blob: 4683fe728c9b4baef4d9f703ab1e8edf11ab6374 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
Paul Jackson029190c2007-10-18 23:40:20 -07007 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
Paul Menage8793d852007-10-18 23:39:39 -07008 * Copyright (C) 2006 Google, Inc
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 *
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 *
Paul Jackson825a46a2006-03-24 03:16:03 -080013 * 2003-10-10 Written by Simon Derr.
Linus Torvalds1da177e2005-04-16 15:20:36 -070014 * 2003-10-22 Updates by Stephen Hemminger.
Paul Jackson825a46a2006-03-24 03:16:03 -080015 * 2004 May-July Rework by Paul Jackson.
Paul Menage8793d852007-10-18 23:39:39 -070016 * 2006 Rework by Paul Menage to use generic cgroups
Max Krasnyanskycf417142008-08-11 14:33:53 -070017 * 2008 Rework of the scheduler domains and CPU hotplug handling
18 * by Max Krasnyansky
Linus Torvalds1da177e2005-04-16 15:20:36 -070019 *
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
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#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>
Paul Jackson68860ec2005-10-30 15:02:36 -080037#include <linux/mempolicy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/mm.h>
Miao Xief4818912008-11-19 15:36:30 -080039#include <linux/memory.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040#include <linux/module.h>
41#include <linux/mount.h>
42#include <linux/namei.h>
43#include <linux/pagemap.h>
44#include <linux/proc_fs.h>
Paul Jackson6b9c2602006-01-08 01:02:02 -080045#include <linux/rcupdate.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046#include <linux/sched.h>
47#include <linux/seq_file.h>
David Quigley22fb52d2006-06-23 02:04:00 -070048#include <linux/security.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050#include <linux/spinlock.h>
51#include <linux/stat.h>
52#include <linux/string.h>
53#include <linux/time.h>
54#include <linux/backing-dev.h>
55#include <linux/sort.h>
56
57#include <asm/uaccess.h>
58#include <asm/atomic.h>
Ingo Molnar3d3f26a2006-03-23 03:00:18 -080059#include <linux/mutex.h>
Cliff Wickman956db3c2008-02-07 00:14:43 -080060#include <linux/workqueue.h>
61#include <linux/cgroup.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Paul Jackson202f72d2006-01-08 01:01:57 -080063/*
Miao Xief90d4112009-01-16 10:24:10 +080064 * Workqueue for cpuset related tasks.
65 *
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
68 */
69static struct workqueue_struct *cpuset_wq;
70
71/*
Paul Jackson202f72d2006-01-08 01:01:57 -080072 * Tracks how many cpusets are currently defined in system.
73 * When there is only one cpuset (the root cpuset) we can
74 * short circuit some hooks.
75 */
Paul Jackson7edc5962006-01-08 01:02:03 -080076int number_of_cpusets __read_mostly;
Paul Jackson202f72d2006-01-08 01:01:57 -080077
Paul Menage2df167a2008-02-07 00:14:45 -080078/* Forward declare cgroup structures */
Paul Menage8793d852007-10-18 23:39:39 -070079struct cgroup_subsys cpuset_subsys;
80struct cpuset;
81
Paul Jackson3e0d98b2006-01-08 01:01:49 -080082/* See "Frequency meter" comments, below. */
83
84struct fmeter {
85 int cnt; /* unprocessed events count */
86 int val; /* most recent output value */
87 time_t time; /* clock (secs) when val computed */
88 spinlock_t lock; /* guards read or write of above */
89};
90
Linus Torvalds1da177e2005-04-16 15:20:36 -070091struct cpuset {
Paul Menage8793d852007-10-18 23:39:39 -070092 struct cgroup_subsys_state css;
93
Linus Torvalds1da177e2005-04-16 15:20:36 -070094 unsigned long flags; /* "unsigned long" so bitops work */
Li Zefan300ed6c2009-01-07 18:08:44 -080095 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
97
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 struct cpuset *parent; /* my parent */
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Paul Jackson3e0d98b2006-01-08 01:01:49 -0800100 struct fmeter fmeter; /* memory_pressure filter */
Paul Jackson029190c2007-10-18 23:40:20 -0700101
102 /* partition number for rebuild_sched_domains() */
103 int pn;
Cliff Wickman956db3c2008-02-07 00:14:43 -0800104
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900105 /* for custom sched domain */
106 int relax_domain_level;
107
Uwe Kleine-König732bee72010-06-11 12:16:59 +0200108 /* used for walking a cpuset hierarchy */
Cliff Wickman956db3c2008-02-07 00:14:43 -0800109 struct list_head stack_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110};
111
Paul Menage8793d852007-10-18 23:39:39 -0700112/* Retrieve the cpuset for a cgroup */
113static inline struct cpuset *cgroup_cs(struct cgroup *cont)
114{
115 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
116 struct cpuset, css);
117}
118
119/* Retrieve the cpuset for a task */
120static inline struct cpuset *task_cs(struct task_struct *task)
121{
122 return container_of(task_subsys_state(task, cpuset_subsys_id),
123 struct cpuset, css);
124}
Paul Menage8793d852007-10-18 23:39:39 -0700125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126/* bits in struct cpuset flags field */
127typedef enum {
128 CS_CPU_EXCLUSIVE,
129 CS_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -0700130 CS_MEM_HARDWALL,
Paul Jackson45b07ef2006-01-08 01:00:56 -0800131 CS_MEMORY_MIGRATE,
Paul Jackson029190c2007-10-18 23:40:20 -0700132 CS_SCHED_LOAD_BALANCE,
Paul Jackson825a46a2006-03-24 03:16:03 -0800133 CS_SPREAD_PAGE,
134 CS_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135} cpuset_flagbits_t;
136
137/* convenient tests for these bits */
138static inline int is_cpu_exclusive(const struct cpuset *cs)
139{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800140 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141}
142
143static inline int is_mem_exclusive(const struct cpuset *cs)
144{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800145 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146}
147
Paul Menage78608362008-04-29 01:00:26 -0700148static inline int is_mem_hardwall(const struct cpuset *cs)
149{
150 return test_bit(CS_MEM_HARDWALL, &cs->flags);
151}
152
Paul Jackson029190c2007-10-18 23:40:20 -0700153static inline int is_sched_load_balance(const struct cpuset *cs)
154{
155 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
156}
157
Paul Jackson45b07ef2006-01-08 01:00:56 -0800158static inline int is_memory_migrate(const struct cpuset *cs)
159{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800160 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
Paul Jackson45b07ef2006-01-08 01:00:56 -0800161}
162
Paul Jackson825a46a2006-03-24 03:16:03 -0800163static inline int is_spread_page(const struct cpuset *cs)
164{
165 return test_bit(CS_SPREAD_PAGE, &cs->flags);
166}
167
168static inline int is_spread_slab(const struct cpuset *cs)
169{
170 return test_bit(CS_SPREAD_SLAB, &cs->flags);
171}
172
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173static struct cpuset top_cpuset = {
174 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175};
176
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177/*
Paul Menage2df167a2008-02-07 00:14:45 -0800178 * There are two global mutexes guarding cpuset structures. The first
179 * is the main control groups cgroup_mutex, accessed via
180 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
181 * callback_mutex, below. They can nest. It is ok to first take
182 * cgroup_mutex, then nest callback_mutex. We also require taking
183 * task_lock() when dereferencing a task's cpuset pointer. See "The
184 * task_lock() exception", at the end of this comment.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800186 * A task must hold both mutexes to modify cpusets. If a task
Paul Menage2df167a2008-02-07 00:14:45 -0800187 * holds cgroup_mutex, then it blocks others wanting that mutex,
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800188 * ensuring that it is the only task able to also acquire callback_mutex
Paul Jackson053199e2005-10-30 15:02:30 -0800189 * and be able to modify cpusets. It can perform various checks on
190 * the cpuset structure first, knowing nothing will change. It can
Paul Menage2df167a2008-02-07 00:14:45 -0800191 * also allocate memory while just holding cgroup_mutex. While it is
Paul Jackson053199e2005-10-30 15:02:30 -0800192 * performing these checks, various callback routines can briefly
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800193 * acquire callback_mutex to query cpusets. Once it is ready to make
194 * the changes, it takes callback_mutex, blocking everyone else.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 *
Paul Jackson053199e2005-10-30 15:02:30 -0800196 * Calls to the kernel memory allocator can not be made while holding
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800197 * callback_mutex, as that would risk double tripping on callback_mutex
Paul Jackson053199e2005-10-30 15:02:30 -0800198 * from one of the callbacks into the cpuset code from within
199 * __alloc_pages().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800201 * If a task is only holding callback_mutex, then it has read-only
Paul Jackson053199e2005-10-30 15:02:30 -0800202 * access to cpusets.
203 *
Miao Xie58568d22009-06-16 15:31:49 -0700204 * Now, the task_struct fields mems_allowed and mempolicy may be changed
205 * by other task, we use alloc_lock in the task_struct fields to protect
206 * them.
Paul Jackson053199e2005-10-30 15:02:30 -0800207 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800208 * The cpuset_common_file_read() handlers only hold callback_mutex across
Paul Jackson053199e2005-10-30 15:02:30 -0800209 * small pieces of code, such as when reading out possibly multi-word
210 * cpumasks and nodemasks.
211 *
Paul Menage2df167a2008-02-07 00:14:45 -0800212 * Accessing a task's cpuset should be done in accordance with the
213 * guidelines for accessing subsystem state in kernel/cgroup.c
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 */
215
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800216static DEFINE_MUTEX(callback_mutex);
Paul Jackson4247bdc2005-09-10 00:26:06 -0700217
Max Krasnyanskycf417142008-08-11 14:33:53 -0700218/*
David Rientjes75aa1992009-01-06 14:39:01 -0800219 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
220 * buffers. They are statically allocated to prevent using excess stack
221 * when calling cpuset_print_task_mems_allowed().
222 */
223#define CPUSET_NAME_LEN (128)
224#define CPUSET_NODELIST_LEN (256)
225static char cpuset_name[CPUSET_NAME_LEN];
226static char cpuset_nodelist[CPUSET_NODELIST_LEN];
227static DEFINE_SPINLOCK(cpuset_buffer_lock);
228
229/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700230 * This is ugly, but preserves the userspace API for existing cpuset
Paul Menage8793d852007-10-18 23:39:39 -0700231 * users. If someone tries to mount the "cpuset" filesystem, we
Max Krasnyanskycf417142008-08-11 14:33:53 -0700232 * silently switch it to mount "cgroup" instead
233 */
Al Virof7e83572010-07-26 13:23:11 +0400234static struct dentry *cpuset_mount(struct file_system_type *fs_type,
235 int flags, const char *unused_dev_name, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
Paul Menage8793d852007-10-18 23:39:39 -0700237 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
Al Virof7e83572010-07-26 13:23:11 +0400238 struct dentry *ret = ERR_PTR(-ENODEV);
Paul Menage8793d852007-10-18 23:39:39 -0700239 if (cgroup_fs) {
240 char mountopts[] =
241 "cpuset,noprefix,"
242 "release_agent=/sbin/cpuset_release_agent";
Al Virof7e83572010-07-26 13:23:11 +0400243 ret = cgroup_fs->mount(cgroup_fs, flags,
244 unused_dev_name, mountopts);
Paul Menage8793d852007-10-18 23:39:39 -0700245 put_filesystem(cgroup_fs);
246 }
247 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248}
249
250static struct file_system_type cpuset_fs_type = {
251 .name = "cpuset",
Al Virof7e83572010-07-26 13:23:11 +0400252 .mount = cpuset_mount,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253};
254
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255/*
Li Zefan300ed6c2009-01-07 18:08:44 -0800256 * Return in pmask the portion of a cpusets's cpus_allowed that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 * are online. If none are online, walk up the cpuset hierarchy
258 * until we find one that does have some online cpus. If we get
259 * all the way to the top and still haven't found any online cpus,
260 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
261 * task, return cpu_online_map.
262 *
263 * One way or another, we guarantee to return some non-empty subset
264 * of cpu_online_map.
265 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800266 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 */
268
Li Zefan6af866a2009-01-07 18:08:45 -0800269static void guarantee_online_cpus(const struct cpuset *cs,
270 struct cpumask *pmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271{
Li Zefan300ed6c2009-01-07 18:08:44 -0800272 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273 cs = cs->parent;
274 if (cs)
Li Zefan300ed6c2009-01-07 18:08:44 -0800275 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 else
Li Zefan300ed6c2009-01-07 18:08:44 -0800277 cpumask_copy(pmask, cpu_online_mask);
278 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279}
280
281/*
282 * Return in *pmask the portion of a cpusets's mems_allowed that
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700283 * are online, with memory. If none are online with memory, walk
284 * up the cpuset hierarchy until we find one that does have some
285 * online mems. If we get all the way to the top and still haven't
286 * found any online mems, return node_states[N_HIGH_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 *
288 * One way or another, we guarantee to return some non-empty subset
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700289 * of node_states[N_HIGH_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800291 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 */
293
294static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
295{
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700296 while (cs && !nodes_intersects(cs->mems_allowed,
297 node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 cs = cs->parent;
299 if (cs)
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700300 nodes_and(*pmask, cs->mems_allowed,
301 node_states[N_HIGH_MEMORY]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 else
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700303 *pmask = node_states[N_HIGH_MEMORY];
304 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305}
306
Miao Xief3b39d42009-06-16 15:31:46 -0700307/*
308 * update task's spread flag if cpuset's page/slab spread flag is set
309 *
310 * Called with callback_mutex/cgroup_mutex held
311 */
312static void cpuset_update_task_spread_flag(struct cpuset *cs,
313 struct task_struct *tsk)
314{
315 if (is_spread_page(cs))
316 tsk->flags |= PF_SPREAD_PAGE;
317 else
318 tsk->flags &= ~PF_SPREAD_PAGE;
319 if (is_spread_slab(cs))
320 tsk->flags |= PF_SPREAD_SLAB;
321 else
322 tsk->flags &= ~PF_SPREAD_SLAB;
323}
324
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325/*
326 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
327 *
328 * One cpuset is a subset of another if all its allowed CPUs and
329 * Memory Nodes are a subset of the other, and its exclusive flags
Paul Menage2df167a2008-02-07 00:14:45 -0800330 * are only set if the other's are set. Call holding cgroup_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331 */
332
333static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
334{
Li Zefan300ed6c2009-01-07 18:08:44 -0800335 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 nodes_subset(p->mems_allowed, q->mems_allowed) &&
337 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
338 is_mem_exclusive(p) <= is_mem_exclusive(q);
339}
340
Li Zefan645fcc92009-01-07 18:08:43 -0800341/**
342 * alloc_trial_cpuset - allocate a trial cpuset
343 * @cs: the cpuset that the trial cpuset duplicates
344 */
345static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
346{
Li Zefan300ed6c2009-01-07 18:08:44 -0800347 struct cpuset *trial;
348
349 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
350 if (!trial)
351 return NULL;
352
353 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
354 kfree(trial);
355 return NULL;
356 }
357 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
358
359 return trial;
Li Zefan645fcc92009-01-07 18:08:43 -0800360}
361
362/**
363 * free_trial_cpuset - free the trial cpuset
364 * @trial: the trial cpuset to be freed
365 */
366static void free_trial_cpuset(struct cpuset *trial)
367{
Li Zefan300ed6c2009-01-07 18:08:44 -0800368 free_cpumask_var(trial->cpus_allowed);
Li Zefan645fcc92009-01-07 18:08:43 -0800369 kfree(trial);
370}
371
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372/*
373 * validate_change() - Used to validate that any proposed cpuset change
374 * follows the structural rules for cpusets.
375 *
376 * If we replaced the flag and mask values of the current cpuset
377 * (cur) with those values in the trial cpuset (trial), would
378 * our various subset and exclusive rules still be valid? Presumes
Paul Menage2df167a2008-02-07 00:14:45 -0800379 * cgroup_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 *
381 * 'cur' is the address of an actual, in-use cpuset. Operations
382 * such as list traversal that depend on the actual address of the
383 * cpuset in the list must use cur below, not trial.
384 *
385 * 'trial' is the address of bulk structure copy of cur, with
386 * perhaps one or more of the fields cpus_allowed, mems_allowed,
387 * or flags changed to new, trial values.
388 *
389 * Return 0 if valid, -errno if not.
390 */
391
392static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
393{
Paul Menage8793d852007-10-18 23:39:39 -0700394 struct cgroup *cont;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 struct cpuset *c, *par;
396
397 /* Each of our child cpusets must be a subset of us */
Paul Menage8793d852007-10-18 23:39:39 -0700398 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
399 if (!is_cpuset_subset(cgroup_cs(cont), trial))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 return -EBUSY;
401 }
402
403 /* Remaining checks don't apply to root cpuset */
Paul Jackson69604062006-12-06 20:36:15 -0800404 if (cur == &top_cpuset)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 return 0;
406
Paul Jackson69604062006-12-06 20:36:15 -0800407 par = cur->parent;
408
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 /* We must be a subset of our parent cpuset */
410 if (!is_cpuset_subset(trial, par))
411 return -EACCES;
412
Paul Menage2df167a2008-02-07 00:14:45 -0800413 /*
414 * If either I or some sibling (!= me) is exclusive, we can't
415 * overlap
416 */
Paul Menage8793d852007-10-18 23:39:39 -0700417 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
418 c = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
420 c != cur &&
Li Zefan300ed6c2009-01-07 18:08:44 -0800421 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 return -EINVAL;
423 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
424 c != cur &&
425 nodes_intersects(trial->mems_allowed, c->mems_allowed))
426 return -EINVAL;
427 }
428
Paul Jackson020958b2007-10-18 23:40:21 -0700429 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
430 if (cgroup_task_count(cur->css.cgroup)) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800431 if (cpumask_empty(trial->cpus_allowed) ||
Paul Jackson020958b2007-10-18 23:40:21 -0700432 nodes_empty(trial->mems_allowed)) {
433 return -ENOSPC;
434 }
435 }
436
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 return 0;
438}
439
Paul Menagedb7f47c2009-04-02 16:57:55 -0700440#ifdef CONFIG_SMP
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700441/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700442 * Helper routine for generate_sched_domains().
Paul Jackson029190c2007-10-18 23:40:20 -0700443 * Do cpusets a, b have overlapping cpus_allowed masks?
444 */
Paul Jackson029190c2007-10-18 23:40:20 -0700445static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
446{
Li Zefan300ed6c2009-01-07 18:08:44 -0800447 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
Paul Jackson029190c2007-10-18 23:40:20 -0700448}
449
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900450static void
451update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
452{
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900453 if (dattr->relax_domain_level < c->relax_domain_level)
454 dattr->relax_domain_level = c->relax_domain_level;
455 return;
456}
457
Lai Jiangshanf5393692008-07-29 22:33:22 -0700458static void
459update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
460{
461 LIST_HEAD(q);
462
463 list_add(&c->stack_list, &q);
464 while (!list_empty(&q)) {
465 struct cpuset *cp;
466 struct cgroup *cont;
467 struct cpuset *child;
468
469 cp = list_first_entry(&q, struct cpuset, stack_list);
470 list_del(q.next);
471
Li Zefan300ed6c2009-01-07 18:08:44 -0800472 if (cpumask_empty(cp->cpus_allowed))
Lai Jiangshanf5393692008-07-29 22:33:22 -0700473 continue;
474
475 if (is_sched_load_balance(cp))
476 update_domain_attr(dattr, cp);
477
478 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
479 child = cgroup_cs(cont);
480 list_add_tail(&child->stack_list, &q);
481 }
482 }
483}
484
Paul Jackson029190c2007-10-18 23:40:20 -0700485/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700486 * generate_sched_domains()
Paul Jackson029190c2007-10-18 23:40:20 -0700487 *
Max Krasnyanskycf417142008-08-11 14:33:53 -0700488 * This function builds a partial partition of the systems CPUs
489 * A 'partial partition' is a set of non-overlapping subsets whose
490 * union is a subset of that set.
491 * The output of this function needs to be passed to kernel/sched.c
492 * partition_sched_domains() routine, which will rebuild the scheduler's
493 * load balancing domains (sched domains) as specified by that partial
494 * partition.
Paul Jackson029190c2007-10-18 23:40:20 -0700495 *
Li Zefan45ce80f2009-01-15 13:50:59 -0800496 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
Paul Jackson029190c2007-10-18 23:40:20 -0700497 * for a background explanation of this.
498 *
499 * Does not return errors, on the theory that the callers of this
500 * routine would rather not worry about failures to rebuild sched
501 * domains when operating in the severe memory shortage situations
502 * that could cause allocation failures below.
503 *
Max Krasnyanskycf417142008-08-11 14:33:53 -0700504 * Must be called with cgroup_lock held.
Paul Jackson029190c2007-10-18 23:40:20 -0700505 *
506 * The three key local variables below are:
Li Zefanaeed6822008-07-29 22:33:24 -0700507 * q - a linked-list queue of cpuset pointers, used to implement a
Paul Jackson029190c2007-10-18 23:40:20 -0700508 * top-down scan of all cpusets. This scan loads a pointer
509 * to each cpuset marked is_sched_load_balance into the
510 * array 'csa'. For our purposes, rebuilding the schedulers
511 * sched domains, we can ignore !is_sched_load_balance cpusets.
512 * csa - (for CpuSet Array) Array of pointers to all the cpusets
513 * that need to be load balanced, for convenient iterative
514 * access by the subsequent code that finds the best partition,
515 * i.e the set of domains (subsets) of CPUs such that the
516 * cpus_allowed of every cpuset marked is_sched_load_balance
517 * is a subset of one of these domains, while there are as
518 * many such domains as possible, each as small as possible.
519 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
520 * the kernel/sched.c routine partition_sched_domains() in a
521 * convenient format, that can be easily compared to the prior
522 * value to determine what partition elements (sched domains)
523 * were changed (added or removed.)
524 *
525 * Finding the best partition (set of domains):
526 * The triple nested loops below over i, j, k scan over the
527 * load balanced cpusets (using the array of cpuset pointers in
528 * csa[]) looking for pairs of cpusets that have overlapping
529 * cpus_allowed, but which don't have the same 'pn' partition
530 * number and gives them in the same partition number. It keeps
531 * looping on the 'restart' label until it can no longer find
532 * any such pairs.
533 *
534 * The union of the cpus_allowed masks from the set of
535 * all cpusets having the same 'pn' value then form the one
536 * element of the partition (one sched domain) to be passed to
537 * partition_sched_domains().
538 */
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030539static int generate_sched_domains(cpumask_var_t **domains,
Max Krasnyanskycf417142008-08-11 14:33:53 -0700540 struct sched_domain_attr **attributes)
Paul Jackson029190c2007-10-18 23:40:20 -0700541{
Max Krasnyanskycf417142008-08-11 14:33:53 -0700542 LIST_HEAD(q); /* queue of cpusets to be scanned */
Paul Jackson029190c2007-10-18 23:40:20 -0700543 struct cpuset *cp; /* scans q */
544 struct cpuset **csa; /* array of all cpuset ptrs */
545 int csn; /* how many cpuset ptrs in csa so far */
546 int i, j, k; /* indices for partition finding loops */
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030547 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900548 struct sched_domain_attr *dattr; /* attributes for custom domains */
Ingo Molnar15837152008-11-25 10:27:49 +0100549 int ndoms = 0; /* number of sched domains in result */
Li Zefan6af866a2009-01-07 18:08:45 -0800550 int nslot; /* next empty doms[] struct cpumask slot */
Paul Jackson029190c2007-10-18 23:40:20 -0700551
Paul Jackson029190c2007-10-18 23:40:20 -0700552 doms = NULL;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900553 dattr = NULL;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700554 csa = NULL;
Paul Jackson029190c2007-10-18 23:40:20 -0700555
556 /* Special case for the 99% of systems with one, full, sched domain */
557 if (is_sched_load_balance(&top_cpuset)) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030558 ndoms = 1;
559 doms = alloc_sched_domains(ndoms);
Paul Jackson029190c2007-10-18 23:40:20 -0700560 if (!doms)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700561 goto done;
562
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900563 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
564 if (dattr) {
565 *dattr = SD_ATTR_INIT;
Li Zefan93a65572008-07-29 22:33:23 -0700566 update_domain_attr_tree(dattr, &top_cpuset);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900567 }
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030568 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700569
Max Krasnyanskycf417142008-08-11 14:33:53 -0700570 goto done;
Paul Jackson029190c2007-10-18 23:40:20 -0700571 }
572
Paul Jackson029190c2007-10-18 23:40:20 -0700573 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
574 if (!csa)
575 goto done;
576 csn = 0;
577
Li Zefanaeed6822008-07-29 22:33:24 -0700578 list_add(&top_cpuset.stack_list, &q);
579 while (!list_empty(&q)) {
Paul Jackson029190c2007-10-18 23:40:20 -0700580 struct cgroup *cont;
581 struct cpuset *child; /* scans child cpusets of cp */
Lai Jiangshan489a5392008-07-25 01:47:23 -0700582
Li Zefanaeed6822008-07-29 22:33:24 -0700583 cp = list_first_entry(&q, struct cpuset, stack_list);
584 list_del(q.next);
585
Li Zefan300ed6c2009-01-07 18:08:44 -0800586 if (cpumask_empty(cp->cpus_allowed))
Lai Jiangshan489a5392008-07-25 01:47:23 -0700587 continue;
588
Lai Jiangshanf5393692008-07-29 22:33:22 -0700589 /*
590 * All child cpusets contain a subset of the parent's cpus, so
591 * just skip them, and then we call update_domain_attr_tree()
592 * to calc relax_domain_level of the corresponding sched
593 * domain.
594 */
595 if (is_sched_load_balance(cp)) {
Paul Jackson029190c2007-10-18 23:40:20 -0700596 csa[csn++] = cp;
Lai Jiangshanf5393692008-07-29 22:33:22 -0700597 continue;
598 }
Lai Jiangshan489a5392008-07-25 01:47:23 -0700599
Paul Jackson029190c2007-10-18 23:40:20 -0700600 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
601 child = cgroup_cs(cont);
Li Zefanaeed6822008-07-29 22:33:24 -0700602 list_add_tail(&child->stack_list, &q);
Paul Jackson029190c2007-10-18 23:40:20 -0700603 }
604 }
605
606 for (i = 0; i < csn; i++)
607 csa[i]->pn = i;
608 ndoms = csn;
609
610restart:
611 /* Find the best partition (set of sched domains) */
612 for (i = 0; i < csn; i++) {
613 struct cpuset *a = csa[i];
614 int apn = a->pn;
615
616 for (j = 0; j < csn; j++) {
617 struct cpuset *b = csa[j];
618 int bpn = b->pn;
619
620 if (apn != bpn && cpusets_overlap(a, b)) {
621 for (k = 0; k < csn; k++) {
622 struct cpuset *c = csa[k];
623
624 if (c->pn == bpn)
625 c->pn = apn;
626 }
627 ndoms--; /* one less element */
628 goto restart;
629 }
630 }
631 }
632
Max Krasnyanskycf417142008-08-11 14:33:53 -0700633 /*
634 * Now we know how many domains to create.
635 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
636 */
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030637 doms = alloc_sched_domains(ndoms);
Li Zefan700018e2008-11-18 14:02:03 +0800638 if (!doms)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700639 goto done;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700640
641 /*
642 * The rest of the code, including the scheduler, can deal with
643 * dattr==NULL case. No need to abort if alloc fails.
644 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900645 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
Paul Jackson029190c2007-10-18 23:40:20 -0700646
647 for (nslot = 0, i = 0; i < csn; i++) {
648 struct cpuset *a = csa[i];
Li Zefan6af866a2009-01-07 18:08:45 -0800649 struct cpumask *dp;
Paul Jackson029190c2007-10-18 23:40:20 -0700650 int apn = a->pn;
651
Max Krasnyanskycf417142008-08-11 14:33:53 -0700652 if (apn < 0) {
653 /* Skip completed partitions */
654 continue;
Paul Jackson029190c2007-10-18 23:40:20 -0700655 }
Max Krasnyanskycf417142008-08-11 14:33:53 -0700656
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030657 dp = doms[nslot];
Max Krasnyanskycf417142008-08-11 14:33:53 -0700658
659 if (nslot == ndoms) {
660 static int warnings = 10;
661 if (warnings) {
662 printk(KERN_WARNING
663 "rebuild_sched_domains confused:"
664 " nslot %d, ndoms %d, csn %d, i %d,"
665 " apn %d\n",
666 nslot, ndoms, csn, i, apn);
667 warnings--;
668 }
669 continue;
670 }
671
Li Zefan6af866a2009-01-07 18:08:45 -0800672 cpumask_clear(dp);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700673 if (dattr)
674 *(dattr + nslot) = SD_ATTR_INIT;
675 for (j = i; j < csn; j++) {
676 struct cpuset *b = csa[j];
677
678 if (apn == b->pn) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800679 cpumask_or(dp, dp, b->cpus_allowed);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700680 if (dattr)
681 update_domain_attr_tree(dattr + nslot, b);
682
683 /* Done with this partition */
684 b->pn = -1;
685 }
686 }
687 nslot++;
Paul Jackson029190c2007-10-18 23:40:20 -0700688 }
689 BUG_ON(nslot != ndoms);
690
Paul Jackson029190c2007-10-18 23:40:20 -0700691done:
Paul Jackson029190c2007-10-18 23:40:20 -0700692 kfree(csa);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700693
Li Zefan700018e2008-11-18 14:02:03 +0800694 /*
695 * Fallback to the default domain if kmalloc() failed.
696 * See comments in partition_sched_domains().
697 */
698 if (doms == NULL)
699 ndoms = 1;
700
Max Krasnyanskycf417142008-08-11 14:33:53 -0700701 *domains = doms;
702 *attributes = dattr;
703 return ndoms;
704}
705
706/*
707 * Rebuild scheduler domains.
708 *
709 * Call with neither cgroup_mutex held nor within get_online_cpus().
710 * Takes both cgroup_mutex and get_online_cpus().
711 *
712 * Cannot be directly called from cpuset code handling changes
713 * to the cpuset pseudo-filesystem, because it cannot be called
714 * from code that already holds cgroup_mutex.
715 */
716static void do_rebuild_sched_domains(struct work_struct *unused)
717{
718 struct sched_domain_attr *attr;
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030719 cpumask_var_t *doms;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700720 int ndoms;
721
722 get_online_cpus();
723
724 /* Generate domain masks and attrs */
725 cgroup_lock();
726 ndoms = generate_sched_domains(&doms, &attr);
727 cgroup_unlock();
728
729 /* Have scheduler rebuild the domains */
730 partition_sched_domains(ndoms, doms, attr);
731
732 put_online_cpus();
733}
Paul Menagedb7f47c2009-04-02 16:57:55 -0700734#else /* !CONFIG_SMP */
735static void do_rebuild_sched_domains(struct work_struct *unused)
736{
737}
738
Geert Uytterhoevene1b80902009-12-06 20:41:16 +0100739static int generate_sched_domains(cpumask_var_t **domains,
Paul Menagedb7f47c2009-04-02 16:57:55 -0700740 struct sched_domain_attr **attributes)
741{
742 *domains = NULL;
743 return 1;
744}
745#endif /* CONFIG_SMP */
Max Krasnyanskycf417142008-08-11 14:33:53 -0700746
747static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
748
749/*
750 * Rebuild scheduler domains, asynchronously via workqueue.
751 *
752 * If the flag 'sched_load_balance' of any cpuset with non-empty
753 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
754 * which has that flag enabled, or if any cpuset with a non-empty
755 * 'cpus' is removed, then call this routine to rebuild the
756 * scheduler's dynamic sched domains.
757 *
758 * The rebuild_sched_domains() and partition_sched_domains()
759 * routines must nest cgroup_lock() inside get_online_cpus(),
760 * but such cpuset changes as these must nest that locking the
761 * other way, holding cgroup_lock() for much of the code.
762 *
763 * So in order to avoid an ABBA deadlock, the cpuset code handling
764 * these user changes delegates the actual sched domain rebuilding
765 * to a separate workqueue thread, which ends up processing the
766 * above do_rebuild_sched_domains() function.
767 */
768static void async_rebuild_sched_domains(void)
769{
Miao Xief90d4112009-01-16 10:24:10 +0800770 queue_work(cpuset_wq, &rebuild_sched_domains_work);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700771}
772
773/*
774 * Accomplishes the same scheduler domain rebuild as the above
775 * async_rebuild_sched_domains(), however it directly calls the
776 * rebuild routine synchronously rather than calling it via an
777 * asynchronous work thread.
778 *
779 * This can only be called from code that is not holding
780 * cgroup_mutex (not nested in a cgroup_lock() call.)
781 */
782void rebuild_sched_domains(void)
783{
784 do_rebuild_sched_domains(NULL);
Paul Jackson029190c2007-10-18 23:40:20 -0700785}
786
Cliff Wickman58f47902008-02-07 00:14:44 -0800787/**
788 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
789 * @tsk: task to test
790 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
791 *
Paul Menage2df167a2008-02-07 00:14:45 -0800792 * Call with cgroup_mutex held. May take callback_mutex during call.
Cliff Wickman58f47902008-02-07 00:14:44 -0800793 * Called for each task in a cgroup by cgroup_scan_tasks().
794 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
795 * words, if its mask is not equal to its cpuset's mask).
Paul Jackson053199e2005-10-30 15:02:30 -0800796 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700797static int cpuset_test_cpumask(struct task_struct *tsk,
798 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800799{
Li Zefan300ed6c2009-01-07 18:08:44 -0800800 return !cpumask_equal(&tsk->cpus_allowed,
Cliff Wickman58f47902008-02-07 00:14:44 -0800801 (cgroup_cs(scan->cg))->cpus_allowed);
802}
Paul Jackson053199e2005-10-30 15:02:30 -0800803
Cliff Wickman58f47902008-02-07 00:14:44 -0800804/**
805 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
806 * @tsk: task to test
807 * @scan: struct cgroup_scanner containing the cgroup of the task
808 *
809 * Called by cgroup_scan_tasks() for each task in a cgroup whose
810 * cpus_allowed mask needs to be changed.
811 *
812 * We don't need to re-check for the cgroup/cpuset membership, since we're
813 * holding cgroup_lock() at this point.
814 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700815static void cpuset_change_cpumask(struct task_struct *tsk,
816 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800817{
Li Zefan300ed6c2009-01-07 18:08:44 -0800818 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
Cliff Wickman58f47902008-02-07 00:14:44 -0800819}
820
821/**
Miao Xie0b2f6302008-07-25 01:47:21 -0700822 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
823 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
Li Zefan4e743392008-09-13 02:33:08 -0700824 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
Miao Xie0b2f6302008-07-25 01:47:21 -0700825 *
826 * Called with cgroup_mutex held
827 *
828 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
829 * calling callback functions for each.
830 *
Li Zefan4e743392008-09-13 02:33:08 -0700831 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
832 * if @heap != NULL.
Miao Xie0b2f6302008-07-25 01:47:21 -0700833 */
Li Zefan4e743392008-09-13 02:33:08 -0700834static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
Miao Xie0b2f6302008-07-25 01:47:21 -0700835{
836 struct cgroup_scanner scan;
Miao Xie0b2f6302008-07-25 01:47:21 -0700837
838 scan.cg = cs->css.cgroup;
839 scan.test_task = cpuset_test_cpumask;
840 scan.process_task = cpuset_change_cpumask;
Li Zefan4e743392008-09-13 02:33:08 -0700841 scan.heap = heap;
842 cgroup_scan_tasks(&scan);
Miao Xie0b2f6302008-07-25 01:47:21 -0700843}
844
845/**
Cliff Wickman58f47902008-02-07 00:14:44 -0800846 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
847 * @cs: the cpuset to consider
848 * @buf: buffer of cpu numbers written to this cpuset
849 */
Li Zefan645fcc92009-01-07 18:08:43 -0800850static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
851 const char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852{
Li Zefan4e743392008-09-13 02:33:08 -0700853 struct ptr_heap heap;
Cliff Wickman58f47902008-02-07 00:14:44 -0800854 int retval;
855 int is_load_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856
Paul Jackson4c4d50f2006-08-27 01:23:51 -0700857 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
858 if (cs == &top_cpuset)
859 return -EACCES;
860
David Rientjes6f7f02e2007-05-08 00:31:43 -0700861 /*
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800862 * An empty cpus_allowed is ok only if the cpuset has no tasks.
Paul Jackson020958b2007-10-18 23:40:21 -0700863 * Since cpulist_parse() fails on an empty mask, we special case
864 * that parsing. The validate_change() call ensures that cpusets
865 * with tasks have cpus.
David Rientjes6f7f02e2007-05-08 00:31:43 -0700866 */
Paul Jackson020958b2007-10-18 23:40:21 -0700867 if (!*buf) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800868 cpumask_clear(trialcs->cpus_allowed);
David Rientjes6f7f02e2007-05-08 00:31:43 -0700869 } else {
Li Zefan300ed6c2009-01-07 18:08:44 -0800870 retval = cpulist_parse(buf, trialcs->cpus_allowed);
David Rientjes6f7f02e2007-05-08 00:31:43 -0700871 if (retval < 0)
872 return retval;
Lai Jiangshan37340742008-06-05 22:46:32 -0700873
Peter Zijlstra6ad4c182009-11-25 13:31:39 +0100874 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
Lai Jiangshan37340742008-06-05 22:46:32 -0700875 return -EINVAL;
David Rientjes6f7f02e2007-05-08 00:31:43 -0700876 }
Li Zefan645fcc92009-01-07 18:08:43 -0800877 retval = validate_change(cs, trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700878 if (retval < 0)
879 return retval;
Paul Jackson029190c2007-10-18 23:40:20 -0700880
Paul Menage8707d8b2007-10-18 23:40:22 -0700881 /* Nothing to do if the cpus didn't change */
Li Zefan300ed6c2009-01-07 18:08:44 -0800882 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
Paul Menage8707d8b2007-10-18 23:40:22 -0700883 return 0;
Cliff Wickman58f47902008-02-07 00:14:44 -0800884
Li Zefan4e743392008-09-13 02:33:08 -0700885 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
886 if (retval)
887 return retval;
888
Li Zefan645fcc92009-01-07 18:08:43 -0800889 is_load_balanced = is_sched_load_balance(trialcs);
Paul Jackson029190c2007-10-18 23:40:20 -0700890
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800891 mutex_lock(&callback_mutex);
Li Zefan300ed6c2009-01-07 18:08:44 -0800892 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800893 mutex_unlock(&callback_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -0700894
Paul Menage8707d8b2007-10-18 23:40:22 -0700895 /*
896 * Scan tasks in the cpuset, and update the cpumasks of any
Cliff Wickman58f47902008-02-07 00:14:44 -0800897 * that need an update.
Paul Menage8707d8b2007-10-18 23:40:22 -0700898 */
Li Zefan4e743392008-09-13 02:33:08 -0700899 update_tasks_cpumask(cs, &heap);
900
901 heap_free(&heap);
Cliff Wickman58f47902008-02-07 00:14:44 -0800902
Paul Menage8707d8b2007-10-18 23:40:22 -0700903 if (is_load_balanced)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700904 async_rebuild_sched_domains();
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700905 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906}
907
Paul Jackson053199e2005-10-30 15:02:30 -0800908/*
Paul Jacksone4e364e2006-03-31 02:30:52 -0800909 * cpuset_migrate_mm
910 *
911 * Migrate memory region from one set of nodes to another.
912 *
913 * Temporarilly set tasks mems_allowed to target nodes of migration,
914 * so that the migration code can allocate pages on these nodes.
915 *
Paul Menage2df167a2008-02-07 00:14:45 -0800916 * Call holding cgroup_mutex, so current's cpuset won't change
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800917 * during this call, as manage_mutex holds off any cpuset_attach()
Paul Jacksone4e364e2006-03-31 02:30:52 -0800918 * calls. Therefore we don't need to take task_lock around the
919 * call to guarantee_online_mems(), as we know no one is changing
Paul Menage2df167a2008-02-07 00:14:45 -0800920 * our task's cpuset.
Paul Jacksone4e364e2006-03-31 02:30:52 -0800921 *
Paul Jacksone4e364e2006-03-31 02:30:52 -0800922 * While the mm_struct we are migrating is typically from some
923 * other task, the task_struct mems_allowed that we are hacking
924 * is for our current task, which must allocate new pages for that
925 * migrating memory region.
Paul Jacksone4e364e2006-03-31 02:30:52 -0800926 */
927
928static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
929 const nodemask_t *to)
930{
931 struct task_struct *tsk = current;
932
Paul Jacksone4e364e2006-03-31 02:30:52 -0800933 tsk->mems_allowed = *to;
Paul Jacksone4e364e2006-03-31 02:30:52 -0800934
935 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
936
Paul Menage8793d852007-10-18 23:39:39 -0700937 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
Paul Jacksone4e364e2006-03-31 02:30:52 -0800938}
939
Li Zefan3b6766f2009-04-02 16:57:51 -0700940/*
Miao Xie58568d22009-06-16 15:31:49 -0700941 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
942 * @tsk: the task to change
943 * @newmems: new nodes that the task will be set
944 *
945 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
946 * we structure updates as setting all new allowed nodes, then clearing newly
947 * disallowed ones.
Miao Xie58568d22009-06-16 15:31:49 -0700948 */
949static void cpuset_change_task_nodemask(struct task_struct *tsk,
950 nodemask_t *newmems)
951{
Miao Xiec0ff7452010-05-24 14:32:08 -0700952repeat:
953 /*
954 * Allow tasks that have access to memory reserves because they have
955 * been OOM killed to get memory anywhere.
956 */
957 if (unlikely(test_thread_flag(TIF_MEMDIE)))
958 return;
959 if (current->flags & PF_EXITING) /* Let dying task have memory */
960 return;
961
962 task_lock(tsk);
Miao Xie58568d22009-06-16 15:31:49 -0700963 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
Miao Xiec0ff7452010-05-24 14:32:08 -0700964 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
965
966
967 /*
968 * ensure checking ->mems_allowed_change_disable after setting all new
969 * allowed nodes.
970 *
971 * the read-side task can see an nodemask with new allowed nodes and
972 * old allowed nodes. and if it allocates page when cpuset clears newly
973 * disallowed ones continuous, it can see the new allowed bits.
974 *
975 * And if setting all new allowed nodes is after the checking, setting
976 * all new allowed nodes and clearing newly disallowed ones will be done
977 * continuous, and the read-side task may find no node to alloc page.
978 */
979 smp_mb();
980
981 /*
982 * Allocation of memory is very fast, we needn't sleep when waiting
983 * for the read-side.
984 */
985 while (ACCESS_ONCE(tsk->mems_allowed_change_disable)) {
986 task_unlock(tsk);
987 if (!task_curr(tsk))
988 yield();
989 goto repeat;
990 }
991
992 /*
993 * ensure checking ->mems_allowed_change_disable before clearing all new
994 * disallowed nodes.
995 *
996 * if clearing newly disallowed bits before the checking, the read-side
997 * task may find no node to alloc page.
998 */
999 smp_mb();
1000
1001 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
Miao Xie58568d22009-06-16 15:31:49 -07001002 tsk->mems_allowed = *newmems;
Miao Xiec0ff7452010-05-24 14:32:08 -07001003 task_unlock(tsk);
Miao Xie58568d22009-06-16 15:31:49 -07001004}
1005
1006/*
1007 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1008 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1009 * memory_migrate flag is set. Called with cgroup_mutex held.
Li Zefan3b6766f2009-04-02 16:57:51 -07001010 */
1011static void cpuset_change_nodemask(struct task_struct *p,
1012 struct cgroup_scanner *scan)
1013{
1014 struct mm_struct *mm;
1015 struct cpuset *cs;
1016 int migrate;
1017 const nodemask_t *oldmem = scan->data;
Miao Xie53feb292010-03-23 13:35:35 -07001018 NODEMASK_ALLOC(nodemask_t, newmems, GFP_KERNEL);
1019
1020 if (!newmems)
1021 return;
Miao Xie58568d22009-06-16 15:31:49 -07001022
1023 cs = cgroup_cs(scan->cg);
Miao Xie53feb292010-03-23 13:35:35 -07001024 guarantee_online_mems(cs, newmems);
Miao Xie58568d22009-06-16 15:31:49 -07001025
Miao Xie53feb292010-03-23 13:35:35 -07001026 cpuset_change_task_nodemask(p, newmems);
Li Zefan3b6766f2009-04-02 16:57:51 -07001027
Miao Xie53feb292010-03-23 13:35:35 -07001028 NODEMASK_FREE(newmems);
1029
Li Zefan3b6766f2009-04-02 16:57:51 -07001030 mm = get_task_mm(p);
1031 if (!mm)
1032 return;
1033
Li Zefan3b6766f2009-04-02 16:57:51 -07001034 migrate = is_memory_migrate(cs);
1035
1036 mpol_rebind_mm(mm, &cs->mems_allowed);
1037 if (migrate)
1038 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1039 mmput(mm);
1040}
1041
Paul Menage8793d852007-10-18 23:39:39 -07001042static void *cpuset_being_rebound;
1043
Miao Xie0b2f6302008-07-25 01:47:21 -07001044/**
1045 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1046 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1047 * @oldmem: old mems_allowed of cpuset cs
Li Zefan010cfac2009-04-02 16:57:52 -07001048 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
Miao Xie0b2f6302008-07-25 01:47:21 -07001049 *
1050 * Called with cgroup_mutex held
Li Zefan010cfac2009-04-02 16:57:52 -07001051 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1052 * if @heap != NULL.
Miao Xie0b2f6302008-07-25 01:47:21 -07001053 */
Li Zefan010cfac2009-04-02 16:57:52 -07001054static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1055 struct ptr_heap *heap)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056{
Li Zefan3b6766f2009-04-02 16:57:51 -07001057 struct cgroup_scanner scan;
Paul Jackson59dac162006-01-08 01:01:52 -08001058
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001059 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
Paul Jackson42253992006-01-08 01:01:59 -08001060
Li Zefan3b6766f2009-04-02 16:57:51 -07001061 scan.cg = cs->css.cgroup;
1062 scan.test_task = NULL;
1063 scan.process_task = cpuset_change_nodemask;
Li Zefan010cfac2009-04-02 16:57:52 -07001064 scan.heap = heap;
Li Zefan3b6766f2009-04-02 16:57:51 -07001065 scan.data = (nodemask_t *)oldmem;
Paul Jackson42253992006-01-08 01:01:59 -08001066
1067 /*
Li Zefan3b6766f2009-04-02 16:57:51 -07001068 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1069 * take while holding tasklist_lock. Forks can happen - the
1070 * mpol_dup() cpuset_being_rebound check will catch such forks,
1071 * and rebind their vma mempolicies too. Because we still hold
1072 * the global cgroup_mutex, we know that no other rebind effort
1073 * will be contending for the global variable cpuset_being_rebound.
Paul Jackson42253992006-01-08 01:01:59 -08001074 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
Paul Jackson04c19fa2006-01-08 01:02:00 -08001075 * is idempotent. Also migrate pages in each mm to new nodes.
Paul Jackson42253992006-01-08 01:01:59 -08001076 */
Li Zefan010cfac2009-04-02 16:57:52 -07001077 cgroup_scan_tasks(&scan);
Paul Jackson42253992006-01-08 01:01:59 -08001078
Paul Menage2df167a2008-02-07 00:14:45 -08001079 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
Paul Menage8793d852007-10-18 23:39:39 -07001080 cpuset_being_rebound = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081}
1082
Miao Xie0b2f6302008-07-25 01:47:21 -07001083/*
1084 * Handle user request to change the 'mems' memory placement
1085 * of a cpuset. Needs to validate the request, update the
Miao Xie58568d22009-06-16 15:31:49 -07001086 * cpusets mems_allowed, and for each task in the cpuset,
1087 * update mems_allowed and rebind task's mempolicy and any vma
1088 * mempolicies and if the cpuset is marked 'memory_migrate',
1089 * migrate the tasks pages to the new memory.
Miao Xie0b2f6302008-07-25 01:47:21 -07001090 *
1091 * Call with cgroup_mutex held. May take callback_mutex during call.
1092 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1093 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1094 * their mempolicies to the cpusets new mems_allowed.
1095 */
Li Zefan645fcc92009-01-07 18:08:43 -08001096static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1097 const char *buf)
Miao Xie0b2f6302008-07-25 01:47:21 -07001098{
Miao Xie53feb292010-03-23 13:35:35 -07001099 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
Miao Xie0b2f6302008-07-25 01:47:21 -07001100 int retval;
Li Zefan010cfac2009-04-02 16:57:52 -07001101 struct ptr_heap heap;
Miao Xie0b2f6302008-07-25 01:47:21 -07001102
Miao Xie53feb292010-03-23 13:35:35 -07001103 if (!oldmem)
1104 return -ENOMEM;
1105
Miao Xie0b2f6302008-07-25 01:47:21 -07001106 /*
1107 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1108 * it's read-only
1109 */
Miao Xie53feb292010-03-23 13:35:35 -07001110 if (cs == &top_cpuset) {
1111 retval = -EACCES;
1112 goto done;
1113 }
Miao Xie0b2f6302008-07-25 01:47:21 -07001114
Miao Xie0b2f6302008-07-25 01:47:21 -07001115 /*
1116 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1117 * Since nodelist_parse() fails on an empty mask, we special case
1118 * that parsing. The validate_change() call ensures that cpusets
1119 * with tasks have memory.
1120 */
1121 if (!*buf) {
Li Zefan645fcc92009-01-07 18:08:43 -08001122 nodes_clear(trialcs->mems_allowed);
Miao Xie0b2f6302008-07-25 01:47:21 -07001123 } else {
Li Zefan645fcc92009-01-07 18:08:43 -08001124 retval = nodelist_parse(buf, trialcs->mems_allowed);
Miao Xie0b2f6302008-07-25 01:47:21 -07001125 if (retval < 0)
1126 goto done;
1127
Li Zefan645fcc92009-01-07 18:08:43 -08001128 if (!nodes_subset(trialcs->mems_allowed,
Miao Xie53feb292010-03-23 13:35:35 -07001129 node_states[N_HIGH_MEMORY])) {
1130 retval = -EINVAL;
1131 goto done;
1132 }
Miao Xie0b2f6302008-07-25 01:47:21 -07001133 }
Miao Xie53feb292010-03-23 13:35:35 -07001134 *oldmem = cs->mems_allowed;
1135 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
Miao Xie0b2f6302008-07-25 01:47:21 -07001136 retval = 0; /* Too easy - nothing to do */
1137 goto done;
1138 }
Li Zefan645fcc92009-01-07 18:08:43 -08001139 retval = validate_change(cs, trialcs);
Miao Xie0b2f6302008-07-25 01:47:21 -07001140 if (retval < 0)
1141 goto done;
1142
Li Zefan010cfac2009-04-02 16:57:52 -07001143 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1144 if (retval < 0)
1145 goto done;
1146
Miao Xie0b2f6302008-07-25 01:47:21 -07001147 mutex_lock(&callback_mutex);
Li Zefan645fcc92009-01-07 18:08:43 -08001148 cs->mems_allowed = trialcs->mems_allowed;
Miao Xie0b2f6302008-07-25 01:47:21 -07001149 mutex_unlock(&callback_mutex);
1150
Miao Xie53feb292010-03-23 13:35:35 -07001151 update_tasks_nodemask(cs, oldmem, &heap);
Li Zefan010cfac2009-04-02 16:57:52 -07001152
1153 heap_free(&heap);
Miao Xie0b2f6302008-07-25 01:47:21 -07001154done:
Miao Xie53feb292010-03-23 13:35:35 -07001155 NODEMASK_FREE(oldmem);
Miao Xie0b2f6302008-07-25 01:47:21 -07001156 return retval;
1157}
1158
Paul Menage8793d852007-10-18 23:39:39 -07001159int current_cpuset_is_being_rebound(void)
1160{
1161 return task_cs(current) == cpuset_being_rebound;
1162}
1163
Paul Menage5be7a472008-05-06 20:42:41 -07001164static int update_relax_domain_level(struct cpuset *cs, s64 val)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001165{
Paul Menagedb7f47c2009-04-02 16:57:55 -07001166#ifdef CONFIG_SMP
Li Zefan30e0e172008-05-13 10:27:17 +08001167 if (val < -1 || val >= SD_LV_MAX)
1168 return -EINVAL;
Paul Menagedb7f47c2009-04-02 16:57:55 -07001169#endif
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001170
1171 if (val != cs->relax_domain_level) {
1172 cs->relax_domain_level = val;
Li Zefan300ed6c2009-01-07 18:08:44 -08001173 if (!cpumask_empty(cs->cpus_allowed) &&
1174 is_sched_load_balance(cs))
Max Krasnyanskycf417142008-08-11 14:33:53 -07001175 async_rebuild_sched_domains();
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001176 }
1177
1178 return 0;
1179}
1180
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001181/*
Miao Xie950592f2009-06-16 15:31:47 -07001182 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1183 * @tsk: task to be updated
1184 * @scan: struct cgroup_scanner containing the cgroup of the task
1185 *
1186 * Called by cgroup_scan_tasks() for each task in a cgroup.
1187 *
1188 * We don't need to re-check for the cgroup/cpuset membership, since we're
1189 * holding cgroup_lock() at this point.
1190 */
1191static void cpuset_change_flag(struct task_struct *tsk,
1192 struct cgroup_scanner *scan)
1193{
1194 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1195}
1196
1197/*
1198 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1199 * @cs: the cpuset in which each task's spread flags needs to be changed
1200 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1201 *
1202 * Called with cgroup_mutex held
1203 *
1204 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1205 * calling callback functions for each.
1206 *
1207 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1208 * if @heap != NULL.
1209 */
1210static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1211{
1212 struct cgroup_scanner scan;
1213
1214 scan.cg = cs->css.cgroup;
1215 scan.test_task = NULL;
1216 scan.process_task = cpuset_change_flag;
1217 scan.heap = heap;
1218 cgroup_scan_tasks(&scan);
1219}
1220
1221/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 * update_flag - read a 0 or a 1 in a file and update associated flag
Paul Menage78608362008-04-29 01:00:26 -07001223 * bit: the bit to update (see cpuset_flagbits_t)
1224 * cs: the cpuset to update
1225 * turning_on: whether the flag is being set or cleared
Paul Jackson053199e2005-10-30 15:02:30 -08001226 *
Paul Menage2df167a2008-02-07 00:14:45 -08001227 * Call with cgroup_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 */
1229
Paul Menage700fe1a2008-04-29 01:00:00 -07001230static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1231 int turning_on)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232{
Li Zefan645fcc92009-01-07 18:08:43 -08001233 struct cpuset *trialcs;
Rakib Mullick40b6a762008-10-18 20:28:18 -07001234 int balance_flag_changed;
Miao Xie950592f2009-06-16 15:31:47 -07001235 int spread_flag_changed;
1236 struct ptr_heap heap;
1237 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238
Li Zefan645fcc92009-01-07 18:08:43 -08001239 trialcs = alloc_trial_cpuset(cs);
1240 if (!trialcs)
1241 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242
Li Zefan645fcc92009-01-07 18:08:43 -08001243 if (turning_on)
1244 set_bit(bit, &trialcs->flags);
1245 else
1246 clear_bit(bit, &trialcs->flags);
1247
1248 err = validate_change(cs, trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001249 if (err < 0)
Li Zefan645fcc92009-01-07 18:08:43 -08001250 goto out;
Paul Jackson029190c2007-10-18 23:40:20 -07001251
Miao Xie950592f2009-06-16 15:31:47 -07001252 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1253 if (err < 0)
1254 goto out;
1255
Paul Jackson029190c2007-10-18 23:40:20 -07001256 balance_flag_changed = (is_sched_load_balance(cs) !=
Li Zefan645fcc92009-01-07 18:08:43 -08001257 is_sched_load_balance(trialcs));
Paul Jackson029190c2007-10-18 23:40:20 -07001258
Miao Xie950592f2009-06-16 15:31:47 -07001259 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1260 || (is_spread_page(cs) != is_spread_page(trialcs)));
1261
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001262 mutex_lock(&callback_mutex);
Li Zefan645fcc92009-01-07 18:08:43 -08001263 cs->flags = trialcs->flags;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001264 mutex_unlock(&callback_mutex);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001265
Li Zefan300ed6c2009-01-07 18:08:44 -08001266 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
Max Krasnyanskycf417142008-08-11 14:33:53 -07001267 async_rebuild_sched_domains();
Paul Jackson029190c2007-10-18 23:40:20 -07001268
Miao Xie950592f2009-06-16 15:31:47 -07001269 if (spread_flag_changed)
1270 update_tasks_flags(cs, &heap);
1271 heap_free(&heap);
Li Zefan645fcc92009-01-07 18:08:43 -08001272out:
1273 free_trial_cpuset(trialcs);
1274 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275}
1276
Paul Jackson053199e2005-10-30 15:02:30 -08001277/*
Adrian Bunk80f72282006-06-30 18:27:16 +02001278 * Frequency meter - How fast is some event occurring?
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001279 *
1280 * These routines manage a digitally filtered, constant time based,
1281 * event frequency meter. There are four routines:
1282 * fmeter_init() - initialize a frequency meter.
1283 * fmeter_markevent() - called each time the event happens.
1284 * fmeter_getrate() - returns the recent rate of such events.
1285 * fmeter_update() - internal routine used to update fmeter.
1286 *
1287 * A common data structure is passed to each of these routines,
1288 * which is used to keep track of the state required to manage the
1289 * frequency meter and its digital filter.
1290 *
1291 * The filter works on the number of events marked per unit time.
1292 * The filter is single-pole low-pass recursive (IIR). The time unit
1293 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1294 * simulate 3 decimal digits of precision (multiplied by 1000).
1295 *
1296 * With an FM_COEF of 933, and a time base of 1 second, the filter
1297 * has a half-life of 10 seconds, meaning that if the events quit
1298 * happening, then the rate returned from the fmeter_getrate()
1299 * will be cut in half each 10 seconds, until it converges to zero.
1300 *
1301 * It is not worth doing a real infinitely recursive filter. If more
1302 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1303 * just compute FM_MAXTICKS ticks worth, by which point the level
1304 * will be stable.
1305 *
1306 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1307 * arithmetic overflow in the fmeter_update() routine.
1308 *
1309 * Given the simple 32 bit integer arithmetic used, this meter works
1310 * best for reporting rates between one per millisecond (msec) and
1311 * one per 32 (approx) seconds. At constant rates faster than one
1312 * per msec it maxes out at values just under 1,000,000. At constant
1313 * rates between one per msec, and one per second it will stabilize
1314 * to a value N*1000, where N is the rate of events per second.
1315 * At constant rates between one per second and one per 32 seconds,
1316 * it will be choppy, moving up on the seconds that have an event,
1317 * and then decaying until the next event. At rates slower than
1318 * about one in 32 seconds, it decays all the way back to zero between
1319 * each event.
1320 */
1321
1322#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1323#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1324#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1325#define FM_SCALE 1000 /* faux fixed point scale */
1326
1327/* Initialize a frequency meter */
1328static void fmeter_init(struct fmeter *fmp)
1329{
1330 fmp->cnt = 0;
1331 fmp->val = 0;
1332 fmp->time = 0;
1333 spin_lock_init(&fmp->lock);
1334}
1335
1336/* Internal meter update - process cnt events and update value */
1337static void fmeter_update(struct fmeter *fmp)
1338{
1339 time_t now = get_seconds();
1340 time_t ticks = now - fmp->time;
1341
1342 if (ticks == 0)
1343 return;
1344
1345 ticks = min(FM_MAXTICKS, ticks);
1346 while (ticks-- > 0)
1347 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1348 fmp->time = now;
1349
1350 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1351 fmp->cnt = 0;
1352}
1353
1354/* Process any previous ticks, then bump cnt by one (times scale). */
1355static void fmeter_markevent(struct fmeter *fmp)
1356{
1357 spin_lock(&fmp->lock);
1358 fmeter_update(fmp);
1359 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1360 spin_unlock(&fmp->lock);
1361}
1362
1363/* Process any previous ticks, then return current value. */
1364static int fmeter_getrate(struct fmeter *fmp)
1365{
1366 int val;
1367
1368 spin_lock(&fmp->lock);
1369 fmeter_update(fmp);
1370 val = fmp->val;
1371 spin_unlock(&fmp->lock);
1372 return val;
1373}
1374
Li Zefan2341d1b2009-01-07 18:08:42 -08001375/* Protected by cgroup_lock */
1376static cpumask_var_t cpus_attach;
1377
Paul Menage2df167a2008-02-07 00:14:45 -08001378/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
Ben Blumbe367d02009-09-23 15:56:31 -07001379static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1380 struct task_struct *tsk, bool threadgroup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381{
Ben Blumbe367d02009-09-23 15:56:31 -07001382 int ret;
Paul Menage8793d852007-10-18 23:39:39 -07001383 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
Li Zefan300ed6c2009-01-07 18:08:44 -08001385 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 return -ENOSPC;
David Rientjes9985b0b2008-06-05 12:57:11 -07001387
David Rientjes6d7b2f52009-04-02 16:57:57 -07001388 /*
1389 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1390 * cannot change their cpu affinity and isolating such threads by their
1391 * set of allowed nodes is unnecessary. Thus, cpusets are not
1392 * applicable for such threads. This prevents checking for success of
1393 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1394 * be changed.
1395 */
1396 if (tsk->flags & PF_THREAD_BOUND)
1397 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398
KOSAKI Motohirob0ae1982010-10-15 04:21:18 +09001399 ret = security_task_setscheduler(tsk);
Ben Blumbe367d02009-09-23 15:56:31 -07001400 if (ret)
1401 return ret;
1402 if (threadgroup) {
1403 struct task_struct *c;
1404
1405 rcu_read_lock();
1406 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
KOSAKI Motohirob0ae1982010-10-15 04:21:18 +09001407 ret = security_task_setscheduler(c);
Ben Blumbe367d02009-09-23 15:56:31 -07001408 if (ret) {
1409 rcu_read_unlock();
1410 return ret;
1411 }
1412 }
1413 rcu_read_unlock();
1414 }
1415 return 0;
Paul Menage8793d852007-10-18 23:39:39 -07001416}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417
Ben Blumbe367d02009-09-23 15:56:31 -07001418static void cpuset_attach_task(struct task_struct *tsk, nodemask_t *to,
1419 struct cpuset *cs)
1420{
1421 int err;
1422 /*
1423 * can_attach beforehand should guarantee that this doesn't fail.
1424 * TODO: have a better way to handle failure here
1425 */
1426 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1427 WARN_ON_ONCE(err);
1428
Ben Blumbe367d02009-09-23 15:56:31 -07001429 cpuset_change_task_nodemask(tsk, to);
Ben Blumbe367d02009-09-23 15:56:31 -07001430 cpuset_update_task_spread_flag(cs, tsk);
1431
1432}
1433
1434static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1435 struct cgroup *oldcont, struct task_struct *tsk,
1436 bool threadgroup)
Paul Menage8793d852007-10-18 23:39:39 -07001437{
Paul Menage8793d852007-10-18 23:39:39 -07001438 struct mm_struct *mm;
1439 struct cpuset *cs = cgroup_cs(cont);
1440 struct cpuset *oldcs = cgroup_cs(oldcont);
Miao Xie53feb292010-03-23 13:35:35 -07001441 NODEMASK_ALLOC(nodemask_t, from, GFP_KERNEL);
1442 NODEMASK_ALLOC(nodemask_t, to, GFP_KERNEL);
1443
1444 if (from == NULL || to == NULL)
1445 goto alloc_fail;
David Quigley22fb52d2006-06-23 02:04:00 -07001446
Miao Xief5813d92009-01-07 18:08:40 -08001447 if (cs == &top_cpuset) {
Li Zefan2341d1b2009-01-07 18:08:42 -08001448 cpumask_copy(cpus_attach, cpu_possible_mask);
Miao Xief5813d92009-01-07 18:08:40 -08001449 } else {
Li Zefan2341d1b2009-01-07 18:08:42 -08001450 guarantee_online_cpus(cs, cpus_attach);
Miao Xief5813d92009-01-07 18:08:40 -08001451 }
Miao Xie53feb292010-03-23 13:35:35 -07001452 guarantee_online_mems(cs, to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453
Ben Blumbe367d02009-09-23 15:56:31 -07001454 /* do per-task migration stuff possibly for each in the threadgroup */
Miao Xie53feb292010-03-23 13:35:35 -07001455 cpuset_attach_task(tsk, to, cs);
Ben Blumbe367d02009-09-23 15:56:31 -07001456 if (threadgroup) {
1457 struct task_struct *c;
1458 rcu_read_lock();
1459 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
Miao Xie53feb292010-03-23 13:35:35 -07001460 cpuset_attach_task(c, to, cs);
Ben Blumbe367d02009-09-23 15:56:31 -07001461 }
1462 rcu_read_unlock();
1463 }
Miao Xie950592f2009-06-16 15:31:47 -07001464
Ben Blumbe367d02009-09-23 15:56:31 -07001465 /* change mm; only needs to be done once even if threadgroup */
Miao Xie53feb292010-03-23 13:35:35 -07001466 *from = oldcs->mems_allowed;
1467 *to = cs->mems_allowed;
Paul Jackson42253992006-01-08 01:01:59 -08001468 mm = get_task_mm(tsk);
1469 if (mm) {
Miao Xie53feb292010-03-23 13:35:35 -07001470 mpol_rebind_mm(mm, to);
Paul Jackson2741a552006-03-31 02:30:51 -08001471 if (is_memory_migrate(cs))
Miao Xie53feb292010-03-23 13:35:35 -07001472 cpuset_migrate_mm(mm, from, to);
Paul Jackson42253992006-01-08 01:01:59 -08001473 mmput(mm);
1474 }
Miao Xie53feb292010-03-23 13:35:35 -07001475
1476alloc_fail:
1477 NODEMASK_FREE(from);
1478 NODEMASK_FREE(to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479}
1480
1481/* The various types of files and directories in a cpuset file system */
1482
1483typedef enum {
Paul Jackson45b07ef2006-01-08 01:00:56 -08001484 FILE_MEMORY_MIGRATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 FILE_CPULIST,
1486 FILE_MEMLIST,
1487 FILE_CPU_EXCLUSIVE,
1488 FILE_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -07001489 FILE_MEM_HARDWALL,
Paul Jackson029190c2007-10-18 23:40:20 -07001490 FILE_SCHED_LOAD_BALANCE,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001491 FILE_SCHED_RELAX_DOMAIN_LEVEL,
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001492 FILE_MEMORY_PRESSURE_ENABLED,
1493 FILE_MEMORY_PRESSURE,
Paul Jackson825a46a2006-03-24 03:16:03 -08001494 FILE_SPREAD_PAGE,
1495 FILE_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496} cpuset_filetype_t;
1497
Paul Menage700fe1a2008-04-29 01:00:00 -07001498static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1499{
1500 int retval = 0;
1501 struct cpuset *cs = cgroup_cs(cgrp);
1502 cpuset_filetype_t type = cft->private;
1503
Paul Menagee3712392008-07-25 01:47:02 -07001504 if (!cgroup_lock_live_group(cgrp))
Paul Menage700fe1a2008-04-29 01:00:00 -07001505 return -ENODEV;
Paul Menage700fe1a2008-04-29 01:00:00 -07001506
1507 switch (type) {
1508 case FILE_CPU_EXCLUSIVE:
1509 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1510 break;
1511 case FILE_MEM_EXCLUSIVE:
1512 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1513 break;
Paul Menage78608362008-04-29 01:00:26 -07001514 case FILE_MEM_HARDWALL:
1515 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1516 break;
Paul Menage700fe1a2008-04-29 01:00:00 -07001517 case FILE_SCHED_LOAD_BALANCE:
1518 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1519 break;
1520 case FILE_MEMORY_MIGRATE:
1521 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1522 break;
1523 case FILE_MEMORY_PRESSURE_ENABLED:
1524 cpuset_memory_pressure_enabled = !!val;
1525 break;
1526 case FILE_MEMORY_PRESSURE:
1527 retval = -EACCES;
1528 break;
1529 case FILE_SPREAD_PAGE:
1530 retval = update_flag(CS_SPREAD_PAGE, cs, val);
Paul Menage700fe1a2008-04-29 01:00:00 -07001531 break;
1532 case FILE_SPREAD_SLAB:
1533 retval = update_flag(CS_SPREAD_SLAB, cs, val);
Paul Menage700fe1a2008-04-29 01:00:00 -07001534 break;
1535 default:
1536 retval = -EINVAL;
1537 break;
1538 }
1539 cgroup_unlock();
1540 return retval;
1541}
1542
Paul Menage5be7a472008-05-06 20:42:41 -07001543static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1544{
1545 int retval = 0;
1546 struct cpuset *cs = cgroup_cs(cgrp);
1547 cpuset_filetype_t type = cft->private;
1548
Paul Menagee3712392008-07-25 01:47:02 -07001549 if (!cgroup_lock_live_group(cgrp))
Paul Menage5be7a472008-05-06 20:42:41 -07001550 return -ENODEV;
Paul Menagee3712392008-07-25 01:47:02 -07001551
Paul Menage5be7a472008-05-06 20:42:41 -07001552 switch (type) {
1553 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1554 retval = update_relax_domain_level(cs, val);
1555 break;
1556 default:
1557 retval = -EINVAL;
1558 break;
1559 }
1560 cgroup_unlock();
1561 return retval;
1562}
1563
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564/*
Paul Menagee3712392008-07-25 01:47:02 -07001565 * Common handling for a write to a "cpus" or "mems" file.
1566 */
1567static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1568 const char *buf)
1569{
1570 int retval = 0;
Li Zefan645fcc92009-01-07 18:08:43 -08001571 struct cpuset *cs = cgroup_cs(cgrp);
1572 struct cpuset *trialcs;
Paul Menagee3712392008-07-25 01:47:02 -07001573
1574 if (!cgroup_lock_live_group(cgrp))
1575 return -ENODEV;
1576
Li Zefan645fcc92009-01-07 18:08:43 -08001577 trialcs = alloc_trial_cpuset(cs);
Li Zefanb75f38d2011-03-04 17:36:21 -08001578 if (!trialcs) {
1579 retval = -ENOMEM;
1580 goto out;
1581 }
Li Zefan645fcc92009-01-07 18:08:43 -08001582
Paul Menagee3712392008-07-25 01:47:02 -07001583 switch (cft->private) {
1584 case FILE_CPULIST:
Li Zefan645fcc92009-01-07 18:08:43 -08001585 retval = update_cpumask(cs, trialcs, buf);
Paul Menagee3712392008-07-25 01:47:02 -07001586 break;
1587 case FILE_MEMLIST:
Li Zefan645fcc92009-01-07 18:08:43 -08001588 retval = update_nodemask(cs, trialcs, buf);
Paul Menagee3712392008-07-25 01:47:02 -07001589 break;
1590 default:
1591 retval = -EINVAL;
1592 break;
1593 }
Li Zefan645fcc92009-01-07 18:08:43 -08001594
1595 free_trial_cpuset(trialcs);
Li Zefanb75f38d2011-03-04 17:36:21 -08001596out:
Paul Menagee3712392008-07-25 01:47:02 -07001597 cgroup_unlock();
1598 return retval;
1599}
1600
1601/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 * These ascii lists should be read in a single call, by using a user
1603 * buffer large enough to hold the entire map. If read in smaller
1604 * chunks, there is no guarantee of atomicity. Since the display format
1605 * used, list of ranges of sequential numbers, is variable length,
1606 * and since these maps can change value dynamically, one could read
1607 * gibberish by doing partial reads while a list was changing.
1608 * A single large read to a buffer that crosses a page boundary is
1609 * ok, because the result being copied to user land is not recomputed
1610 * across a page fault.
1611 */
1612
Li Zefan9303e0c2011-03-23 16:42:45 -07001613static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614{
Li Zefan9303e0c2011-03-23 16:42:45 -07001615 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001617 mutex_lock(&callback_mutex);
Li Zefan9303e0c2011-03-23 16:42:45 -07001618 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001619 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620
Li Zefan9303e0c2011-03-23 16:42:45 -07001621 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622}
1623
Li Zefan9303e0c2011-03-23 16:42:45 -07001624static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625{
Li Zefan9303e0c2011-03-23 16:42:45 -07001626 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001628 mutex_lock(&callback_mutex);
Li Zefan9303e0c2011-03-23 16:42:45 -07001629 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001630 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631
Li Zefan9303e0c2011-03-23 16:42:45 -07001632 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633}
1634
Paul Menage8793d852007-10-18 23:39:39 -07001635static ssize_t cpuset_common_file_read(struct cgroup *cont,
1636 struct cftype *cft,
1637 struct file *file,
1638 char __user *buf,
1639 size_t nbytes, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640{
Paul Menage8793d852007-10-18 23:39:39 -07001641 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 cpuset_filetype_t type = cft->private;
1643 char *page;
1644 ssize_t retval = 0;
1645 char *s;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646
Mel Gormane12ba742007-10-16 01:25:52 -07001647 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 return -ENOMEM;
1649
1650 s = page;
1651
1652 switch (type) {
1653 case FILE_CPULIST:
1654 s += cpuset_sprintf_cpulist(s, cs);
1655 break;
1656 case FILE_MEMLIST:
1657 s += cpuset_sprintf_memlist(s, cs);
1658 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 default:
1660 retval = -EINVAL;
1661 goto out;
1662 }
1663 *s++ = '\n';
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664
Al Viroeacaa1f2005-09-30 03:26:43 +01001665 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666out:
1667 free_page((unsigned long)page);
1668 return retval;
1669}
1670
Paul Menage700fe1a2008-04-29 01:00:00 -07001671static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1672{
1673 struct cpuset *cs = cgroup_cs(cont);
1674 cpuset_filetype_t type = cft->private;
1675 switch (type) {
1676 case FILE_CPU_EXCLUSIVE:
1677 return is_cpu_exclusive(cs);
1678 case FILE_MEM_EXCLUSIVE:
1679 return is_mem_exclusive(cs);
Paul Menage78608362008-04-29 01:00:26 -07001680 case FILE_MEM_HARDWALL:
1681 return is_mem_hardwall(cs);
Paul Menage700fe1a2008-04-29 01:00:00 -07001682 case FILE_SCHED_LOAD_BALANCE:
1683 return is_sched_load_balance(cs);
1684 case FILE_MEMORY_MIGRATE:
1685 return is_memory_migrate(cs);
1686 case FILE_MEMORY_PRESSURE_ENABLED:
1687 return cpuset_memory_pressure_enabled;
1688 case FILE_MEMORY_PRESSURE:
1689 return fmeter_getrate(&cs->fmeter);
1690 case FILE_SPREAD_PAGE:
1691 return is_spread_page(cs);
1692 case FILE_SPREAD_SLAB:
1693 return is_spread_slab(cs);
1694 default:
1695 BUG();
1696 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07001697
1698 /* Unreachable but makes gcc happy */
1699 return 0;
Paul Menage700fe1a2008-04-29 01:00:00 -07001700}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701
Paul Menage5be7a472008-05-06 20:42:41 -07001702static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1703{
1704 struct cpuset *cs = cgroup_cs(cont);
1705 cpuset_filetype_t type = cft->private;
1706 switch (type) {
1707 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1708 return cs->relax_domain_level;
1709 default:
1710 BUG();
1711 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07001712
1713 /* Unrechable but makes gcc happy */
1714 return 0;
Paul Menage5be7a472008-05-06 20:42:41 -07001715}
1716
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717
1718/*
1719 * for the common functions, 'private' gives the type of file
1720 */
1721
Paul Menageaddf2c72008-04-29 01:00:26 -07001722static struct cftype files[] = {
1723 {
1724 .name = "cpus",
1725 .read = cpuset_common_file_read,
Paul Menagee3712392008-07-25 01:47:02 -07001726 .write_string = cpuset_write_resmask,
1727 .max_write_len = (100U + 6 * NR_CPUS),
Paul Menageaddf2c72008-04-29 01:00:26 -07001728 .private = FILE_CPULIST,
1729 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730
Paul Menageaddf2c72008-04-29 01:00:26 -07001731 {
1732 .name = "mems",
1733 .read = cpuset_common_file_read,
Paul Menagee3712392008-07-25 01:47:02 -07001734 .write_string = cpuset_write_resmask,
1735 .max_write_len = (100U + 6 * MAX_NUMNODES),
Paul Menageaddf2c72008-04-29 01:00:26 -07001736 .private = FILE_MEMLIST,
1737 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738
Paul Menageaddf2c72008-04-29 01:00:26 -07001739 {
1740 .name = "cpu_exclusive",
1741 .read_u64 = cpuset_read_u64,
1742 .write_u64 = cpuset_write_u64,
1743 .private = FILE_CPU_EXCLUSIVE,
1744 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745
Paul Menageaddf2c72008-04-29 01:00:26 -07001746 {
1747 .name = "mem_exclusive",
1748 .read_u64 = cpuset_read_u64,
1749 .write_u64 = cpuset_write_u64,
1750 .private = FILE_MEM_EXCLUSIVE,
1751 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Paul Menageaddf2c72008-04-29 01:00:26 -07001753 {
Paul Menage78608362008-04-29 01:00:26 -07001754 .name = "mem_hardwall",
1755 .read_u64 = cpuset_read_u64,
1756 .write_u64 = cpuset_write_u64,
1757 .private = FILE_MEM_HARDWALL,
1758 },
1759
1760 {
Paul Menageaddf2c72008-04-29 01:00:26 -07001761 .name = "sched_load_balance",
1762 .read_u64 = cpuset_read_u64,
1763 .write_u64 = cpuset_write_u64,
1764 .private = FILE_SCHED_LOAD_BALANCE,
1765 },
Paul Jackson029190c2007-10-18 23:40:20 -07001766
Paul Menageaddf2c72008-04-29 01:00:26 -07001767 {
1768 .name = "sched_relax_domain_level",
Paul Menage5be7a472008-05-06 20:42:41 -07001769 .read_s64 = cpuset_read_s64,
1770 .write_s64 = cpuset_write_s64,
Paul Menageaddf2c72008-04-29 01:00:26 -07001771 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1772 },
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001773
Paul Menageaddf2c72008-04-29 01:00:26 -07001774 {
1775 .name = "memory_migrate",
1776 .read_u64 = cpuset_read_u64,
1777 .write_u64 = cpuset_write_u64,
1778 .private = FILE_MEMORY_MIGRATE,
1779 },
1780
1781 {
1782 .name = "memory_pressure",
1783 .read_u64 = cpuset_read_u64,
1784 .write_u64 = cpuset_write_u64,
1785 .private = FILE_MEMORY_PRESSURE,
Li Zefan099fca32009-04-02 16:57:29 -07001786 .mode = S_IRUGO,
Paul Menageaddf2c72008-04-29 01:00:26 -07001787 },
1788
1789 {
1790 .name = "memory_spread_page",
1791 .read_u64 = cpuset_read_u64,
1792 .write_u64 = cpuset_write_u64,
1793 .private = FILE_SPREAD_PAGE,
1794 },
1795
1796 {
1797 .name = "memory_spread_slab",
1798 .read_u64 = cpuset_read_u64,
1799 .write_u64 = cpuset_write_u64,
1800 .private = FILE_SPREAD_SLAB,
1801 },
Paul Jackson45b07ef2006-01-08 01:00:56 -08001802};
1803
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001804static struct cftype cft_memory_pressure_enabled = {
1805 .name = "memory_pressure_enabled",
Paul Menage700fe1a2008-04-29 01:00:00 -07001806 .read_u64 = cpuset_read_u64,
1807 .write_u64 = cpuset_write_u64,
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001808 .private = FILE_MEMORY_PRESSURE_ENABLED,
1809};
1810
Paul Menage8793d852007-10-18 23:39:39 -07001811static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812{
1813 int err;
1814
Paul Menageaddf2c72008-04-29 01:00:26 -07001815 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1816 if (err)
Paul Jackson825a46a2006-03-24 03:16:03 -08001817 return err;
Paul Menage8793d852007-10-18 23:39:39 -07001818 /* memory_pressure_enabled is in root cpuset only */
Paul Menageaddf2c72008-04-29 01:00:26 -07001819 if (!cont->parent)
Paul Menage8793d852007-10-18 23:39:39 -07001820 err = cgroup_add_file(cont, ss,
Paul Menageaddf2c72008-04-29 01:00:26 -07001821 &cft_memory_pressure_enabled);
1822 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823}
1824
1825/*
Paul Menage8793d852007-10-18 23:39:39 -07001826 * post_clone() is called at the end of cgroup_clone().
1827 * 'cgroup' was just created automatically as a result of
1828 * a cgroup_clone(), and the current task is about to
1829 * be moved into 'cgroup'.
1830 *
1831 * Currently we refuse to set up the cgroup - thereby
1832 * refusing the task to be entered, and as a result refusing
1833 * the sys_unshare() or clone() which initiated it - if any
1834 * sibling cpusets have exclusive cpus or mem.
1835 *
1836 * If this becomes a problem for some users who wish to
1837 * allow that scenario, then cpuset_post_clone() could be
1838 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
Paul Menage2df167a2008-02-07 00:14:45 -08001839 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1840 * held.
Paul Menage8793d852007-10-18 23:39:39 -07001841 */
1842static void cpuset_post_clone(struct cgroup_subsys *ss,
1843 struct cgroup *cgroup)
1844{
1845 struct cgroup *parent, *child;
1846 struct cpuset *cs, *parent_cs;
1847
1848 parent = cgroup->parent;
1849 list_for_each_entry(child, &parent->children, sibling) {
1850 cs = cgroup_cs(child);
1851 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1852 return;
1853 }
1854 cs = cgroup_cs(cgroup);
1855 parent_cs = cgroup_cs(parent);
1856
1857 cs->mems_allowed = parent_cs->mems_allowed;
Li Zefan300ed6c2009-01-07 18:08:44 -08001858 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
Paul Menage8793d852007-10-18 23:39:39 -07001859 return;
1860}
1861
1862/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 * cpuset_create - create a cpuset
Paul Menage2df167a2008-02-07 00:14:45 -08001864 * ss: cpuset cgroup subsystem
1865 * cont: control group that the new cpuset will be part of
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 */
1867
Paul Menage8793d852007-10-18 23:39:39 -07001868static struct cgroup_subsys_state *cpuset_create(
1869 struct cgroup_subsys *ss,
1870 struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871{
1872 struct cpuset *cs;
Paul Menage8793d852007-10-18 23:39:39 -07001873 struct cpuset *parent;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874
Paul Menage8793d852007-10-18 23:39:39 -07001875 if (!cont->parent) {
Paul Menage8793d852007-10-18 23:39:39 -07001876 return &top_cpuset.css;
1877 }
1878 parent = cgroup_cs(cont->parent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1880 if (!cs)
Paul Menage8793d852007-10-18 23:39:39 -07001881 return ERR_PTR(-ENOMEM);
Li Zefan300ed6c2009-01-07 18:08:44 -08001882 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1883 kfree(cs);
1884 return ERR_PTR(-ENOMEM);
1885 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887 cs->flags = 0;
Paul Jackson825a46a2006-03-24 03:16:03 -08001888 if (is_spread_page(parent))
1889 set_bit(CS_SPREAD_PAGE, &cs->flags);
1890 if (is_spread_slab(parent))
1891 set_bit(CS_SPREAD_SLAB, &cs->flags);
Paul Jackson029190c2007-10-18 23:40:20 -07001892 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
Li Zefan300ed6c2009-01-07 18:08:44 -08001893 cpumask_clear(cs->cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001894 nodes_clear(cs->mems_allowed);
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001895 fmeter_init(&cs->fmeter);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001896 cs->relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897
1898 cs->parent = parent;
Paul Jackson202f72d2006-01-08 01:01:57 -08001899 number_of_cpusets++;
Paul Menage8793d852007-10-18 23:39:39 -07001900 return &cs->css ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901}
1902
Paul Jackson029190c2007-10-18 23:40:20 -07001903/*
Paul Jackson029190c2007-10-18 23:40:20 -07001904 * If the cpuset being removed has its flag 'sched_load_balance'
1905 * enabled, then simulate turning sched_load_balance off, which
Max Krasnyanskycf417142008-08-11 14:33:53 -07001906 * will call async_rebuild_sched_domains().
Paul Jackson029190c2007-10-18 23:40:20 -07001907 */
1908
Paul Menage8793d852007-10-18 23:39:39 -07001909static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910{
Paul Menage8793d852007-10-18 23:39:39 -07001911 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912
Paul Jackson029190c2007-10-18 23:40:20 -07001913 if (is_sched_load_balance(cs))
Paul Menage700fe1a2008-04-29 01:00:00 -07001914 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
Paul Jackson029190c2007-10-18 23:40:20 -07001915
Paul Jackson202f72d2006-01-08 01:01:57 -08001916 number_of_cpusets--;
Li Zefan300ed6c2009-01-07 18:08:44 -08001917 free_cpumask_var(cs->cpus_allowed);
Paul Menage8793d852007-10-18 23:39:39 -07001918 kfree(cs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919}
1920
Paul Menage8793d852007-10-18 23:39:39 -07001921struct cgroup_subsys cpuset_subsys = {
1922 .name = "cpuset",
1923 .create = cpuset_create,
Max Krasnyanskycf417142008-08-11 14:33:53 -07001924 .destroy = cpuset_destroy,
Paul Menage8793d852007-10-18 23:39:39 -07001925 .can_attach = cpuset_can_attach,
1926 .attach = cpuset_attach,
1927 .populate = cpuset_populate,
1928 .post_clone = cpuset_post_clone,
1929 .subsys_id = cpuset_subsys_id,
1930 .early_init = 1,
1931};
1932
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933/**
1934 * cpuset_init - initialize cpusets at system boot
1935 *
1936 * Description: Initialize top_cpuset and the cpuset internal file system,
1937 **/
1938
1939int __init cpuset_init(void)
1940{
Paul Menage8793d852007-10-18 23:39:39 -07001941 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942
Miao Xie58568d22009-06-16 15:31:49 -07001943 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1944 BUG();
1945
Li Zefan300ed6c2009-01-07 18:08:44 -08001946 cpumask_setall(top_cpuset.cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001947 nodes_setall(top_cpuset.mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001949 fmeter_init(&top_cpuset.fmeter);
Paul Jackson029190c2007-10-18 23:40:20 -07001950 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001951 top_cpuset.relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 err = register_filesystem(&cpuset_fs_type);
1954 if (err < 0)
Paul Menage8793d852007-10-18 23:39:39 -07001955 return err;
1956
Li Zefan2341d1b2009-01-07 18:08:42 -08001957 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1958 BUG();
1959
Paul Jackson202f72d2006-01-08 01:01:57 -08001960 number_of_cpusets = 1;
Paul Menage8793d852007-10-18 23:39:39 -07001961 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962}
1963
Cliff Wickman956db3c2008-02-07 00:14:43 -08001964/**
1965 * cpuset_do_move_task - move a given task to another cpuset
1966 * @tsk: pointer to task_struct the task to move
1967 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1968 *
1969 * Called by cgroup_scan_tasks() for each task in a cgroup.
1970 * Return nonzero to stop the walk through the tasks.
1971 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -07001972static void cpuset_do_move_task(struct task_struct *tsk,
1973 struct cgroup_scanner *scan)
Cliff Wickman956db3c2008-02-07 00:14:43 -08001974{
Li Zefan7f81b1a2009-04-02 16:57:53 -07001975 struct cgroup *new_cgroup = scan->data;
Cliff Wickman956db3c2008-02-07 00:14:43 -08001976
Li Zefan7f81b1a2009-04-02 16:57:53 -07001977 cgroup_attach_task(new_cgroup, tsk);
Cliff Wickman956db3c2008-02-07 00:14:43 -08001978}
1979
1980/**
1981 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1982 * @from: cpuset in which the tasks currently reside
1983 * @to: cpuset to which the tasks will be moved
1984 *
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001985 * Called with cgroup_mutex held
1986 * callback_mutex must not be held, as cpuset_attach() will take it.
Cliff Wickman956db3c2008-02-07 00:14:43 -08001987 *
1988 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1989 * calling callback functions for each.
1990 */
1991static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1992{
Li Zefan7f81b1a2009-04-02 16:57:53 -07001993 struct cgroup_scanner scan;
Cliff Wickman956db3c2008-02-07 00:14:43 -08001994
Li Zefan7f81b1a2009-04-02 16:57:53 -07001995 scan.cg = from->css.cgroup;
1996 scan.test_task = NULL; /* select all tasks in cgroup */
1997 scan.process_task = cpuset_do_move_task;
1998 scan.heap = NULL;
1999 scan.data = to->css.cgroup;
Cliff Wickman956db3c2008-02-07 00:14:43 -08002000
Li Zefan7f81b1a2009-04-02 16:57:53 -07002001 if (cgroup_scan_tasks(&scan))
Cliff Wickman956db3c2008-02-07 00:14:43 -08002002 printk(KERN_ERR "move_member_tasks_to_cpuset: "
2003 "cgroup_scan_tasks failed\n");
2004}
2005
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002006/*
Max Krasnyanskycf417142008-08-11 14:33:53 -07002007 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002008 * or memory nodes, we need to walk over the cpuset hierarchy,
2009 * removing that CPU or node from all cpusets. If this removes the
Cliff Wickman956db3c2008-02-07 00:14:43 -08002010 * last CPU or node from a cpuset, then move the tasks in the empty
2011 * cpuset to its next-highest non-empty parent.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002012 *
Paul Jacksonc8d9c902008-02-07 00:14:46 -08002013 * Called with cgroup_mutex held
2014 * callback_mutex must not be held, as cpuset_attach() will take it.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002015 */
Cliff Wickman956db3c2008-02-07 00:14:43 -08002016static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002017{
Cliff Wickman956db3c2008-02-07 00:14:43 -08002018 struct cpuset *parent;
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002019
Paul Jacksonc8d9c902008-02-07 00:14:46 -08002020 /*
2021 * The cgroup's css_sets list is in use if there are tasks
2022 * in the cpuset; the list is empty if there are none;
2023 * the cs->css.refcnt seems always 0.
2024 */
Cliff Wickman956db3c2008-02-07 00:14:43 -08002025 if (list_empty(&cs->css.cgroup->css_sets))
2026 return;
2027
2028 /*
2029 * Find its next-highest non-empty parent, (top cpuset
2030 * has online cpus, so can't be empty).
2031 */
2032 parent = cs->parent;
Li Zefan300ed6c2009-01-07 18:08:44 -08002033 while (cpumask_empty(parent->cpus_allowed) ||
Paul Jacksonb4501292008-02-07 00:14:47 -08002034 nodes_empty(parent->mems_allowed))
Cliff Wickman956db3c2008-02-07 00:14:43 -08002035 parent = parent->parent;
Cliff Wickman956db3c2008-02-07 00:14:43 -08002036
2037 move_member_tasks_to_cpuset(cs, parent);
2038}
2039
2040/*
2041 * Walk the specified cpuset subtree and look for empty cpusets.
2042 * The tasks of such cpuset must be moved to a parent cpuset.
2043 *
Paul Menage2df167a2008-02-07 00:14:45 -08002044 * Called with cgroup_mutex held. We take callback_mutex to modify
Cliff Wickman956db3c2008-02-07 00:14:43 -08002045 * cpus_allowed and mems_allowed.
2046 *
2047 * This walk processes the tree from top to bottom, completing one layer
2048 * before dropping down to the next. It always processes a node before
2049 * any of its children.
2050 *
2051 * For now, since we lack memory hot unplug, we'll never see a cpuset
2052 * that has tasks along with an empty 'mems'. But if we did see such
2053 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2054 */
Frederic Weisbeckerd294eb82008-10-03 12:10:10 +02002055static void scan_for_empty_cpusets(struct cpuset *root)
Cliff Wickman956db3c2008-02-07 00:14:43 -08002056{
Li Zefan8d1e6262008-07-29 22:33:21 -07002057 LIST_HEAD(queue);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002058 struct cpuset *cp; /* scans cpusets being updated */
2059 struct cpuset *child; /* scans child cpusets of cp */
Cliff Wickman956db3c2008-02-07 00:14:43 -08002060 struct cgroup *cont;
Miao Xie53feb292010-03-23 13:35:35 -07002061 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2062
2063 if (oldmems == NULL)
2064 return;
Cliff Wickman956db3c2008-02-07 00:14:43 -08002065
Cliff Wickman956db3c2008-02-07 00:14:43 -08002066 list_add_tail((struct list_head *)&root->stack_list, &queue);
2067
Cliff Wickman956db3c2008-02-07 00:14:43 -08002068 while (!list_empty(&queue)) {
Li Zefan8d1e6262008-07-29 22:33:21 -07002069 cp = list_first_entry(&queue, struct cpuset, stack_list);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002070 list_del(queue.next);
2071 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2072 child = cgroup_cs(cont);
2073 list_add_tail(&child->stack_list, &queue);
2074 }
Paul Jacksonb4501292008-02-07 00:14:47 -08002075
2076 /* Continue past cpusets with all cpus, mems online */
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01002077 if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
Paul Jacksonb4501292008-02-07 00:14:47 -08002078 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2079 continue;
2080
Miao Xie53feb292010-03-23 13:35:35 -07002081 *oldmems = cp->mems_allowed;
Miao Xief9b4fb82008-07-25 01:47:22 -07002082
Cliff Wickman956db3c2008-02-07 00:14:43 -08002083 /* Remove offline cpus and mems from this cpuset. */
Paul Jacksonb4501292008-02-07 00:14:47 -08002084 mutex_lock(&callback_mutex);
Li Zefan300ed6c2009-01-07 18:08:44 -08002085 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01002086 cpu_active_mask);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002087 nodes_and(cp->mems_allowed, cp->mems_allowed,
2088 node_states[N_HIGH_MEMORY]);
Paul Jacksonb4501292008-02-07 00:14:47 -08002089 mutex_unlock(&callback_mutex);
2090
2091 /* Move tasks from the empty cpuset to a parent */
Li Zefan300ed6c2009-01-07 18:08:44 -08002092 if (cpumask_empty(cp->cpus_allowed) ||
Paul Jacksonb4501292008-02-07 00:14:47 -08002093 nodes_empty(cp->mems_allowed))
Cliff Wickman956db3c2008-02-07 00:14:43 -08002094 remove_tasks_in_empty_cpuset(cp);
Miao Xief9b4fb82008-07-25 01:47:22 -07002095 else {
Li Zefan4e743392008-09-13 02:33:08 -07002096 update_tasks_cpumask(cp, NULL);
Miao Xie53feb292010-03-23 13:35:35 -07002097 update_tasks_nodemask(cp, oldmems, NULL);
Miao Xief9b4fb82008-07-25 01:47:22 -07002098 }
Cliff Wickman956db3c2008-02-07 00:14:43 -08002099 }
Miao Xie53feb292010-03-23 13:35:35 -07002100 NODEMASK_FREE(oldmems);
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002101}
2102
2103/*
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002104 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2105 * period. This is necessary in order to make cpusets transparent
2106 * (of no affect) on systems that are actively using CPU hotplug
2107 * but making no active use of cpusets.
2108 *
Paul Jackson38837fc2006-09-29 02:01:16 -07002109 * This routine ensures that top_cpuset.cpus_allowed tracks
Tejun Heo3a101d02010-06-08 21:40:36 +02002110 * cpu_active_mask on each CPU hotplug (cpuhp) event.
Max Krasnyanskycf417142008-08-11 14:33:53 -07002111 *
2112 * Called within get_online_cpus(). Needs to call cgroup_lock()
2113 * before calling generate_sched_domains().
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002114 */
Tejun Heo0b2e9182010-06-21 23:53:31 +02002115void cpuset_update_active_cpus(void)
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002116{
Max Krasnyanskycf417142008-08-11 14:33:53 -07002117 struct sched_domain_attr *attr;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10302118 cpumask_var_t *doms;
Max Krasnyanskycf417142008-08-11 14:33:53 -07002119 int ndoms;
2120
Max Krasnyanskycf417142008-08-11 14:33:53 -07002121 cgroup_lock();
Li Zefan0b4217b2009-04-02 16:57:49 -07002122 mutex_lock(&callback_mutex);
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01002123 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
Li Zefan0b4217b2009-04-02 16:57:49 -07002124 mutex_unlock(&callback_mutex);
Max Krasnyanskycf417142008-08-11 14:33:53 -07002125 scan_for_empty_cpusets(&top_cpuset);
2126 ndoms = generate_sched_domains(&doms, &attr);
2127 cgroup_unlock();
2128
2129 /* Have scheduler rebuild the domains */
2130 partition_sched_domains(ndoms, doms, attr);
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002131}
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002132
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002133#ifdef CONFIG_MEMORY_HOTPLUG
Paul Jackson38837fc2006-09-29 02:01:16 -07002134/*
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07002135 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
Max Krasnyanskycf417142008-08-11 14:33:53 -07002136 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2137 * See also the previous routine cpuset_track_online_cpus().
Paul Jackson38837fc2006-09-29 02:01:16 -07002138 */
Miao Xief4818912008-11-19 15:36:30 -08002139static int cpuset_track_online_nodes(struct notifier_block *self,
2140 unsigned long action, void *arg)
Paul Jackson38837fc2006-09-29 02:01:16 -07002141{
Miao Xie53feb292010-03-23 13:35:35 -07002142 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2143
2144 if (oldmems == NULL)
2145 return NOTIFY_DONE;
Miao Xie5ab116c2010-03-23 13:35:34 -07002146
Max Krasnyanskycf417142008-08-11 14:33:53 -07002147 cgroup_lock();
Miao Xief4818912008-11-19 15:36:30 -08002148 switch (action) {
2149 case MEM_ONLINE:
Miao Xie53feb292010-03-23 13:35:35 -07002150 *oldmems = top_cpuset.mems_allowed;
Li Zefan0b4217b2009-04-02 16:57:49 -07002151 mutex_lock(&callback_mutex);
Miao Xief4818912008-11-19 15:36:30 -08002152 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
Li Zefan0b4217b2009-04-02 16:57:49 -07002153 mutex_unlock(&callback_mutex);
Miao Xie53feb292010-03-23 13:35:35 -07002154 update_tasks_nodemask(&top_cpuset, oldmems, NULL);
Miao Xie5ab116c2010-03-23 13:35:34 -07002155 break;
2156 case MEM_OFFLINE:
2157 /*
2158 * needn't update top_cpuset.mems_allowed explicitly because
2159 * scan_for_empty_cpusets() will update it.
2160 */
2161 scan_for_empty_cpusets(&top_cpuset);
Miao Xief4818912008-11-19 15:36:30 -08002162 break;
2163 default:
2164 break;
2165 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07002166 cgroup_unlock();
Miao Xie53feb292010-03-23 13:35:35 -07002167
2168 NODEMASK_FREE(oldmems);
Miao Xief4818912008-11-19 15:36:30 -08002169 return NOTIFY_OK;
Paul Jackson38837fc2006-09-29 02:01:16 -07002170}
2171#endif
2172
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173/**
2174 * cpuset_init_smp - initialize cpus_allowed
2175 *
2176 * Description: Finish top cpuset after cpu, node maps are initialized
2177 **/
2178
2179void __init cpuset_init_smp(void)
2180{
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01002181 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07002182 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002183
Miao Xief4818912008-11-19 15:36:30 -08002184 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
Miao Xief90d4112009-01-16 10:24:10 +08002185
2186 cpuset_wq = create_singlethread_workqueue("cpuset");
2187 BUG_ON(!cpuset_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188}
2189
2190/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2192 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
Li Zefan6af866a2009-01-07 18:08:45 -08002193 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 *
Li Zefan300ed6c2009-01-07 18:08:44 -08002195 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 * attached to the specified @tsk. Guaranteed to return some non-empty
2197 * subset of cpu_online_map, even if this means going outside the
2198 * tasks cpuset.
2199 **/
2200
Li Zefan6af866a2009-01-07 18:08:45 -08002201void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002203 mutex_lock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002204 task_lock(tsk);
Mike Travisf9a86fc2008-04-04 18:11:07 -07002205 guarantee_online_cpus(task_cs(tsk), pmask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002206 task_unlock(tsk);
Oleg Nesterov897f0b32010-03-15 10:10:03 +01002207 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208}
2209
Oleg Nesterov9084bb82010-03-15 10:10:27 +01002210int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2211{
2212 const struct cpuset *cs;
2213 int cpu;
2214
2215 rcu_read_lock();
2216 cs = task_cs(tsk);
2217 if (cs)
2218 cpumask_copy(&tsk->cpus_allowed, cs->cpus_allowed);
2219 rcu_read_unlock();
2220
2221 /*
2222 * We own tsk->cpus_allowed, nobody can change it under us.
2223 *
2224 * But we used cs && cs->cpus_allowed lockless and thus can
2225 * race with cgroup_attach_task() or update_cpumask() and get
2226 * the wrong tsk->cpus_allowed. However, both cases imply the
2227 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2228 * which takes task_rq_lock().
2229 *
2230 * If we are called after it dropped the lock we must see all
2231 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2232 * set any mask even if it is not right from task_cs() pov,
2233 * the pending set_cpus_allowed_ptr() will fix things.
2234 */
2235
2236 cpu = cpumask_any_and(&tsk->cpus_allowed, cpu_active_mask);
2237 if (cpu >= nr_cpu_ids) {
2238 /*
2239 * Either tsk->cpus_allowed is wrong (see above) or it
2240 * is actually empty. The latter case is only possible
2241 * if we are racing with remove_tasks_in_empty_cpuset().
2242 * Like above we can temporary set any mask and rely on
2243 * set_cpus_allowed_ptr() as synchronization point.
2244 */
2245 cpumask_copy(&tsk->cpus_allowed, cpu_possible_mask);
2246 cpu = cpumask_any(cpu_active_mask);
2247 }
2248
2249 return cpu;
2250}
2251
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252void cpuset_init_current_mems_allowed(void)
2253{
Mike Travisf9a86fc2008-04-04 18:11:07 -07002254 nodes_setall(current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255}
2256
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002257/**
Paul Jackson909d75a2006-01-08 01:01:55 -08002258 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2259 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2260 *
2261 * Description: Returns the nodemask_t mems_allowed of the cpuset
2262 * attached to the specified @tsk. Guaranteed to return some non-empty
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07002263 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
Paul Jackson909d75a2006-01-08 01:01:55 -08002264 * tasks cpuset.
2265 **/
2266
2267nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2268{
2269 nodemask_t mask;
2270
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002271 mutex_lock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002272 task_lock(tsk);
Paul Menage8793d852007-10-18 23:39:39 -07002273 guarantee_online_mems(task_cs(tsk), &mask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002274 task_unlock(tsk);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002275 mutex_unlock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002276
2277 return mask;
2278}
2279
2280/**
Mel Gorman19770b32008-04-28 02:12:18 -07002281 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2282 * @nodemask: the nodemask to be checked
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002283 *
Mel Gorman19770b32008-04-28 02:12:18 -07002284 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 */
Mel Gorman19770b32008-04-28 02:12:18 -07002286int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287{
Mel Gorman19770b32008-04-28 02:12:18 -07002288 return nodes_intersects(*nodemask, current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289}
2290
Paul Jackson9bf22292005-09-06 15:18:12 -07002291/*
Paul Menage78608362008-04-29 01:00:26 -07002292 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2293 * mem_hardwall ancestor to the specified cpuset. Call holding
2294 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2295 * (an unusual configuration), then returns the root cpuset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 */
Paul Menage78608362008-04-29 01:00:26 -07002297static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298{
Paul Menage78608362008-04-29 01:00:26 -07002299 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
Paul Jackson9bf22292005-09-06 15:18:12 -07002300 cs = cs->parent;
2301 return cs;
2302}
2303
2304/**
David Rientjesa1bc5a42009-04-02 16:57:54 -07002305 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2306 * @node: is this an allowed node?
Paul Jackson02a0e532006-12-13 00:34:25 -08002307 * @gfp_mask: memory allocation flags
Paul Jackson9bf22292005-09-06 15:18:12 -07002308 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002309 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2310 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2311 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2312 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2313 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2314 * flag, yes.
Paul Jackson9bf22292005-09-06 15:18:12 -07002315 * Otherwise, no.
2316 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002317 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2318 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2319 * might sleep, and might allow a node from an enclosing cpuset.
Paul Jackson02a0e532006-12-13 00:34:25 -08002320 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002321 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2322 * cpusets, and never sleeps.
Paul Jackson02a0e532006-12-13 00:34:25 -08002323 *
2324 * The __GFP_THISNODE placement logic is really handled elsewhere,
2325 * by forcibly using a zonelist starting at a specified node, and by
2326 * (in get_page_from_freelist()) refusing to consider the zones for
2327 * any node on the zonelist except the first. By the time any such
2328 * calls get to this routine, we should just shut up and say 'yes'.
2329 *
Paul Jackson9bf22292005-09-06 15:18:12 -07002330 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
David Rientjesc596d9f2007-05-06 14:49:32 -07002331 * and do not allow allocations outside the current tasks cpuset
2332 * unless the task has been OOM killed as is marked TIF_MEMDIE.
Paul Jackson9bf22292005-09-06 15:18:12 -07002333 * GFP_KERNEL allocations are not so marked, so can escape to the
Paul Menage78608362008-04-29 01:00:26 -07002334 * nearest enclosing hardwalled ancestor cpuset.
Paul Jackson9bf22292005-09-06 15:18:12 -07002335 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002336 * Scanning up parent cpusets requires callback_mutex. The
2337 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2338 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2339 * current tasks mems_allowed came up empty on the first pass over
2340 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2341 * cpuset are short of memory, might require taking the callback_mutex
2342 * mutex.
Paul Jackson9bf22292005-09-06 15:18:12 -07002343 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002344 * The first call here from mm/page_alloc:get_page_from_freelist()
Paul Jackson02a0e532006-12-13 00:34:25 -08002345 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2346 * so no allocation on a node outside the cpuset is allowed (unless
2347 * in interrupt, of course).
Paul Jackson9bf22292005-09-06 15:18:12 -07002348 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002349 * The second pass through get_page_from_freelist() doesn't even call
2350 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2351 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2352 * in alloc_flags. That logic and the checks below have the combined
2353 * affect that:
Paul Jackson9bf22292005-09-06 15:18:12 -07002354 * in_interrupt - any node ok (current task context irrelevant)
2355 * GFP_ATOMIC - any node ok
David Rientjesc596d9f2007-05-06 14:49:32 -07002356 * TIF_MEMDIE - any node ok
Paul Menage78608362008-04-29 01:00:26 -07002357 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
Paul Jackson9bf22292005-09-06 15:18:12 -07002358 * GFP_USER - only nodes in current tasks mems allowed ok.
Paul Jackson36be57f2006-05-20 15:00:10 -07002359 *
2360 * Rule:
David Rientjesa1bc5a42009-04-02 16:57:54 -07002361 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
Paul Jackson36be57f2006-05-20 15:00:10 -07002362 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2363 * the code that might scan up ancestor cpusets and sleep.
Paul Jackson02a0e532006-12-13 00:34:25 -08002364 */
David Rientjesa1bc5a42009-04-02 16:57:54 -07002365int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
Paul Jackson9bf22292005-09-06 15:18:12 -07002366{
Paul Jackson9bf22292005-09-06 15:18:12 -07002367 const struct cpuset *cs; /* current cpuset ancestors */
Paul Jackson29afd492006-03-24 03:16:12 -08002368 int allowed; /* is allocation in zone z allowed? */
Paul Jackson9bf22292005-09-06 15:18:12 -07002369
Christoph Lameter9b819d22006-09-25 23:31:40 -07002370 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
Paul Jackson9bf22292005-09-06 15:18:12 -07002371 return 1;
Paul Jackson92d1dbd2006-05-20 15:00:11 -07002372 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
Paul Jackson9bf22292005-09-06 15:18:12 -07002373 if (node_isset(node, current->mems_allowed))
2374 return 1;
David Rientjesc596d9f2007-05-06 14:49:32 -07002375 /*
2376 * Allow tasks that have access to memory reserves because they have
2377 * been OOM killed to get memory anywhere.
2378 */
2379 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2380 return 1;
Paul Jackson9bf22292005-09-06 15:18:12 -07002381 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2382 return 0;
2383
Bob Picco5563e772005-11-13 16:06:35 -08002384 if (current->flags & PF_EXITING) /* Let dying task have memory */
2385 return 1;
2386
Paul Jackson9bf22292005-09-06 15:18:12 -07002387 /* Not hardwall and node outside mems_allowed: scan up cpusets */
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002388 mutex_lock(&callback_mutex);
Paul Jackson053199e2005-10-30 15:02:30 -08002389
Paul Jackson053199e2005-10-30 15:02:30 -08002390 task_lock(current);
Paul Menage78608362008-04-29 01:00:26 -07002391 cs = nearest_hardwall_ancestor(task_cs(current));
Paul Jackson053199e2005-10-30 15:02:30 -08002392 task_unlock(current);
2393
Paul Jackson9bf22292005-09-06 15:18:12 -07002394 allowed = node_isset(node, cs->mems_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002395 mutex_unlock(&callback_mutex);
Paul Jackson9bf22292005-09-06 15:18:12 -07002396 return allowed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397}
2398
Paul Jackson02a0e532006-12-13 00:34:25 -08002399/*
David Rientjesa1bc5a42009-04-02 16:57:54 -07002400 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2401 * @node: is this an allowed node?
Paul Jackson02a0e532006-12-13 00:34:25 -08002402 * @gfp_mask: memory allocation flags
2403 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002404 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2405 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2406 * yes. If the task has been OOM killed and has access to memory reserves as
2407 * specified by the TIF_MEMDIE flag, yes.
2408 * Otherwise, no.
Paul Jackson02a0e532006-12-13 00:34:25 -08002409 *
2410 * The __GFP_THISNODE placement logic is really handled elsewhere,
2411 * by forcibly using a zonelist starting at a specified node, and by
2412 * (in get_page_from_freelist()) refusing to consider the zones for
2413 * any node on the zonelist except the first. By the time any such
2414 * calls get to this routine, we should just shut up and say 'yes'.
2415 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002416 * Unlike the cpuset_node_allowed_softwall() variant, above,
2417 * this variant requires that the node be in the current task's
Paul Jackson02a0e532006-12-13 00:34:25 -08002418 * mems_allowed or that we're in interrupt. It does not scan up the
2419 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2420 * It never sleeps.
2421 */
David Rientjesa1bc5a42009-04-02 16:57:54 -07002422int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
Paul Jackson02a0e532006-12-13 00:34:25 -08002423{
Paul Jackson02a0e532006-12-13 00:34:25 -08002424 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2425 return 1;
Paul Jackson02a0e532006-12-13 00:34:25 -08002426 if (node_isset(node, current->mems_allowed))
2427 return 1;
Daniel Walkerdedf8b72007-10-18 03:06:04 -07002428 /*
2429 * Allow tasks that have access to memory reserves because they have
2430 * been OOM killed to get memory anywhere.
2431 */
2432 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2433 return 1;
Paul Jackson02a0e532006-12-13 00:34:25 -08002434 return 0;
2435}
2436
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002437/**
Paul Jackson505970b2006-01-14 13:21:06 -08002438 * cpuset_unlock - release lock on cpuset changes
2439 *
2440 * Undo the lock taken in a previous cpuset_lock() call.
2441 */
2442
2443void cpuset_unlock(void)
2444{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002445 mutex_unlock(&callback_mutex);
Paul Jackson505970b2006-01-14 13:21:06 -08002446}
2447
2448/**
Jack Steiner6adef3e2010-05-26 14:42:49 -07002449 * cpuset_mem_spread_node() - On which node to begin search for a file page
2450 * cpuset_slab_spread_node() - On which node to begin search for a slab page
Paul Jackson825a46a2006-03-24 03:16:03 -08002451 *
2452 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2453 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2454 * and if the memory allocation used cpuset_mem_spread_node()
2455 * to determine on which node to start looking, as it will for
2456 * certain page cache or slab cache pages such as used for file
2457 * system buffers and inode caches, then instead of starting on the
2458 * local node to look for a free page, rather spread the starting
2459 * node around the tasks mems_allowed nodes.
2460 *
2461 * We don't have to worry about the returned node being offline
2462 * because "it can't happen", and even if it did, it would be ok.
2463 *
2464 * The routines calling guarantee_online_mems() are careful to
2465 * only set nodes in task->mems_allowed that are online. So it
2466 * should not be possible for the following code to return an
2467 * offline node. But if it did, that would be ok, as this routine
2468 * is not returning the node where the allocation must be, only
2469 * the node where the search should start. The zonelist passed to
2470 * __alloc_pages() will include all nodes. If the slab allocator
2471 * is passed an offline node, it will fall back to the local node.
2472 * See kmem_cache_alloc_node().
2473 */
2474
Jack Steiner6adef3e2010-05-26 14:42:49 -07002475static int cpuset_spread_node(int *rotor)
Paul Jackson825a46a2006-03-24 03:16:03 -08002476{
2477 int node;
2478
Jack Steiner6adef3e2010-05-26 14:42:49 -07002479 node = next_node(*rotor, current->mems_allowed);
Paul Jackson825a46a2006-03-24 03:16:03 -08002480 if (node == MAX_NUMNODES)
2481 node = first_node(current->mems_allowed);
Jack Steiner6adef3e2010-05-26 14:42:49 -07002482 *rotor = node;
Paul Jackson825a46a2006-03-24 03:16:03 -08002483 return node;
2484}
Jack Steiner6adef3e2010-05-26 14:42:49 -07002485
2486int cpuset_mem_spread_node(void)
2487{
2488 return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2489}
2490
2491int cpuset_slab_spread_node(void)
2492{
2493 return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2494}
2495
Paul Jackson825a46a2006-03-24 03:16:03 -08002496EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2497
2498/**
David Rientjesbbe373f2007-10-16 23:25:58 -07002499 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2500 * @tsk1: pointer to task_struct of some task.
2501 * @tsk2: pointer to task_struct of some other task.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002502 *
David Rientjesbbe373f2007-10-16 23:25:58 -07002503 * Description: Return true if @tsk1's mems_allowed intersects the
2504 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2505 * one of the task's memory usage might impact the memory available
2506 * to the other.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002507 **/
2508
David Rientjesbbe373f2007-10-16 23:25:58 -07002509int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2510 const struct task_struct *tsk2)
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002511{
David Rientjesbbe373f2007-10-16 23:25:58 -07002512 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002513}
2514
David Rientjes75aa1992009-01-06 14:39:01 -08002515/**
2516 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2517 * @task: pointer to task_struct of some task.
2518 *
2519 * Description: Prints @task's name, cpuset name, and cached copy of its
2520 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2521 * dereferencing task_cs(task).
2522 */
2523void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2524{
2525 struct dentry *dentry;
2526
2527 dentry = task_cs(tsk)->css.cgroup->dentry;
2528 spin_lock(&cpuset_buffer_lock);
2529 snprintf(cpuset_name, CPUSET_NAME_LEN,
2530 dentry ? (const char *)dentry->d_name.name : "/");
2531 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2532 tsk->mems_allowed);
2533 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2534 tsk->comm, cpuset_name, cpuset_nodelist);
2535 spin_unlock(&cpuset_buffer_lock);
2536}
2537
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538/*
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002539 * Collection of memory_pressure is suppressed unless
2540 * this flag is enabled by writing "1" to the special
2541 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2542 */
2543
Paul Jacksonc5b2aff2006-01-08 01:01:51 -08002544int cpuset_memory_pressure_enabled __read_mostly;
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002545
2546/**
2547 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2548 *
2549 * Keep a running average of the rate of synchronous (direct)
2550 * page reclaim efforts initiated by tasks in each cpuset.
2551 *
2552 * This represents the rate at which some task in the cpuset
2553 * ran low on memory on all nodes it was allowed to use, and
2554 * had to enter the kernels page reclaim code in an effort to
2555 * create more free memory by tossing clean pages or swapping
2556 * or writing dirty pages.
2557 *
2558 * Display to user space in the per-cpuset read-only file
2559 * "memory_pressure". Value displayed is an integer
2560 * representing the recent rate of entry into the synchronous
2561 * (direct) page reclaim by any task attached to the cpuset.
2562 **/
2563
2564void __cpuset_memory_pressure_bump(void)
2565{
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002566 task_lock(current);
Paul Menage8793d852007-10-18 23:39:39 -07002567 fmeter_markevent(&task_cs(current)->fmeter);
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002568 task_unlock(current);
2569}
2570
Paul Menage8793d852007-10-18 23:39:39 -07002571#ifdef CONFIG_PROC_PID_CPUSET
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002572/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 * proc_cpuset_show()
2574 * - Print tasks cpuset path into seq_file.
2575 * - Used for /proc/<pid>/cpuset.
Paul Jackson053199e2005-10-30 15:02:30 -08002576 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2577 * doesn't really matter if tsk->cpuset changes after we read it,
Paul Jacksonc8d9c902008-02-07 00:14:46 -08002578 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
Paul Menage2df167a2008-02-07 00:14:45 -08002579 * anyway.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 */
Paul Jackson029190c2007-10-18 23:40:20 -07002581static int proc_cpuset_show(struct seq_file *m, void *unused_v)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582{
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002583 struct pid *pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584 struct task_struct *tsk;
2585 char *buf;
Paul Menage8793d852007-10-18 23:39:39 -07002586 struct cgroup_subsys_state *css;
Eric W. Biederman99f89552006-06-26 00:25:55 -07002587 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588
Eric W. Biederman99f89552006-06-26 00:25:55 -07002589 retval = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2591 if (!buf)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002592 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593
Eric W. Biederman99f89552006-06-26 00:25:55 -07002594 retval = -ESRCH;
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002595 pid = m->private;
2596 tsk = get_pid_task(pid, PIDTYPE_PID);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002597 if (!tsk)
2598 goto out_free;
2599
2600 retval = -EINVAL;
Paul Menage8793d852007-10-18 23:39:39 -07002601 cgroup_lock();
2602 css = task_subsys_state(tsk, cpuset_subsys_id);
2603 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 if (retval < 0)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002605 goto out_unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 seq_puts(m, buf);
2607 seq_putc(m, '\n');
Eric W. Biederman99f89552006-06-26 00:25:55 -07002608out_unlock:
Paul Menage8793d852007-10-18 23:39:39 -07002609 cgroup_unlock();
Eric W. Biederman99f89552006-06-26 00:25:55 -07002610 put_task_struct(tsk);
2611out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 kfree(buf);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002613out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 return retval;
2615}
2616
2617static int cpuset_open(struct inode *inode, struct file *file)
2618{
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002619 struct pid *pid = PROC_I(inode)->pid;
2620 return single_open(file, proc_cpuset_show, pid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621}
2622
Arjan van de Ven9a321442007-02-12 00:55:35 -08002623const struct file_operations proc_cpuset_operations = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 .open = cpuset_open,
2625 .read = seq_read,
2626 .llseek = seq_lseek,
2627 .release = single_release,
2628};
Paul Menage8793d852007-10-18 23:39:39 -07002629#endif /* CONFIG_PROC_PID_CPUSET */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630
Heiko Carstensd01d4822009-09-21 11:06:27 +02002631/* Display task mems_allowed in /proc/<pid>/status file. */
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002632void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633{
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002634 seq_printf(m, "Mems_allowed:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002635 seq_nodemask(m, &task->mems_allowed);
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002636 seq_printf(m, "\n");
Mike Travis39106dc2008-04-08 11:43:03 -07002637 seq_printf(m, "Mems_allowed_list:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002638 seq_nodemask_list(m, &task->mems_allowed);
Mike Travis39106dc2008-04-08 11:43:03 -07002639 seq_printf(m, "\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640}