Merge remote-tracking branch 'remotes/mst/tags/for_upstream' into staging
[qemu/ar7.git] / util / thread-pool.c
blobd763cea505b68575af8e1b39cd95d02caa73801a
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/osdep.h"
18 #include "qemu/queue.h"
19 #include "qemu/thread.h"
20 #include "qemu/coroutine.h"
21 #include "trace.h"
22 #include "block/thread-pool.h"
23 #include "qemu/main-loop.h"
25 static void do_spawn_thread(ThreadPool *pool);
27 typedef struct ThreadPoolElement ThreadPoolElement;
29 enum ThreadState {
30     THREAD_QUEUED,
31     THREAD_ACTIVE,
32     THREAD_DONE,
35 struct ThreadPoolElement {
36     BlockAIOCB common;
37     ThreadPool *pool;
38     ThreadPoolFunc *func;
39     void *arg;
41     /* Moving state out of THREAD_QUEUED is protected by lock.  After
42      * that, only the worker thread can write to it.  Reads and writes
43      * of state and ret are ordered with memory barriers.
44      */
45     enum ThreadState state;
46     int ret;
48     /* Access to this list is protected by lock.  */
49     QTAILQ_ENTRY(ThreadPoolElement) reqs;
51     /* Access to this list is protected by the global mutex.  */
52     QLIST_ENTRY(ThreadPoolElement) all;
55 struct ThreadPool {
56     AioContext *ctx;
57     QEMUBH *completion_bh;
58     QemuMutex lock;
59     QemuCond worker_stopped;
60     QemuSemaphore sem;
61     int max_threads;
62     QEMUBH *new_thread_bh;
64     /* The following variables are only accessed from one AioContext. */
65     QLIST_HEAD(, ThreadPoolElement) head;
67     /* The following variables are protected by lock.  */
68     QTAILQ_HEAD(, ThreadPoolElement) request_list;
69     int cur_threads;
70     int idle_threads;
71     int new_threads;     /* backlog of threads we need to create */
72     int pending_threads; /* threads created but not running yet */
73     bool stopping;
76 static void *worker_thread(void *opaque)
78     ThreadPool *pool = opaque;
80     qemu_mutex_lock(&pool->lock);
81     pool->pending_threads--;
82     do_spawn_thread(pool);
84     while (!pool->stopping) {
85         ThreadPoolElement *req;
86         int ret;
88         do {
89             pool->idle_threads++;
90             qemu_mutex_unlock(&pool->lock);
91             ret = qemu_sem_timedwait(&pool->sem, 10000);
92             qemu_mutex_lock(&pool->lock);
93             pool->idle_threads--;
94         } while (ret == -1 && !QTAILQ_EMPTY(&pool->request_list));
95         if (ret == -1 || pool->stopping) {
96             break;
97         }
99         req = QTAILQ_FIRST(&pool->request_list);
100         QTAILQ_REMOVE(&pool->request_list, req, reqs);
101         req->state = THREAD_ACTIVE;
102         qemu_mutex_unlock(&pool->lock);
104         ret = req->func(req->arg);
106         req->ret = ret;
107         /* Write ret before state.  */
108         smp_wmb();
109         req->state = THREAD_DONE;
111         qemu_mutex_lock(&pool->lock);
113         qemu_bh_schedule(pool->completion_bh);
114     }
116     pool->cur_threads--;
117     qemu_cond_signal(&pool->worker_stopped);
118     qemu_mutex_unlock(&pool->lock);
119     return NULL;
122 static void do_spawn_thread(ThreadPool *pool)
124     QemuThread t;
126     /* Runs with lock taken.  */
127     if (!pool->new_threads) {
128         return;
129     }
131     pool->new_threads--;
132     pool->pending_threads++;
134     qemu_thread_create(&t, "worker", worker_thread, pool, QEMU_THREAD_DETACHED);
137 static void spawn_thread_bh_fn(void *opaque)
139     ThreadPool *pool = opaque;
141     qemu_mutex_lock(&pool->lock);
142     do_spawn_thread(pool);
143     qemu_mutex_unlock(&pool->lock);
146 static void spawn_thread(ThreadPool *pool)
148     pool->cur_threads++;
149     pool->new_threads++;
150     /* If there are threads being created, they will spawn new workers, so
151      * we don't spend time creating many threads in a loop holding a mutex or
152      * starving the current vcpu.
153      *
154      * If there are no idle threads, ask the main thread to create one, so we
155      * inherit the correct affinity instead of the vcpu affinity.
156      */
157     if (!pool->pending_threads) {
158         qemu_bh_schedule(pool->new_thread_bh);
159     }
162 static void thread_pool_completion_bh(void *opaque)
164     ThreadPool *pool = opaque;
165     ThreadPoolElement *elem, *next;
167     aio_context_acquire(pool->ctx);
168 restart:
169     QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
170         if (elem->state != THREAD_DONE) {
171             continue;
172         }
174         trace_thread_pool_complete(pool, elem, elem->common.opaque,
175                                    elem->ret);
176         QLIST_REMOVE(elem, all);
178         if (elem->common.cb) {
179             /* Read state before ret.  */
180             smp_rmb();
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);
187             aio_context_release(pool->ctx);
188             elem->common.cb(elem->common.opaque, elem->ret);
189             aio_context_acquire(pool->ctx);
191             /* We can safely cancel the completion_bh here regardless of someone
192              * else having scheduled it meanwhile because we reenter the
193              * completion function anyway (goto restart).
194              */
195             qemu_bh_cancel(pool->completion_bh);
197             qemu_aio_unref(elem);
198             goto restart;
199         } else {
200             qemu_aio_unref(elem);
201         }
202     }
203     aio_context_release(pool->ctx);
206 static void thread_pool_cancel(BlockAIOCB *acb)
208     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
209     ThreadPool *pool = elem->pool;
211     trace_thread_pool_cancel(elem, elem->common.opaque);
213     QEMU_LOCK_GUARD(&pool->lock);
214     if (elem->state == THREAD_QUEUED &&
215         /* No thread has yet started working on elem. we can try to "steal"
216          * the item from the worker if we can get a signal from the
217          * semaphore.  Because this is non-blocking, we can do it with
218          * the lock taken and ensure that elem will remain THREAD_QUEUED.
219          */
220         qemu_sem_timedwait(&pool->sem, 0) == 0) {
221         QTAILQ_REMOVE(&pool->request_list, elem, reqs);
222         qemu_bh_schedule(pool->completion_bh);
224         elem->state = THREAD_DONE;
225         elem->ret = -ECANCELED;
226     }
230 static AioContext *thread_pool_get_aio_context(BlockAIOCB *acb)
232     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
233     ThreadPool *pool = elem->pool;
234     return pool->ctx;
237 static const AIOCBInfo thread_pool_aiocb_info = {
238     .aiocb_size         = sizeof(ThreadPoolElement),
239     .cancel_async       = thread_pool_cancel,
240     .get_aio_context    = thread_pool_get_aio_context,
243 BlockAIOCB *thread_pool_submit_aio(ThreadPool *pool,
244         ThreadPoolFunc *func, void *arg,
245         BlockCompletionFunc *cb, void *opaque)
247     ThreadPoolElement *req;
249     req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque);
250     req->func = func;
251     req->arg = arg;
252     req->state = THREAD_QUEUED;
253     req->pool = pool;
255     QLIST_INSERT_HEAD(&pool->head, req, all);
257     trace_thread_pool_submit(pool, req, arg);
259     qemu_mutex_lock(&pool->lock);
260     if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) {
261         spawn_thread(pool);
262     }
263     QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs);
264     qemu_mutex_unlock(&pool->lock);
265     qemu_sem_post(&pool->sem);
266     return &req->common;
269 typedef struct ThreadPoolCo {
270     Coroutine *co;
271     int ret;
272 } ThreadPoolCo;
274 static void thread_pool_co_cb(void *opaque, int ret)
276     ThreadPoolCo *co = opaque;
278     co->ret = ret;
279     aio_co_wake(co->co);
282 int coroutine_fn thread_pool_submit_co(ThreadPool *pool, ThreadPoolFunc *func,
283                                        void *arg)
285     ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
286     assert(qemu_in_coroutine());
287     thread_pool_submit_aio(pool, func, arg, thread_pool_co_cb, &tpc);
288     qemu_coroutine_yield();
289     return tpc.ret;
292 void thread_pool_submit(ThreadPool *pool, ThreadPoolFunc *func, void *arg)
294     thread_pool_submit_aio(pool, func, arg, NULL, NULL);
297 static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx)
299     if (!ctx) {
300         ctx = qemu_get_aio_context();
301     }
303     memset(pool, 0, sizeof(*pool));
304     pool->ctx = ctx;
305     pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool);
306     qemu_mutex_init(&pool->lock);
307     qemu_cond_init(&pool->worker_stopped);
308     qemu_sem_init(&pool->sem, 0);
309     pool->max_threads = 64;
310     pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool);
312     QLIST_INIT(&pool->head);
313     QTAILQ_INIT(&pool->request_list);
316 ThreadPool *thread_pool_new(AioContext *ctx)
318     ThreadPool *pool = g_new(ThreadPool, 1);
319     thread_pool_init_one(pool, ctx);
320     return pool;
323 void thread_pool_free(ThreadPool *pool)
325     if (!pool) {
326         return;
327     }
329     assert(QLIST_EMPTY(&pool->head));
331     qemu_mutex_lock(&pool->lock);
333     /* Stop new threads from spawning */
334     qemu_bh_delete(pool->new_thread_bh);
335     pool->cur_threads -= pool->new_threads;
336     pool->new_threads = 0;
338     /* Wait for worker threads to terminate */
339     pool->stopping = true;
340     while (pool->cur_threads > 0) {
341         qemu_sem_post(&pool->sem);
342         qemu_cond_wait(&pool->worker_stopped, &pool->lock);
343     }
345     qemu_mutex_unlock(&pool->lock);
347     qemu_bh_delete(pool->completion_bh);
348     qemu_sem_destroy(&pool->sem);
349     qemu_cond_destroy(&pool->worker_stopped);
350     qemu_mutex_destroy(&pool->lock);
351     g_free(pool);