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-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
  43     /* Moving state out of THREAD_QUEUED is protected by lock.  After
  44      * that, only the worker thread can write to it.  Reads and writes
  45      * of state and ret are ordered with memory barriers.
  46      */
  47     enum ThreadState state;
  48     int ret;
  49
  50     /* Access to this list is protected by lock.  */
  51     QTAILQ_ENTRY(ThreadPoolElement) reqs;
  52
  53     /* Access to this list is protected by the global mutex.  */
  54     QLIST_ENTRY(ThreadPoolElement) all;
  55 };
  56
  57 static EventNotifier notifier;
  58 static QemuMutex lock;
  59 static QemuCond check_cancel;
  60 static QemuSemaphore sem;
  61 static int max_threads = 64;
  62 static QEMUBH *new_thread_bh;
  63
  64 /* The following variables are protected by the global mutex.  */
  65 static QLIST_HEAD(, ThreadPoolElement) head;
  66
  67 /* The following variables are protected by lock.  */
  68 static QTAILQ_HEAD(, ThreadPoolElement) request_list;
  69 static int cur_threads;
  70 static int idle_threads;
  71 static int new_threads;     /* backlog of threads we need to create */
  72 static int pending_threads; /* threads created but not running yet */
  73 static int pending_cancellations; /* whether we need a cond_broadcast */
  74
  75 static void *worker_thread(void *unused)
  76 {
  77     qemu_mutex_lock(&lock);
  78     pending_threads--;
  79     do_spawn_thread();
  80
  81     while (1) {
  82         ThreadPoolElement *req;
  83         int ret;
  84
  85         do {
  86             idle_threads++;
  87             qemu_mutex_unlock(&lock);
  88             ret = qemu_sem_timedwait(&sem, 10000);
  89             qemu_mutex_lock(&lock);
  90             idle_threads--;
  91         } while (ret == -1 && !QTAILQ_EMPTY(&request_list));
  92         if (ret == -1) {
  93             break;
  94         }
  95
  96         req = QTAILQ_FIRST(&request_list);
  97         QTAILQ_REMOVE(&request_list, req, reqs);
  98         req->state = THREAD_ACTIVE;
  99         qemu_mutex_unlock(&lock);
 100
 101         ret = req->func(req->arg);
 102
 103         req->ret = ret;
 104         /* Write ret before state.  */
 105         smp_wmb();
 106         req->state = THREAD_DONE;
 107
 108         qemu_mutex_lock(&lock);
 109         if (pending_cancellations) {
 110             qemu_cond_broadcast(&check_cancel);
 111         }
 112
 113         event_notifier_set(&notifier);
 114     }
 115
 116     cur_threads--;
 117     qemu_mutex_unlock(&lock);
 118     return NULL;
 119 }
 120
 121 static void do_spawn_thread(void)
 122 {
 123     QemuThread t;
 124
 125     /* Runs with lock taken.  */
 126     if (!new_threads) {
 127         return;
 128     }
 129
 130     new_threads--;
 131     pending_threads++;
 132
 133     qemu_thread_create(&t, worker_thread, NULL, QEMU_THREAD_DETACHED);
 134 }
 135
 136 static void spawn_thread_bh_fn(void *opaque)
 137 {
 138     qemu_mutex_lock(&lock);
 139     do_spawn_thread();
 140     qemu_mutex_unlock(&lock);
 141 }
 142
 143 static void spawn_thread(void)
 144 {
 145     cur_threads++;
 146     new_threads++;
 147     /* If there are threads being created, they will spawn new workers, so
 148      * we don't spend time creating many threads in a loop holding a mutex or
 149      * starving the current vcpu.
 150      *
 151      * If there are no idle threads, ask the main thread to create one, so we
 152      * inherit the correct affinity instead of the vcpu affinity.
 153      */
 154     if (!pending_threads) {
 155         qemu_bh_schedule(new_thread_bh);
 156     }
 157 }
 158
 159 static void event_notifier_ready(EventNotifier *notifier)
 160 {
 161     ThreadPoolElement *elem, *next;
 162
 163     event_notifier_test_and_clear(notifier);
 164 restart:
 165     QLIST_FOREACH_SAFE(elem, &head, all, next) {
 166         if (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) {
 167             continue;
 168         }
 169         if (elem->state == THREAD_DONE) {
 170             trace_thread_pool_complete(elem, elem->common.opaque, elem->ret);
 171         }
 172         if (elem->state == THREAD_DONE && elem->common.cb) {
 173             QLIST_REMOVE(elem, all);
 174             /* Read state before ret.  */
 175             smp_rmb();
 176             elem->common.cb(elem->common.opaque, elem->ret);
 177             qemu_aio_release(elem);
 178             goto restart;
 179         } else {
 180             /* remove the request */
 181             QLIST_REMOVE(elem, all);
 182             qemu_aio_release(elem);
 183         }
 184     }
 185 }
 186
 187 static int thread_pool_active(EventNotifier *notifier)
 188 {
 189     return !QLIST_EMPTY(&head);
 190 }
 191
 192 static void thread_pool_cancel(BlockDriverAIOCB *acb)
 193 {
 194     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
 195
 196     trace_thread_pool_cancel(elem, elem->common.opaque);
 197
 198     qemu_mutex_lock(&lock);
 199     if (elem->state == THREAD_QUEUED &&
 200         /* No thread has yet started working on elem. we can try to "steal"
 201          * the item from the worker if we can get a signal from the
 202          * semaphore.  Because this is non-blocking, we can do it with
 203          * the lock taken and ensure that elem will remain THREAD_QUEUED.
 204          */
 205         qemu_sem_timedwait(&sem, 0) == 0) {
 206         QTAILQ_REMOVE(&request_list, elem, reqs);
 207         elem->state = THREAD_CANCELED;
 208         event_notifier_set(&notifier);
 209     } else {
 210         pending_cancellations++;
 211         while (elem->state != THREAD_CANCELED && elem->state != THREAD_DONE) {
 212             qemu_cond_wait(&check_cancel, &lock);
 213         }
 214         pending_cancellations--;
 215     }
 216     qemu_mutex_unlock(&lock);
 217 }
 218
 219 static AIOPool thread_pool_cb_pool = {
 220     .aiocb_size         = sizeof(ThreadPoolElement),
 221     .cancel             = thread_pool_cancel,
 222 };
 223
 224 BlockDriverAIOCB *thread_pool_submit_aio(ThreadPoolFunc *func, void *arg,
 225         BlockDriverCompletionFunc *cb, void *opaque)
 226 {
 227     ThreadPoolElement *req;
 228
 229     req = qemu_aio_get(&thread_pool_cb_pool, NULL, cb, opaque);
 230     req->func = func;
 231     req->arg = arg;
 232     req->state = THREAD_QUEUED;
 233
 234     QLIST_INSERT_HEAD(&head, req, all);
 235
 236     trace_thread_pool_submit(req, arg);
 237
 238     qemu_mutex_lock(&lock);
 239     if (idle_threads == 0 && cur_threads < max_threads) {
 240         spawn_thread();
 241     }
 242     QTAILQ_INSERT_TAIL(&request_list, req, reqs);
 243     qemu_mutex_unlock(&lock);
 244     qemu_sem_post(&sem);
 245     return &req->common;
 246 }
 247
 248 typedef struct ThreadPoolCo {
 249     Coroutine *co;
 250     int ret;
 251 } ThreadPoolCo;
 252
 253 static void thread_pool_co_cb(void *opaque, int ret)
 254 {
 255     ThreadPoolCo *co = opaque;
 256
 257     co->ret = ret;
 258     qemu_coroutine_enter(co->co, NULL);
 259 }
 260
 261 int coroutine_fn thread_pool_submit_co(ThreadPoolFunc *func, void *arg)
 262 {
 263     ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
 264     assert(qemu_in_coroutine());
 265     thread_pool_submit_aio(func, arg, thread_pool_co_cb, &tpc);
 266     qemu_coroutine_yield();
 267     return tpc.ret;
 268 }
 269
 270 void thread_pool_submit(ThreadPoolFunc *func, void *arg)
 271 {
 272     thread_pool_submit_aio(func, arg, NULL, NULL);
 273 }
 274
 275 static void thread_pool_init(void)
 276 {
 277     QLIST_INIT(&head);
 278     event_notifier_init(&notifier, false);
 279     qemu_mutex_init(&lock);
 280     qemu_cond_init(&check_cancel);
 281     qemu_sem_init(&sem, 0);
 282     qemu_aio_set_event_notifier(&notifier, event_notifier_ready,
 283                                 thread_pool_active);
 284
 285     QTAILQ_INIT(&request_list);
 286     new_thread_bh = qemu_bh_new(spawn_thread_bh_fn, NULL);
 287 }
 288
 289 block_init(thread_pool_init)