]>
Commit | Line | Data |
---|---|---|
d354c7ec PB |
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 <[email protected]> | |
9 | * Paolo Bonzini <[email protected]> | |
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" | |
1de7afc9 PB |
18 | #include "qemu/queue.h" |
19 | #include "qemu/thread.h" | |
20 | #include "qemu/osdep.h" | |
737e150e | 21 | #include "block/coroutine.h" |
d354c7ec | 22 | #include "trace.h" |
737e150e | 23 | #include "block/thread-pool.h" |
6a1751b7 | 24 | #include "qemu/main-loop.h" |
d354c7ec | 25 | |
b811203c | 26 | static void do_spawn_thread(ThreadPool *pool); |
d354c7ec PB |
27 | |
28 | typedef struct ThreadPoolElement ThreadPoolElement; | |
29 | ||
30 | enum ThreadState { | |
31 | THREAD_QUEUED, | |
32 | THREAD_ACTIVE, | |
33 | THREAD_DONE, | |
34 | THREAD_CANCELED, | |
35 | }; | |
36 | ||
37 | struct ThreadPoolElement { | |
38 | BlockDriverAIOCB common; | |
b811203c | 39 | ThreadPool *pool; |
d354c7ec PB |
40 | ThreadPoolFunc *func; |
41 | void *arg; | |
19d092cf PB |
42 | |
43 | /* Moving state out of THREAD_QUEUED is protected by lock. After | |
44 | * that, only the worker thread can write to it. Reads and writes | |
45 | * of state and ret are ordered with memory barriers. | |
46 | */ | |
d354c7ec PB |
47 | enum ThreadState state; |
48 | int ret; | |
49 | ||
50 | /* Access to this list is protected by lock. */ | |
51 | QTAILQ_ENTRY(ThreadPoolElement) reqs; | |
52 | ||
53 | /* Access to this list is protected by the global mutex. */ | |
54 | QLIST_ENTRY(ThreadPoolElement) all; | |
55 | }; | |
56 | ||
b811203c | 57 | struct ThreadPool { |
f7311ccc | 58 | AioContext *ctx; |
c2e50e3d | 59 | QEMUBH *completion_bh; |
b811203c SH |
60 | QemuMutex lock; |
61 | QemuCond check_cancel; | |
f7311ccc | 62 | QemuCond worker_stopped; |
b811203c SH |
63 | QemuSemaphore sem; |
64 | int max_threads; | |
65 | QEMUBH *new_thread_bh; | |
66 | ||
67 | /* The following variables are only accessed from one AioContext. */ | |
68 | QLIST_HEAD(, ThreadPoolElement) head; | |
69 | ||
70 | /* The following variables are protected by lock. */ | |
71 | QTAILQ_HEAD(, ThreadPoolElement) request_list; | |
72 | int cur_threads; | |
73 | int idle_threads; | |
74 | int new_threads; /* backlog of threads we need to create */ | |
75 | int pending_threads; /* threads created but not running yet */ | |
76 | int pending_cancellations; /* whether we need a cond_broadcast */ | |
f7311ccc | 77 | bool stopping; |
b811203c SH |
78 | }; |
79 | ||
b811203c | 80 | static void *worker_thread(void *opaque) |
d354c7ec | 81 | { |
b811203c SH |
82 | ThreadPool *pool = opaque; |
83 | ||
84 | qemu_mutex_lock(&pool->lock); | |
85 | pool->pending_threads--; | |
86 | do_spawn_thread(pool); | |
d354c7ec | 87 | |
f7311ccc | 88 | while (!pool->stopping) { |
d354c7ec PB |
89 | ThreadPoolElement *req; |
90 | int ret; | |
91 | ||
92 | do { | |
b811203c SH |
93 | pool->idle_threads++; |
94 | qemu_mutex_unlock(&pool->lock); | |
95 | ret = qemu_sem_timedwait(&pool->sem, 10000); | |
96 | qemu_mutex_lock(&pool->lock); | |
97 | pool->idle_threads--; | |
98 | } while (ret == -1 && !QTAILQ_EMPTY(&pool->request_list)); | |
f7311ccc | 99 | if (ret == -1 || pool->stopping) { |
d354c7ec PB |
100 | break; |
101 | } | |
102 | ||
b811203c SH |
103 | req = QTAILQ_FIRST(&pool->request_list); |
104 | QTAILQ_REMOVE(&pool->request_list, req, reqs); | |
d354c7ec | 105 | req->state = THREAD_ACTIVE; |
b811203c | 106 | qemu_mutex_unlock(&pool->lock); |
d354c7ec PB |
107 | |
108 | ret = req->func(req->arg); | |
109 | ||
d354c7ec | 110 | req->ret = ret; |
19d092cf PB |
111 | /* Write ret before state. */ |
112 | smp_wmb(); | |
113 | req->state = THREAD_DONE; | |
114 | ||
b811203c SH |
115 | qemu_mutex_lock(&pool->lock); |
116 | if (pool->pending_cancellations) { | |
117 | qemu_cond_broadcast(&pool->check_cancel); | |
d354c7ec PB |
118 | } |
119 | ||
c2e50e3d | 120 | qemu_bh_schedule(pool->completion_bh); |
d354c7ec PB |
121 | } |
122 | ||
b811203c | 123 | pool->cur_threads--; |
f7311ccc | 124 | qemu_cond_signal(&pool->worker_stopped); |
b811203c | 125 | qemu_mutex_unlock(&pool->lock); |
d354c7ec PB |
126 | return NULL; |
127 | } | |
128 | ||
b811203c | 129 | static void do_spawn_thread(ThreadPool *pool) |
d354c7ec PB |
130 | { |
131 | QemuThread t; | |
132 | ||
133 | /* Runs with lock taken. */ | |
b811203c | 134 | if (!pool->new_threads) { |
d354c7ec PB |
135 | return; |
136 | } | |
137 | ||
b811203c SH |
138 | pool->new_threads--; |
139 | pool->pending_threads++; | |
d354c7ec | 140 | |
4900116e | 141 | qemu_thread_create(&t, "worker", worker_thread, pool, QEMU_THREAD_DETACHED); |
d354c7ec PB |
142 | } |
143 | ||
144 | static void spawn_thread_bh_fn(void *opaque) | |
145 | { | |
b811203c SH |
146 | ThreadPool *pool = opaque; |
147 | ||
148 | qemu_mutex_lock(&pool->lock); | |
149 | do_spawn_thread(pool); | |
150 | qemu_mutex_unlock(&pool->lock); | |
d354c7ec PB |
151 | } |
152 | ||
b811203c | 153 | static void spawn_thread(ThreadPool *pool) |
d354c7ec | 154 | { |
b811203c SH |
155 | pool->cur_threads++; |
156 | pool->new_threads++; | |
d354c7ec PB |
157 | /* If there are threads being created, they will spawn new workers, so |
158 | * we don't spend time creating many threads in a loop holding a mutex or | |
159 | * starving the current vcpu. | |
160 | * | |
161 | * If there are no idle threads, ask the main thread to create one, so we | |
162 | * inherit the correct affinity instead of the vcpu affinity. | |
163 | */ | |
b811203c SH |
164 | if (!pool->pending_threads) { |
165 | qemu_bh_schedule(pool->new_thread_bh); | |
d354c7ec PB |
166 | } |
167 | } | |
168 | ||
c2e50e3d | 169 | static void thread_pool_completion_bh(void *opaque) |
d354c7ec | 170 | { |
c2e50e3d | 171 | ThreadPool *pool = opaque; |
d354c7ec PB |
172 | ThreadPoolElement *elem, *next; |
173 | ||
d354c7ec | 174 | restart: |
b811203c | 175 | QLIST_FOREACH_SAFE(elem, &pool->head, all, next) { |
d354c7ec PB |
176 | if (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) { |
177 | continue; | |
178 | } | |
179 | if (elem->state == THREAD_DONE) { | |
b811203c SH |
180 | trace_thread_pool_complete(pool, elem, elem->common.opaque, |
181 | elem->ret); | |
d354c7ec PB |
182 | } |
183 | if (elem->state == THREAD_DONE && elem->common.cb) { | |
d354c7ec | 184 | QLIST_REMOVE(elem, all); |
19d092cf PB |
185 | /* Read state before ret. */ |
186 | smp_rmb(); | |
3c80ca15 SH |
187 | |
188 | /* Schedule ourselves in case elem->common.cb() calls aio_poll() to | |
189 | * wait for another request that completed at the same time. | |
190 | */ | |
191 | qemu_bh_schedule(pool->completion_bh); | |
192 | ||
19d092cf | 193 | elem->common.cb(elem->common.opaque, elem->ret); |
d354c7ec PB |
194 | qemu_aio_release(elem); |
195 | goto restart; | |
196 | } else { | |
197 | /* remove the request */ | |
198 | QLIST_REMOVE(elem, all); | |
199 | qemu_aio_release(elem); | |
200 | } | |
201 | } | |
202 | } | |
203 | ||
d354c7ec PB |
204 | static void thread_pool_cancel(BlockDriverAIOCB *acb) |
205 | { | |
206 | ThreadPoolElement *elem = (ThreadPoolElement *)acb; | |
b811203c | 207 | ThreadPool *pool = elem->pool; |
d354c7ec PB |
208 | |
209 | trace_thread_pool_cancel(elem, elem->common.opaque); | |
210 | ||
b811203c | 211 | qemu_mutex_lock(&pool->lock); |
d354c7ec PB |
212 | if (elem->state == THREAD_QUEUED && |
213 | /* No thread has yet started working on elem. we can try to "steal" | |
214 | * the item from the worker if we can get a signal from the | |
215 | * semaphore. Because this is non-blocking, we can do it with | |
216 | * the lock taken and ensure that elem will remain THREAD_QUEUED. | |
217 | */ | |
b811203c SH |
218 | qemu_sem_timedwait(&pool->sem, 0) == 0) { |
219 | QTAILQ_REMOVE(&pool->request_list, elem, reqs); | |
d354c7ec | 220 | elem->state = THREAD_CANCELED; |
c2e50e3d | 221 | qemu_bh_schedule(pool->completion_bh); |
d354c7ec | 222 | } else { |
b811203c | 223 | pool->pending_cancellations++; |
d354c7ec | 224 | while (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) { |
b811203c | 225 | qemu_cond_wait(&pool->check_cancel, &pool->lock); |
d354c7ec | 226 | } |
b811203c | 227 | pool->pending_cancellations--; |
d354c7ec | 228 | } |
b811203c | 229 | qemu_mutex_unlock(&pool->lock); |
c2e50e3d | 230 | thread_pool_completion_bh(pool); |
d354c7ec PB |
231 | } |
232 | ||
d7331bed | 233 | static const AIOCBInfo thread_pool_aiocb_info = { |
d354c7ec PB |
234 | .aiocb_size = sizeof(ThreadPoolElement), |
235 | .cancel = thread_pool_cancel, | |
236 | }; | |
237 | ||
c4d9d196 SH |
238 | BlockDriverAIOCB *thread_pool_submit_aio(ThreadPool *pool, |
239 | ThreadPoolFunc *func, void *arg, | |
d354c7ec PB |
240 | BlockDriverCompletionFunc *cb, void *opaque) |
241 | { | |
242 | ThreadPoolElement *req; | |
243 | ||
d7331bed | 244 | req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque); |
d354c7ec PB |
245 | req->func = func; |
246 | req->arg = arg; | |
247 | req->state = THREAD_QUEUED; | |
b811203c | 248 | req->pool = pool; |
d354c7ec | 249 | |
b811203c | 250 | QLIST_INSERT_HEAD(&pool->head, req, all); |
d354c7ec | 251 | |
b811203c | 252 | trace_thread_pool_submit(pool, req, arg); |
d354c7ec | 253 | |
b811203c SH |
254 | qemu_mutex_lock(&pool->lock); |
255 | if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) { | |
256 | spawn_thread(pool); | |
d354c7ec | 257 | } |
b811203c SH |
258 | QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs); |
259 | qemu_mutex_unlock(&pool->lock); | |
260 | qemu_sem_post(&pool->sem); | |
d354c7ec PB |
261 | return &req->common; |
262 | } | |
263 | ||
264 | typedef struct ThreadPoolCo { | |
265 | Coroutine *co; | |
266 | int ret; | |
267 | } ThreadPoolCo; | |
268 | ||
269 | static void thread_pool_co_cb(void *opaque, int ret) | |
270 | { | |
271 | ThreadPoolCo *co = opaque; | |
272 | ||
273 | co->ret = ret; | |
274 | qemu_coroutine_enter(co->co, NULL); | |
275 | } | |
276 | ||
c4d9d196 SH |
277 | int coroutine_fn thread_pool_submit_co(ThreadPool *pool, ThreadPoolFunc *func, |
278 | void *arg) | |
d354c7ec PB |
279 | { |
280 | ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS }; | |
281 | assert(qemu_in_coroutine()); | |
c4d9d196 | 282 | thread_pool_submit_aio(pool, func, arg, thread_pool_co_cb, &tpc); |
d354c7ec PB |
283 | qemu_coroutine_yield(); |
284 | return tpc.ret; | |
285 | } | |
286 | ||
c4d9d196 | 287 | void thread_pool_submit(ThreadPool *pool, ThreadPoolFunc *func, void *arg) |
d354c7ec | 288 | { |
c4d9d196 | 289 | thread_pool_submit_aio(pool, func, arg, NULL, NULL); |
d354c7ec PB |
290 | } |
291 | ||
b811203c SH |
292 | static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx) |
293 | { | |
294 | if (!ctx) { | |
295 | ctx = qemu_get_aio_context(); | |
296 | } | |
297 | ||
298 | memset(pool, 0, sizeof(*pool)); | |
f7311ccc | 299 | pool->ctx = ctx; |
c2e50e3d | 300 | pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool); |
b811203c SH |
301 | qemu_mutex_init(&pool->lock); |
302 | qemu_cond_init(&pool->check_cancel); | |
f7311ccc | 303 | qemu_cond_init(&pool->worker_stopped); |
b811203c SH |
304 | qemu_sem_init(&pool->sem, 0); |
305 | pool->max_threads = 64; | |
306 | pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool); | |
307 | ||
308 | QLIST_INIT(&pool->head); | |
309 | QTAILQ_INIT(&pool->request_list); | |
b811203c SH |
310 | } |
311 | ||
f7311ccc SH |
312 | ThreadPool *thread_pool_new(AioContext *ctx) |
313 | { | |
314 | ThreadPool *pool = g_new(ThreadPool, 1); | |
315 | thread_pool_init_one(pool, ctx); | |
316 | return pool; | |
317 | } | |
318 | ||
319 | void thread_pool_free(ThreadPool *pool) | |
320 | { | |
321 | if (!pool) { | |
322 | return; | |
323 | } | |
324 | ||
325 | assert(QLIST_EMPTY(&pool->head)); | |
326 | ||
327 | qemu_mutex_lock(&pool->lock); | |
328 | ||
329 | /* Stop new threads from spawning */ | |
330 | qemu_bh_delete(pool->new_thread_bh); | |
331 | pool->cur_threads -= pool->new_threads; | |
332 | pool->new_threads = 0; | |
333 | ||
334 | /* Wait for worker threads to terminate */ | |
335 | pool->stopping = true; | |
336 | while (pool->cur_threads > 0) { | |
337 | qemu_sem_post(&pool->sem); | |
338 | qemu_cond_wait(&pool->worker_stopped, &pool->lock); | |
339 | } | |
340 | ||
341 | qemu_mutex_unlock(&pool->lock); | |
342 | ||
c2e50e3d | 343 | qemu_bh_delete(pool->completion_bh); |
f7311ccc SH |
344 | qemu_sem_destroy(&pool->sem); |
345 | qemu_cond_destroy(&pool->check_cancel); | |
346 | qemu_cond_destroy(&pool->worker_stopped); | |
347 | qemu_mutex_destroy(&pool->lock); | |
f7311ccc SH |
348 | g_free(pool); |
349 | } |