blob: 64b3f791bbe595905b00e9cf8ecbee763cbacf7d [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>
Paul Gortmaker9984de12011-05-23 14:51:41 -040040#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041#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>
Arun Sharma600634972011-07-26 16:09:06 -070058#include <linux/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/*
64 * Tracks how many cpusets are currently defined in system.
65 * When there is only one cpuset (the root cpuset) we can
66 * short circuit some hooks.
67 */
Paul Jackson7edc5962006-01-08 01:02:03 -080068int number_of_cpusets __read_mostly;
Paul Jackson202f72d2006-01-08 01:01:57 -080069
Paul Menage2df167a2008-02-07 00:14:45 -080070/* Forward declare cgroup structures */
Paul Menage8793d852007-10-18 23:39:39 -070071struct cgroup_subsys cpuset_subsys;
72struct cpuset;
73
Paul Jackson3e0d98b2006-01-08 01:01:49 -080074/* See "Frequency meter" comments, below. */
75
76struct fmeter {
77 int cnt; /* unprocessed events count */
78 int val; /* most recent output value */
79 time_t time; /* clock (secs) when val computed */
80 spinlock_t lock; /* guards read or write of above */
81};
82
Linus Torvalds1da177e2005-04-16 15:20:36 -070083struct cpuset {
Paul Menage8793d852007-10-18 23:39:39 -070084 struct cgroup_subsys_state css;
85
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 unsigned long flags; /* "unsigned long" so bitops work */
Li Zefan300ed6c2009-01-07 18:08:44 -080087 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
89
Paul Jackson3e0d98b2006-01-08 01:01:49 -080090 struct fmeter fmeter; /* memory_pressure filter */
Paul Jackson029190c2007-10-18 23:40:20 -070091
Tejun Heo452477f2013-01-07 08:51:07 -080092 /*
93 * Tasks are being attached to this cpuset. Used to prevent
94 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
95 */
96 int attach_in_progress;
97
Paul Jackson029190c2007-10-18 23:40:20 -070098 /* partition number for rebuild_sched_domains() */
99 int pn;
Cliff Wickman956db3c2008-02-07 00:14:43 -0800100
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900101 /* for custom sched domain */
102 int relax_domain_level;
103
Tejun Heo8d033942013-01-07 08:51:07 -0800104 struct work_struct hotplug_work;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105};
106
Paul Menage8793d852007-10-18 23:39:39 -0700107/* Retrieve the cpuset for a cgroup */
108static inline struct cpuset *cgroup_cs(struct cgroup *cont)
109{
110 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
111 struct cpuset, css);
112}
113
114/* Retrieve the cpuset for a task */
115static inline struct cpuset *task_cs(struct task_struct *task)
116{
117 return container_of(task_subsys_state(task, cpuset_subsys_id),
118 struct cpuset, css);
119}
Paul Menage8793d852007-10-18 23:39:39 -0700120
Tejun Heoc4310692013-01-07 08:51:08 -0800121static inline struct cpuset *parent_cs(const struct cpuset *cs)
122{
123 struct cgroup *pcgrp = cs->css.cgroup->parent;
124
125 if (pcgrp)
126 return cgroup_cs(pcgrp);
127 return NULL;
128}
129
David Rientjesb2462722011-12-19 17:11:52 -0800130#ifdef CONFIG_NUMA
131static inline bool task_has_mempolicy(struct task_struct *task)
132{
133 return task->mempolicy;
134}
135#else
136static inline bool task_has_mempolicy(struct task_struct *task)
137{
138 return false;
139}
140#endif
141
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143/* bits in struct cpuset flags field */
144typedef enum {
Tejun Heoefeb77b2013-01-07 08:51:07 -0800145 CS_ONLINE,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 CS_CPU_EXCLUSIVE,
147 CS_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -0700148 CS_MEM_HARDWALL,
Paul Jackson45b07ef2006-01-08 01:00:56 -0800149 CS_MEMORY_MIGRATE,
Paul Jackson029190c2007-10-18 23:40:20 -0700150 CS_SCHED_LOAD_BALANCE,
Paul Jackson825a46a2006-03-24 03:16:03 -0800151 CS_SPREAD_PAGE,
152 CS_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153} cpuset_flagbits_t;
154
155/* convenient tests for these bits */
Tejun Heoefeb77b2013-01-07 08:51:07 -0800156static inline bool is_cpuset_online(const struct cpuset *cs)
157{
158 return test_bit(CS_ONLINE, &cs->flags);
159}
160
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161static inline int is_cpu_exclusive(const struct cpuset *cs)
162{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800163 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164}
165
166static inline int is_mem_exclusive(const struct cpuset *cs)
167{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800168 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
Paul Menage78608362008-04-29 01:00:26 -0700171static inline int is_mem_hardwall(const struct cpuset *cs)
172{
173 return test_bit(CS_MEM_HARDWALL, &cs->flags);
174}
175
Paul Jackson029190c2007-10-18 23:40:20 -0700176static inline int is_sched_load_balance(const struct cpuset *cs)
177{
178 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
179}
180
Paul Jackson45b07ef2006-01-08 01:00:56 -0800181static inline int is_memory_migrate(const struct cpuset *cs)
182{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800183 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
Paul Jackson45b07ef2006-01-08 01:00:56 -0800184}
185
Paul Jackson825a46a2006-03-24 03:16:03 -0800186static inline int is_spread_page(const struct cpuset *cs)
187{
188 return test_bit(CS_SPREAD_PAGE, &cs->flags);
189}
190
191static inline int is_spread_slab(const struct cpuset *cs)
192{
193 return test_bit(CS_SPREAD_SLAB, &cs->flags);
194}
195
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196static struct cpuset top_cpuset = {
Tejun Heoefeb77b2013-01-07 08:51:07 -0800197 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
198 (1 << CS_MEM_EXCLUSIVE)),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199};
200
Tejun Heoae8086c2013-01-07 08:51:07 -0800201/**
202 * cpuset_for_each_child - traverse online children of a cpuset
203 * @child_cs: loop cursor pointing to the current child
204 * @pos_cgrp: used for iteration
205 * @parent_cs: target cpuset to walk children of
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 *
Tejun Heoae8086c2013-01-07 08:51:07 -0800207 * Walk @child_cs through the online children of @parent_cs. Must be used
208 * with RCU read locked.
209 */
210#define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
211 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
212 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
213
Tejun Heofc560a22013-01-07 08:51:08 -0800214/**
215 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
216 * @des_cs: loop cursor pointing to the current descendant
217 * @pos_cgrp: used for iteration
218 * @root_cs: target cpuset to walk ancestor of
219 *
220 * Walk @des_cs through the online descendants of @root_cs. Must be used
221 * with RCU read locked. The caller may modify @pos_cgrp by calling
222 * cgroup_rightmost_descendant() to skip subtree.
223 */
224#define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \
225 cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
226 if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
227
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228/*
Tejun Heo5d21cc22013-01-07 08:51:08 -0800229 * There are two global mutexes guarding cpuset structures - cpuset_mutex
230 * and callback_mutex. The latter may nest inside the former. We also
231 * require taking task_lock() when dereferencing a task's cpuset pointer.
232 * See "The task_lock() exception", at the end of this comment.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 *
Tejun Heo5d21cc22013-01-07 08:51:08 -0800234 * A task must hold both mutexes to modify cpusets. If a task holds
235 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
236 * is the only task able to also acquire callback_mutex and be able to
237 * modify cpusets. It can perform various checks on the cpuset structure
238 * first, knowing nothing will change. It can also allocate memory while
239 * just holding cpuset_mutex. While it is performing these checks, various
240 * callback routines can briefly acquire callback_mutex to query cpusets.
241 * Once it is ready to make the changes, it takes callback_mutex, blocking
242 * everyone else.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 *
Paul Jackson053199e2005-10-30 15:02:30 -0800244 * Calls to the kernel memory allocator can not be made while holding
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800245 * callback_mutex, as that would risk double tripping on callback_mutex
Paul Jackson053199e2005-10-30 15:02:30 -0800246 * from one of the callbacks into the cpuset code from within
247 * __alloc_pages().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800249 * If a task is only holding callback_mutex, then it has read-only
Paul Jackson053199e2005-10-30 15:02:30 -0800250 * access to cpusets.
251 *
Miao Xie58568d22009-06-16 15:31:49 -0700252 * Now, the task_struct fields mems_allowed and mempolicy may be changed
253 * by other task, we use alloc_lock in the task_struct fields to protect
254 * them.
Paul Jackson053199e2005-10-30 15:02:30 -0800255 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800256 * The cpuset_common_file_read() handlers only hold callback_mutex across
Paul Jackson053199e2005-10-30 15:02:30 -0800257 * small pieces of code, such as when reading out possibly multi-word
258 * cpumasks and nodemasks.
259 *
Paul Menage2df167a2008-02-07 00:14:45 -0800260 * Accessing a task's cpuset should be done in accordance with the
261 * guidelines for accessing subsystem state in kernel/cgroup.c
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 */
263
Tejun Heo5d21cc22013-01-07 08:51:08 -0800264static DEFINE_MUTEX(cpuset_mutex);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800265static DEFINE_MUTEX(callback_mutex);
Paul Jackson4247bdc2005-09-10 00:26:06 -0700266
Max Krasnyanskycf417142008-08-11 14:33:53 -0700267/*
Tejun Heo3a5a6d02013-01-07 08:51:07 -0800268 * CPU / memory hotplug is handled asynchronously.
269 */
Tejun Heo8d033942013-01-07 08:51:07 -0800270static struct workqueue_struct *cpuset_propagate_hotplug_wq;
271
Tejun Heo3a5a6d02013-01-07 08:51:07 -0800272static void cpuset_hotplug_workfn(struct work_struct *work);
Tejun Heo8d033942013-01-07 08:51:07 -0800273static void cpuset_propagate_hotplug_workfn(struct work_struct *work);
Tejun Heo02bb5862013-01-07 08:51:08 -0800274static void schedule_cpuset_propagate_hotplug(struct cpuset *cs);
Tejun Heo3a5a6d02013-01-07 08:51:07 -0800275
276static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
277
278/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700279 * This is ugly, but preserves the userspace API for existing cpuset
Paul Menage8793d852007-10-18 23:39:39 -0700280 * users. If someone tries to mount the "cpuset" filesystem, we
Max Krasnyanskycf417142008-08-11 14:33:53 -0700281 * silently switch it to mount "cgroup" instead
282 */
Al Virof7e83572010-07-26 13:23:11 +0400283static struct dentry *cpuset_mount(struct file_system_type *fs_type,
284 int flags, const char *unused_dev_name, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285{
Paul Menage8793d852007-10-18 23:39:39 -0700286 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
Al Virof7e83572010-07-26 13:23:11 +0400287 struct dentry *ret = ERR_PTR(-ENODEV);
Paul Menage8793d852007-10-18 23:39:39 -0700288 if (cgroup_fs) {
289 char mountopts[] =
290 "cpuset,noprefix,"
291 "release_agent=/sbin/cpuset_release_agent";
Al Virof7e83572010-07-26 13:23:11 +0400292 ret = cgroup_fs->mount(cgroup_fs, flags,
293 unused_dev_name, mountopts);
Paul Menage8793d852007-10-18 23:39:39 -0700294 put_filesystem(cgroup_fs);
295 }
296 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297}
298
299static struct file_system_type cpuset_fs_type = {
300 .name = "cpuset",
Al Virof7e83572010-07-26 13:23:11 +0400301 .mount = cpuset_mount,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304/*
Li Zefan300ed6c2009-01-07 18:08:44 -0800305 * Return in pmask the portion of a cpusets's cpus_allowed that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 * are online. If none are online, walk up the cpuset hierarchy
307 * until we find one that does have some online cpus. If we get
308 * all the way to the top and still haven't found any online cpus,
Rusty Russell5f054e32012-03-29 15:38:31 +1030309 * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
310 * task, return cpu_online_mask.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 *
312 * One way or another, we guarantee to return some non-empty subset
Rusty Russell5f054e32012-03-29 15:38:31 +1030313 * of cpu_online_mask.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800315 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 */
317
Li Zefan6af866a2009-01-07 18:08:45 -0800318static void guarantee_online_cpus(const struct cpuset *cs,
319 struct cpumask *pmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320{
Li Zefan300ed6c2009-01-07 18:08:44 -0800321 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
Tejun Heoc4310692013-01-07 08:51:08 -0800322 cs = parent_cs(cs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 if (cs)
Li Zefan300ed6c2009-01-07 18:08:44 -0800324 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 else
Li Zefan300ed6c2009-01-07 18:08:44 -0800326 cpumask_copy(pmask, cpu_online_mask);
327 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328}
329
330/*
331 * Return in *pmask the portion of a cpusets's mems_allowed that
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700332 * are online, with memory. If none are online with memory, walk
333 * up the cpuset hierarchy until we find one that does have some
334 * online mems. If we get all the way to the top and still haven't
Lai Jiangshan38d7bee2012-12-12 13:51:24 -0800335 * found any online mems, return node_states[N_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 *
337 * One way or another, we guarantee to return some non-empty subset
Lai Jiangshan38d7bee2012-12-12 13:51:24 -0800338 * of node_states[N_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800340 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 */
342
343static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
344{
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700345 while (cs && !nodes_intersects(cs->mems_allowed,
Lai Jiangshan38d7bee2012-12-12 13:51:24 -0800346 node_states[N_MEMORY]))
Tejun Heoc4310692013-01-07 08:51:08 -0800347 cs = parent_cs(cs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 if (cs)
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700349 nodes_and(*pmask, cs->mems_allowed,
Lai Jiangshan38d7bee2012-12-12 13:51:24 -0800350 node_states[N_MEMORY]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351 else
Lai Jiangshan38d7bee2012-12-12 13:51:24 -0800352 *pmask = node_states[N_MEMORY];
353 BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354}
355
Miao Xief3b39d42009-06-16 15:31:46 -0700356/*
357 * update task's spread flag if cpuset's page/slab spread flag is set
358 *
Tejun Heo5d21cc22013-01-07 08:51:08 -0800359 * Called with callback_mutex/cpuset_mutex held
Miao Xief3b39d42009-06-16 15:31:46 -0700360 */
361static void cpuset_update_task_spread_flag(struct cpuset *cs,
362 struct task_struct *tsk)
363{
364 if (is_spread_page(cs))
365 tsk->flags |= PF_SPREAD_PAGE;
366 else
367 tsk->flags &= ~PF_SPREAD_PAGE;
368 if (is_spread_slab(cs))
369 tsk->flags |= PF_SPREAD_SLAB;
370 else
371 tsk->flags &= ~PF_SPREAD_SLAB;
372}
373
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374/*
375 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
376 *
377 * One cpuset is a subset of another if all its allowed CPUs and
378 * Memory Nodes are a subset of the other, and its exclusive flags
Tejun Heo5d21cc22013-01-07 08:51:08 -0800379 * are only set if the other's are set. Call holding cpuset_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 */
381
382static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
383{
Li Zefan300ed6c2009-01-07 18:08:44 -0800384 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 nodes_subset(p->mems_allowed, q->mems_allowed) &&
386 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
387 is_mem_exclusive(p) <= is_mem_exclusive(q);
388}
389
Li Zefan645fcc92009-01-07 18:08:43 -0800390/**
391 * alloc_trial_cpuset - allocate a trial cpuset
392 * @cs: the cpuset that the trial cpuset duplicates
393 */
394static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
395{
Li Zefan300ed6c2009-01-07 18:08:44 -0800396 struct cpuset *trial;
397
398 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
399 if (!trial)
400 return NULL;
401
402 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
403 kfree(trial);
404 return NULL;
405 }
406 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
407
408 return trial;
Li Zefan645fcc92009-01-07 18:08:43 -0800409}
410
411/**
412 * free_trial_cpuset - free the trial cpuset
413 * @trial: the trial cpuset to be freed
414 */
415static void free_trial_cpuset(struct cpuset *trial)
416{
Li Zefan300ed6c2009-01-07 18:08:44 -0800417 free_cpumask_var(trial->cpus_allowed);
Li Zefan645fcc92009-01-07 18:08:43 -0800418 kfree(trial);
419}
420
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421/*
422 * validate_change() - Used to validate that any proposed cpuset change
423 * follows the structural rules for cpusets.
424 *
425 * If we replaced the flag and mask values of the current cpuset
426 * (cur) with those values in the trial cpuset (trial), would
427 * our various subset and exclusive rules still be valid? Presumes
Tejun Heo5d21cc22013-01-07 08:51:08 -0800428 * cpuset_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 *
430 * 'cur' is the address of an actual, in-use cpuset. Operations
431 * such as list traversal that depend on the actual address of the
432 * cpuset in the list must use cur below, not trial.
433 *
434 * 'trial' is the address of bulk structure copy of cur, with
435 * perhaps one or more of the fields cpus_allowed, mems_allowed,
436 * or flags changed to new, trial values.
437 *
438 * Return 0 if valid, -errno if not.
439 */
440
441static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
442{
Paul Menage8793d852007-10-18 23:39:39 -0700443 struct cgroup *cont;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 struct cpuset *c, *par;
Tejun Heoae8086c2013-01-07 08:51:07 -0800445 int ret;
446
447 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448
449 /* Each of our child cpusets must be a subset of us */
Tejun Heoae8086c2013-01-07 08:51:07 -0800450 ret = -EBUSY;
451 cpuset_for_each_child(c, cont, cur)
452 if (!is_cpuset_subset(c, trial))
453 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454
455 /* Remaining checks don't apply to root cpuset */
Tejun Heoae8086c2013-01-07 08:51:07 -0800456 ret = 0;
Paul Jackson69604062006-12-06 20:36:15 -0800457 if (cur == &top_cpuset)
Tejun Heoae8086c2013-01-07 08:51:07 -0800458 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459
Tejun Heoc4310692013-01-07 08:51:08 -0800460 par = parent_cs(cur);
Paul Jackson69604062006-12-06 20:36:15 -0800461
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 /* We must be a subset of our parent cpuset */
Tejun Heoae8086c2013-01-07 08:51:07 -0800463 ret = -EACCES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 if (!is_cpuset_subset(trial, par))
Tejun Heoae8086c2013-01-07 08:51:07 -0800465 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466
Paul Menage2df167a2008-02-07 00:14:45 -0800467 /*
468 * If either I or some sibling (!= me) is exclusive, we can't
469 * overlap
470 */
Tejun Heoae8086c2013-01-07 08:51:07 -0800471 ret = -EINVAL;
472 cpuset_for_each_child(c, cont, par) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
474 c != cur &&
Li Zefan300ed6c2009-01-07 18:08:44 -0800475 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
Tejun Heoae8086c2013-01-07 08:51:07 -0800476 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
478 c != cur &&
479 nodes_intersects(trial->mems_allowed, c->mems_allowed))
Tejun Heoae8086c2013-01-07 08:51:07 -0800480 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 }
482
Tejun Heo452477f2013-01-07 08:51:07 -0800483 /*
484 * Cpusets with tasks - existing or newly being attached - can't
485 * have empty cpus_allowed or mems_allowed.
486 */
Tejun Heoae8086c2013-01-07 08:51:07 -0800487 ret = -ENOSPC;
Tejun Heo452477f2013-01-07 08:51:07 -0800488 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
Tejun Heoae8086c2013-01-07 08:51:07 -0800489 (cpumask_empty(trial->cpus_allowed) ||
490 nodes_empty(trial->mems_allowed)))
491 goto out;
Paul Jackson020958b2007-10-18 23:40:21 -0700492
Tejun Heoae8086c2013-01-07 08:51:07 -0800493 ret = 0;
494out:
495 rcu_read_unlock();
496 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497}
498
Paul Menagedb7f47c2009-04-02 16:57:55 -0700499#ifdef CONFIG_SMP
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700500/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700501 * Helper routine for generate_sched_domains().
Paul Jackson029190c2007-10-18 23:40:20 -0700502 * Do cpusets a, b have overlapping cpus_allowed masks?
503 */
Paul Jackson029190c2007-10-18 23:40:20 -0700504static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
505{
Li Zefan300ed6c2009-01-07 18:08:44 -0800506 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
Paul Jackson029190c2007-10-18 23:40:20 -0700507}
508
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900509static void
510update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
511{
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900512 if (dattr->relax_domain_level < c->relax_domain_level)
513 dattr->relax_domain_level = c->relax_domain_level;
514 return;
515}
516
Tejun Heofc560a22013-01-07 08:51:08 -0800517static void update_domain_attr_tree(struct sched_domain_attr *dattr,
518 struct cpuset *root_cs)
Lai Jiangshanf5393692008-07-29 22:33:22 -0700519{
Tejun Heofc560a22013-01-07 08:51:08 -0800520 struct cpuset *cp;
521 struct cgroup *pos_cgrp;
Lai Jiangshanf5393692008-07-29 22:33:22 -0700522
Tejun Heofc560a22013-01-07 08:51:08 -0800523 rcu_read_lock();
524 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
525 /* skip the whole subtree if @cp doesn't have any CPU */
526 if (cpumask_empty(cp->cpus_allowed)) {
527 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
Lai Jiangshanf5393692008-07-29 22:33:22 -0700528 continue;
Tejun Heofc560a22013-01-07 08:51:08 -0800529 }
Lai Jiangshanf5393692008-07-29 22:33:22 -0700530
531 if (is_sched_load_balance(cp))
532 update_domain_attr(dattr, cp);
Lai Jiangshanf5393692008-07-29 22:33:22 -0700533 }
Tejun Heofc560a22013-01-07 08:51:08 -0800534 rcu_read_unlock();
Lai Jiangshanf5393692008-07-29 22:33:22 -0700535}
536
Paul Jackson029190c2007-10-18 23:40:20 -0700537/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700538 * generate_sched_domains()
Paul Jackson029190c2007-10-18 23:40:20 -0700539 *
Max Krasnyanskycf417142008-08-11 14:33:53 -0700540 * This function builds a partial partition of the systems CPUs
541 * A 'partial partition' is a set of non-overlapping subsets whose
542 * union is a subset of that set.
543 * The output of this function needs to be passed to kernel/sched.c
544 * partition_sched_domains() routine, which will rebuild the scheduler's
545 * load balancing domains (sched domains) as specified by that partial
546 * partition.
Paul Jackson029190c2007-10-18 23:40:20 -0700547 *
Li Zefan45ce80f2009-01-15 13:50:59 -0800548 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
Paul Jackson029190c2007-10-18 23:40:20 -0700549 * for a background explanation of this.
550 *
551 * Does not return errors, on the theory that the callers of this
552 * routine would rather not worry about failures to rebuild sched
553 * domains when operating in the severe memory shortage situations
554 * that could cause allocation failures below.
555 *
Tejun Heo5d21cc22013-01-07 08:51:08 -0800556 * Must be called with cpuset_mutex held.
Paul Jackson029190c2007-10-18 23:40:20 -0700557 *
558 * The three key local variables below are:
Li Zefanaeed6822008-07-29 22:33:24 -0700559 * q - a linked-list queue of cpuset pointers, used to implement a
Paul Jackson029190c2007-10-18 23:40:20 -0700560 * top-down scan of all cpusets. This scan loads a pointer
561 * to each cpuset marked is_sched_load_balance into the
562 * array 'csa'. For our purposes, rebuilding the schedulers
563 * sched domains, we can ignore !is_sched_load_balance cpusets.
564 * csa - (for CpuSet Array) Array of pointers to all the cpusets
565 * that need to be load balanced, for convenient iterative
566 * access by the subsequent code that finds the best partition,
567 * i.e the set of domains (subsets) of CPUs such that the
568 * cpus_allowed of every cpuset marked is_sched_load_balance
569 * is a subset of one of these domains, while there are as
570 * many such domains as possible, each as small as possible.
571 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
572 * the kernel/sched.c routine partition_sched_domains() in a
573 * convenient format, that can be easily compared to the prior
574 * value to determine what partition elements (sched domains)
575 * were changed (added or removed.)
576 *
577 * Finding the best partition (set of domains):
578 * The triple nested loops below over i, j, k scan over the
579 * load balanced cpusets (using the array of cpuset pointers in
580 * csa[]) looking for pairs of cpusets that have overlapping
581 * cpus_allowed, but which don't have the same 'pn' partition
582 * number and gives them in the same partition number. It keeps
583 * looping on the 'restart' label until it can no longer find
584 * any such pairs.
585 *
586 * The union of the cpus_allowed masks from the set of
587 * all cpusets having the same 'pn' value then form the one
588 * element of the partition (one sched domain) to be passed to
589 * partition_sched_domains().
590 */
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030591static int generate_sched_domains(cpumask_var_t **domains,
Max Krasnyanskycf417142008-08-11 14:33:53 -0700592 struct sched_domain_attr **attributes)
Paul Jackson029190c2007-10-18 23:40:20 -0700593{
Paul Jackson029190c2007-10-18 23:40:20 -0700594 struct cpuset *cp; /* scans q */
595 struct cpuset **csa; /* array of all cpuset ptrs */
596 int csn; /* how many cpuset ptrs in csa so far */
597 int i, j, k; /* indices for partition finding loops */
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030598 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900599 struct sched_domain_attr *dattr; /* attributes for custom domains */
Ingo Molnar15837152008-11-25 10:27:49 +0100600 int ndoms = 0; /* number of sched domains in result */
Li Zefan6af866a2009-01-07 18:08:45 -0800601 int nslot; /* next empty doms[] struct cpumask slot */
Tejun Heofc560a22013-01-07 08:51:08 -0800602 struct cgroup *pos_cgrp;
Paul Jackson029190c2007-10-18 23:40:20 -0700603
Paul Jackson029190c2007-10-18 23:40:20 -0700604 doms = NULL;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900605 dattr = NULL;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700606 csa = NULL;
Paul Jackson029190c2007-10-18 23:40:20 -0700607
608 /* Special case for the 99% of systems with one, full, sched domain */
609 if (is_sched_load_balance(&top_cpuset)) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030610 ndoms = 1;
611 doms = alloc_sched_domains(ndoms);
Paul Jackson029190c2007-10-18 23:40:20 -0700612 if (!doms)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700613 goto done;
614
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900615 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
616 if (dattr) {
617 *dattr = SD_ATTR_INIT;
Li Zefan93a65572008-07-29 22:33:23 -0700618 update_domain_attr_tree(dattr, &top_cpuset);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900619 }
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030620 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700621
Max Krasnyanskycf417142008-08-11 14:33:53 -0700622 goto done;
Paul Jackson029190c2007-10-18 23:40:20 -0700623 }
624
Paul Jackson029190c2007-10-18 23:40:20 -0700625 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
626 if (!csa)
627 goto done;
628 csn = 0;
629
Tejun Heofc560a22013-01-07 08:51:08 -0800630 rcu_read_lock();
631 cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
Lai Jiangshanf5393692008-07-29 22:33:22 -0700632 /*
Tejun Heofc560a22013-01-07 08:51:08 -0800633 * Continue traversing beyond @cp iff @cp has some CPUs and
634 * isn't load balancing. The former is obvious. The
635 * latter: All child cpusets contain a subset of the
636 * parent's cpus, so just skip them, and then we call
637 * update_domain_attr_tree() to calc relax_domain_level of
638 * the corresponding sched domain.
Lai Jiangshanf5393692008-07-29 22:33:22 -0700639 */
Tejun Heofc560a22013-01-07 08:51:08 -0800640 if (!cpumask_empty(cp->cpus_allowed) &&
641 !is_sched_load_balance(cp))
Lai Jiangshanf5393692008-07-29 22:33:22 -0700642 continue;
Lai Jiangshan489a5392008-07-25 01:47:23 -0700643
Tejun Heofc560a22013-01-07 08:51:08 -0800644 if (is_sched_load_balance(cp))
645 csa[csn++] = cp;
646
647 /* skip @cp's subtree */
648 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
649 }
650 rcu_read_unlock();
Paul Jackson029190c2007-10-18 23:40:20 -0700651
652 for (i = 0; i < csn; i++)
653 csa[i]->pn = i;
654 ndoms = csn;
655
656restart:
657 /* Find the best partition (set of sched domains) */
658 for (i = 0; i < csn; i++) {
659 struct cpuset *a = csa[i];
660 int apn = a->pn;
661
662 for (j = 0; j < csn; j++) {
663 struct cpuset *b = csa[j];
664 int bpn = b->pn;
665
666 if (apn != bpn && cpusets_overlap(a, b)) {
667 for (k = 0; k < csn; k++) {
668 struct cpuset *c = csa[k];
669
670 if (c->pn == bpn)
671 c->pn = apn;
672 }
673 ndoms--; /* one less element */
674 goto restart;
675 }
676 }
677 }
678
Max Krasnyanskycf417142008-08-11 14:33:53 -0700679 /*
680 * Now we know how many domains to create.
681 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
682 */
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030683 doms = alloc_sched_domains(ndoms);
Li Zefan700018e2008-11-18 14:02:03 +0800684 if (!doms)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700685 goto done;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700686
687 /*
688 * The rest of the code, including the scheduler, can deal with
689 * dattr==NULL case. No need to abort if alloc fails.
690 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900691 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
Paul Jackson029190c2007-10-18 23:40:20 -0700692
693 for (nslot = 0, i = 0; i < csn; i++) {
694 struct cpuset *a = csa[i];
Li Zefan6af866a2009-01-07 18:08:45 -0800695 struct cpumask *dp;
Paul Jackson029190c2007-10-18 23:40:20 -0700696 int apn = a->pn;
697
Max Krasnyanskycf417142008-08-11 14:33:53 -0700698 if (apn < 0) {
699 /* Skip completed partitions */
700 continue;
Paul Jackson029190c2007-10-18 23:40:20 -0700701 }
Max Krasnyanskycf417142008-08-11 14:33:53 -0700702
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030703 dp = doms[nslot];
Max Krasnyanskycf417142008-08-11 14:33:53 -0700704
705 if (nslot == ndoms) {
706 static int warnings = 10;
707 if (warnings) {
708 printk(KERN_WARNING
709 "rebuild_sched_domains confused:"
710 " nslot %d, ndoms %d, csn %d, i %d,"
711 " apn %d\n",
712 nslot, ndoms, csn, i, apn);
713 warnings--;
714 }
715 continue;
716 }
717
Li Zefan6af866a2009-01-07 18:08:45 -0800718 cpumask_clear(dp);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700719 if (dattr)
720 *(dattr + nslot) = SD_ATTR_INIT;
721 for (j = i; j < csn; j++) {
722 struct cpuset *b = csa[j];
723
724 if (apn == b->pn) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800725 cpumask_or(dp, dp, b->cpus_allowed);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700726 if (dattr)
727 update_domain_attr_tree(dattr + nslot, b);
728
729 /* Done with this partition */
730 b->pn = -1;
731 }
732 }
733 nslot++;
Paul Jackson029190c2007-10-18 23:40:20 -0700734 }
735 BUG_ON(nslot != ndoms);
736
Paul Jackson029190c2007-10-18 23:40:20 -0700737done:
Paul Jackson029190c2007-10-18 23:40:20 -0700738 kfree(csa);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700739
Li Zefan700018e2008-11-18 14:02:03 +0800740 /*
741 * Fallback to the default domain if kmalloc() failed.
742 * See comments in partition_sched_domains().
743 */
744 if (doms == NULL)
745 ndoms = 1;
746
Max Krasnyanskycf417142008-08-11 14:33:53 -0700747 *domains = doms;
748 *attributes = dattr;
749 return ndoms;
750}
751
752/*
753 * Rebuild scheduler domains.
754 *
Tejun Heo699140b2013-01-07 08:51:07 -0800755 * If the flag 'sched_load_balance' of any cpuset with non-empty
756 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
757 * which has that flag enabled, or if any cpuset with a non-empty
758 * 'cpus' is removed, then call this routine to rebuild the
759 * scheduler's dynamic sched domains.
Max Krasnyanskycf417142008-08-11 14:33:53 -0700760 *
Tejun Heo5d21cc22013-01-07 08:51:08 -0800761 * Call with cpuset_mutex held. Takes get_online_cpus().
Max Krasnyanskycf417142008-08-11 14:33:53 -0700762 */
Tejun Heo699140b2013-01-07 08:51:07 -0800763static void rebuild_sched_domains_locked(void)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700764{
765 struct sched_domain_attr *attr;
Rusty Russellacc3f5d2009-11-03 14:53:40 +1030766 cpumask_var_t *doms;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700767 int ndoms;
768
Tejun Heo5d21cc22013-01-07 08:51:08 -0800769 lockdep_assert_held(&cpuset_mutex);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700770 get_online_cpus();
771
Li Zefan5b16c2a2013-04-27 06:52:43 -0700772 /*
773 * We have raced with CPU hotplug. Don't do anything to avoid
774 * passing doms with offlined cpu to partition_sched_domains().
775 * Anyways, hotplug work item will rebuild sched domains.
776 */
777 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
778 goto out;
779
Max Krasnyanskycf417142008-08-11 14:33:53 -0700780 /* Generate domain masks and attrs */
Max Krasnyanskycf417142008-08-11 14:33:53 -0700781 ndoms = generate_sched_domains(&doms, &attr);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700782
783 /* Have scheduler rebuild the domains */
784 partition_sched_domains(ndoms, doms, attr);
Li Zefan5b16c2a2013-04-27 06:52:43 -0700785out:
Max Krasnyanskycf417142008-08-11 14:33:53 -0700786 put_online_cpus();
787}
Paul Menagedb7f47c2009-04-02 16:57:55 -0700788#else /* !CONFIG_SMP */
Tejun Heo699140b2013-01-07 08:51:07 -0800789static void rebuild_sched_domains_locked(void)
Paul Menagedb7f47c2009-04-02 16:57:55 -0700790{
791}
Paul Menagedb7f47c2009-04-02 16:57:55 -0700792#endif /* CONFIG_SMP */
Max Krasnyanskycf417142008-08-11 14:33:53 -0700793
Max Krasnyanskycf417142008-08-11 14:33:53 -0700794void rebuild_sched_domains(void)
795{
Tejun Heo5d21cc22013-01-07 08:51:08 -0800796 mutex_lock(&cpuset_mutex);
Tejun Heo699140b2013-01-07 08:51:07 -0800797 rebuild_sched_domains_locked();
Tejun Heo5d21cc22013-01-07 08:51:08 -0800798 mutex_unlock(&cpuset_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -0700799}
800
Cliff Wickman58f47902008-02-07 00:14:44 -0800801/**
802 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
803 * @tsk: task to test
804 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
805 *
Tejun Heo5d21cc22013-01-07 08:51:08 -0800806 * Call with cpuset_mutex held. May take callback_mutex during call.
Cliff Wickman58f47902008-02-07 00:14:44 -0800807 * Called for each task in a cgroup by cgroup_scan_tasks().
808 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
809 * words, if its mask is not equal to its cpuset's mask).
Paul Jackson053199e2005-10-30 15:02:30 -0800810 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700811static int cpuset_test_cpumask(struct task_struct *tsk,
812 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800813{
Li Zefan300ed6c2009-01-07 18:08:44 -0800814 return !cpumask_equal(&tsk->cpus_allowed,
Cliff Wickman58f47902008-02-07 00:14:44 -0800815 (cgroup_cs(scan->cg))->cpus_allowed);
816}
Paul Jackson053199e2005-10-30 15:02:30 -0800817
Cliff Wickman58f47902008-02-07 00:14:44 -0800818/**
819 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
820 * @tsk: task to test
821 * @scan: struct cgroup_scanner containing the cgroup of the task
822 *
823 * Called by cgroup_scan_tasks() for each task in a cgroup whose
824 * cpus_allowed mask needs to be changed.
825 *
826 * We don't need to re-check for the cgroup/cpuset membership, since we're
Tejun Heo5d21cc22013-01-07 08:51:08 -0800827 * holding cpuset_mutex at this point.
Cliff Wickman58f47902008-02-07 00:14:44 -0800828 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700829static void cpuset_change_cpumask(struct task_struct *tsk,
830 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800831{
Li Zefan300ed6c2009-01-07 18:08:44 -0800832 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
Cliff Wickman58f47902008-02-07 00:14:44 -0800833}
834
835/**
Miao Xie0b2f6302008-07-25 01:47:21 -0700836 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
837 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
Li Zefan4e743392008-09-13 02:33:08 -0700838 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
Miao Xie0b2f6302008-07-25 01:47:21 -0700839 *
Tejun Heo5d21cc22013-01-07 08:51:08 -0800840 * Called with cpuset_mutex held
Miao Xie0b2f6302008-07-25 01:47:21 -0700841 *
842 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
843 * calling callback functions for each.
844 *
Li Zefan4e743392008-09-13 02:33:08 -0700845 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
846 * if @heap != NULL.
Miao Xie0b2f6302008-07-25 01:47:21 -0700847 */
Li Zefan4e743392008-09-13 02:33:08 -0700848static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
Miao Xie0b2f6302008-07-25 01:47:21 -0700849{
850 struct cgroup_scanner scan;
Miao Xie0b2f6302008-07-25 01:47:21 -0700851
852 scan.cg = cs->css.cgroup;
853 scan.test_task = cpuset_test_cpumask;
854 scan.process_task = cpuset_change_cpumask;
Li Zefan4e743392008-09-13 02:33:08 -0700855 scan.heap = heap;
856 cgroup_scan_tasks(&scan);
Miao Xie0b2f6302008-07-25 01:47:21 -0700857}
858
859/**
Cliff Wickman58f47902008-02-07 00:14:44 -0800860 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
861 * @cs: the cpuset to consider
862 * @buf: buffer of cpu numbers written to this cpuset
863 */
Li Zefan645fcc92009-01-07 18:08:43 -0800864static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
865 const char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866{
Li Zefan4e743392008-09-13 02:33:08 -0700867 struct ptr_heap heap;
Cliff Wickman58f47902008-02-07 00:14:44 -0800868 int retval;
869 int is_load_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
Rusty Russell5f054e32012-03-29 15:38:31 +1030871 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
Paul Jackson4c4d50f2006-08-27 01:23:51 -0700872 if (cs == &top_cpuset)
873 return -EACCES;
874
David Rientjes6f7f02e2007-05-08 00:31:43 -0700875 /*
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800876 * An empty cpus_allowed is ok only if the cpuset has no tasks.
Paul Jackson020958b2007-10-18 23:40:21 -0700877 * Since cpulist_parse() fails on an empty mask, we special case
878 * that parsing. The validate_change() call ensures that cpusets
879 * with tasks have cpus.
David Rientjes6f7f02e2007-05-08 00:31:43 -0700880 */
Paul Jackson020958b2007-10-18 23:40:21 -0700881 if (!*buf) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800882 cpumask_clear(trialcs->cpus_allowed);
David Rientjes6f7f02e2007-05-08 00:31:43 -0700883 } else {
Li Zefan300ed6c2009-01-07 18:08:44 -0800884 retval = cpulist_parse(buf, trialcs->cpus_allowed);
David Rientjes6f7f02e2007-05-08 00:31:43 -0700885 if (retval < 0)
886 return retval;
Lai Jiangshan37340742008-06-05 22:46:32 -0700887
Peter Zijlstra6ad4c182009-11-25 13:31:39 +0100888 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
Lai Jiangshan37340742008-06-05 22:46:32 -0700889 return -EINVAL;
David Rientjes6f7f02e2007-05-08 00:31:43 -0700890 }
Li Zefan645fcc92009-01-07 18:08:43 -0800891 retval = validate_change(cs, trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700892 if (retval < 0)
893 return retval;
Paul Jackson029190c2007-10-18 23:40:20 -0700894
Paul Menage8707d8b2007-10-18 23:40:22 -0700895 /* Nothing to do if the cpus didn't change */
Li Zefan300ed6c2009-01-07 18:08:44 -0800896 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
Paul Menage8707d8b2007-10-18 23:40:22 -0700897 return 0;
Cliff Wickman58f47902008-02-07 00:14:44 -0800898
Li Zefan4e743392008-09-13 02:33:08 -0700899 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
900 if (retval)
901 return retval;
902
Li Zefan645fcc92009-01-07 18:08:43 -0800903 is_load_balanced = is_sched_load_balance(trialcs);
Paul Jackson029190c2007-10-18 23:40:20 -0700904
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800905 mutex_lock(&callback_mutex);
Li Zefan300ed6c2009-01-07 18:08:44 -0800906 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800907 mutex_unlock(&callback_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -0700908
Paul Menage8707d8b2007-10-18 23:40:22 -0700909 /*
910 * Scan tasks in the cpuset, and update the cpumasks of any
Cliff Wickman58f47902008-02-07 00:14:44 -0800911 * that need an update.
Paul Menage8707d8b2007-10-18 23:40:22 -0700912 */
Li Zefan4e743392008-09-13 02:33:08 -0700913 update_tasks_cpumask(cs, &heap);
914
915 heap_free(&heap);
Cliff Wickman58f47902008-02-07 00:14:44 -0800916
Paul Menage8707d8b2007-10-18 23:40:22 -0700917 if (is_load_balanced)
Tejun Heo699140b2013-01-07 08:51:07 -0800918 rebuild_sched_domains_locked();
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700919 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920}
921
Paul Jackson053199e2005-10-30 15:02:30 -0800922/*
Paul Jacksone4e364e2006-03-31 02:30:52 -0800923 * cpuset_migrate_mm
924 *
925 * Migrate memory region from one set of nodes to another.
926 *
927 * Temporarilly set tasks mems_allowed to target nodes of migration,
928 * so that the migration code can allocate pages on these nodes.
929 *
Tejun Heo5d21cc22013-01-07 08:51:08 -0800930 * Call holding cpuset_mutex, so current's cpuset won't change
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800931 * during this call, as manage_mutex holds off any cpuset_attach()
Paul Jacksone4e364e2006-03-31 02:30:52 -0800932 * calls. Therefore we don't need to take task_lock around the
933 * call to guarantee_online_mems(), as we know no one is changing
Paul Menage2df167a2008-02-07 00:14:45 -0800934 * our task's cpuset.
Paul Jacksone4e364e2006-03-31 02:30:52 -0800935 *
Paul Jacksone4e364e2006-03-31 02:30:52 -0800936 * While the mm_struct we are migrating is typically from some
937 * other task, the task_struct mems_allowed that we are hacking
938 * is for our current task, which must allocate new pages for that
939 * migrating memory region.
Paul Jacksone4e364e2006-03-31 02:30:52 -0800940 */
941
942static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
943 const nodemask_t *to)
944{
945 struct task_struct *tsk = current;
946
Paul Jacksone4e364e2006-03-31 02:30:52 -0800947 tsk->mems_allowed = *to;
Paul Jacksone4e364e2006-03-31 02:30:52 -0800948
949 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
950
Paul Menage8793d852007-10-18 23:39:39 -0700951 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
Paul Jacksone4e364e2006-03-31 02:30:52 -0800952}
953
Li Zefan3b6766f2009-04-02 16:57:51 -0700954/*
Miao Xie58568d22009-06-16 15:31:49 -0700955 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
956 * @tsk: the task to change
957 * @newmems: new nodes that the task will be set
958 *
959 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
960 * we structure updates as setting all new allowed nodes, then clearing newly
961 * disallowed ones.
Miao Xie58568d22009-06-16 15:31:49 -0700962 */
963static void cpuset_change_task_nodemask(struct task_struct *tsk,
964 nodemask_t *newmems)
965{
David Rientjesb2462722011-12-19 17:11:52 -0800966 bool need_loop;
David Rientjes89e8a242011-11-02 13:38:39 -0700967
Miao Xiec0ff7452010-05-24 14:32:08 -0700968 /*
969 * Allow tasks that have access to memory reserves because they have
970 * been OOM killed to get memory anywhere.
971 */
972 if (unlikely(test_thread_flag(TIF_MEMDIE)))
973 return;
974 if (current->flags & PF_EXITING) /* Let dying task have memory */
975 return;
976
977 task_lock(tsk);
David Rientjesb2462722011-12-19 17:11:52 -0800978 /*
979 * Determine if a loop is necessary if another thread is doing
980 * get_mems_allowed(). If at least one node remains unchanged and
981 * tsk does not have a mempolicy, then an empty nodemask will not be
982 * possible when mems_allowed is larger than a word.
983 */
984 need_loop = task_has_mempolicy(tsk) ||
985 !nodes_intersects(*newmems, tsk->mems_allowed);
Mel Gormancc9a6c82012-03-21 16:34:11 -0700986
987 if (need_loop)
988 write_seqcount_begin(&tsk->mems_allowed_seq);
989
Miao Xie58568d22009-06-16 15:31:49 -0700990 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
Miao Xiec0ff7452010-05-24 14:32:08 -0700991 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
992
Miao Xiec0ff7452010-05-24 14:32:08 -0700993 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
Miao Xie58568d22009-06-16 15:31:49 -0700994 tsk->mems_allowed = *newmems;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700995
996 if (need_loop)
997 write_seqcount_end(&tsk->mems_allowed_seq);
998
Miao Xiec0ff7452010-05-24 14:32:08 -0700999 task_unlock(tsk);
Miao Xie58568d22009-06-16 15:31:49 -07001000}
1001
1002/*
1003 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1004 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
Tejun Heo5d21cc22013-01-07 08:51:08 -08001005 * memory_migrate flag is set. Called with cpuset_mutex held.
Li Zefan3b6766f2009-04-02 16:57:51 -07001006 */
1007static void cpuset_change_nodemask(struct task_struct *p,
1008 struct cgroup_scanner *scan)
1009{
1010 struct mm_struct *mm;
1011 struct cpuset *cs;
1012 int migrate;
1013 const nodemask_t *oldmem = scan->data;
Tejun Heo5d21cc22013-01-07 08:51:08 -08001014 static nodemask_t newmems; /* protected by cpuset_mutex */
Miao Xie58568d22009-06-16 15:31:49 -07001015
1016 cs = cgroup_cs(scan->cg);
Li Zefanee24d372011-03-23 16:42:47 -07001017 guarantee_online_mems(cs, &newmems);
Miao Xie58568d22009-06-16 15:31:49 -07001018
Li Zefanee24d372011-03-23 16:42:47 -07001019 cpuset_change_task_nodemask(p, &newmems);
Miao Xie53feb292010-03-23 13:35:35 -07001020
Li Zefan3b6766f2009-04-02 16:57:51 -07001021 mm = get_task_mm(p);
1022 if (!mm)
1023 return;
1024
Li Zefan3b6766f2009-04-02 16:57:51 -07001025 migrate = is_memory_migrate(cs);
1026
1027 mpol_rebind_mm(mm, &cs->mems_allowed);
1028 if (migrate)
1029 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1030 mmput(mm);
1031}
1032
Paul Menage8793d852007-10-18 23:39:39 -07001033static void *cpuset_being_rebound;
1034
Miao Xie0b2f6302008-07-25 01:47:21 -07001035/**
1036 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1037 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1038 * @oldmem: old mems_allowed of cpuset cs
Li Zefan010cfac2009-04-02 16:57:52 -07001039 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
Miao Xie0b2f6302008-07-25 01:47:21 -07001040 *
Tejun Heo5d21cc22013-01-07 08:51:08 -08001041 * Called with cpuset_mutex held
Li Zefan010cfac2009-04-02 16:57:52 -07001042 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1043 * if @heap != NULL.
Miao Xie0b2f6302008-07-25 01:47:21 -07001044 */
Li Zefan010cfac2009-04-02 16:57:52 -07001045static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1046 struct ptr_heap *heap)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047{
Li Zefan3b6766f2009-04-02 16:57:51 -07001048 struct cgroup_scanner scan;
Paul Jackson59dac162006-01-08 01:01:52 -08001049
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001050 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
Paul Jackson42253992006-01-08 01:01:59 -08001051
Li Zefan3b6766f2009-04-02 16:57:51 -07001052 scan.cg = cs->css.cgroup;
1053 scan.test_task = NULL;
1054 scan.process_task = cpuset_change_nodemask;
Li Zefan010cfac2009-04-02 16:57:52 -07001055 scan.heap = heap;
Li Zefan3b6766f2009-04-02 16:57:51 -07001056 scan.data = (nodemask_t *)oldmem;
Paul Jackson42253992006-01-08 01:01:59 -08001057
1058 /*
Li Zefan3b6766f2009-04-02 16:57:51 -07001059 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1060 * take while holding tasklist_lock. Forks can happen - the
1061 * mpol_dup() cpuset_being_rebound check will catch such forks,
1062 * and rebind their vma mempolicies too. Because we still hold
Tejun Heo5d21cc22013-01-07 08:51:08 -08001063 * the global cpuset_mutex, we know that no other rebind effort
Li Zefan3b6766f2009-04-02 16:57:51 -07001064 * will be contending for the global variable cpuset_being_rebound.
Paul Jackson42253992006-01-08 01:01:59 -08001065 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
Paul Jackson04c19fa2006-01-08 01:02:00 -08001066 * is idempotent. Also migrate pages in each mm to new nodes.
Paul Jackson42253992006-01-08 01:01:59 -08001067 */
Li Zefan010cfac2009-04-02 16:57:52 -07001068 cgroup_scan_tasks(&scan);
Paul Jackson42253992006-01-08 01:01:59 -08001069
Paul Menage2df167a2008-02-07 00:14:45 -08001070 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
Paul Menage8793d852007-10-18 23:39:39 -07001071 cpuset_being_rebound = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072}
1073
Miao Xie0b2f6302008-07-25 01:47:21 -07001074/*
1075 * Handle user request to change the 'mems' memory placement
1076 * of a cpuset. Needs to validate the request, update the
Miao Xie58568d22009-06-16 15:31:49 -07001077 * cpusets mems_allowed, and for each task in the cpuset,
1078 * update mems_allowed and rebind task's mempolicy and any vma
1079 * mempolicies and if the cpuset is marked 'memory_migrate',
1080 * migrate the tasks pages to the new memory.
Miao Xie0b2f6302008-07-25 01:47:21 -07001081 *
Tejun Heo5d21cc22013-01-07 08:51:08 -08001082 * Call with cpuset_mutex held. May take callback_mutex during call.
Miao Xie0b2f6302008-07-25 01:47:21 -07001083 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1084 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1085 * their mempolicies to the cpusets new mems_allowed.
1086 */
Li Zefan645fcc92009-01-07 18:08:43 -08001087static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1088 const char *buf)
Miao Xie0b2f6302008-07-25 01:47:21 -07001089{
Miao Xie53feb292010-03-23 13:35:35 -07001090 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
Miao Xie0b2f6302008-07-25 01:47:21 -07001091 int retval;
Li Zefan010cfac2009-04-02 16:57:52 -07001092 struct ptr_heap heap;
Miao Xie0b2f6302008-07-25 01:47:21 -07001093
Miao Xie53feb292010-03-23 13:35:35 -07001094 if (!oldmem)
1095 return -ENOMEM;
1096
Miao Xie0b2f6302008-07-25 01:47:21 -07001097 /*
Lai Jiangshan38d7bee2012-12-12 13:51:24 -08001098 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
Miao Xie0b2f6302008-07-25 01:47:21 -07001099 * it's read-only
1100 */
Miao Xie53feb292010-03-23 13:35:35 -07001101 if (cs == &top_cpuset) {
1102 retval = -EACCES;
1103 goto done;
1104 }
Miao Xie0b2f6302008-07-25 01:47:21 -07001105
Miao Xie0b2f6302008-07-25 01:47:21 -07001106 /*
1107 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1108 * Since nodelist_parse() fails on an empty mask, we special case
1109 * that parsing. The validate_change() call ensures that cpusets
1110 * with tasks have memory.
1111 */
1112 if (!*buf) {
Li Zefan645fcc92009-01-07 18:08:43 -08001113 nodes_clear(trialcs->mems_allowed);
Miao Xie0b2f6302008-07-25 01:47:21 -07001114 } else {
Li Zefan645fcc92009-01-07 18:08:43 -08001115 retval = nodelist_parse(buf, trialcs->mems_allowed);
Miao Xie0b2f6302008-07-25 01:47:21 -07001116 if (retval < 0)
1117 goto done;
1118
Li Zefan645fcc92009-01-07 18:08:43 -08001119 if (!nodes_subset(trialcs->mems_allowed,
Lai Jiangshan38d7bee2012-12-12 13:51:24 -08001120 node_states[N_MEMORY])) {
Miao Xie53feb292010-03-23 13:35:35 -07001121 retval = -EINVAL;
1122 goto done;
1123 }
Miao Xie0b2f6302008-07-25 01:47:21 -07001124 }
Miao Xie53feb292010-03-23 13:35:35 -07001125 *oldmem = cs->mems_allowed;
1126 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
Miao Xie0b2f6302008-07-25 01:47:21 -07001127 retval = 0; /* Too easy - nothing to do */
1128 goto done;
1129 }
Li Zefan645fcc92009-01-07 18:08:43 -08001130 retval = validate_change(cs, trialcs);
Miao Xie0b2f6302008-07-25 01:47:21 -07001131 if (retval < 0)
1132 goto done;
1133
Li Zefan010cfac2009-04-02 16:57:52 -07001134 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1135 if (retval < 0)
1136 goto done;
1137
Miao Xie0b2f6302008-07-25 01:47:21 -07001138 mutex_lock(&callback_mutex);
Li Zefan645fcc92009-01-07 18:08:43 -08001139 cs->mems_allowed = trialcs->mems_allowed;
Miao Xie0b2f6302008-07-25 01:47:21 -07001140 mutex_unlock(&callback_mutex);
1141
Miao Xie53feb292010-03-23 13:35:35 -07001142 update_tasks_nodemask(cs, oldmem, &heap);
Li Zefan010cfac2009-04-02 16:57:52 -07001143
1144 heap_free(&heap);
Miao Xie0b2f6302008-07-25 01:47:21 -07001145done:
Miao Xie53feb292010-03-23 13:35:35 -07001146 NODEMASK_FREE(oldmem);
Miao Xie0b2f6302008-07-25 01:47:21 -07001147 return retval;
1148}
1149
Paul Menage8793d852007-10-18 23:39:39 -07001150int current_cpuset_is_being_rebound(void)
1151{
1152 return task_cs(current) == cpuset_being_rebound;
1153}
1154
Paul Menage5be7a472008-05-06 20:42:41 -07001155static int update_relax_domain_level(struct cpuset *cs, s64 val)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001156{
Paul Menagedb7f47c2009-04-02 16:57:55 -07001157#ifdef CONFIG_SMP
Peter Zijlstra60495e72011-04-07 14:10:04 +02001158 if (val < -1 || val >= sched_domain_level_max)
Li Zefan30e0e172008-05-13 10:27:17 +08001159 return -EINVAL;
Paul Menagedb7f47c2009-04-02 16:57:55 -07001160#endif
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001161
1162 if (val != cs->relax_domain_level) {
1163 cs->relax_domain_level = val;
Li Zefan300ed6c2009-01-07 18:08:44 -08001164 if (!cpumask_empty(cs->cpus_allowed) &&
1165 is_sched_load_balance(cs))
Tejun Heo699140b2013-01-07 08:51:07 -08001166 rebuild_sched_domains_locked();
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001167 }
1168
1169 return 0;
1170}
1171
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001172/*
Miao Xie950592f2009-06-16 15:31:47 -07001173 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1174 * @tsk: task to be updated
1175 * @scan: struct cgroup_scanner containing the cgroup of the task
1176 *
1177 * Called by cgroup_scan_tasks() for each task in a cgroup.
1178 *
1179 * We don't need to re-check for the cgroup/cpuset membership, since we're
Tejun Heo5d21cc22013-01-07 08:51:08 -08001180 * holding cpuset_mutex at this point.
Miao Xie950592f2009-06-16 15:31:47 -07001181 */
1182static void cpuset_change_flag(struct task_struct *tsk,
1183 struct cgroup_scanner *scan)
1184{
1185 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1186}
1187
1188/*
1189 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1190 * @cs: the cpuset in which each task's spread flags needs to be changed
1191 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1192 *
Tejun Heo5d21cc22013-01-07 08:51:08 -08001193 * Called with cpuset_mutex held
Miao Xie950592f2009-06-16 15:31:47 -07001194 *
1195 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1196 * calling callback functions for each.
1197 *
1198 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1199 * if @heap != NULL.
1200 */
1201static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1202{
1203 struct cgroup_scanner scan;
1204
1205 scan.cg = cs->css.cgroup;
1206 scan.test_task = NULL;
1207 scan.process_task = cpuset_change_flag;
1208 scan.heap = heap;
1209 cgroup_scan_tasks(&scan);
1210}
1211
1212/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 * update_flag - read a 0 or a 1 in a file and update associated flag
Paul Menage78608362008-04-29 01:00:26 -07001214 * bit: the bit to update (see cpuset_flagbits_t)
1215 * cs: the cpuset to update
1216 * turning_on: whether the flag is being set or cleared
Paul Jackson053199e2005-10-30 15:02:30 -08001217 *
Tejun Heo5d21cc22013-01-07 08:51:08 -08001218 * Call with cpuset_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 */
1220
Paul Menage700fe1a2008-04-29 01:00:00 -07001221static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1222 int turning_on)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
Li Zefan645fcc92009-01-07 18:08:43 -08001224 struct cpuset *trialcs;
Rakib Mullick40b6a762008-10-18 20:28:18 -07001225 int balance_flag_changed;
Miao Xie950592f2009-06-16 15:31:47 -07001226 int spread_flag_changed;
1227 struct ptr_heap heap;
1228 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229
Li Zefan645fcc92009-01-07 18:08:43 -08001230 trialcs = alloc_trial_cpuset(cs);
1231 if (!trialcs)
1232 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233
Li Zefan645fcc92009-01-07 18:08:43 -08001234 if (turning_on)
1235 set_bit(bit, &trialcs->flags);
1236 else
1237 clear_bit(bit, &trialcs->flags);
1238
1239 err = validate_change(cs, trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001240 if (err < 0)
Li Zefan645fcc92009-01-07 18:08:43 -08001241 goto out;
Paul Jackson029190c2007-10-18 23:40:20 -07001242
Miao Xie950592f2009-06-16 15:31:47 -07001243 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1244 if (err < 0)
1245 goto out;
1246
Paul Jackson029190c2007-10-18 23:40:20 -07001247 balance_flag_changed = (is_sched_load_balance(cs) !=
Li Zefan645fcc92009-01-07 18:08:43 -08001248 is_sched_load_balance(trialcs));
Paul Jackson029190c2007-10-18 23:40:20 -07001249
Miao Xie950592f2009-06-16 15:31:47 -07001250 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1251 || (is_spread_page(cs) != is_spread_page(trialcs)));
1252
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001253 mutex_lock(&callback_mutex);
Li Zefan645fcc92009-01-07 18:08:43 -08001254 cs->flags = trialcs->flags;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001255 mutex_unlock(&callback_mutex);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001256
Li Zefan300ed6c2009-01-07 18:08:44 -08001257 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
Tejun Heo699140b2013-01-07 08:51:07 -08001258 rebuild_sched_domains_locked();
Paul Jackson029190c2007-10-18 23:40:20 -07001259
Miao Xie950592f2009-06-16 15:31:47 -07001260 if (spread_flag_changed)
1261 update_tasks_flags(cs, &heap);
1262 heap_free(&heap);
Li Zefan645fcc92009-01-07 18:08:43 -08001263out:
1264 free_trial_cpuset(trialcs);
1265 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266}
1267
Paul Jackson053199e2005-10-30 15:02:30 -08001268/*
Adrian Bunk80f72282006-06-30 18:27:16 +02001269 * Frequency meter - How fast is some event occurring?
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001270 *
1271 * These routines manage a digitally filtered, constant time based,
1272 * event frequency meter. There are four routines:
1273 * fmeter_init() - initialize a frequency meter.
1274 * fmeter_markevent() - called each time the event happens.
1275 * fmeter_getrate() - returns the recent rate of such events.
1276 * fmeter_update() - internal routine used to update fmeter.
1277 *
1278 * A common data structure is passed to each of these routines,
1279 * which is used to keep track of the state required to manage the
1280 * frequency meter and its digital filter.
1281 *
1282 * The filter works on the number of events marked per unit time.
1283 * The filter is single-pole low-pass recursive (IIR). The time unit
1284 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1285 * simulate 3 decimal digits of precision (multiplied by 1000).
1286 *
1287 * With an FM_COEF of 933, and a time base of 1 second, the filter
1288 * has a half-life of 10 seconds, meaning that if the events quit
1289 * happening, then the rate returned from the fmeter_getrate()
1290 * will be cut in half each 10 seconds, until it converges to zero.
1291 *
1292 * It is not worth doing a real infinitely recursive filter. If more
1293 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1294 * just compute FM_MAXTICKS ticks worth, by which point the level
1295 * will be stable.
1296 *
1297 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1298 * arithmetic overflow in the fmeter_update() routine.
1299 *
1300 * Given the simple 32 bit integer arithmetic used, this meter works
1301 * best for reporting rates between one per millisecond (msec) and
1302 * one per 32 (approx) seconds. At constant rates faster than one
1303 * per msec it maxes out at values just under 1,000,000. At constant
1304 * rates between one per msec, and one per second it will stabilize
1305 * to a value N*1000, where N is the rate of events per second.
1306 * At constant rates between one per second and one per 32 seconds,
1307 * it will be choppy, moving up on the seconds that have an event,
1308 * and then decaying until the next event. At rates slower than
1309 * about one in 32 seconds, it decays all the way back to zero between
1310 * each event.
1311 */
1312
1313#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1314#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1315#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1316#define FM_SCALE 1000 /* faux fixed point scale */
1317
1318/* Initialize a frequency meter */
1319static void fmeter_init(struct fmeter *fmp)
1320{
1321 fmp->cnt = 0;
1322 fmp->val = 0;
1323 fmp->time = 0;
1324 spin_lock_init(&fmp->lock);
1325}
1326
1327/* Internal meter update - process cnt events and update value */
1328static void fmeter_update(struct fmeter *fmp)
1329{
1330 time_t now = get_seconds();
1331 time_t ticks = now - fmp->time;
1332
1333 if (ticks == 0)
1334 return;
1335
1336 ticks = min(FM_MAXTICKS, ticks);
1337 while (ticks-- > 0)
1338 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1339 fmp->time = now;
1340
1341 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1342 fmp->cnt = 0;
1343}
1344
1345/* Process any previous ticks, then bump cnt by one (times scale). */
1346static void fmeter_markevent(struct fmeter *fmp)
1347{
1348 spin_lock(&fmp->lock);
1349 fmeter_update(fmp);
1350 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1351 spin_unlock(&fmp->lock);
1352}
1353
1354/* Process any previous ticks, then return current value. */
1355static int fmeter_getrate(struct fmeter *fmp)
1356{
1357 int val;
1358
1359 spin_lock(&fmp->lock);
1360 fmeter_update(fmp);
1361 val = fmp->val;
1362 spin_unlock(&fmp->lock);
1363 return val;
1364}
1365
Tejun Heo5d21cc22013-01-07 08:51:08 -08001366/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
Li Zefan761b3ef2012-01-31 13:47:36 +08001367static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
Ben Blumf780bdb2011-05-26 16:25:19 -07001368{
Tejun Heo2f7ee562011-12-12 18:12:21 -08001369 struct cpuset *cs = cgroup_cs(cgrp);
Tejun Heobb9d97b2011-12-12 18:12:21 -08001370 struct task_struct *task;
1371 int ret;
Ben Blumf780bdb2011-05-26 16:25:19 -07001372
Tejun Heo5d21cc22013-01-07 08:51:08 -08001373 mutex_lock(&cpuset_mutex);
1374
1375 ret = -ENOSPC;
Ben Blumbe367d02009-09-23 15:56:31 -07001376 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
Tejun Heo5d21cc22013-01-07 08:51:08 -08001377 goto out_unlock;
Ben Blumbe367d02009-09-23 15:56:31 -07001378
Tejun Heobb9d97b2011-12-12 18:12:21 -08001379 cgroup_taskset_for_each(task, cgrp, tset) {
1380 /*
Tejun Heo14a40ff2013-03-19 13:45:20 -07001381 * Kthreads which disallow setaffinity shouldn't be moved
1382 * to a new cpuset; we don't want to change their cpu
1383 * affinity and isolating such threads by their set of
1384 * allowed nodes is unnecessary. Thus, cpusets are not
1385 * applicable for such threads. This prevents checking for
1386 * success of set_cpus_allowed_ptr() on all attached tasks
1387 * before cpus_allowed may be changed.
Tejun Heobb9d97b2011-12-12 18:12:21 -08001388 */
Tejun Heo5d21cc22013-01-07 08:51:08 -08001389 ret = -EINVAL;
Tejun Heo14a40ff2013-03-19 13:45:20 -07001390 if (task->flags & PF_NO_SETAFFINITY)
Tejun Heo5d21cc22013-01-07 08:51:08 -08001391 goto out_unlock;
1392 ret = security_task_setscheduler(task);
1393 if (ret)
1394 goto out_unlock;
Tejun Heobb9d97b2011-12-12 18:12:21 -08001395 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
Tejun Heo452477f2013-01-07 08:51:07 -08001397 /*
1398 * Mark attach is in progress. This makes validate_change() fail
1399 * changes which zero cpus/mems_allowed.
1400 */
1401 cs->attach_in_progress++;
Tejun Heo5d21cc22013-01-07 08:51:08 -08001402 ret = 0;
1403out_unlock:
1404 mutex_unlock(&cpuset_mutex);
1405 return ret;
Tejun Heo4e4c9a12013-01-07 08:51:07 -08001406}
1407
Tejun Heo452477f2013-01-07 08:51:07 -08001408static void cpuset_cancel_attach(struct cgroup *cgrp,
1409 struct cgroup_taskset *tset)
1410{
Tejun Heo5d21cc22013-01-07 08:51:08 -08001411 mutex_lock(&cpuset_mutex);
Tejun Heo452477f2013-01-07 08:51:07 -08001412 cgroup_cs(cgrp)->attach_in_progress--;
Tejun Heo5d21cc22013-01-07 08:51:08 -08001413 mutex_unlock(&cpuset_mutex);
Tejun Heo452477f2013-01-07 08:51:07 -08001414}
1415
Tejun Heo4e4c9a12013-01-07 08:51:07 -08001416/*
Tejun Heo5d21cc22013-01-07 08:51:08 -08001417 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
Tejun Heo4e4c9a12013-01-07 08:51:07 -08001418 * but we can't allocate it dynamically there. Define it global and
1419 * allocate from cpuset_init().
1420 */
1421static cpumask_var_t cpus_attach;
1422
1423static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1424{
Tejun Heo5d21cc22013-01-07 08:51:08 -08001425 /* static bufs protected by cpuset_mutex */
Tejun Heo4e4c9a12013-01-07 08:51:07 -08001426 static nodemask_t cpuset_attach_nodemask_from;
1427 static nodemask_t cpuset_attach_nodemask_to;
1428 struct mm_struct *mm;
1429 struct task_struct *task;
1430 struct task_struct *leader = cgroup_taskset_first(tset);
1431 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1432 struct cpuset *cs = cgroup_cs(cgrp);
1433 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1434
Tejun Heo5d21cc22013-01-07 08:51:08 -08001435 mutex_lock(&cpuset_mutex);
1436
Tejun Heo94196f52011-12-12 18:12:22 -08001437 /* prepare for attach */
Ben Blumf780bdb2011-05-26 16:25:19 -07001438 if (cs == &top_cpuset)
1439 cpumask_copy(cpus_attach, cpu_possible_mask);
1440 else
1441 guarantee_online_cpus(cs, cpus_attach);
1442
1443 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
Tejun Heo94196f52011-12-12 18:12:22 -08001444
Tejun Heobb9d97b2011-12-12 18:12:21 -08001445 cgroup_taskset_for_each(task, cgrp, tset) {
1446 /*
1447 * can_attach beforehand should guarantee that this doesn't
1448 * fail. TODO: have a better way to handle failure here
1449 */
1450 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1451
1452 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1453 cpuset_update_task_spread_flag(cs, task);
1454 }
David Quigley22fb52d2006-06-23 02:04:00 -07001455
Ben Blumf780bdb2011-05-26 16:25:19 -07001456 /*
1457 * Change mm, possibly for multiple threads in a threadgroup. This is
1458 * expensive and may sleep.
1459 */
1460 cpuset_attach_nodemask_from = oldcs->mems_allowed;
1461 cpuset_attach_nodemask_to = cs->mems_allowed;
Tejun Heobb9d97b2011-12-12 18:12:21 -08001462 mm = get_task_mm(leader);
Paul Jackson42253992006-01-08 01:01:59 -08001463 if (mm) {
Ben Blumf780bdb2011-05-26 16:25:19 -07001464 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
Paul Jackson2741a552006-03-31 02:30:51 -08001465 if (is_memory_migrate(cs))
Ben Blumf780bdb2011-05-26 16:25:19 -07001466 cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1467 &cpuset_attach_nodemask_to);
Paul Jackson42253992006-01-08 01:01:59 -08001468 mmput(mm);
1469 }
Tejun Heo452477f2013-01-07 08:51:07 -08001470
1471 cs->attach_in_progress--;
Tejun Heo02bb5862013-01-07 08:51:08 -08001472
1473 /*
1474 * We may have raced with CPU/memory hotunplug. Trigger hotplug
1475 * propagation if @cs doesn't have any CPU or memory. It will move
1476 * the newly added tasks to the nearest parent which can execute.
1477 */
1478 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1479 schedule_cpuset_propagate_hotplug(cs);
Tejun Heo5d21cc22013-01-07 08:51:08 -08001480
1481 mutex_unlock(&cpuset_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482}
1483
1484/* The various types of files and directories in a cpuset file system */
1485
1486typedef enum {
Paul Jackson45b07ef2006-01-08 01:00:56 -08001487 FILE_MEMORY_MIGRATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 FILE_CPULIST,
1489 FILE_MEMLIST,
1490 FILE_CPU_EXCLUSIVE,
1491 FILE_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -07001492 FILE_MEM_HARDWALL,
Paul Jackson029190c2007-10-18 23:40:20 -07001493 FILE_SCHED_LOAD_BALANCE,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001494 FILE_SCHED_RELAX_DOMAIN_LEVEL,
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001495 FILE_MEMORY_PRESSURE_ENABLED,
1496 FILE_MEMORY_PRESSURE,
Paul Jackson825a46a2006-03-24 03:16:03 -08001497 FILE_SPREAD_PAGE,
1498 FILE_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499} cpuset_filetype_t;
1500
Paul Menage700fe1a2008-04-29 01:00:00 -07001501static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1502{
Paul Menage700fe1a2008-04-29 01:00:00 -07001503 struct cpuset *cs = cgroup_cs(cgrp);
1504 cpuset_filetype_t type = cft->private;
Tejun Heo5d21cc22013-01-07 08:51:08 -08001505 int retval = -ENODEV;
Paul Menage700fe1a2008-04-29 01:00:00 -07001506
Tejun Heo5d21cc22013-01-07 08:51:08 -08001507 mutex_lock(&cpuset_mutex);
1508 if (!is_cpuset_online(cs))
1509 goto out_unlock;
Paul Menage700fe1a2008-04-29 01:00:00 -07001510
1511 switch (type) {
1512 case FILE_CPU_EXCLUSIVE:
1513 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1514 break;
1515 case FILE_MEM_EXCLUSIVE:
1516 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1517 break;
Paul Menage78608362008-04-29 01:00:26 -07001518 case FILE_MEM_HARDWALL:
1519 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1520 break;
Paul Menage700fe1a2008-04-29 01:00:00 -07001521 case FILE_SCHED_LOAD_BALANCE:
1522 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1523 break;
1524 case FILE_MEMORY_MIGRATE:
1525 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1526 break;
1527 case FILE_MEMORY_PRESSURE_ENABLED:
1528 cpuset_memory_pressure_enabled = !!val;
1529 break;
1530 case FILE_MEMORY_PRESSURE:
1531 retval = -EACCES;
1532 break;
1533 case FILE_SPREAD_PAGE:
1534 retval = update_flag(CS_SPREAD_PAGE, cs, val);
Paul Menage700fe1a2008-04-29 01:00:00 -07001535 break;
1536 case FILE_SPREAD_SLAB:
1537 retval = update_flag(CS_SPREAD_SLAB, cs, val);
Paul Menage700fe1a2008-04-29 01:00:00 -07001538 break;
1539 default:
1540 retval = -EINVAL;
1541 break;
1542 }
Tejun Heo5d21cc22013-01-07 08:51:08 -08001543out_unlock:
1544 mutex_unlock(&cpuset_mutex);
Paul Menage700fe1a2008-04-29 01:00:00 -07001545 return retval;
1546}
1547
Paul Menage5be7a472008-05-06 20:42:41 -07001548static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1549{
Paul Menage5be7a472008-05-06 20:42:41 -07001550 struct cpuset *cs = cgroup_cs(cgrp);
1551 cpuset_filetype_t type = cft->private;
Tejun Heo5d21cc22013-01-07 08:51:08 -08001552 int retval = -ENODEV;
Paul Menage5be7a472008-05-06 20:42:41 -07001553
Tejun Heo5d21cc22013-01-07 08:51:08 -08001554 mutex_lock(&cpuset_mutex);
1555 if (!is_cpuset_online(cs))
1556 goto out_unlock;
Paul Menagee3712392008-07-25 01:47:02 -07001557
Paul Menage5be7a472008-05-06 20:42:41 -07001558 switch (type) {
1559 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1560 retval = update_relax_domain_level(cs, val);
1561 break;
1562 default:
1563 retval = -EINVAL;
1564 break;
1565 }
Tejun Heo5d21cc22013-01-07 08:51:08 -08001566out_unlock:
1567 mutex_unlock(&cpuset_mutex);
Paul Menage5be7a472008-05-06 20:42:41 -07001568 return retval;
1569}
1570
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571/*
Paul Menagee3712392008-07-25 01:47:02 -07001572 * Common handling for a write to a "cpus" or "mems" file.
1573 */
1574static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1575 const char *buf)
1576{
Li Zefan645fcc92009-01-07 18:08:43 -08001577 struct cpuset *cs = cgroup_cs(cgrp);
1578 struct cpuset *trialcs;
Tejun Heo5d21cc22013-01-07 08:51:08 -08001579 int retval = -ENODEV;
Paul Menagee3712392008-07-25 01:47:02 -07001580
Tejun Heo3a5a6d02013-01-07 08:51:07 -08001581 /*
1582 * CPU or memory hotunplug may leave @cs w/o any execution
1583 * resources, in which case the hotplug code asynchronously updates
1584 * configuration and transfers all tasks to the nearest ancestor
1585 * which can execute.
1586 *
1587 * As writes to "cpus" or "mems" may restore @cs's execution
1588 * resources, wait for the previously scheduled operations before
1589 * proceeding, so that we don't end up keep removing tasks added
1590 * after execution capability is restored.
Tejun Heo02bb5862013-01-07 08:51:08 -08001591 *
1592 * Flushing cpuset_hotplug_work is enough to synchronize against
1593 * hotplug hanlding; however, cpuset_attach() may schedule
1594 * propagation work directly. Flush the workqueue too.
Tejun Heo3a5a6d02013-01-07 08:51:07 -08001595 */
1596 flush_work(&cpuset_hotplug_work);
Tejun Heo02bb5862013-01-07 08:51:08 -08001597 flush_workqueue(cpuset_propagate_hotplug_wq);
Tejun Heo3a5a6d02013-01-07 08:51:07 -08001598
Tejun Heo5d21cc22013-01-07 08:51:08 -08001599 mutex_lock(&cpuset_mutex);
1600 if (!is_cpuset_online(cs))
1601 goto out_unlock;
Paul Menagee3712392008-07-25 01:47:02 -07001602
Li Zefan645fcc92009-01-07 18:08:43 -08001603 trialcs = alloc_trial_cpuset(cs);
Li Zefanb75f38d2011-03-04 17:36:21 -08001604 if (!trialcs) {
1605 retval = -ENOMEM;
Tejun Heo5d21cc22013-01-07 08:51:08 -08001606 goto out_unlock;
Li Zefanb75f38d2011-03-04 17:36:21 -08001607 }
Li Zefan645fcc92009-01-07 18:08:43 -08001608
Paul Menagee3712392008-07-25 01:47:02 -07001609 switch (cft->private) {
1610 case FILE_CPULIST:
Li Zefan645fcc92009-01-07 18:08:43 -08001611 retval = update_cpumask(cs, trialcs, buf);
Paul Menagee3712392008-07-25 01:47:02 -07001612 break;
1613 case FILE_MEMLIST:
Li Zefan645fcc92009-01-07 18:08:43 -08001614 retval = update_nodemask(cs, trialcs, buf);
Paul Menagee3712392008-07-25 01:47:02 -07001615 break;
1616 default:
1617 retval = -EINVAL;
1618 break;
1619 }
Li Zefan645fcc92009-01-07 18:08:43 -08001620
1621 free_trial_cpuset(trialcs);
Tejun Heo5d21cc22013-01-07 08:51:08 -08001622out_unlock:
1623 mutex_unlock(&cpuset_mutex);
Paul Menagee3712392008-07-25 01:47:02 -07001624 return retval;
1625}
1626
1627/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 * These ascii lists should be read in a single call, by using a user
1629 * buffer large enough to hold the entire map. If read in smaller
1630 * chunks, there is no guarantee of atomicity. Since the display format
1631 * used, list of ranges of sequential numbers, is variable length,
1632 * and since these maps can change value dynamically, one could read
1633 * gibberish by doing partial reads while a list was changing.
1634 * A single large read to a buffer that crosses a page boundary is
1635 * ok, because the result being copied to user land is not recomputed
1636 * across a page fault.
1637 */
1638
Li Zefan9303e0c2011-03-23 16:42:45 -07001639static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640{
Li Zefan9303e0c2011-03-23 16:42:45 -07001641 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001643 mutex_lock(&callback_mutex);
Li Zefan9303e0c2011-03-23 16:42:45 -07001644 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001645 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646
Li Zefan9303e0c2011-03-23 16:42:45 -07001647 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648}
1649
Li Zefan9303e0c2011-03-23 16:42:45 -07001650static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651{
Li Zefan9303e0c2011-03-23 16:42:45 -07001652 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001654 mutex_lock(&callback_mutex);
Li Zefan9303e0c2011-03-23 16:42:45 -07001655 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001656 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657
Li Zefan9303e0c2011-03-23 16:42:45 -07001658 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659}
1660
Paul Menage8793d852007-10-18 23:39:39 -07001661static ssize_t cpuset_common_file_read(struct cgroup *cont,
1662 struct cftype *cft,
1663 struct file *file,
1664 char __user *buf,
1665 size_t nbytes, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666{
Paul Menage8793d852007-10-18 23:39:39 -07001667 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 cpuset_filetype_t type = cft->private;
1669 char *page;
1670 ssize_t retval = 0;
1671 char *s;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672
Mel Gormane12ba742007-10-16 01:25:52 -07001673 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 return -ENOMEM;
1675
1676 s = page;
1677
1678 switch (type) {
1679 case FILE_CPULIST:
1680 s += cpuset_sprintf_cpulist(s, cs);
1681 break;
1682 case FILE_MEMLIST:
1683 s += cpuset_sprintf_memlist(s, cs);
1684 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 default:
1686 retval = -EINVAL;
1687 goto out;
1688 }
1689 *s++ = '\n';
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690
Al Viroeacaa1f2005-09-30 03:26:43 +01001691 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692out:
1693 free_page((unsigned long)page);
1694 return retval;
1695}
1696
Paul Menage700fe1a2008-04-29 01:00:00 -07001697static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1698{
1699 struct cpuset *cs = cgroup_cs(cont);
1700 cpuset_filetype_t type = cft->private;
1701 switch (type) {
1702 case FILE_CPU_EXCLUSIVE:
1703 return is_cpu_exclusive(cs);
1704 case FILE_MEM_EXCLUSIVE:
1705 return is_mem_exclusive(cs);
Paul Menage78608362008-04-29 01:00:26 -07001706 case FILE_MEM_HARDWALL:
1707 return is_mem_hardwall(cs);
Paul Menage700fe1a2008-04-29 01:00:00 -07001708 case FILE_SCHED_LOAD_BALANCE:
1709 return is_sched_load_balance(cs);
1710 case FILE_MEMORY_MIGRATE:
1711 return is_memory_migrate(cs);
1712 case FILE_MEMORY_PRESSURE_ENABLED:
1713 return cpuset_memory_pressure_enabled;
1714 case FILE_MEMORY_PRESSURE:
1715 return fmeter_getrate(&cs->fmeter);
1716 case FILE_SPREAD_PAGE:
1717 return is_spread_page(cs);
1718 case FILE_SPREAD_SLAB:
1719 return is_spread_slab(cs);
1720 default:
1721 BUG();
1722 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07001723
1724 /* Unreachable but makes gcc happy */
1725 return 0;
Paul Menage700fe1a2008-04-29 01:00:00 -07001726}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727
Paul Menage5be7a472008-05-06 20:42:41 -07001728static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1729{
1730 struct cpuset *cs = cgroup_cs(cont);
1731 cpuset_filetype_t type = cft->private;
1732 switch (type) {
1733 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1734 return cs->relax_domain_level;
1735 default:
1736 BUG();
1737 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07001738
1739 /* Unrechable but makes gcc happy */
1740 return 0;
Paul Menage5be7a472008-05-06 20:42:41 -07001741}
1742
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
1744/*
1745 * for the common functions, 'private' gives the type of file
1746 */
1747
Paul Menageaddf2c72008-04-29 01:00:26 -07001748static struct cftype files[] = {
1749 {
1750 .name = "cpus",
1751 .read = cpuset_common_file_read,
Paul Menagee3712392008-07-25 01:47:02 -07001752 .write_string = cpuset_write_resmask,
1753 .max_write_len = (100U + 6 * NR_CPUS),
Paul Menageaddf2c72008-04-29 01:00:26 -07001754 .private = FILE_CPULIST,
1755 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756
Paul Menageaddf2c72008-04-29 01:00:26 -07001757 {
1758 .name = "mems",
1759 .read = cpuset_common_file_read,
Paul Menagee3712392008-07-25 01:47:02 -07001760 .write_string = cpuset_write_resmask,
1761 .max_write_len = (100U + 6 * MAX_NUMNODES),
Paul Menageaddf2c72008-04-29 01:00:26 -07001762 .private = FILE_MEMLIST,
1763 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764
Paul Menageaddf2c72008-04-29 01:00:26 -07001765 {
1766 .name = "cpu_exclusive",
1767 .read_u64 = cpuset_read_u64,
1768 .write_u64 = cpuset_write_u64,
1769 .private = FILE_CPU_EXCLUSIVE,
1770 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771
Paul Menageaddf2c72008-04-29 01:00:26 -07001772 {
1773 .name = "mem_exclusive",
1774 .read_u64 = cpuset_read_u64,
1775 .write_u64 = cpuset_write_u64,
1776 .private = FILE_MEM_EXCLUSIVE,
1777 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778
Paul Menageaddf2c72008-04-29 01:00:26 -07001779 {
Paul Menage78608362008-04-29 01:00:26 -07001780 .name = "mem_hardwall",
1781 .read_u64 = cpuset_read_u64,
1782 .write_u64 = cpuset_write_u64,
1783 .private = FILE_MEM_HARDWALL,
1784 },
1785
1786 {
Paul Menageaddf2c72008-04-29 01:00:26 -07001787 .name = "sched_load_balance",
1788 .read_u64 = cpuset_read_u64,
1789 .write_u64 = cpuset_write_u64,
1790 .private = FILE_SCHED_LOAD_BALANCE,
1791 },
Paul Jackson029190c2007-10-18 23:40:20 -07001792
Paul Menageaddf2c72008-04-29 01:00:26 -07001793 {
1794 .name = "sched_relax_domain_level",
Paul Menage5be7a472008-05-06 20:42:41 -07001795 .read_s64 = cpuset_read_s64,
1796 .write_s64 = cpuset_write_s64,
Paul Menageaddf2c72008-04-29 01:00:26 -07001797 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1798 },
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001799
Paul Menageaddf2c72008-04-29 01:00:26 -07001800 {
1801 .name = "memory_migrate",
1802 .read_u64 = cpuset_read_u64,
1803 .write_u64 = cpuset_write_u64,
1804 .private = FILE_MEMORY_MIGRATE,
1805 },
1806
1807 {
1808 .name = "memory_pressure",
1809 .read_u64 = cpuset_read_u64,
1810 .write_u64 = cpuset_write_u64,
1811 .private = FILE_MEMORY_PRESSURE,
Li Zefan099fca32009-04-02 16:57:29 -07001812 .mode = S_IRUGO,
Paul Menageaddf2c72008-04-29 01:00:26 -07001813 },
1814
1815 {
1816 .name = "memory_spread_page",
1817 .read_u64 = cpuset_read_u64,
1818 .write_u64 = cpuset_write_u64,
1819 .private = FILE_SPREAD_PAGE,
1820 },
1821
1822 {
1823 .name = "memory_spread_slab",
1824 .read_u64 = cpuset_read_u64,
1825 .write_u64 = cpuset_write_u64,
1826 .private = FILE_SPREAD_SLAB,
1827 },
Tejun Heo4baf6e32012-04-01 12:09:55 -07001828
1829 {
1830 .name = "memory_pressure_enabled",
1831 .flags = CFTYPE_ONLY_ON_ROOT,
1832 .read_u64 = cpuset_read_u64,
1833 .write_u64 = cpuset_write_u64,
1834 .private = FILE_MEMORY_PRESSURE_ENABLED,
1835 },
1836
1837 { } /* terminate */
Paul Jackson45b07ef2006-01-08 01:00:56 -08001838};
1839
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840/*
Tejun Heo92fb9742012-11-19 08:13:38 -08001841 * cpuset_css_alloc - allocate a cpuset css
Paul Menage2df167a2008-02-07 00:14:45 -08001842 * cont: control group that the new cpuset will be part of
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 */
1844
Tejun Heo92fb9742012-11-19 08:13:38 -08001845static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846{
Tejun Heoc8f699b2013-01-07 08:51:07 -08001847 struct cpuset *cs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848
Tejun Heoc8f699b2013-01-07 08:51:07 -08001849 if (!cont->parent)
Paul Menage8793d852007-10-18 23:39:39 -07001850 return &top_cpuset.css;
Tejun Heo033fa1c2012-11-19 08:13:39 -08001851
Tejun Heoc8f699b2013-01-07 08:51:07 -08001852 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 if (!cs)
Paul Menage8793d852007-10-18 23:39:39 -07001854 return ERR_PTR(-ENOMEM);
Li Zefan300ed6c2009-01-07 18:08:44 -08001855 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1856 kfree(cs);
1857 return ERR_PTR(-ENOMEM);
1858 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859
Paul Jackson029190c2007-10-18 23:40:20 -07001860 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
Li Zefan300ed6c2009-01-07 18:08:44 -08001861 cpumask_clear(cs->cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001862 nodes_clear(cs->mems_allowed);
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001863 fmeter_init(&cs->fmeter);
Tejun Heo8d033942013-01-07 08:51:07 -08001864 INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001865 cs->relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866
Tejun Heoc8f699b2013-01-07 08:51:07 -08001867 return &cs->css;
1868}
1869
1870static int cpuset_css_online(struct cgroup *cgrp)
1871{
1872 struct cpuset *cs = cgroup_cs(cgrp);
Tejun Heoc4310692013-01-07 08:51:08 -08001873 struct cpuset *parent = parent_cs(cs);
Tejun Heoae8086c2013-01-07 08:51:07 -08001874 struct cpuset *tmp_cs;
1875 struct cgroup *pos_cg;
Tejun Heoc8f699b2013-01-07 08:51:07 -08001876
1877 if (!parent)
1878 return 0;
1879
Tejun Heo5d21cc22013-01-07 08:51:08 -08001880 mutex_lock(&cpuset_mutex);
1881
Tejun Heoefeb77b2013-01-07 08:51:07 -08001882 set_bit(CS_ONLINE, &cs->flags);
Tejun Heoc8f699b2013-01-07 08:51:07 -08001883 if (is_spread_page(parent))
1884 set_bit(CS_SPREAD_PAGE, &cs->flags);
1885 if (is_spread_slab(parent))
1886 set_bit(CS_SPREAD_SLAB, &cs->flags);
1887
Paul Jackson202f72d2006-01-08 01:01:57 -08001888 number_of_cpusets++;
Tejun Heo033fa1c2012-11-19 08:13:39 -08001889
Tejun Heoc8f699b2013-01-07 08:51:07 -08001890 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
Tejun Heo5d21cc22013-01-07 08:51:08 -08001891 goto out_unlock;
Tejun Heo033fa1c2012-11-19 08:13:39 -08001892
1893 /*
1894 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1895 * set. This flag handling is implemented in cgroup core for
1896 * histrical reasons - the flag may be specified during mount.
1897 *
1898 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1899 * refuse to clone the configuration - thereby refusing the task to
1900 * be entered, and as a result refusing the sys_unshare() or
1901 * clone() which initiated it. If this becomes a problem for some
1902 * users who wish to allow that scenario, then this could be
1903 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1904 * (and likewise for mems) to the new cgroup.
1905 */
Tejun Heoae8086c2013-01-07 08:51:07 -08001906 rcu_read_lock();
1907 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1908 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1909 rcu_read_unlock();
Tejun Heo5d21cc22013-01-07 08:51:08 -08001910 goto out_unlock;
Tejun Heoae8086c2013-01-07 08:51:07 -08001911 }
Tejun Heo033fa1c2012-11-19 08:13:39 -08001912 }
Tejun Heoae8086c2013-01-07 08:51:07 -08001913 rcu_read_unlock();
Tejun Heo033fa1c2012-11-19 08:13:39 -08001914
1915 mutex_lock(&callback_mutex);
1916 cs->mems_allowed = parent->mems_allowed;
1917 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1918 mutex_unlock(&callback_mutex);
Tejun Heo5d21cc22013-01-07 08:51:08 -08001919out_unlock:
1920 mutex_unlock(&cpuset_mutex);
Tejun Heoc8f699b2013-01-07 08:51:07 -08001921 return 0;
1922}
1923
1924static void cpuset_css_offline(struct cgroup *cgrp)
1925{
1926 struct cpuset *cs = cgroup_cs(cgrp);
1927
Tejun Heo5d21cc22013-01-07 08:51:08 -08001928 mutex_lock(&cpuset_mutex);
Tejun Heoc8f699b2013-01-07 08:51:07 -08001929
1930 if (is_sched_load_balance(cs))
1931 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1932
1933 number_of_cpusets--;
Tejun Heoefeb77b2013-01-07 08:51:07 -08001934 clear_bit(CS_ONLINE, &cs->flags);
Tejun Heoc8f699b2013-01-07 08:51:07 -08001935
Tejun Heo5d21cc22013-01-07 08:51:08 -08001936 mutex_unlock(&cpuset_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937}
1938
Paul Jackson029190c2007-10-18 23:40:20 -07001939/*
Paul Jackson029190c2007-10-18 23:40:20 -07001940 * If the cpuset being removed has its flag 'sched_load_balance'
1941 * enabled, then simulate turning sched_load_balance off, which
Tejun Heo699140b2013-01-07 08:51:07 -08001942 * will call rebuild_sched_domains_locked().
Paul Jackson029190c2007-10-18 23:40:20 -07001943 */
1944
Tejun Heo92fb9742012-11-19 08:13:38 -08001945static void cpuset_css_free(struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946{
Paul Menage8793d852007-10-18 23:39:39 -07001947 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948
Li Zefan300ed6c2009-01-07 18:08:44 -08001949 free_cpumask_var(cs->cpus_allowed);
Paul Menage8793d852007-10-18 23:39:39 -07001950 kfree(cs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951}
1952
Paul Menage8793d852007-10-18 23:39:39 -07001953struct cgroup_subsys cpuset_subsys = {
1954 .name = "cpuset",
Tejun Heo92fb9742012-11-19 08:13:38 -08001955 .css_alloc = cpuset_css_alloc,
Tejun Heoc8f699b2013-01-07 08:51:07 -08001956 .css_online = cpuset_css_online,
1957 .css_offline = cpuset_css_offline,
Tejun Heo92fb9742012-11-19 08:13:38 -08001958 .css_free = cpuset_css_free,
Paul Menage8793d852007-10-18 23:39:39 -07001959 .can_attach = cpuset_can_attach,
Tejun Heo452477f2013-01-07 08:51:07 -08001960 .cancel_attach = cpuset_cancel_attach,
Paul Menage8793d852007-10-18 23:39:39 -07001961 .attach = cpuset_attach,
Paul Menage8793d852007-10-18 23:39:39 -07001962 .subsys_id = cpuset_subsys_id,
Tejun Heo4baf6e32012-04-01 12:09:55 -07001963 .base_cftypes = files,
Paul Menage8793d852007-10-18 23:39:39 -07001964 .early_init = 1,
1965};
1966
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967/**
1968 * cpuset_init - initialize cpusets at system boot
1969 *
1970 * Description: Initialize top_cpuset and the cpuset internal file system,
1971 **/
1972
1973int __init cpuset_init(void)
1974{
Paul Menage8793d852007-10-18 23:39:39 -07001975 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976
Miao Xie58568d22009-06-16 15:31:49 -07001977 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1978 BUG();
1979
Li Zefan300ed6c2009-01-07 18:08:44 -08001980 cpumask_setall(top_cpuset.cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001981 nodes_setall(top_cpuset.mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001983 fmeter_init(&top_cpuset.fmeter);
Paul Jackson029190c2007-10-18 23:40:20 -07001984 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001985 top_cpuset.relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987 err = register_filesystem(&cpuset_fs_type);
1988 if (err < 0)
Paul Menage8793d852007-10-18 23:39:39 -07001989 return err;
1990
Li Zefan2341d1b2009-01-07 18:08:42 -08001991 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1992 BUG();
1993
Paul Jackson202f72d2006-01-08 01:01:57 -08001994 number_of_cpusets = 1;
Paul Menage8793d852007-10-18 23:39:39 -07001995 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996}
1997
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001998/*
Max Krasnyanskycf417142008-08-11 14:33:53 -07001999 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002000 * or memory nodes, we need to walk over the cpuset hierarchy,
2001 * removing that CPU or node from all cpusets. If this removes the
Cliff Wickman956db3c2008-02-07 00:14:43 -08002002 * last CPU or node from a cpuset, then move the tasks in the empty
2003 * cpuset to its next-highest non-empty parent.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002004 */
Cliff Wickman956db3c2008-02-07 00:14:43 -08002005static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002006{
Cliff Wickman956db3c2008-02-07 00:14:43 -08002007 struct cpuset *parent;
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002008
Paul Jacksonc8d9c902008-02-07 00:14:46 -08002009 /*
Cliff Wickman956db3c2008-02-07 00:14:43 -08002010 * Find its next-highest non-empty parent, (top cpuset
2011 * has online cpus, so can't be empty).
2012 */
Tejun Heoc4310692013-01-07 08:51:08 -08002013 parent = parent_cs(cs);
Li Zefan300ed6c2009-01-07 18:08:44 -08002014 while (cpumask_empty(parent->cpus_allowed) ||
Paul Jacksonb4501292008-02-07 00:14:47 -08002015 nodes_empty(parent->mems_allowed))
Tejun Heoc4310692013-01-07 08:51:08 -08002016 parent = parent_cs(parent);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002017
Tejun Heo8cc99342013-04-07 09:29:50 -07002018 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2019 rcu_read_lock();
2020 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2021 cgroup_name(cs->css.cgroup));
2022 rcu_read_unlock();
2023 }
Cliff Wickman956db3c2008-02-07 00:14:43 -08002024}
2025
Tejun Heodeb7aa32013-01-07 08:51:07 -08002026/**
Tejun Heo8d033942013-01-07 08:51:07 -08002027 * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
Tejun Heodeb7aa32013-01-07 08:51:07 -08002028 * @cs: cpuset in interest
Cliff Wickman956db3c2008-02-07 00:14:43 -08002029 *
Tejun Heodeb7aa32013-01-07 08:51:07 -08002030 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2031 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2032 * all its tasks are moved to the nearest ancestor with both resources.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002033 */
Tejun Heo8d033942013-01-07 08:51:07 -08002034static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002035{
Tejun Heodeb7aa32013-01-07 08:51:07 -08002036 static cpumask_t off_cpus;
2037 static nodemask_t off_mems, tmp_mems;
Tejun Heo8d033942013-01-07 08:51:07 -08002038 struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
Tejun Heo5d21cc22013-01-07 08:51:08 -08002039 bool is_empty;
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002040
Tejun Heo5d21cc22013-01-07 08:51:08 -08002041 mutex_lock(&cpuset_mutex);
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002042
Tejun Heodeb7aa32013-01-07 08:51:07 -08002043 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2044 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
Paul Jacksonb4501292008-02-07 00:14:47 -08002045
Tejun Heodeb7aa32013-01-07 08:51:07 -08002046 /* remove offline cpus from @cs */
2047 if (!cpumask_empty(&off_cpus)) {
2048 mutex_lock(&callback_mutex);
2049 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2050 mutex_unlock(&callback_mutex);
2051 update_tasks_cpumask(cs, NULL);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002052 }
2053
Tejun Heodeb7aa32013-01-07 08:51:07 -08002054 /* remove offline mems from @cs */
2055 if (!nodes_empty(off_mems)) {
2056 tmp_mems = cs->mems_allowed;
2057 mutex_lock(&callback_mutex);
2058 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2059 mutex_unlock(&callback_mutex);
2060 update_tasks_nodemask(cs, &tmp_mems, NULL);
2061 }
Miao Xief9b4fb82008-07-25 01:47:22 -07002062
Tejun Heo5d21cc22013-01-07 08:51:08 -08002063 is_empty = cpumask_empty(cs->cpus_allowed) ||
2064 nodes_empty(cs->mems_allowed);
Tejun Heo8d033942013-01-07 08:51:07 -08002065
Tejun Heo5d21cc22013-01-07 08:51:08 -08002066 mutex_unlock(&cpuset_mutex);
2067
2068 /*
2069 * If @cs became empty, move tasks to the nearest ancestor with
2070 * execution resources. This is full cgroup operation which will
2071 * also call back into cpuset. Should be done outside any lock.
2072 */
2073 if (is_empty)
2074 remove_tasks_in_empty_cpuset(cs);
Tejun Heo8d033942013-01-07 08:51:07 -08002075
2076 /* the following may free @cs, should be the last operation */
2077 css_put(&cs->css);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002078}
2079
Tejun Heo8d033942013-01-07 08:51:07 -08002080/**
2081 * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
2082 * @cs: cpuset of interest
Cliff Wickman956db3c2008-02-07 00:14:43 -08002083 *
Tejun Heo8d033942013-01-07 08:51:07 -08002084 * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
2085 * memory masks according to top_cpuset.
Paul Jacksonb4501292008-02-07 00:14:47 -08002086 */
Tejun Heo8d033942013-01-07 08:51:07 -08002087static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
Paul Jacksonb4501292008-02-07 00:14:47 -08002088{
Tejun Heo8d033942013-01-07 08:51:07 -08002089 /*
2090 * Pin @cs. The refcnt will be released when the work item
2091 * finishes executing.
2092 */
2093 if (!css_tryget(&cs->css))
2094 return;
Paul Jacksonb4501292008-02-07 00:14:47 -08002095
Tejun Heo8d033942013-01-07 08:51:07 -08002096 /*
2097 * Queue @cs->hotplug_work. If already pending, lose the css ref.
2098 * cpuset_propagate_hotplug_wq is ordered and propagation will
2099 * happen in the order this function is called.
2100 */
2101 if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
2102 css_put(&cs->css);
Tejun Heodeb7aa32013-01-07 08:51:07 -08002103}
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01002104
Tejun Heodeb7aa32013-01-07 08:51:07 -08002105/**
Tejun Heo3a5a6d02013-01-07 08:51:07 -08002106 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
Tejun Heodeb7aa32013-01-07 08:51:07 -08002107 *
2108 * This function is called after either CPU or memory configuration has
2109 * changed and updates cpuset accordingly. The top_cpuset is always
2110 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2111 * order to make cpusets transparent (of no affect) on systems that are
2112 * actively using CPU hotplug but making no active use of cpusets.
2113 *
2114 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2115 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2116 * descendants.
2117 *
2118 * Note that CPU offlining during suspend is ignored. We don't modify
2119 * cpusets across suspend/resume cycles at all.
2120 */
Tejun Heo3a5a6d02013-01-07 08:51:07 -08002121static void cpuset_hotplug_workfn(struct work_struct *work)
Tejun Heodeb7aa32013-01-07 08:51:07 -08002122{
2123 static cpumask_t new_cpus, tmp_cpus;
2124 static nodemask_t new_mems, tmp_mems;
2125 bool cpus_updated, mems_updated;
2126 bool cpus_offlined, mems_offlined;
Paul Jacksonb4501292008-02-07 00:14:47 -08002127
Tejun Heo5d21cc22013-01-07 08:51:08 -08002128 mutex_lock(&cpuset_mutex);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002129
Tejun Heodeb7aa32013-01-07 08:51:07 -08002130 /* fetch the available cpus/mems and find out which changed how */
2131 cpumask_copy(&new_cpus, cpu_active_mask);
2132 new_mems = node_states[N_MEMORY];
Cliff Wickman956db3c2008-02-07 00:14:43 -08002133
Tejun Heodeb7aa32013-01-07 08:51:07 -08002134 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2135 cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2136 &new_cpus);
Srivatsa S. Bhat7ddf96b2012-05-24 19:46:55 +05302137
Tejun Heodeb7aa32013-01-07 08:51:07 -08002138 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2139 nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2140 mems_offlined = !nodes_empty(tmp_mems);
Srivatsa S. Bhat7ddf96b2012-05-24 19:46:55 +05302141
Tejun Heodeb7aa32013-01-07 08:51:07 -08002142 /* synchronize cpus_allowed to cpu_active_mask */
2143 if (cpus_updated) {
2144 mutex_lock(&callback_mutex);
2145 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2146 mutex_unlock(&callback_mutex);
2147 /* we don't mess with cpumasks of tasks in top_cpuset */
2148 }
Srivatsa S. Bhat7ddf96b2012-05-24 19:46:55 +05302149
Tejun Heodeb7aa32013-01-07 08:51:07 -08002150 /* synchronize mems_allowed to N_MEMORY */
2151 if (mems_updated) {
2152 tmp_mems = top_cpuset.mems_allowed;
2153 mutex_lock(&callback_mutex);
2154 top_cpuset.mems_allowed = new_mems;
2155 mutex_unlock(&callback_mutex);
2156 update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
2157 }
Srivatsa S. Bhat7ddf96b2012-05-24 19:46:55 +05302158
Tejun Heodeb7aa32013-01-07 08:51:07 -08002159 /* if cpus or mems went down, we need to propagate to descendants */
2160 if (cpus_offlined || mems_offlined) {
2161 struct cpuset *cs;
Tejun Heofc560a22013-01-07 08:51:08 -08002162 struct cgroup *pos_cgrp;
Paul Jacksonb4501292008-02-07 00:14:47 -08002163
Tejun Heofc560a22013-01-07 08:51:08 -08002164 rcu_read_lock();
2165 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset)
2166 schedule_cpuset_propagate_hotplug(cs);
2167 rcu_read_unlock();
Tejun Heodeb7aa32013-01-07 08:51:07 -08002168 }
2169
Tejun Heo5d21cc22013-01-07 08:51:08 -08002170 mutex_unlock(&cpuset_mutex);
Tejun Heodeb7aa32013-01-07 08:51:07 -08002171
Tejun Heo8d033942013-01-07 08:51:07 -08002172 /* wait for propagations to finish */
2173 flush_workqueue(cpuset_propagate_hotplug_wq);
2174
Tejun Heodeb7aa32013-01-07 08:51:07 -08002175 /* rebuild sched domains if cpus_allowed has changed */
Li Zhonge0e80a02013-04-27 06:52:43 -07002176 if (cpus_updated)
2177 rebuild_sched_domains();
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002178}
2179
Srivatsa S. Bhat7ddf96b2012-05-24 19:46:55 +05302180void cpuset_update_active_cpus(bool cpu_online)
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002181{
Tejun Heo3a5a6d02013-01-07 08:51:07 -08002182 /*
2183 * We're inside cpu hotplug critical region which usually nests
2184 * inside cgroup synchronization. Bounce actual hotplug processing
2185 * to a work item to avoid reverse locking order.
2186 *
2187 * We still need to do partition_sched_domains() synchronously;
2188 * otherwise, the scheduler will get confused and put tasks to the
2189 * dead CPU. Fall back to the default single domain.
2190 * cpuset_hotplug_workfn() will rebuild it as necessary.
2191 */
2192 partition_sched_domains(1, NULL, NULL);
2193 schedule_work(&cpuset_hotplug_work);
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002194}
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002195
Paul Jackson38837fc2006-09-29 02:01:16 -07002196/*
Lai Jiangshan38d7bee2012-12-12 13:51:24 -08002197 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2198 * Call this routine anytime after node_states[N_MEMORY] changes.
Srivatsa S. Bhata1cd2b12012-05-24 19:47:03 +05302199 * See cpuset_update_active_cpus() for CPU hotplug handling.
Paul Jackson38837fc2006-09-29 02:01:16 -07002200 */
Miao Xief4818912008-11-19 15:36:30 -08002201static int cpuset_track_online_nodes(struct notifier_block *self,
2202 unsigned long action, void *arg)
Paul Jackson38837fc2006-09-29 02:01:16 -07002203{
Tejun Heo3a5a6d02013-01-07 08:51:07 -08002204 schedule_work(&cpuset_hotplug_work);
Miao Xief4818912008-11-19 15:36:30 -08002205 return NOTIFY_OK;
Paul Jackson38837fc2006-09-29 02:01:16 -07002206}
Andrew Mortond8f10cb2013-04-29 15:08:08 -07002207
2208static struct notifier_block cpuset_track_online_nodes_nb = {
2209 .notifier_call = cpuset_track_online_nodes,
2210 .priority = 10, /* ??! */
2211};
Paul Jackson38837fc2006-09-29 02:01:16 -07002212
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213/**
2214 * cpuset_init_smp - initialize cpus_allowed
2215 *
2216 * Description: Finish top cpuset after cpu, node maps are initialized
Andrew Mortond8f10cb2013-04-29 15:08:08 -07002217 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218void __init cpuset_init_smp(void)
2219{
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01002220 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
Lai Jiangshan38d7bee2012-12-12 13:51:24 -08002221 top_cpuset.mems_allowed = node_states[N_MEMORY];
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002222
Andrew Mortond8f10cb2013-04-29 15:08:08 -07002223 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
Miao Xief90d4112009-01-16 10:24:10 +08002224
Tejun Heo8d033942013-01-07 08:51:07 -08002225 cpuset_propagate_hotplug_wq =
2226 alloc_ordered_workqueue("cpuset_hotplug", 0);
2227 BUG_ON(!cpuset_propagate_hotplug_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228}
2229
2230/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2232 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
Li Zefan6af866a2009-01-07 18:08:45 -08002233 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 *
Li Zefan300ed6c2009-01-07 18:08:44 -08002235 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 * attached to the specified @tsk. Guaranteed to return some non-empty
Rusty Russell5f054e32012-03-29 15:38:31 +10302237 * subset of cpu_online_mask, even if this means going outside the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 * tasks cpuset.
2239 **/
2240
Li Zefan6af866a2009-01-07 18:08:45 -08002241void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002243 mutex_lock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002244 task_lock(tsk);
Mike Travisf9a86fc2008-04-04 18:11:07 -07002245 guarantee_online_cpus(task_cs(tsk), pmask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002246 task_unlock(tsk);
Oleg Nesterov897f0b32010-03-15 10:10:03 +01002247 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248}
2249
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01002250void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
Oleg Nesterov9084bb82010-03-15 10:10:27 +01002251{
2252 const struct cpuset *cs;
Oleg Nesterov9084bb82010-03-15 10:10:27 +01002253
2254 rcu_read_lock();
2255 cs = task_cs(tsk);
2256 if (cs)
KOSAKI Motohiro1e1b6c52011-05-19 15:08:58 +09002257 do_set_cpus_allowed(tsk, cs->cpus_allowed);
Oleg Nesterov9084bb82010-03-15 10:10:27 +01002258 rcu_read_unlock();
2259
2260 /*
2261 * We own tsk->cpus_allowed, nobody can change it under us.
2262 *
2263 * But we used cs && cs->cpus_allowed lockless and thus can
2264 * race with cgroup_attach_task() or update_cpumask() and get
2265 * the wrong tsk->cpus_allowed. However, both cases imply the
2266 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2267 * which takes task_rq_lock().
2268 *
2269 * If we are called after it dropped the lock we must see all
2270 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2271 * set any mask even if it is not right from task_cs() pov,
2272 * the pending set_cpus_allowed_ptr() will fix things.
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01002273 *
2274 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2275 * if required.
Oleg Nesterov9084bb82010-03-15 10:10:27 +01002276 */
Oleg Nesterov9084bb82010-03-15 10:10:27 +01002277}
2278
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279void cpuset_init_current_mems_allowed(void)
2280{
Mike Travisf9a86fc2008-04-04 18:11:07 -07002281 nodes_setall(current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282}
2283
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002284/**
Paul Jackson909d75a2006-01-08 01:01:55 -08002285 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2286 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2287 *
2288 * Description: Returns the nodemask_t mems_allowed of the cpuset
2289 * attached to the specified @tsk. Guaranteed to return some non-empty
Lai Jiangshan38d7bee2012-12-12 13:51:24 -08002290 * subset of node_states[N_MEMORY], even if this means going outside the
Paul Jackson909d75a2006-01-08 01:01:55 -08002291 * tasks cpuset.
2292 **/
2293
2294nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2295{
2296 nodemask_t mask;
2297
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002298 mutex_lock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002299 task_lock(tsk);
Paul Menage8793d852007-10-18 23:39:39 -07002300 guarantee_online_mems(task_cs(tsk), &mask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002301 task_unlock(tsk);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002302 mutex_unlock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002303
2304 return mask;
2305}
2306
2307/**
Mel Gorman19770b32008-04-28 02:12:18 -07002308 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2309 * @nodemask: the nodemask to be checked
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002310 *
Mel Gorman19770b32008-04-28 02:12:18 -07002311 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 */
Mel Gorman19770b32008-04-28 02:12:18 -07002313int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314{
Mel Gorman19770b32008-04-28 02:12:18 -07002315 return nodes_intersects(*nodemask, current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316}
2317
Paul Jackson9bf22292005-09-06 15:18:12 -07002318/*
Paul Menage78608362008-04-29 01:00:26 -07002319 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2320 * mem_hardwall ancestor to the specified cpuset. Call holding
2321 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2322 * (an unusual configuration), then returns the root cpuset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 */
Paul Menage78608362008-04-29 01:00:26 -07002324static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325{
Tejun Heoc4310692013-01-07 08:51:08 -08002326 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2327 cs = parent_cs(cs);
Paul Jackson9bf22292005-09-06 15:18:12 -07002328 return cs;
2329}
2330
2331/**
David Rientjesa1bc5a42009-04-02 16:57:54 -07002332 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2333 * @node: is this an allowed node?
Paul Jackson02a0e532006-12-13 00:34:25 -08002334 * @gfp_mask: memory allocation flags
Paul Jackson9bf22292005-09-06 15:18:12 -07002335 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002336 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2337 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2338 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2339 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2340 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2341 * flag, yes.
Paul Jackson9bf22292005-09-06 15:18:12 -07002342 * Otherwise, no.
2343 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002344 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2345 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2346 * might sleep, and might allow a node from an enclosing cpuset.
Paul Jackson02a0e532006-12-13 00:34:25 -08002347 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002348 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2349 * cpusets, and never sleeps.
Paul Jackson02a0e532006-12-13 00:34:25 -08002350 *
2351 * The __GFP_THISNODE placement logic is really handled elsewhere,
2352 * by forcibly using a zonelist starting at a specified node, and by
2353 * (in get_page_from_freelist()) refusing to consider the zones for
2354 * any node on the zonelist except the first. By the time any such
2355 * calls get to this routine, we should just shut up and say 'yes'.
2356 *
Paul Jackson9bf22292005-09-06 15:18:12 -07002357 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
David Rientjesc596d9f2007-05-06 14:49:32 -07002358 * and do not allow allocations outside the current tasks cpuset
2359 * unless the task has been OOM killed as is marked TIF_MEMDIE.
Paul Jackson9bf22292005-09-06 15:18:12 -07002360 * GFP_KERNEL allocations are not so marked, so can escape to the
Paul Menage78608362008-04-29 01:00:26 -07002361 * nearest enclosing hardwalled ancestor cpuset.
Paul Jackson9bf22292005-09-06 15:18:12 -07002362 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002363 * Scanning up parent cpusets requires callback_mutex. The
2364 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2365 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2366 * current tasks mems_allowed came up empty on the first pass over
2367 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2368 * cpuset are short of memory, might require taking the callback_mutex
2369 * mutex.
Paul Jackson9bf22292005-09-06 15:18:12 -07002370 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002371 * The first call here from mm/page_alloc:get_page_from_freelist()
Paul Jackson02a0e532006-12-13 00:34:25 -08002372 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2373 * so no allocation on a node outside the cpuset is allowed (unless
2374 * in interrupt, of course).
Paul Jackson9bf22292005-09-06 15:18:12 -07002375 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002376 * The second pass through get_page_from_freelist() doesn't even call
2377 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2378 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2379 * in alloc_flags. That logic and the checks below have the combined
2380 * affect that:
Paul Jackson9bf22292005-09-06 15:18:12 -07002381 * in_interrupt - any node ok (current task context irrelevant)
2382 * GFP_ATOMIC - any node ok
David Rientjesc596d9f2007-05-06 14:49:32 -07002383 * TIF_MEMDIE - any node ok
Paul Menage78608362008-04-29 01:00:26 -07002384 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
Paul Jackson9bf22292005-09-06 15:18:12 -07002385 * GFP_USER - only nodes in current tasks mems allowed ok.
Paul Jackson36be57f2006-05-20 15:00:10 -07002386 *
2387 * Rule:
David Rientjesa1bc5a42009-04-02 16:57:54 -07002388 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
Paul Jackson36be57f2006-05-20 15:00:10 -07002389 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2390 * the code that might scan up ancestor cpusets and sleep.
Paul Jackson02a0e532006-12-13 00:34:25 -08002391 */
David Rientjesa1bc5a42009-04-02 16:57:54 -07002392int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
Paul Jackson9bf22292005-09-06 15:18:12 -07002393{
Paul Jackson9bf22292005-09-06 15:18:12 -07002394 const struct cpuset *cs; /* current cpuset ancestors */
Paul Jackson29afd492006-03-24 03:16:12 -08002395 int allowed; /* is allocation in zone z allowed? */
Paul Jackson9bf22292005-09-06 15:18:12 -07002396
Christoph Lameter9b819d22006-09-25 23:31:40 -07002397 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
Paul Jackson9bf22292005-09-06 15:18:12 -07002398 return 1;
Paul Jackson92d1dbd2006-05-20 15:00:11 -07002399 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
Paul Jackson9bf22292005-09-06 15:18:12 -07002400 if (node_isset(node, current->mems_allowed))
2401 return 1;
David Rientjesc596d9f2007-05-06 14:49:32 -07002402 /*
2403 * Allow tasks that have access to memory reserves because they have
2404 * been OOM killed to get memory anywhere.
2405 */
2406 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2407 return 1;
Paul Jackson9bf22292005-09-06 15:18:12 -07002408 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2409 return 0;
2410
Bob Picco5563e772005-11-13 16:06:35 -08002411 if (current->flags & PF_EXITING) /* Let dying task have memory */
2412 return 1;
2413
Paul Jackson9bf22292005-09-06 15:18:12 -07002414 /* Not hardwall and node outside mems_allowed: scan up cpusets */
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002415 mutex_lock(&callback_mutex);
Paul Jackson053199e2005-10-30 15:02:30 -08002416
Paul Jackson053199e2005-10-30 15:02:30 -08002417 task_lock(current);
Paul Menage78608362008-04-29 01:00:26 -07002418 cs = nearest_hardwall_ancestor(task_cs(current));
Paul Jackson053199e2005-10-30 15:02:30 -08002419 task_unlock(current);
2420
Paul Jackson9bf22292005-09-06 15:18:12 -07002421 allowed = node_isset(node, cs->mems_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002422 mutex_unlock(&callback_mutex);
Paul Jackson9bf22292005-09-06 15:18:12 -07002423 return allowed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424}
2425
Paul Jackson02a0e532006-12-13 00:34:25 -08002426/*
David Rientjesa1bc5a42009-04-02 16:57:54 -07002427 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2428 * @node: is this an allowed node?
Paul Jackson02a0e532006-12-13 00:34:25 -08002429 * @gfp_mask: memory allocation flags
2430 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002431 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2432 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2433 * yes. If the task has been OOM killed and has access to memory reserves as
2434 * specified by the TIF_MEMDIE flag, yes.
2435 * Otherwise, no.
Paul Jackson02a0e532006-12-13 00:34:25 -08002436 *
2437 * The __GFP_THISNODE placement logic is really handled elsewhere,
2438 * by forcibly using a zonelist starting at a specified node, and by
2439 * (in get_page_from_freelist()) refusing to consider the zones for
2440 * any node on the zonelist except the first. By the time any such
2441 * calls get to this routine, we should just shut up and say 'yes'.
2442 *
David Rientjesa1bc5a42009-04-02 16:57:54 -07002443 * Unlike the cpuset_node_allowed_softwall() variant, above,
2444 * this variant requires that the node be in the current task's
Paul Jackson02a0e532006-12-13 00:34:25 -08002445 * mems_allowed or that we're in interrupt. It does not scan up the
2446 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2447 * It never sleeps.
2448 */
David Rientjesa1bc5a42009-04-02 16:57:54 -07002449int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
Paul Jackson02a0e532006-12-13 00:34:25 -08002450{
Paul Jackson02a0e532006-12-13 00:34:25 -08002451 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2452 return 1;
Paul Jackson02a0e532006-12-13 00:34:25 -08002453 if (node_isset(node, current->mems_allowed))
2454 return 1;
Daniel Walkerdedf8b72007-10-18 03:06:04 -07002455 /*
2456 * Allow tasks that have access to memory reserves because they have
2457 * been OOM killed to get memory anywhere.
2458 */
2459 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2460 return 1;
Paul Jackson02a0e532006-12-13 00:34:25 -08002461 return 0;
2462}
2463
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002464/**
Jack Steiner6adef3e2010-05-26 14:42:49 -07002465 * cpuset_mem_spread_node() - On which node to begin search for a file page
2466 * cpuset_slab_spread_node() - On which node to begin search for a slab page
Paul Jackson825a46a2006-03-24 03:16:03 -08002467 *
2468 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2469 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2470 * and if the memory allocation used cpuset_mem_spread_node()
2471 * to determine on which node to start looking, as it will for
2472 * certain page cache or slab cache pages such as used for file
2473 * system buffers and inode caches, then instead of starting on the
2474 * local node to look for a free page, rather spread the starting
2475 * node around the tasks mems_allowed nodes.
2476 *
2477 * We don't have to worry about the returned node being offline
2478 * because "it can't happen", and even if it did, it would be ok.
2479 *
2480 * The routines calling guarantee_online_mems() are careful to
2481 * only set nodes in task->mems_allowed that are online. So it
2482 * should not be possible for the following code to return an
2483 * offline node. But if it did, that would be ok, as this routine
2484 * is not returning the node where the allocation must be, only
2485 * the node where the search should start. The zonelist passed to
2486 * __alloc_pages() will include all nodes. If the slab allocator
2487 * is passed an offline node, it will fall back to the local node.
2488 * See kmem_cache_alloc_node().
2489 */
2490
Jack Steiner6adef3e2010-05-26 14:42:49 -07002491static int cpuset_spread_node(int *rotor)
Paul Jackson825a46a2006-03-24 03:16:03 -08002492{
2493 int node;
2494
Jack Steiner6adef3e2010-05-26 14:42:49 -07002495 node = next_node(*rotor, current->mems_allowed);
Paul Jackson825a46a2006-03-24 03:16:03 -08002496 if (node == MAX_NUMNODES)
2497 node = first_node(current->mems_allowed);
Jack Steiner6adef3e2010-05-26 14:42:49 -07002498 *rotor = node;
Paul Jackson825a46a2006-03-24 03:16:03 -08002499 return node;
2500}
Jack Steiner6adef3e2010-05-26 14:42:49 -07002501
2502int cpuset_mem_spread_node(void)
2503{
Michal Hocko778d3b02011-07-26 16:08:30 -07002504 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2505 current->cpuset_mem_spread_rotor =
2506 node_random(&current->mems_allowed);
2507
Jack Steiner6adef3e2010-05-26 14:42:49 -07002508 return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2509}
2510
2511int cpuset_slab_spread_node(void)
2512{
Michal Hocko778d3b02011-07-26 16:08:30 -07002513 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2514 current->cpuset_slab_spread_rotor =
2515 node_random(&current->mems_allowed);
2516
Jack Steiner6adef3e2010-05-26 14:42:49 -07002517 return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2518}
2519
Paul Jackson825a46a2006-03-24 03:16:03 -08002520EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2521
2522/**
David Rientjesbbe373f2007-10-16 23:25:58 -07002523 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2524 * @tsk1: pointer to task_struct of some task.
2525 * @tsk2: pointer to task_struct of some other task.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002526 *
David Rientjesbbe373f2007-10-16 23:25:58 -07002527 * Description: Return true if @tsk1's mems_allowed intersects the
2528 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2529 * one of the task's memory usage might impact the memory available
2530 * to the other.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002531 **/
2532
David Rientjesbbe373f2007-10-16 23:25:58 -07002533int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2534 const struct task_struct *tsk2)
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002535{
David Rientjesbbe373f2007-10-16 23:25:58 -07002536 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002537}
2538
Li Zefanf440d982013-03-01 15:02:15 +08002539#define CPUSET_NODELIST_LEN (256)
2540
David Rientjes75aa1992009-01-06 14:39:01 -08002541/**
2542 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2543 * @task: pointer to task_struct of some task.
2544 *
2545 * Description: Prints @task's name, cpuset name, and cached copy of its
2546 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2547 * dereferencing task_cs(task).
2548 */
2549void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2550{
Li Zefanf440d982013-03-01 15:02:15 +08002551 /* Statically allocated to prevent using excess stack. */
2552 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2553 static DEFINE_SPINLOCK(cpuset_buffer_lock);
David Rientjes75aa1992009-01-06 14:39:01 -08002554
Li Zefanf440d982013-03-01 15:02:15 +08002555 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2556
Li Zefancfb59662013-03-12 10:28:39 +08002557 rcu_read_lock();
David Rientjes75aa1992009-01-06 14:39:01 -08002558 spin_lock(&cpuset_buffer_lock);
Li Zefan63f43f52013-01-25 16:08:01 +08002559
David Rientjes75aa1992009-01-06 14:39:01 -08002560 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2561 tsk->mems_allowed);
2562 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
Li Zefanf440d982013-03-01 15:02:15 +08002563 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2564
David Rientjes75aa1992009-01-06 14:39:01 -08002565 spin_unlock(&cpuset_buffer_lock);
Li Zefancfb59662013-03-12 10:28:39 +08002566 rcu_read_unlock();
David Rientjes75aa1992009-01-06 14:39:01 -08002567}
2568
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569/*
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002570 * Collection of memory_pressure is suppressed unless
2571 * this flag is enabled by writing "1" to the special
2572 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2573 */
2574
Paul Jacksonc5b2aff2006-01-08 01:01:51 -08002575int cpuset_memory_pressure_enabled __read_mostly;
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002576
2577/**
2578 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2579 *
2580 * Keep a running average of the rate of synchronous (direct)
2581 * page reclaim efforts initiated by tasks in each cpuset.
2582 *
2583 * This represents the rate at which some task in the cpuset
2584 * ran low on memory on all nodes it was allowed to use, and
2585 * had to enter the kernels page reclaim code in an effort to
2586 * create more free memory by tossing clean pages or swapping
2587 * or writing dirty pages.
2588 *
2589 * Display to user space in the per-cpuset read-only file
2590 * "memory_pressure". Value displayed is an integer
2591 * representing the recent rate of entry into the synchronous
2592 * (direct) page reclaim by any task attached to the cpuset.
2593 **/
2594
2595void __cpuset_memory_pressure_bump(void)
2596{
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002597 task_lock(current);
Paul Menage8793d852007-10-18 23:39:39 -07002598 fmeter_markevent(&task_cs(current)->fmeter);
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002599 task_unlock(current);
2600}
2601
Paul Menage8793d852007-10-18 23:39:39 -07002602#ifdef CONFIG_PROC_PID_CPUSET
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002603/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 * proc_cpuset_show()
2605 * - Print tasks cpuset path into seq_file.
2606 * - Used for /proc/<pid>/cpuset.
Paul Jackson053199e2005-10-30 15:02:30 -08002607 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2608 * doesn't really matter if tsk->cpuset changes after we read it,
Tejun Heo5d21cc22013-01-07 08:51:08 -08002609 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
Paul Menage2df167a2008-02-07 00:14:45 -08002610 * anyway.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 */
Al Viro8d8b97b2013-04-19 23:11:24 -04002612int proc_cpuset_show(struct seq_file *m, void *unused_v)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613{
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002614 struct pid *pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 struct task_struct *tsk;
2616 char *buf;
Paul Menage8793d852007-10-18 23:39:39 -07002617 struct cgroup_subsys_state *css;
Eric W. Biederman99f89552006-06-26 00:25:55 -07002618 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619
Eric W. Biederman99f89552006-06-26 00:25:55 -07002620 retval = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2622 if (!buf)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002623 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624
Eric W. Biederman99f89552006-06-26 00:25:55 -07002625 retval = -ESRCH;
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002626 pid = m->private;
2627 tsk = get_pid_task(pid, PIDTYPE_PID);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002628 if (!tsk)
2629 goto out_free;
2630
Li Zefan27e89ae2013-01-15 14:10:57 +08002631 rcu_read_lock();
Paul Menage8793d852007-10-18 23:39:39 -07002632 css = task_subsys_state(tsk, cpuset_subsys_id);
2633 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
Li Zefan27e89ae2013-01-15 14:10:57 +08002634 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 if (retval < 0)
Li Zefan27e89ae2013-01-15 14:10:57 +08002636 goto out_put_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 seq_puts(m, buf);
2638 seq_putc(m, '\n');
Li Zefan27e89ae2013-01-15 14:10:57 +08002639out_put_task:
Eric W. Biederman99f89552006-06-26 00:25:55 -07002640 put_task_struct(tsk);
2641out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 kfree(buf);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002643out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 return retval;
2645}
Paul Menage8793d852007-10-18 23:39:39 -07002646#endif /* CONFIG_PROC_PID_CPUSET */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647
Heiko Carstensd01d4822009-09-21 11:06:27 +02002648/* Display task mems_allowed in /proc/<pid>/status file. */
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002649void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650{
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002651 seq_printf(m, "Mems_allowed:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002652 seq_nodemask(m, &task->mems_allowed);
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002653 seq_printf(m, "\n");
Mike Travis39106dc2008-04-08 11:43:03 -07002654 seq_printf(m, "Mems_allowed_list:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002655 seq_nodemask_list(m, &task->mems_allowed);
Mike Travis39106dc2008-04-08 11:43:03 -07002656 seq_printf(m, "\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657}