block/linux-aio.c

   1 /*
   2  * Linux native AIO support.
   3  *
   4  * Copyright (C) 2009 IBM, Corp.
   5  * Copyright (C) 2009 Red Hat, Inc.
   6  *
   7  * This work is licensed under the terms of the GNU GPL, version 2 or later.
   8  * See the COPYING file in the top-level directory.
   9  */
  10 #include "qemu/osdep.h"
  11 #include "qemu-common.h"
  12 #include "block/aio.h"
  13 #include "qemu/queue.h"
  14 #include "block/block.h"
  15 #include "block/raw-aio.h"
  16 #include "qemu/event_notifier.h"
  17 #include "qemu/coroutine.h"
  18
  19 #include <libaio.h>
  20
  21 /*
  22  * Queue size (per-device).
  23  *
  24  * XXX: eventually we need to communicate this to the guest and/or make it
  25  *      tunable by the guest.  If we get more outstanding requests at a time
  26  *      than this we will get EAGAIN from io_submit which is communicated to
  27  *      the guest as an I/O error.
  28  */
  29 #define MAX_EVENTS 128
  30
  31 struct qemu_laiocb {
  32     BlockAIOCB common;
  33     Coroutine *co;
  34     LinuxAioState *ctx;
  35     struct iocb iocb;
  36     ssize_t ret;
  37     size_t nbytes;
  38     QEMUIOVector *qiov;
  39     bool is_read;
  40     QSIMPLEQ_ENTRY(qemu_laiocb) next;
  41 };
  42
  43 typedef struct {
  44     int plugged;
  45     unsigned int in_queue;
  46     unsigned int in_flight;
  47     bool blocked;
  48     QSIMPLEQ_HEAD(, qemu_laiocb) pending;
  49 } LaioQueue;
  50
  51 struct LinuxAioState {
  52     AioContext *aio_context;
  53
  54     io_context_t ctx;
  55     EventNotifier e;
  56
  57     /* io queue for submit at batch */
  58     LaioQueue io_q;
  59
  60     /* I/O completion processing */
  61     QEMUBH *completion_bh;
  62     int event_idx;
  63     int event_max;
  64 };
  65
  66 static void ioq_submit(LinuxAioState *s);
  67
  68 static inline ssize_t io_event_ret(struct io_event *ev)
  69 {
  70     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
  71 }
  72
  73 /*
  74  * Completes an AIO request (calls the callback and frees the ACB).
  75  */
  76 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
  77 {
  78     int ret;
  79
  80     ret = laiocb->ret;
  81     if (ret != -ECANCELED) {
  82         if (ret == laiocb->nbytes) {
  83             ret = 0;
  84         } else if (ret >= 0) {
  85             /* Short reads mean EOF, pad with zeros. */
  86             if (laiocb->is_read) {
  87                 qemu_iovec_memset(laiocb->qiov, ret, 0,
  88                     laiocb->qiov->size - ret);
  89             } else {
  90                 ret = -ENOSPC;
  91             }
  92         }
  93     }
  94
  95     laiocb->ret = ret;
  96     if (laiocb->co) {
  97         /* If the coroutine is already entered it must be in ioq_submit() and
  98          * will notice laio->ret has been filled in when it eventually runs
  99          * later.  Coroutines cannot be entered recursively so avoid doing
 100          * that!
 101          */
 102         if (!qemu_coroutine_entered(laiocb->co)) {
 103             qemu_coroutine_enter(laiocb->co);
 104         }
 105     } else {
 106         laiocb->common.cb(laiocb->common.opaque, ret);
 107         qemu_aio_unref(laiocb);
 108     }
 109 }
 110
 111 /**
 112  * aio_ring buffer which is shared between userspace and kernel.
 113  *
 114  * This copied from linux/fs/aio.c, common header does not exist
 115  * but AIO exists for ages so we assume ABI is stable.
 116  */
 117 struct aio_ring {
 118     unsigned    id;    /* kernel internal index number */
 119     unsigned    nr;    /* number of io_events */
 120     unsigned    head;  /* Written to by userland or by kernel. */
 121     unsigned    tail;
 122
 123     unsigned    magic;
 124     unsigned    compat_features;
 125     unsigned    incompat_features;
 126     unsigned    header_length;  /* size of aio_ring */
 127
 128     struct io_event io_events[0];
 129 };
 130
 131 /**
 132  * io_getevents_peek:
 133  * @ctx: AIO context
 134  * @events: pointer on events array, output value
 135
 136  * Returns the number of completed events and sets a pointer
 137  * on events array.  This function does not update the internal
 138  * ring buffer, only reads head and tail.  When @events has been
 139  * processed io_getevents_commit() must be called.
 140  */
 141 static inline unsigned int io_getevents_peek(io_context_t ctx,
 142                                              struct io_event **events)
 143 {
 144     struct aio_ring *ring = (struct aio_ring *)ctx;
 145     unsigned int head = ring->head, tail = ring->tail;
 146     unsigned int nr;
 147
 148     nr = tail >= head ? tail - head : ring->nr - head;
 149     *events = ring->io_events + head;
 150     /* To avoid speculative loads of s->events[i] before observing tail.
 151        Paired with smp_wmb() inside linux/fs/aio.c: aio_complete(). */
 152     smp_rmb();
 153
 154     return nr;
 155 }
 156
 157 /**
 158  * io_getevents_commit:
 159  * @ctx: AIO context
 160  * @nr: the number of events on which head should be advanced
 161  *
 162  * Advances head of a ring buffer.
 163  */
 164 static inline void io_getevents_commit(io_context_t ctx, unsigned int nr)
 165 {
 166     struct aio_ring *ring = (struct aio_ring *)ctx;
 167
 168     if (nr) {
 169         ring->head = (ring->head + nr) % ring->nr;
 170     }
 171 }
 172
 173 /**
 174  * io_getevents_advance_and_peek:
 175  * @ctx: AIO context
 176  * @events: pointer on events array, output value
 177  * @nr: the number of events on which head should be advanced
 178  *
 179  * Advances head of a ring buffer and returns number of elements left.
 180  */
 181 static inline unsigned int
 182 io_getevents_advance_and_peek(io_context_t ctx,
 183                               struct io_event **events,
 184                               unsigned int nr)
 185 {
 186     io_getevents_commit(ctx, nr);
 187     return io_getevents_peek(ctx, events);
 188 }
 189
 190 /**
 191  * qemu_laio_process_completions:
 192  * @s: AIO state
 193  *
 194  * Fetches completed I/O requests and invokes their callbacks.
 195  *
 196  * The function is somewhat tricky because it supports nested event loops, for
 197  * example when a request callback invokes aio_poll().  In order to do this,
 198  * indices are kept in LinuxAioState.  Function schedules BH completion so it
 199  * can be called again in a nested event loop.  When there are no events left
 200  * to complete the BH is being canceled.
 201  */
 202 static void qemu_laio_process_completions(LinuxAioState *s)
 203 {
 204     struct io_event *events;
 205
 206     /* Reschedule so nested event loops see currently pending completions */
 207     qemu_bh_schedule(s->completion_bh);
 208
 209     while ((s->event_max = io_getevents_advance_and_peek(s->ctx, &events,
 210                                                          s->event_idx))) {
 211         for (s->event_idx = 0; s->event_idx < s->event_max; ) {
 212             struct iocb *iocb = events[s->event_idx].obj;
 213             struct qemu_laiocb *laiocb =
 214                 container_of(iocb, struct qemu_laiocb, iocb);
 215
 216             laiocb->ret = io_event_ret(&events[s->event_idx]);
 217
 218             /* Change counters one-by-one because we can be nested. */
 219             s->io_q.in_flight--;
 220             s->event_idx++;
 221             qemu_laio_process_completion(laiocb);
 222         }
 223     }
 224
 225     qemu_bh_cancel(s->completion_bh);
 226
 227     /* If we are nested we have to notify the level above that we are done
 228      * by setting event_max to zero, upper level will then jump out of it's
 229      * own `for` loop.  If we are the last all counters droped to zero. */
 230     s->event_max = 0;
 231     s->event_idx = 0;
 232 }
 233
 234 static void qemu_laio_process_completions_and_submit(LinuxAioState *s)
 235 {
 236     qemu_laio_process_completions(s);
 237     if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 238         ioq_submit(s);
 239     }
 240 }
 241
 242 static void qemu_laio_completion_bh(void *opaque)
 243 {
 244     LinuxAioState *s = opaque;
 245
 246     qemu_laio_process_completions_and_submit(s);
 247 }
 248
 249 static void qemu_laio_completion_cb(EventNotifier *e)
 250 {
 251     LinuxAioState *s = container_of(e, LinuxAioState, e);
 252
 253     if (event_notifier_test_and_clear(&s->e)) {
 254         qemu_laio_process_completions_and_submit(s);
 255     }
 256 }
 257
 258 static void laio_cancel(BlockAIOCB *blockacb)
 259 {
 260     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 261     struct io_event event;
 262     int ret;
 263
 264     if (laiocb->ret != -EINPROGRESS) {
 265         return;
 266     }
 267     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 268     laiocb->ret = -ECANCELED;
 269     if (ret != 0) {
 270         /* iocb is not cancelled, cb will be called by the event loop later */
 271         return;
 272     }
 273
 274     laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
 275 }
 276
 277 static const AIOCBInfo laio_aiocb_info = {
 278     .aiocb_size         = sizeof(struct qemu_laiocb),
 279     .cancel_async       = laio_cancel,
 280 };
 281
 282 static void ioq_init(LaioQueue *io_q)
 283 {
 284     QSIMPLEQ_INIT(&io_q->pending);
 285     io_q->plugged = 0;
 286     io_q->in_queue = 0;
 287     io_q->in_flight = 0;
 288     io_q->blocked = false;
 289 }
 290
 291 static void ioq_submit(LinuxAioState *s)
 292 {
 293     int ret, len;
 294     struct qemu_laiocb *aiocb;
 295     struct iocb *iocbs[MAX_EVENTS];
 296     QSIMPLEQ_HEAD(, qemu_laiocb) completed;
 297
 298     do {
 299         if (s->io_q.in_flight >= MAX_EVENTS) {
 300             break;
 301         }
 302         len = 0;
 303         QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
 304             iocbs[len++] = &aiocb->iocb;
 305             if (s->io_q.in_flight + len >= MAX_EVENTS) {
 306                 break;
 307             }
 308         }
 309
 310         ret = io_submit(s->ctx, len, iocbs);
 311         if (ret == -EAGAIN) {
 312             break;
 313         }
 314         if (ret < 0) {
 315             /* Fail the first request, retry the rest */
 316             aiocb = QSIMPLEQ_FIRST(&s->io_q.pending);
 317             QSIMPLEQ_REMOVE_HEAD(&s->io_q.pending, next);
 318             s->io_q.in_queue--;
 319             aiocb->ret = ret;
 320             qemu_laio_process_completion(aiocb);
 321             continue;
 322         }
 323
 324         s->io_q.in_flight += ret;
 325         s->io_q.in_queue  -= ret;
 326         aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
 327         QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
 328     } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
 329     s->io_q.blocked = (s->io_q.in_queue > 0);
 330
 331     if (s->io_q.in_flight) {
 332         /* We can try to complete something just right away if there are
 333          * still requests in-flight. */
 334         qemu_laio_process_completions(s);
 335         /*
 336          * Even we have completed everything (in_flight == 0), the queue can
 337          * have still pended requests (in_queue > 0).  We do not attempt to
 338          * repeat submission to avoid IO hang.  The reason is simple: s->e is
 339          * still set and completion callback will be called shortly and all
 340          * pended requests will be submitted from there.
 341          */
 342     }
 343 }
 344
 345 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
 346 {
 347     s->io_q.plugged++;
 348 }
 349
 350 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s)
 351 {
 352     assert(s->io_q.plugged);
 353     if (--s->io_q.plugged == 0 &&
 354         !s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 355         ioq_submit(s);
 356     }
 357 }
 358
 359 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
 360                           int type)
 361 {
 362     LinuxAioState *s = laiocb->ctx;
 363     struct iocb *iocbs = &laiocb->iocb;
 364     QEMUIOVector *qiov = laiocb->qiov;
 365
 366     switch (type) {
 367     case QEMU_AIO_WRITE:
 368         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 369         break;
 370     case QEMU_AIO_READ:
 371         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 372         break;
 373     /* Currently Linux kernel does not support other operations */
 374     default:
 375         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 376                         __func__, type);
 377         return -EIO;
 378     }
 379     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
 380
 381     QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
 382     s->io_q.in_queue++;
 383     if (!s->io_q.blocked &&
 384         (!s->io_q.plugged ||
 385          s->io_q.in_flight + s->io_q.in_queue >= MAX_EVENTS)) {
 386         ioq_submit(s);
 387     }
 388
 389     return 0;
 390 }
 391
 392 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
 393                                 uint64_t offset, QEMUIOVector *qiov, int type)
 394 {
 395     int ret;
 396     struct qemu_laiocb laiocb = {
 397         .co         = qemu_coroutine_self(),
 398         .nbytes     = qiov->size,
 399         .ctx        = s,
 400         .ret        = -EINPROGRESS,
 401         .is_read    = (type == QEMU_AIO_READ),
 402         .qiov       = qiov,
 403     };
 404
 405     ret = laio_do_submit(fd, &laiocb, offset, type);
 406     if (ret < 0) {
 407         return ret;
 408     }
 409
 410     if (laiocb.ret == -EINPROGRESS) {
 411         qemu_coroutine_yield();
 412     }
 413     return laiocb.ret;
 414 }
 415
 416 BlockAIOCB *laio_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
 417         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 418         BlockCompletionFunc *cb, void *opaque, int type)
 419 {
 420     struct qemu_laiocb *laiocb;
 421     off_t offset = sector_num * BDRV_SECTOR_SIZE;
 422     int ret;
 423
 424     laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
 425     laiocb->nbytes = nb_sectors * BDRV_SECTOR_SIZE;
 426     laiocb->ctx = s;
 427     laiocb->ret = -EINPROGRESS;
 428     laiocb->is_read = (type == QEMU_AIO_READ);
 429     laiocb->qiov = qiov;
 430
 431     ret = laio_do_submit(fd, laiocb, offset, type);
 432     if (ret < 0) {
 433         qemu_aio_unref(laiocb);
 434         return NULL;
 435     }
 436
 437     return &laiocb->common;
 438 }
 439
 440 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
 441 {
 442     aio_set_event_notifier(old_context, &s->e, false, NULL);
 443     qemu_bh_delete(s->completion_bh);
 444 }
 445
 446 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
 447 {
 448     s->aio_context = new_context;
 449     s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
 450     aio_set_event_notifier(new_context, &s->e, false,
 451                            qemu_laio_completion_cb);
 452 }
 453
 454 LinuxAioState *laio_init(void)
 455 {
 456     LinuxAioState *s;
 457
 458     s = g_malloc0(sizeof(*s));
 459     if (event_notifier_init(&s->e, false) < 0) {
 460         goto out_free_state;
 461     }
 462
 463     if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
 464         goto out_close_efd;
 465     }
 466
 467     ioq_init(&s->io_q);
 468
 469     return s;
 470
 471 out_close_efd:
 472     event_notifier_cleanup(&s->e);
 473 out_free_state:
 474     g_free(s);
 475     return NULL;
 476 }
 477
 478 void laio_cleanup(LinuxAioState *s)
 479 {
 480     event_notifier_cleanup(&s->e);
 481
 482     if (io_destroy(s->ctx) != 0) {
 483         fprintf(stderr, "%s: destroy AIO context %p failed\n",
 484                         __func__, &s->ctx);
 485     }
 486     g_free(s);
 487 }