xref: /qemu/util/thread-pool.c (revision a6f4d2ec)
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"
24 
25 static void do_spawn_thread(ThreadPool *pool);
26 
27 typedef struct ThreadPoolElement ThreadPoolElement;
28 
29 enum ThreadState {
30     THREAD_QUEUED,
31     THREAD_ACTIVE,
32     THREAD_DONE,
33 };
34 
35 struct ThreadPoolElement {
36     BlockAIOCB common;
37     ThreadPool *pool;
38     ThreadPoolFunc *func;
39     void *arg;
40 
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;
47 
48     /* Access to this list is protected by lock.  */
49     QTAILQ_ENTRY(ThreadPoolElement) reqs;
50 
51     /* This list is only written by the thread pool's mother thread.  */
52     QLIST_ENTRY(ThreadPoolElement) all;
53 };
54 
55 struct ThreadPool {
56     AioContext *ctx;
57     QEMUBH *completion_bh;
58     QemuMutex lock;
59     QemuCond worker_stopped;
60     QemuCond request_cond;
61     QEMUBH *new_thread_bh;
62 
63     /* The following variables are only accessed from one AioContext. */
64     QLIST_HEAD(, ThreadPoolElement) head;
65 
66     /* The following variables are protected by lock.  */
67     QTAILQ_HEAD(, ThreadPoolElement) request_list;
68     int cur_threads;
69     int idle_threads;
70     int new_threads;     /* backlog of threads we need to create */
71     int pending_threads; /* threads created but not running yet */
72     int min_threads;
73     int max_threads;
74 };
75 
76 static void *worker_thread(void *opaque)
77 {
78     ThreadPool *pool = opaque;
79 
80     qemu_mutex_lock(&pool->lock);
81     pool->pending_threads--;
82     do_spawn_thread(pool);
83 
84     while (pool->cur_threads <= pool->max_threads) {
85         ThreadPoolElement *req;
86         int ret;
87 
88         if (QTAILQ_EMPTY(&pool->request_list)) {
89             pool->idle_threads++;
90             ret = qemu_cond_timedwait(&pool->request_cond, &pool->lock, 10000);
91             pool->idle_threads--;
92             if (ret == 0 &&
93                 QTAILQ_EMPTY(&pool->request_list) &&
94                 pool->cur_threads > pool->min_threads) {
95                 /* Timed out + no work to do + no need for warm threads = exit.  */
96                 break;
97             }
98             /*
99              * Even if there was some work to do, check if there aren't
100              * too many worker threads before picking it up.
101              */
102             continue;
103         }
104 
105         req = QTAILQ_FIRST(&pool->request_list);
106         QTAILQ_REMOVE(&pool->request_list, req, reqs);
107         req->state = THREAD_ACTIVE;
108         qemu_mutex_unlock(&pool->lock);
109 
110         ret = req->func(req->arg);
111 
112         req->ret = ret;
113         /* Write ret before state.  */
114         smp_wmb();
115         req->state = THREAD_DONE;
116 
117         qemu_bh_schedule(pool->completion_bh);
118         qemu_mutex_lock(&pool->lock);
119     }
120 
121     pool->cur_threads--;
122     qemu_cond_signal(&pool->worker_stopped);
123 
124     /*
125      * Wake up another thread, in case we got a wakeup but decided
126      * to exit due to pool->cur_threads > pool->max_threads.
127      */
128     qemu_cond_signal(&pool->request_cond);
129     qemu_mutex_unlock(&pool->lock);
130     return NULL;
131 }
132 
133 static void do_spawn_thread(ThreadPool *pool)
134 {
135     QemuThread t;
136 
137     /* Runs with lock taken.  */
138     if (!pool->new_threads) {
139         return;
140     }
141 
142     pool->new_threads--;
143     pool->pending_threads++;
144 
145     qemu_thread_create(&t, "worker", worker_thread, pool, QEMU_THREAD_DETACHED);
146 }
147 
148 static void spawn_thread_bh_fn(void *opaque)
149 {
150     ThreadPool *pool = opaque;
151 
152     qemu_mutex_lock(&pool->lock);
153     do_spawn_thread(pool);
154     qemu_mutex_unlock(&pool->lock);
155 }
156 
157 static void spawn_thread(ThreadPool *pool)
158 {
159     pool->cur_threads++;
160     pool->new_threads++;
161     /* If there are threads being created, they will spawn new workers, so
162      * we don't spend time creating many threads in a loop holding a mutex or
163      * starving the current vcpu.
164      *
165      * If there are no idle threads, ask the main thread to create one, so we
166      * inherit the correct affinity instead of the vcpu affinity.
167      */
168     if (!pool->pending_threads) {
169         qemu_bh_schedule(pool->new_thread_bh);
170     }
171 }
172 
173 static void thread_pool_completion_bh(void *opaque)
174 {
175     ThreadPool *pool = opaque;
176     ThreadPoolElement *elem, *next;
177 
178 restart:
179     QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
180         if (elem->state != THREAD_DONE) {
181             continue;
182         }
183 
184         trace_thread_pool_complete(pool, elem, elem->common.opaque,
185                                    elem->ret);
186         QLIST_REMOVE(elem, all);
187 
188         if (elem->common.cb) {
189             /* Read state before ret.  */
190             smp_rmb();
191 
192             /* Schedule ourselves in case elem->common.cb() calls aio_poll() to
193              * wait for another request that completed at the same time.
194              */
195             qemu_bh_schedule(pool->completion_bh);
196 
197             elem->common.cb(elem->common.opaque, elem->ret);
198 
199             /* We can safely cancel the completion_bh here regardless of someone
200              * else having scheduled it meanwhile because we reenter the
201              * completion function anyway (goto restart).
202              */
203             qemu_bh_cancel(pool->completion_bh);
204 
205             qemu_aio_unref(elem);
206             goto restart;
207         } else {
208             qemu_aio_unref(elem);
209         }
210     }
211 }
212 
213 static void thread_pool_cancel(BlockAIOCB *acb)
214 {
215     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
216     ThreadPool *pool = elem->pool;
217 
218     trace_thread_pool_cancel(elem, elem->common.opaque);
219 
220     QEMU_LOCK_GUARD(&pool->lock);
221     if (elem->state == THREAD_QUEUED) {
222         QTAILQ_REMOVE(&pool->request_list, elem, reqs);
223         qemu_bh_schedule(pool->completion_bh);
224 
225         elem->state = THREAD_DONE;
226         elem->ret = -ECANCELED;
227     }
228 
229 }
230 
231 static const AIOCBInfo thread_pool_aiocb_info = {
232     .aiocb_size         = sizeof(ThreadPoolElement),
233     .cancel_async       = thread_pool_cancel,
234 };
235 
236 BlockAIOCB *thread_pool_submit_aio(ThreadPoolFunc *func, void *arg,
237                                    BlockCompletionFunc *cb, void *opaque)
238 {
239     ThreadPoolElement *req;
240     AioContext *ctx = qemu_get_current_aio_context();
241     ThreadPool *pool = aio_get_thread_pool(ctx);
242 
243     /* Assert that the thread submitting work is the same running the pool */
244     assert(pool->ctx == qemu_get_current_aio_context());
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_cond_signal(&pool->request_cond);
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     aio_co_wake(co->co);
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 void thread_pool_update_params(ThreadPool *pool, AioContext *ctx)
294 {
295     qemu_mutex_lock(&pool->lock);
296 
297     pool->min_threads = ctx->thread_pool_min;
298     pool->max_threads = ctx->thread_pool_max;
299 
300     /*
301      * We either have to:
302      *  - Increase the number available of threads until over the min_threads
303      *    threshold.
304      *  - Bump the worker threads so that they exit, until under the max_threads
305      *    threshold.
306      *  - Do nothing. The current number of threads fall in between the min and
307      *    max thresholds. We'll let the pool manage itself.
308      */
309     for (int i = pool->cur_threads; i < pool->min_threads; i++) {
310         spawn_thread(pool);
311     }
312 
313     for (int i = pool->cur_threads; i > pool->max_threads; i--) {
314         qemu_cond_signal(&pool->request_cond);
315     }
316 
317     qemu_mutex_unlock(&pool->lock);
318 }
319 
320 static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx)
321 {
322     if (!ctx) {
323         ctx = qemu_get_aio_context();
324     }
325 
326     memset(pool, 0, sizeof(*pool));
327     pool->ctx = ctx;
328     pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool);
329     qemu_mutex_init(&pool->lock);
330     qemu_cond_init(&pool->worker_stopped);
331     qemu_cond_init(&pool->request_cond);
332     pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool);
333 
334     QLIST_INIT(&pool->head);
335     QTAILQ_INIT(&pool->request_list);
336 
337     thread_pool_update_params(pool, ctx);
338 }
339 
340 ThreadPool *thread_pool_new(AioContext *ctx)
341 {
342     ThreadPool *pool = g_new(ThreadPool, 1);
343     thread_pool_init_one(pool, ctx);
344     return pool;
345 }
346 
347 void thread_pool_free(ThreadPool *pool)
348 {
349     if (!pool) {
350         return;
351     }
352 
353     assert(QLIST_EMPTY(&pool->head));
354 
355     qemu_mutex_lock(&pool->lock);
356 
357     /* Stop new threads from spawning */
358     qemu_bh_delete(pool->new_thread_bh);
359     pool->cur_threads -= pool->new_threads;
360     pool->new_threads = 0;
361 
362     /* Wait for worker threads to terminate */
363     pool->max_threads = 0;
364     qemu_cond_broadcast(&pool->request_cond);
365     while (pool->cur_threads > 0) {
366         qemu_cond_wait(&pool->worker_stopped, &pool->lock);
367     }
368 
369     qemu_mutex_unlock(&pool->lock);
370 
371     qemu_bh_delete(pool->completion_bh);
372     qemu_cond_destroy(&pool->request_cond);
373     qemu_cond_destroy(&pool->worker_stopped);
374     qemu_mutex_destroy(&pool->lock);
375     g_free(pool);
376 }
377