root / thread-pool.c @ b811203c
History | View | Annotate | Download (8.7 kB)
1 |
/*
|
---|---|
2 |
* QEMU block layer thread pool
|
3 |
*
|
4 |
* Copyright IBM, Corp. 2008
|
5 |
* Copyright Red Hat, Inc. 2012
|
6 |
*
|
7 |
* Authors:
|
8 |
* Anthony Liguori <aliguori@us.ibm.com>
|
9 |
* Paolo Bonzini <pbonzini@redhat.com>
|
10 |
*
|
11 |
* This work is licensed under the terms of the GNU GPL, version 2. See
|
12 |
* the COPYING file in the top-level directory.
|
13 |
*
|
14 |
* Contributions after 2012-01-13 are licensed under the terms of the
|
15 |
* GNU GPL, version 2 or (at your option) any later version.
|
16 |
*/
|
17 |
#include "qemu-common.h" |
18 |
#include "qemu/queue.h" |
19 |
#include "qemu/thread.h" |
20 |
#include "qemu/osdep.h" |
21 |
#include "block/coroutine.h" |
22 |
#include "trace.h" |
23 |
#include "block/block_int.h" |
24 |
#include "qemu/event_notifier.h" |
25 |
#include "block/thread-pool.h" |
26 |
|
27 |
typedef struct ThreadPool ThreadPool; |
28 |
|
29 |
static void do_spawn_thread(ThreadPool *pool); |
30 |
|
31 |
typedef struct ThreadPoolElement ThreadPoolElement; |
32 |
|
33 |
enum ThreadState {
|
34 |
THREAD_QUEUED, |
35 |
THREAD_ACTIVE, |
36 |
THREAD_DONE, |
37 |
THREAD_CANCELED, |
38 |
}; |
39 |
|
40 |
struct ThreadPoolElement {
|
41 |
BlockDriverAIOCB common; |
42 |
ThreadPool *pool; |
43 |
ThreadPoolFunc *func; |
44 |
void *arg;
|
45 |
|
46 |
/* Moving state out of THREAD_QUEUED is protected by lock. After
|
47 |
* that, only the worker thread can write to it. Reads and writes
|
48 |
* of state and ret are ordered with memory barriers.
|
49 |
*/
|
50 |
enum ThreadState state;
|
51 |
int ret;
|
52 |
|
53 |
/* Access to this list is protected by lock. */
|
54 |
QTAILQ_ENTRY(ThreadPoolElement) reqs; |
55 |
|
56 |
/* Access to this list is protected by the global mutex. */
|
57 |
QLIST_ENTRY(ThreadPoolElement) all; |
58 |
}; |
59 |
|
60 |
struct ThreadPool {
|
61 |
EventNotifier notifier; |
62 |
QemuMutex lock; |
63 |
QemuCond check_cancel; |
64 |
QemuSemaphore sem; |
65 |
int max_threads;
|
66 |
QEMUBH *new_thread_bh; |
67 |
|
68 |
/* The following variables are only accessed from one AioContext. */
|
69 |
QLIST_HEAD(, ThreadPoolElement) head; |
70 |
|
71 |
/* The following variables are protected by lock. */
|
72 |
QTAILQ_HEAD(, ThreadPoolElement) request_list; |
73 |
int cur_threads;
|
74 |
int idle_threads;
|
75 |
int new_threads; /* backlog of threads we need to create */ |
76 |
int pending_threads; /* threads created but not running yet */ |
77 |
int pending_cancellations; /* whether we need a cond_broadcast */ |
78 |
}; |
79 |
|
80 |
/* Currently there is only one thread pool instance. */
|
81 |
static ThreadPool global_pool;
|
82 |
|
83 |
static void *worker_thread(void *opaque) |
84 |
{ |
85 |
ThreadPool *pool = opaque; |
86 |
|
87 |
qemu_mutex_lock(&pool->lock); |
88 |
pool->pending_threads--; |
89 |
do_spawn_thread(pool); |
90 |
|
91 |
while (1) { |
92 |
ThreadPoolElement *req; |
93 |
int ret;
|
94 |
|
95 |
do {
|
96 |
pool->idle_threads++; |
97 |
qemu_mutex_unlock(&pool->lock); |
98 |
ret = qemu_sem_timedwait(&pool->sem, 10000);
|
99 |
qemu_mutex_lock(&pool->lock); |
100 |
pool->idle_threads--; |
101 |
} while (ret == -1 && !QTAILQ_EMPTY(&pool->request_list)); |
102 |
if (ret == -1) { |
103 |
break;
|
104 |
} |
105 |
|
106 |
req = QTAILQ_FIRST(&pool->request_list); |
107 |
QTAILQ_REMOVE(&pool->request_list, req, reqs); |
108 |
req->state = THREAD_ACTIVE; |
109 |
qemu_mutex_unlock(&pool->lock); |
110 |
|
111 |
ret = req->func(req->arg); |
112 |
|
113 |
req->ret = ret; |
114 |
/* Write ret before state. */
|
115 |
smp_wmb(); |
116 |
req->state = THREAD_DONE; |
117 |
|
118 |
qemu_mutex_lock(&pool->lock); |
119 |
if (pool->pending_cancellations) {
|
120 |
qemu_cond_broadcast(&pool->check_cancel); |
121 |
} |
122 |
|
123 |
event_notifier_set(&pool->notifier); |
124 |
} |
125 |
|
126 |
pool->cur_threads--; |
127 |
qemu_mutex_unlock(&pool->lock); |
128 |
return NULL; |
129 |
} |
130 |
|
131 |
static void do_spawn_thread(ThreadPool *pool) |
132 |
{ |
133 |
QemuThread t; |
134 |
|
135 |
/* Runs with lock taken. */
|
136 |
if (!pool->new_threads) {
|
137 |
return;
|
138 |
} |
139 |
|
140 |
pool->new_threads--; |
141 |
pool->pending_threads++; |
142 |
|
143 |
qemu_thread_create(&t, worker_thread, pool, QEMU_THREAD_DETACHED); |
144 |
} |
145 |
|
146 |
static void spawn_thread_bh_fn(void *opaque) |
147 |
{ |
148 |
ThreadPool *pool = opaque; |
149 |
|
150 |
qemu_mutex_lock(&pool->lock); |
151 |
do_spawn_thread(pool); |
152 |
qemu_mutex_unlock(&pool->lock); |
153 |
} |
154 |
|
155 |
static void spawn_thread(ThreadPool *pool) |
156 |
{ |
157 |
pool->cur_threads++; |
158 |
pool->new_threads++; |
159 |
/* If there are threads being created, they will spawn new workers, so
|
160 |
* we don't spend time creating many threads in a loop holding a mutex or
|
161 |
* starving the current vcpu.
|
162 |
*
|
163 |
* If there are no idle threads, ask the main thread to create one, so we
|
164 |
* inherit the correct affinity instead of the vcpu affinity.
|
165 |
*/
|
166 |
if (!pool->pending_threads) {
|
167 |
qemu_bh_schedule(pool->new_thread_bh); |
168 |
} |
169 |
} |
170 |
|
171 |
static void event_notifier_ready(EventNotifier *notifier) |
172 |
{ |
173 |
ThreadPool *pool = container_of(notifier, ThreadPool, notifier); |
174 |
ThreadPoolElement *elem, *next; |
175 |
|
176 |
event_notifier_test_and_clear(notifier); |
177 |
restart:
|
178 |
QLIST_FOREACH_SAFE(elem, &pool->head, all, next) { |
179 |
if (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) {
|
180 |
continue;
|
181 |
} |
182 |
if (elem->state == THREAD_DONE) {
|
183 |
trace_thread_pool_complete(pool, elem, elem->common.opaque, |
184 |
elem->ret); |
185 |
} |
186 |
if (elem->state == THREAD_DONE && elem->common.cb) {
|
187 |
QLIST_REMOVE(elem, all); |
188 |
/* Read state before ret. */
|
189 |
smp_rmb(); |
190 |
elem->common.cb(elem->common.opaque, elem->ret); |
191 |
qemu_aio_release(elem); |
192 |
goto restart;
|
193 |
} else {
|
194 |
/* remove the request */
|
195 |
QLIST_REMOVE(elem, all); |
196 |
qemu_aio_release(elem); |
197 |
} |
198 |
} |
199 |
} |
200 |
|
201 |
static int thread_pool_active(EventNotifier *notifier) |
202 |
{ |
203 |
ThreadPool *pool = container_of(notifier, ThreadPool, notifier); |
204 |
return !QLIST_EMPTY(&pool->head);
|
205 |
} |
206 |
|
207 |
static void thread_pool_cancel(BlockDriverAIOCB *acb) |
208 |
{ |
209 |
ThreadPoolElement *elem = (ThreadPoolElement *)acb; |
210 |
ThreadPool *pool = elem->pool; |
211 |
|
212 |
trace_thread_pool_cancel(elem, elem->common.opaque); |
213 |
|
214 |
qemu_mutex_lock(&pool->lock); |
215 |
if (elem->state == THREAD_QUEUED &&
|
216 |
/* No thread has yet started working on elem. we can try to "steal"
|
217 |
* the item from the worker if we can get a signal from the
|
218 |
* semaphore. Because this is non-blocking, we can do it with
|
219 |
* the lock taken and ensure that elem will remain THREAD_QUEUED.
|
220 |
*/
|
221 |
qemu_sem_timedwait(&pool->sem, 0) == 0) { |
222 |
QTAILQ_REMOVE(&pool->request_list, elem, reqs); |
223 |
elem->state = THREAD_CANCELED; |
224 |
event_notifier_set(&pool->notifier); |
225 |
} else {
|
226 |
pool->pending_cancellations++; |
227 |
while (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) {
|
228 |
qemu_cond_wait(&pool->check_cancel, &pool->lock); |
229 |
} |
230 |
pool->pending_cancellations--; |
231 |
} |
232 |
qemu_mutex_unlock(&pool->lock); |
233 |
} |
234 |
|
235 |
static const AIOCBInfo thread_pool_aiocb_info = { |
236 |
.aiocb_size = sizeof(ThreadPoolElement),
|
237 |
.cancel = thread_pool_cancel, |
238 |
}; |
239 |
|
240 |
BlockDriverAIOCB *thread_pool_submit_aio(ThreadPoolFunc *func, void *arg,
|
241 |
BlockDriverCompletionFunc *cb, void *opaque)
|
242 |
{ |
243 |
ThreadPool *pool = &global_pool; |
244 |
ThreadPoolElement *req; |
245 |
|
246 |
req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque);
|
247 |
req->func = func; |
248 |
req->arg = arg; |
249 |
req->state = THREAD_QUEUED; |
250 |
req->pool = pool; |
251 |
|
252 |
QLIST_INSERT_HEAD(&pool->head, req, all); |
253 |
|
254 |
trace_thread_pool_submit(pool, req, arg); |
255 |
|
256 |
qemu_mutex_lock(&pool->lock); |
257 |
if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) { |
258 |
spawn_thread(pool); |
259 |
} |
260 |
QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs); |
261 |
qemu_mutex_unlock(&pool->lock); |
262 |
qemu_sem_post(&pool->sem); |
263 |
return &req->common;
|
264 |
} |
265 |
|
266 |
typedef struct ThreadPoolCo { |
267 |
Coroutine *co; |
268 |
int ret;
|
269 |
} ThreadPoolCo; |
270 |
|
271 |
static void thread_pool_co_cb(void *opaque, int ret) |
272 |
{ |
273 |
ThreadPoolCo *co = opaque; |
274 |
|
275 |
co->ret = ret; |
276 |
qemu_coroutine_enter(co->co, NULL);
|
277 |
} |
278 |
|
279 |
int coroutine_fn thread_pool_submit_co(ThreadPoolFunc *func, void *arg) |
280 |
{ |
281 |
ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS }; |
282 |
assert(qemu_in_coroutine()); |
283 |
thread_pool_submit_aio(func, arg, thread_pool_co_cb, &tpc); |
284 |
qemu_coroutine_yield(); |
285 |
return tpc.ret;
|
286 |
} |
287 |
|
288 |
void thread_pool_submit(ThreadPoolFunc *func, void *arg) |
289 |
{ |
290 |
thread_pool_submit_aio(func, arg, NULL, NULL); |
291 |
} |
292 |
|
293 |
static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx) |
294 |
{ |
295 |
if (!ctx) {
|
296 |
ctx = qemu_get_aio_context(); |
297 |
} |
298 |
|
299 |
memset(pool, 0, sizeof(*pool)); |
300 |
event_notifier_init(&pool->notifier, false);
|
301 |
qemu_mutex_init(&pool->lock); |
302 |
qemu_cond_init(&pool->check_cancel); |
303 |
qemu_sem_init(&pool->sem, 0);
|
304 |
pool->max_threads = 64;
|
305 |
pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool); |
306 |
|
307 |
QLIST_INIT(&pool->head); |
308 |
QTAILQ_INIT(&pool->request_list); |
309 |
|
310 |
aio_set_event_notifier(ctx, &pool->notifier, event_notifier_ready, |
311 |
thread_pool_active); |
312 |
} |
313 |
|
314 |
static void thread_pool_init(void) |
315 |
{ |
316 |
thread_pool_init_one(&global_pool, NULL);
|
317 |
} |
318 |
|
319 |
block_init(thread_pool_init) |