1 |
3 |
xianfeng |
/*
|
2 |
|
|
* kernel/cpuset.c
|
3 |
|
|
*
|
4 |
|
|
* Processor and Memory placement constraints for sets of tasks.
|
5 |
|
|
*
|
6 |
|
|
* Copyright (C) 2003 BULL SA.
|
7 |
|
|
* Copyright (C) 2004-2007 Silicon Graphics, Inc.
|
8 |
|
|
* Copyright (C) 2006 Google, Inc
|
9 |
|
|
*
|
10 |
|
|
* Portions derived from Patrick Mochel's sysfs code.
|
11 |
|
|
* sysfs is Copyright (c) 2001-3 Patrick Mochel
|
12 |
|
|
*
|
13 |
|
|
* 2003-10-10 Written by Simon Derr.
|
14 |
|
|
* 2003-10-22 Updates by Stephen Hemminger.
|
15 |
|
|
* 2004 May-July Rework by Paul Jackson.
|
16 |
|
|
* 2006 Rework by Paul Menage to use generic cgroups
|
17 |
|
|
*
|
18 |
|
|
* This file is subject to the terms and conditions of the GNU General Public
|
19 |
|
|
* License. See the file COPYING in the main directory of the Linux
|
20 |
|
|
* distribution for more details.
|
21 |
|
|
*/
|
22 |
|
|
|
23 |
|
|
#include <linux/cpu.h>
|
24 |
|
|
#include <linux/cpumask.h>
|
25 |
|
|
#include <linux/cpuset.h>
|
26 |
|
|
#include <linux/err.h>
|
27 |
|
|
#include <linux/errno.h>
|
28 |
|
|
#include <linux/file.h>
|
29 |
|
|
#include <linux/fs.h>
|
30 |
|
|
#include <linux/init.h>
|
31 |
|
|
#include <linux/interrupt.h>
|
32 |
|
|
#include <linux/kernel.h>
|
33 |
|
|
#include <linux/kmod.h>
|
34 |
|
|
#include <linux/list.h>
|
35 |
|
|
#include <linux/mempolicy.h>
|
36 |
|
|
#include <linux/mm.h>
|
37 |
|
|
#include <linux/module.h>
|
38 |
|
|
#include <linux/mount.h>
|
39 |
|
|
#include <linux/namei.h>
|
40 |
|
|
#include <linux/pagemap.h>
|
41 |
|
|
#include <linux/prio_heap.h>
|
42 |
|
|
#include <linux/proc_fs.h>
|
43 |
|
|
#include <linux/rcupdate.h>
|
44 |
|
|
#include <linux/sched.h>
|
45 |
|
|
#include <linux/seq_file.h>
|
46 |
|
|
#include <linux/security.h>
|
47 |
|
|
#include <linux/slab.h>
|
48 |
|
|
#include <linux/spinlock.h>
|
49 |
|
|
#include <linux/stat.h>
|
50 |
|
|
#include <linux/string.h>
|
51 |
|
|
#include <linux/time.h>
|
52 |
|
|
#include <linux/backing-dev.h>
|
53 |
|
|
#include <linux/sort.h>
|
54 |
|
|
|
55 |
|
|
#include <asm/uaccess.h>
|
56 |
|
|
#include <asm/atomic.h>
|
57 |
|
|
#include <linux/mutex.h>
|
58 |
|
|
#include <linux/kfifo.h>
|
59 |
|
|
|
60 |
|
|
/*
|
61 |
|
|
* Tracks how many cpusets are currently defined in system.
|
62 |
|
|
* When there is only one cpuset (the root cpuset) we can
|
63 |
|
|
* short circuit some hooks.
|
64 |
|
|
*/
|
65 |
|
|
int number_of_cpusets __read_mostly;
|
66 |
|
|
|
67 |
|
|
/* Retrieve the cpuset from a cgroup */
|
68 |
|
|
struct cgroup_subsys cpuset_subsys;
|
69 |
|
|
struct cpuset;
|
70 |
|
|
|
71 |
|
|
/* See "Frequency meter" comments, below. */
|
72 |
|
|
|
73 |
|
|
struct fmeter {
|
74 |
|
|
int cnt; /* unprocessed events count */
|
75 |
|
|
int val; /* most recent output value */
|
76 |
|
|
time_t time; /* clock (secs) when val computed */
|
77 |
|
|
spinlock_t lock; /* guards read or write of above */
|
78 |
|
|
};
|
79 |
|
|
|
80 |
|
|
struct cpuset {
|
81 |
|
|
struct cgroup_subsys_state css;
|
82 |
|
|
|
83 |
|
|
unsigned long flags; /* "unsigned long" so bitops work */
|
84 |
|
|
cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
|
85 |
|
|
nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
|
86 |
|
|
|
87 |
|
|
struct cpuset *parent; /* my parent */
|
88 |
|
|
|
89 |
|
|
/*
|
90 |
|
|
* Copy of global cpuset_mems_generation as of the most
|
91 |
|
|
* recent time this cpuset changed its mems_allowed.
|
92 |
|
|
*/
|
93 |
|
|
int mems_generation;
|
94 |
|
|
|
95 |
|
|
struct fmeter fmeter; /* memory_pressure filter */
|
96 |
|
|
|
97 |
|
|
/* partition number for rebuild_sched_domains() */
|
98 |
|
|
int pn;
|
99 |
|
|
};
|
100 |
|
|
|
101 |
|
|
/* Retrieve the cpuset for a cgroup */
|
102 |
|
|
static inline struct cpuset *cgroup_cs(struct cgroup *cont)
|
103 |
|
|
{
|
104 |
|
|
return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
|
105 |
|
|
struct cpuset, css);
|
106 |
|
|
}
|
107 |
|
|
|
108 |
|
|
/* Retrieve the cpuset for a task */
|
109 |
|
|
static inline struct cpuset *task_cs(struct task_struct *task)
|
110 |
|
|
{
|
111 |
|
|
return container_of(task_subsys_state(task, cpuset_subsys_id),
|
112 |
|
|
struct cpuset, css);
|
113 |
|
|
}
|
114 |
|
|
|
115 |
|
|
|
116 |
|
|
/* bits in struct cpuset flags field */
|
117 |
|
|
typedef enum {
|
118 |
|
|
CS_CPU_EXCLUSIVE,
|
119 |
|
|
CS_MEM_EXCLUSIVE,
|
120 |
|
|
CS_MEMORY_MIGRATE,
|
121 |
|
|
CS_SCHED_LOAD_BALANCE,
|
122 |
|
|
CS_SPREAD_PAGE,
|
123 |
|
|
CS_SPREAD_SLAB,
|
124 |
|
|
} cpuset_flagbits_t;
|
125 |
|
|
|
126 |
|
|
/* convenient tests for these bits */
|
127 |
|
|
static inline int is_cpu_exclusive(const struct cpuset *cs)
|
128 |
|
|
{
|
129 |
|
|
return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
|
130 |
|
|
}
|
131 |
|
|
|
132 |
|
|
static inline int is_mem_exclusive(const struct cpuset *cs)
|
133 |
|
|
{
|
134 |
|
|
return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
|
135 |
|
|
}
|
136 |
|
|
|
137 |
|
|
static inline int is_sched_load_balance(const struct cpuset *cs)
|
138 |
|
|
{
|
139 |
|
|
return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
140 |
|
|
}
|
141 |
|
|
|
142 |
|
|
static inline int is_memory_migrate(const struct cpuset *cs)
|
143 |
|
|
{
|
144 |
|
|
return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
|
145 |
|
|
}
|
146 |
|
|
|
147 |
|
|
static inline int is_spread_page(const struct cpuset *cs)
|
148 |
|
|
{
|
149 |
|
|
return test_bit(CS_SPREAD_PAGE, &cs->flags);
|
150 |
|
|
}
|
151 |
|
|
|
152 |
|
|
static inline int is_spread_slab(const struct cpuset *cs)
|
153 |
|
|
{
|
154 |
|
|
return test_bit(CS_SPREAD_SLAB, &cs->flags);
|
155 |
|
|
}
|
156 |
|
|
|
157 |
|
|
/*
|
158 |
|
|
* Increment this integer everytime any cpuset changes its
|
159 |
|
|
* mems_allowed value. Users of cpusets can track this generation
|
160 |
|
|
* number, and avoid having to lock and reload mems_allowed unless
|
161 |
|
|
* the cpuset they're using changes generation.
|
162 |
|
|
*
|
163 |
|
|
* A single, global generation is needed because attach_task() could
|
164 |
|
|
* reattach a task to a different cpuset, which must not have its
|
165 |
|
|
* generation numbers aliased with those of that tasks previous cpuset.
|
166 |
|
|
*
|
167 |
|
|
* Generations are needed for mems_allowed because one task cannot
|
168 |
|
|
* modify anothers memory placement. So we must enable every task,
|
169 |
|
|
* on every visit to __alloc_pages(), to efficiently check whether
|
170 |
|
|
* its current->cpuset->mems_allowed has changed, requiring an update
|
171 |
|
|
* of its current->mems_allowed.
|
172 |
|
|
*
|
173 |
|
|
* Since cpuset_mems_generation is guarded by manage_mutex,
|
174 |
|
|
* there is no need to mark it atomic.
|
175 |
|
|
*/
|
176 |
|
|
static int cpuset_mems_generation;
|
177 |
|
|
|
178 |
|
|
static struct cpuset top_cpuset = {
|
179 |
|
|
.flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
|
180 |
|
|
.cpus_allowed = CPU_MASK_ALL,
|
181 |
|
|
.mems_allowed = NODE_MASK_ALL,
|
182 |
|
|
};
|
183 |
|
|
|
184 |
|
|
/*
|
185 |
|
|
* We have two global cpuset mutexes below. They can nest.
|
186 |
|
|
* It is ok to first take manage_mutex, then nest callback_mutex. We also
|
187 |
|
|
* require taking task_lock() when dereferencing a tasks cpuset pointer.
|
188 |
|
|
* See "The task_lock() exception", at the end of this comment.
|
189 |
|
|
*
|
190 |
|
|
* A task must hold both mutexes to modify cpusets. If a task
|
191 |
|
|
* holds manage_mutex, then it blocks others wanting that mutex,
|
192 |
|
|
* ensuring that it is the only task able to also acquire callback_mutex
|
193 |
|
|
* and be able to modify cpusets. It can perform various checks on
|
194 |
|
|
* the cpuset structure first, knowing nothing will change. It can
|
195 |
|
|
* also allocate memory while just holding manage_mutex. While it is
|
196 |
|
|
* performing these checks, various callback routines can briefly
|
197 |
|
|
* acquire callback_mutex to query cpusets. Once it is ready to make
|
198 |
|
|
* the changes, it takes callback_mutex, blocking everyone else.
|
199 |
|
|
*
|
200 |
|
|
* Calls to the kernel memory allocator can not be made while holding
|
201 |
|
|
* callback_mutex, as that would risk double tripping on callback_mutex
|
202 |
|
|
* from one of the callbacks into the cpuset code from within
|
203 |
|
|
* __alloc_pages().
|
204 |
|
|
*
|
205 |
|
|
* If a task is only holding callback_mutex, then it has read-only
|
206 |
|
|
* access to cpusets.
|
207 |
|
|
*
|
208 |
|
|
* The task_struct fields mems_allowed and mems_generation may only
|
209 |
|
|
* be accessed in the context of that task, so require no locks.
|
210 |
|
|
*
|
211 |
|
|
* Any task can increment and decrement the count field without lock.
|
212 |
|
|
* So in general, code holding manage_mutex or callback_mutex can't rely
|
213 |
|
|
* on the count field not changing. However, if the count goes to
|
214 |
|
|
* zero, then only attach_task(), which holds both mutexes, can
|
215 |
|
|
* increment it again. Because a count of zero means that no tasks
|
216 |
|
|
* are currently attached, therefore there is no way a task attached
|
217 |
|
|
* to that cpuset can fork (the other way to increment the count).
|
218 |
|
|
* So code holding manage_mutex or callback_mutex can safely assume that
|
219 |
|
|
* if the count is zero, it will stay zero. Similarly, if a task
|
220 |
|
|
* holds manage_mutex or callback_mutex on a cpuset with zero count, it
|
221 |
|
|
* knows that the cpuset won't be removed, as cpuset_rmdir() needs
|
222 |
|
|
* both of those mutexes.
|
223 |
|
|
*
|
224 |
|
|
* The cpuset_common_file_write handler for operations that modify
|
225 |
|
|
* the cpuset hierarchy holds manage_mutex across the entire operation,
|
226 |
|
|
* single threading all such cpuset modifications across the system.
|
227 |
|
|
*
|
228 |
|
|
* The cpuset_common_file_read() handlers only hold callback_mutex across
|
229 |
|
|
* small pieces of code, such as when reading out possibly multi-word
|
230 |
|
|
* cpumasks and nodemasks.
|
231 |
|
|
*
|
232 |
|
|
* The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
|
233 |
|
|
* (usually) take either mutex. These are the two most performance
|
234 |
|
|
* critical pieces of code here. The exception occurs on cpuset_exit(),
|
235 |
|
|
* when a task in a notify_on_release cpuset exits. Then manage_mutex
|
236 |
|
|
* is taken, and if the cpuset count is zero, a usermode call made
|
237 |
|
|
* to /sbin/cpuset_release_agent with the name of the cpuset (path
|
238 |
|
|
* relative to the root of cpuset file system) as the argument.
|
239 |
|
|
*
|
240 |
|
|
* A cpuset can only be deleted if both its 'count' of using tasks
|
241 |
|
|
* is zero, and its list of 'children' cpusets is empty. Since all
|
242 |
|
|
* tasks in the system use _some_ cpuset, and since there is always at
|
243 |
|
|
* least one task in the system (init), therefore, top_cpuset
|
244 |
|
|
* always has either children cpusets and/or using tasks. So we don't
|
245 |
|
|
* need a special hack to ensure that top_cpuset cannot be deleted.
|
246 |
|
|
*
|
247 |
|
|
* The above "Tale of Two Semaphores" would be complete, but for:
|
248 |
|
|
*
|
249 |
|
|
* The task_lock() exception
|
250 |
|
|
*
|
251 |
|
|
* The need for this exception arises from the action of attach_task(),
|
252 |
|
|
* which overwrites one tasks cpuset pointer with another. It does
|
253 |
|
|
* so using both mutexes, however there are several performance
|
254 |
|
|
* critical places that need to reference task->cpuset without the
|
255 |
|
|
* expense of grabbing a system global mutex. Therefore except as
|
256 |
|
|
* noted below, when dereferencing or, as in attach_task(), modifying
|
257 |
|
|
* a tasks cpuset pointer we use task_lock(), which acts on a spinlock
|
258 |
|
|
* (task->alloc_lock) already in the task_struct routinely used for
|
259 |
|
|
* such matters.
|
260 |
|
|
*
|
261 |
|
|
* P.S. One more locking exception. RCU is used to guard the
|
262 |
|
|
* update of a tasks cpuset pointer by attach_task() and the
|
263 |
|
|
* access of task->cpuset->mems_generation via that pointer in
|
264 |
|
|
* the routine cpuset_update_task_memory_state().
|
265 |
|
|
*/
|
266 |
|
|
|
267 |
|
|
static DEFINE_MUTEX(callback_mutex);
|
268 |
|
|
|
269 |
|
|
/* This is ugly, but preserves the userspace API for existing cpuset
|
270 |
|
|
* users. If someone tries to mount the "cpuset" filesystem, we
|
271 |
|
|
* silently switch it to mount "cgroup" instead */
|
272 |
|
|
static int cpuset_get_sb(struct file_system_type *fs_type,
|
273 |
|
|
int flags, const char *unused_dev_name,
|
274 |
|
|
void *data, struct vfsmount *mnt)
|
275 |
|
|
{
|
276 |
|
|
struct file_system_type *cgroup_fs = get_fs_type("cgroup");
|
277 |
|
|
int ret = -ENODEV;
|
278 |
|
|
if (cgroup_fs) {
|
279 |
|
|
char mountopts[] =
|
280 |
|
|
"cpuset,noprefix,"
|
281 |
|
|
"release_agent=/sbin/cpuset_release_agent";
|
282 |
|
|
ret = cgroup_fs->get_sb(cgroup_fs, flags,
|
283 |
|
|
unused_dev_name, mountopts, mnt);
|
284 |
|
|
put_filesystem(cgroup_fs);
|
285 |
|
|
}
|
286 |
|
|
return ret;
|
287 |
|
|
}
|
288 |
|
|
|
289 |
|
|
static struct file_system_type cpuset_fs_type = {
|
290 |
|
|
.name = "cpuset",
|
291 |
|
|
.get_sb = cpuset_get_sb,
|
292 |
|
|
};
|
293 |
|
|
|
294 |
|
|
/*
|
295 |
|
|
* Return in *pmask the portion of a cpusets's cpus_allowed that
|
296 |
|
|
* are online. If none are online, walk up the cpuset hierarchy
|
297 |
|
|
* until we find one that does have some online cpus. If we get
|
298 |
|
|
* all the way to the top and still haven't found any online cpus,
|
299 |
|
|
* return cpu_online_map. Or if passed a NULL cs from an exit'ing
|
300 |
|
|
* task, return cpu_online_map.
|
301 |
|
|
*
|
302 |
|
|
* One way or another, we guarantee to return some non-empty subset
|
303 |
|
|
* of cpu_online_map.
|
304 |
|
|
*
|
305 |
|
|
* Call with callback_mutex held.
|
306 |
|
|
*/
|
307 |
|
|
|
308 |
|
|
static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
|
309 |
|
|
{
|
310 |
|
|
while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
|
311 |
|
|
cs = cs->parent;
|
312 |
|
|
if (cs)
|
313 |
|
|
cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
|
314 |
|
|
else
|
315 |
|
|
*pmask = cpu_online_map;
|
316 |
|
|
BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
|
317 |
|
|
}
|
318 |
|
|
|
319 |
|
|
/*
|
320 |
|
|
* Return in *pmask the portion of a cpusets's mems_allowed that
|
321 |
|
|
* are online, with memory. If none are online with memory, walk
|
322 |
|
|
* up the cpuset hierarchy until we find one that does have some
|
323 |
|
|
* online mems. If we get all the way to the top and still haven't
|
324 |
|
|
* found any online mems, return node_states[N_HIGH_MEMORY].
|
325 |
|
|
*
|
326 |
|
|
* One way or another, we guarantee to return some non-empty subset
|
327 |
|
|
* of node_states[N_HIGH_MEMORY].
|
328 |
|
|
*
|
329 |
|
|
* Call with callback_mutex held.
|
330 |
|
|
*/
|
331 |
|
|
|
332 |
|
|
static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
|
333 |
|
|
{
|
334 |
|
|
while (cs && !nodes_intersects(cs->mems_allowed,
|
335 |
|
|
node_states[N_HIGH_MEMORY]))
|
336 |
|
|
cs = cs->parent;
|
337 |
|
|
if (cs)
|
338 |
|
|
nodes_and(*pmask, cs->mems_allowed,
|
339 |
|
|
node_states[N_HIGH_MEMORY]);
|
340 |
|
|
else
|
341 |
|
|
*pmask = node_states[N_HIGH_MEMORY];
|
342 |
|
|
BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
|
343 |
|
|
}
|
344 |
|
|
|
345 |
|
|
/**
|
346 |
|
|
* cpuset_update_task_memory_state - update task memory placement
|
347 |
|
|
*
|
348 |
|
|
* If the current tasks cpusets mems_allowed changed behind our
|
349 |
|
|
* backs, update current->mems_allowed, mems_generation and task NUMA
|
350 |
|
|
* mempolicy to the new value.
|
351 |
|
|
*
|
352 |
|
|
* Task mempolicy is updated by rebinding it relative to the
|
353 |
|
|
* current->cpuset if a task has its memory placement changed.
|
354 |
|
|
* Do not call this routine if in_interrupt().
|
355 |
|
|
*
|
356 |
|
|
* Call without callback_mutex or task_lock() held. May be
|
357 |
|
|
* called with or without manage_mutex held. Thanks in part to
|
358 |
|
|
* 'the_top_cpuset_hack', the tasks cpuset pointer will never
|
359 |
|
|
* be NULL. This routine also might acquire callback_mutex and
|
360 |
|
|
* current->mm->mmap_sem during call.
|
361 |
|
|
*
|
362 |
|
|
* Reading current->cpuset->mems_generation doesn't need task_lock
|
363 |
|
|
* to guard the current->cpuset derefence, because it is guarded
|
364 |
|
|
* from concurrent freeing of current->cpuset by attach_task(),
|
365 |
|
|
* using RCU.
|
366 |
|
|
*
|
367 |
|
|
* The rcu_dereference() is technically probably not needed,
|
368 |
|
|
* as I don't actually mind if I see a new cpuset pointer but
|
369 |
|
|
* an old value of mems_generation. However this really only
|
370 |
|
|
* matters on alpha systems using cpusets heavily. If I dropped
|
371 |
|
|
* that rcu_dereference(), it would save them a memory barrier.
|
372 |
|
|
* For all other arch's, rcu_dereference is a no-op anyway, and for
|
373 |
|
|
* alpha systems not using cpusets, another planned optimization,
|
374 |
|
|
* avoiding the rcu critical section for tasks in the root cpuset
|
375 |
|
|
* which is statically allocated, so can't vanish, will make this
|
376 |
|
|
* irrelevant. Better to use RCU as intended, than to engage in
|
377 |
|
|
* some cute trick to save a memory barrier that is impossible to
|
378 |
|
|
* test, for alpha systems using cpusets heavily, which might not
|
379 |
|
|
* even exist.
|
380 |
|
|
*
|
381 |
|
|
* This routine is needed to update the per-task mems_allowed data,
|
382 |
|
|
* within the tasks context, when it is trying to allocate memory
|
383 |
|
|
* (in various mm/mempolicy.c routines) and notices that some other
|
384 |
|
|
* task has been modifying its cpuset.
|
385 |
|
|
*/
|
386 |
|
|
|
387 |
|
|
void cpuset_update_task_memory_state(void)
|
388 |
|
|
{
|
389 |
|
|
int my_cpusets_mem_gen;
|
390 |
|
|
struct task_struct *tsk = current;
|
391 |
|
|
struct cpuset *cs;
|
392 |
|
|
|
393 |
|
|
if (task_cs(tsk) == &top_cpuset) {
|
394 |
|
|
/* Don't need rcu for top_cpuset. It's never freed. */
|
395 |
|
|
my_cpusets_mem_gen = top_cpuset.mems_generation;
|
396 |
|
|
} else {
|
397 |
|
|
rcu_read_lock();
|
398 |
|
|
my_cpusets_mem_gen = task_cs(current)->mems_generation;
|
399 |
|
|
rcu_read_unlock();
|
400 |
|
|
}
|
401 |
|
|
|
402 |
|
|
if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
|
403 |
|
|
mutex_lock(&callback_mutex);
|
404 |
|
|
task_lock(tsk);
|
405 |
|
|
cs = task_cs(tsk); /* Maybe changed when task not locked */
|
406 |
|
|
guarantee_online_mems(cs, &tsk->mems_allowed);
|
407 |
|
|
tsk->cpuset_mems_generation = cs->mems_generation;
|
408 |
|
|
if (is_spread_page(cs))
|
409 |
|
|
tsk->flags |= PF_SPREAD_PAGE;
|
410 |
|
|
else
|
411 |
|
|
tsk->flags &= ~PF_SPREAD_PAGE;
|
412 |
|
|
if (is_spread_slab(cs))
|
413 |
|
|
tsk->flags |= PF_SPREAD_SLAB;
|
414 |
|
|
else
|
415 |
|
|
tsk->flags &= ~PF_SPREAD_SLAB;
|
416 |
|
|
task_unlock(tsk);
|
417 |
|
|
mutex_unlock(&callback_mutex);
|
418 |
|
|
mpol_rebind_task(tsk, &tsk->mems_allowed);
|
419 |
|
|
}
|
420 |
|
|
}
|
421 |
|
|
|
422 |
|
|
/*
|
423 |
|
|
* is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
|
424 |
|
|
*
|
425 |
|
|
* One cpuset is a subset of another if all its allowed CPUs and
|
426 |
|
|
* Memory Nodes are a subset of the other, and its exclusive flags
|
427 |
|
|
* are only set if the other's are set. Call holding manage_mutex.
|
428 |
|
|
*/
|
429 |
|
|
|
430 |
|
|
static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
|
431 |
|
|
{
|
432 |
|
|
return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
|
433 |
|
|
nodes_subset(p->mems_allowed, q->mems_allowed) &&
|
434 |
|
|
is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
|
435 |
|
|
is_mem_exclusive(p) <= is_mem_exclusive(q);
|
436 |
|
|
}
|
437 |
|
|
|
438 |
|
|
/*
|
439 |
|
|
* validate_change() - Used to validate that any proposed cpuset change
|
440 |
|
|
* follows the structural rules for cpusets.
|
441 |
|
|
*
|
442 |
|
|
* If we replaced the flag and mask values of the current cpuset
|
443 |
|
|
* (cur) with those values in the trial cpuset (trial), would
|
444 |
|
|
* our various subset and exclusive rules still be valid? Presumes
|
445 |
|
|
* manage_mutex held.
|
446 |
|
|
*
|
447 |
|
|
* 'cur' is the address of an actual, in-use cpuset. Operations
|
448 |
|
|
* such as list traversal that depend on the actual address of the
|
449 |
|
|
* cpuset in the list must use cur below, not trial.
|
450 |
|
|
*
|
451 |
|
|
* 'trial' is the address of bulk structure copy of cur, with
|
452 |
|
|
* perhaps one or more of the fields cpus_allowed, mems_allowed,
|
453 |
|
|
* or flags changed to new, trial values.
|
454 |
|
|
*
|
455 |
|
|
* Return 0 if valid, -errno if not.
|
456 |
|
|
*/
|
457 |
|
|
|
458 |
|
|
static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
|
459 |
|
|
{
|
460 |
|
|
struct cgroup *cont;
|
461 |
|
|
struct cpuset *c, *par;
|
462 |
|
|
|
463 |
|
|
/* Each of our child cpusets must be a subset of us */
|
464 |
|
|
list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
|
465 |
|
|
if (!is_cpuset_subset(cgroup_cs(cont), trial))
|
466 |
|
|
return -EBUSY;
|
467 |
|
|
}
|
468 |
|
|
|
469 |
|
|
/* Remaining checks don't apply to root cpuset */
|
470 |
|
|
if (cur == &top_cpuset)
|
471 |
|
|
return 0;
|
472 |
|
|
|
473 |
|
|
par = cur->parent;
|
474 |
|
|
|
475 |
|
|
/* We must be a subset of our parent cpuset */
|
476 |
|
|
if (!is_cpuset_subset(trial, par))
|
477 |
|
|
return -EACCES;
|
478 |
|
|
|
479 |
|
|
/* If either I or some sibling (!= me) is exclusive, we can't overlap */
|
480 |
|
|
list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
|
481 |
|
|
c = cgroup_cs(cont);
|
482 |
|
|
if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
|
483 |
|
|
c != cur &&
|
484 |
|
|
cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
|
485 |
|
|
return -EINVAL;
|
486 |
|
|
if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
|
487 |
|
|
c != cur &&
|
488 |
|
|
nodes_intersects(trial->mems_allowed, c->mems_allowed))
|
489 |
|
|
return -EINVAL;
|
490 |
|
|
}
|
491 |
|
|
|
492 |
|
|
/* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
|
493 |
|
|
if (cgroup_task_count(cur->css.cgroup)) {
|
494 |
|
|
if (cpus_empty(trial->cpus_allowed) ||
|
495 |
|
|
nodes_empty(trial->mems_allowed)) {
|
496 |
|
|
return -ENOSPC;
|
497 |
|
|
}
|
498 |
|
|
}
|
499 |
|
|
|
500 |
|
|
return 0;
|
501 |
|
|
}
|
502 |
|
|
|
503 |
|
|
/*
|
504 |
|
|
* Helper routine for rebuild_sched_domains().
|
505 |
|
|
* Do cpusets a, b have overlapping cpus_allowed masks?
|
506 |
|
|
*/
|
507 |
|
|
|
508 |
|
|
static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
|
509 |
|
|
{
|
510 |
|
|
return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
|
511 |
|
|
}
|
512 |
|
|
|
513 |
|
|
/*
|
514 |
|
|
* rebuild_sched_domains()
|
515 |
|
|
*
|
516 |
|
|
* If the flag 'sched_load_balance' of any cpuset with non-empty
|
517 |
|
|
* 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
|
518 |
|
|
* which has that flag enabled, or if any cpuset with a non-empty
|
519 |
|
|
* 'cpus' is removed, then call this routine to rebuild the
|
520 |
|
|
* scheduler's dynamic sched domains.
|
521 |
|
|
*
|
522 |
|
|
* This routine builds a partial partition of the systems CPUs
|
523 |
|
|
* (the set of non-overlappping cpumask_t's in the array 'part'
|
524 |
|
|
* below), and passes that partial partition to the kernel/sched.c
|
525 |
|
|
* partition_sched_domains() routine, which will rebuild the
|
526 |
|
|
* schedulers load balancing domains (sched domains) as specified
|
527 |
|
|
* by that partial partition. A 'partial partition' is a set of
|
528 |
|
|
* non-overlapping subsets whose union is a subset of that set.
|
529 |
|
|
*
|
530 |
|
|
* See "What is sched_load_balance" in Documentation/cpusets.txt
|
531 |
|
|
* for a background explanation of this.
|
532 |
|
|
*
|
533 |
|
|
* Does not return errors, on the theory that the callers of this
|
534 |
|
|
* routine would rather not worry about failures to rebuild sched
|
535 |
|
|
* domains when operating in the severe memory shortage situations
|
536 |
|
|
* that could cause allocation failures below.
|
537 |
|
|
*
|
538 |
|
|
* Call with cgroup_mutex held. May take callback_mutex during
|
539 |
|
|
* call due to the kfifo_alloc() and kmalloc() calls. May nest
|
540 |
|
|
* a call to the lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
|
541 |
|
|
* Must not be called holding callback_mutex, because we must not
|
542 |
|
|
* call lock_cpu_hotplug() while holding callback_mutex. Elsewhere
|
543 |
|
|
* the kernel nests callback_mutex inside lock_cpu_hotplug() calls.
|
544 |
|
|
* So the reverse nesting would risk an ABBA deadlock.
|
545 |
|
|
*
|
546 |
|
|
* The three key local variables below are:
|
547 |
|
|
* q - a kfifo queue of cpuset pointers, used to implement a
|
548 |
|
|
* top-down scan of all cpusets. This scan loads a pointer
|
549 |
|
|
* to each cpuset marked is_sched_load_balance into the
|
550 |
|
|
* array 'csa'. For our purposes, rebuilding the schedulers
|
551 |
|
|
* sched domains, we can ignore !is_sched_load_balance cpusets.
|
552 |
|
|
* csa - (for CpuSet Array) Array of pointers to all the cpusets
|
553 |
|
|
* that need to be load balanced, for convenient iterative
|
554 |
|
|
* access by the subsequent code that finds the best partition,
|
555 |
|
|
* i.e the set of domains (subsets) of CPUs such that the
|
556 |
|
|
* cpus_allowed of every cpuset marked is_sched_load_balance
|
557 |
|
|
* is a subset of one of these domains, while there are as
|
558 |
|
|
* many such domains as possible, each as small as possible.
|
559 |
|
|
* doms - Conversion of 'csa' to an array of cpumasks, for passing to
|
560 |
|
|
* the kernel/sched.c routine partition_sched_domains() in a
|
561 |
|
|
* convenient format, that can be easily compared to the prior
|
562 |
|
|
* value to determine what partition elements (sched domains)
|
563 |
|
|
* were changed (added or removed.)
|
564 |
|
|
*
|
565 |
|
|
* Finding the best partition (set of domains):
|
566 |
|
|
* The triple nested loops below over i, j, k scan over the
|
567 |
|
|
* load balanced cpusets (using the array of cpuset pointers in
|
568 |
|
|
* csa[]) looking for pairs of cpusets that have overlapping
|
569 |
|
|
* cpus_allowed, but which don't have the same 'pn' partition
|
570 |
|
|
* number and gives them in the same partition number. It keeps
|
571 |
|
|
* looping on the 'restart' label until it can no longer find
|
572 |
|
|
* any such pairs.
|
573 |
|
|
*
|
574 |
|
|
* The union of the cpus_allowed masks from the set of
|
575 |
|
|
* all cpusets having the same 'pn' value then form the one
|
576 |
|
|
* element of the partition (one sched domain) to be passed to
|
577 |
|
|
* partition_sched_domains().
|
578 |
|
|
*/
|
579 |
|
|
|
580 |
|
|
static void rebuild_sched_domains(void)
|
581 |
|
|
{
|
582 |
|
|
struct kfifo *q; /* queue of cpusets to be scanned */
|
583 |
|
|
struct cpuset *cp; /* scans q */
|
584 |
|
|
struct cpuset **csa; /* array of all cpuset ptrs */
|
585 |
|
|
int csn; /* how many cpuset ptrs in csa so far */
|
586 |
|
|
int i, j, k; /* indices for partition finding loops */
|
587 |
|
|
cpumask_t *doms; /* resulting partition; i.e. sched domains */
|
588 |
|
|
int ndoms; /* number of sched domains in result */
|
589 |
|
|
int nslot; /* next empty doms[] cpumask_t slot */
|
590 |
|
|
|
591 |
|
|
q = NULL;
|
592 |
|
|
csa = NULL;
|
593 |
|
|
doms = NULL;
|
594 |
|
|
|
595 |
|
|
/* Special case for the 99% of systems with one, full, sched domain */
|
596 |
|
|
if (is_sched_load_balance(&top_cpuset)) {
|
597 |
|
|
ndoms = 1;
|
598 |
|
|
doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
|
599 |
|
|
if (!doms)
|
600 |
|
|
goto rebuild;
|
601 |
|
|
*doms = top_cpuset.cpus_allowed;
|
602 |
|
|
goto rebuild;
|
603 |
|
|
}
|
604 |
|
|
|
605 |
|
|
q = kfifo_alloc(number_of_cpusets * sizeof(cp), GFP_KERNEL, NULL);
|
606 |
|
|
if (IS_ERR(q))
|
607 |
|
|
goto done;
|
608 |
|
|
csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
|
609 |
|
|
if (!csa)
|
610 |
|
|
goto done;
|
611 |
|
|
csn = 0;
|
612 |
|
|
|
613 |
|
|
cp = &top_cpuset;
|
614 |
|
|
__kfifo_put(q, (void *)&cp, sizeof(cp));
|
615 |
|
|
while (__kfifo_get(q, (void *)&cp, sizeof(cp))) {
|
616 |
|
|
struct cgroup *cont;
|
617 |
|
|
struct cpuset *child; /* scans child cpusets of cp */
|
618 |
|
|
if (is_sched_load_balance(cp))
|
619 |
|
|
csa[csn++] = cp;
|
620 |
|
|
list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
|
621 |
|
|
child = cgroup_cs(cont);
|
622 |
|
|
__kfifo_put(q, (void *)&child, sizeof(cp));
|
623 |
|
|
}
|
624 |
|
|
}
|
625 |
|
|
|
626 |
|
|
for (i = 0; i < csn; i++)
|
627 |
|
|
csa[i]->pn = i;
|
628 |
|
|
ndoms = csn;
|
629 |
|
|
|
630 |
|
|
restart:
|
631 |
|
|
/* Find the best partition (set of sched domains) */
|
632 |
|
|
for (i = 0; i < csn; i++) {
|
633 |
|
|
struct cpuset *a = csa[i];
|
634 |
|
|
int apn = a->pn;
|
635 |
|
|
|
636 |
|
|
for (j = 0; j < csn; j++) {
|
637 |
|
|
struct cpuset *b = csa[j];
|
638 |
|
|
int bpn = b->pn;
|
639 |
|
|
|
640 |
|
|
if (apn != bpn && cpusets_overlap(a, b)) {
|
641 |
|
|
for (k = 0; k < csn; k++) {
|
642 |
|
|
struct cpuset *c = csa[k];
|
643 |
|
|
|
644 |
|
|
if (c->pn == bpn)
|
645 |
|
|
c->pn = apn;
|
646 |
|
|
}
|
647 |
|
|
ndoms--; /* one less element */
|
648 |
|
|
goto restart;
|
649 |
|
|
}
|
650 |
|
|
}
|
651 |
|
|
}
|
652 |
|
|
|
653 |
|
|
/* Convert <csn, csa> to <ndoms, doms> */
|
654 |
|
|
doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
|
655 |
|
|
if (!doms)
|
656 |
|
|
goto rebuild;
|
657 |
|
|
|
658 |
|
|
for (nslot = 0, i = 0; i < csn; i++) {
|
659 |
|
|
struct cpuset *a = csa[i];
|
660 |
|
|
int apn = a->pn;
|
661 |
|
|
|
662 |
|
|
if (apn >= 0) {
|
663 |
|
|
cpumask_t *dp = doms + nslot;
|
664 |
|
|
|
665 |
|
|
if (nslot == ndoms) {
|
666 |
|
|
static int warnings = 10;
|
667 |
|
|
if (warnings) {
|
668 |
|
|
printk(KERN_WARNING
|
669 |
|
|
"rebuild_sched_domains confused:"
|
670 |
|
|
" nslot %d, ndoms %d, csn %d, i %d,"
|
671 |
|
|
" apn %d\n",
|
672 |
|
|
nslot, ndoms, csn, i, apn);
|
673 |
|
|
warnings--;
|
674 |
|
|
}
|
675 |
|
|
continue;
|
676 |
|
|
}
|
677 |
|
|
|
678 |
|
|
cpus_clear(*dp);
|
679 |
|
|
for (j = i; j < csn; j++) {
|
680 |
|
|
struct cpuset *b = csa[j];
|
681 |
|
|
|
682 |
|
|
if (apn == b->pn) {
|
683 |
|
|
cpus_or(*dp, *dp, b->cpus_allowed);
|
684 |
|
|
b->pn = -1;
|
685 |
|
|
}
|
686 |
|
|
}
|
687 |
|
|
nslot++;
|
688 |
|
|
}
|
689 |
|
|
}
|
690 |
|
|
BUG_ON(nslot != ndoms);
|
691 |
|
|
|
692 |
|
|
rebuild:
|
693 |
|
|
/* Have scheduler rebuild sched domains */
|
694 |
|
|
lock_cpu_hotplug();
|
695 |
|
|
partition_sched_domains(ndoms, doms);
|
696 |
|
|
unlock_cpu_hotplug();
|
697 |
|
|
|
698 |
|
|
done:
|
699 |
|
|
if (q && !IS_ERR(q))
|
700 |
|
|
kfifo_free(q);
|
701 |
|
|
kfree(csa);
|
702 |
|
|
/* Don't kfree(doms) -- partition_sched_domains() does that. */
|
703 |
|
|
}
|
704 |
|
|
|
705 |
|
|
static inline int started_after_time(struct task_struct *t1,
|
706 |
|
|
struct timespec *time,
|
707 |
|
|
struct task_struct *t2)
|
708 |
|
|
{
|
709 |
|
|
int start_diff = timespec_compare(&t1->start_time, time);
|
710 |
|
|
if (start_diff > 0) {
|
711 |
|
|
return 1;
|
712 |
|
|
} else if (start_diff < 0) {
|
713 |
|
|
return 0;
|
714 |
|
|
} else {
|
715 |
|
|
/*
|
716 |
|
|
* Arbitrarily, if two processes started at the same
|
717 |
|
|
* time, we'll say that the lower pointer value
|
718 |
|
|
* started first. Note that t2 may have exited by now
|
719 |
|
|
* so this may not be a valid pointer any longer, but
|
720 |
|
|
* that's fine - it still serves to distinguish
|
721 |
|
|
* between two tasks started (effectively)
|
722 |
|
|
* simultaneously.
|
723 |
|
|
*/
|
724 |
|
|
return t1 > t2;
|
725 |
|
|
}
|
726 |
|
|
}
|
727 |
|
|
|
728 |
|
|
static inline int started_after(void *p1, void *p2)
|
729 |
|
|
{
|
730 |
|
|
struct task_struct *t1 = p1;
|
731 |
|
|
struct task_struct *t2 = p2;
|
732 |
|
|
return started_after_time(t1, &t2->start_time, t2);
|
733 |
|
|
}
|
734 |
|
|
|
735 |
|
|
/*
|
736 |
|
|
* Call with manage_mutex held. May take callback_mutex during call.
|
737 |
|
|
*/
|
738 |
|
|
|
739 |
|
|
static int update_cpumask(struct cpuset *cs, char *buf)
|
740 |
|
|
{
|
741 |
|
|
struct cpuset trialcs;
|
742 |
|
|
int retval, i;
|
743 |
|
|
int is_load_balanced;
|
744 |
|
|
struct cgroup_iter it;
|
745 |
|
|
struct cgroup *cgrp = cs->css.cgroup;
|
746 |
|
|
struct task_struct *p, *dropped;
|
747 |
|
|
/* Never dereference latest_task, since it's not refcounted */
|
748 |
|
|
struct task_struct *latest_task = NULL;
|
749 |
|
|
struct ptr_heap heap;
|
750 |
|
|
struct timespec latest_time = { 0, 0 };
|
751 |
|
|
|
752 |
|
|
/* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
|
753 |
|
|
if (cs == &top_cpuset)
|
754 |
|
|
return -EACCES;
|
755 |
|
|
|
756 |
|
|
trialcs = *cs;
|
757 |
|
|
|
758 |
|
|
/*
|
759 |
|
|
* An empty cpus_allowed is ok iff there are no tasks in the cpuset.
|
760 |
|
|
* Since cpulist_parse() fails on an empty mask, we special case
|
761 |
|
|
* that parsing. The validate_change() call ensures that cpusets
|
762 |
|
|
* with tasks have cpus.
|
763 |
|
|
*/
|
764 |
|
|
buf = strstrip(buf);
|
765 |
|
|
if (!*buf) {
|
766 |
|
|
cpus_clear(trialcs.cpus_allowed);
|
767 |
|
|
} else {
|
768 |
|
|
retval = cpulist_parse(buf, trialcs.cpus_allowed);
|
769 |
|
|
if (retval < 0)
|
770 |
|
|
return retval;
|
771 |
|
|
}
|
772 |
|
|
cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
|
773 |
|
|
retval = validate_change(cs, &trialcs);
|
774 |
|
|
if (retval < 0)
|
775 |
|
|
return retval;
|
776 |
|
|
|
777 |
|
|
/* Nothing to do if the cpus didn't change */
|
778 |
|
|
if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
|
779 |
|
|
return 0;
|
780 |
|
|
retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, &started_after);
|
781 |
|
|
if (retval)
|
782 |
|
|
return retval;
|
783 |
|
|
|
784 |
|
|
is_load_balanced = is_sched_load_balance(&trialcs);
|
785 |
|
|
|
786 |
|
|
mutex_lock(&callback_mutex);
|
787 |
|
|
cs->cpus_allowed = trialcs.cpus_allowed;
|
788 |
|
|
mutex_unlock(&callback_mutex);
|
789 |
|
|
|
790 |
|
|
again:
|
791 |
|
|
/*
|
792 |
|
|
* Scan tasks in the cpuset, and update the cpumasks of any
|
793 |
|
|
* that need an update. Since we can't call set_cpus_allowed()
|
794 |
|
|
* while holding tasklist_lock, gather tasks to be processed
|
795 |
|
|
* in a heap structure. If the statically-sized heap fills up,
|
796 |
|
|
* overflow tasks that started later, and in future iterations
|
797 |
|
|
* only consider tasks that started after the latest task in
|
798 |
|
|
* the previous pass. This guarantees forward progress and
|
799 |
|
|
* that we don't miss any tasks
|
800 |
|
|
*/
|
801 |
|
|
heap.size = 0;
|
802 |
|
|
cgroup_iter_start(cgrp, &it);
|
803 |
|
|
while ((p = cgroup_iter_next(cgrp, &it))) {
|
804 |
|
|
/* Only affect tasks that don't have the right cpus_allowed */
|
805 |
|
|
if (cpus_equal(p->cpus_allowed, cs->cpus_allowed))
|
806 |
|
|
continue;
|
807 |
|
|
/*
|
808 |
|
|
* Only process tasks that started after the last task
|
809 |
|
|
* we processed
|
810 |
|
|
*/
|
811 |
|
|
if (!started_after_time(p, &latest_time, latest_task))
|
812 |
|
|
continue;
|
813 |
|
|
dropped = heap_insert(&heap, p);
|
814 |
|
|
if (dropped == NULL) {
|
815 |
|
|
get_task_struct(p);
|
816 |
|
|
} else if (dropped != p) {
|
817 |
|
|
get_task_struct(p);
|
818 |
|
|
put_task_struct(dropped);
|
819 |
|
|
}
|
820 |
|
|
}
|
821 |
|
|
cgroup_iter_end(cgrp, &it);
|
822 |
|
|
if (heap.size) {
|
823 |
|
|
for (i = 0; i < heap.size; i++) {
|
824 |
|
|
struct task_struct *p = heap.ptrs[i];
|
825 |
|
|
if (i == 0) {
|
826 |
|
|
latest_time = p->start_time;
|
827 |
|
|
latest_task = p;
|
828 |
|
|
}
|
829 |
|
|
set_cpus_allowed(p, cs->cpus_allowed);
|
830 |
|
|
put_task_struct(p);
|
831 |
|
|
}
|
832 |
|
|
/*
|
833 |
|
|
* If we had to process any tasks at all, scan again
|
834 |
|
|
* in case some of them were in the middle of forking
|
835 |
|
|
* children that didn't notice the new cpumask
|
836 |
|
|
* restriction. Not the most efficient way to do it,
|
837 |
|
|
* but it avoids having to take callback_mutex in the
|
838 |
|
|
* fork path
|
839 |
|
|
*/
|
840 |
|
|
goto again;
|
841 |
|
|
}
|
842 |
|
|
heap_free(&heap);
|
843 |
|
|
if (is_load_balanced)
|
844 |
|
|
rebuild_sched_domains();
|
845 |
|
|
|
846 |
|
|
return 0;
|
847 |
|
|
}
|
848 |
|
|
|
849 |
|
|
/*
|
850 |
|
|
* cpuset_migrate_mm
|
851 |
|
|
*
|
852 |
|
|
* Migrate memory region from one set of nodes to another.
|
853 |
|
|
*
|
854 |
|
|
* Temporarilly set tasks mems_allowed to target nodes of migration,
|
855 |
|
|
* so that the migration code can allocate pages on these nodes.
|
856 |
|
|
*
|
857 |
|
|
* Call holding manage_mutex, so our current->cpuset won't change
|
858 |
|
|
* during this call, as manage_mutex holds off any attach_task()
|
859 |
|
|
* calls. Therefore we don't need to take task_lock around the
|
860 |
|
|
* call to guarantee_online_mems(), as we know no one is changing
|
861 |
|
|
* our tasks cpuset.
|
862 |
|
|
*
|
863 |
|
|
* Hold callback_mutex around the two modifications of our tasks
|
864 |
|
|
* mems_allowed to synchronize with cpuset_mems_allowed().
|
865 |
|
|
*
|
866 |
|
|
* While the mm_struct we are migrating is typically from some
|
867 |
|
|
* other task, the task_struct mems_allowed that we are hacking
|
868 |
|
|
* is for our current task, which must allocate new pages for that
|
869 |
|
|
* migrating memory region.
|
870 |
|
|
*
|
871 |
|
|
* We call cpuset_update_task_memory_state() before hacking
|
872 |
|
|
* our tasks mems_allowed, so that we are assured of being in
|
873 |
|
|
* sync with our tasks cpuset, and in particular, callbacks to
|
874 |
|
|
* cpuset_update_task_memory_state() from nested page allocations
|
875 |
|
|
* won't see any mismatch of our cpuset and task mems_generation
|
876 |
|
|
* values, so won't overwrite our hacked tasks mems_allowed
|
877 |
|
|
* nodemask.
|
878 |
|
|
*/
|
879 |
|
|
|
880 |
|
|
static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
|
881 |
|
|
const nodemask_t *to)
|
882 |
|
|
{
|
883 |
|
|
struct task_struct *tsk = current;
|
884 |
|
|
|
885 |
|
|
cpuset_update_task_memory_state();
|
886 |
|
|
|
887 |
|
|
mutex_lock(&callback_mutex);
|
888 |
|
|
tsk->mems_allowed = *to;
|
889 |
|
|
mutex_unlock(&callback_mutex);
|
890 |
|
|
|
891 |
|
|
do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
|
892 |
|
|
|
893 |
|
|
mutex_lock(&callback_mutex);
|
894 |
|
|
guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
|
895 |
|
|
mutex_unlock(&callback_mutex);
|
896 |
|
|
}
|
897 |
|
|
|
898 |
|
|
/*
|
899 |
|
|
* Handle user request to change the 'mems' memory placement
|
900 |
|
|
* of a cpuset. Needs to validate the request, update the
|
901 |
|
|
* cpusets mems_allowed and mems_generation, and for each
|
902 |
|
|
* task in the cpuset, rebind any vma mempolicies and if
|
903 |
|
|
* the cpuset is marked 'memory_migrate', migrate the tasks
|
904 |
|
|
* pages to the new memory.
|
905 |
|
|
*
|
906 |
|
|
* Call with manage_mutex held. May take callback_mutex during call.
|
907 |
|
|
* Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
|
908 |
|
|
* lock each such tasks mm->mmap_sem, scan its vma's and rebind
|
909 |
|
|
* their mempolicies to the cpusets new mems_allowed.
|
910 |
|
|
*/
|
911 |
|
|
|
912 |
|
|
static void *cpuset_being_rebound;
|
913 |
|
|
|
914 |
|
|
static int update_nodemask(struct cpuset *cs, char *buf)
|
915 |
|
|
{
|
916 |
|
|
struct cpuset trialcs;
|
917 |
|
|
nodemask_t oldmem;
|
918 |
|
|
struct task_struct *p;
|
919 |
|
|
struct mm_struct **mmarray;
|
920 |
|
|
int i, n, ntasks;
|
921 |
|
|
int migrate;
|
922 |
|
|
int fudge;
|
923 |
|
|
int retval;
|
924 |
|
|
struct cgroup_iter it;
|
925 |
|
|
|
926 |
|
|
/*
|
927 |
|
|
* top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
|
928 |
|
|
* it's read-only
|
929 |
|
|
*/
|
930 |
|
|
if (cs == &top_cpuset)
|
931 |
|
|
return -EACCES;
|
932 |
|
|
|
933 |
|
|
trialcs = *cs;
|
934 |
|
|
|
935 |
|
|
/*
|
936 |
|
|
* An empty mems_allowed is ok iff there are no tasks in the cpuset.
|
937 |
|
|
* Since nodelist_parse() fails on an empty mask, we special case
|
938 |
|
|
* that parsing. The validate_change() call ensures that cpusets
|
939 |
|
|
* with tasks have memory.
|
940 |
|
|
*/
|
941 |
|
|
buf = strstrip(buf);
|
942 |
|
|
if (!*buf) {
|
943 |
|
|
nodes_clear(trialcs.mems_allowed);
|
944 |
|
|
} else {
|
945 |
|
|
retval = nodelist_parse(buf, trialcs.mems_allowed);
|
946 |
|
|
if (retval < 0)
|
947 |
|
|
goto done;
|
948 |
|
|
}
|
949 |
|
|
nodes_and(trialcs.mems_allowed, trialcs.mems_allowed,
|
950 |
|
|
node_states[N_HIGH_MEMORY]);
|
951 |
|
|
oldmem = cs->mems_allowed;
|
952 |
|
|
if (nodes_equal(oldmem, trialcs.mems_allowed)) {
|
953 |
|
|
retval = 0; /* Too easy - nothing to do */
|
954 |
|
|
goto done;
|
955 |
|
|
}
|
956 |
|
|
retval = validate_change(cs, &trialcs);
|
957 |
|
|
if (retval < 0)
|
958 |
|
|
goto done;
|
959 |
|
|
|
960 |
|
|
mutex_lock(&callback_mutex);
|
961 |
|
|
cs->mems_allowed = trialcs.mems_allowed;
|
962 |
|
|
cs->mems_generation = cpuset_mems_generation++;
|
963 |
|
|
mutex_unlock(&callback_mutex);
|
964 |
|
|
|
965 |
|
|
cpuset_being_rebound = cs; /* causes mpol_copy() rebind */
|
966 |
|
|
|
967 |
|
|
fudge = 10; /* spare mmarray[] slots */
|
968 |
|
|
fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
|
969 |
|
|
retval = -ENOMEM;
|
970 |
|
|
|
971 |
|
|
/*
|
972 |
|
|
* Allocate mmarray[] to hold mm reference for each task
|
973 |
|
|
* in cpuset cs. Can't kmalloc GFP_KERNEL while holding
|
974 |
|
|
* tasklist_lock. We could use GFP_ATOMIC, but with a
|
975 |
|
|
* few more lines of code, we can retry until we get a big
|
976 |
|
|
* enough mmarray[] w/o using GFP_ATOMIC.
|
977 |
|
|
*/
|
978 |
|
|
while (1) {
|
979 |
|
|
ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
|
980 |
|
|
ntasks += fudge;
|
981 |
|
|
mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
|
982 |
|
|
if (!mmarray)
|
983 |
|
|
goto done;
|
984 |
|
|
read_lock(&tasklist_lock); /* block fork */
|
985 |
|
|
if (cgroup_task_count(cs->css.cgroup) <= ntasks)
|
986 |
|
|
break; /* got enough */
|
987 |
|
|
read_unlock(&tasklist_lock); /* try again */
|
988 |
|
|
kfree(mmarray);
|
989 |
|
|
}
|
990 |
|
|
|
991 |
|
|
n = 0;
|
992 |
|
|
|
993 |
|
|
/* Load up mmarray[] with mm reference for each task in cpuset. */
|
994 |
|
|
cgroup_iter_start(cs->css.cgroup, &it);
|
995 |
|
|
while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
|
996 |
|
|
struct mm_struct *mm;
|
997 |
|
|
|
998 |
|
|
if (n >= ntasks) {
|
999 |
|
|
printk(KERN_WARNING
|
1000 |
|
|
"Cpuset mempolicy rebind incomplete.\n");
|
1001 |
|
|
break;
|
1002 |
|
|
}
|
1003 |
|
|
mm = get_task_mm(p);
|
1004 |
|
|
if (!mm)
|
1005 |
|
|
continue;
|
1006 |
|
|
mmarray[n++] = mm;
|
1007 |
|
|
}
|
1008 |
|
|
cgroup_iter_end(cs->css.cgroup, &it);
|
1009 |
|
|
read_unlock(&tasklist_lock);
|
1010 |
|
|
|
1011 |
|
|
/*
|
1012 |
|
|
* Now that we've dropped the tasklist spinlock, we can
|
1013 |
|
|
* rebind the vma mempolicies of each mm in mmarray[] to their
|
1014 |
|
|
* new cpuset, and release that mm. The mpol_rebind_mm()
|
1015 |
|
|
* call takes mmap_sem, which we couldn't take while holding
|
1016 |
|
|
* tasklist_lock. Forks can happen again now - the mpol_copy()
|
1017 |
|
|
* cpuset_being_rebound check will catch such forks, and rebind
|
1018 |
|
|
* their vma mempolicies too. Because we still hold the global
|
1019 |
|
|
* cpuset manage_mutex, we know that no other rebind effort will
|
1020 |
|
|
* be contending for the global variable cpuset_being_rebound.
|
1021 |
|
|
* It's ok if we rebind the same mm twice; mpol_rebind_mm()
|
1022 |
|
|
* is idempotent. Also migrate pages in each mm to new nodes.
|
1023 |
|
|
*/
|
1024 |
|
|
migrate = is_memory_migrate(cs);
|
1025 |
|
|
for (i = 0; i < n; i++) {
|
1026 |
|
|
struct mm_struct *mm = mmarray[i];
|
1027 |
|
|
|
1028 |
|
|
mpol_rebind_mm(mm, &cs->mems_allowed);
|
1029 |
|
|
if (migrate)
|
1030 |
|
|
cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed);
|
1031 |
|
|
mmput(mm);
|
1032 |
|
|
}
|
1033 |
|
|
|
1034 |
|
|
/* We're done rebinding vma's to this cpusets new mems_allowed. */
|
1035 |
|
|
kfree(mmarray);
|
1036 |
|
|
cpuset_being_rebound = NULL;
|
1037 |
|
|
retval = 0;
|
1038 |
|
|
done:
|
1039 |
|
|
return retval;
|
1040 |
|
|
}
|
1041 |
|
|
|
1042 |
|
|
int current_cpuset_is_being_rebound(void)
|
1043 |
|
|
{
|
1044 |
|
|
return task_cs(current) == cpuset_being_rebound;
|
1045 |
|
|
}
|
1046 |
|
|
|
1047 |
|
|
/*
|
1048 |
|
|
* Call with manage_mutex held.
|
1049 |
|
|
*/
|
1050 |
|
|
|
1051 |
|
|
static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
|
1052 |
|
|
{
|
1053 |
|
|
if (simple_strtoul(buf, NULL, 10) != 0)
|
1054 |
|
|
cpuset_memory_pressure_enabled = 1;
|
1055 |
|
|
else
|
1056 |
|
|
cpuset_memory_pressure_enabled = 0;
|
1057 |
|
|
return 0;
|
1058 |
|
|
}
|
1059 |
|
|
|
1060 |
|
|
/*
|
1061 |
|
|
* update_flag - read a 0 or a 1 in a file and update associated flag
|
1062 |
|
|
* bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
|
1063 |
|
|
* CS_SCHED_LOAD_BALANCE,
|
1064 |
|
|
* CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
|
1065 |
|
|
* CS_SPREAD_PAGE, CS_SPREAD_SLAB)
|
1066 |
|
|
* cs: the cpuset to update
|
1067 |
|
|
* buf: the buffer where we read the 0 or 1
|
1068 |
|
|
*
|
1069 |
|
|
* Call with manage_mutex held.
|
1070 |
|
|
*/
|
1071 |
|
|
|
1072 |
|
|
static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
|
1073 |
|
|
{
|
1074 |
|
|
int turning_on;
|
1075 |
|
|
struct cpuset trialcs;
|
1076 |
|
|
int err;
|
1077 |
|
|
int cpus_nonempty, balance_flag_changed;
|
1078 |
|
|
|
1079 |
|
|
turning_on = (simple_strtoul(buf, NULL, 10) != 0);
|
1080 |
|
|
|
1081 |
|
|
trialcs = *cs;
|
1082 |
|
|
if (turning_on)
|
1083 |
|
|
set_bit(bit, &trialcs.flags);
|
1084 |
|
|
else
|
1085 |
|
|
clear_bit(bit, &trialcs.flags);
|
1086 |
|
|
|
1087 |
|
|
err = validate_change(cs, &trialcs);
|
1088 |
|
|
if (err < 0)
|
1089 |
|
|
return err;
|
1090 |
|
|
|
1091 |
|
|
cpus_nonempty = !cpus_empty(trialcs.cpus_allowed);
|
1092 |
|
|
balance_flag_changed = (is_sched_load_balance(cs) !=
|
1093 |
|
|
is_sched_load_balance(&trialcs));
|
1094 |
|
|
|
1095 |
|
|
mutex_lock(&callback_mutex);
|
1096 |
|
|
cs->flags = trialcs.flags;
|
1097 |
|
|
mutex_unlock(&callback_mutex);
|
1098 |
|
|
|
1099 |
|
|
if (cpus_nonempty && balance_flag_changed)
|
1100 |
|
|
rebuild_sched_domains();
|
1101 |
|
|
|
1102 |
|
|
return 0;
|
1103 |
|
|
}
|
1104 |
|
|
|
1105 |
|
|
/*
|
1106 |
|
|
* Frequency meter - How fast is some event occurring?
|
1107 |
|
|
*
|
1108 |
|
|
* These routines manage a digitally filtered, constant time based,
|
1109 |
|
|
* event frequency meter. There are four routines:
|
1110 |
|
|
* fmeter_init() - initialize a frequency meter.
|
1111 |
|
|
* fmeter_markevent() - called each time the event happens.
|
1112 |
|
|
* fmeter_getrate() - returns the recent rate of such events.
|
1113 |
|
|
* fmeter_update() - internal routine used to update fmeter.
|
1114 |
|
|
*
|
1115 |
|
|
* A common data structure is passed to each of these routines,
|
1116 |
|
|
* which is used to keep track of the state required to manage the
|
1117 |
|
|
* frequency meter and its digital filter.
|
1118 |
|
|
*
|
1119 |
|
|
* The filter works on the number of events marked per unit time.
|
1120 |
|
|
* The filter is single-pole low-pass recursive (IIR). The time unit
|
1121 |
|
|
* is 1 second. Arithmetic is done using 32-bit integers scaled to
|
1122 |
|
|
* simulate 3 decimal digits of precision (multiplied by 1000).
|
1123 |
|
|
*
|
1124 |
|
|
* With an FM_COEF of 933, and a time base of 1 second, the filter
|
1125 |
|
|
* has a half-life of 10 seconds, meaning that if the events quit
|
1126 |
|
|
* happening, then the rate returned from the fmeter_getrate()
|
1127 |
|
|
* will be cut in half each 10 seconds, until it converges to zero.
|
1128 |
|
|
*
|
1129 |
|
|
* It is not worth doing a real infinitely recursive filter. If more
|
1130 |
|
|
* than FM_MAXTICKS ticks have elapsed since the last filter event,
|
1131 |
|
|
* just compute FM_MAXTICKS ticks worth, by which point the level
|
1132 |
|
|
* will be stable.
|
1133 |
|
|
*
|
1134 |
|
|
* Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
|
1135 |
|
|
* arithmetic overflow in the fmeter_update() routine.
|
1136 |
|
|
*
|
1137 |
|
|
* Given the simple 32 bit integer arithmetic used, this meter works
|
1138 |
|
|
* best for reporting rates between one per millisecond (msec) and
|
1139 |
|
|
* one per 32 (approx) seconds. At constant rates faster than one
|
1140 |
|
|
* per msec it maxes out at values just under 1,000,000. At constant
|
1141 |
|
|
* rates between one per msec, and one per second it will stabilize
|
1142 |
|
|
* to a value N*1000, where N is the rate of events per second.
|
1143 |
|
|
* At constant rates between one per second and one per 32 seconds,
|
1144 |
|
|
* it will be choppy, moving up on the seconds that have an event,
|
1145 |
|
|
* and then decaying until the next event. At rates slower than
|
1146 |
|
|
* about one in 32 seconds, it decays all the way back to zero between
|
1147 |
|
|
* each event.
|
1148 |
|
|
*/
|
1149 |
|
|
|
1150 |
|
|
#define FM_COEF 933 /* coefficient for half-life of 10 secs */
|
1151 |
|
|
#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
|
1152 |
|
|
#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
|
1153 |
|
|
#define FM_SCALE 1000 /* faux fixed point scale */
|
1154 |
|
|
|
1155 |
|
|
/* Initialize a frequency meter */
|
1156 |
|
|
static void fmeter_init(struct fmeter *fmp)
|
1157 |
|
|
{
|
1158 |
|
|
fmp->cnt = 0;
|
1159 |
|
|
fmp->val = 0;
|
1160 |
|
|
fmp->time = 0;
|
1161 |
|
|
spin_lock_init(&fmp->lock);
|
1162 |
|
|
}
|
1163 |
|
|
|
1164 |
|
|
/* Internal meter update - process cnt events and update value */
|
1165 |
|
|
static void fmeter_update(struct fmeter *fmp)
|
1166 |
|
|
{
|
1167 |
|
|
time_t now = get_seconds();
|
1168 |
|
|
time_t ticks = now - fmp->time;
|
1169 |
|
|
|
1170 |
|
|
if (ticks == 0)
|
1171 |
|
|
return;
|
1172 |
|
|
|
1173 |
|
|
ticks = min(FM_MAXTICKS, ticks);
|
1174 |
|
|
while (ticks-- > 0)
|
1175 |
|
|
fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
|
1176 |
|
|
fmp->time = now;
|
1177 |
|
|
|
1178 |
|
|
fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
|
1179 |
|
|
fmp->cnt = 0;
|
1180 |
|
|
}
|
1181 |
|
|
|
1182 |
|
|
/* Process any previous ticks, then bump cnt by one (times scale). */
|
1183 |
|
|
static void fmeter_markevent(struct fmeter *fmp)
|
1184 |
|
|
{
|
1185 |
|
|
spin_lock(&fmp->lock);
|
1186 |
|
|
fmeter_update(fmp);
|
1187 |
|
|
fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
|
1188 |
|
|
spin_unlock(&fmp->lock);
|
1189 |
|
|
}
|
1190 |
|
|
|
1191 |
|
|
/* Process any previous ticks, then return current value. */
|
1192 |
|
|
static int fmeter_getrate(struct fmeter *fmp)
|
1193 |
|
|
{
|
1194 |
|
|
int val;
|
1195 |
|
|
|
1196 |
|
|
spin_lock(&fmp->lock);
|
1197 |
|
|
fmeter_update(fmp);
|
1198 |
|
|
val = fmp->val;
|
1199 |
|
|
spin_unlock(&fmp->lock);
|
1200 |
|
|
return val;
|
1201 |
|
|
}
|
1202 |
|
|
|
1203 |
|
|
static int cpuset_can_attach(struct cgroup_subsys *ss,
|
1204 |
|
|
struct cgroup *cont, struct task_struct *tsk)
|
1205 |
|
|
{
|
1206 |
|
|
struct cpuset *cs = cgroup_cs(cont);
|
1207 |
|
|
|
1208 |
|
|
if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
|
1209 |
|
|
return -ENOSPC;
|
1210 |
|
|
|
1211 |
|
|
return security_task_setscheduler(tsk, 0, NULL);
|
1212 |
|
|
}
|
1213 |
|
|
|
1214 |
|
|
static void cpuset_attach(struct cgroup_subsys *ss,
|
1215 |
|
|
struct cgroup *cont, struct cgroup *oldcont,
|
1216 |
|
|
struct task_struct *tsk)
|
1217 |
|
|
{
|
1218 |
|
|
cpumask_t cpus;
|
1219 |
|
|
nodemask_t from, to;
|
1220 |
|
|
struct mm_struct *mm;
|
1221 |
|
|
struct cpuset *cs = cgroup_cs(cont);
|
1222 |
|
|
struct cpuset *oldcs = cgroup_cs(oldcont);
|
1223 |
|
|
|
1224 |
|
|
mutex_lock(&callback_mutex);
|
1225 |
|
|
guarantee_online_cpus(cs, &cpus);
|
1226 |
|
|
set_cpus_allowed(tsk, cpus);
|
1227 |
|
|
mutex_unlock(&callback_mutex);
|
1228 |
|
|
|
1229 |
|
|
from = oldcs->mems_allowed;
|
1230 |
|
|
to = cs->mems_allowed;
|
1231 |
|
|
mm = get_task_mm(tsk);
|
1232 |
|
|
if (mm) {
|
1233 |
|
|
mpol_rebind_mm(mm, &to);
|
1234 |
|
|
if (is_memory_migrate(cs))
|
1235 |
|
|
cpuset_migrate_mm(mm, &from, &to);
|
1236 |
|
|
mmput(mm);
|
1237 |
|
|
}
|
1238 |
|
|
|
1239 |
|
|
}
|
1240 |
|
|
|
1241 |
|
|
/* The various types of files and directories in a cpuset file system */
|
1242 |
|
|
|
1243 |
|
|
typedef enum {
|
1244 |
|
|
FILE_MEMORY_MIGRATE,
|
1245 |
|
|
FILE_CPULIST,
|
1246 |
|
|
FILE_MEMLIST,
|
1247 |
|
|
FILE_CPU_EXCLUSIVE,
|
1248 |
|
|
FILE_MEM_EXCLUSIVE,
|
1249 |
|
|
FILE_SCHED_LOAD_BALANCE,
|
1250 |
|
|
FILE_MEMORY_PRESSURE_ENABLED,
|
1251 |
|
|
FILE_MEMORY_PRESSURE,
|
1252 |
|
|
FILE_SPREAD_PAGE,
|
1253 |
|
|
FILE_SPREAD_SLAB,
|
1254 |
|
|
} cpuset_filetype_t;
|
1255 |
|
|
|
1256 |
|
|
static ssize_t cpuset_common_file_write(struct cgroup *cont,
|
1257 |
|
|
struct cftype *cft,
|
1258 |
|
|
struct file *file,
|
1259 |
|
|
const char __user *userbuf,
|
1260 |
|
|
size_t nbytes, loff_t *unused_ppos)
|
1261 |
|
|
{
|
1262 |
|
|
struct cpuset *cs = cgroup_cs(cont);
|
1263 |
|
|
cpuset_filetype_t type = cft->private;
|
1264 |
|
|
char *buffer;
|
1265 |
|
|
int retval = 0;
|
1266 |
|
|
|
1267 |
|
|
/* Crude upper limit on largest legitimate cpulist user might write. */
|
1268 |
|
|
if (nbytes > 100U + 6 * max(NR_CPUS, MAX_NUMNODES))
|
1269 |
|
|
return -E2BIG;
|
1270 |
|
|
|
1271 |
|
|
/* +1 for nul-terminator */
|
1272 |
|
|
if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
|
1273 |
|
|
return -ENOMEM;
|
1274 |
|
|
|
1275 |
|
|
if (copy_from_user(buffer, userbuf, nbytes)) {
|
1276 |
|
|
retval = -EFAULT;
|
1277 |
|
|
goto out1;
|
1278 |
|
|
}
|
1279 |
|
|
buffer[nbytes] = 0; /* nul-terminate */
|
1280 |
|
|
|
1281 |
|
|
cgroup_lock();
|
1282 |
|
|
|
1283 |
|
|
if (cgroup_is_removed(cont)) {
|
1284 |
|
|
retval = -ENODEV;
|
1285 |
|
|
goto out2;
|
1286 |
|
|
}
|
1287 |
|
|
|
1288 |
|
|
switch (type) {
|
1289 |
|
|
case FILE_CPULIST:
|
1290 |
|
|
retval = update_cpumask(cs, buffer);
|
1291 |
|
|
break;
|
1292 |
|
|
case FILE_MEMLIST:
|
1293 |
|
|
retval = update_nodemask(cs, buffer);
|
1294 |
|
|
break;
|
1295 |
|
|
case FILE_CPU_EXCLUSIVE:
|
1296 |
|
|
retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
|
1297 |
|
|
break;
|
1298 |
|
|
case FILE_MEM_EXCLUSIVE:
|
1299 |
|
|
retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
|
1300 |
|
|
break;
|
1301 |
|
|
case FILE_SCHED_LOAD_BALANCE:
|
1302 |
|
|
retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, buffer);
|
1303 |
|
|
break;
|
1304 |
|
|
case FILE_MEMORY_MIGRATE:
|
1305 |
|
|
retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
|
1306 |
|
|
break;
|
1307 |
|
|
case FILE_MEMORY_PRESSURE_ENABLED:
|
1308 |
|
|
retval = update_memory_pressure_enabled(cs, buffer);
|
1309 |
|
|
break;
|
1310 |
|
|
case FILE_MEMORY_PRESSURE:
|
1311 |
|
|
retval = -EACCES;
|
1312 |
|
|
break;
|
1313 |
|
|
case FILE_SPREAD_PAGE:
|
1314 |
|
|
retval = update_flag(CS_SPREAD_PAGE, cs, buffer);
|
1315 |
|
|
cs->mems_generation = cpuset_mems_generation++;
|
1316 |
|
|
break;
|
1317 |
|
|
case FILE_SPREAD_SLAB:
|
1318 |
|
|
retval = update_flag(CS_SPREAD_SLAB, cs, buffer);
|
1319 |
|
|
cs->mems_generation = cpuset_mems_generation++;
|
1320 |
|
|
break;
|
1321 |
|
|
default:
|
1322 |
|
|
retval = -EINVAL;
|
1323 |
|
|
goto out2;
|
1324 |
|
|
}
|
1325 |
|
|
|
1326 |
|
|
if (retval == 0)
|
1327 |
|
|
retval = nbytes;
|
1328 |
|
|
out2:
|
1329 |
|
|
cgroup_unlock();
|
1330 |
|
|
out1:
|
1331 |
|
|
kfree(buffer);
|
1332 |
|
|
return retval;
|
1333 |
|
|
}
|
1334 |
|
|
|
1335 |
|
|
/*
|
1336 |
|
|
* These ascii lists should be read in a single call, by using a user
|
1337 |
|
|
* buffer large enough to hold the entire map. If read in smaller
|
1338 |
|
|
* chunks, there is no guarantee of atomicity. Since the display format
|
1339 |
|
|
* used, list of ranges of sequential numbers, is variable length,
|
1340 |
|
|
* and since these maps can change value dynamically, one could read
|
1341 |
|
|
* gibberish by doing partial reads while a list was changing.
|
1342 |
|
|
* A single large read to a buffer that crosses a page boundary is
|
1343 |
|
|
* ok, because the result being copied to user land is not recomputed
|
1344 |
|
|
* across a page fault.
|
1345 |
|
|
*/
|
1346 |
|
|
|
1347 |
|
|
static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
|
1348 |
|
|
{
|
1349 |
|
|
cpumask_t mask;
|
1350 |
|
|
|
1351 |
|
|
mutex_lock(&callback_mutex);
|
1352 |
|
|
mask = cs->cpus_allowed;
|
1353 |
|
|
mutex_unlock(&callback_mutex);
|
1354 |
|
|
|
1355 |
|
|
return cpulist_scnprintf(page, PAGE_SIZE, mask);
|
1356 |
|
|
}
|
1357 |
|
|
|
1358 |
|
|
static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
|
1359 |
|
|
{
|
1360 |
|
|
nodemask_t mask;
|
1361 |
|
|
|
1362 |
|
|
mutex_lock(&callback_mutex);
|
1363 |
|
|
mask = cs->mems_allowed;
|
1364 |
|
|
mutex_unlock(&callback_mutex);
|
1365 |
|
|
|
1366 |
|
|
return nodelist_scnprintf(page, PAGE_SIZE, mask);
|
1367 |
|
|
}
|
1368 |
|
|
|
1369 |
|
|
static ssize_t cpuset_common_file_read(struct cgroup *cont,
|
1370 |
|
|
struct cftype *cft,
|
1371 |
|
|
struct file *file,
|
1372 |
|
|
char __user *buf,
|
1373 |
|
|
size_t nbytes, loff_t *ppos)
|
1374 |
|
|
{
|
1375 |
|
|
struct cpuset *cs = cgroup_cs(cont);
|
1376 |
|
|
cpuset_filetype_t type = cft->private;
|
1377 |
|
|
char *page;
|
1378 |
|
|
ssize_t retval = 0;
|
1379 |
|
|
char *s;
|
1380 |
|
|
|
1381 |
|
|
if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
|
1382 |
|
|
return -ENOMEM;
|
1383 |
|
|
|
1384 |
|
|
s = page;
|
1385 |
|
|
|
1386 |
|
|
switch (type) {
|
1387 |
|
|
case FILE_CPULIST:
|
1388 |
|
|
s += cpuset_sprintf_cpulist(s, cs);
|
1389 |
|
|
break;
|
1390 |
|
|
case FILE_MEMLIST:
|
1391 |
|
|
s += cpuset_sprintf_memlist(s, cs);
|
1392 |
|
|
break;
|
1393 |
|
|
case FILE_CPU_EXCLUSIVE:
|
1394 |
|
|
*s++ = is_cpu_exclusive(cs) ? '1' : '0';
|
1395 |
|
|
break;
|
1396 |
|
|
case FILE_MEM_EXCLUSIVE:
|
1397 |
|
|
*s++ = is_mem_exclusive(cs) ? '1' : '0';
|
1398 |
|
|
break;
|
1399 |
|
|
case FILE_SCHED_LOAD_BALANCE:
|
1400 |
|
|
*s++ = is_sched_load_balance(cs) ? '1' : '0';
|
1401 |
|
|
break;
|
1402 |
|
|
case FILE_MEMORY_MIGRATE:
|
1403 |
|
|
*s++ = is_memory_migrate(cs) ? '1' : '0';
|
1404 |
|
|
break;
|
1405 |
|
|
case FILE_MEMORY_PRESSURE_ENABLED:
|
1406 |
|
|
*s++ = cpuset_memory_pressure_enabled ? '1' : '0';
|
1407 |
|
|
break;
|
1408 |
|
|
case FILE_MEMORY_PRESSURE:
|
1409 |
|
|
s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
|
1410 |
|
|
break;
|
1411 |
|
|
case FILE_SPREAD_PAGE:
|
1412 |
|
|
*s++ = is_spread_page(cs) ? '1' : '0';
|
1413 |
|
|
break;
|
1414 |
|
|
case FILE_SPREAD_SLAB:
|
1415 |
|
|
*s++ = is_spread_slab(cs) ? '1' : '0';
|
1416 |
|
|
break;
|
1417 |
|
|
default:
|
1418 |
|
|
retval = -EINVAL;
|
1419 |
|
|
goto out;
|
1420 |
|
|
}
|
1421 |
|
|
*s++ = '\n';
|
1422 |
|
|
|
1423 |
|
|
retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
|
1424 |
|
|
out:
|
1425 |
|
|
free_page((unsigned long)page);
|
1426 |
|
|
return retval;
|
1427 |
|
|
}
|
1428 |
|
|
|
1429 |
|
|
|
1430 |
|
|
|
1431 |
|
|
|
1432 |
|
|
|
1433 |
|
|
/*
|
1434 |
|
|
* for the common functions, 'private' gives the type of file
|
1435 |
|
|
*/
|
1436 |
|
|
|
1437 |
|
|
static struct cftype cft_cpus = {
|
1438 |
|
|
.name = "cpus",
|
1439 |
|
|
.read = cpuset_common_file_read,
|
1440 |
|
|
.write = cpuset_common_file_write,
|
1441 |
|
|
.private = FILE_CPULIST,
|
1442 |
|
|
};
|
1443 |
|
|
|
1444 |
|
|
static struct cftype cft_mems = {
|
1445 |
|
|
.name = "mems",
|
1446 |
|
|
.read = cpuset_common_file_read,
|
1447 |
|
|
.write = cpuset_common_file_write,
|
1448 |
|
|
.private = FILE_MEMLIST,
|
1449 |
|
|
};
|
1450 |
|
|
|
1451 |
|
|
static struct cftype cft_cpu_exclusive = {
|
1452 |
|
|
.name = "cpu_exclusive",
|
1453 |
|
|
.read = cpuset_common_file_read,
|
1454 |
|
|
.write = cpuset_common_file_write,
|
1455 |
|
|
.private = FILE_CPU_EXCLUSIVE,
|
1456 |
|
|
};
|
1457 |
|
|
|
1458 |
|
|
static struct cftype cft_mem_exclusive = {
|
1459 |
|
|
.name = "mem_exclusive",
|
1460 |
|
|
.read = cpuset_common_file_read,
|
1461 |
|
|
.write = cpuset_common_file_write,
|
1462 |
|
|
.private = FILE_MEM_EXCLUSIVE,
|
1463 |
|
|
};
|
1464 |
|
|
|
1465 |
|
|
static struct cftype cft_sched_load_balance = {
|
1466 |
|
|
.name = "sched_load_balance",
|
1467 |
|
|
.read = cpuset_common_file_read,
|
1468 |
|
|
.write = cpuset_common_file_write,
|
1469 |
|
|
.private = FILE_SCHED_LOAD_BALANCE,
|
1470 |
|
|
};
|
1471 |
|
|
|
1472 |
|
|
static struct cftype cft_memory_migrate = {
|
1473 |
|
|
.name = "memory_migrate",
|
1474 |
|
|
.read = cpuset_common_file_read,
|
1475 |
|
|
.write = cpuset_common_file_write,
|
1476 |
|
|
.private = FILE_MEMORY_MIGRATE,
|
1477 |
|
|
};
|
1478 |
|
|
|
1479 |
|
|
static struct cftype cft_memory_pressure_enabled = {
|
1480 |
|
|
.name = "memory_pressure_enabled",
|
1481 |
|
|
.read = cpuset_common_file_read,
|
1482 |
|
|
.write = cpuset_common_file_write,
|
1483 |
|
|
.private = FILE_MEMORY_PRESSURE_ENABLED,
|
1484 |
|
|
};
|
1485 |
|
|
|
1486 |
|
|
static struct cftype cft_memory_pressure = {
|
1487 |
|
|
.name = "memory_pressure",
|
1488 |
|
|
.read = cpuset_common_file_read,
|
1489 |
|
|
.write = cpuset_common_file_write,
|
1490 |
|
|
.private = FILE_MEMORY_PRESSURE,
|
1491 |
|
|
};
|
1492 |
|
|
|
1493 |
|
|
static struct cftype cft_spread_page = {
|
1494 |
|
|
.name = "memory_spread_page",
|
1495 |
|
|
.read = cpuset_common_file_read,
|
1496 |
|
|
.write = cpuset_common_file_write,
|
1497 |
|
|
.private = FILE_SPREAD_PAGE,
|
1498 |
|
|
};
|
1499 |
|
|
|
1500 |
|
|
static struct cftype cft_spread_slab = {
|
1501 |
|
|
.name = "memory_spread_slab",
|
1502 |
|
|
.read = cpuset_common_file_read,
|
1503 |
|
|
.write = cpuset_common_file_write,
|
1504 |
|
|
.private = FILE_SPREAD_SLAB,
|
1505 |
|
|
};
|
1506 |
|
|
|
1507 |
|
|
static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
|
1508 |
|
|
{
|
1509 |
|
|
int err;
|
1510 |
|
|
|
1511 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_cpus)) < 0)
|
1512 |
|
|
return err;
|
1513 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_mems)) < 0)
|
1514 |
|
|
return err;
|
1515 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_cpu_exclusive)) < 0)
|
1516 |
|
|
return err;
|
1517 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_mem_exclusive)) < 0)
|
1518 |
|
|
return err;
|
1519 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_memory_migrate)) < 0)
|
1520 |
|
|
return err;
|
1521 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_sched_load_balance)) < 0)
|
1522 |
|
|
return err;
|
1523 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_memory_pressure)) < 0)
|
1524 |
|
|
return err;
|
1525 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_spread_page)) < 0)
|
1526 |
|
|
return err;
|
1527 |
|
|
if ((err = cgroup_add_file(cont, ss, &cft_spread_slab)) < 0)
|
1528 |
|
|
return err;
|
1529 |
|
|
/* memory_pressure_enabled is in root cpuset only */
|
1530 |
|
|
if (err == 0 && !cont->parent)
|
1531 |
|
|
err = cgroup_add_file(cont, ss,
|
1532 |
|
|
&cft_memory_pressure_enabled);
|
1533 |
|
|
return 0;
|
1534 |
|
|
}
|
1535 |
|
|
|
1536 |
|
|
/*
|
1537 |
|
|
* post_clone() is called at the end of cgroup_clone().
|
1538 |
|
|
* 'cgroup' was just created automatically as a result of
|
1539 |
|
|
* a cgroup_clone(), and the current task is about to
|
1540 |
|
|
* be moved into 'cgroup'.
|
1541 |
|
|
*
|
1542 |
|
|
* Currently we refuse to set up the cgroup - thereby
|
1543 |
|
|
* refusing the task to be entered, and as a result refusing
|
1544 |
|
|
* the sys_unshare() or clone() which initiated it - if any
|
1545 |
|
|
* sibling cpusets have exclusive cpus or mem.
|
1546 |
|
|
*
|
1547 |
|
|
* If this becomes a problem for some users who wish to
|
1548 |
|
|
* allow that scenario, then cpuset_post_clone() could be
|
1549 |
|
|
* changed to grant parent->cpus_allowed-sibling_cpus_exclusive
|
1550 |
|
|
* (and likewise for mems) to the new cgroup.
|
1551 |
|
|
*/
|
1552 |
|
|
static void cpuset_post_clone(struct cgroup_subsys *ss,
|
1553 |
|
|
struct cgroup *cgroup)
|
1554 |
|
|
{
|
1555 |
|
|
struct cgroup *parent, *child;
|
1556 |
|
|
struct cpuset *cs, *parent_cs;
|
1557 |
|
|
|
1558 |
|
|
parent = cgroup->parent;
|
1559 |
|
|
list_for_each_entry(child, &parent->children, sibling) {
|
1560 |
|
|
cs = cgroup_cs(child);
|
1561 |
|
|
if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
|
1562 |
|
|
return;
|
1563 |
|
|
}
|
1564 |
|
|
cs = cgroup_cs(cgroup);
|
1565 |
|
|
parent_cs = cgroup_cs(parent);
|
1566 |
|
|
|
1567 |
|
|
cs->mems_allowed = parent_cs->mems_allowed;
|
1568 |
|
|
cs->cpus_allowed = parent_cs->cpus_allowed;
|
1569 |
|
|
return;
|
1570 |
|
|
}
|
1571 |
|
|
|
1572 |
|
|
/*
|
1573 |
|
|
* cpuset_create - create a cpuset
|
1574 |
|
|
* parent: cpuset that will be parent of the new cpuset.
|
1575 |
|
|
* name: name of the new cpuset. Will be strcpy'ed.
|
1576 |
|
|
* mode: mode to set on new inode
|
1577 |
|
|
*
|
1578 |
|
|
* Must be called with the mutex on the parent inode held
|
1579 |
|
|
*/
|
1580 |
|
|
|
1581 |
|
|
static struct cgroup_subsys_state *cpuset_create(
|
1582 |
|
|
struct cgroup_subsys *ss,
|
1583 |
|
|
struct cgroup *cont)
|
1584 |
|
|
{
|
1585 |
|
|
struct cpuset *cs;
|
1586 |
|
|
struct cpuset *parent;
|
1587 |
|
|
|
1588 |
|
|
if (!cont->parent) {
|
1589 |
|
|
/* This is early initialization for the top cgroup */
|
1590 |
|
|
top_cpuset.mems_generation = cpuset_mems_generation++;
|
1591 |
|
|
return &top_cpuset.css;
|
1592 |
|
|
}
|
1593 |
|
|
parent = cgroup_cs(cont->parent);
|
1594 |
|
|
cs = kmalloc(sizeof(*cs), GFP_KERNEL);
|
1595 |
|
|
if (!cs)
|
1596 |
|
|
return ERR_PTR(-ENOMEM);
|
1597 |
|
|
|
1598 |
|
|
cpuset_update_task_memory_state();
|
1599 |
|
|
cs->flags = 0;
|
1600 |
|
|
if (is_spread_page(parent))
|
1601 |
|
|
set_bit(CS_SPREAD_PAGE, &cs->flags);
|
1602 |
|
|
if (is_spread_slab(parent))
|
1603 |
|
|
set_bit(CS_SPREAD_SLAB, &cs->flags);
|
1604 |
|
|
set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
|
1605 |
|
|
cs->cpus_allowed = CPU_MASK_NONE;
|
1606 |
|
|
cs->mems_allowed = NODE_MASK_NONE;
|
1607 |
|
|
cs->mems_generation = cpuset_mems_generation++;
|
1608 |
|
|
fmeter_init(&cs->fmeter);
|
1609 |
|
|
|
1610 |
|
|
cs->parent = parent;
|
1611 |
|
|
number_of_cpusets++;
|
1612 |
|
|
return &cs->css ;
|
1613 |
|
|
}
|
1614 |
|
|
|
1615 |
|
|
/*
|
1616 |
|
|
* Locking note on the strange update_flag() call below:
|
1617 |
|
|
*
|
1618 |
|
|
* If the cpuset being removed has its flag 'sched_load_balance'
|
1619 |
|
|
* enabled, then simulate turning sched_load_balance off, which
|
1620 |
|
|
* will call rebuild_sched_domains(). The lock_cpu_hotplug()
|
1621 |
|
|
* call in rebuild_sched_domains() must not be made while holding
|
1622 |
|
|
* callback_mutex. Elsewhere the kernel nests callback_mutex inside
|
1623 |
|
|
* lock_cpu_hotplug() calls. So the reverse nesting would risk an
|
1624 |
|
|
* ABBA deadlock.
|
1625 |
|
|
*/
|
1626 |
|
|
|
1627 |
|
|
static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
|
1628 |
|
|
{
|
1629 |
|
|
struct cpuset *cs = cgroup_cs(cont);
|
1630 |
|
|
|
1631 |
|
|
cpuset_update_task_memory_state();
|
1632 |
|
|
|
1633 |
|
|
if (is_sched_load_balance(cs))
|
1634 |
|
|
update_flag(CS_SCHED_LOAD_BALANCE, cs, "0");
|
1635 |
|
|
|
1636 |
|
|
number_of_cpusets--;
|
1637 |
|
|
kfree(cs);
|
1638 |
|
|
}
|
1639 |
|
|
|
1640 |
|
|
struct cgroup_subsys cpuset_subsys = {
|
1641 |
|
|
.name = "cpuset",
|
1642 |
|
|
.create = cpuset_create,
|
1643 |
|
|
.destroy = cpuset_destroy,
|
1644 |
|
|
.can_attach = cpuset_can_attach,
|
1645 |
|
|
.attach = cpuset_attach,
|
1646 |
|
|
.populate = cpuset_populate,
|
1647 |
|
|
.post_clone = cpuset_post_clone,
|
1648 |
|
|
.subsys_id = cpuset_subsys_id,
|
1649 |
|
|
.early_init = 1,
|
1650 |
|
|
};
|
1651 |
|
|
|
1652 |
|
|
/*
|
1653 |
|
|
* cpuset_init_early - just enough so that the calls to
|
1654 |
|
|
* cpuset_update_task_memory_state() in early init code
|
1655 |
|
|
* are harmless.
|
1656 |
|
|
*/
|
1657 |
|
|
|
1658 |
|
|
int __init cpuset_init_early(void)
|
1659 |
|
|
{
|
1660 |
|
|
top_cpuset.mems_generation = cpuset_mems_generation++;
|
1661 |
|
|
return 0;
|
1662 |
|
|
}
|
1663 |
|
|
|
1664 |
|
|
|
1665 |
|
|
/**
|
1666 |
|
|
* cpuset_init - initialize cpusets at system boot
|
1667 |
|
|
*
|
1668 |
|
|
* Description: Initialize top_cpuset and the cpuset internal file system,
|
1669 |
|
|
**/
|
1670 |
|
|
|
1671 |
|
|
int __init cpuset_init(void)
|
1672 |
|
|
{
|
1673 |
|
|
int err = 0;
|
1674 |
|
|
|
1675 |
|
|
top_cpuset.cpus_allowed = CPU_MASK_ALL;
|
1676 |
|
|
top_cpuset.mems_allowed = NODE_MASK_ALL;
|
1677 |
|
|
|
1678 |
|
|
fmeter_init(&top_cpuset.fmeter);
|
1679 |
|
|
top_cpuset.mems_generation = cpuset_mems_generation++;
|
1680 |
|
|
set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
|
1681 |
|
|
|
1682 |
|
|
err = register_filesystem(&cpuset_fs_type);
|
1683 |
|
|
if (err < 0)
|
1684 |
|
|
return err;
|
1685 |
|
|
|
1686 |
|
|
number_of_cpusets = 1;
|
1687 |
|
|
return 0;
|
1688 |
|
|
}
|
1689 |
|
|
|
1690 |
|
|
/*
|
1691 |
|
|
* If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
|
1692 |
|
|
* or memory nodes, we need to walk over the cpuset hierarchy,
|
1693 |
|
|
* removing that CPU or node from all cpusets. If this removes the
|
1694 |
|
|
* last CPU or node from a cpuset, then the guarantee_online_cpus()
|
1695 |
|
|
* or guarantee_online_mems() code will use that emptied cpusets
|
1696 |
|
|
* parent online CPUs or nodes. Cpusets that were already empty of
|
1697 |
|
|
* CPUs or nodes are left empty.
|
1698 |
|
|
*
|
1699 |
|
|
* This routine is intentionally inefficient in a couple of regards.
|
1700 |
|
|
* It will check all cpusets in a subtree even if the top cpuset of
|
1701 |
|
|
* the subtree has no offline CPUs or nodes. It checks both CPUs and
|
1702 |
|
|
* nodes, even though the caller could have been coded to know that
|
1703 |
|
|
* only one of CPUs or nodes needed to be checked on a given call.
|
1704 |
|
|
* This was done to minimize text size rather than cpu cycles.
|
1705 |
|
|
*
|
1706 |
|
|
* Call with both manage_mutex and callback_mutex held.
|
1707 |
|
|
*
|
1708 |
|
|
* Recursive, on depth of cpuset subtree.
|
1709 |
|
|
*/
|
1710 |
|
|
|
1711 |
|
|
static void guarantee_online_cpus_mems_in_subtree(const struct cpuset *cur)
|
1712 |
|
|
{
|
1713 |
|
|
struct cgroup *cont;
|
1714 |
|
|
struct cpuset *c;
|
1715 |
|
|
|
1716 |
|
|
/* Each of our child cpusets mems must be online */
|
1717 |
|
|
list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
|
1718 |
|
|
c = cgroup_cs(cont);
|
1719 |
|
|
guarantee_online_cpus_mems_in_subtree(c);
|
1720 |
|
|
if (!cpus_empty(c->cpus_allowed))
|
1721 |
|
|
guarantee_online_cpus(c, &c->cpus_allowed);
|
1722 |
|
|
if (!nodes_empty(c->mems_allowed))
|
1723 |
|
|
guarantee_online_mems(c, &c->mems_allowed);
|
1724 |
|
|
}
|
1725 |
|
|
}
|
1726 |
|
|
|
1727 |
|
|
/*
|
1728 |
|
|
* The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
|
1729 |
|
|
* cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
|
1730 |
|
|
* track what's online after any CPU or memory node hotplug or unplug
|
1731 |
|
|
* event.
|
1732 |
|
|
*
|
1733 |
|
|
* To ensure that we don't remove a CPU or node from the top cpuset
|
1734 |
|
|
* that is currently in use by a child cpuset (which would violate
|
1735 |
|
|
* the rule that cpusets must be subsets of their parent), we first
|
1736 |
|
|
* call the recursive routine guarantee_online_cpus_mems_in_subtree().
|
1737 |
|
|
*
|
1738 |
|
|
* Since there are two callers of this routine, one for CPU hotplug
|
1739 |
|
|
* events and one for memory node hotplug events, we could have coded
|
1740 |
|
|
* two separate routines here. We code it as a single common routine
|
1741 |
|
|
* in order to minimize text size.
|
1742 |
|
|
*/
|
1743 |
|
|
|
1744 |
|
|
static void common_cpu_mem_hotplug_unplug(void)
|
1745 |
|
|
{
|
1746 |
|
|
cgroup_lock();
|
1747 |
|
|
mutex_lock(&callback_mutex);
|
1748 |
|
|
|
1749 |
|
|
guarantee_online_cpus_mems_in_subtree(&top_cpuset);
|
1750 |
|
|
top_cpuset.cpus_allowed = cpu_online_map;
|
1751 |
|
|
top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
|
1752 |
|
|
|
1753 |
|
|
mutex_unlock(&callback_mutex);
|
1754 |
|
|
cgroup_unlock();
|
1755 |
|
|
}
|
1756 |
|
|
|
1757 |
|
|
/*
|
1758 |
|
|
* The top_cpuset tracks what CPUs and Memory Nodes are online,
|
1759 |
|
|
* period. This is necessary in order to make cpusets transparent
|
1760 |
|
|
* (of no affect) on systems that are actively using CPU hotplug
|
1761 |
|
|
* but making no active use of cpusets.
|
1762 |
|
|
*
|
1763 |
|
|
* This routine ensures that top_cpuset.cpus_allowed tracks
|
1764 |
|
|
* cpu_online_map on each CPU hotplug (cpuhp) event.
|
1765 |
|
|
*/
|
1766 |
|
|
|
1767 |
|
|
static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
|
1768 |
|
|
unsigned long phase, void *unused_cpu)
|
1769 |
|
|
{
|
1770 |
|
|
if (phase == CPU_DYING || phase == CPU_DYING_FROZEN)
|
1771 |
|
|
return NOTIFY_DONE;
|
1772 |
|
|
|
1773 |
|
|
common_cpu_mem_hotplug_unplug();
|
1774 |
|
|
return 0;
|
1775 |
|
|
}
|
1776 |
|
|
|
1777 |
|
|
#ifdef CONFIG_MEMORY_HOTPLUG
|
1778 |
|
|
/*
|
1779 |
|
|
* Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
|
1780 |
|
|
* Call this routine anytime after you change
|
1781 |
|
|
* node_states[N_HIGH_MEMORY].
|
1782 |
|
|
* See also the previous routine cpuset_handle_cpuhp().
|
1783 |
|
|
*/
|
1784 |
|
|
|
1785 |
|
|
void cpuset_track_online_nodes(void)
|
1786 |
|
|
{
|
1787 |
|
|
common_cpu_mem_hotplug_unplug();
|
1788 |
|
|
}
|
1789 |
|
|
#endif
|
1790 |
|
|
|
1791 |
|
|
/**
|
1792 |
|
|
* cpuset_init_smp - initialize cpus_allowed
|
1793 |
|
|
*
|
1794 |
|
|
* Description: Finish top cpuset after cpu, node maps are initialized
|
1795 |
|
|
**/
|
1796 |
|
|
|
1797 |
|
|
void __init cpuset_init_smp(void)
|
1798 |
|
|
{
|
1799 |
|
|
top_cpuset.cpus_allowed = cpu_online_map;
|
1800 |
|
|
top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
|
1801 |
|
|
|
1802 |
|
|
hotcpu_notifier(cpuset_handle_cpuhp, 0);
|
1803 |
|
|
}
|
1804 |
|
|
|
1805 |
|
|
/**
|
1806 |
|
|
|
1807 |
|
|
* cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
|
1808 |
|
|
* @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
|
1809 |
|
|
*
|
1810 |
|
|
* Description: Returns the cpumask_t cpus_allowed of the cpuset
|
1811 |
|
|
* attached to the specified @tsk. Guaranteed to return some non-empty
|
1812 |
|
|
* subset of cpu_online_map, even if this means going outside the
|
1813 |
|
|
* tasks cpuset.
|
1814 |
|
|
**/
|
1815 |
|
|
|
1816 |
|
|
cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
|
1817 |
|
|
{
|
1818 |
|
|
cpumask_t mask;
|
1819 |
|
|
|
1820 |
|
|
mutex_lock(&callback_mutex);
|
1821 |
|
|
mask = cpuset_cpus_allowed_locked(tsk);
|
1822 |
|
|
mutex_unlock(&callback_mutex);
|
1823 |
|
|
|
1824 |
|
|
return mask;
|
1825 |
|
|
}
|
1826 |
|
|
|
1827 |
|
|
/**
|
1828 |
|
|
* cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
|
1829 |
|
|
* Must be called with callback_mutex held.
|
1830 |
|
|
**/
|
1831 |
|
|
cpumask_t cpuset_cpus_allowed_locked(struct task_struct *tsk)
|
1832 |
|
|
{
|
1833 |
|
|
cpumask_t mask;
|
1834 |
|
|
|
1835 |
|
|
task_lock(tsk);
|
1836 |
|
|
guarantee_online_cpus(task_cs(tsk), &mask);
|
1837 |
|
|
task_unlock(tsk);
|
1838 |
|
|
|
1839 |
|
|
return mask;
|
1840 |
|
|
}
|
1841 |
|
|
|
1842 |
|
|
void cpuset_init_current_mems_allowed(void)
|
1843 |
|
|
{
|
1844 |
|
|
current->mems_allowed = NODE_MASK_ALL;
|
1845 |
|
|
}
|
1846 |
|
|
|
1847 |
|
|
/**
|
1848 |
|
|
* cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
|
1849 |
|
|
* @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
|
1850 |
|
|
*
|
1851 |
|
|
* Description: Returns the nodemask_t mems_allowed of the cpuset
|
1852 |
|
|
* attached to the specified @tsk. Guaranteed to return some non-empty
|
1853 |
|
|
* subset of node_states[N_HIGH_MEMORY], even if this means going outside the
|
1854 |
|
|
* tasks cpuset.
|
1855 |
|
|
**/
|
1856 |
|
|
|
1857 |
|
|
nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
|
1858 |
|
|
{
|
1859 |
|
|
nodemask_t mask;
|
1860 |
|
|
|
1861 |
|
|
mutex_lock(&callback_mutex);
|
1862 |
|
|
task_lock(tsk);
|
1863 |
|
|
guarantee_online_mems(task_cs(tsk), &mask);
|
1864 |
|
|
task_unlock(tsk);
|
1865 |
|
|
mutex_unlock(&callback_mutex);
|
1866 |
|
|
|
1867 |
|
|
return mask;
|
1868 |
|
|
}
|
1869 |
|
|
|
1870 |
|
|
/**
|
1871 |
|
|
* cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
|
1872 |
|
|
* @zl: the zonelist to be checked
|
1873 |
|
|
*
|
1874 |
|
|
* Are any of the nodes on zonelist zl allowed in current->mems_allowed?
|
1875 |
|
|
*/
|
1876 |
|
|
int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
|
1877 |
|
|
{
|
1878 |
|
|
int i;
|
1879 |
|
|
|
1880 |
|
|
for (i = 0; zl->zones[i]; i++) {
|
1881 |
|
|
int nid = zone_to_nid(zl->zones[i]);
|
1882 |
|
|
|
1883 |
|
|
if (node_isset(nid, current->mems_allowed))
|
1884 |
|
|
return 1;
|
1885 |
|
|
}
|
1886 |
|
|
return 0;
|
1887 |
|
|
}
|
1888 |
|
|
|
1889 |
|
|
/*
|
1890 |
|
|
* nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
|
1891 |
|
|
* ancestor to the specified cpuset. Call holding callback_mutex.
|
1892 |
|
|
* If no ancestor is mem_exclusive (an unusual configuration), then
|
1893 |
|
|
* returns the root cpuset.
|
1894 |
|
|
*/
|
1895 |
|
|
static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
|
1896 |
|
|
{
|
1897 |
|
|
while (!is_mem_exclusive(cs) && cs->parent)
|
1898 |
|
|
cs = cs->parent;
|
1899 |
|
|
return cs;
|
1900 |
|
|
}
|
1901 |
|
|
|
1902 |
|
|
/**
|
1903 |
|
|
* cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
|
1904 |
|
|
* @z: is this zone on an allowed node?
|
1905 |
|
|
* @gfp_mask: memory allocation flags
|
1906 |
|
|
*
|
1907 |
|
|
* If we're in interrupt, yes, we can always allocate. If
|
1908 |
|
|
* __GFP_THISNODE is set, yes, we can always allocate. If zone
|
1909 |
|
|
* z's node is in our tasks mems_allowed, yes. If it's not a
|
1910 |
|
|
* __GFP_HARDWALL request and this zone's nodes is in the nearest
|
1911 |
|
|
* mem_exclusive cpuset ancestor to this tasks cpuset, yes.
|
1912 |
|
|
* If the task has been OOM killed and has access to memory reserves
|
1913 |
|
|
* as specified by the TIF_MEMDIE flag, yes.
|
1914 |
|
|
* Otherwise, no.
|
1915 |
|
|
*
|
1916 |
|
|
* If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
|
1917 |
|
|
* reduces to cpuset_zone_allowed_hardwall(). Otherwise,
|
1918 |
|
|
* cpuset_zone_allowed_softwall() might sleep, and might allow a zone
|
1919 |
|
|
* from an enclosing cpuset.
|
1920 |
|
|
*
|
1921 |
|
|
* cpuset_zone_allowed_hardwall() only handles the simpler case of
|
1922 |
|
|
* hardwall cpusets, and never sleeps.
|
1923 |
|
|
*
|
1924 |
|
|
* The __GFP_THISNODE placement logic is really handled elsewhere,
|
1925 |
|
|
* by forcibly using a zonelist starting at a specified node, and by
|
1926 |
|
|
* (in get_page_from_freelist()) refusing to consider the zones for
|
1927 |
|
|
* any node on the zonelist except the first. By the time any such
|
1928 |
|
|
* calls get to this routine, we should just shut up and say 'yes'.
|
1929 |
|
|
*
|
1930 |
|
|
* GFP_USER allocations are marked with the __GFP_HARDWALL bit,
|
1931 |
|
|
* and do not allow allocations outside the current tasks cpuset
|
1932 |
|
|
* unless the task has been OOM killed as is marked TIF_MEMDIE.
|
1933 |
|
|
* GFP_KERNEL allocations are not so marked, so can escape to the
|
1934 |
|
|
* nearest enclosing mem_exclusive ancestor cpuset.
|
1935 |
|
|
*
|
1936 |
|
|
* Scanning up parent cpusets requires callback_mutex. The
|
1937 |
|
|
* __alloc_pages() routine only calls here with __GFP_HARDWALL bit
|
1938 |
|
|
* _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
|
1939 |
|
|
* current tasks mems_allowed came up empty on the first pass over
|
1940 |
|
|
* the zonelist. So only GFP_KERNEL allocations, if all nodes in the
|
1941 |
|
|
* cpuset are short of memory, might require taking the callback_mutex
|
1942 |
|
|
* mutex.
|
1943 |
|
|
*
|
1944 |
|
|
* The first call here from mm/page_alloc:get_page_from_freelist()
|
1945 |
|
|
* has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
|
1946 |
|
|
* so no allocation on a node outside the cpuset is allowed (unless
|
1947 |
|
|
* in interrupt, of course).
|
1948 |
|
|
*
|
1949 |
|
|
* The second pass through get_page_from_freelist() doesn't even call
|
1950 |
|
|
* here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
|
1951 |
|
|
* variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
|
1952 |
|
|
* in alloc_flags. That logic and the checks below have the combined
|
1953 |
|
|
* affect that:
|
1954 |
|
|
* in_interrupt - any node ok (current task context irrelevant)
|
1955 |
|
|
* GFP_ATOMIC - any node ok
|
1956 |
|
|
* TIF_MEMDIE - any node ok
|
1957 |
|
|
* GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
|
1958 |
|
|
* GFP_USER - only nodes in current tasks mems allowed ok.
|
1959 |
|
|
*
|
1960 |
|
|
* Rule:
|
1961 |
|
|
* Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
|
1962 |
|
|
* pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
|
1963 |
|
|
* the code that might scan up ancestor cpusets and sleep.
|
1964 |
|
|
*/
|
1965 |
|
|
|
1966 |
|
|
int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
|
1967 |
|
|
{
|
1968 |
|
|
int node; /* node that zone z is on */
|
1969 |
|
|
const struct cpuset *cs; /* current cpuset ancestors */
|
1970 |
|
|
int allowed; /* is allocation in zone z allowed? */
|
1971 |
|
|
|
1972 |
|
|
if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
|
1973 |
|
|
return 1;
|
1974 |
|
|
node = zone_to_nid(z);
|
1975 |
|
|
might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
|
1976 |
|
|
if (node_isset(node, current->mems_allowed))
|
1977 |
|
|
return 1;
|
1978 |
|
|
/*
|
1979 |
|
|
* Allow tasks that have access to memory reserves because they have
|
1980 |
|
|
* been OOM killed to get memory anywhere.
|
1981 |
|
|
*/
|
1982 |
|
|
if (unlikely(test_thread_flag(TIF_MEMDIE)))
|
1983 |
|
|
return 1;
|
1984 |
|
|
if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
|
1985 |
|
|
return 0;
|
1986 |
|
|
|
1987 |
|
|
if (current->flags & PF_EXITING) /* Let dying task have memory */
|
1988 |
|
|
return 1;
|
1989 |
|
|
|
1990 |
|
|
/* Not hardwall and node outside mems_allowed: scan up cpusets */
|
1991 |
|
|
mutex_lock(&callback_mutex);
|
1992 |
|
|
|
1993 |
|
|
task_lock(current);
|
1994 |
|
|
cs = nearest_exclusive_ancestor(task_cs(current));
|
1995 |
|
|
task_unlock(current);
|
1996 |
|
|
|
1997 |
|
|
allowed = node_isset(node, cs->mems_allowed);
|
1998 |
|
|
mutex_unlock(&callback_mutex);
|
1999 |
|
|
return allowed;
|
2000 |
|
|
}
|
2001 |
|
|
|
2002 |
|
|
/*
|
2003 |
|
|
* cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
|
2004 |
|
|
* @z: is this zone on an allowed node?
|
2005 |
|
|
* @gfp_mask: memory allocation flags
|
2006 |
|
|
*
|
2007 |
|
|
* If we're in interrupt, yes, we can always allocate.
|
2008 |
|
|
* If __GFP_THISNODE is set, yes, we can always allocate. If zone
|
2009 |
|
|
* z's node is in our tasks mems_allowed, yes. If the task has been
|
2010 |
|
|
* OOM killed and has access to memory reserves as specified by the
|
2011 |
|
|
* TIF_MEMDIE flag, yes. Otherwise, no.
|
2012 |
|
|
*
|
2013 |
|
|
* The __GFP_THISNODE placement logic is really handled elsewhere,
|
2014 |
|
|
* by forcibly using a zonelist starting at a specified node, and by
|
2015 |
|
|
* (in get_page_from_freelist()) refusing to consider the zones for
|
2016 |
|
|
* any node on the zonelist except the first. By the time any such
|
2017 |
|
|
* calls get to this routine, we should just shut up and say 'yes'.
|
2018 |
|
|
*
|
2019 |
|
|
* Unlike the cpuset_zone_allowed_softwall() variant, above,
|
2020 |
|
|
* this variant requires that the zone be in the current tasks
|
2021 |
|
|
* mems_allowed or that we're in interrupt. It does not scan up the
|
2022 |
|
|
* cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
|
2023 |
|
|
* It never sleeps.
|
2024 |
|
|
*/
|
2025 |
|
|
|
2026 |
|
|
int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
|
2027 |
|
|
{
|
2028 |
|
|
int node; /* node that zone z is on */
|
2029 |
|
|
|
2030 |
|
|
if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
|
2031 |
|
|
return 1;
|
2032 |
|
|
node = zone_to_nid(z);
|
2033 |
|
|
if (node_isset(node, current->mems_allowed))
|
2034 |
|
|
return 1;
|
2035 |
|
|
/*
|
2036 |
|
|
* Allow tasks that have access to memory reserves because they have
|
2037 |
|
|
* been OOM killed to get memory anywhere.
|
2038 |
|
|
*/
|
2039 |
|
|
if (unlikely(test_thread_flag(TIF_MEMDIE)))
|
2040 |
|
|
return 1;
|
2041 |
|
|
return 0;
|
2042 |
|
|
}
|
2043 |
|
|
|
2044 |
|
|
/**
|
2045 |
|
|
* cpuset_lock - lock out any changes to cpuset structures
|
2046 |
|
|
*
|
2047 |
|
|
* The out of memory (oom) code needs to mutex_lock cpusets
|
2048 |
|
|
* from being changed while it scans the tasklist looking for a
|
2049 |
|
|
* task in an overlapping cpuset. Expose callback_mutex via this
|
2050 |
|
|
* cpuset_lock() routine, so the oom code can lock it, before
|
2051 |
|
|
* locking the task list. The tasklist_lock is a spinlock, so
|
2052 |
|
|
* must be taken inside callback_mutex.
|
2053 |
|
|
*/
|
2054 |
|
|
|
2055 |
|
|
void cpuset_lock(void)
|
2056 |
|
|
{
|
2057 |
|
|
mutex_lock(&callback_mutex);
|
2058 |
|
|
}
|
2059 |
|
|
|
2060 |
|
|
/**
|
2061 |
|
|
* cpuset_unlock - release lock on cpuset changes
|
2062 |
|
|
*
|
2063 |
|
|
* Undo the lock taken in a previous cpuset_lock() call.
|
2064 |
|
|
*/
|
2065 |
|
|
|
2066 |
|
|
void cpuset_unlock(void)
|
2067 |
|
|
{
|
2068 |
|
|
mutex_unlock(&callback_mutex);
|
2069 |
|
|
}
|
2070 |
|
|
|
2071 |
|
|
/**
|
2072 |
|
|
* cpuset_mem_spread_node() - On which node to begin search for a page
|
2073 |
|
|
*
|
2074 |
|
|
* If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
|
2075 |
|
|
* tasks in a cpuset with is_spread_page or is_spread_slab set),
|
2076 |
|
|
* and if the memory allocation used cpuset_mem_spread_node()
|
2077 |
|
|
* to determine on which node to start looking, as it will for
|
2078 |
|
|
* certain page cache or slab cache pages such as used for file
|
2079 |
|
|
* system buffers and inode caches, then instead of starting on the
|
2080 |
|
|
* local node to look for a free page, rather spread the starting
|
2081 |
|
|
* node around the tasks mems_allowed nodes.
|
2082 |
|
|
*
|
2083 |
|
|
* We don't have to worry about the returned node being offline
|
2084 |
|
|
* because "it can't happen", and even if it did, it would be ok.
|
2085 |
|
|
*
|
2086 |
|
|
* The routines calling guarantee_online_mems() are careful to
|
2087 |
|
|
* only set nodes in task->mems_allowed that are online. So it
|
2088 |
|
|
* should not be possible for the following code to return an
|
2089 |
|
|
* offline node. But if it did, that would be ok, as this routine
|
2090 |
|
|
* is not returning the node where the allocation must be, only
|
2091 |
|
|
* the node where the search should start. The zonelist passed to
|
2092 |
|
|
* __alloc_pages() will include all nodes. If the slab allocator
|
2093 |
|
|
* is passed an offline node, it will fall back to the local node.
|
2094 |
|
|
* See kmem_cache_alloc_node().
|
2095 |
|
|
*/
|
2096 |
|
|
|
2097 |
|
|
int cpuset_mem_spread_node(void)
|
2098 |
|
|
{
|
2099 |
|
|
int node;
|
2100 |
|
|
|
2101 |
|
|
node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
|
2102 |
|
|
if (node == MAX_NUMNODES)
|
2103 |
|
|
node = first_node(current->mems_allowed);
|
2104 |
|
|
current->cpuset_mem_spread_rotor = node;
|
2105 |
|
|
return node;
|
2106 |
|
|
}
|
2107 |
|
|
EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
|
2108 |
|
|
|
2109 |
|
|
/**
|
2110 |
|
|
* cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
|
2111 |
|
|
* @tsk1: pointer to task_struct of some task.
|
2112 |
|
|
* @tsk2: pointer to task_struct of some other task.
|
2113 |
|
|
*
|
2114 |
|
|
* Description: Return true if @tsk1's mems_allowed intersects the
|
2115 |
|
|
* mems_allowed of @tsk2. Used by the OOM killer to determine if
|
2116 |
|
|
* one of the task's memory usage might impact the memory available
|
2117 |
|
|
* to the other.
|
2118 |
|
|
**/
|
2119 |
|
|
|
2120 |
|
|
int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
|
2121 |
|
|
const struct task_struct *tsk2)
|
2122 |
|
|
{
|
2123 |
|
|
return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
|
2124 |
|
|
}
|
2125 |
|
|
|
2126 |
|
|
/*
|
2127 |
|
|
* Collection of memory_pressure is suppressed unless
|
2128 |
|
|
* this flag is enabled by writing "1" to the special
|
2129 |
|
|
* cpuset file 'memory_pressure_enabled' in the root cpuset.
|
2130 |
|
|
*/
|
2131 |
|
|
|
2132 |
|
|
int cpuset_memory_pressure_enabled __read_mostly;
|
2133 |
|
|
|
2134 |
|
|
/**
|
2135 |
|
|
* cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
|
2136 |
|
|
*
|
2137 |
|
|
* Keep a running average of the rate of synchronous (direct)
|
2138 |
|
|
* page reclaim efforts initiated by tasks in each cpuset.
|
2139 |
|
|
*
|
2140 |
|
|
* This represents the rate at which some task in the cpuset
|
2141 |
|
|
* ran low on memory on all nodes it was allowed to use, and
|
2142 |
|
|
* had to enter the kernels page reclaim code in an effort to
|
2143 |
|
|
* create more free memory by tossing clean pages or swapping
|
2144 |
|
|
* or writing dirty pages.
|
2145 |
|
|
*
|
2146 |
|
|
* Display to user space in the per-cpuset read-only file
|
2147 |
|
|
* "memory_pressure". Value displayed is an integer
|
2148 |
|
|
* representing the recent rate of entry into the synchronous
|
2149 |
|
|
* (direct) page reclaim by any task attached to the cpuset.
|
2150 |
|
|
**/
|
2151 |
|
|
|
2152 |
|
|
void __cpuset_memory_pressure_bump(void)
|
2153 |
|
|
{
|
2154 |
|
|
task_lock(current);
|
2155 |
|
|
fmeter_markevent(&task_cs(current)->fmeter);
|
2156 |
|
|
task_unlock(current);
|
2157 |
|
|
}
|
2158 |
|
|
|
2159 |
|
|
#ifdef CONFIG_PROC_PID_CPUSET
|
2160 |
|
|
/*
|
2161 |
|
|
* proc_cpuset_show()
|
2162 |
|
|
* - Print tasks cpuset path into seq_file.
|
2163 |
|
|
* - Used for /proc/<pid>/cpuset.
|
2164 |
|
|
* - No need to task_lock(tsk) on this tsk->cpuset reference, as it
|
2165 |
|
|
* doesn't really matter if tsk->cpuset changes after we read it,
|
2166 |
|
|
* and we take manage_mutex, keeping attach_task() from changing it
|
2167 |
|
|
* anyway. No need to check that tsk->cpuset != NULL, thanks to
|
2168 |
|
|
* the_top_cpuset_hack in cpuset_exit(), which sets an exiting tasks
|
2169 |
|
|
* cpuset to top_cpuset.
|
2170 |
|
|
*/
|
2171 |
|
|
static int proc_cpuset_show(struct seq_file *m, void *unused_v)
|
2172 |
|
|
{
|
2173 |
|
|
struct pid *pid;
|
2174 |
|
|
struct task_struct *tsk;
|
2175 |
|
|
char *buf;
|
2176 |
|
|
struct cgroup_subsys_state *css;
|
2177 |
|
|
int retval;
|
2178 |
|
|
|
2179 |
|
|
retval = -ENOMEM;
|
2180 |
|
|
buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
|
2181 |
|
|
if (!buf)
|
2182 |
|
|
goto out;
|
2183 |
|
|
|
2184 |
|
|
retval = -ESRCH;
|
2185 |
|
|
pid = m->private;
|
2186 |
|
|
tsk = get_pid_task(pid, PIDTYPE_PID);
|
2187 |
|
|
if (!tsk)
|
2188 |
|
|
goto out_free;
|
2189 |
|
|
|
2190 |
|
|
retval = -EINVAL;
|
2191 |
|
|
cgroup_lock();
|
2192 |
|
|
css = task_subsys_state(tsk, cpuset_subsys_id);
|
2193 |
|
|
retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
|
2194 |
|
|
if (retval < 0)
|
2195 |
|
|
goto out_unlock;
|
2196 |
|
|
seq_puts(m, buf);
|
2197 |
|
|
seq_putc(m, '\n');
|
2198 |
|
|
out_unlock:
|
2199 |
|
|
cgroup_unlock();
|
2200 |
|
|
put_task_struct(tsk);
|
2201 |
|
|
out_free:
|
2202 |
|
|
kfree(buf);
|
2203 |
|
|
out:
|
2204 |
|
|
return retval;
|
2205 |
|
|
}
|
2206 |
|
|
|
2207 |
|
|
static int cpuset_open(struct inode *inode, struct file *file)
|
2208 |
|
|
{
|
2209 |
|
|
struct pid *pid = PROC_I(inode)->pid;
|
2210 |
|
|
return single_open(file, proc_cpuset_show, pid);
|
2211 |
|
|
}
|
2212 |
|
|
|
2213 |
|
|
const struct file_operations proc_cpuset_operations = {
|
2214 |
|
|
.open = cpuset_open,
|
2215 |
|
|
.read = seq_read,
|
2216 |
|
|
.llseek = seq_lseek,
|
2217 |
|
|
.release = single_release,
|
2218 |
|
|
};
|
2219 |
|
|
#endif /* CONFIG_PROC_PID_CPUSET */
|
2220 |
|
|
|
2221 |
|
|
/* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
|
2222 |
|
|
char *cpuset_task_status_allowed(struct task_struct *task, char *buffer)
|
2223 |
|
|
{
|
2224 |
|
|
buffer += sprintf(buffer, "Cpus_allowed:\t");
|
2225 |
|
|
buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed);
|
2226 |
|
|
buffer += sprintf(buffer, "\n");
|
2227 |
|
|
buffer += sprintf(buffer, "Mems_allowed:\t");
|
2228 |
|
|
buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed);
|
2229 |
|
|
buffer += sprintf(buffer, "\n");
|
2230 |
|
|
return buffer;
|
2231 |
|
|
}
|