util/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/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     /* Access to this list is protected by the global mutex.  */
  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     bool stopping;
  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->stopping && 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     qemu_mutex_unlock(&pool->lock);
 125
 126     /*
 127      * Wake up another thread, in case we got a wakeup but decided
 128      * to exit due to pool->cur_threads > pool->max_threads.
 129      */
 130     qemu_cond_signal(&pool->request_cond);
 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     aio_context_acquire(pool->ctx);
 180 restart:
 181     QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
 182         if (elem->state != THREAD_DONE) {
 183             continue;
 184         }
 185
 186         trace_thread_pool_complete(pool, elem, elem->common.opaque,
 187                                    elem->ret);
 188         QLIST_REMOVE(elem, all);
 189
 190         if (elem->common.cb) {
 191             /* Read state before ret.  */
 192             smp_rmb();
 193
 194             /* Schedule ourselves in case elem->common.cb() calls aio_poll() to
 195              * wait for another request that completed at the same time.
 196              */
 197             qemu_bh_schedule(pool->completion_bh);
 198
 199             aio_context_release(pool->ctx);
 200             elem->common.cb(elem->common.opaque, elem->ret);
 201             aio_context_acquire(pool->ctx);
 202
 203             /* We can safely cancel the completion_bh here regardless of someone
 204              * else having scheduled it meanwhile because we reenter the
 205              * completion function anyway (goto restart).
 206              */
 207             qemu_bh_cancel(pool->completion_bh);
 208
 209             qemu_aio_unref(elem);
 210             goto restart;
 211         } else {
 212             qemu_aio_unref(elem);
 213         }
 214     }
 215     aio_context_release(pool->ctx);
 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 AioContext *thread_pool_get_aio_context(BlockAIOCB *acb)
 237 {
 238     ThreadPoolElement *elem = (ThreadPoolElement *)acb;
 239     ThreadPool *pool = elem->pool;
 240     return pool->ctx;
 241 }
 242
 243 static const AIOCBInfo thread_pool_aiocb_info = {
 244     .aiocb_size         = sizeof(ThreadPoolElement),
 245     .cancel_async       = thread_pool_cancel,
 246     .get_aio_context    = thread_pool_get_aio_context,
 247 };
 248
 249 BlockAIOCB *thread_pool_submit_aio(ThreadPool *pool,
 250         ThreadPoolFunc *func, void *arg,
 251         BlockCompletionFunc *cb, void *opaque)
 252 {
 253     ThreadPoolElement *req;
 254
 255     req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque);
 256     req->func = func;
 257     req->arg = arg;
 258     req->state = THREAD_QUEUED;
 259     req->pool = pool;
 260
 261     QLIST_INSERT_HEAD(&pool->head, req, all);
 262
 263     trace_thread_pool_submit(pool, req, arg);
 264
 265     qemu_mutex_lock(&pool->lock);
 266     if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) {
 267         spawn_thread(pool);
 268     }
 269     QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs);
 270     qemu_mutex_unlock(&pool->lock);
 271     qemu_cond_signal(&pool->request_cond);
 272     return &req->common;
 273 }
 274
 275 typedef struct ThreadPoolCo {
 276     Coroutine *co;
 277     int ret;
 278 } ThreadPoolCo;
 279
 280 static void thread_pool_co_cb(void *opaque, int ret)
 281 {
 282     ThreadPoolCo *co = opaque;
 283
 284     co->ret = ret;
 285     aio_co_wake(co->co);
 286 }
 287
 288 int coroutine_fn thread_pool_submit_co(ThreadPool *pool, ThreadPoolFunc *func,
 289                                        void *arg)
 290 {
 291     ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
 292     assert(qemu_in_coroutine());
 293     thread_pool_submit_aio(pool, func, arg, thread_pool_co_cb, &tpc);
 294     qemu_coroutine_yield();
 295     return tpc.ret;
 296 }
 297
 298 void thread_pool_submit(ThreadPool *pool, ThreadPoolFunc *func, void *arg)
 299 {
 300     thread_pool_submit_aio(pool, func, arg, NULL, NULL);
 301 }
 302
 303 void thread_pool_update_params(ThreadPool *pool, AioContext *ctx)
 304 {
 305     qemu_mutex_lock(&pool->lock);
 306
 307     pool->min_threads = ctx->thread_pool_min;
 308     pool->max_threads = ctx->thread_pool_max;
 309
 310     /*
 311      * We either have to:
 312      *  - Increase the number available of threads until over the min_threads
 313      *    threshold.
 314      *  - Bump the worker threads so that they exit, until under the max_threads
 315      *    threshold.
 316      *  - Do nothing. The current number of threads fall in between the min and
 317      *    max thresholds. We'll let the pool manage itself.
 318      */
 319     for (int i = pool->cur_threads; i < pool->min_threads; i++) {
 320         spawn_thread(pool);
 321     }
 322
 323     for (int i = pool->cur_threads; i > pool->max_threads; i--) {
 324         qemu_cond_signal(&pool->request_cond);
 325     }
 326
 327     qemu_mutex_unlock(&pool->lock);
 328 }
 329
 330 static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx)
 331 {
 332     if (!ctx) {
 333         ctx = qemu_get_aio_context();
 334     }
 335
 336     memset(pool, 0, sizeof(*pool));
 337     pool->ctx = ctx;
 338     pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool);
 339     qemu_mutex_init(&pool->lock);
 340     qemu_cond_init(&pool->worker_stopped);
 341     qemu_cond_init(&pool->request_cond);
 342     pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool);
 343
 344     QLIST_INIT(&pool->head);
 345     QTAILQ_INIT(&pool->request_list);
 346
 347     thread_pool_update_params(pool, ctx);
 348 }
 349
 350 ThreadPool *thread_pool_new(AioContext *ctx)
 351 {
 352     ThreadPool *pool = g_new(ThreadPool, 1);
 353     thread_pool_init_one(pool, ctx);
 354     return pool;
 355 }
 356
 357 void thread_pool_free(ThreadPool *pool)
 358 {
 359     if (!pool) {
 360         return;
 361     }
 362
 363     assert(QLIST_EMPTY(&pool->head));
 364
 365     qemu_mutex_lock(&pool->lock);
 366
 367     /* Stop new threads from spawning */
 368     qemu_bh_delete(pool->new_thread_bh);
 369     pool->cur_threads -= pool->new_threads;
 370     pool->new_threads = 0;
 371
 372     /* Wait for worker threads to terminate */
 373     pool->stopping = true;
 374     qemu_cond_broadcast(&pool->request_cond);
 375     while (pool->cur_threads > 0) {
 376         qemu_cond_wait(&pool->worker_stopped, &pool->lock);
 377     }
 378
 379     qemu_mutex_unlock(&pool->lock);
 380
 381     qemu_bh_delete(pool->completion_bh);
 382     qemu_cond_destroy(&pool->request_cond);
 383     qemu_cond_destroy(&pool->worker_stopped);
 384     qemu_mutex_destroy(&pool->lock);
 385     g_free(pool);
 386 }