|
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"
|
|
18 |
#include "qemu-queue.h"
|
|
19 |
#include "qemu-thread.h"
|
|
20 |
#include "osdep.h"
|
|
21 |
#include "qemu-coroutine.h"
|
|
22 |
#include "trace.h"
|
|
23 |
#include "block_int.h"
|
|
24 |
#include "event_notifier.h"
|
|
25 |
#include "thread-pool.h"
|
|
26 |
|
|
27 |
static void do_spawn_thread(void);
|
|
28 |
|
|
29 |
typedef struct ThreadPoolElement ThreadPoolElement;
|
|
30 |
|
|
31 |
enum ThreadState {
|
|
32 |
THREAD_QUEUED,
|
|
33 |
THREAD_ACTIVE,
|
|
34 |
THREAD_DONE,
|
|
35 |
THREAD_CANCELED,
|
|
36 |
};
|
|
37 |
|
|
38 |
struct ThreadPoolElement {
|
|
39 |
BlockDriverAIOCB common;
|
|
40 |
ThreadPoolFunc *func;
|
|
41 |
void *arg;
|
|
42 |
enum ThreadState state;
|
|
43 |
int ret;
|
|
44 |
|
|
45 |
/* Access to this list is protected by lock. */
|
|
46 |
QTAILQ_ENTRY(ThreadPoolElement) reqs;
|
|
47 |
|
|
48 |
/* Access to this list is protected by the global mutex. */
|
|
49 |
QLIST_ENTRY(ThreadPoolElement) all;
|
|
50 |
};
|
|
51 |
|
|
52 |
static EventNotifier notifier;
|
|
53 |
static QemuMutex lock;
|
|
54 |
static QemuCond check_cancel;
|
|
55 |
static QemuSemaphore sem;
|
|
56 |
static int max_threads = 64;
|
|
57 |
static QEMUBH *new_thread_bh;
|
|
58 |
|
|
59 |
/* The following variables are protected by the global mutex. */
|
|
60 |
static QLIST_HEAD(, ThreadPoolElement) head;
|
|
61 |
|
|
62 |
/* The following variables are protected by lock. */
|
|
63 |
static QTAILQ_HEAD(, ThreadPoolElement) request_list;
|
|
64 |
static int cur_threads;
|
|
65 |
static int idle_threads;
|
|
66 |
static int new_threads; /* backlog of threads we need to create */
|
|
67 |
static int pending_threads; /* threads created but not running yet */
|
|
68 |
static int pending_cancellations; /* whether we need a cond_broadcast */
|
|
69 |
|
|
70 |
static void *worker_thread(void *unused)
|
|
71 |
{
|
|
72 |
qemu_mutex_lock(&lock);
|
|
73 |
pending_threads--;
|
|
74 |
do_spawn_thread();
|
|
75 |
|
|
76 |
while (1) {
|
|
77 |
ThreadPoolElement *req;
|
|
78 |
int ret;
|
|
79 |
|
|
80 |
do {
|
|
81 |
idle_threads++;
|
|
82 |
qemu_mutex_unlock(&lock);
|
|
83 |
ret = qemu_sem_timedwait(&sem, 10000);
|
|
84 |
qemu_mutex_lock(&lock);
|
|
85 |
idle_threads--;
|
|
86 |
} while (ret == -1 && !QTAILQ_EMPTY(&request_list));
|
|
87 |
if (ret == -1) {
|
|
88 |
break;
|
|
89 |
}
|
|
90 |
|
|
91 |
req = QTAILQ_FIRST(&request_list);
|
|
92 |
QTAILQ_REMOVE(&request_list, req, reqs);
|
|
93 |
req->state = THREAD_ACTIVE;
|
|
94 |
qemu_mutex_unlock(&lock);
|
|
95 |
|
|
96 |
ret = req->func(req->arg);
|
|
97 |
|
|
98 |
qemu_mutex_lock(&lock);
|
|
99 |
req->state = THREAD_DONE;
|
|
100 |
req->ret = ret;
|
|
101 |
if (pending_cancellations) {
|
|
102 |
qemu_cond_broadcast(&check_cancel);
|
|
103 |
}
|
|
104 |
|
|
105 |
event_notifier_set(¬ifier);
|
|
106 |
}
|
|
107 |
|
|
108 |
cur_threads--;
|
|
109 |
qemu_mutex_unlock(&lock);
|
|
110 |
return NULL;
|
|
111 |
}
|
|
112 |
|
|
113 |
static void do_spawn_thread(void)
|
|
114 |
{
|
|
115 |
QemuThread t;
|
|
116 |
|
|
117 |
/* Runs with lock taken. */
|
|
118 |
if (!new_threads) {
|
|
119 |
return;
|
|
120 |
}
|
|
121 |
|
|
122 |
new_threads--;
|
|
123 |
pending_threads++;
|
|
124 |
|
|
125 |
qemu_thread_create(&t, worker_thread, NULL, QEMU_THREAD_DETACHED);
|
|
126 |
}
|
|
127 |
|
|
128 |
static void spawn_thread_bh_fn(void *opaque)
|
|
129 |
{
|
|
130 |
qemu_mutex_lock(&lock);
|
|
131 |
do_spawn_thread();
|
|
132 |
qemu_mutex_unlock(&lock);
|
|
133 |
}
|
|
134 |
|
|
135 |
static void spawn_thread(void)
|
|
136 |
{
|
|
137 |
cur_threads++;
|
|
138 |
new_threads++;
|
|
139 |
/* If there are threads being created, they will spawn new workers, so
|
|
140 |
* we don't spend time creating many threads in a loop holding a mutex or
|
|
141 |
* starving the current vcpu.
|
|
142 |
*
|
|
143 |
* If there are no idle threads, ask the main thread to create one, so we
|
|
144 |
* inherit the correct affinity instead of the vcpu affinity.
|
|
145 |
*/
|
|
146 |
if (!pending_threads) {
|
|
147 |
qemu_bh_schedule(new_thread_bh);
|
|
148 |
}
|
|
149 |
}
|
|
150 |
|
|
151 |
static void event_notifier_ready(EventNotifier *notifier)
|
|
152 |
{
|
|
153 |
ThreadPoolElement *elem, *next;
|
|
154 |
|
|
155 |
event_notifier_test_and_clear(notifier);
|
|
156 |
restart:
|
|
157 |
QLIST_FOREACH_SAFE(elem, &head, all, next) {
|
|
158 |
if (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) {
|
|
159 |
continue;
|
|
160 |
}
|
|
161 |
if (elem->state == THREAD_DONE) {
|
|
162 |
trace_thread_pool_complete(elem, elem->common.opaque, elem->ret);
|
|
163 |
}
|
|
164 |
if (elem->state == THREAD_DONE && elem->common.cb) {
|
|
165 |
qemu_mutex_lock(&lock);
|
|
166 |
int ret = elem->ret;
|
|
167 |
qemu_mutex_unlock(&lock);
|
|
168 |
QLIST_REMOVE(elem, all);
|
|
169 |
elem->common.cb(elem->common.opaque, ret);
|
|
170 |
qemu_aio_release(elem);
|
|
171 |
goto restart;
|
|
172 |
} else {
|
|
173 |
/* remove the request */
|
|
174 |
QLIST_REMOVE(elem, all);
|
|
175 |
qemu_aio_release(elem);
|
|
176 |
}
|
|
177 |
}
|
|
178 |
}
|
|
179 |
|
|
180 |
static int thread_pool_active(EventNotifier *notifier)
|
|
181 |
{
|
|
182 |
return !QLIST_EMPTY(&head);
|
|
183 |
}
|
|
184 |
|
|
185 |
static void thread_pool_cancel(BlockDriverAIOCB *acb)
|
|
186 |
{
|
|
187 |
ThreadPoolElement *elem = (ThreadPoolElement *)acb;
|
|
188 |
|
|
189 |
trace_thread_pool_cancel(elem, elem->common.opaque);
|
|
190 |
|
|
191 |
qemu_mutex_lock(&lock);
|
|
192 |
if (elem->state == THREAD_QUEUED &&
|
|
193 |
/* No thread has yet started working on elem. we can try to "steal"
|
|
194 |
* the item from the worker if we can get a signal from the
|
|
195 |
* semaphore. Because this is non-blocking, we can do it with
|
|
196 |
* the lock taken and ensure that elem will remain THREAD_QUEUED.
|
|
197 |
*/
|
|
198 |
qemu_sem_timedwait(&sem, 0) == 0) {
|
|
199 |
QTAILQ_REMOVE(&request_list, elem, reqs);
|
|
200 |
elem->state = THREAD_CANCELED;
|
|
201 |
event_notifier_set(¬ifier);
|
|
202 |
} else {
|
|
203 |
pending_cancellations++;
|
|
204 |
while (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) {
|
|
205 |
qemu_cond_wait(&check_cancel, &lock);
|
|
206 |
}
|
|
207 |
pending_cancellations--;
|
|
208 |
}
|
|
209 |
qemu_mutex_unlock(&lock);
|
|
210 |
}
|
|
211 |
|
|
212 |
static AIOPool thread_pool_cb_pool = {
|
|
213 |
.aiocb_size = sizeof(ThreadPoolElement),
|
|
214 |
.cancel = thread_pool_cancel,
|
|
215 |
};
|
|
216 |
|
|
217 |
BlockDriverAIOCB *thread_pool_submit_aio(ThreadPoolFunc *func, void *arg,
|
|
218 |
BlockDriverCompletionFunc *cb, void *opaque)
|
|
219 |
{
|
|
220 |
ThreadPoolElement *req;
|
|
221 |
|
|
222 |
req = qemu_aio_get(&thread_pool_cb_pool, NULL, cb, opaque);
|
|
223 |
req->func = func;
|
|
224 |
req->arg = arg;
|
|
225 |
req->state = THREAD_QUEUED;
|
|
226 |
|
|
227 |
QLIST_INSERT_HEAD(&head, req, all);
|
|
228 |
|
|
229 |
trace_thread_pool_submit(req, arg);
|
|
230 |
|
|
231 |
qemu_mutex_lock(&lock);
|
|
232 |
if (idle_threads == 0 && cur_threads < max_threads) {
|
|
233 |
spawn_thread();
|
|
234 |
}
|
|
235 |
QTAILQ_INSERT_TAIL(&request_list, req, reqs);
|
|
236 |
qemu_mutex_unlock(&lock);
|
|
237 |
qemu_sem_post(&sem);
|
|
238 |
return &req->common;
|
|
239 |
}
|
|
240 |
|
|
241 |
typedef struct ThreadPoolCo {
|
|
242 |
Coroutine *co;
|
|
243 |
int ret;
|
|
244 |
} ThreadPoolCo;
|
|
245 |
|
|
246 |
static void thread_pool_co_cb(void *opaque, int ret)
|
|
247 |
{
|
|
248 |
ThreadPoolCo *co = opaque;
|
|
249 |
|
|
250 |
co->ret = ret;
|
|
251 |
qemu_coroutine_enter(co->co, NULL);
|
|
252 |
}
|
|
253 |
|
|
254 |
int coroutine_fn thread_pool_submit_co(ThreadPoolFunc *func, void *arg)
|
|
255 |
{
|
|
256 |
ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
|
|
257 |
assert(qemu_in_coroutine());
|
|
258 |
thread_pool_submit_aio(func, arg, thread_pool_co_cb, &tpc);
|
|
259 |
qemu_coroutine_yield();
|
|
260 |
return tpc.ret;
|
|
261 |
}
|
|
262 |
|
|
263 |
void thread_pool_submit(ThreadPoolFunc *func, void *arg)
|
|
264 |
{
|
|
265 |
thread_pool_submit_aio(func, arg, NULL, NULL);
|
|
266 |
}
|
|
267 |
|
|
268 |
static void thread_pool_init(void)
|
|
269 |
{
|
|
270 |
QLIST_INIT(&head);
|
|
271 |
event_notifier_init(¬ifier, false);
|
|
272 |
qemu_mutex_init(&lock);
|
|
273 |
qemu_cond_init(&check_cancel);
|
|
274 |
qemu_sem_init(&sem, 0);
|
|
275 |
qemu_aio_set_event_notifier(¬ifier, event_notifier_ready,
|
|
276 |
thread_pool_active);
|
|
277 |
|
|
278 |
QTAILQ_INIT(&request_list);
|
|
279 |
new_thread_bh = qemu_bh_new(spawn_thread_bh_fn, NULL);
|
|
280 |
}
|
|
281 |
|
|
282 |
block_init(thread_pool_init)
|