async.c

   1 /*
   2  * QEMU System Emulator
   3  *
   4  * Copyright (c) 2003-2008 Fabrice Bellard
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24
  25 #include "qemu-common.h"
  26 #include "block/aio.h"
  27 #include "block/thread-pool.h"
  28 #include "qemu/main-loop.h"
  29 #include "qemu/atomic.h"
  30
  31 /***********************************************************/
  32 /* bottom halves (can be seen as timers which expire ASAP) */
  33
  34 struct QEMUBH {
  35     AioContext *ctx;
  36     QEMUBHFunc *cb;
  37     void *opaque;
  38     QEMUBH *next;
  39     bool scheduled;
  40     bool idle;
  41     bool deleted;
  42 };
  43
  44 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
  45 {
  46     QEMUBH *bh;
  47     bh = g_new(QEMUBH, 1);
  48     *bh = (QEMUBH){
  49         .ctx = ctx,
  50         .cb = cb,
  51         .opaque = opaque,
  52     };
  53     qemu_mutex_lock(&ctx->bh_lock);
  54     bh->next = ctx->first_bh;
  55     /* Make sure that the members are ready before putting bh into list */
  56     smp_wmb();
  57     ctx->first_bh = bh;
  58     qemu_mutex_unlock(&ctx->bh_lock);
  59     return bh;
  60 }
  61
  62 void aio_bh_call(QEMUBH *bh)
  63 {
  64     bh->cb(bh->opaque);
  65 }
  66
  67 /* Multiple occurrences of aio_bh_poll cannot be called concurrently */
  68 int aio_bh_poll(AioContext *ctx)
  69 {
  70     QEMUBH *bh, **bhp, *next;
  71     int ret;
  72
  73     ctx->walking_bh++;
  74
  75     ret = 0;
  76     for (bh = ctx->first_bh; bh; bh = next) {
  77         /* Make sure that fetching bh happens before accessing its members */
  78         smp_read_barrier_depends();
  79         next = bh->next;
  80         /* The atomic_xchg is paired with the one in qemu_bh_schedule.  The
  81          * implicit memory barrier ensures that the callback sees all writes
  82          * done by the scheduling thread.  It also ensures that the scheduling
  83          * thread sees the zero before bh->cb has run, and thus will call
  84          * aio_notify again if necessary.
  85          */
  86         if (!bh->deleted && atomic_xchg(&bh->scheduled, 0)) {
  87             /* Idle BHs and the notify BH don't count as progress */
  88             if (!bh->idle && bh != ctx->notify_dummy_bh) {
  89                 ret = 1;
  90             }
  91             bh->idle = 0;
  92             aio_bh_call(bh);
  93         }
  94     }
  95
  96     ctx->walking_bh--;
  97
  98     /* remove deleted bhs */
  99     if (!ctx->walking_bh) {
 100         qemu_mutex_lock(&ctx->bh_lock);
 101         bhp = &ctx->first_bh;
 102         while (*bhp) {
 103             bh = *bhp;
 104             if (bh->deleted) {
 105                 *bhp = bh->next;
 106                 g_free(bh);
 107             } else {
 108                 bhp = &bh->next;
 109             }
 110         }
 111         qemu_mutex_unlock(&ctx->bh_lock);
 112     }
 113
 114     return ret;
 115 }
 116
 117 void qemu_bh_schedule_idle(QEMUBH *bh)
 118 {
 119     bh->idle = 1;
 120     /* Make sure that idle & any writes needed by the callback are done
 121      * before the locations are read in the aio_bh_poll.
 122      */
 123     atomic_mb_set(&bh->scheduled, 1);
 124 }
 125
 126 void qemu_bh_schedule(QEMUBH *bh)
 127 {
 128     AioContext *ctx;
 129
 130     ctx = bh->ctx;
 131     bh->idle = 0;
 132     /* The memory barrier implicit in atomic_xchg makes sure that:
 133      * 1. idle & any writes needed by the callback are done before the
 134      *    locations are read in the aio_bh_poll.
 135      * 2. ctx is loaded before scheduled is set and the callback has a chance
 136      *    to execute.
 137      */
 138     if (atomic_xchg(&bh->scheduled, 1) == 0) {
 139         aio_notify(ctx);
 140     }
 141 }
 142
 143
 144 /* This func is async.
 145  */
 146 void qemu_bh_cancel(QEMUBH *bh)
 147 {
 148     bh->scheduled = 0;
 149 }
 150
 151 /* This func is async.The bottom half will do the delete action at the finial
 152  * end.
 153  */
 154 void qemu_bh_delete(QEMUBH *bh)
 155 {
 156     bh->scheduled = 0;
 157     bh->deleted = 1;
 158 }
 159
 160 int64_t
 161 aio_compute_timeout(AioContext *ctx)
 162 {
 163     int64_t deadline;
 164     int timeout = -1;
 165     QEMUBH *bh;
 166
 167     for (bh = ctx->first_bh; bh; bh = bh->next) {
 168         if (!bh->deleted && bh->scheduled) {
 169             if (bh->idle) {
 170                 /* idle bottom halves will be polled at least
 171                  * every 10ms */
 172                 timeout = 10000000;
 173             } else {
 174                 /* non-idle bottom halves will be executed
 175                  * immediately */
 176                 return 0;
 177             }
 178         }
 179     }
 180
 181     deadline = timerlistgroup_deadline_ns(&ctx->tlg);
 182     if (deadline == 0) {
 183         return 0;
 184     } else {
 185         return qemu_soonest_timeout(timeout, deadline);
 186     }
 187 }
 188
 189 static gboolean
 190 aio_ctx_prepare(GSource *source, gint    *timeout)
 191 {
 192     AioContext *ctx = (AioContext *) source;
 193
 194     atomic_or(&ctx->notify_me, 1);
 195
 196     /* We assume there is no timeout already supplied */
 197     *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
 198
 199     if (aio_prepare(ctx)) {
 200         *timeout = 0;
 201     }
 202
 203     return *timeout == 0;
 204 }
 205
 206 static gboolean
 207 aio_ctx_check(GSource *source)
 208 {
 209     AioContext *ctx = (AioContext *) source;
 210     QEMUBH *bh;
 211
 212     atomic_and(&ctx->notify_me, ~1);
 213     aio_notify_accept(ctx);
 214
 215     for (bh = ctx->first_bh; bh; bh = bh->next) {
 216         if (!bh->deleted && bh->scheduled) {
 217             return true;
 218         }
 219     }
 220     return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
 221 }
 222
 223 static gboolean
 224 aio_ctx_dispatch(GSource     *source,
 225                  GSourceFunc  callback,
 226                  gpointer     user_data)
 227 {
 228     AioContext *ctx = (AioContext *) source;
 229
 230     assert(callback == NULL);
 231     aio_dispatch(ctx);
 232     return true;
 233 }
 234
 235 static void
 236 aio_ctx_finalize(GSource     *source)
 237 {
 238     AioContext *ctx = (AioContext *) source;
 239
 240     qemu_bh_delete(ctx->notify_dummy_bh);
 241     thread_pool_free(ctx->thread_pool);
 242
 243     qemu_mutex_lock(&ctx->bh_lock);
 244     while (ctx->first_bh) {
 245         QEMUBH *next = ctx->first_bh->next;
 246
 247         /* qemu_bh_delete() must have been called on BHs in this AioContext */
 248         assert(ctx->first_bh->deleted);
 249
 250         g_free(ctx->first_bh);
 251         ctx->first_bh = next;
 252     }
 253     qemu_mutex_unlock(&ctx->bh_lock);
 254
 255     aio_set_event_notifier(ctx, &ctx->notifier, false, NULL);
 256     event_notifier_cleanup(&ctx->notifier);
 257     rfifolock_destroy(&ctx->lock);
 258     qemu_mutex_destroy(&ctx->bh_lock);
 259     timerlistgroup_deinit(&ctx->tlg);
 260 }
 261
 262 static GSourceFuncs aio_source_funcs = {
 263     aio_ctx_prepare,
 264     aio_ctx_check,
 265     aio_ctx_dispatch,
 266     aio_ctx_finalize
 267 };
 268
 269 GSource *aio_get_g_source(AioContext *ctx)
 270 {
 271     g_source_ref(&ctx->source);
 272     return &ctx->source;
 273 }
 274
 275 ThreadPool *aio_get_thread_pool(AioContext *ctx)
 276 {
 277     if (!ctx->thread_pool) {
 278         ctx->thread_pool = thread_pool_new(ctx);
 279     }
 280     return ctx->thread_pool;
 281 }
 282
 283 void aio_notify(AioContext *ctx)
 284 {
 285     /* Write e.g. bh->scheduled before reading ctx->notify_me.  Pairs
 286      * with atomic_or in aio_ctx_prepare or atomic_add in aio_poll.
 287      */
 288     smp_mb();
 289     if (ctx->notify_me) {
 290         event_notifier_set(&ctx->notifier);
 291         atomic_mb_set(&ctx->notified, true);
 292     }
 293 }
 294
 295 void aio_notify_accept(AioContext *ctx)
 296 {
 297     if (atomic_xchg(&ctx->notified, false)) {
 298         event_notifier_test_and_clear(&ctx->notifier);
 299     }
 300 }
 301
 302 static void aio_timerlist_notify(void *opaque)
 303 {
 304     aio_notify(opaque);
 305 }
 306
 307 static void aio_rfifolock_cb(void *opaque)
 308 {
 309     AioContext *ctx = opaque;
 310
 311     /* Kick owner thread in case they are blocked in aio_poll() */
 312     qemu_bh_schedule(ctx->notify_dummy_bh);
 313 }
 314
 315 static void notify_dummy_bh(void *opaque)
 316 {
 317     /* Do nothing, we were invoked just to force the event loop to iterate */
 318 }
 319
 320 static void event_notifier_dummy_cb(EventNotifier *e)
 321 {
 322 }
 323
 324 AioContext *aio_context_new(Error **errp)
 325 {
 326     int ret;
 327     AioContext *ctx;
 328     Error *local_err = NULL;
 329
 330     ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
 331     aio_context_setup(ctx, &local_err);
 332     if (local_err) {
 333         error_propagate(errp, local_err);
 334         goto fail;
 335     }
 336     ret = event_notifier_init(&ctx->notifier, false);
 337     if (ret < 0) {
 338         error_setg_errno(errp, -ret, "Failed to initialize event notifier");
 339         goto fail;
 340     }
 341     g_source_set_can_recurse(&ctx->source, true);
 342     aio_set_event_notifier(ctx, &ctx->notifier,
 343                            false,
 344                            (EventNotifierHandler *)
 345                            event_notifier_dummy_cb);
 346     ctx->thread_pool = NULL;
 347     qemu_mutex_init(&ctx->bh_lock);
 348     rfifolock_init(&ctx->lock, aio_rfifolock_cb, ctx);
 349     timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
 350
 351     ctx->notify_dummy_bh = aio_bh_new(ctx, notify_dummy_bh, NULL);
 352
 353     return ctx;
 354 fail:
 355     g_source_destroy(&ctx->source);
 356     return NULL;
 357 }
 358
 359 void aio_context_ref(AioContext *ctx)
 360 {
 361     g_source_ref(&ctx->source);
 362 }
 363
 364 void aio_context_unref(AioContext *ctx)
 365 {
 366     g_source_unref(&ctx->source);
 367 }
 368
 369 void aio_context_acquire(AioContext *ctx)
 370 {
 371     rfifolock_lock(&ctx->lock);
 372 }
 373
 374 void aio_context_release(AioContext *ctx)
 375 {
 376     rfifolock_unlock(&ctx->lock);
 377 }