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/defer-call.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     /* This list is only written by the thread pool's mother thread.  */
  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     QemuCond request_cond;
  62     QEMUBH *new_thread_bh;
  63
  64     /* The following variables are only accessed from one AioContext. */
  65     QLIST_HEAD(, ThreadPoolElement) head;
  66
  67     /* The following variables are protected by lock.  */
  68     QTAILQ_HEAD(, ThreadPoolElement) request_list;
  69     int cur_threads;
  70     int idle_threads;
  71     int new_threads;     /* backlog of threads we need to create */
  72     int pending_threads; /* threads created but not running yet */
  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->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
 125     /*
 126      * Wake up another thread, in case we got a wakeup but decided
 127      * to exit due to pool->cur_threads > pool->max_threads.
 128      */
 129     qemu_cond_signal(&pool->request_cond);
 130     qemu_mutex_unlock(&pool->lock);
 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     defer_call_begin(); /* cb() may use defer_call() to coalesce work */
 180
 181 restart:
 182     QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
 183         if (elem->state != THREAD_DONE) {
 184             continue;
 185         }
 186
 187         trace_thread_pool_complete(pool, elem, elem->common.opaque,
 188                                    elem->ret);
 189         QLIST_REMOVE(elem, all);
 190
 191         if (elem->common.cb) {
 192             /* Read state before ret.  */
 193             smp_rmb();
 194
 195             /* Schedule ourselves in case elem->common.cb() calls aio_poll() to
 196              * wait for another request that completed at the same time.
 197              */
 198             qemu_bh_schedule(pool->completion_bh);
 199
 200             elem->common.cb(elem->common.opaque, elem->ret);
 201
 202             /* We can safely cancel the completion_bh here regardless of someone
 203              * else having scheduled it meanwhile because we reenter the
 204              * completion function anyway (goto restart).
 205              */
 206             qemu_bh_cancel(pool->completion_bh);
 207
 208             qemu_aio_unref(elem);
 209             goto restart;
 210         } else {
 211             qemu_aio_unref(elem);
 212         }
 213     }
 214
 215     defer_call_end();
 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 const AIOCBInfo thread_pool_aiocb_info = {
 237     .aiocb_size         = sizeof(ThreadPoolElement),
 238     .cancel_async       = thread_pool_cancel,
 239 };
 240
 241 BlockAIOCB *thread_pool_submit_aio(ThreadPoolFunc *func, void *arg,
 242                                    BlockCompletionFunc *cb, void *opaque)
 243 {
 244     ThreadPoolElement *req;
 245     AioContext *ctx = qemu_get_current_aio_context();
 246     ThreadPool *pool = aio_get_thread_pool(ctx);
 247
 248     /* Assert that the thread submitting work is the same running the pool */
 249     assert(pool->ctx == qemu_get_current_aio_context());
 250
 251     req = qemu_aio_get(&thread_pool_aiocb_info, NULL, cb, opaque);
 252     req->func = func;
 253     req->arg = arg;
 254     req->state = THREAD_QUEUED;
 255     req->pool = pool;
 256
 257     QLIST_INSERT_HEAD(&pool->head, req, all);
 258
 259     trace_thread_pool_submit(pool, req, arg);
 260
 261     qemu_mutex_lock(&pool->lock);
 262     if (pool->idle_threads == 0 && pool->cur_threads < pool->max_threads) {
 263         spawn_thread(pool);
 264     }
 265     QTAILQ_INSERT_TAIL(&pool->request_list, req, reqs);
 266     qemu_mutex_unlock(&pool->lock);
 267     qemu_cond_signal(&pool->request_cond);
 268     return &req->common;
 269 }
 270
 271 typedef struct ThreadPoolCo {
 272     Coroutine *co;
 273     int ret;
 274 } ThreadPoolCo;
 275
 276 static void thread_pool_co_cb(void *opaque, int ret)
 277 {
 278     ThreadPoolCo *co = opaque;
 279
 280     co->ret = ret;
 281     aio_co_wake(co->co);
 282 }
 283
 284 int coroutine_fn thread_pool_submit_co(ThreadPoolFunc *func, void *arg)
 285 {
 286     ThreadPoolCo tpc = { .co = qemu_coroutine_self(), .ret = -EINPROGRESS };
 287     assert(qemu_in_coroutine());
 288     thread_pool_submit_aio(func, arg, thread_pool_co_cb, &tpc);
 289     qemu_coroutine_yield();
 290     return tpc.ret;
 291 }
 292
 293 void thread_pool_submit(ThreadPoolFunc *func, void *arg)
 294 {
 295     thread_pool_submit_aio(func, arg, NULL, NULL);
 296 }
 297
 298 void thread_pool_update_params(ThreadPool *pool, AioContext *ctx)
 299 {
 300     qemu_mutex_lock(&pool->lock);
 301
 302     pool->min_threads = ctx->thread_pool_min;
 303     pool->max_threads = ctx->thread_pool_max;
 304
 305     /*
 306      * We either have to:
 307      *  - Increase the number available of threads until over the min_threads
 308      *    threshold.
 309      *  - Bump the worker threads so that they exit, until under the max_threads
 310      *    threshold.
 311      *  - Do nothing. The current number of threads fall in between the min and
 312      *    max thresholds. We'll let the pool manage itself.
 313      */
 314     for (int i = pool->cur_threads; i < pool->min_threads; i++) {
 315         spawn_thread(pool);
 316     }
 317
 318     for (int i = pool->cur_threads; i > pool->max_threads; i--) {
 319         qemu_cond_signal(&pool->request_cond);
 320     }
 321
 322     qemu_mutex_unlock(&pool->lock);
 323 }
 324
 325 static void thread_pool_init_one(ThreadPool *pool, AioContext *ctx)
 326 {
 327     if (!ctx) {
 328         ctx = qemu_get_aio_context();
 329     }
 330
 331     memset(pool, 0, sizeof(*pool));
 332     pool->ctx = ctx;
 333     pool->completion_bh = aio_bh_new(ctx, thread_pool_completion_bh, pool);
 334     qemu_mutex_init(&pool->lock);
 335     qemu_cond_init(&pool->worker_stopped);
 336     qemu_cond_init(&pool->request_cond);
 337     pool->new_thread_bh = aio_bh_new(ctx, spawn_thread_bh_fn, pool);
 338
 339     QLIST_INIT(&pool->head);
 340     QTAILQ_INIT(&pool->request_list);
 341
 342     thread_pool_update_params(pool, ctx);
 343 }
 344
 345 ThreadPool *thread_pool_new(AioContext *ctx)
 346 {
 347     ThreadPool *pool = g_new(ThreadPool, 1);
 348     thread_pool_init_one(pool, ctx);
 349     return pool;
 350 }
 351
 352 void thread_pool_free(ThreadPool *pool)
 353 {
 354     if (!pool) {
 355         return;
 356     }
 357
 358     assert(QLIST_EMPTY(&pool->head));
 359
 360     qemu_mutex_lock(&pool->lock);
 361
 362     /* Stop new threads from spawning */
 363     qemu_bh_delete(pool->new_thread_bh);
 364     pool->cur_threads -= pool->new_threads;
 365     pool->new_threads = 0;
 366
 367     /* Wait for worker threads to terminate */
 368     pool->max_threads = 0;
 369     qemu_cond_broadcast(&pool->request_cond);
 370     while (pool->cur_threads > 0) {
 371         qemu_cond_wait(&pool->worker_stopped, &pool->lock);
 372     }
 373
 374     qemu_mutex_unlock(&pool->lock);
 375
 376     qemu_bh_delete(pool->completion_bh);
 377     qemu_cond_destroy(&pool->request_cond);
 378     qemu_cond_destroy(&pool->worker_stopped);
 379     qemu_mutex_destroy(&pool->lock);
 380     g_free(pool);
 381 }