xref: /qemu/util/thread-pool.c (revision b49f4755)
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/defer-call.h"
19 #include "qemu/queue.h"
20 #include "qemu/thread.h"
21 #include "qemu/coroutine.h"
22 #include "trace.h"
23 #include "block/thread-pool.h"
24 #include "qemu/main-loop.h"
25 
26 static void do_spawn_thread(ThreadPool *pool);
27 
28 typedef struct ThreadPoolElement ThreadPoolElement;
29 
30 enum ThreadState {
31     THREAD_QUEUED,
32     THREAD_ACTIVE,
33     THREAD_DONE,
34 };
35 
36 struct ThreadPoolElement {
37     BlockAIOCB common;
38     ThreadPool *pool;
39     ThreadPoolFunc *func;
40     void *arg;
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      */
46     enum ThreadState state;
47     int ret;
48 
49     /* Access to this list is protected by lock.  */
50     QTAILQ_ENTRY(ThreadPoolElement) reqs;
51 
52     /* This list is only written by the thread pool's mother thread.  */
53     QLIST_ENTRY(ThreadPoolElement) all;
54 };
55 
56 struct ThreadPool {
57     AioContext *ctx;
58     QEMUBH *completion_bh;
59     QemuMutex lock;
60     QemuCond worker_stopped;
61     QemuCond request_cond;
62     QEMUBH *new_thread_bh;
63 
64     /* The following variables are only accessed from one AioContext. */
65     QLIST_HEAD(, ThreadPoolElement) head;
66 
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     int min_threads;
74     int max_threads;
75 };
76 
77 static void *worker_thread(void *opaque)
78 {
79     ThreadPool *pool = opaque;
80 
81     qemu_mutex_lock(&pool->lock);
82     pool->pending_threads--;
83     do_spawn_thread(pool);
84 
85     while (pool->cur_threads <= pool->max_threads) {
86         ThreadPoolElement *req;
87         int ret;
88 
89         if (QTAILQ_EMPTY(&pool->request_list)) {
90             pool->idle_threads++;
91             ret = qemu_cond_timedwait(&pool->request_cond, &pool->lock, 10000);
92             pool->idle_threads--;
93             if (ret == 0 &&
94                 QTAILQ_EMPTY(&pool->request_list) &&
95                 pool->cur_threads > pool->min_threads) {
96                 /* Timed out + no work to do + no need for warm threads = exit.  */
97                 break;
98             }
99             /*
100              * Even if there was some work to do, check if there aren't
101              * too many worker threads before picking it up.
102              */
103             continue;
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_bh_schedule(pool->completion_bh);
119         qemu_mutex_lock(&pool->lock);
120     }
121 
122     pool->cur_threads--;
123     qemu_cond_signal(&pool->worker_stopped);
124 
125     /*
126      * Wake up another thread, in case we got a wakeup but decided
127      * to exit due to pool->cur_threads > pool->max_threads.
128      */
129     qemu_cond_signal(&pool->request_cond);
130     qemu_mutex_unlock(&pool->lock);
131     return NULL;
132 }
133 
134 static void do_spawn_thread(ThreadPool *pool)
135 {
136     QemuThread t;
137 
138     /* Runs with lock taken.  */
139     if (!pool->new_threads) {
140         return;
141     }
142 
143     pool->new_threads--;
144     pool->pending_threads++;
145 
146     qemu_thread_create(&t, "worker", worker_thread, pool, QEMU_THREAD_DETACHED);
147 }
148 
149 static void spawn_thread_bh_fn(void *opaque)
150 {
151     ThreadPool *pool = opaque;
152 
153     qemu_mutex_lock(&pool->lock);
154     do_spawn_thread(pool);
155     qemu_mutex_unlock(&pool->lock);
156 }
157 
158 static void spawn_thread(ThreadPool *pool)
159 {
160     pool->cur_threads++;
161     pool->new_threads++;
162     /* If there are threads being created, they will spawn new workers, so
163      * we don't spend time creating many threads in a loop holding a mutex or
164      * starving the current vcpu.
165      *
166      * If there are no idle threads, ask the main thread to create one, so we
167      * inherit the correct affinity instead of the vcpu affinity.
168      */
169     if (!pool->pending_threads) {
170         qemu_bh_schedule(pool->new_thread_bh);
171     }
172 }
173 
174 static void thread_pool_completion_bh(void *opaque)
175 {
176     ThreadPool *pool = opaque;
177     ThreadPoolElement *elem, *next;
178 
179     defer_call_begin(); /* cb() may use defer_call() to coalesce work */
180 
181 restart:
182     QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
183         if (elem->state != THREAD_DONE) {
184             continue;
185         }
186 
187         trace_thread_pool_complete(pool, elem, elem->common.opaque,
188                                    elem->ret);
189         QLIST_REMOVE(elem, all);
190 
191         if (elem->common.cb) {
192             /* Read state before ret.  */
193             smp_rmb();
194 
195             /* Schedule ourselves in case elem->common.cb() calls aio_poll() to
196              * wait for another request that completed at the same time.
197              */
198             qemu_bh_schedule(pool->completion_bh);
199 
200             elem->common.cb(elem->common.opaque, elem->ret);
201 
202             /* We can safely cancel the completion_bh here regardless of someone
203              * else having scheduled it meanwhile because we reenter the
204              * completion function anyway (goto restart).
205              */
206             qemu_bh_cancel(pool->completion_bh);
207 
208             qemu_aio_unref(elem);
209             goto restart;
210         } else {
211             qemu_aio_unref(elem);
212         }
213     }
214 
215     defer_call_end();
216 }
217 
218 static void thread_pool_cancel(BlockAIOCB *acb)
219 {
220     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
221     ThreadPool *pool = elem->pool;
222 
223     trace_thread_pool_cancel(elem, elem->common.opaque);
224 
225     QEMU_LOCK_GUARD(&pool->lock);
226     if (elem->state == THREAD_QUEUED) {
227         QTAILQ_REMOVE(&pool->request_list, elem, reqs);
228         qemu_bh_schedule(pool->completion_bh);
229 
230         elem->state = THREAD_DONE;
231         elem->ret = -ECANCELED;
232     }
233 
234 }
235 
236 static const AIOCBInfo thread_pool_aiocb_info = {
237     .aiocb_size         = sizeof(ThreadPoolElement),
238     .cancel_async       = thread_pool_cancel,
239 };
240 
241 BlockAIOCB *thread_pool_submit_aio(ThreadPoolFunc *func, void *arg,
242                                    BlockCompletionFunc *cb, void *opaque)
243 {
244     ThreadPoolElement *req;
245     AioContext *ctx = qemu_get_current_aio_context();
246     ThreadPool *pool = aio_get_thread_pool(ctx);
247 
248     /* Assert that the thread submitting work is the same running the pool */
249     assert(pool->ctx == qemu_get_current_aio_context());
250 
251     req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque);
252     req->func = func;
253     req->arg = arg;
254     req->state = THREAD_QUEUED;
255     req->pool = pool;
256 
257     QLIST_INSERT_HEAD(&pool->head, req, all);
258 
259     trace_thread_pool_submit(pool, req, arg);
260 
261     qemu_mutex_lock(&pool->lock);
262     if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) {
263         spawn_thread(pool);
264     }
265     QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs);
266     qemu_mutex_unlock(&pool->lock);
267     qemu_cond_signal(&pool->request_cond);
268     return &req->common;
269 }
270 
271 typedef struct ThreadPoolCo {
272     Coroutine *co;
273     int ret;
274 } ThreadPoolCo;
275 
276 static void thread_pool_co_cb(void *opaque, int ret)
277 {
278     ThreadPoolCo *co = opaque;
279 
280     co->ret = ret;
281     aio_co_wake(co->co);
282 }
283 
284 int coroutine_fn thread_pool_submit_co(ThreadPoolFunc *func, void *arg)
285 {
286     ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
287     assert(qemu_in_coroutine());
288     thread_pool_submit_aio(func, arg, thread_pool_co_cb, &tpc);
289     qemu_coroutine_yield();
290     return tpc.ret;
291 }
292 
293 void thread_pool_submit(ThreadPoolFunc *func, void *arg)
294 {
295     thread_pool_submit_aio(func, arg, NULL, NULL);
296 }
297 
298 void thread_pool_update_params(ThreadPool *pool, AioContext *ctx)
299 {
300     qemu_mutex_lock(&pool->lock);
301 
302     pool->min_threads = ctx->thread_pool_min;
303     pool->max_threads = ctx->thread_pool_max;
304 
305     /*
306      * We either have to:
307      *  - Increase the number available of threads until over the min_threads
308      *    threshold.
309      *  - Bump the worker threads so that they exit, until under the max_threads
310      *    threshold.
311      *  - Do nothing. The current number of threads fall in between the min and
312      *    max thresholds. We'll let the pool manage itself.
313      */
314     for (int i = pool->cur_threads; i < pool->min_threads; i++) {
315         spawn_thread(pool);
316     }
317 
318     for (int i = pool->cur_threads; i > pool->max_threads; i--) {
319         qemu_cond_signal(&pool->request_cond);
320     }
321 
322     qemu_mutex_unlock(&pool->lock);
323 }
324 
325 static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx)
326 {
327     if (!ctx) {
328         ctx = qemu_get_aio_context();
329     }
330 
331     memset(pool, 0, sizeof(*pool));
332     pool->ctx = ctx;
333     pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool);
334     qemu_mutex_init(&pool->lock);
335     qemu_cond_init(&pool->worker_stopped);
336     qemu_cond_init(&pool->request_cond);
337     pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool);
338 
339     QLIST_INIT(&pool->head);
340     QTAILQ_INIT(&pool->request_list);
341 
342     thread_pool_update_params(pool, ctx);
343 }
344 
345 ThreadPool *thread_pool_new(AioContext *ctx)
346 {
347     ThreadPool *pool = g_new(ThreadPool, 1);
348     thread_pool_init_one(pool, ctx);
349     return pool;
350 }
351 
352 void thread_pool_free(ThreadPool *pool)
353 {
354     if (!pool) {
355         return;
356     }
357 
358     assert(QLIST_EMPTY(&pool->head));
359 
360     qemu_mutex_lock(&pool->lock);
361 
362     /* Stop new threads from spawning */
363     qemu_bh_delete(pool->new_thread_bh);
364     pool->cur_threads -= pool->new_threads;
365     pool->new_threads = 0;
366 
367     /* Wait for worker threads to terminate */
368     pool->max_threads = 0;
369     qemu_cond_broadcast(&pool->request_cond);
370     while (pool->cur_threads > 0) {
371         qemu_cond_wait(&pool->worker_stopped, &pool->lock);
372     }
373 
374     qemu_mutex_unlock(&pool->lock);
375 
376     qemu_bh_delete(pool->completion_bh);
377     qemu_cond_destroy(&pool->request_cond);
378     qemu_cond_destroy(&pool->worker_stopped);
379     qemu_mutex_destroy(&pool->lock);
380     g_free(pool);
381 }
382