]> git.proxmox.com Git - mirror_qemu.git/blame - thread-pool.c
virtio: minor cleanup
[mirror_qemu.git] / thread-pool.c
CommitLineData
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 <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"
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 26static void do_spawn_thread(ThreadPool *pool);
d354c7ec
PB
27
28typedef struct ThreadPoolElement ThreadPoolElement;
29
30enum ThreadState {
31 THREAD_QUEUED,
32 THREAD_ACTIVE,
33 THREAD_DONE,
d354c7ec
PB
34};
35
36struct ThreadPoolElement {
7c84b1b8 37 BlockAIOCB common;
b811203c 38 ThreadPool *pool;
d354c7ec
PB
39 ThreadPoolFunc *func;
40 void *arg;
19d092cf
PB
41
42 /* Moving state out of THREAD_QUEUED is protected by lock. After
43 * that, only the worker thread can write to it. Reads and writes
44 * of state and ret are ordered with memory barriers.
45 */
d354c7ec
PB
46 enum ThreadState state;
47 int ret;
48
49 /* Access to this list is protected by lock. */
50 QTAILQ_ENTRY(ThreadPoolElement) reqs;
51
52 /* Access to this list is protected by the global mutex. */
53 QLIST_ENTRY(ThreadPoolElement) all;
54};
55
b811203c 56struct ThreadPool {
f7311ccc 57 AioContext *ctx;
c2e50e3d 58 QEMUBH *completion_bh;
b811203c 59 QemuMutex lock;
f7311ccc 60 QemuCond worker_stopped;
b811203c
SH
61 QemuSemaphore sem;
62 int max_threads;
63 QEMUBH *new_thread_bh;
64
65 /* The following variables are only accessed from one AioContext. */
66 QLIST_HEAD(, ThreadPoolElement) head;
67
68 /* The following variables are protected by lock. */
69 QTAILQ_HEAD(, ThreadPoolElement) request_list;
70 int cur_threads;
71 int idle_threads;
72 int new_threads; /* backlog of threads we need to create */
73 int pending_threads; /* threads created but not running yet */
f7311ccc 74 bool stopping;
b811203c
SH
75};
76
b811203c 77static void *worker_thread(void *opaque)
d354c7ec 78{
b811203c
SH
79 ThreadPool *pool = opaque;
80
81 qemu_mutex_lock(&pool->lock);
82 pool->pending_threads--;
83 do_spawn_thread(pool);
d354c7ec 84
f7311ccc 85 while (!pool->stopping) {
d354c7ec
PB
86 ThreadPoolElement *req;
87 int ret;
88
89 do {
b811203c
SH
90 pool->idle_threads++;
91 qemu_mutex_unlock(&pool->lock);
92 ret = qemu_sem_timedwait(&pool->sem, 10000);
93 qemu_mutex_lock(&pool->lock);
94 pool->idle_threads--;
95 } while (ret == -1 && !QTAILQ_EMPTY(&pool->request_list));
f7311ccc 96 if (ret == -1 || pool->stopping) {
d354c7ec
PB
97 break;
98 }
99
b811203c
SH
100 req = QTAILQ_FIRST(&pool->request_list);
101 QTAILQ_REMOVE(&pool->request_list, req, reqs);
d354c7ec 102 req->state = THREAD_ACTIVE;
b811203c 103 qemu_mutex_unlock(&pool->lock);
d354c7ec
PB
104
105 ret = req->func(req->arg);
106
d354c7ec 107 req->ret = ret;
19d092cf
PB
108 /* Write ret before state. */
109 smp_wmb();
110 req->state = THREAD_DONE;
111
b811203c 112 qemu_mutex_lock(&pool->lock);
d354c7ec 113
c2e50e3d 114 qemu_bh_schedule(pool->completion_bh);
d354c7ec
PB
115 }
116
b811203c 117 pool->cur_threads--;
f7311ccc 118 qemu_cond_signal(&pool->worker_stopped);
b811203c 119 qemu_mutex_unlock(&pool->lock);
d354c7ec
PB
120 return NULL;
121}
122
b811203c 123static void do_spawn_thread(ThreadPool *pool)
d354c7ec
PB
124{
125 QemuThread t;
126
127 /* Runs with lock taken. */
b811203c 128 if (!pool->new_threads) {
d354c7ec
PB
129 return;
130 }
131
b811203c
SH
132 pool->new_threads--;
133 pool->pending_threads++;
d354c7ec 134
4900116e 135 qemu_thread_create(&t, "worker", worker_thread, pool, QEMU_THREAD_DETACHED);
d354c7ec
PB
136}
137
138static void spawn_thread_bh_fn(void *opaque)
139{
b811203c
SH
140 ThreadPool *pool = opaque;
141
142 qemu_mutex_lock(&pool->lock);
143 do_spawn_thread(pool);
144 qemu_mutex_unlock(&pool->lock);
d354c7ec
PB
145}
146
b811203c 147static void spawn_thread(ThreadPool *pool)
d354c7ec 148{
b811203c
SH
149 pool->cur_threads++;
150 pool->new_threads++;
d354c7ec
PB
151 /* If there are threads being created, they will spawn new workers, so
152 * we don't spend time creating many threads in a loop holding a mutex or
153 * starving the current vcpu.
154 *
155 * If there are no idle threads, ask the main thread to create one, so we
156 * inherit the correct affinity instead of the vcpu affinity.
157 */
b811203c
SH
158 if (!pool->pending_threads) {
159 qemu_bh_schedule(pool->new_thread_bh);
d354c7ec
PB
160 }
161}
162
c2e50e3d 163static void thread_pool_completion_bh(void *opaque)
d354c7ec 164{
c2e50e3d 165 ThreadPool *pool = opaque;
d354c7ec
PB
166 ThreadPoolElement *elem, *next;
167
d354c7ec 168restart:
b811203c 169 QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
3391f5e5 170 if (elem->state != THREAD_DONE) {
d354c7ec
PB
171 continue;
172 }
1faa5bb7
SH
173
174 trace_thread_pool_complete(pool, elem, elem->common.opaque,
175 elem->ret);
176 QLIST_REMOVE(elem, all);
177
178 if (elem->common.cb) {
19d092cf
PB
179 /* Read state before ret. */
180 smp_rmb();
3c80ca15
SH
181
182 /* Schedule ourselves in case elem->common.cb() calls aio_poll() to
183 * wait for another request that completed at the same time.
184 */
185 qemu_bh_schedule(pool->completion_bh);
186
19d092cf 187 elem->common.cb(elem->common.opaque, elem->ret);
8007429a 188 qemu_aio_unref(elem);
d354c7ec
PB
189 goto restart;
190 } else {
8007429a 191 qemu_aio_unref(elem);
d354c7ec
PB
192 }
193 }
194}
195
7c84b1b8 196static void thread_pool_cancel(BlockAIOCB *acb)
d354c7ec
PB
197{
198 ThreadPoolElement *elem = (ThreadPoolElement *)acb;
b811203c 199 ThreadPool *pool = elem->pool;
d354c7ec
PB
200
201 trace_thread_pool_cancel(elem, elem->common.opaque);
202
b811203c 203 qemu_mutex_lock(&pool->lock);
d354c7ec
PB
204 if (elem->state == THREAD_QUEUED &&
205 /* No thread has yet started working on elem. we can try to "steal"
206 * the item from the worker if we can get a signal from the
207 * semaphore. Because this is non-blocking, we can do it with
208 * the lock taken and ensure that elem will remain THREAD_QUEUED.
209 */
b811203c
SH
210 qemu_sem_timedwait(&pool->sem, 0) == 0) {
211 QTAILQ_REMOVE(&pool->request_list, elem, reqs);
c2e50e3d 212 qemu_bh_schedule(pool->completion_bh);
3391f5e5
FZ
213
214 elem->state = THREAD_DONE;
215 elem->ret = -ECANCELED;
d354c7ec 216 }
3391f5e5 217
b811203c 218 qemu_mutex_unlock(&pool->lock);
3391f5e5
FZ
219}
220
7c84b1b8 221static AioContext *thread_pool_get_aio_context(BlockAIOCB *acb)
3391f5e5
FZ
222{
223 ThreadPoolElement *elem = (ThreadPoolElement *)acb;
224 ThreadPool *pool = elem->pool;
225 return pool->ctx;
d354c7ec
PB
226}
227
d7331bed 228static const AIOCBInfo thread_pool_aiocb_info = {
d354c7ec 229 .aiocb_size = sizeof(ThreadPoolElement),
3391f5e5
FZ
230 .cancel_async = thread_pool_cancel,
231 .get_aio_context = thread_pool_get_aio_context,
d354c7ec
PB
232};
233
7c84b1b8 234BlockAIOCB *thread_pool_submit_aio(ThreadPool *pool,
c4d9d196 235 ThreadPoolFunc *func, void *arg,
097310b5 236 BlockCompletionFunc *cb, void *opaque)
d354c7ec
PB
237{
238 ThreadPoolElement *req;
239
d7331bed 240 req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque);
d354c7ec
PB
241 req->func = func;
242 req->arg = arg;
243 req->state = THREAD_QUEUED;
b811203c 244 req->pool = pool;
d354c7ec 245
b811203c 246 QLIST_INSERT_HEAD(&pool->head, req, all);
d354c7ec 247
b811203c 248 trace_thread_pool_submit(pool, req, arg);
d354c7ec 249
b811203c
SH
250 qemu_mutex_lock(&pool->lock);
251 if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) {
252 spawn_thread(pool);
d354c7ec 253 }
b811203c
SH
254 QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs);
255 qemu_mutex_unlock(&pool->lock);
256 qemu_sem_post(&pool->sem);
d354c7ec
PB
257 return &req->common;
258}
259
260typedef struct ThreadPoolCo {
261 Coroutine *co;
262 int ret;
263} ThreadPoolCo;
264
265static void thread_pool_co_cb(void *opaque, int ret)
266{
267 ThreadPoolCo *co = opaque;
268
269 co->ret = ret;
270 qemu_coroutine_enter(co->co, NULL);
271}
272
c4d9d196
SH
273int coroutine_fn thread_pool_submit_co(ThreadPool *pool, ThreadPoolFunc *func,
274 void *arg)
d354c7ec
PB
275{
276 ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
277 assert(qemu_in_coroutine());
c4d9d196 278 thread_pool_submit_aio(pool, func, arg, thread_pool_co_cb, &tpc);
d354c7ec
PB
279 qemu_coroutine_yield();
280 return tpc.ret;
281}
282
c4d9d196 283void thread_pool_submit(ThreadPool *pool, ThreadPoolFunc *func, void *arg)
d354c7ec 284{
c4d9d196 285 thread_pool_submit_aio(pool, func, arg, NULL, NULL);
d354c7ec
PB
286}
287
b811203c
SH
288static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx)
289{
290 if (!ctx) {
291 ctx = qemu_get_aio_context();
292 }
293
294 memset(pool, 0, sizeof(*pool));
f7311ccc 295 pool->ctx = ctx;
c2e50e3d 296 pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool);
b811203c 297 qemu_mutex_init(&pool->lock);
f7311ccc 298 qemu_cond_init(&pool->worker_stopped);
b811203c
SH
299 qemu_sem_init(&pool->sem, 0);
300 pool->max_threads = 64;
301 pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool);
302
303 QLIST_INIT(&pool->head);
304 QTAILQ_INIT(&pool->request_list);
b811203c
SH
305}
306
f7311ccc
SH
307ThreadPool *thread_pool_new(AioContext *ctx)
308{
309 ThreadPool *pool = g_new(ThreadPool, 1);
310 thread_pool_init_one(pool, ctx);
311 return pool;
312}
313
314void thread_pool_free(ThreadPool *pool)
315{
316 if (!pool) {
317 return;
318 }
319
320 assert(QLIST_EMPTY(&pool->head));
321
322 qemu_mutex_lock(&pool->lock);
323
324 /* Stop new threads from spawning */
325 qemu_bh_delete(pool->new_thread_bh);
326 pool->cur_threads -= pool->new_threads;
327 pool->new_threads = 0;
328
329 /* Wait for worker threads to terminate */
330 pool->stopping = true;
331 while (pool->cur_threads > 0) {
332 qemu_sem_post(&pool->sem);
333 qemu_cond_wait(&pool->worker_stopped, &pool->lock);
334 }
335
336 qemu_mutex_unlock(&pool->lock);
337
c2e50e3d 338 qemu_bh_delete(pool->completion_bh);
f7311ccc 339 qemu_sem_destroy(&pool->sem);
f7311ccc
SH
340 qemu_cond_destroy(&pool->worker_stopped);
341 qemu_mutex_destroy(&pool->lock);
f7311ccc
SH
342 g_free(pool);
343}