]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - kernel/cpuset.c
Input: atmel_mxt_ts - add additional bootloader addresses
[mirror_ubuntu-artful-kernel.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>
e44193d3 62#include <linux/wait.h>
1da177e4 63
202f72d5
PJ
64/*
65 * Tracks how many cpusets are currently defined in system.
66 * When there is only one cpuset (the root cpuset) we can
67 * short circuit some hooks.
68 */
7edc5962 69int number_of_cpusets __read_mostly;
202f72d5 70
3e0d98b9
PJ
71/* See "Frequency meter" comments, below. */
72
73struct fmeter {
74 int cnt; /* unprocessed events count */
75 int val; /* most recent output value */
76 time_t time; /* clock (secs) when val computed */
77 spinlock_t lock; /* guards read or write of above */
78};
79
1da177e4 80struct cpuset {
8793d854
PM
81 struct cgroup_subsys_state css;
82
1da177e4 83 unsigned long flags; /* "unsigned long" so bitops work */
300ed6cb 84 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
1da177e4
LT
85 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
86
33ad801d
LZ
87 /*
88 * This is old Memory Nodes tasks took on.
89 *
90 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
91 * - A new cpuset's old_mems_allowed is initialized when some
92 * task is moved into it.
93 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
94 * cpuset.mems_allowed and have tasks' nodemask updated, and
95 * then old_mems_allowed is updated to mems_allowed.
96 */
97 nodemask_t old_mems_allowed;
98
3e0d98b9 99 struct fmeter fmeter; /* memory_pressure filter */
029190c5 100
452477fa
TH
101 /*
102 * Tasks are being attached to this cpuset. Used to prevent
103 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
104 */
105 int attach_in_progress;
106
029190c5
PJ
107 /* partition number for rebuild_sched_domains() */
108 int pn;
956db3ca 109
1d3504fc
HS
110 /* for custom sched domain */
111 int relax_domain_level;
1da177e4
LT
112};
113
a7c6d554 114static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
8793d854 115{
a7c6d554 116 return css ? container_of(css, struct cpuset, css) : NULL;
8793d854
PM
117}
118
119/* Retrieve the cpuset for a task */
120static inline struct cpuset *task_cs(struct task_struct *task)
121{
073219e9 122 return css_cs(task_css(task, cpuset_cgrp_id));
8793d854 123}
8793d854 124
c9710d80 125static inline struct cpuset *parent_cs(struct cpuset *cs)
c431069f 126{
63876986 127 return css_cs(css_parent(&cs->css));
c431069f
TH
128}
129
b246272e
DR
130#ifdef CONFIG_NUMA
131static inline bool task_has_mempolicy(struct task_struct *task)
132{
133 return task->mempolicy;
134}
135#else
136static inline bool task_has_mempolicy(struct task_struct *task)
137{
138 return false;
139}
140#endif
141
142
1da177e4
LT
143/* bits in struct cpuset flags field */
144typedef enum {
efeb77b2 145 CS_ONLINE,
1da177e4
LT
146 CS_CPU_EXCLUSIVE,
147 CS_MEM_EXCLUSIVE,
78608366 148 CS_MEM_HARDWALL,
45b07ef3 149 CS_MEMORY_MIGRATE,
029190c5 150 CS_SCHED_LOAD_BALANCE,
825a46af
PJ
151 CS_SPREAD_PAGE,
152 CS_SPREAD_SLAB,
1da177e4
LT
153} cpuset_flagbits_t;
154
155/* convenient tests for these bits */
efeb77b2
TH
156static inline bool is_cpuset_online(const struct cpuset *cs)
157{
158 return test_bit(CS_ONLINE, &cs->flags);
159}
160
1da177e4
LT
161static inline int is_cpu_exclusive(const struct cpuset *cs)
162{
7b5b9ef0 163 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
1da177e4
LT
164}
165
166static inline int is_mem_exclusive(const struct cpuset *cs)
167{
7b5b9ef0 168 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
1da177e4
LT
169}
170
78608366
PM
171static inline int is_mem_hardwall(const struct cpuset *cs)
172{
173 return test_bit(CS_MEM_HARDWALL, &cs->flags);
174}
175
029190c5
PJ
176static inline int is_sched_load_balance(const struct cpuset *cs)
177{
178 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
179}
180
45b07ef3
PJ
181static inline int is_memory_migrate(const struct cpuset *cs)
182{
7b5b9ef0 183 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
45b07ef3
PJ
184}
185
825a46af
PJ
186static inline int is_spread_page(const struct cpuset *cs)
187{
188 return test_bit(CS_SPREAD_PAGE, &cs->flags);
189}
190
191static inline int is_spread_slab(const struct cpuset *cs)
192{
193 return test_bit(CS_SPREAD_SLAB, &cs->flags);
194}
195
1da177e4 196static struct cpuset top_cpuset = {
efeb77b2
TH
197 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
198 (1 << CS_MEM_EXCLUSIVE)),
1da177e4
LT
199};
200
ae8086ce
TH
201/**
202 * cpuset_for_each_child - traverse online children of a cpuset
203 * @child_cs: loop cursor pointing to the current child
492eb21b 204 * @pos_css: used for iteration
ae8086ce
TH
205 * @parent_cs: target cpuset to walk children of
206 *
207 * Walk @child_cs through the online children of @parent_cs. Must be used
208 * with RCU read locked.
209 */
492eb21b
TH
210#define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
211 css_for_each_child((pos_css), &(parent_cs)->css) \
212 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
ae8086ce 213
fc560a26
TH
214/**
215 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
216 * @des_cs: loop cursor pointing to the current descendant
492eb21b 217 * @pos_css: used for iteration
fc560a26
TH
218 * @root_cs: target cpuset to walk ancestor of
219 *
220 * Walk @des_cs through the online descendants of @root_cs. Must be used
492eb21b 221 * with RCU read locked. The caller may modify @pos_css by calling
bd8815a6
TH
222 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
223 * iteration and the first node to be visited.
fc560a26 224 */
492eb21b
TH
225#define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
226 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
227 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
fc560a26 228
1da177e4 229/*
5d21cc2d
TH
230 * There are two global mutexes guarding cpuset structures - cpuset_mutex
231 * and callback_mutex. The latter may nest inside the former. We also
232 * require taking task_lock() when dereferencing a task's cpuset pointer.
233 * See "The task_lock() exception", at the end of this comment.
234 *
235 * A task must hold both mutexes to modify cpusets. If a task holds
236 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
237 * is the only task able to also acquire callback_mutex and be able to
238 * modify cpusets. It can perform various checks on the cpuset structure
239 * first, knowing nothing will change. It can also allocate memory while
240 * just holding cpuset_mutex. While it is performing these checks, various
241 * callback routines can briefly acquire callback_mutex to query cpusets.
242 * Once it is ready to make the changes, it takes callback_mutex, blocking
243 * everyone else.
053199ed
PJ
244 *
245 * Calls to the kernel memory allocator can not be made while holding
3d3f26a7 246 * callback_mutex, as that would risk double tripping on callback_mutex
053199ed
PJ
247 * from one of the callbacks into the cpuset code from within
248 * __alloc_pages().
249 *
3d3f26a7 250 * If a task is only holding callback_mutex, then it has read-only
053199ed
PJ
251 * access to cpusets.
252 *
58568d2a
MX
253 * Now, the task_struct fields mems_allowed and mempolicy may be changed
254 * by other task, we use alloc_lock in the task_struct fields to protect
255 * them.
053199ed 256 *
3d3f26a7 257 * The cpuset_common_file_read() handlers only hold callback_mutex across
053199ed
PJ
258 * small pieces of code, such as when reading out possibly multi-word
259 * cpumasks and nodemasks.
260 *
2df167a3
PM
261 * Accessing a task's cpuset should be done in accordance with the
262 * guidelines for accessing subsystem state in kernel/cgroup.c
1da177e4
LT
263 */
264
5d21cc2d 265static DEFINE_MUTEX(cpuset_mutex);
3d3f26a7 266static DEFINE_MUTEX(callback_mutex);
4247bdc6 267
3a5a6d0c
TH
268/*
269 * CPU / memory hotplug is handled asynchronously.
270 */
271static void cpuset_hotplug_workfn(struct work_struct *work);
3a5a6d0c
TH
272static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
273
e44193d3
LZ
274static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
275
cf417141
MK
276/*
277 * This is ugly, but preserves the userspace API for existing cpuset
8793d854 278 * users. If someone tries to mount the "cpuset" filesystem, we
cf417141
MK
279 * silently switch it to mount "cgroup" instead
280 */
f7e83571
AV
281static struct dentry *cpuset_mount(struct file_system_type *fs_type,
282 int flags, const char *unused_dev_name, void *data)
1da177e4 283{
8793d854 284 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
f7e83571 285 struct dentry *ret = ERR_PTR(-ENODEV);
8793d854
PM
286 if (cgroup_fs) {
287 char mountopts[] =
288 "cpuset,noprefix,"
289 "release_agent=/sbin/cpuset_release_agent";
f7e83571
AV
290 ret = cgroup_fs->mount(cgroup_fs, flags,
291 unused_dev_name, mountopts);
8793d854
PM
292 put_filesystem(cgroup_fs);
293 }
294 return ret;
1da177e4
LT
295}
296
297static struct file_system_type cpuset_fs_type = {
298 .name = "cpuset",
f7e83571 299 .mount = cpuset_mount,
1da177e4
LT
300};
301
1da177e4 302/*
300ed6cb 303 * Return in pmask the portion of a cpusets's cpus_allowed that
1da177e4 304 * are online. If none are online, walk up the cpuset hierarchy
40df2deb
LZ
305 * until we find one that does have some online cpus. The top
306 * cpuset always has some cpus online.
1da177e4
LT
307 *
308 * One way or another, we guarantee to return some non-empty subset
5f054e31 309 * of cpu_online_mask.
1da177e4 310 *
3d3f26a7 311 * Call with callback_mutex held.
1da177e4 312 */
c9710d80 313static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
1da177e4 314{
40df2deb 315 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
c431069f 316 cs = parent_cs(cs);
40df2deb 317 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
1da177e4
LT
318}
319
320/*
321 * Return in *pmask the portion of a cpusets's mems_allowed that
0e1e7c7a
CL
322 * are online, with memory. If none are online with memory, walk
323 * up the cpuset hierarchy until we find one that does have some
40df2deb 324 * online mems. The top cpuset always has some mems online.
1da177e4
LT
325 *
326 * One way or another, we guarantee to return some non-empty subset
38d7bee9 327 * of node_states[N_MEMORY].
1da177e4 328 *
3d3f26a7 329 * Call with callback_mutex held.
1da177e4 330 */
c9710d80 331static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
1da177e4 332{
40df2deb 333 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
c431069f 334 cs = parent_cs(cs);
40df2deb 335 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
1da177e4
LT
336}
337
f3b39d47
MX
338/*
339 * update task's spread flag if cpuset's page/slab spread flag is set
340 *
5d21cc2d 341 * Called with callback_mutex/cpuset_mutex held
f3b39d47
MX
342 */
343static void cpuset_update_task_spread_flag(struct cpuset *cs,
344 struct task_struct *tsk)
345{
346 if (is_spread_page(cs))
347 tsk->flags |= PF_SPREAD_PAGE;
348 else
349 tsk->flags &= ~PF_SPREAD_PAGE;
350 if (is_spread_slab(cs))
351 tsk->flags |= PF_SPREAD_SLAB;
352 else
353 tsk->flags &= ~PF_SPREAD_SLAB;
354}
355
1da177e4
LT
356/*
357 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
358 *
359 * One cpuset is a subset of another if all its allowed CPUs and
360 * Memory Nodes are a subset of the other, and its exclusive flags
5d21cc2d 361 * are only set if the other's are set. Call holding cpuset_mutex.
1da177e4
LT
362 */
363
364static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
365{
300ed6cb 366 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
1da177e4
LT
367 nodes_subset(p->mems_allowed, q->mems_allowed) &&
368 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
369 is_mem_exclusive(p) <= is_mem_exclusive(q);
370}
371
645fcc9d
LZ
372/**
373 * alloc_trial_cpuset - allocate a trial cpuset
374 * @cs: the cpuset that the trial cpuset duplicates
375 */
c9710d80 376static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
645fcc9d 377{
300ed6cb
LZ
378 struct cpuset *trial;
379
380 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
381 if (!trial)
382 return NULL;
383
384 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
385 kfree(trial);
386 return NULL;
387 }
388 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
389
390 return trial;
645fcc9d
LZ
391}
392
393/**
394 * free_trial_cpuset - free the trial cpuset
395 * @trial: the trial cpuset to be freed
396 */
397static void free_trial_cpuset(struct cpuset *trial)
398{
300ed6cb 399 free_cpumask_var(trial->cpus_allowed);
645fcc9d
LZ
400 kfree(trial);
401}
402
1da177e4
LT
403/*
404 * validate_change() - Used to validate that any proposed cpuset change
405 * follows the structural rules for cpusets.
406 *
407 * If we replaced the flag and mask values of the current cpuset
408 * (cur) with those values in the trial cpuset (trial), would
409 * our various subset and exclusive rules still be valid? Presumes
5d21cc2d 410 * cpuset_mutex held.
1da177e4
LT
411 *
412 * 'cur' is the address of an actual, in-use cpuset. Operations
413 * such as list traversal that depend on the actual address of the
414 * cpuset in the list must use cur below, not trial.
415 *
416 * 'trial' is the address of bulk structure copy of cur, with
417 * perhaps one or more of the fields cpus_allowed, mems_allowed,
418 * or flags changed to new, trial values.
419 *
420 * Return 0 if valid, -errno if not.
421 */
422
c9710d80 423static int validate_change(struct cpuset *cur, struct cpuset *trial)
1da177e4 424{
492eb21b 425 struct cgroup_subsys_state *css;
1da177e4 426 struct cpuset *c, *par;
ae8086ce
TH
427 int ret;
428
429 rcu_read_lock();
1da177e4
LT
430
431 /* Each of our child cpusets must be a subset of us */
ae8086ce 432 ret = -EBUSY;
492eb21b 433 cpuset_for_each_child(c, css, cur)
ae8086ce
TH
434 if (!is_cpuset_subset(c, trial))
435 goto out;
1da177e4
LT
436
437 /* Remaining checks don't apply to root cpuset */
ae8086ce 438 ret = 0;
69604067 439 if (cur == &top_cpuset)
ae8086ce 440 goto out;
1da177e4 441
c431069f 442 par = parent_cs(cur);
69604067 443
1da177e4 444 /* We must be a subset of our parent cpuset */
ae8086ce 445 ret = -EACCES;
1da177e4 446 if (!is_cpuset_subset(trial, par))
ae8086ce 447 goto out;
1da177e4 448
2df167a3
PM
449 /*
450 * If either I or some sibling (!= me) is exclusive, we can't
451 * overlap
452 */
ae8086ce 453 ret = -EINVAL;
492eb21b 454 cpuset_for_each_child(c, css, par) {
1da177e4
LT
455 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
456 c != cur &&
300ed6cb 457 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
ae8086ce 458 goto out;
1da177e4
LT
459 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
460 c != cur &&
461 nodes_intersects(trial->mems_allowed, c->mems_allowed))
ae8086ce 462 goto out;
1da177e4
LT
463 }
464
452477fa
TH
465 /*
466 * Cpusets with tasks - existing or newly being attached - can't
1c09b195 467 * be changed to have empty cpus_allowed or mems_allowed.
452477fa 468 */
ae8086ce 469 ret = -ENOSPC;
07bc356e 470 if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
1c09b195
LZ
471 if (!cpumask_empty(cur->cpus_allowed) &&
472 cpumask_empty(trial->cpus_allowed))
473 goto out;
474 if (!nodes_empty(cur->mems_allowed) &&
475 nodes_empty(trial->mems_allowed))
476 goto out;
477 }
020958b6 478
ae8086ce
TH
479 ret = 0;
480out:
481 rcu_read_unlock();
482 return ret;
1da177e4
LT
483}
484
db7f47cf 485#ifdef CONFIG_SMP
029190c5 486/*
cf417141 487 * Helper routine for generate_sched_domains().
029190c5
PJ
488 * Do cpusets a, b have overlapping cpus_allowed masks?
489 */
029190c5
PJ
490static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
491{
300ed6cb 492 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
029190c5
PJ
493}
494
1d3504fc
HS
495static void
496update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
497{
1d3504fc
HS
498 if (dattr->relax_domain_level < c->relax_domain_level)
499 dattr->relax_domain_level = c->relax_domain_level;
500 return;
501}
502
fc560a26
TH
503static void update_domain_attr_tree(struct sched_domain_attr *dattr,
504 struct cpuset *root_cs)
f5393693 505{
fc560a26 506 struct cpuset *cp;
492eb21b 507 struct cgroup_subsys_state *pos_css;
f5393693 508
fc560a26 509 rcu_read_lock();
492eb21b 510 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
bd8815a6
TH
511 if (cp == root_cs)
512 continue;
513
fc560a26
TH
514 /* skip the whole subtree if @cp doesn't have any CPU */
515 if (cpumask_empty(cp->cpus_allowed)) {
492eb21b 516 pos_css = css_rightmost_descendant(pos_css);
f5393693 517 continue;
fc560a26 518 }
f5393693
LJ
519
520 if (is_sched_load_balance(cp))
521 update_domain_attr(dattr, cp);
f5393693 522 }
fc560a26 523 rcu_read_unlock();
f5393693
LJ
524}
525
029190c5 526/*
cf417141
MK
527 * generate_sched_domains()
528 *
529 * This function builds a partial partition of the systems CPUs
530 * A 'partial partition' is a set of non-overlapping subsets whose
531 * union is a subset of that set.
0a0fca9d 532 * The output of this function needs to be passed to kernel/sched/core.c
cf417141
MK
533 * partition_sched_domains() routine, which will rebuild the scheduler's
534 * load balancing domains (sched domains) as specified by that partial
535 * partition.
029190c5 536 *
45ce80fb 537 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
029190c5
PJ
538 * for a background explanation of this.
539 *
540 * Does not return errors, on the theory that the callers of this
541 * routine would rather not worry about failures to rebuild sched
542 * domains when operating in the severe memory shortage situations
543 * that could cause allocation failures below.
544 *
5d21cc2d 545 * Must be called with cpuset_mutex held.
029190c5
PJ
546 *
547 * The three key local variables below are:
aeed6824 548 * q - a linked-list queue of cpuset pointers, used to implement a
029190c5
PJ
549 * top-down scan of all cpusets. This scan loads a pointer
550 * to each cpuset marked is_sched_load_balance into the
551 * array 'csa'. For our purposes, rebuilding the schedulers
552 * sched domains, we can ignore !is_sched_load_balance cpusets.
553 * csa - (for CpuSet Array) Array of pointers to all the cpusets
554 * that need to be load balanced, for convenient iterative
555 * access by the subsequent code that finds the best partition,
556 * i.e the set of domains (subsets) of CPUs such that the
557 * cpus_allowed of every cpuset marked is_sched_load_balance
558 * is a subset of one of these domains, while there are as
559 * many such domains as possible, each as small as possible.
560 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
0a0fca9d 561 * the kernel/sched/core.c routine partition_sched_domains() in a
029190c5
PJ
562 * convenient format, that can be easily compared to the prior
563 * value to determine what partition elements (sched domains)
564 * were changed (added or removed.)
565 *
566 * Finding the best partition (set of domains):
567 * The triple nested loops below over i, j, k scan over the
568 * load balanced cpusets (using the array of cpuset pointers in
569 * csa[]) looking for pairs of cpusets that have overlapping
570 * cpus_allowed, but which don't have the same 'pn' partition
571 * number and gives them in the same partition number. It keeps
572 * looping on the 'restart' label until it can no longer find
573 * any such pairs.
574 *
575 * The union of the cpus_allowed masks from the set of
576 * all cpusets having the same 'pn' value then form the one
577 * element of the partition (one sched domain) to be passed to
578 * partition_sched_domains().
579 */
acc3f5d7 580static int generate_sched_domains(cpumask_var_t **domains,
cf417141 581 struct sched_domain_attr **attributes)
029190c5 582{
029190c5
PJ
583 struct cpuset *cp; /* scans q */
584 struct cpuset **csa; /* array of all cpuset ptrs */
585 int csn; /* how many cpuset ptrs in csa so far */
586 int i, j, k; /* indices for partition finding loops */
acc3f5d7 587 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
1d3504fc 588 struct sched_domain_attr *dattr; /* attributes for custom domains */
1583715d 589 int ndoms = 0; /* number of sched domains in result */
6af866af 590 int nslot; /* next empty doms[] struct cpumask slot */
492eb21b 591 struct cgroup_subsys_state *pos_css;
029190c5 592
029190c5 593 doms = NULL;
1d3504fc 594 dattr = NULL;
cf417141 595 csa = NULL;
029190c5
PJ
596
597 /* Special case for the 99% of systems with one, full, sched domain */
598 if (is_sched_load_balance(&top_cpuset)) {
acc3f5d7
RR
599 ndoms = 1;
600 doms = alloc_sched_domains(ndoms);
029190c5 601 if (!doms)
cf417141
MK
602 goto done;
603
1d3504fc
HS
604 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
605 if (dattr) {
606 *dattr = SD_ATTR_INIT;
93a65575 607 update_domain_attr_tree(dattr, &top_cpuset);
1d3504fc 608 }
acc3f5d7 609 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
cf417141 610
cf417141 611 goto done;
029190c5
PJ
612 }
613
029190c5
PJ
614 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
615 if (!csa)
616 goto done;
617 csn = 0;
618
fc560a26 619 rcu_read_lock();
492eb21b 620 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
bd8815a6
TH
621 if (cp == &top_cpuset)
622 continue;
f5393693 623 /*
fc560a26
TH
624 * Continue traversing beyond @cp iff @cp has some CPUs and
625 * isn't load balancing. The former is obvious. The
626 * latter: All child cpusets contain a subset of the
627 * parent's cpus, so just skip them, and then we call
628 * update_domain_attr_tree() to calc relax_domain_level of
629 * the corresponding sched domain.
f5393693 630 */
fc560a26
TH
631 if (!cpumask_empty(cp->cpus_allowed) &&
632 !is_sched_load_balance(cp))
f5393693 633 continue;
489a5393 634
fc560a26
TH
635 if (is_sched_load_balance(cp))
636 csa[csn++] = cp;
637
638 /* skip @cp's subtree */
492eb21b 639 pos_css = css_rightmost_descendant(pos_css);
fc560a26
TH
640 }
641 rcu_read_unlock();
029190c5
PJ
642
643 for (i = 0; i < csn; i++)
644 csa[i]->pn = i;
645 ndoms = csn;
646
647restart:
648 /* Find the best partition (set of sched domains) */
649 for (i = 0; i < csn; i++) {
650 struct cpuset *a = csa[i];
651 int apn = a->pn;
652
653 for (j = 0; j < csn; j++) {
654 struct cpuset *b = csa[j];
655 int bpn = b->pn;
656
657 if (apn != bpn && cpusets_overlap(a, b)) {
658 for (k = 0; k < csn; k++) {
659 struct cpuset *c = csa[k];
660
661 if (c->pn == bpn)
662 c->pn = apn;
663 }
664 ndoms--; /* one less element */
665 goto restart;
666 }
667 }
668 }
669
cf417141
MK
670 /*
671 * Now we know how many domains to create.
672 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
673 */
acc3f5d7 674 doms = alloc_sched_domains(ndoms);
700018e0 675 if (!doms)
cf417141 676 goto done;
cf417141
MK
677
678 /*
679 * The rest of the code, including the scheduler, can deal with
680 * dattr==NULL case. No need to abort if alloc fails.
681 */
1d3504fc 682 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
029190c5
PJ
683
684 for (nslot = 0, i = 0; i < csn; i++) {
685 struct cpuset *a = csa[i];
6af866af 686 struct cpumask *dp;
029190c5
PJ
687 int apn = a->pn;
688
cf417141
MK
689 if (apn < 0) {
690 /* Skip completed partitions */
691 continue;
692 }
693
acc3f5d7 694 dp = doms[nslot];
cf417141
MK
695
696 if (nslot == ndoms) {
697 static int warnings = 10;
698 if (warnings) {
699 printk(KERN_WARNING
700 "rebuild_sched_domains confused:"
701 " nslot %d, ndoms %d, csn %d, i %d,"
702 " apn %d\n",
703 nslot, ndoms, csn, i, apn);
704 warnings--;
029190c5 705 }
cf417141
MK
706 continue;
707 }
029190c5 708
6af866af 709 cpumask_clear(dp);
cf417141
MK
710 if (dattr)
711 *(dattr + nslot) = SD_ATTR_INIT;
712 for (j = i; j < csn; j++) {
713 struct cpuset *b = csa[j];
714
715 if (apn == b->pn) {
300ed6cb 716 cpumask_or(dp, dp, b->cpus_allowed);
cf417141
MK
717 if (dattr)
718 update_domain_attr_tree(dattr + nslot, b);
719
720 /* Done with this partition */
721 b->pn = -1;
029190c5 722 }
029190c5 723 }
cf417141 724 nslot++;
029190c5
PJ
725 }
726 BUG_ON(nslot != ndoms);
727
cf417141
MK
728done:
729 kfree(csa);
730
700018e0
LZ
731 /*
732 * Fallback to the default domain if kmalloc() failed.
733 * See comments in partition_sched_domains().
734 */
735 if (doms == NULL)
736 ndoms = 1;
737
cf417141
MK
738 *domains = doms;
739 *attributes = dattr;
740 return ndoms;
741}
742
743/*
744 * Rebuild scheduler domains.
745 *
699140ba
TH
746 * If the flag 'sched_load_balance' of any cpuset with non-empty
747 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
748 * which has that flag enabled, or if any cpuset with a non-empty
749 * 'cpus' is removed, then call this routine to rebuild the
750 * scheduler's dynamic sched domains.
cf417141 751 *
5d21cc2d 752 * Call with cpuset_mutex held. Takes get_online_cpus().
cf417141 753 */
699140ba 754static void rebuild_sched_domains_locked(void)
cf417141
MK
755{
756 struct sched_domain_attr *attr;
acc3f5d7 757 cpumask_var_t *doms;
cf417141
MK
758 int ndoms;
759
5d21cc2d 760 lockdep_assert_held(&cpuset_mutex);
86ef5c9a 761 get_online_cpus();
cf417141 762
5b16c2a4
LZ
763 /*
764 * We have raced with CPU hotplug. Don't do anything to avoid
765 * passing doms with offlined cpu to partition_sched_domains().
766 * Anyways, hotplug work item will rebuild sched domains.
767 */
768 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
769 goto out;
770
cf417141 771 /* Generate domain masks and attrs */
cf417141 772 ndoms = generate_sched_domains(&doms, &attr);
cf417141
MK
773
774 /* Have scheduler rebuild the domains */
775 partition_sched_domains(ndoms, doms, attr);
5b16c2a4 776out:
86ef5c9a 777 put_online_cpus();
cf417141 778}
db7f47cf 779#else /* !CONFIG_SMP */
699140ba 780static void rebuild_sched_domains_locked(void)
db7f47cf
PM
781{
782}
db7f47cf 783#endif /* CONFIG_SMP */
029190c5 784
cf417141
MK
785void rebuild_sched_domains(void)
786{
5d21cc2d 787 mutex_lock(&cpuset_mutex);
699140ba 788 rebuild_sched_domains_locked();
5d21cc2d 789 mutex_unlock(&cpuset_mutex);
029190c5
PJ
790}
791
070b57fc
LZ
792/*
793 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
794 * @cs: the cpuset in interest
58f4790b 795 *
070b57fc
LZ
796 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
797 * with non-empty cpus. We use effective cpumask whenever:
798 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
799 * if the cpuset they reside in has no cpus)
800 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
801 *
802 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
803 * exception. See comments there.
804 */
805static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
806{
807 while (cpumask_empty(cs->cpus_allowed))
808 cs = parent_cs(cs);
809 return cs;
810}
811
812/*
813 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
814 * @cs: the cpuset in interest
815 *
816 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
817 * with non-empty memss. We use effective nodemask whenever:
818 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
819 * if the cpuset they reside in has no mems)
820 * - we want to retrieve task_cs(tsk)'s mems_allowed.
821 *
822 * Called with cpuset_mutex held.
053199ed 823 */
070b57fc 824static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
58f4790b 825{
070b57fc
LZ
826 while (nodes_empty(cs->mems_allowed))
827 cs = parent_cs(cs);
828 return cs;
58f4790b 829}
053199ed 830
0b2f630a
MX
831/**
832 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
833 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
0b2f630a 834 *
d66393e5
TH
835 * Iterate through each task of @cs updating its cpus_allowed to the
836 * effective cpuset's. As this function is called with cpuset_mutex held,
837 * cpuset membership stays stable.
0b2f630a 838 */
d66393e5 839static void update_tasks_cpumask(struct cpuset *cs)
0b2f630a 840{
d66393e5
TH
841 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
842 struct css_task_iter it;
843 struct task_struct *task;
844
845 css_task_iter_start(&cs->css, &it);
846 while ((task = css_task_iter_next(&it)))
847 set_cpus_allowed_ptr(task, cpus_cs->cpus_allowed);
848 css_task_iter_end(&it);
0b2f630a
MX
849}
850
5c5cc623
LZ
851/*
852 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
853 * @root_cs: the root cpuset of the hierarchy
854 * @update_root: update root cpuset or not?
5c5cc623
LZ
855 *
856 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
857 * which take on cpumask of @root_cs.
858 *
859 * Called with cpuset_mutex held
860 */
d66393e5 861static void update_tasks_cpumask_hier(struct cpuset *root_cs, bool update_root)
5c5cc623
LZ
862{
863 struct cpuset *cp;
492eb21b 864 struct cgroup_subsys_state *pos_css;
5c5cc623
LZ
865
866 rcu_read_lock();
492eb21b 867 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
bd8815a6
TH
868 if (cp == root_cs) {
869 if (!update_root)
870 continue;
871 } else {
872 /* skip the whole subtree if @cp have some CPU */
873 if (!cpumask_empty(cp->cpus_allowed)) {
874 pos_css = css_rightmost_descendant(pos_css);
875 continue;
876 }
5c5cc623
LZ
877 }
878 if (!css_tryget(&cp->css))
879 continue;
880 rcu_read_unlock();
881
d66393e5 882 update_tasks_cpumask(cp);
5c5cc623
LZ
883
884 rcu_read_lock();
885 css_put(&cp->css);
886 }
887 rcu_read_unlock();
888}
889
58f4790b
CW
890/**
891 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
892 * @cs: the cpuset to consider
893 * @buf: buffer of cpu numbers written to this cpuset
894 */
645fcc9d
LZ
895static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
896 const char *buf)
1da177e4 897{
58f4790b
CW
898 int retval;
899 int is_load_balanced;
1da177e4 900
5f054e31 901 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
4c4d50f7
PJ
902 if (cs == &top_cpuset)
903 return -EACCES;
904
6f7f02e7 905 /*
c8d9c90c 906 * An empty cpus_allowed is ok only if the cpuset has no tasks.
020958b6
PJ
907 * Since cpulist_parse() fails on an empty mask, we special case
908 * that parsing. The validate_change() call ensures that cpusets
909 * with tasks have cpus.
6f7f02e7 910 */
020958b6 911 if (!*buf) {
300ed6cb 912 cpumask_clear(trialcs->cpus_allowed);
6f7f02e7 913 } else {
300ed6cb 914 retval = cpulist_parse(buf, trialcs->cpus_allowed);
6f7f02e7
DR
915 if (retval < 0)
916 return retval;
37340746 917
6ad4c188 918 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
37340746 919 return -EINVAL;
6f7f02e7 920 }
029190c5 921
8707d8b8 922 /* Nothing to do if the cpus didn't change */
300ed6cb 923 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
8707d8b8 924 return 0;
58f4790b 925
a73456f3
LZ
926 retval = validate_change(cs, trialcs);
927 if (retval < 0)
928 return retval;
929
645fcc9d 930 is_load_balanced = is_sched_load_balance(trialcs);
029190c5 931
3d3f26a7 932 mutex_lock(&callback_mutex);
300ed6cb 933 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
3d3f26a7 934 mutex_unlock(&callback_mutex);
029190c5 935
d66393e5 936 update_tasks_cpumask_hier(cs, true);
58f4790b 937
8707d8b8 938 if (is_load_balanced)
699140ba 939 rebuild_sched_domains_locked();
85d7b949 940 return 0;
1da177e4
LT
941}
942
e4e364e8
PJ
943/*
944 * cpuset_migrate_mm
945 *
946 * Migrate memory region from one set of nodes to another.
947 *
948 * Temporarilly set tasks mems_allowed to target nodes of migration,
949 * so that the migration code can allocate pages on these nodes.
950 *
e4e364e8
PJ
951 * While the mm_struct we are migrating is typically from some
952 * other task, the task_struct mems_allowed that we are hacking
953 * is for our current task, which must allocate new pages for that
954 * migrating memory region.
e4e364e8
PJ
955 */
956
957static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
958 const nodemask_t *to)
959{
960 struct task_struct *tsk = current;
070b57fc 961 struct cpuset *mems_cs;
e4e364e8 962
e4e364e8 963 tsk->mems_allowed = *to;
e4e364e8
PJ
964
965 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
966
47295830 967 rcu_read_lock();
070b57fc
LZ
968 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
969 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
47295830 970 rcu_read_unlock();
e4e364e8
PJ
971}
972
3b6766fe 973/*
58568d2a
MX
974 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
975 * @tsk: the task to change
976 * @newmems: new nodes that the task will be set
977 *
978 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
979 * we structure updates as setting all new allowed nodes, then clearing newly
980 * disallowed ones.
58568d2a
MX
981 */
982static void cpuset_change_task_nodemask(struct task_struct *tsk,
983 nodemask_t *newmems)
984{
b246272e 985 bool need_loop;
89e8a244 986
c0ff7453
MX
987 /*
988 * Allow tasks that have access to memory reserves because they have
989 * been OOM killed to get memory anywhere.
990 */
991 if (unlikely(test_thread_flag(TIF_MEMDIE)))
992 return;
993 if (current->flags & PF_EXITING) /* Let dying task have memory */
994 return;
995
996 task_lock(tsk);
b246272e
DR
997 /*
998 * Determine if a loop is necessary if another thread is doing
d26914d1 999 * read_mems_allowed_begin(). If at least one node remains unchanged and
b246272e
DR
1000 * tsk does not have a mempolicy, then an empty nodemask will not be
1001 * possible when mems_allowed is larger than a word.
1002 */
1003 need_loop = task_has_mempolicy(tsk) ||
1004 !nodes_intersects(*newmems, tsk->mems_allowed);
c0ff7453 1005
0fc0287c
PZ
1006 if (need_loop) {
1007 local_irq_disable();
cc9a6c87 1008 write_seqcount_begin(&tsk->mems_allowed_seq);
0fc0287c 1009 }
c0ff7453 1010
cc9a6c87
MG
1011 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1012 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
c0ff7453
MX
1013
1014 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
58568d2a 1015 tsk->mems_allowed = *newmems;
cc9a6c87 1016
0fc0287c 1017 if (need_loop) {
cc9a6c87 1018 write_seqcount_end(&tsk->mems_allowed_seq);
0fc0287c
PZ
1019 local_irq_enable();
1020 }
cc9a6c87 1021
c0ff7453 1022 task_unlock(tsk);
58568d2a
MX
1023}
1024
8793d854
PM
1025static void *cpuset_being_rebound;
1026
0b2f630a
MX
1027/**
1028 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1029 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
0b2f630a 1030 *
d66393e5
TH
1031 * Iterate through each task of @cs updating its mems_allowed to the
1032 * effective cpuset's. As this function is called with cpuset_mutex held,
1033 * cpuset membership stays stable.
0b2f630a 1034 */
d66393e5 1035static void update_tasks_nodemask(struct cpuset *cs)
1da177e4 1036{
33ad801d 1037 static nodemask_t newmems; /* protected by cpuset_mutex */
070b57fc 1038 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
d66393e5
TH
1039 struct css_task_iter it;
1040 struct task_struct *task;
59dac16f 1041
846a16bf 1042 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
4225399a 1043
070b57fc 1044 guarantee_online_mems(mems_cs, &newmems);
33ad801d 1045
4225399a 1046 /*
3b6766fe
LZ
1047 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1048 * take while holding tasklist_lock. Forks can happen - the
1049 * mpol_dup() cpuset_being_rebound check will catch such forks,
1050 * and rebind their vma mempolicies too. Because we still hold
5d21cc2d 1051 * the global cpuset_mutex, we know that no other rebind effort
3b6766fe 1052 * will be contending for the global variable cpuset_being_rebound.
4225399a 1053 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
04c19fa6 1054 * is idempotent. Also migrate pages in each mm to new nodes.
4225399a 1055 */
d66393e5
TH
1056 css_task_iter_start(&cs->css, &it);
1057 while ((task = css_task_iter_next(&it))) {
1058 struct mm_struct *mm;
1059 bool migrate;
1060
1061 cpuset_change_task_nodemask(task, &newmems);
1062
1063 mm = get_task_mm(task);
1064 if (!mm)
1065 continue;
1066
1067 migrate = is_memory_migrate(cs);
1068
1069 mpol_rebind_mm(mm, &cs->mems_allowed);
1070 if (migrate)
1071 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1072 mmput(mm);
1073 }
1074 css_task_iter_end(&it);
4225399a 1075
33ad801d
LZ
1076 /*
1077 * All the tasks' nodemasks have been updated, update
1078 * cs->old_mems_allowed.
1079 */
1080 cs->old_mems_allowed = newmems;
1081
2df167a3 1082 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
8793d854 1083 cpuset_being_rebound = NULL;
1da177e4
LT
1084}
1085
5c5cc623
LZ
1086/*
1087 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1088 * @cs: the root cpuset of the hierarchy
1089 * @update_root: update the root cpuset or not?
5c5cc623
LZ
1090 *
1091 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1092 * which take on nodemask of @root_cs.
1093 *
1094 * Called with cpuset_mutex held
1095 */
d66393e5 1096static void update_tasks_nodemask_hier(struct cpuset *root_cs, bool update_root)
5c5cc623
LZ
1097{
1098 struct cpuset *cp;
492eb21b 1099 struct cgroup_subsys_state *pos_css;
5c5cc623
LZ
1100
1101 rcu_read_lock();
492eb21b 1102 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
bd8815a6
TH
1103 if (cp == root_cs) {
1104 if (!update_root)
1105 continue;
1106 } else {
1107 /* skip the whole subtree if @cp have some CPU */
1108 if (!nodes_empty(cp->mems_allowed)) {
1109 pos_css = css_rightmost_descendant(pos_css);
1110 continue;
1111 }
5c5cc623
LZ
1112 }
1113 if (!css_tryget(&cp->css))
1114 continue;
1115 rcu_read_unlock();
1116
d66393e5 1117 update_tasks_nodemask(cp);
5c5cc623
LZ
1118
1119 rcu_read_lock();
1120 css_put(&cp->css);
1121 }
1122 rcu_read_unlock();
1123}
1124
0b2f630a
MX
1125/*
1126 * Handle user request to change the 'mems' memory placement
1127 * of a cpuset. Needs to validate the request, update the
58568d2a
MX
1128 * cpusets mems_allowed, and for each task in the cpuset,
1129 * update mems_allowed and rebind task's mempolicy and any vma
1130 * mempolicies and if the cpuset is marked 'memory_migrate',
1131 * migrate the tasks pages to the new memory.
0b2f630a 1132 *
5d21cc2d 1133 * Call with cpuset_mutex held. May take callback_mutex during call.
0b2f630a
MX
1134 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1135 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1136 * their mempolicies to the cpusets new mems_allowed.
1137 */
645fcc9d
LZ
1138static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1139 const char *buf)
0b2f630a 1140{
0b2f630a
MX
1141 int retval;
1142
1143 /*
38d7bee9 1144 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
0b2f630a
MX
1145 * it's read-only
1146 */
53feb297
MX
1147 if (cs == &top_cpuset) {
1148 retval = -EACCES;
1149 goto done;
1150 }
0b2f630a 1151
0b2f630a
MX
1152 /*
1153 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1154 * Since nodelist_parse() fails on an empty mask, we special case
1155 * that parsing. The validate_change() call ensures that cpusets
1156 * with tasks have memory.
1157 */
1158 if (!*buf) {
645fcc9d 1159 nodes_clear(trialcs->mems_allowed);
0b2f630a 1160 } else {
645fcc9d 1161 retval = nodelist_parse(buf, trialcs->mems_allowed);
0b2f630a
MX
1162 if (retval < 0)
1163 goto done;
1164
645fcc9d 1165 if (!nodes_subset(trialcs->mems_allowed,
38d7bee9 1166 node_states[N_MEMORY])) {
53feb297
MX
1167 retval = -EINVAL;
1168 goto done;
1169 }
0b2f630a 1170 }
33ad801d
LZ
1171
1172 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
0b2f630a
MX
1173 retval = 0; /* Too easy - nothing to do */
1174 goto done;
1175 }
645fcc9d 1176 retval = validate_change(cs, trialcs);
0b2f630a
MX
1177 if (retval < 0)
1178 goto done;
1179
1180 mutex_lock(&callback_mutex);
645fcc9d 1181 cs->mems_allowed = trialcs->mems_allowed;
0b2f630a
MX
1182 mutex_unlock(&callback_mutex);
1183
d66393e5 1184 update_tasks_nodemask_hier(cs, true);
0b2f630a
MX
1185done:
1186 return retval;
1187}
1188
8793d854
PM
1189int current_cpuset_is_being_rebound(void)
1190{
1191 return task_cs(current) == cpuset_being_rebound;
1192}
1193
5be7a479 1194static int update_relax_domain_level(struct cpuset *cs, s64 val)
1d3504fc 1195{
db7f47cf 1196#ifdef CONFIG_SMP
60495e77 1197 if (val < -1 || val >= sched_domain_level_max)
30e0e178 1198 return -EINVAL;
db7f47cf 1199#endif
1d3504fc
HS
1200
1201 if (val != cs->relax_domain_level) {
1202 cs->relax_domain_level = val;
300ed6cb
LZ
1203 if (!cpumask_empty(cs->cpus_allowed) &&
1204 is_sched_load_balance(cs))
699140ba 1205 rebuild_sched_domains_locked();
1d3504fc
HS
1206 }
1207
1208 return 0;
1209}
1210
72ec7029 1211/**
950592f7
MX
1212 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1213 * @cs: the cpuset in which each task's spread flags needs to be changed
950592f7 1214 *
d66393e5
TH
1215 * Iterate through each task of @cs updating its spread flags. As this
1216 * function is called with cpuset_mutex held, cpuset membership stays
1217 * stable.
950592f7 1218 */
d66393e5 1219static void update_tasks_flags(struct cpuset *cs)
950592f7 1220{
d66393e5
TH
1221 struct css_task_iter it;
1222 struct task_struct *task;
1223
1224 css_task_iter_start(&cs->css, &it);
1225 while ((task = css_task_iter_next(&it)))
1226 cpuset_update_task_spread_flag(cs, task);
1227 css_task_iter_end(&it);
950592f7
MX
1228}
1229
1da177e4
LT
1230/*
1231 * update_flag - read a 0 or a 1 in a file and update associated flag
78608366
PM
1232 * bit: the bit to update (see cpuset_flagbits_t)
1233 * cs: the cpuset to update
1234 * turning_on: whether the flag is being set or cleared
053199ed 1235 *
5d21cc2d 1236 * Call with cpuset_mutex held.
1da177e4
LT
1237 */
1238
700fe1ab
PM
1239static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1240 int turning_on)
1da177e4 1241{
645fcc9d 1242 struct cpuset *trialcs;
40b6a762 1243 int balance_flag_changed;
950592f7 1244 int spread_flag_changed;
950592f7 1245 int err;
1da177e4 1246
645fcc9d
LZ
1247 trialcs = alloc_trial_cpuset(cs);
1248 if (!trialcs)
1249 return -ENOMEM;
1250
1da177e4 1251 if (turning_on)
645fcc9d 1252 set_bit(bit, &trialcs->flags);
1da177e4 1253 else
645fcc9d 1254 clear_bit(bit, &trialcs->flags);
1da177e4 1255
645fcc9d 1256 err = validate_change(cs, trialcs);
85d7b949 1257 if (err < 0)
645fcc9d 1258 goto out;
029190c5 1259
029190c5 1260 balance_flag_changed = (is_sched_load_balance(cs) !=
645fcc9d 1261 is_sched_load_balance(trialcs));
029190c5 1262
950592f7
MX
1263 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1264 || (is_spread_page(cs) != is_spread_page(trialcs)));
1265
3d3f26a7 1266 mutex_lock(&callback_mutex);
645fcc9d 1267 cs->flags = trialcs->flags;
3d3f26a7 1268 mutex_unlock(&callback_mutex);
85d7b949 1269
300ed6cb 1270 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
699140ba 1271 rebuild_sched_domains_locked();
029190c5 1272
950592f7 1273 if (spread_flag_changed)
d66393e5 1274 update_tasks_flags(cs);
645fcc9d
LZ
1275out:
1276 free_trial_cpuset(trialcs);
1277 return err;
1da177e4
LT
1278}
1279
3e0d98b9 1280/*
80f7228b 1281 * Frequency meter - How fast is some event occurring?
3e0d98b9
PJ
1282 *
1283 * These routines manage a digitally filtered, constant time based,
1284 * event frequency meter. There are four routines:
1285 * fmeter_init() - initialize a frequency meter.
1286 * fmeter_markevent() - called each time the event happens.
1287 * fmeter_getrate() - returns the recent rate of such events.
1288 * fmeter_update() - internal routine used to update fmeter.
1289 *
1290 * A common data structure is passed to each of these routines,
1291 * which is used to keep track of the state required to manage the
1292 * frequency meter and its digital filter.
1293 *
1294 * The filter works on the number of events marked per unit time.
1295 * The filter is single-pole low-pass recursive (IIR). The time unit
1296 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1297 * simulate 3 decimal digits of precision (multiplied by 1000).
1298 *
1299 * With an FM_COEF of 933, and a time base of 1 second, the filter
1300 * has a half-life of 10 seconds, meaning that if the events quit
1301 * happening, then the rate returned from the fmeter_getrate()
1302 * will be cut in half each 10 seconds, until it converges to zero.
1303 *
1304 * It is not worth doing a real infinitely recursive filter. If more
1305 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1306 * just compute FM_MAXTICKS ticks worth, by which point the level
1307 * will be stable.
1308 *
1309 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1310 * arithmetic overflow in the fmeter_update() routine.
1311 *
1312 * Given the simple 32 bit integer arithmetic used, this meter works
1313 * best for reporting rates between one per millisecond (msec) and
1314 * one per 32 (approx) seconds. At constant rates faster than one
1315 * per msec it maxes out at values just under 1,000,000. At constant
1316 * rates between one per msec, and one per second it will stabilize
1317 * to a value N*1000, where N is the rate of events per second.
1318 * At constant rates between one per second and one per 32 seconds,
1319 * it will be choppy, moving up on the seconds that have an event,
1320 * and then decaying until the next event. At rates slower than
1321 * about one in 32 seconds, it decays all the way back to zero between
1322 * each event.
1323 */
1324
1325#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1326#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1327#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1328#define FM_SCALE 1000 /* faux fixed point scale */
1329
1330/* Initialize a frequency meter */
1331static void fmeter_init(struct fmeter *fmp)
1332{
1333 fmp->cnt = 0;
1334 fmp->val = 0;
1335 fmp->time = 0;
1336 spin_lock_init(&fmp->lock);
1337}
1338
1339/* Internal meter update - process cnt events and update value */
1340static void fmeter_update(struct fmeter *fmp)
1341{
1342 time_t now = get_seconds();
1343 time_t ticks = now - fmp->time;
1344
1345 if (ticks == 0)
1346 return;
1347
1348 ticks = min(FM_MAXTICKS, ticks);
1349 while (ticks-- > 0)
1350 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1351 fmp->time = now;
1352
1353 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1354 fmp->cnt = 0;
1355}
1356
1357/* Process any previous ticks, then bump cnt by one (times scale). */
1358static void fmeter_markevent(struct fmeter *fmp)
1359{
1360 spin_lock(&fmp->lock);
1361 fmeter_update(fmp);
1362 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1363 spin_unlock(&fmp->lock);
1364}
1365
1366/* Process any previous ticks, then return current value. */
1367static int fmeter_getrate(struct fmeter *fmp)
1368{
1369 int val;
1370
1371 spin_lock(&fmp->lock);
1372 fmeter_update(fmp);
1373 val = fmp->val;
1374 spin_unlock(&fmp->lock);
1375 return val;
1376}
1377
57fce0a6
TH
1378static struct cpuset *cpuset_attach_old_cs;
1379
5d21cc2d 1380/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
eb95419b
TH
1381static int cpuset_can_attach(struct cgroup_subsys_state *css,
1382 struct cgroup_taskset *tset)
f780bdb7 1383{
eb95419b 1384 struct cpuset *cs = css_cs(css);
bb9d97b6
TH
1385 struct task_struct *task;
1386 int ret;
1da177e4 1387
57fce0a6
TH
1388 /* used later by cpuset_attach() */
1389 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1390
5d21cc2d
TH
1391 mutex_lock(&cpuset_mutex);
1392
88fa523b
LZ
1393 /*
1394 * We allow to move tasks into an empty cpuset if sane_behavior
1395 * flag is set.
1396 */
5d21cc2d 1397 ret = -ENOSPC;
eb95419b 1398 if (!cgroup_sane_behavior(css->cgroup) &&
88fa523b 1399 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
5d21cc2d 1400 goto out_unlock;
9985b0ba 1401
924f0d9a 1402 cgroup_taskset_for_each(task, tset) {
bb9d97b6 1403 /*
14a40ffc
TH
1404 * Kthreads which disallow setaffinity shouldn't be moved
1405 * to a new cpuset; we don't want to change their cpu
1406 * affinity and isolating such threads by their set of
1407 * allowed nodes is unnecessary. Thus, cpusets are not
1408 * applicable for such threads. This prevents checking for
1409 * success of set_cpus_allowed_ptr() on all attached tasks
1410 * before cpus_allowed may be changed.
bb9d97b6 1411 */
5d21cc2d 1412 ret = -EINVAL;
14a40ffc 1413 if (task->flags & PF_NO_SETAFFINITY)
5d21cc2d
TH
1414 goto out_unlock;
1415 ret = security_task_setscheduler(task);
1416 if (ret)
1417 goto out_unlock;
bb9d97b6 1418 }
f780bdb7 1419
452477fa
TH
1420 /*
1421 * Mark attach is in progress. This makes validate_change() fail
1422 * changes which zero cpus/mems_allowed.
1423 */
1424 cs->attach_in_progress++;
5d21cc2d
TH
1425 ret = 0;
1426out_unlock:
1427 mutex_unlock(&cpuset_mutex);
1428 return ret;
8793d854 1429}
f780bdb7 1430
eb95419b 1431static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
452477fa
TH
1432 struct cgroup_taskset *tset)
1433{
5d21cc2d 1434 mutex_lock(&cpuset_mutex);
eb95419b 1435 css_cs(css)->attach_in_progress--;
5d21cc2d 1436 mutex_unlock(&cpuset_mutex);
8793d854 1437}
1da177e4 1438
4e4c9a14 1439/*
5d21cc2d 1440 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
4e4c9a14
TH
1441 * but we can't allocate it dynamically there. Define it global and
1442 * allocate from cpuset_init().
1443 */
1444static cpumask_var_t cpus_attach;
1445
eb95419b
TH
1446static void cpuset_attach(struct cgroup_subsys_state *css,
1447 struct cgroup_taskset *tset)
8793d854 1448{
67bd2c59 1449 /* static buf protected by cpuset_mutex */
4e4c9a14 1450 static nodemask_t cpuset_attach_nodemask_to;
8793d854 1451 struct mm_struct *mm;
bb9d97b6
TH
1452 struct task_struct *task;
1453 struct task_struct *leader = cgroup_taskset_first(tset);
eb95419b 1454 struct cpuset *cs = css_cs(css);
57fce0a6 1455 struct cpuset *oldcs = cpuset_attach_old_cs;
070b57fc
LZ
1456 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1457 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
22fb52dd 1458
5d21cc2d
TH
1459 mutex_lock(&cpuset_mutex);
1460
4e4c9a14
TH
1461 /* prepare for attach */
1462 if (cs == &top_cpuset)
1463 cpumask_copy(cpus_attach, cpu_possible_mask);
1464 else
070b57fc 1465 guarantee_online_cpus(cpus_cs, cpus_attach);
4e4c9a14 1466
070b57fc 1467 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
4e4c9a14 1468
924f0d9a 1469 cgroup_taskset_for_each(task, tset) {
bb9d97b6
TH
1470 /*
1471 * can_attach beforehand should guarantee that this doesn't
1472 * fail. TODO: have a better way to handle failure here
1473 */
1474 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1475
1476 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1477 cpuset_update_task_spread_flag(cs, task);
1478 }
22fb52dd 1479
f780bdb7
BB
1480 /*
1481 * Change mm, possibly for multiple threads in a threadgroup. This is
1482 * expensive and may sleep.
1483 */
f780bdb7 1484 cpuset_attach_nodemask_to = cs->mems_allowed;
bb9d97b6 1485 mm = get_task_mm(leader);
4225399a 1486 if (mm) {
070b57fc
LZ
1487 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1488
f780bdb7 1489 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
f047cecf
LZ
1490
1491 /*
1492 * old_mems_allowed is the same with mems_allowed here, except
1493 * if this task is being moved automatically due to hotplug.
1494 * In that case @mems_allowed has been updated and is empty,
1495 * so @old_mems_allowed is the right nodesets that we migrate
1496 * mm from.
1497 */
1498 if (is_memory_migrate(cs)) {
1499 cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
f780bdb7 1500 &cpuset_attach_nodemask_to);
f047cecf 1501 }
4225399a
PJ
1502 mmput(mm);
1503 }
452477fa 1504
33ad801d 1505 cs->old_mems_allowed = cpuset_attach_nodemask_to;
02bb5863 1506
452477fa 1507 cs->attach_in_progress--;
e44193d3
LZ
1508 if (!cs->attach_in_progress)
1509 wake_up(&cpuset_attach_wq);
5d21cc2d
TH
1510
1511 mutex_unlock(&cpuset_mutex);
1da177e4
LT
1512}
1513
1514/* The various types of files and directories in a cpuset file system */
1515
1516typedef enum {
45b07ef3 1517 FILE_MEMORY_MIGRATE,
1da177e4
LT
1518 FILE_CPULIST,
1519 FILE_MEMLIST,
1520 FILE_CPU_EXCLUSIVE,
1521 FILE_MEM_EXCLUSIVE,
78608366 1522 FILE_MEM_HARDWALL,
029190c5 1523 FILE_SCHED_LOAD_BALANCE,
1d3504fc 1524 FILE_SCHED_RELAX_DOMAIN_LEVEL,
3e0d98b9
PJ
1525 FILE_MEMORY_PRESSURE_ENABLED,
1526 FILE_MEMORY_PRESSURE,
825a46af
PJ
1527 FILE_SPREAD_PAGE,
1528 FILE_SPREAD_SLAB,
1da177e4
LT
1529} cpuset_filetype_t;
1530
182446d0
TH
1531static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1532 u64 val)
700fe1ab 1533{
182446d0 1534 struct cpuset *cs = css_cs(css);
700fe1ab 1535 cpuset_filetype_t type = cft->private;
a903f086 1536 int retval = 0;
700fe1ab 1537
5d21cc2d 1538 mutex_lock(&cpuset_mutex);
a903f086
LZ
1539 if (!is_cpuset_online(cs)) {
1540 retval = -ENODEV;
5d21cc2d 1541 goto out_unlock;
a903f086 1542 }
700fe1ab
PM
1543
1544 switch (type) {
1da177e4 1545 case FILE_CPU_EXCLUSIVE:
700fe1ab 1546 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1da177e4
LT
1547 break;
1548 case FILE_MEM_EXCLUSIVE:
700fe1ab 1549 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1da177e4 1550 break;
78608366
PM
1551 case FILE_MEM_HARDWALL:
1552 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1553 break;
029190c5 1554 case FILE_SCHED_LOAD_BALANCE:
700fe1ab 1555 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1d3504fc 1556 break;
45b07ef3 1557 case FILE_MEMORY_MIGRATE:
700fe1ab 1558 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
45b07ef3 1559 break;
3e0d98b9 1560 case FILE_MEMORY_PRESSURE_ENABLED:
700fe1ab 1561 cpuset_memory_pressure_enabled = !!val;
3e0d98b9
PJ
1562 break;
1563 case FILE_MEMORY_PRESSURE:
1564 retval = -EACCES;
1565 break;
825a46af 1566 case FILE_SPREAD_PAGE:
700fe1ab 1567 retval = update_flag(CS_SPREAD_PAGE, cs, val);
825a46af
PJ
1568 break;
1569 case FILE_SPREAD_SLAB:
700fe1ab 1570 retval = update_flag(CS_SPREAD_SLAB, cs, val);
825a46af 1571 break;
1da177e4
LT
1572 default:
1573 retval = -EINVAL;
700fe1ab 1574 break;
1da177e4 1575 }
5d21cc2d
TH
1576out_unlock:
1577 mutex_unlock(&cpuset_mutex);
1da177e4
LT
1578 return retval;
1579}
1580
182446d0
TH
1581static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1582 s64 val)
5be7a479 1583{
182446d0 1584 struct cpuset *cs = css_cs(css);
5be7a479 1585 cpuset_filetype_t type = cft->private;
5d21cc2d 1586 int retval = -ENODEV;
5be7a479 1587
5d21cc2d
TH
1588 mutex_lock(&cpuset_mutex);
1589 if (!is_cpuset_online(cs))
1590 goto out_unlock;
e3712395 1591
5be7a479
PM
1592 switch (type) {
1593 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1594 retval = update_relax_domain_level(cs, val);
1595 break;
1596 default:
1597 retval = -EINVAL;
1598 break;
1599 }
5d21cc2d
TH
1600out_unlock:
1601 mutex_unlock(&cpuset_mutex);
5be7a479
PM
1602 return retval;
1603}
1604
e3712395
PM
1605/*
1606 * Common handling for a write to a "cpus" or "mems" file.
1607 */
182446d0 1608static int cpuset_write_resmask(struct cgroup_subsys_state *css,
4d3bb511 1609 struct cftype *cft, char *buf)
e3712395 1610{
182446d0 1611 struct cpuset *cs = css_cs(css);
645fcc9d 1612 struct cpuset *trialcs;
5d21cc2d 1613 int retval = -ENODEV;
e3712395 1614
3a5a6d0c
TH
1615 /*
1616 * CPU or memory hotunplug may leave @cs w/o any execution
1617 * resources, in which case the hotplug code asynchronously updates
1618 * configuration and transfers all tasks to the nearest ancestor
1619 * which can execute.
1620 *
1621 * As writes to "cpus" or "mems" may restore @cs's execution
1622 * resources, wait for the previously scheduled operations before
1623 * proceeding, so that we don't end up keep removing tasks added
1624 * after execution capability is restored.
1625 */
1626 flush_work(&cpuset_hotplug_work);
1627
5d21cc2d
TH
1628 mutex_lock(&cpuset_mutex);
1629 if (!is_cpuset_online(cs))
1630 goto out_unlock;
e3712395 1631
645fcc9d 1632 trialcs = alloc_trial_cpuset(cs);
b75f38d6
LZ
1633 if (!trialcs) {
1634 retval = -ENOMEM;
5d21cc2d 1635 goto out_unlock;
b75f38d6 1636 }
645fcc9d 1637
e3712395
PM
1638 switch (cft->private) {
1639 case FILE_CPULIST:
645fcc9d 1640 retval = update_cpumask(cs, trialcs, buf);
e3712395
PM
1641 break;
1642 case FILE_MEMLIST:
645fcc9d 1643 retval = update_nodemask(cs, trialcs, buf);
e3712395
PM
1644 break;
1645 default:
1646 retval = -EINVAL;
1647 break;
1648 }
645fcc9d
LZ
1649
1650 free_trial_cpuset(trialcs);
5d21cc2d
TH
1651out_unlock:
1652 mutex_unlock(&cpuset_mutex);
e3712395
PM
1653 return retval;
1654}
1655
1da177e4
LT
1656/*
1657 * These ascii lists should be read in a single call, by using a user
1658 * buffer large enough to hold the entire map. If read in smaller
1659 * chunks, there is no guarantee of atomicity. Since the display format
1660 * used, list of ranges of sequential numbers, is variable length,
1661 * and since these maps can change value dynamically, one could read
1662 * gibberish by doing partial reads while a list was changing.
1da177e4 1663 */
2da8ca82 1664static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1da177e4 1665{
2da8ca82
TH
1666 struct cpuset *cs = css_cs(seq_css(sf));
1667 cpuset_filetype_t type = seq_cft(sf)->private;
51ffe411
TH
1668 ssize_t count;
1669 char *buf, *s;
1670 int ret = 0;
1da177e4 1671
51ffe411
TH
1672 count = seq_get_buf(sf, &buf);
1673 s = buf;
1da177e4 1674
51ffe411 1675 mutex_lock(&callback_mutex);
1da177e4
LT
1676
1677 switch (type) {
1678 case FILE_CPULIST:
51ffe411 1679 s += cpulist_scnprintf(s, count, cs->cpus_allowed);
1da177e4
LT
1680 break;
1681 case FILE_MEMLIST:
51ffe411 1682 s += nodelist_scnprintf(s, count, cs->mems_allowed);
1da177e4 1683 break;
1da177e4 1684 default:
51ffe411
TH
1685 ret = -EINVAL;
1686 goto out_unlock;
1da177e4 1687 }
1da177e4 1688
51ffe411
TH
1689 if (s < buf + count - 1) {
1690 *s++ = '\n';
1691 seq_commit(sf, s - buf);
1692 } else {
1693 seq_commit(sf, -1);
1694 }
1695out_unlock:
1696 mutex_unlock(&callback_mutex);
1697 return ret;
1da177e4
LT
1698}
1699
182446d0 1700static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
700fe1ab 1701{
182446d0 1702 struct cpuset *cs = css_cs(css);
700fe1ab
PM
1703 cpuset_filetype_t type = cft->private;
1704 switch (type) {
1705 case FILE_CPU_EXCLUSIVE:
1706 return is_cpu_exclusive(cs);
1707 case FILE_MEM_EXCLUSIVE:
1708 return is_mem_exclusive(cs);
78608366
PM
1709 case FILE_MEM_HARDWALL:
1710 return is_mem_hardwall(cs);
700fe1ab
PM
1711 case FILE_SCHED_LOAD_BALANCE:
1712 return is_sched_load_balance(cs);
1713 case FILE_MEMORY_MIGRATE:
1714 return is_memory_migrate(cs);
1715 case FILE_MEMORY_PRESSURE_ENABLED:
1716 return cpuset_memory_pressure_enabled;
1717 case FILE_MEMORY_PRESSURE:
1718 return fmeter_getrate(&cs->fmeter);
1719 case FILE_SPREAD_PAGE:
1720 return is_spread_page(cs);
1721 case FILE_SPREAD_SLAB:
1722 return is_spread_slab(cs);
1723 default:
1724 BUG();
1725 }
cf417141
MK
1726
1727 /* Unreachable but makes gcc happy */
1728 return 0;
700fe1ab 1729}
1da177e4 1730
182446d0 1731static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
5be7a479 1732{
182446d0 1733 struct cpuset *cs = css_cs(css);
5be7a479
PM
1734 cpuset_filetype_t type = cft->private;
1735 switch (type) {
1736 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1737 return cs->relax_domain_level;
1738 default:
1739 BUG();
1740 }
cf417141
MK
1741
1742 /* Unrechable but makes gcc happy */
1743 return 0;
5be7a479
PM
1744}
1745
1da177e4
LT
1746
1747/*
1748 * for the common functions, 'private' gives the type of file
1749 */
1750
addf2c73
PM
1751static struct cftype files[] = {
1752 {
1753 .name = "cpus",
2da8ca82 1754 .seq_show = cpuset_common_seq_show,
e3712395
PM
1755 .write_string = cpuset_write_resmask,
1756 .max_write_len = (100U + 6 * NR_CPUS),
addf2c73
PM
1757 .private = FILE_CPULIST,
1758 },
1759
1760 {
1761 .name = "mems",
2da8ca82 1762 .seq_show = cpuset_common_seq_show,
e3712395
PM
1763 .write_string = cpuset_write_resmask,
1764 .max_write_len = (100U + 6 * MAX_NUMNODES),
addf2c73
PM
1765 .private = FILE_MEMLIST,
1766 },
1767
1768 {
1769 .name = "cpu_exclusive",
1770 .read_u64 = cpuset_read_u64,
1771 .write_u64 = cpuset_write_u64,
1772 .private = FILE_CPU_EXCLUSIVE,
1773 },
1774
1775 {
1776 .name = "mem_exclusive",
1777 .read_u64 = cpuset_read_u64,
1778 .write_u64 = cpuset_write_u64,
1779 .private = FILE_MEM_EXCLUSIVE,
1780 },
1781
78608366
PM
1782 {
1783 .name = "mem_hardwall",
1784 .read_u64 = cpuset_read_u64,
1785 .write_u64 = cpuset_write_u64,
1786 .private = FILE_MEM_HARDWALL,
1787 },
1788
addf2c73
PM
1789 {
1790 .name = "sched_load_balance",
1791 .read_u64 = cpuset_read_u64,
1792 .write_u64 = cpuset_write_u64,
1793 .private = FILE_SCHED_LOAD_BALANCE,
1794 },
1795
1796 {
1797 .name = "sched_relax_domain_level",
5be7a479
PM
1798 .read_s64 = cpuset_read_s64,
1799 .write_s64 = cpuset_write_s64,
addf2c73
PM
1800 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1801 },
1802
1803 {
1804 .name = "memory_migrate",
1805 .read_u64 = cpuset_read_u64,
1806 .write_u64 = cpuset_write_u64,
1807 .private = FILE_MEMORY_MIGRATE,
1808 },
1809
1810 {
1811 .name = "memory_pressure",
1812 .read_u64 = cpuset_read_u64,
1813 .write_u64 = cpuset_write_u64,
1814 .private = FILE_MEMORY_PRESSURE,
099fca32 1815 .mode = S_IRUGO,
addf2c73
PM
1816 },
1817
1818 {
1819 .name = "memory_spread_page",
1820 .read_u64 = cpuset_read_u64,
1821 .write_u64 = cpuset_write_u64,
1822 .private = FILE_SPREAD_PAGE,
1823 },
1824
1825 {
1826 .name = "memory_spread_slab",
1827 .read_u64 = cpuset_read_u64,
1828 .write_u64 = cpuset_write_u64,
1829 .private = FILE_SPREAD_SLAB,
1830 },
3e0d98b9 1831
4baf6e33
TH
1832 {
1833 .name = "memory_pressure_enabled",
1834 .flags = CFTYPE_ONLY_ON_ROOT,
1835 .read_u64 = cpuset_read_u64,
1836 .write_u64 = cpuset_write_u64,
1837 .private = FILE_MEMORY_PRESSURE_ENABLED,
1838 },
1da177e4 1839
4baf6e33
TH
1840 { } /* terminate */
1841};
1da177e4
LT
1842
1843/*
92fb9748 1844 * cpuset_css_alloc - allocate a cpuset css
c9e5fe66 1845 * cgrp: control group that the new cpuset will be part of
1da177e4
LT
1846 */
1847
eb95419b
TH
1848static struct cgroup_subsys_state *
1849cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1da177e4 1850{
c8f699bb 1851 struct cpuset *cs;
1da177e4 1852
eb95419b 1853 if (!parent_css)
8793d854 1854 return &top_cpuset.css;
033fa1c5 1855
c8f699bb 1856 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1da177e4 1857 if (!cs)
8793d854 1858 return ERR_PTR(-ENOMEM);
300ed6cb
LZ
1859 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1860 kfree(cs);
1861 return ERR_PTR(-ENOMEM);
1862 }
1da177e4 1863
029190c5 1864 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
300ed6cb 1865 cpumask_clear(cs->cpus_allowed);
f9a86fcb 1866 nodes_clear(cs->mems_allowed);
3e0d98b9 1867 fmeter_init(&cs->fmeter);
1d3504fc 1868 cs->relax_domain_level = -1;
1da177e4 1869
c8f699bb
TH
1870 return &cs->css;
1871}
1872
eb95419b 1873static int cpuset_css_online(struct cgroup_subsys_state *css)
c8f699bb 1874{
eb95419b 1875 struct cpuset *cs = css_cs(css);
c431069f 1876 struct cpuset *parent = parent_cs(cs);
ae8086ce 1877 struct cpuset *tmp_cs;
492eb21b 1878 struct cgroup_subsys_state *pos_css;
c8f699bb
TH
1879
1880 if (!parent)
1881 return 0;
1882
5d21cc2d
TH
1883 mutex_lock(&cpuset_mutex);
1884
efeb77b2 1885 set_bit(CS_ONLINE, &cs->flags);
c8f699bb
TH
1886 if (is_spread_page(parent))
1887 set_bit(CS_SPREAD_PAGE, &cs->flags);
1888 if (is_spread_slab(parent))
1889 set_bit(CS_SPREAD_SLAB, &cs->flags);
1da177e4 1890
202f72d5 1891 number_of_cpusets++;
033fa1c5 1892
eb95419b 1893 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
5d21cc2d 1894 goto out_unlock;
033fa1c5
TH
1895
1896 /*
1897 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1898 * set. This flag handling is implemented in cgroup core for
1899 * histrical reasons - the flag may be specified during mount.
1900 *
1901 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1902 * refuse to clone the configuration - thereby refusing the task to
1903 * be entered, and as a result refusing the sys_unshare() or
1904 * clone() which initiated it. If this becomes a problem for some
1905 * users who wish to allow that scenario, then this could be
1906 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1907 * (and likewise for mems) to the new cgroup.
1908 */
ae8086ce 1909 rcu_read_lock();
492eb21b 1910 cpuset_for_each_child(tmp_cs, pos_css, parent) {
ae8086ce
TH
1911 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1912 rcu_read_unlock();
5d21cc2d 1913 goto out_unlock;
ae8086ce 1914 }
033fa1c5 1915 }
ae8086ce 1916 rcu_read_unlock();
033fa1c5
TH
1917
1918 mutex_lock(&callback_mutex);
1919 cs->mems_allowed = parent->mems_allowed;
1920 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1921 mutex_unlock(&callback_mutex);
5d21cc2d
TH
1922out_unlock:
1923 mutex_unlock(&cpuset_mutex);
c8f699bb
TH
1924 return 0;
1925}
1926
0b9e6965
ZH
1927/*
1928 * If the cpuset being removed has its flag 'sched_load_balance'
1929 * enabled, then simulate turning sched_load_balance off, which
1930 * will call rebuild_sched_domains_locked().
1931 */
1932
eb95419b 1933static void cpuset_css_offline(struct cgroup_subsys_state *css)
c8f699bb 1934{
eb95419b 1935 struct cpuset *cs = css_cs(css);
c8f699bb 1936
5d21cc2d 1937 mutex_lock(&cpuset_mutex);
c8f699bb
TH
1938
1939 if (is_sched_load_balance(cs))
1940 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1941
1942 number_of_cpusets--;
efeb77b2 1943 clear_bit(CS_ONLINE, &cs->flags);
c8f699bb 1944
5d21cc2d 1945 mutex_unlock(&cpuset_mutex);
1da177e4
LT
1946}
1947
eb95419b 1948static void cpuset_css_free(struct cgroup_subsys_state *css)
1da177e4 1949{
eb95419b 1950 struct cpuset *cs = css_cs(css);
1da177e4 1951
300ed6cb 1952 free_cpumask_var(cs->cpus_allowed);
8793d854 1953 kfree(cs);
1da177e4
LT
1954}
1955
073219e9 1956struct cgroup_subsys cpuset_cgrp_subsys = {
92fb9748 1957 .css_alloc = cpuset_css_alloc,
c8f699bb
TH
1958 .css_online = cpuset_css_online,
1959 .css_offline = cpuset_css_offline,
92fb9748 1960 .css_free = cpuset_css_free,
8793d854 1961 .can_attach = cpuset_can_attach,
452477fa 1962 .cancel_attach = cpuset_cancel_attach,
8793d854 1963 .attach = cpuset_attach,
4baf6e33 1964 .base_cftypes = files,
8793d854
PM
1965 .early_init = 1,
1966};
1967
1da177e4
LT
1968/**
1969 * cpuset_init - initialize cpusets at system boot
1970 *
1971 * Description: Initialize top_cpuset and the cpuset internal file system,
1972 **/
1973
1974int __init cpuset_init(void)
1975{
8793d854 1976 int err = 0;
1da177e4 1977
58568d2a
MX
1978 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1979 BUG();
1980
300ed6cb 1981 cpumask_setall(top_cpuset.cpus_allowed);
f9a86fcb 1982 nodes_setall(top_cpuset.mems_allowed);
1da177e4 1983
3e0d98b9 1984 fmeter_init(&top_cpuset.fmeter);
029190c5 1985 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1d3504fc 1986 top_cpuset.relax_domain_level = -1;
1da177e4 1987
1da177e4
LT
1988 err = register_filesystem(&cpuset_fs_type);
1989 if (err < 0)
8793d854
PM
1990 return err;
1991
2341d1b6
LZ
1992 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1993 BUG();
1994
202f72d5 1995 number_of_cpusets = 1;
8793d854 1996 return 0;
1da177e4
LT
1997}
1998
b1aac8bb 1999/*
cf417141 2000 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
b1aac8bb
PJ
2001 * or memory nodes, we need to walk over the cpuset hierarchy,
2002 * removing that CPU or node from all cpusets. If this removes the
956db3ca
CW
2003 * last CPU or node from a cpuset, then move the tasks in the empty
2004 * cpuset to its next-highest non-empty parent.
b1aac8bb 2005 */
956db3ca
CW
2006static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2007{
2008 struct cpuset *parent;
2009
956db3ca
CW
2010 /*
2011 * Find its next-highest non-empty parent, (top cpuset
2012 * has online cpus, so can't be empty).
2013 */
c431069f 2014 parent = parent_cs(cs);
300ed6cb 2015 while (cpumask_empty(parent->cpus_allowed) ||
b4501295 2016 nodes_empty(parent->mems_allowed))
c431069f 2017 parent = parent_cs(parent);
956db3ca 2018
8cc99345 2019 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
e61734c5
TH
2020 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset ");
2021 pr_cont_cgroup_name(cs->css.cgroup);
2022 pr_cont("\n");
8cc99345 2023 }
956db3ca
CW
2024}
2025
deb7aa30 2026/**
388afd85 2027 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
deb7aa30 2028 * @cs: cpuset in interest
956db3ca 2029 *
deb7aa30
TH
2030 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2031 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2032 * all its tasks are moved to the nearest ancestor with both resources.
80d1fa64 2033 */
388afd85 2034static void cpuset_hotplug_update_tasks(struct cpuset *cs)
80d1fa64 2035{
deb7aa30 2036 static cpumask_t off_cpus;
33ad801d 2037 static nodemask_t off_mems;
5d21cc2d 2038 bool is_empty;
5c5cc623 2039 bool sane = cgroup_sane_behavior(cs->css.cgroup);
80d1fa64 2040
e44193d3
LZ
2041retry:
2042 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
80d1fa64 2043
5d21cc2d 2044 mutex_lock(&cpuset_mutex);
7ddf96b0 2045
e44193d3
LZ
2046 /*
2047 * We have raced with task attaching. We wait until attaching
2048 * is finished, so we won't attach a task to an empty cpuset.
2049 */
2050 if (cs->attach_in_progress) {
2051 mutex_unlock(&cpuset_mutex);
2052 goto retry;
2053 }
2054
deb7aa30
TH
2055 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2056 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
80d1fa64 2057
5c5cc623
LZ
2058 mutex_lock(&callback_mutex);
2059 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2060 mutex_unlock(&callback_mutex);
2061
2062 /*
2063 * If sane_behavior flag is set, we need to update tasks' cpumask
f047cecf
LZ
2064 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2065 * call update_tasks_cpumask() if the cpuset becomes empty, as
2066 * the tasks in it will be migrated to an ancestor.
5c5cc623
LZ
2067 */
2068 if ((sane && cpumask_empty(cs->cpus_allowed)) ||
f047cecf 2069 (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
d66393e5 2070 update_tasks_cpumask(cs);
80d1fa64 2071
5c5cc623
LZ
2072 mutex_lock(&callback_mutex);
2073 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2074 mutex_unlock(&callback_mutex);
2075
2076 /*
2077 * If sane_behavior flag is set, we need to update tasks' nodemask
f047cecf
LZ
2078 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2079 * call update_tasks_nodemask() if the cpuset becomes empty, as
2080 * the tasks in it will be migratd to an ancestor.
5c5cc623
LZ
2081 */
2082 if ((sane && nodes_empty(cs->mems_allowed)) ||
f047cecf 2083 (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
d66393e5 2084 update_tasks_nodemask(cs);
deb7aa30 2085
5d21cc2d
TH
2086 is_empty = cpumask_empty(cs->cpus_allowed) ||
2087 nodes_empty(cs->mems_allowed);
8d033948 2088
5d21cc2d
TH
2089 mutex_unlock(&cpuset_mutex);
2090
2091 /*
5c5cc623
LZ
2092 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2093 *
2094 * Otherwise move tasks to the nearest ancestor with execution
2095 * resources. This is full cgroup operation which will
5d21cc2d
TH
2096 * also call back into cpuset. Should be done outside any lock.
2097 */
5c5cc623 2098 if (!sane && is_empty)
5d21cc2d 2099 remove_tasks_in_empty_cpuset(cs);
b1aac8bb
PJ
2100}
2101
deb7aa30 2102/**
3a5a6d0c 2103 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
956db3ca 2104 *
deb7aa30
TH
2105 * This function is called after either CPU or memory configuration has
2106 * changed and updates cpuset accordingly. The top_cpuset is always
2107 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2108 * order to make cpusets transparent (of no affect) on systems that are
2109 * actively using CPU hotplug but making no active use of cpusets.
956db3ca 2110 *
deb7aa30 2111 * Non-root cpusets are only affected by offlining. If any CPUs or memory
388afd85
LZ
2112 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2113 * all descendants.
956db3ca 2114 *
deb7aa30
TH
2115 * Note that CPU offlining during suspend is ignored. We don't modify
2116 * cpusets across suspend/resume cycles at all.
956db3ca 2117 */
3a5a6d0c 2118static void cpuset_hotplug_workfn(struct work_struct *work)
b1aac8bb 2119{
5c5cc623
LZ
2120 static cpumask_t new_cpus;
2121 static nodemask_t new_mems;
deb7aa30 2122 bool cpus_updated, mems_updated;
b1aac8bb 2123
5d21cc2d 2124 mutex_lock(&cpuset_mutex);
956db3ca 2125
deb7aa30
TH
2126 /* fetch the available cpus/mems and find out which changed how */
2127 cpumask_copy(&new_cpus, cpu_active_mask);
2128 new_mems = node_states[N_MEMORY];
7ddf96b0 2129
deb7aa30 2130 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
deb7aa30 2131 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
7ddf96b0 2132
deb7aa30
TH
2133 /* synchronize cpus_allowed to cpu_active_mask */
2134 if (cpus_updated) {
2135 mutex_lock(&callback_mutex);
2136 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2137 mutex_unlock(&callback_mutex);
2138 /* we don't mess with cpumasks of tasks in top_cpuset */
2139 }
b4501295 2140
deb7aa30
TH
2141 /* synchronize mems_allowed to N_MEMORY */
2142 if (mems_updated) {
deb7aa30
TH
2143 mutex_lock(&callback_mutex);
2144 top_cpuset.mems_allowed = new_mems;
2145 mutex_unlock(&callback_mutex);
d66393e5 2146 update_tasks_nodemask(&top_cpuset);
deb7aa30 2147 }
b4501295 2148
388afd85
LZ
2149 mutex_unlock(&cpuset_mutex);
2150
5c5cc623
LZ
2151 /* if cpus or mems changed, we need to propagate to descendants */
2152 if (cpus_updated || mems_updated) {
deb7aa30 2153 struct cpuset *cs;
492eb21b 2154 struct cgroup_subsys_state *pos_css;
f9b4fb8d 2155
fc560a26 2156 rcu_read_lock();
492eb21b 2157 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
bd8815a6 2158 if (cs == &top_cpuset || !css_tryget(&cs->css))
388afd85
LZ
2159 continue;
2160 rcu_read_unlock();
7ddf96b0 2161
388afd85 2162 cpuset_hotplug_update_tasks(cs);
b4501295 2163
388afd85
LZ
2164 rcu_read_lock();
2165 css_put(&cs->css);
2166 }
2167 rcu_read_unlock();
2168 }
8d033948 2169
deb7aa30 2170 /* rebuild sched domains if cpus_allowed has changed */
e0e80a02
LZ
2171 if (cpus_updated)
2172 rebuild_sched_domains();
b1aac8bb
PJ
2173}
2174
7ddf96b0 2175void cpuset_update_active_cpus(bool cpu_online)
4c4d50f7 2176{
3a5a6d0c
TH
2177 /*
2178 * We're inside cpu hotplug critical region which usually nests
2179 * inside cgroup synchronization. Bounce actual hotplug processing
2180 * to a work item to avoid reverse locking order.
2181 *
2182 * We still need to do partition_sched_domains() synchronously;
2183 * otherwise, the scheduler will get confused and put tasks to the
2184 * dead CPU. Fall back to the default single domain.
2185 * cpuset_hotplug_workfn() will rebuild it as necessary.
2186 */
2187 partition_sched_domains(1, NULL, NULL);
2188 schedule_work(&cpuset_hotplug_work);
4c4d50f7 2189}
4c4d50f7 2190
38837fc7 2191/*
38d7bee9
LJ
2192 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2193 * Call this routine anytime after node_states[N_MEMORY] changes.
a1cd2b13 2194 * See cpuset_update_active_cpus() for CPU hotplug handling.
38837fc7 2195 */
f481891f
MX
2196static int cpuset_track_online_nodes(struct notifier_block *self,
2197 unsigned long action, void *arg)
38837fc7 2198{
3a5a6d0c 2199 schedule_work(&cpuset_hotplug_work);
f481891f 2200 return NOTIFY_OK;
38837fc7 2201}
d8f10cb3
AM
2202
2203static struct notifier_block cpuset_track_online_nodes_nb = {
2204 .notifier_call = cpuset_track_online_nodes,
2205 .priority = 10, /* ??! */
2206};
38837fc7 2207
1da177e4
LT
2208/**
2209 * cpuset_init_smp - initialize cpus_allowed
2210 *
2211 * Description: Finish top cpuset after cpu, node maps are initialized
d8f10cb3 2212 */
1da177e4
LT
2213void __init cpuset_init_smp(void)
2214{
6ad4c188 2215 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
38d7bee9 2216 top_cpuset.mems_allowed = node_states[N_MEMORY];
33ad801d 2217 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
4c4d50f7 2218
d8f10cb3 2219 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
1da177e4
LT
2220}
2221
2222/**
1da177e4
LT
2223 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2224 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
6af866af 2225 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
1da177e4 2226 *
300ed6cb 2227 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
1da177e4 2228 * attached to the specified @tsk. Guaranteed to return some non-empty
5f054e31 2229 * subset of cpu_online_mask, even if this means going outside the
1da177e4
LT
2230 * tasks cpuset.
2231 **/
2232
6af866af 2233void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
1da177e4 2234{
070b57fc
LZ
2235 struct cpuset *cpus_cs;
2236
3d3f26a7 2237 mutex_lock(&callback_mutex);
b8dadcb5 2238 rcu_read_lock();
070b57fc
LZ
2239 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2240 guarantee_online_cpus(cpus_cs, pmask);
b8dadcb5 2241 rcu_read_unlock();
897f0b3c 2242 mutex_unlock(&callback_mutex);
1da177e4
LT
2243}
2244
2baab4e9 2245void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
9084bb82 2246{
c9710d80 2247 struct cpuset *cpus_cs;
9084bb82
ON
2248
2249 rcu_read_lock();
070b57fc
LZ
2250 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2251 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
9084bb82
ON
2252 rcu_read_unlock();
2253
2254 /*
2255 * We own tsk->cpus_allowed, nobody can change it under us.
2256 *
2257 * But we used cs && cs->cpus_allowed lockless and thus can
2258 * race with cgroup_attach_task() or update_cpumask() and get
2259 * the wrong tsk->cpus_allowed. However, both cases imply the
2260 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2261 * which takes task_rq_lock().
2262 *
2263 * If we are called after it dropped the lock we must see all
2264 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2265 * set any mask even if it is not right from task_cs() pov,
2266 * the pending set_cpus_allowed_ptr() will fix things.
2baab4e9
PZ
2267 *
2268 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2269 * if required.
9084bb82 2270 */
9084bb82
ON
2271}
2272
1da177e4
LT
2273void cpuset_init_current_mems_allowed(void)
2274{
f9a86fcb 2275 nodes_setall(current->mems_allowed);
1da177e4
LT
2276}
2277
909d75a3
PJ
2278/**
2279 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2280 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2281 *
2282 * Description: Returns the nodemask_t mems_allowed of the cpuset
2283 * attached to the specified @tsk. Guaranteed to return some non-empty
38d7bee9 2284 * subset of node_states[N_MEMORY], even if this means going outside the
909d75a3
PJ
2285 * tasks cpuset.
2286 **/
2287
2288nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2289{
070b57fc 2290 struct cpuset *mems_cs;
909d75a3
PJ
2291 nodemask_t mask;
2292
3d3f26a7 2293 mutex_lock(&callback_mutex);
b8dadcb5 2294 rcu_read_lock();
070b57fc
LZ
2295 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2296 guarantee_online_mems(mems_cs, &mask);
b8dadcb5 2297 rcu_read_unlock();
3d3f26a7 2298 mutex_unlock(&callback_mutex);
909d75a3
PJ
2299
2300 return mask;
2301}
2302
d9fd8a6d 2303/**
19770b32
MG
2304 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2305 * @nodemask: the nodemask to be checked
d9fd8a6d 2306 *
19770b32 2307 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
1da177e4 2308 */
19770b32 2309int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
1da177e4 2310{
19770b32 2311 return nodes_intersects(*nodemask, current->mems_allowed);
1da177e4
LT
2312}
2313
9bf2229f 2314/*
78608366
PM
2315 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2316 * mem_hardwall ancestor to the specified cpuset. Call holding
2317 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2318 * (an unusual configuration), then returns the root cpuset.
9bf2229f 2319 */
c9710d80 2320static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
9bf2229f 2321{
c431069f
TH
2322 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2323 cs = parent_cs(cs);
9bf2229f
PJ
2324 return cs;
2325}
2326
d9fd8a6d 2327/**
a1bc5a4e
DR
2328 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2329 * @node: is this an allowed node?
02a0e53d 2330 * @gfp_mask: memory allocation flags
d9fd8a6d 2331 *
a1bc5a4e
DR
2332 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2333 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2334 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2335 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2336 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2337 * flag, yes.
9bf2229f
PJ
2338 * Otherwise, no.
2339 *
a1bc5a4e
DR
2340 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2341 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2342 * might sleep, and might allow a node from an enclosing cpuset.
02a0e53d 2343 *
a1bc5a4e
DR
2344 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2345 * cpusets, and never sleeps.
02a0e53d
PJ
2346 *
2347 * The __GFP_THISNODE placement logic is really handled elsewhere,
2348 * by forcibly using a zonelist starting at a specified node, and by
2349 * (in get_page_from_freelist()) refusing to consider the zones for
2350 * any node on the zonelist except the first. By the time any such
2351 * calls get to this routine, we should just shut up and say 'yes'.
2352 *
9bf2229f 2353 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
c596d9f3
DR
2354 * and do not allow allocations outside the current tasks cpuset
2355 * unless the task has been OOM killed as is marked TIF_MEMDIE.
9bf2229f 2356 * GFP_KERNEL allocations are not so marked, so can escape to the
78608366 2357 * nearest enclosing hardwalled ancestor cpuset.
9bf2229f 2358 *
02a0e53d
PJ
2359 * Scanning up parent cpusets requires callback_mutex. The
2360 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2361 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2362 * current tasks mems_allowed came up empty on the first pass over
2363 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2364 * cpuset are short of memory, might require taking the callback_mutex
2365 * mutex.
9bf2229f 2366 *
36be57ff 2367 * The first call here from mm/page_alloc:get_page_from_freelist()
02a0e53d
PJ
2368 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2369 * so no allocation on a node outside the cpuset is allowed (unless
2370 * in interrupt, of course).
36be57ff
PJ
2371 *
2372 * The second pass through get_page_from_freelist() doesn't even call
2373 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2374 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2375 * in alloc_flags. That logic and the checks below have the combined
2376 * affect that:
9bf2229f
PJ
2377 * in_interrupt - any node ok (current task context irrelevant)
2378 * GFP_ATOMIC - any node ok
c596d9f3 2379 * TIF_MEMDIE - any node ok
78608366 2380 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
9bf2229f 2381 * GFP_USER - only nodes in current tasks mems allowed ok.
36be57ff
PJ
2382 *
2383 * Rule:
a1bc5a4e 2384 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
36be57ff
PJ
2385 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2386 * the code that might scan up ancestor cpusets and sleep.
02a0e53d 2387 */
a1bc5a4e 2388int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
1da177e4 2389{
c9710d80 2390 struct cpuset *cs; /* current cpuset ancestors */
29afd49b 2391 int allowed; /* is allocation in zone z allowed? */
9bf2229f 2392
9b819d20 2393 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
9bf2229f 2394 return 1;
92d1dbd2 2395 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
9bf2229f
PJ
2396 if (node_isset(node, current->mems_allowed))
2397 return 1;
c596d9f3
DR
2398 /*
2399 * Allow tasks that have access to memory reserves because they have
2400 * been OOM killed to get memory anywhere.
2401 */
2402 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2403 return 1;
9bf2229f
PJ
2404 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2405 return 0;
2406
5563e770
BP
2407 if (current->flags & PF_EXITING) /* Let dying task have memory */
2408 return 1;
2409
9bf2229f 2410 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3d3f26a7 2411 mutex_lock(&callback_mutex);
053199ed 2412
b8dadcb5 2413 rcu_read_lock();
78608366 2414 cs = nearest_hardwall_ancestor(task_cs(current));
99afb0fd 2415 allowed = node_isset(node, cs->mems_allowed);
b8dadcb5 2416 rcu_read_unlock();
053199ed 2417
3d3f26a7 2418 mutex_unlock(&callback_mutex);
9bf2229f 2419 return allowed;
1da177e4
LT
2420}
2421
02a0e53d 2422/*
a1bc5a4e
DR
2423 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2424 * @node: is this an allowed node?
02a0e53d
PJ
2425 * @gfp_mask: memory allocation flags
2426 *
a1bc5a4e
DR
2427 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2428 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2429 * yes. If the task has been OOM killed and has access to memory reserves as
2430 * specified by the TIF_MEMDIE flag, yes.
2431 * Otherwise, no.
02a0e53d
PJ
2432 *
2433 * The __GFP_THISNODE placement logic is really handled elsewhere,
2434 * by forcibly using a zonelist starting at a specified node, and by
2435 * (in get_page_from_freelist()) refusing to consider the zones for
2436 * any node on the zonelist except the first. By the time any such
2437 * calls get to this routine, we should just shut up and say 'yes'.
2438 *
a1bc5a4e
DR
2439 * Unlike the cpuset_node_allowed_softwall() variant, above,
2440 * this variant requires that the node be in the current task's
02a0e53d
PJ
2441 * mems_allowed or that we're in interrupt. It does not scan up the
2442 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2443 * It never sleeps.
2444 */
a1bc5a4e 2445int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
02a0e53d 2446{
02a0e53d
PJ
2447 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2448 return 1;
02a0e53d
PJ
2449 if (node_isset(node, current->mems_allowed))
2450 return 1;
dedf8b79
DW
2451 /*
2452 * Allow tasks that have access to memory reserves because they have
2453 * been OOM killed to get memory anywhere.
2454 */
2455 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2456 return 1;
02a0e53d
PJ
2457 return 0;
2458}
2459
825a46af 2460/**
6adef3eb
JS
2461 * cpuset_mem_spread_node() - On which node to begin search for a file page
2462 * cpuset_slab_spread_node() - On which node to begin search for a slab page
825a46af
PJ
2463 *
2464 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2465 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2466 * and if the memory allocation used cpuset_mem_spread_node()
2467 * to determine on which node to start looking, as it will for
2468 * certain page cache or slab cache pages such as used for file
2469 * system buffers and inode caches, then instead of starting on the
2470 * local node to look for a free page, rather spread the starting
2471 * node around the tasks mems_allowed nodes.
2472 *
2473 * We don't have to worry about the returned node being offline
2474 * because "it can't happen", and even if it did, it would be ok.
2475 *
2476 * The routines calling guarantee_online_mems() are careful to
2477 * only set nodes in task->mems_allowed that are online. So it
2478 * should not be possible for the following code to return an
2479 * offline node. But if it did, that would be ok, as this routine
2480 * is not returning the node where the allocation must be, only
2481 * the node where the search should start. The zonelist passed to
2482 * __alloc_pages() will include all nodes. If the slab allocator
2483 * is passed an offline node, it will fall back to the local node.
2484 * See kmem_cache_alloc_node().
2485 */
2486
6adef3eb 2487static int cpuset_spread_node(int *rotor)
825a46af
PJ
2488{
2489 int node;
2490
6adef3eb 2491 node = next_node(*rotor, current->mems_allowed);
825a46af
PJ
2492 if (node == MAX_NUMNODES)
2493 node = first_node(current->mems_allowed);
6adef3eb 2494 *rotor = node;
825a46af
PJ
2495 return node;
2496}
6adef3eb
JS
2497
2498int cpuset_mem_spread_node(void)
2499{
778d3b0f
MH
2500 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2501 current->cpuset_mem_spread_rotor =
2502 node_random(&current->mems_allowed);
2503
6adef3eb
JS
2504 return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2505}
2506
2507int cpuset_slab_spread_node(void)
2508{
778d3b0f
MH
2509 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2510 current->cpuset_slab_spread_rotor =
2511 node_random(&current->mems_allowed);
2512
6adef3eb
JS
2513 return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2514}
2515
825a46af
PJ
2516EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2517
ef08e3b4 2518/**
bbe373f2
DR
2519 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2520 * @tsk1: pointer to task_struct of some task.
2521 * @tsk2: pointer to task_struct of some other task.
2522 *
2523 * Description: Return true if @tsk1's mems_allowed intersects the
2524 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2525 * one of the task's memory usage might impact the memory available
2526 * to the other.
ef08e3b4
PJ
2527 **/
2528
bbe373f2
DR
2529int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2530 const struct task_struct *tsk2)
ef08e3b4 2531{
bbe373f2 2532 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
ef08e3b4
PJ
2533}
2534
f440d98f
LZ
2535#define CPUSET_NODELIST_LEN (256)
2536
75aa1994
DR
2537/**
2538 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2539 * @task: pointer to task_struct of some task.
2540 *
2541 * Description: Prints @task's name, cpuset name, and cached copy of its
b8dadcb5 2542 * mems_allowed to the kernel log.
75aa1994
DR
2543 */
2544void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2545{
f440d98f
LZ
2546 /* Statically allocated to prevent using excess stack. */
2547 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2548 static DEFINE_SPINLOCK(cpuset_buffer_lock);
b8dadcb5 2549 struct cgroup *cgrp;
75aa1994 2550
f440d98f 2551 spin_lock(&cpuset_buffer_lock);
b8dadcb5 2552 rcu_read_lock();
63f43f55 2553
b8dadcb5 2554 cgrp = task_cs(tsk)->css.cgroup;
75aa1994
DR
2555 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2556 tsk->mems_allowed);
e61734c5
TH
2557 printk(KERN_INFO "%s cpuset=", tsk->comm);
2558 pr_cont_cgroup_name(cgrp);
2559 pr_cont(" mems_allowed=%s\n", cpuset_nodelist);
f440d98f 2560
cfb5966b 2561 rcu_read_unlock();
75aa1994
DR
2562 spin_unlock(&cpuset_buffer_lock);
2563}
2564
3e0d98b9
PJ
2565/*
2566 * Collection of memory_pressure is suppressed unless
2567 * this flag is enabled by writing "1" to the special
2568 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2569 */
2570
c5b2aff8 2571int cpuset_memory_pressure_enabled __read_mostly;
3e0d98b9
PJ
2572
2573/**
2574 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2575 *
2576 * Keep a running average of the rate of synchronous (direct)
2577 * page reclaim efforts initiated by tasks in each cpuset.
2578 *
2579 * This represents the rate at which some task in the cpuset
2580 * ran low on memory on all nodes it was allowed to use, and
2581 * had to enter the kernels page reclaim code in an effort to
2582 * create more free memory by tossing clean pages or swapping
2583 * or writing dirty pages.
2584 *
2585 * Display to user space in the per-cpuset read-only file
2586 * "memory_pressure". Value displayed is an integer
2587 * representing the recent rate of entry into the synchronous
2588 * (direct) page reclaim by any task attached to the cpuset.
2589 **/
2590
2591void __cpuset_memory_pressure_bump(void)
2592{
b8dadcb5 2593 rcu_read_lock();
8793d854 2594 fmeter_markevent(&task_cs(current)->fmeter);
b8dadcb5 2595 rcu_read_unlock();
3e0d98b9
PJ
2596}
2597
8793d854 2598#ifdef CONFIG_PROC_PID_CPUSET
1da177e4
LT
2599/*
2600 * proc_cpuset_show()
2601 * - Print tasks cpuset path into seq_file.
2602 * - Used for /proc/<pid>/cpuset.
053199ed
PJ
2603 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2604 * doesn't really matter if tsk->cpuset changes after we read it,
5d21cc2d 2605 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2df167a3 2606 * anyway.
1da177e4 2607 */
8d8b97ba 2608int proc_cpuset_show(struct seq_file *m, void *unused_v)
1da177e4 2609{
13b41b09 2610 struct pid *pid;
1da177e4 2611 struct task_struct *tsk;
e61734c5 2612 char *buf, *p;
8793d854 2613 struct cgroup_subsys_state *css;
99f89551 2614 int retval;
1da177e4 2615
99f89551 2616 retval = -ENOMEM;
e61734c5 2617 buf = kmalloc(PATH_MAX, GFP_KERNEL);
1da177e4 2618 if (!buf)
99f89551
EB
2619 goto out;
2620
2621 retval = -ESRCH;
13b41b09
EB
2622 pid = m->private;
2623 tsk = get_pid_task(pid, PIDTYPE_PID);
99f89551
EB
2624 if (!tsk)
2625 goto out_free;
1da177e4 2626
e61734c5 2627 retval = -ENAMETOOLONG;
27e89ae5 2628 rcu_read_lock();
073219e9 2629 css = task_css(tsk, cpuset_cgrp_id);
e61734c5 2630 p = cgroup_path(css->cgroup, buf, PATH_MAX);
27e89ae5 2631 rcu_read_unlock();
e61734c5 2632 if (!p)
27e89ae5 2633 goto out_put_task;
e61734c5 2634 seq_puts(m, p);
1da177e4 2635 seq_putc(m, '\n');
e61734c5 2636 retval = 0;
27e89ae5 2637out_put_task:
99f89551
EB
2638 put_task_struct(tsk);
2639out_free:
1da177e4 2640 kfree(buf);
99f89551 2641out:
1da177e4
LT
2642 return retval;
2643}
8793d854 2644#endif /* CONFIG_PROC_PID_CPUSET */
1da177e4 2645
d01d4827 2646/* Display task mems_allowed in /proc/<pid>/status file. */
df5f8314
EB
2647void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2648{
df5f8314 2649 seq_printf(m, "Mems_allowed:\t");
30e8e136 2650 seq_nodemask(m, &task->mems_allowed);
df5f8314 2651 seq_printf(m, "\n");
39106dcf 2652 seq_printf(m, "Mems_allowed_list:\t");
30e8e136 2653 seq_nodemask_list(m, &task->mems_allowed);
39106dcf 2654 seq_printf(m, "\n");
1da177e4 2655}