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/osdep.h"
  26 #include "qapi/error.h"
  27 #include "qemu-common.h"
  28 #include "block/aio.h"
  29 #include "block/thread-pool.h"
  30 #include "qemu/main-loop.h"
  31 #include "qemu/atomic.h"
  32 #include "block/raw-aio.h"
  33
  34 /***********************************************************/
  35 /* bottom halves (can be seen as timers which expire ASAP) */
  36
  37 struct QEMUBH {
  38     AioContext *ctx;
  39     QEMUBHFunc *cb;
  40     void *opaque;
  41     QEMUBH *next;
  42     bool scheduled;
  43     bool idle;
  44     bool deleted;
  45 };
  46
  47 void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
  48 {
  49     QEMUBH *bh;
  50     bh = g_new(QEMUBH, 1);
  51     *bh = (QEMUBH){
  52         .ctx = ctx,
  53         .cb = cb,
  54         .opaque = opaque,
  55     };
  56     qemu_mutex_lock(&ctx->bh_lock);
  57     bh->next = ctx->first_bh;
  58     bh->scheduled = 1;
  59     bh->deleted = 1;
  60     /* Make sure that the members are ready before putting bh into list */
  61     smp_wmb();
  62     ctx->first_bh = bh;
  63     qemu_mutex_unlock(&ctx->bh_lock);
  64     aio_notify(ctx);
  65 }
  66
  67 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
  68 {
  69     QEMUBH *bh;
  70     bh = g_new(QEMUBH, 1);
  71     *bh = (QEMUBH){
  72         .ctx = ctx,
  73         .cb = cb,
  74         .opaque = opaque,
  75     };
  76     qemu_mutex_lock(&ctx->bh_lock);
  77     bh->next = ctx->first_bh;
  78     /* Make sure that the members are ready before putting bh into list */
  79     smp_wmb();
  80     ctx->first_bh = bh;
  81     qemu_mutex_unlock(&ctx->bh_lock);
  82     return bh;
  83 }
  84
  85 void aio_bh_call(QEMUBH *bh)
  86 {
  87     bh->cb(bh->opaque);
  88 }
  89
  90 /* Multiple occurrences of aio_bh_poll cannot be called concurrently */
  91 int aio_bh_poll(AioContext *ctx)
  92 {
  93     QEMUBH *bh, **bhp, *next;
  94     int ret;
  95
  96     ctx->walking_bh++;
  97
  98     ret = 0;
  99     for (bh = ctx->first_bh; bh; bh = next) {
 100         /* Make sure that fetching bh happens before accessing its members */
 101         smp_read_barrier_depends();
 102         next = bh->next;
 103         /* The atomic_xchg is paired with the one in qemu_bh_schedule.  The
 104          * implicit memory barrier ensures that the callback sees all writes
 105          * done by the scheduling thread.  It also ensures that the scheduling
 106          * thread sees the zero before bh->cb has run, and thus will call
 107          * aio_notify again if necessary.
 108          */
 109         if (atomic_xchg(&bh->scheduled, 0)) {
 110             /* Idle BHs and the notify BH don't count as progress */
 111             if (!bh->idle && bh != ctx->notify_dummy_bh) {
 112                 ret = 1;
 113             }
 114             bh->idle = 0;
 115             aio_bh_call(bh);
 116         }
 117     }
 118
 119     ctx->walking_bh--;
 120
 121     /* remove deleted bhs */
 122     if (!ctx->walking_bh) {
 123         qemu_mutex_lock(&ctx->bh_lock);
 124         bhp = &ctx->first_bh;
 125         while (*bhp) {
 126             bh = *bhp;
 127             if (bh->deleted && !bh->scheduled) {
 128                 *bhp = bh->next;
 129                 g_free(bh);
 130             } else {
 131                 bhp = &bh->next;
 132             }
 133         }
 134         qemu_mutex_unlock(&ctx->bh_lock);
 135     }
 136
 137     return ret;
 138 }
 139
 140 void qemu_bh_schedule_idle(QEMUBH *bh)
 141 {
 142     bh->idle = 1;
 143     /* Make sure that idle & any writes needed by the callback are done
 144      * before the locations are read in the aio_bh_poll.
 145      */
 146     atomic_mb_set(&bh->scheduled, 1);
 147 }
 148
 149 void qemu_bh_schedule(QEMUBH *bh)
 150 {
 151     AioContext *ctx;
 152
 153     ctx = bh->ctx;
 154     bh->idle = 0;
 155     /* The memory barrier implicit in atomic_xchg makes sure that:
 156      * 1. idle & any writes needed by the callback are done before the
 157      *    locations are read in the aio_bh_poll.
 158      * 2. ctx is loaded before scheduled is set and the callback has a chance
 159      *    to execute.
 160      */
 161     if (atomic_xchg(&bh->scheduled, 1) == 0) {
 162         aio_notify(ctx);
 163     }
 164 }
 165
 166
 167 /* This func is async.
 168  */
 169 void qemu_bh_cancel(QEMUBH *bh)
 170 {
 171     bh->scheduled = 0;
 172 }
 173
 174 /* This func is async.The bottom half will do the delete action at the finial
 175  * end.
 176  */
 177 void qemu_bh_delete(QEMUBH *bh)
 178 {
 179     bh->scheduled = 0;
 180     bh->deleted = 1;
 181 }
 182
 183 int64_t
 184 aio_compute_timeout(AioContext *ctx)
 185 {
 186     int64_t deadline;
 187     int timeout = -1;
 188     QEMUBH *bh;
 189
 190     for (bh = ctx->first_bh; bh; bh = bh->next) {
 191         if (bh->scheduled) {
 192             if (bh->idle) {
 193                 /* idle bottom halves will be polled at least
 194                  * every 10ms */
 195                 timeout = 10000000;
 196             } else {
 197                 /* non-idle bottom halves will be executed
 198                  * immediately */
 199                 return 0;
 200             }
 201         }
 202     }
 203
 204     deadline = timerlistgroup_deadline_ns(&ctx->tlg);
 205     if (deadline == 0) {
 206         return 0;
 207     } else {
 208         return qemu_soonest_timeout(timeout, deadline);
 209     }
 210 }
 211
 212 static gboolean
 213 aio_ctx_prepare(GSource *source, gint    *timeout)
 214 {
 215     AioContext *ctx = (AioContext *) source;
 216
 217     atomic_or(&ctx->notify_me, 1);
 218
 219     /* We assume there is no timeout already supplied */
 220     *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
 221
 222     if (aio_prepare(ctx)) {
 223         *timeout = 0;
 224     }
 225
 226     return *timeout == 0;
 227 }
 228
 229 static gboolean
 230 aio_ctx_check(GSource *source)
 231 {
 232     AioContext *ctx = (AioContext *) source;
 233     QEMUBH *bh;
 234
 235     atomic_and(&ctx->notify_me, ~1);
 236     aio_notify_accept(ctx);
 237
 238     for (bh = ctx->first_bh; bh; bh = bh->next) {
 239         if (bh->scheduled) {
 240             return true;
 241         }
 242     }
 243     return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
 244 }
 245
 246 static gboolean
 247 aio_ctx_dispatch(GSource     *source,
 248                  GSourceFunc  callback,
 249                  gpointer     user_data)
 250 {
 251     AioContext *ctx = (AioContext *) source;
 252
 253     assert(callback == NULL);
 254     aio_dispatch(ctx);
 255     return true;
 256 }
 257
 258 static void
 259 aio_ctx_finalize(GSource     *source)
 260 {
 261     AioContext *ctx = (AioContext *) source;
 262
 263     qemu_bh_delete(ctx->notify_dummy_bh);
 264     thread_pool_free(ctx->thread_pool);
 265
 266 #ifdef CONFIG_LINUX_AIO
 267     if (ctx->linux_aio) {
 268         laio_detach_aio_context(ctx->linux_aio, ctx);
 269         laio_cleanup(ctx->linux_aio);
 270         ctx->linux_aio = NULL;
 271     }
 272 #endif
 273
 274     qemu_mutex_lock(&ctx->bh_lock);
 275     while (ctx->first_bh) {
 276         QEMUBH *next = ctx->first_bh->next;
 277
 278         /* qemu_bh_delete() must have been called on BHs in this AioContext */
 279         assert(ctx->first_bh->deleted);
 280
 281         g_free(ctx->first_bh);
 282         ctx->first_bh = next;
 283     }
 284     qemu_mutex_unlock(&ctx->bh_lock);
 285
 286     aio_set_event_notifier(ctx, &ctx->notifier, false, NULL);
 287     event_notifier_cleanup(&ctx->notifier);
 288     rfifolock_destroy(&ctx->lock);
 289     qemu_mutex_destroy(&ctx->bh_lock);
 290     timerlistgroup_deinit(&ctx->tlg);
 291 }
 292
 293 static GSourceFuncs aio_source_funcs = {
 294     aio_ctx_prepare,
 295     aio_ctx_check,
 296     aio_ctx_dispatch,
 297     aio_ctx_finalize
 298 };
 299
 300 GSource *aio_get_g_source(AioContext *ctx)
 301 {
 302     g_source_ref(&ctx->source);
 303     return &ctx->source;
 304 }
 305
 306 ThreadPool *aio_get_thread_pool(AioContext *ctx)
 307 {
 308     if (!ctx->thread_pool) {
 309         ctx->thread_pool = thread_pool_new(ctx);
 310     }
 311     return ctx->thread_pool;
 312 }
 313
 314 #ifdef CONFIG_LINUX_AIO
 315 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
 316 {
 317     if (!ctx->linux_aio) {
 318         ctx->linux_aio = laio_init();
 319         laio_attach_aio_context(ctx->linux_aio, ctx);
 320     }
 321     return ctx->linux_aio;
 322 }
 323 #endif
 324
 325 void aio_notify(AioContext *ctx)
 326 {
 327     /* Write e.g. bh->scheduled before reading ctx->notify_me.  Pairs
 328      * with atomic_or in aio_ctx_prepare or atomic_add in aio_poll.
 329      */
 330     smp_mb();
 331     if (ctx->notify_me) {
 332         event_notifier_set(&ctx->notifier);
 333         atomic_mb_set(&ctx->notified, true);
 334     }
 335 }
 336
 337 void aio_notify_accept(AioContext *ctx)
 338 {
 339     if (atomic_xchg(&ctx->notified, false)) {
 340         event_notifier_test_and_clear(&ctx->notifier);
 341     }
 342 }
 343
 344 static void aio_timerlist_notify(void *opaque)
 345 {
 346     aio_notify(opaque);
 347 }
 348
 349 static void aio_rfifolock_cb(void *opaque)
 350 {
 351     AioContext *ctx = opaque;
 352
 353     /* Kick owner thread in case they are blocked in aio_poll() */
 354     qemu_bh_schedule(ctx->notify_dummy_bh);
 355 }
 356
 357 static void notify_dummy_bh(void *opaque)
 358 {
 359     /* Do nothing, we were invoked just to force the event loop to iterate */
 360 }
 361
 362 static void event_notifier_dummy_cb(EventNotifier *e)
 363 {
 364 }
 365
 366 AioContext *aio_context_new(Error **errp)
 367 {
 368     int ret;
 369     AioContext *ctx;
 370
 371     ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
 372     aio_context_setup(ctx);
 373
 374     ret = event_notifier_init(&ctx->notifier, false);
 375     if (ret < 0) {
 376         error_setg_errno(errp, -ret, "Failed to initialize event notifier");
 377         goto fail;
 378     }
 379     g_source_set_can_recurse(&ctx->source, true);
 380     aio_set_event_notifier(ctx, &ctx->notifier,
 381                            false,
 382                            (EventNotifierHandler *)
 383                            event_notifier_dummy_cb);
 384 #ifdef CONFIG_LINUX_AIO
 385     ctx->linux_aio = NULL;
 386 #endif
 387     ctx->thread_pool = NULL;
 388     qemu_mutex_init(&ctx->bh_lock);
 389     rfifolock_init(&ctx->lock, aio_rfifolock_cb, ctx);
 390     timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
 391
 392     ctx->notify_dummy_bh = aio_bh_new(ctx, notify_dummy_bh, NULL);
 393
 394     return ctx;
 395 fail:
 396     g_source_destroy(&ctx->source);
 397     return NULL;
 398 }
 399
 400 void aio_context_ref(AioContext *ctx)
 401 {
 402     g_source_ref(&ctx->source);
 403 }
 404
 405 void aio_context_unref(AioContext *ctx)
 406 {
 407     g_source_unref(&ctx->source);
 408 }
 409
 410 void aio_context_acquire(AioContext *ctx)
 411 {
 412     rfifolock_lock(&ctx->lock);
 413 }
 414
 415 void aio_context_release(AioContext *ctx)
 416 {
 417     rfifolock_unlock(&ctx->lock);
 418 }