thread-pool.c

   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-common.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     /* Access to this list is protected by the global mutex.  */
  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     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 */
  74     bool stopping;
  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->stopping) {
  86         ThreadPoolElement *req;
  87         int ret;
  88
  89         do {
  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));
  96         if (ret == -1 || pool->stopping) {
  97             break;
  98         }
  99
 100         req = QTAILQ_FIRST(&pool->request_list);
 101         QTAILQ_REMOVE(&pool->request_list, req, reqs);
 102         req->state = THREAD_ACTIVE;
 103         qemu_mutex_unlock(&pool->lock);
 104
 105         ret = req->func(req->arg);
 106
 107         req->ret = ret;
 108         /* Write ret before state.  */
 109         smp_wmb();
 110         req->state = THREAD_DONE;
 111
 112         qemu_mutex_lock(&pool->lock);
 113
 114         qemu_bh_schedule(pool->completion_bh);
 115     }
 116
 117     pool->cur_threads--;
 118     qemu_cond_signal(&pool->worker_stopped);
 119     qemu_mutex_unlock(&pool->lock);
 120     return NULL;
 121 }
 122
 123 static void do_spawn_thread(ThreadPool *pool)
 124 {
 125     QemuThread t;
 126
 127     /* Runs with lock taken.  */
 128     if (!pool->new_threads) {
 129         return;
 130     }
 131
 132     pool->new_threads--;
 133     pool->pending_threads++;
 134
 135     qemu_thread_create(&t, "worker", worker_thread, pool, QEMU_THREAD_DETACHED);
 136 }
 137
 138 static void spawn_thread_bh_fn(void *opaque)
 139 {
 140     ThreadPool *pool = opaque;
 141
 142     qemu_mutex_lock(&pool->lock);
 143     do_spawn_thread(pool);
 144     qemu_mutex_unlock(&pool->lock);
 145 }
 146
 147 static void spawn_thread(ThreadPool *pool)
 148 {
 149     pool->cur_threads++;
 150     pool->new_threads++;
 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      */
 158     if (!pool->pending_threads) {
 159         qemu_bh_schedule(pool->new_thread_bh);
 160     }
 161 }
 162
 163 static void thread_pool_completion_bh(void *opaque)
 164 {
 165     ThreadPool *pool = opaque;
 166     ThreadPoolElement *elem, *next;
 167
 168 restart:
 169     QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
 170         if (elem->state != THREAD_DONE) {
 171             continue;
 172         }
 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) {
 179             /* Read state before ret.  */
 180             smp_rmb();
 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
 187             elem->common.cb(elem->common.opaque, elem->ret);
 188
 189             /* We can safely cancel the completion_bh here regardless of someone
 190              * else having scheduled it meanwhile because we reenter the
 191              * completion function anyway (goto restart).
 192              */
 193             qemu_bh_cancel(pool->completion_bh);
 194
 195             qemu_aio_unref(elem);
 196             goto restart;
 197         } else {
 198             qemu_aio_unref(elem);
 199         }
 200     }
 201 }
 202
 203 static void thread_pool_cancel(BlockAIOCB *acb)
 204 {
 205     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
 206     ThreadPool *pool = elem->pool;
 207
 208     trace_thread_pool_cancel(elem, elem->common.opaque);
 209
 210     qemu_mutex_lock(&pool->lock);
 211     if (elem->state == THREAD_QUEUED &&
 212         /* No thread has yet started working on elem. we can try to "steal"
 213          * the item from the worker if we can get a signal from the
 214          * semaphore.  Because this is non-blocking, we can do it with
 215          * the lock taken and ensure that elem will remain THREAD_QUEUED.
 216          */
 217         qemu_sem_timedwait(&pool->sem, 0) == 0) {
 218         QTAILQ_REMOVE(&pool->request_list, elem, reqs);
 219         qemu_bh_schedule(pool->completion_bh);
 220
 221         elem->state = THREAD_DONE;
 222         elem->ret = -ECANCELED;
 223     }
 224
 225     qemu_mutex_unlock(&pool->lock);
 226 }
 227
 228 static AioContext *thread_pool_get_aio_context(BlockAIOCB *acb)
 229 {
 230     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
 231     ThreadPool *pool = elem->pool;
 232     return pool->ctx;
 233 }
 234
 235 static const AIOCBInfo thread_pool_aiocb_info = {
 236     .aiocb_size         = sizeof(ThreadPoolElement),
 237     .cancel_async       = thread_pool_cancel,
 238     .get_aio_context    = thread_pool_get_aio_context,
 239 };
 240
 241 BlockAIOCB *thread_pool_submit_aio(ThreadPool *pool,
 242         ThreadPoolFunc *func, void *arg,
 243         BlockCompletionFunc *cb, void *opaque)
 244 {
 245     ThreadPoolElement *req;
 246
 247     req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque);
 248     req->func = func;
 249     req->arg = arg;
 250     req->state = THREAD_QUEUED;
 251     req->pool = pool;
 252
 253     QLIST_INSERT_HEAD(&pool->head, req, all);
 254
 255     trace_thread_pool_submit(pool, req, arg);
 256
 257     qemu_mutex_lock(&pool->lock);
 258     if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) {
 259         spawn_thread(pool);
 260     }
 261     QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs);
 262     qemu_mutex_unlock(&pool->lock);
 263     qemu_sem_post(&pool->sem);
 264     return &req->common;
 265 }
 266
 267 typedef struct ThreadPoolCo {
 268     Coroutine *co;
 269     int ret;
 270 } ThreadPoolCo;
 271
 272 static void thread_pool_co_cb(void *opaque, int ret)
 273 {
 274     ThreadPoolCo *co = opaque;
 275
 276     co->ret = ret;
 277     qemu_coroutine_enter(co->co);
 278 }
 279
 280 int coroutine_fn thread_pool_submit_co(ThreadPool *pool, ThreadPoolFunc *func,
 281                                        void *arg)
 282 {
 283     ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
 284     assert(qemu_in_coroutine());
 285     thread_pool_submit_aio(pool, func, arg, thread_pool_co_cb, &tpc);
 286     qemu_coroutine_yield();
 287     return tpc.ret;
 288 }
 289
 290 void thread_pool_submit(ThreadPool *pool, ThreadPoolFunc *func, void *arg)
 291 {
 292     thread_pool_submit_aio(pool, func, arg, NULL, NULL);
 293 }
 294
 295 static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx)
 296 {
 297     if (!ctx) {
 298         ctx = qemu_get_aio_context();
 299     }
 300
 301     memset(pool, 0, sizeof(*pool));
 302     pool->ctx = ctx;
 303     pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool);
 304     qemu_mutex_init(&pool->lock);
 305     qemu_cond_init(&pool->worker_stopped);
 306     qemu_sem_init(&pool->sem, 0);
 307     pool->max_threads = 64;
 308     pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool);
 309
 310     QLIST_INIT(&pool->head);
 311     QTAILQ_INIT(&pool->request_list);
 312 }
 313
 314 ThreadPool *thread_pool_new(AioContext *ctx)
 315 {
 316     ThreadPool *pool = g_new(ThreadPool, 1);
 317     thread_pool_init_one(pool, ctx);
 318     return pool;
 319 }
 320
 321 void thread_pool_free(ThreadPool *pool)
 322 {
 323     if (!pool) {
 324         return;
 325     }
 326
 327     assert(QLIST_EMPTY(&pool->head));
 328
 329     qemu_mutex_lock(&pool->lock);
 330
 331     /* Stop new threads from spawning */
 332     qemu_bh_delete(pool->new_thread_bh);
 333     pool->cur_threads -= pool->new_threads;
 334     pool->new_threads = 0;
 335
 336     /* Wait for worker threads to terminate */
 337     pool->stopping = true;
 338     while (pool->cur_threads > 0) {
 339         qemu_sem_post(&pool->sem);
 340         qemu_cond_wait(&pool->worker_stopped, &pool->lock);
 341     }
 342
 343     qemu_mutex_unlock(&pool->lock);
 344
 345     qemu_bh_delete(pool->completion_bh);
 346     qemu_sem_destroy(&pool->sem);
 347     qemu_cond_destroy(&pool->worker_stopped);
 348     qemu_mutex_destroy(&pool->lock);
 349     g_free(pool);
 350 }