f630a 1098 int retval;
010cfac4 1099 struct ptr_heap heap;
0b2f630a 1100
53feb297
MX
1101 if (!oldmem)
1102 return -ENOMEM;
1103
0b2f630a
MX
1104 /*
1105 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1106 * it's read-only
1107 */
53feb297
MX
1108 if (cs == &top_cpuset) {
1109 retval = -EACCES;
1110 goto done;
1111 }
0b2f630a 1112
0b2f630a
MX
1113 /*
1114 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1115 * Since nodelist_parse() fails on an empty mask, we special case
1116 * that parsing. The validate_change() call ensures that cpusets
1117 * with tasks have memory.
1118 */
1119 if (!*buf) {
645fcc9d 1120 nodes_clear(trialcs->mems_allowed);
0b2f630a 1121 } else {
645fcc9d 1122 retval = nodelist_parse(buf, trialcs->mems_allowed);
0b2f630a
MX
1123 if (retval < 0)
1124 goto done;
1125
645fcc9d 1126 if (!nodes_subset(trialcs->mems_allowed,
53feb297
MX
1127 node_states[N_HIGH_MEMORY])) {
1128 retval = -EINVAL;
1129 goto done;
1130 }
0b2f630a 1131 }
53feb297
MX
1132 *oldmem = cs->mems_allowed;
1133 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
0b2f630a
MX
1134 retval = 0; /* Too easy - nothing to do */
1135 goto done;
1136 }
645fcc9d 1137 retval = validate_change(cs, trialcs);
0b2f630a
MX
1138 if (retval < 0)
1139 goto done;
1140
010cfac4
LZ
1141 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1142 if (retval < 0)
1143 goto done;
1144
0b2f630a 1145 mutex_lock(&callback_mutex);
645fcc9d 1146 cs->mems_allowed = trialcs->mems_allowed;
0b2f630a
MX
1147 mutex_unlock(&callback_mutex);
1148
53feb297 1149 update_tasks_nodemask(cs, oldmem, &heap);
010cfac4
LZ
1150
1151 heap_free(&heap);
0b2f630a 1152done:
53feb297 1153 NODEMASK_FREE(oldmem);
0b2f630a
MX
1154 return retval;
1155}
1156
8793d854
PM
1157int current_cpuset_is_being_rebound(void)
1158{
1159 return task_cs(current) == cpuset_being_rebound;
1160}
1161
5be7a479 1162static int update_relax_domain_level(struct cpuset *cs, s64 val)
1d3504fc 1163{
db7f47cf 1164#ifdef CONFIG_SMP
60495e77 1165 if (val < -1 || val >= sched_domain_level_max)
30e0e178 1166 return -EINVAL;
db7f47cf 1167#endif
1d3504fc
HS
1168
1169 if (val != cs->relax_domain_level) {
1170 cs->relax_domain_level = val;
300ed6cb
LZ
1171 if (!cpumask_empty(cs->cpus_allowed) &&
1172 is_sched_load_balance(cs))
cf417141 1173 async_rebuild_sched_domains();
1d3504fc
HS
1174 }
1175
1176 return 0;
1177}
1178
950592f7
MX
1179/*
1180 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1181 * @tsk: task to be updated
1182 * @scan: struct cgroup_scanner containing the cgroup of the task
1183 *
1184 * Called by cgroup_scan_tasks() for each task in a cgroup.
1185 *
1186 * We don't need to re-check for the cgroup/cpuset membership, since we're
1187 * holding cgroup_lock() at this point.
1188 */
1189static void cpuset_change_flag(struct task_struct *tsk,
1190 struct cgroup_scanner *scan)
1191{
1192 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1193}
1194
1195/*
1196 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1197 * @cs: the cpuset in which each task's spread flags needs to be changed
1198 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1199 *
1200 * Called with cgroup_mutex held
1201 *
1202 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1203 * calling callback functions for each.
1204 *
1205 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1206 * if @heap != NULL.
1207 */
1208static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1209{
1210 struct cgroup_scanner scan;
1211
1212 scan.cg = cs->css.cgroup;
1213 scan.test_task = NULL;
1214 scan.process_task = cpuset_change_flag;
1215 scan.heap = heap;
1216 cgroup_scan_tasks(&scan);
1217}
1218
1da177e4
LT
1219/*
1220 * update_flag - read a 0 or a 1 in a file and update associated flag
78608366
PM
1221 * bit: the bit to update (see cpuset_flagbits_t)
1222 * cs: the cpuset to update
1223 * turning_on: whether the flag is being set or cleared
053199ed 1224 *
2df167a3 1225 * Call with cgroup_mutex held.
1da177e4
LT
1226 */
1227
700fe1ab
PM
1228static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1229 int turning_on)
1da177e4 1230{
645fcc9d 1231 struct cpuset *trialcs;
40b6a762 1232 int balance_flag_changed;
950592f7
MX
1233 int spread_flag_changed;
1234 struct ptr_heap heap;
1235 int err;
1da177e4 1236
645fcc9d
LZ
1237 trialcs = alloc_trial_cpuset(cs);
1238 if (!trialcs)
1239 return -ENOMEM;
1240
1da177e4 1241 if (turning_on)
645fcc9d 1242 set_bit(bit, &trialcs->flags);
1da177e4 1243 else
645fcc9d 1244 clear_bit(bit, &trialcs->flags);
1da177e4 1245
645fcc9d 1246 err = validate_change(cs, trialcs);
85d7b949 1247 if (err < 0)
645fcc9d 1248 goto out;
029190c5 1249
950592f7
MX
1250 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1251 if (err < 0)
1252 goto out;
1253
029190c5 1254 balance_flag_changed = (is_sched_load_balance(cs) !=
645fcc9d 1255 is_sched_load_balance(trialcs));
029190c5 1256
950592f7
MX
1257 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1258 || (is_spread_page(cs) != is_spread_page(trialcs)));
1259
3d3f26a7 1260 mutex_lock(&callback_mutex);
645fcc9d 1261 cs->flags = trialcs->flags;
3d3f26a7 1262 mutex_unlock(&callback_mutex);
85d7b949 1263
300ed6cb 1264 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
cf417141 1265 async_rebuild_sched_domains();
029190c5 1266
950592f7
MX
1267 if (spread_flag_changed)
1268 update_tasks_flags(cs, &heap);
1269 heap_free(&heap);
645fcc9d
LZ
1270out:
1271 free_trial_cpuset(trialcs);
1272 return err;
1da177e4
LT
1273}
1274
3e0d98b9 1275/*
80f7228b 1276 * Frequency meter - How fast is some event occurring?
3e0d98b9
PJ
1277 *
1278 * These routines manage a digitally filtered, constant time based,
1279 * event frequency meter. There are four routines:
1280 * fmeter_init() - initialize a frequency meter.
1281 * fmeter_markevent() - called each time the event happens.
1282 * fmeter_getrate() - returns the recent rate of such events.
1283 * fmeter_update() - internal routine used to update fmeter.
1284 *
1285 * A common data structure is passed to each of these routines,
1286 * which is used to keep track of the state required to manage the
1287 * frequency meter and its digital filter.
1288 *
1289 * The filter works on the number of events marked per unit time.
1290 * The filter is single-pole low-pass recursive (IIR). The time unit
1291 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1292 * simulate 3 decimal digits of precision (multiplied by 1000).
1293 *
1294 * With an FM_COEF of 933, and a time base of 1 second, the filter
1295 * has a half-life of 10 seconds, meaning that if the events quit
1296 * happening, then the rate returned from the fmeter_getrate()
1297 * will be cut in half each 10 seconds, until it converges to zero.
1298 *
1299 * It is not worth doing a real infinitely recursive filter. If more
1300 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1301 * just compute FM_MAXTICKS ticks worth, by which point the level
1302 * will be stable.
1303 *
1304 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1305 * arithmetic overflow in the fmeter_update() routine.
1306 *
1307 * Given the simple 32 bit integer arithmetic used, this meter works
1308 * best for reporting rates between one per millisecond (msec) and
1309 * one per 32 (approx) seconds. At constant rates faster than one
1310 * per msec it maxes out at values just under 1,000,000. At constant
1311 * rates between one per msec, and one per second it will stabilize
1312 * to a value N*1000, where N is the rate of events per second.
1313 * At constant rates between one per second and one per 32 seconds,
1314 * it will be choppy, moving up on the seconds that have an event,
1315 * and then decaying until the next event. At rates slower than
1316 * about one in 32 seconds, it decays all the way back to zero between
1317 * each event.
1318 */
1319
1320#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1321#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1322#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1323#define FM_SCALE 1000 /* faux fixed point scale */
1324
1325/* Initialize a frequency meter */
1326static void fmeter_init(struct fmeter *fmp)
1327{
1328 fmp->cnt = 0;
1329 fmp->val = 0;
1330 fmp->time = 0;
1331 spin_lock_init(&fmp->lock);
1332}
1333
1334/* Internal meter update - process cnt events and update value */
1335static void fmeter_update(struct fmeter *fmp)
1336{
1337 time_t now = get_seconds();
1338 time_t ticks = now - fmp->time;
1339
1340 if (ticks == 0)
1341 return;
1342
1343 ticks = min(FM_MAXTICKS, ticks);
1344 while (ticks-- > 0)
1345 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1346 fmp->time = now;
1347
1348 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1349 fmp->cnt = 0;
1350}
1351
1352/* Process any previous ticks, then bump cnt by one (times scale). */
1353static void fmeter_markevent(struct fmeter *fmp)
1354{
1355 spin_lock(&fmp->lock);
1356 fmeter_update(fmp);
1357 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1358 spin_unlock(&fmp->lock);
1359}
1360
1361/* Process any previous ticks, then return current value. */
1362static int fmeter_getrate(struct fmeter *fmp)
1363{
1364 int val;
1365
1366 spin_lock(&fmp->lock);
1367 fmeter_update(fmp);
1368 val = fmp->val;
1369 spin_unlock(&fmp->lock);
1370 return val;
1371}
1372
2df167a3 1373/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
2f7ee569
TH
1374static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
1375 struct cgroup_taskset *tset)
1da177e4 1376{
2f7ee569 1377 struct cpuset *cs = cgroup_cs(cgrp);
1da177e4 1378
300ed6cb 1379 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1da177e4 1380 return -ENOSPC;
9985b0ba 1381
6d7b2f5f
DR
1382 /*
1383 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1384 * cannot change their cpu affinity and isolating such threads by their
1385 * set of allowed nodes is unnecessary. Thus, cpusets are not
1386 * applicable for such threads. This prevents checking for success of
1387 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1388 * be changed.
1389 */
2f7ee569 1390 if (cgroup_taskset_first(tset)->flags & PF_THREAD_BOUND)
6d7b2f5f 1391 return -EINVAL;
1da177e4 1392
be367d09
BB
1393 return 0;
1394}
1395
f780bdb7
BB
1396static int cpuset_can_attach_task(struct cgroup *cgrp, struct task_struct *task)
1397{
1398 return security_task_setscheduler(task);
1399}
1400
1401/*
1402 * Protected by cgroup_lock. The nodemasks must be stored globally because
1403 * dynamically allocating them is not allowed in pre_attach, and they must
1404 * persist among pre_attach, attach_task, and attach.
1405 */
1406static cpumask_var_t cpus_attach;
1407static nodemask_t cpuset_attach_nodemask_from;
1408static nodemask_t cpuset_attach_nodemask_to;
1409
1410/* Set-up work for before attaching each task. */
1411static void cpuset_pre_attach(struct cgroup *cont)
1412{
1413 struct cpuset *cs = cgroup_cs(cont);
1414
1415 if (cs == &top_cpuset)
1416 cpumask_copy(cpus_attach, cpu_possible_mask);
1417 else
1418 guarantee_online_cpus(cs, cpus_attach);
1419
1420 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1421}
1422
1423/* Per-thread attachment work. */
1424static void cpuset_attach_task(struct cgroup *cont, struct task_struct *tsk)
be367d09
BB
1425{
1426 int err;
f780bdb7
BB
1427 struct cpuset *cs = cgroup_cs(cont);
1428
be367d09
BB
1429 /*
1430 * can_attach beforehand should guarantee that this doesn't fail.
1431 * TODO: have a better way to handle failure here
1432 */
1433 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1434 WARN_ON_ONCE(err);
1435
f780bdb7 1436 cpuset_change_task_nodemask(tsk, &cpuset_attach_nodemask_to);
be367d09 1437 cpuset_update_task_spread_flag(cs, tsk);
8793d854 1438}
1da177e4 1439
2f7ee569
TH
1440static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
1441 struct cgroup_taskset *tset)
8793d854 1442{
8793d854 1443 struct mm_struct *mm;
2f7ee569
TH
1444 struct task_struct *tsk = cgroup_taskset_first(tset);
1445 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1446 struct cpuset *cs = cgroup_cs(cgrp);
1447 struct cpuset *oldcs = cgroup_cs(oldcgrp);
22fb52dd 1448
f780bdb7
BB
1449 /*
1450 * Change mm, possibly for multiple threads in a threadgroup. This is
1451 * expensive and may sleep.
1452 */
1453 cpuset_attach_nodemask_from = oldcs->mems_allowed;
1454 cpuset_attach_nodemask_to = cs->mems_allowed;
4225399a
PJ
1455 mm = get_task_mm(tsk);
1456 if (mm) {
f780bdb7 1457 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2741a559 1458 if (is_memory_migrate(cs))
f780bdb7
BB
1459 cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1460 &cpuset_attach_nodemask_to);
4225399a
PJ
1461 mmput(mm);
1462 }
1da177e4
LT
1463}
1464
1465/* The various types of files and directories in a cpuset file system */
1466
1467typedef enum {
45b07ef3 1468 FILE_MEMORY_MIGRATE,
1da177e4
LT
1469 FILE_CPULIST,
1470 FILE_MEMLIST,
1471 FILE_CPU_EXCLUSIVE,
1472 FILE_MEM_EXCLUSIVE,
78608366 1473 FILE_MEM_HARDWALL,
029190c5 1474 FILE_SCHED_LOAD_BALANCE,
1d3504fc 1475 FILE_SCHED_RELAX_DOMAIN_LEVEL,
3e0d98b9
PJ
1476 FILE_MEMORY_PRESSURE_ENABLED,
1477 FILE_MEMORY_PRESSURE,
825a46af
PJ
1478 FILE_SPREAD_PAGE,
1479 FILE_SPREAD_SLAB,
1da177e4
LT
1480} cpuset_filetype_t;
1481
700fe1ab
PM
1482static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1483{
1484 int retval = 0;
1485 struct cpuset *cs = cgroup_cs(cgrp);
1486 cpuset_filetype_t type = cft->private;
1487
e3712395 1488 if (!cgroup_lock_live_group(cgrp))
700fe1ab 1489 return -ENODEV;
700fe1ab
PM
1490
1491 switch (type) {
1da177e4 1492 case FILE_CPU_EXCLUSIVE:
700fe1ab 1493 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1da177e4
LT
1494 break;
1495 case FILE_MEM_EXCLUSIVE:
700fe1ab 1496 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1da177e4 1497 break;
78608366
PM
1498 case FILE_MEM_HARDWALL:
1499 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1500 break;
029190c5 1501 case FILE_SCHED_LOAD_BALANCE:
700fe1ab 1502 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1d3504fc 1503 break;
45b07ef3 1504 case FILE_MEMORY_MIGRATE:
700fe1ab 1505 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
45b07ef3 1506 break;
3e0d98b9 1507 case FILE_MEMORY_PRESSURE_ENABLED:
700fe1ab 1508 cpuset_memory_pressure_enabled = !!val;
3e0d98b9
PJ
1509 break;
1510 case FILE_MEMORY_PRESSURE:
1511 retval = -EACCES;
1512 break;
825a46af 1513 case FILE_SPREAD_PAGE:
700fe1ab 1514 retval = update_flag(CS_SPREAD_PAGE, cs, val);
825a46af
PJ
1515 break;
1516 case FILE_SPREAD_SLAB:
700fe1ab 1517 retval = update_flag(CS_SPREAD_SLAB, cs, val);
825a46af 1518 break;
1da177e4
LT
1519 default:
1520 retval = -EINVAL;
700fe1ab 1521 break;
1da177e4 1522 }
8793d854 1523 cgroup_unlock();
1da177e4
LT
1524 return retval;
1525}
1526
5be7a479
PM
1527static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1528{
1529 int retval = 0;
1530 struct cpuset *cs = cgroup_cs(cgrp);
1531 cpuset_filetype_t type = cft->private;
1532
e3712395 1533 if (!cgroup_lock_live_group(cgrp))
5be7a479 1534 return -ENODEV;
e3712395 1535
5be7a479
PM
1536 switch (type) {
1537 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1538 retval = update_relax_domain_level(cs, val);
1539 break;
1540 default:
1541 retval = -EINVAL;
1542 break;
1543 }
1544 cgroup_unlock();
1545 return retval;
1546}
1547
e3712395
PM
1548/*
1549 * Common handling for a write to a "cpus" or "mems" file.
1550 */
1551static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1552 const char *buf)
1553{
1554 int retval = 0;
645fcc9d
LZ
1555 struct cpuset *cs = cgroup_cs(cgrp);
1556 struct cpuset *trialcs;
e3712395
PM
1557
1558 if (!cgroup_lock_live_group(cgrp))
1559 return -ENODEV;
1560
645fcc9d 1561 trialcs = alloc_trial_cpuset(cs);
b75f38d6
LZ
1562 if (!trialcs) {
1563 retval = -ENOMEM;
1564 goto out;
1565 }
645fcc9d 1566
e3712395
PM
1567 switch (cft->private) {
1568 case FILE_CPULIST:
645fcc9d 1569 retval = update_cpumask(cs, trialcs, buf);
e3712395
PM
1570 break;
1571 case FILE_MEMLIST:
645fcc9d 1572 retval = update_nodemask(cs, trialcs, buf);
e3712395
PM
1573 break;
1574 default:
1575 retval = -EINVAL;
1576 break;
1577 }
645fcc9d
LZ
1578
1579 free_trial_cpuset(trialcs);
b75f38d6 1580out:
e3712395
PM
1581 cgroup_unlock();
1582 return retval;
1583}
1584
1da177e4
LT
1585/*
1586 * These ascii lists should be read in a single call, by using a user
1587 * buffer large enough to hold the entire map. If read in smaller
1588 * chunks, there is no guarantee of atomicity. Since the display format
1589 * used, list of ranges of sequential numbers, is variable length,
1590 * and since these maps can change value dynamically, one could read
1591 * gibberish by doing partial reads while a list was changing.
1592 * A single large read to a buffer that crosses a page boundary is
1593 * ok, because the result being copied to user land is not recomputed
1594 * across a page fault.
1595 */
1596
9303e0c4 1597static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1da177e4 1598{
9303e0c4 1599 size_t count;
1da177e4 1600
3d3f26a7 1601 mutex_lock(&callback_mutex);
9303e0c4 1602 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
3d3f26a7 1603 mutex_unlock(&callback_mutex);
1da177e4 1604
9303e0c4 1605 return count;
1da177e4
LT
1606}
1607
9303e0c4 1608static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1da177e4 1609{
9303e0c4 1610 size_t count;
1da177e4 1611
3d3f26a7 1612 mutex_lock(&callback_mutex);
9303e0c4 1613 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
3d3f26a7 1614 mutex_unlock(&callback_mutex);
1da177e4 1615
9303e0c4 1616 return count;
1da177e4
LT
1617}
1618
8793d854
PM
1619static ssize_t cpuset_common_file_read(struct cgroup *cont,
1620 struct cftype *cft,
1621 struct file *file,
1622 char __user *buf,
1623 size_t nbytes, loff_t *ppos)
1da177e4 1624{
8793d854 1625 struct cpuset *cs = cgroup_cs(cont);
1da177e4
LT
1626 cpuset_filetype_t type = cft->private;
1627 char *page;
1628 ssize_t retval = 0;
1629 char *s;
1da177e4 1630
e12ba74d 1631 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1da177e4
LT
1632 return -ENOMEM;
1633
1634 s = page;
1635
1636 switch (type) {
1637 case FILE_CPULIST:
1638 s += cpuset_sprintf_cpulist(s, cs);
1639 break;
1640 case FILE_MEMLIST:
1641 s += cpuset_sprintf_memlist(s, cs);
1642 break;
1da177e4
LT
1643 default:
1644 retval = -EINVAL;
1645 goto out;
1646 }
1647 *s++ = '\n';
1da177e4 1648
eacaa1f5 1649 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1da177e4
LT
1650out:
1651 free_page((unsigned long)page);
1652 return retval;
1653}
1654
700fe1ab
PM
1655static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1656{
1657 struct cpuset *cs = cgroup_cs(cont);
1658 cpuset_filetype_t type = cft->private;
1659 switch (type) {
1660 case FILE_CPU_EXCLUSIVE:
1661 return is_cpu_exclusive(cs);
1662 case FILE_MEM_EXCLUSIVE:
1663 return is_mem_exclusive(cs);
78608366
PM
1664 case FILE_MEM_HARDWALL:
1665 return is_mem_hardwall(cs);
700fe1ab
PM
1666 case FILE_SCHED_LOAD_BALANCE:
1667 return is_sched_load_balance(cs);
1668 case FILE_MEMORY_MIGRATE:
1669 return is_memory_migrate(cs);
1670 case FILE_MEMORY_PRESSURE_ENABLED:
1671 return cpuset_memory_pressure_enabled;
1672 case FILE_MEMORY_PRESSURE:
1673 return fmeter_getrate(&cs->fmeter);
1674 case FILE_SPREAD_PAGE:
1675 return is_spread_page(cs);
1676 case FILE_SPREAD_SLAB:
1677 return is_spread_slab(cs);
1678 default:
1679 BUG();
1680 }
cf417141
MK
1681
1682 /* Unreachable but makes gcc happy */
1683 return 0;
700fe1ab 1684}
1da177e4 1685
5be7a479
PM
1686static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1687{
1688 struct cpuset *cs = cgroup_cs(cont);
1689 cpuset_filetype_t type = cft->private;
1690 switch (type) {
1691 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1692 return cs->relax_domain_level;
1693 default:
1694 BUG();
1695 }
cf417141
MK
1696
1697 /* Unrechable but makes gcc happy */
1698 return 0;
5be7a479
PM
1699}
1700
1da177e4
LT
1701
1702/*
1703 * for the common functions, 'private' gives the type of file
1704 */
1705
addf2c73
PM
1706static struct cftype files[] = {
1707 {
1708 .name = "cpus",
1709 .read = cpuset_common_file_read,
e3712395
PM
1710 .write_string = cpuset_write_resmask,
1711 .max_write_len = (100U + 6 * NR_CPUS),
addf2c73
PM
1712 .private = FILE_CPULIST,
1713 },
1714
1715 {
1716 .name = "mems",
1717 .read = cpuset_common_file_read,
e3712395
PM
1718 .write_string = cpuset_write_resmask,
1719 .max_write_len = (100U + 6 * MAX_NUMNODES),
addf2c73
PM
1720 .private = FILE_MEMLIST,
1721 },
1722
1723 {
1724 .name = "cpu_exclusive",
1725 .read_u64 = cpuset_read_u64,
1726 .write_u64 = cpuset_write_u64,
1727 .private = FILE_CPU_EXCLUSIVE,
1728 },
1729
1730 {
1731 .name = "mem_exclusive",
1732 .read_u64 = cpuset_read_u64,
1733 .write_u64 = cpuset_write_u64,
1734 .private = FILE_MEM_EXCLUSIVE,
1735 },
1736
78608366
PM
1737 {
1738 .name = "mem_hardwall",
1739 .read_u64 = cpuset_read_u64,
1740 .write_u64 = cpuset_write_u64,
1741 .private = FILE_MEM_HARDWALL,
1742 },
1743
addf2c73
PM
1744 {
1745 .name = "sched_load_balance",
1746 .read_u64 = cpuset_read_u64,
1747 .write_u64 = cpuset_write_u64,
1748 .private = FILE_SCHED_LOAD_BALANCE,
1749 },
1750
1751 {
1752 .name = "sched_relax_domain_level",
5be7a479
PM
1753 .read_s64 = cpuset_read_s64,
1754 .write_s64 = cpuset_write_s64,
addf2c73
PM
1755 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1756 },
1757
1758 {
1759 .name = "memory_migrate",
1760 .read_u64 = cpuset_read_u64,
1761 .write_u64 = cpuset_write_u64,
1762 .private = FILE_MEMORY_MIGRATE,
1763 },
1764
1765 {
1766 .name = "memory_pressure",
1767 .read_u64 = cpuset_read_u64,
1768 .write_u64 = cpuset_write_u64,
1769 .private = FILE_MEMORY_PRESSURE,
099fca32 1770 .mode = S_IRUGO,
addf2c73
PM
1771 },
1772
1773 {
1774 .name = "memory_spread_page",
1775 .read_u64 = cpuset_read_u64,
1776 .write_u64 = cpuset_write_u64,
1777 .private = FILE_SPREAD_PAGE,
1778 },
1779
1780 {
1781 .name = "memory_spread_slab",
1782 .read_u64 = cpuset_read_u64,
1783 .write_u64 = cpuset_write_u64,
1784 .private = FILE_SPREAD_SLAB,
1785 },
45b07ef3
PJ
1786};
1787
3e0d98b9
PJ
1788static struct cftype cft_memory_pressure_enabled = {
1789 .name = "memory_pressure_enabled",
700fe1ab
PM
1790 .read_u64 = cpuset_read_u64,
1791 .write_u64 = cpuset_write_u64,
3e0d98b9
PJ
1792 .private = FILE_MEMORY_PRESSURE_ENABLED,
1793};
1794
8793d854 1795static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1da177e4
LT
1796{
1797 int err;
1798
addf2c73
PM
1799 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1800 if (err)
1da177e4 1801 return err;
8793d854 1802 /* memory_pressure_enabled is in root cpuset only */
addf2c73 1803 if (!cont->parent)
8793d854 1804 err = cgroup_add_file(cont, ss,
addf2c73
PM
1805 &cft_memory_pressure_enabled);
1806 return err;
1da177e4
LT
1807}
1808
8793d854 1809/*
a77aea92
DL
1810 * post_clone() is called during cgroup_create() when the
1811 * clone_children mount argument was specified. The cgroup
1812 * can not yet have any tasks.
8793d854
PM
1813 *
1814 * Currently we refuse to set up the cgroup - thereby
1815 * refusing the task to be entered, and as a result refusing
1816 * the sys_unshare() or clone() which initiated it - if any
1817 * sibling cpusets have exclusive cpus or mem.
1818 *
1819 * If this becomes a problem for some users who wish to
1820 * allow that scenario, then cpuset_post_clone() could be
1821 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2df167a3
PM
1822 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1823 * held.
8793d854
PM
1824 */
1825static void cpuset_post_clone(struct cgroup_subsys *ss,
1826 struct cgroup *cgroup)
1827{
1828 struct cgroup *parent, *child;
1829 struct cpuset *cs, *parent_cs;
1830
1831 parent = cgroup->parent;
1832 list_for_each_entry(child, &parent->children, sibling) {
1833 cs = cgroup_cs(child);
1834 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1835 return;
1836 }
1837 cs = cgroup_cs(cgroup);
1838 parent_cs = cgroup_cs(parent);
1839
523fb486 1840 mutex_lock(&callback_mutex);
8793d854 1841 cs->mems_allowed = parent_cs->mems_allowed;
300ed6cb 1842 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
523fb486 1843 mutex_unlock(&callback_mutex);
8793d854
PM
1844 return;
1845}
1846
1da177e4
LT
1847/*
1848 * cpuset_create - create a cpuset
2df167a3
PM
1849 * ss: cpuset cgroup subsystem
1850 * cont: control group that the new cpuset will be part of
1da177e4
LT
1851 */
1852
8793d854
PM
1853static struct cgroup_subsys_state *cpuset_create(
1854 struct cgroup_subsys *ss,
1855 struct cgroup *cont)
1da177e4
LT
1856{
1857 struct cpuset *cs;
8793d854 1858 struct cpuset *parent;
1da177e4 1859
8793d854 1860 if (!cont->parent) {
8793d854
PM
1861 return &top_cpuset.css;
1862 }
1863 parent = cgroup_cs(cont->parent);
1da177e4
LT
1864 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1865 if (!cs)
8793d854 1866 return ERR_PTR(-ENOMEM);
300ed6cb
LZ
1867 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1868 kfree(cs);
1869 return ERR_PTR(-ENOMEM);
1870 }
1da177e4 1871
1da177e4 1872 cs->flags = 0;
825a46af
PJ
1873 if (is_spread_page(parent))
1874 set_bit(CS_SPREAD_PAGE, &cs->flags);
1875 if (is_spread_slab(parent))
1876 set_bit(CS_SPREAD_SLAB, &cs->flags);
029190c5 1877 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
300ed6cb 1878 cpumask_clear(cs->cpus_allowed);
f9a86fcb 1879 nodes_clear(cs->mems_allowed);
3e0d98b9 1880 fmeter_init(&cs->fmeter);
1d3504fc 1881 cs->relax_domain_level = -1;
1da177e4
LT
1882
1883 cs->parent = parent;
202f72d5 1884 number_of_cpusets++;
8793d854 1885 return &cs->css ;
1da177e4
LT
1886}
1887
029190c5 1888/*
029190c5
PJ
1889 * If the cpuset being removed has its flag 'sched_load_balance'
1890 * enabled, then simulate turning sched_load_balance off, which
cf417141 1891 * will call async_rebuild_sched_domains().
029190c5
PJ
1892 */
1893
8793d854 1894static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1da177e4 1895{
8793d854 1896 struct cpuset *cs = cgroup_cs(cont);
1da177e4 1897
029190c5 1898 if (is_sched_load_balance(cs))
700fe1ab 1899 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
029190c5 1900
202f72d5 1901 number_of_cpusets--;
300ed6cb 1902 free_cpumask_var(cs->cpus_allowed);
8793d854 1903 kfree(cs);
1da177e4
LT
1904}
1905
8793d854
PM
1906struct cgroup_subsys cpuset_subsys = {
1907 .name = "cpuset",
1908 .create = cpuset_create,
cf417141 1909 .destroy = cpuset_destroy,
8793d854 1910 .can_attach = cpuset_can_attach,
f780bdb7
BB
1911 .can_attach_task = cpuset_can_attach_task,
1912 .pre_attach = cpuset_pre_attach,
1913 .attach_task = cpuset_attach_task,
8793d854
PM
1914 .attach = cpuset_attach,
1915 .populate = cpuset_populate,
1916 .post_clone = cpuset_post_clone,
1917 .subsys_id = cpuset_subsys_id,
1918 .early_init = 1,
1919};
1920
1da177e4
LT
1921/**
1922 * cpuset_init - initialize cpusets at system boot
1923 *
1924 * Description: Initialize top_cpuset and the cpuset internal file system,
1925 **/
1926
1927int __init cpuset_init(void)
1928{
8793d854 1929 int err = 0;
1da177e4 1930
58568d2a
MX
1931 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1932 BUG();
1933
300ed6cb 1934 cpumask_setall(top_cpuset.cpus_allowed);
f9a86fcb 1935 nodes_setall(top_cpuset.mems_allowed);
1da177e4 1936
3e0d98b9 1937 fmeter_init(&top_cpuset.fmeter);
029190c5 1938 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1d3504fc 1939 top_cpuset.relax_domain_level = -1;
1da177e4 1940
1da177e4
LT
1941 err = register_filesystem(&cpuset_fs_type);
1942 if (err < 0)
8793d854
PM
1943 return err;
1944
2341d1b6
LZ
1945 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1946 BUG();
1947
202f72d5 1948 number_of_cpusets = 1;
8793d854 1949 return 0;
1da177e4
LT
1950}
1951
956db3ca
CW
1952/**
1953 * cpuset_do_move_task - move a given task to another cpuset
1954 * @tsk: pointer to task_struct the task to move
1955 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1956 *
1957 * Called by cgroup_scan_tasks() for each task in a cgroup.
1958 * Return nonzero to stop the walk through the tasks.
1959 */
9e0c914c
AB
1960static void cpuset_do_move_task(struct task_struct *tsk,
1961 struct cgroup_scanner *scan)
956db3ca 1962{
7f81b1ae 1963 struct cgroup *new_cgroup = scan->data;
956db3ca 1964
7f81b1ae 1965 cgroup_attach_task(new_cgroup, tsk);
956db3ca
CW
1966}
1967
1968/**
1969 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1970 * @from: cpuset in which the tasks currently reside
1971 * @to: cpuset to which the tasks will be moved
1972 *
c8d9c90c
PJ
1973 * Called with cgroup_mutex held
1974 * callback_mutex must not be held, as cpuset_attach() will take it.
956db3ca
CW
1975 *
1976 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1977 * calling callback functions for each.
1978 */
1979static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1980{
7f81b1ae 1981 struct cgroup_scanner scan;
956db3ca 1982
7f81b1ae
LZ
1983 scan.cg = from->css.cgroup;
1984 scan.test_task = NULL; /* select all tasks in cgroup */
1985 scan.process_task = cpuset_do_move_task;
1986 scan.heap = NULL;
1987 scan.data = to->css.cgroup;
956db3ca 1988
7f81b1ae 1989 if (cgroup_scan_tasks(&scan))
956db3ca
CW
1990 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1991 "cgroup_scan_tasks failed\n");
1992}
1993
b1aac8bb 1994/*
cf417141 1995 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
b1aac8bb
PJ
1996 * or memory nodes, we need to walk over the cpuset hierarchy,
1997 * removing that CPU or node from all cpusets. If this removes the
956db3ca
CW
1998 * last CPU or node from a cpuset, then move the tasks in the empty
1999 * cpuset to its next-highest non-empty parent.
b1aac8bb 2000 *
c8d9c90c
PJ
2001 * Called with cgroup_mutex held
2002 * callback_mutex must not be held, as cpuset_attach() will take it.
b1aac8bb 2003 */
956db3ca
CW
2004static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2005{
2006 struct cpuset *parent;
2007
c8d9c90c
PJ
2008 /*
2009 * The cgroup's css_sets list is in use if there are tasks
2010 * in the cpuset; the list is empty if there are none;
2011 * the cs->css.refcnt seems always 0.
2012 */
956db3ca
CW
2013 if (list_empty(&cs->css.cgroup->css_sets))
2014 return;
b1aac8bb 2015
956db3ca
CW
2016 /*
2017 * Find its next-highest non-empty parent, (top cpuset
2018 * has online cpus, so can't be empty).
2019 */
2020 parent = cs->parent;
300ed6cb 2021 while (cpumask_empty(parent->cpus_allowed) ||
b4501295 2022 nodes_empty(parent->mems_allowed))
956db3ca 2023 parent = parent->parent;
956db3ca
CW
2024
2025 move_member_tasks_to_cpuset(cs, parent);
2026}
2027
2028/*
2029 * Walk the specified cpuset subtree and look for empty cpusets.
2030 * The tasks of such cpuset must be moved to a parent cpuset.
2031 *
2df167a3 2032 * Called with cgroup_mutex held. We take callback_mutex to modify
956db3ca
CW
2033 * cpus_allowed and mems_allowed.
2034 *
2035 * This walk processes the tree from top to bottom, completing one layer
2036 * before dropping down to the next. It always processes a node before
2037 * any of its children.
2038 *
2039 * For now, since we lack memory hot unplug, we'll never see a cpuset
2040 * that has tasks along with an empty 'mems'. But if we did see such
2041 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2042 */
d294eb83 2043static void scan_for_empty_cpusets(struct cpuset *root)
b1aac8bb 2044{
8d1e6266 2045 LIST_HEAD(queue);
956db3ca
CW
2046 struct cpuset *cp; /* scans cpusets being updated */
2047 struct cpuset *child; /* scans child cpusets of cp */
8793d854 2048 struct cgroup *cont;
ee24d379 2049 static nodemask_t oldmems; /* protected by cgroup_mutex */
b1aac8bb 2050
956db3ca
CW
2051 list_add_tail((struct list_head *)&root->stack_list, &queue);
2052
956db3ca 2053 while (!list_empty(&queue)) {
8d1e6266 2054 cp = list_first_entry(&queue, struct cpuset, stack_list);
956db3ca
CW
2055 list_del(queue.next);
2056 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2057 child = cgroup_cs(cont);
2058 list_add_tail(&child->stack_list, &queue);
2059 }
b4501295
PJ
2060
2061 /* Continue past cpusets with all cpus, mems online */
6ad4c188 2062 if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
b4501295
PJ
2063 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2064 continue;
2065
ee24d379 2066 oldmems = cp->mems_allowed;
f9b4fb8d 2067
956db3ca 2068 /* Remove offline cpus and mems from this cpuset. */
b4501295 2069 mutex_lock(&callback_mutex);
300ed6cb 2070 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
6ad4c188 2071 cpu_active_mask);
956db3ca
CW
2072 nodes_and(cp->mems_allowed, cp->mems_allowed,
2073 node_states[N_HIGH_MEMORY]);
b4501295
PJ
2074 mutex_unlock(&callback_mutex);
2075
2076 /* Move tasks from the empty cpuset to a parent */
300ed6cb 2077 if (cpumask_empty(cp->cpus_allowed) ||
b4501295 2078 nodes_empty(cp->mems_allowed))
956db3ca 2079 remove_tasks_in_empty_cpuset(cp);
f9b4fb8d 2080 else {
4e74339a 2081 update_tasks_cpumask(cp, NULL);
ee24d379 2082 update_tasks_nodemask(cp, &oldmems, NULL);
f9b4fb8d 2083 }
b1aac8bb
PJ
2084 }
2085}
2086
4c4d50f7
PJ
2087/*
2088 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2089 * period. This is necessary in order to make cpusets transparent
2090 * (of no affect) on systems that are actively using CPU hotplug
2091 * but making no active use of cpusets.
2092 *
38837fc7 2093 * This routine ensures that top_cpuset.cpus_allowed tracks
3a101d05 2094 * cpu_active_mask on each CPU hotplug (cpuhp) event.
cf417141
MK
2095 *
2096 * Called within get_online_cpus(). Needs to call cgroup_lock()
2097 * before calling generate_sched_domains().
4c4d50f7 2098 */
0b2e918a 2099void cpuset_update_active_cpus(void)
4c4d50f7 2100{
cf417141 2101 struct sched_domain_attr *attr;
acc3f5d7 2102 cpumask_var_t *doms;
cf417141
MK
2103 int ndoms;
2104
cf417141 2105 cgroup_lock();
0b4217b3 2106 mutex_lock(&callback_mutex);
6ad4c188 2107 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
0b4217b3 2108 mutex_unlock(&callback_mutex);
cf417141
MK
2109 scan_for_empty_cpusets(&top_cpuset);
2110 ndoms = generate_sched_domains(&doms, &attr);
2111 cgroup_unlock();
2112
2113 /* Have scheduler rebuild the domains */
2114 partition_sched_domains(ndoms, doms, attr);
4c4d50f7 2115}
4c4d50f7 2116
b1aac8bb 2117#ifdef CONFIG_MEMORY_HOTPLUG
38837fc7 2118/*
0e1e7c7a 2119 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
cf417141
MK
2120 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2121 * See also the previous routine cpuset_track_online_cpus().
38837fc7 2122 */
f481891f
MX
2123static int cpuset_track_online_nodes(struct notifier_block *self,
2124 unsigned long action, void *arg)
38837fc7 2125{
ee24d379 2126 static nodemask_t oldmems; /* protected by cgroup_mutex */
5ab116c9 2127
cf417141 2128 cgroup_lock();
f481891f
MX
2129 switch (action) {
2130 case MEM_ONLINE:
ee24d379 2131 oldmems = top_cpuset.mems_allowed;
0b4217b3 2132 mutex_lock(&callback_mutex);
f481891f 2133 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
0b4217b3 2134 mutex_unlock(&callback_mutex);
ee24d379 2135 update_tasks_nodemask(&top_cpuset, &oldmems, NULL);
5ab116c9
MX
2136 break;
2137 case MEM_OFFLINE:
2138 /*
2139 * needn't update top_cpuset.mems_allowed explicitly because
2140 * scan_for_empty_cpusets() will update it.
2141 */
2142 scan_for_empty_cpusets(&top_cpuset);
f481891f
MX
2143 break;
2144 default:
2145 break;
2146 }
cf417141 2147 cgroup_unlock();
53feb297 2148
f481891f 2149 return NOTIFY_OK;
38837fc7
PJ
2150}
2151#endif
2152
1da177e4
LT
2153/**
2154 * cpuset_init_smp - initialize cpus_allowed
2155 *
2156 * Description: Finish top cpuset after cpu, node maps are initialized
2157 **/
2158
2159void __init cpuset_init_smp(void)
2160{
6ad4c188 2161 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
0e1e7c7a 2162 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
4c4d50f7 2163
f481891f 2164 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
f90d4118
MX
2165
2166 cpuset_wq = create_singlethread_workqueue("cpuset");
2167 BUG_ON(!cpuset_wq);
1da177e4
LT
2168}
2169
2170/**
1da177e4
LT
2171 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2172 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
6af866af 2173 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
1da177e4 2174 *
300ed6cb 2175 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
1da177e4
LT
2176 * attached to the specified @tsk. Guaranteed to return some non-empty
2177 * subset of cpu_online_map, even if this means going outside the
2178 * tasks cpuset.
2179 **/
2180
6af866af 2181void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
1da177e4 2182{
3d3f26a7 2183 mutex_lock(&callback_mutex);
909d75a3 2184 task_lock(tsk);
f9a86fcb 2185 guarantee_online_cpus(task_cs(tsk), pmask);
909d75a3 2186 task_unlock(tsk);
897f0b3c 2187 mutex_unlock(&callback_mutex);
1da177e4
LT
2188}
2189
9084bb82
ON
2190int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2191{
2192 const struct cpuset *cs;
2193 int cpu;
2194
2195 rcu_read_lock();
2196 cs = task_cs(tsk);
2197 if (cs)
1e1b6c51 2198 do_set_cpus_allowed(tsk, cs->cpus_allowed);
9084bb82
ON
2199 rcu_read_unlock();
2200
2201 /*
2202 * We own tsk->cpus_allowed, nobody can change it under us.
2203 *
2204 * But we used cs && cs->cpus_allowed lockless and thus can
2205 * race with cgroup_attach_task() or update_cpumask() and get
2206 * the wrong tsk->cpus_allowed. However, both cases imply the
2207 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2208 * which takes task_rq_lock().
2209 *
2210 * If we are called after it dropped the lock we must see all
2211 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2212 * set any mask even if it is not right from task_cs() pov,
2213 * the pending set_cpus_allowed_ptr() will fix things.
2214 */
2215
2216 cpu = cpumask_any_and(&tsk->cpus_allowed, cpu_active_mask);
2217 if (cpu >= nr_cpu_ids) {
2218 /*
2219 * Either tsk->cpus_allowed is wrong (see above) or it
2220 * is actually empty. The latter case is only possible
2221 * if we are racing with remove_tasks_in_empty_cpuset().
2222 * Like above we can temporary set any mask and rely on
2223 * set_cpus_allowed_ptr() as synchronization point.
2224 */
1e1b6c51 2225 do_set_cpus_allowed(tsk, cpu_possible_mask);
9084bb82
ON
2226 cpu = cpumask_any(cpu_active_mask);
2227 }
2228
2229 return cpu;
2230}
2231
1da177e4
LT
2232void cpuset_init_current_mems_allowed(void)
2233{
f9a86fcb 2234 nodes_setall(current->mems_allowed);
1da177e4
LT
2235}
2236
909d75a3
PJ
2237/**
2238 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2239 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2240 *
2241 * Description: Returns the nodemask_t mems_allowed of the cpuset
2242 * attached to the specified @tsk. Guaranteed to return some non-empty
0e1e7c7a 2243 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
909d75a3
PJ
2244 * tasks cpuset.
2245 **/
2246
2247nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2248{
2249 nodemask_t mask;
2250
3d3f26a7 2251 mutex_lock(&callback_mutex);
909d75a3 2252 task_lock(tsk);
8793d854 2253 guarantee_online_mems(task_cs(tsk), &mask);
909d75a3 2254 task_unlock(tsk);
3d3f26a7 2255 mutex_unlock(&callback_mutex);
909d75a3
PJ
2256
2257 return mask;
2258}
2259
d9fd8a6d 2260/**
19770b32
MG
2261 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2262 * @nodemask: the nodemask to be checked
d9fd8a6d 2263 *
19770b32 2264 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
1da177e4 2265 */
19770b32 2266int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
1da177e4 2267{
19770b32 2268 return nodes_intersects(*nodemask, current->mems_allowed);
1da177e4
LT
2269}
2270
9bf2229f 2271/*
78608366
PM
2272 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2273 * mem_hardwall ancestor to the specified cpuset. Call holding
2274 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2275 * (an unusual configuration), then returns the root cpuset.
9bf2229f 2276 */
78608366 2277static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
9bf2229f 2278{
78608366 2279 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
9bf2229f
PJ
2280 cs = cs->parent;
2281 return cs;
2282}
2283
d9fd8a6d 2284/**
a1bc5a4e
DR
2285 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2286 * @node: is this an allowed node?
02a0e53d 2287 * @gfp_mask: memory allocation flags
d9fd8a6d 2288 *
a1bc5a4e
DR
2289 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2290 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2291 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2292 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2293 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2294 * flag, yes.
9bf2229f
PJ
2295 * Otherwise, no.
2296 *
a1bc5a4e
DR
2297 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2298 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2299 * might sleep, and might allow a node from an enclosing cpuset.
02a0e53d 2300 *
a1bc5a4e
DR
2301 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2302 * cpusets, and never sleeps.
02a0e53d
PJ
2303 *
2304 * The __GFP_THISNODE placement logic is really handled elsewhere,
2305 * by forcibly using a zonelist starting at a specified node, and by
2306 * (in get_page_from_freelist()) refusing to consider the zones for
2307 * any node on the zonelist except the first. By the time any such
2308 * calls get to this routine, we should just shut up and say 'yes'.
2309 *
9bf2229f 2310 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
c596d9f3
DR
2311 * and do not allow allocations outside the current tasks cpuset
2312 * unless the task has been OOM killed as is marked TIF_MEMDIE.
9bf2229f 2313 * GFP_KERNEL allocations are not so marked, so can escape to the
78608366 2314 * nearest enclosing hardwalled ancestor cpuset.
9bf2229f 2315 *
02a0e53d
PJ
2316 * Scanning up parent cpusets requires callback_mutex. The
2317 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2318 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2319 * current tasks mems_allowed came up empty on the first pass over
2320 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2321 * cpuset are short of memory, might require taking the callback_mutex
2322 * mutex.
9bf2229f 2323 *
36be57ff 2324 * The first call here from mm/page_alloc:get_page_from_freelist()
02a0e53d
PJ
2325 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2326 * so no allocation on a node outside the cpuset is allowed (unless
2327 * in interrupt, of course).
36be57ff
PJ
2328 *
2329 * The second pass through get_page_from_freelist() doesn't even call
2330 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2331 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2332 * in alloc_flags. That logic and the checks below have the combined
2333 * affect that:
9bf2229f
PJ
2334 * in_interrupt - any node ok (current task context irrelevant)
2335 * GFP_ATOMIC - any node ok
c596d9f3 2336 * TIF_MEMDIE - any node ok
78608366 2337 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
9bf2229f 2338 * GFP_USER - only nodes in current tasks mems allowed ok.
36be57ff
PJ
2339 *
2340 * Rule:
a1bc5a4e 2341 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
36be57ff
PJ
2342 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2343 * the code that might scan up ancestor cpusets and sleep.
02a0e53d 2344 */
a1bc5a4e 2345int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
1da177e4 2346{
9bf2229f 2347 const struct cpuset *cs; /* current cpuset ancestors */
29afd49b 2348 int allowed; /* is allocation in zone z allowed? */
9bf2229f 2349
9b819d20 2350 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
9bf2229f 2351 return 1;
92d1dbd2 2352 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
9bf2229f
PJ
2353 if (node_isset(node, current->mems_allowed))
2354 return 1;
c596d9f3
DR
2355 /*
2356 * Allow tasks that have access to memory reserves because they have
2357 * been OOM killed to get memory anywhere.
2358 */
2359 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2360 return 1;
9bf2229f
PJ
2361 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2362 return 0;
2363
5563e770
BP
2364 if (current->flags & PF_EXITING) /* Let dying task have memory */
2365 return 1;
2366
9bf2229f 2367 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3d3f26a7 2368 mutex_lock(&callback_mutex);
053199ed 2369
053199ed 2370 task_lock(current);
78608366 2371 cs = nearest_hardwall_ancestor(task_cs(current));
053199ed
PJ
2372 task_unlock(current);
2373
9bf2229f 2374 allowed = node_isset(node, cs->mems_allowed);
3d3f26a7 2375 mutex_unlock(&callback_mutex);
9bf2229f 2376 return allowed;
1da177e4
LT
2377}
2378
02a0e53d 2379/*
a1bc5a4e
DR
2380 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2381 * @node: is this an allowed node?
02a0e53d
PJ
2382 * @gfp_mask: memory allocation flags
2383 *
a1bc5a4e
DR
2384 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2385 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2386 * yes. If the task has been OOM killed and has access to memory reserves as
2387 * specified by the TIF_MEMDIE flag, yes.
2388 * Otherwise, no.
02a0e53d
PJ
2389 *
2390 * The __GFP_THISNODE placement logic is really handled elsewhere,
2391 * by forcibly using a zonelist starting at a specified node, and by
2392 * (in get_page_from_freelist()) refusing to consider the zones for
2393 * any node on the zonelist except the first. By the time any such
2394 * calls get to this routine, we should just shut up and say 'yes'.
2395 *
a1bc5a4e
DR
2396 * Unlike the cpuset_node_allowed_softwall() variant, above,
2397 * this variant requires that the node be in the current task's
02a0e53d
PJ
2398 * mems_allowed or that we're in interrupt. It does not scan up the
2399 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2400 * It never sleeps.
2401 */
a1bc5a4e 2402int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
02a0e53d 2403{
02a0e53d
PJ
2404 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2405 return 1;
02a0e53d
PJ
2406 if (node_isset(node, current->mems_allowed))
2407 return 1;
dedf8b79
DW
2408 /*
2409 * Allow tasks that have access to memory reserves because they have
2410 * been OOM killed to get memory anywhere.
2411 */
2412 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2413 return 1;
02a0e53d
PJ
2414 return 0;
2415}
2416
505970b9
PJ
2417/**
2418 * cpuset_unlock - release lock on cpuset changes
2419 *
2420 * Undo the lock taken in a previous cpuset_lock() call.
2421 */
2422
2423void cpuset_unlock(void)
2424{
3d3f26a7 2425 mutex_unlock(&callback_mutex);
505970b9
PJ
2426}
2427
825a46af 2428/**
6adef3eb
JS
2429 * cpuset_mem_spread_node() - On which node to begin search for a file page
2430 * cpuset_slab_spread_node() - On which node to begin search for a slab page
825a46af
PJ
2431 *
2432 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2433 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2434 * and if the memory allocation used cpuset_mem_spread_node()
2435 * to determine on which node to start looking, as it will for
2436 * certain page cache or slab cache pages such as used for file
2437 * system buffers and inode caches, then instead of starting on the
2438 * local node to look for a free page, rather spread the starting
2439 * node around the tasks mems_allowed nodes.
2440 *
2441 * We don't have to worry about the returned node being offline
2442 * because "it can't happen", and even if it did, it would be ok.
2443 *
2444 * The routines calling guarantee_online_mems() are careful to
2445 * only set nodes in task->mems_allowed that are online. So it
2446 * should not be possible for the following code to return an
2447 * offline node. But if it did, that would be ok, as this routine
2448 * is not returning the node where the allocation must be, only
2449 * the node where the search should start. The zonelist passed to
2450 * __alloc_pages() will include all nodes. If the slab allocator
2451 * is passed an offline node, it will fall back to the local node.
2452 * See kmem_cache_alloc_node().
2453 */
2454
6adef3eb 2455static int cpuset_spread_node(int *rotor)
825a46af
PJ
2456{
2457 int node;
2458
6adef3eb 2459 node = next_node(*rotor, current->mems_allowed);
825a46af
PJ
2460 if (node == MAX_NUMNODES)
2461 node = first_node(current->mems_allowed);
6adef3eb 2462 *rotor = node;
825a46af
PJ
2463 return node;
2464}
6adef3eb
JS
2465
2466int cpuset_mem_spread_node(void)
2467{
778d3b0f
MH
2468 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2469 current->cpuset_mem_spread_rotor =
2470 node_random(&current->mems_allowed);
2471
6adef3eb
JS
2472 return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2473}
2474
2475int cpuset_slab_spread_node(void)
2476{
778d3b0f
MH
2477 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2478 current->cpuset_slab_spread_rotor =
2479 node_random(&current->mems_allowed);
2480
6adef3eb
JS
2481 return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2482}
2483
825a46af
PJ
2484EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2485
ef08e3b4 2486/**
bbe373f2
DR
2487 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2488 * @tsk1: pointer to task_struct of some task.
2489 * @tsk2: pointer to task_struct of some other task.
2490 *
2491 * Description: Return true if @tsk1's mems_allowed intersects the
2492 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2493 * one of the task's memory usage might impact the memory available
2494 * to the other.
ef08e3b4
PJ
2495 **/
2496
bbe373f2
DR
2497int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2498 const struct task_struct *tsk2)
ef08e3b4 2499{
bbe373f2 2500 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
ef08e3b4
PJ
2501}
2502
75aa1994
DR
2503/**
2504 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2505 * @task: pointer to task_struct of some task.
2506 *
2507 * Description: Prints @task's name, cpuset name, and cached copy of its
2508 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2509 * dereferencing task_cs(task).
2510 */
2511void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2512{
2513 struct dentry *dentry;
2514
2515 dentry = task_cs(tsk)->css.cgroup->dentry;
2516 spin_lock(&cpuset_buffer_lock);
2517 snprintf(cpuset_name, CPUSET_NAME_LEN,
2518 dentry ? (const char *)dentry->d_name.name : "/");
2519 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2520 tsk->mems_allowed);
2521 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2522 tsk->comm, cpuset_name, cpuset_nodelist);
2523 spin_unlock(&cpuset_buffer_lock);
2524}
2525
3e0d98b9
PJ
2526/*
2527 * Collection of memory_pressure is suppressed unless
2528 * this flag is enabled by writing "1" to the special
2529 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2530 */
2531
c5b2aff8 2532int cpuset_memory_pressure_enabled __read_mostly;
3e0d98b9
PJ
2533
2534/**
2535 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2536 *
2537 * Keep a running average of the rate of synchronous (direct)
2538 * page reclaim efforts initiated by tasks in each cpuset.
2539 *
2540 * This represents the rate at which some task in the cpuset
2541 * ran low on memory on all nodes it was allowed to use, and
2542 * had to enter the kernels page reclaim code in an effort to
2543 * create more free memory by tossing clean pages or swapping
2544 * or writing dirty pages.
2545 *
2546 * Display to user space in the per-cpuset read-only file
2547 * "memory_pressure". Value displayed is an integer
2548 * representing the recent rate of entry into the synchronous
2549 * (direct) page reclaim by any task attached to the cpuset.
2550 **/
2551
2552void __cpuset_memory_pressure_bump(void)
2553{
3e0d98b9 2554 task_lock(current);
8793d854 2555 fmeter_markevent(&task_cs(current)->fmeter);
3e0d98b9
PJ
2556 task_unlock(current);
2557}
2558
8793d854 2559#ifdef CONFIG_PROC_PID_CPUSET
1da177e4
LT
2560/*
2561 * proc_cpuset_show()
2562 * - Print tasks cpuset path into seq_file.
2563 * - Used for /proc/<pid>/cpuset.
053199ed
PJ
2564 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2565 * doesn't really matter if tsk->cpuset changes after we read it,
c8d9c90c 2566 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2df167a3 2567 * anyway.
1da177e4 2568 */
029190c5 2569static int proc_cpuset_show(struct seq_file *m, void *unused_v)
1da177e4 2570{
13b41b09 2571 struct pid *pid;
1da177e4
LT
2572 struct task_struct *tsk;
2573 char *buf;
8793d854 2574 struct cgroup_subsys_state *css;
99f89551 2575 int retval;
1da177e4 2576
99f89551 2577 retval = -ENOMEM;
1da177e4
LT
2578 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2579 if (!buf)
99f89551
EB
2580 goto out;
2581
2582 retval = -ESRCH;
13b41b09
EB
2583 pid = m->private;
2584 tsk = get_pid_task(pid, PIDTYPE_PID);
99f89551
EB
2585 if (!tsk)
2586 goto out_free;
1da177e4 2587
99f89551 2588 retval = -EINVAL;
8793d854
PM
2589 cgroup_lock();
2590 css = task_subsys_state(tsk, cpuset_subsys_id);
2591 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
1da177e4 2592 if (retval < 0)
99f89551 2593 goto out_unlock;
1da177e4
LT
2594 seq_puts(m, buf);
2595 seq_putc(m, '\n');
99f89551 2596out_unlock:
8793d854 2597 cgroup_unlock();
99f89551
EB
2598 put_task_struct(tsk);
2599out_free:
1da177e4 2600 kfree(buf);
99f89551 2601out:
1da177e4
LT
2602 return retval;
2603}
2604
2605static int cpuset_open(struct inode *inode, struct file *file)
2606{
13b41b09
EB
2607 struct pid *pid = PROC_I(inode)->pid;
2608 return single_open(file, proc_cpuset_show, pid);
1da177e4
LT
2609}
2610
9a32144e 2611const struct file_operations proc_cpuset_operations = {
1da177e4
LT
2612 .open = cpuset_open,
2613 .read = seq_read,
2614 .llseek = seq_lseek,
2615 .release = single_release,
2616};
8793d854 2617#endif /* CONFIG_PROC_PID_CPUSET */
1da177e4 2618
d01d4827 2619/* Display task mems_allowed in /proc/<pid>/status file. */
df5f8314
EB
2620void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2621{
df5f8314 2622 seq_printf(m, "Mems_allowed:\t");
30e8e136 2623 seq_nodemask(m, &task->mems_allowed);
df5f8314 2624 seq_printf(m, "\n");
39106dcf 2625 seq_printf(m, "Mems_allowed_list:\t");
30e8e136 2626 seq_nodemask_list(m, &task->mems_allowed);
39106dcf 2627 seq_printf(m, "\n");
1da177e4 2628}