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 #define MAX_QUEUED_IO  128
  32
  33 struct qemu_laiocb {
  34     BlockAIOCB common;
  35     Coroutine *co;
  36     LinuxAioState *ctx;
  37     struct iocb iocb;
  38     ssize_t ret;
  39     size_t nbytes;
  40     QEMUIOVector *qiov;
  41     bool is_read;
  42     QSIMPLEQ_ENTRY(qemu_laiocb) next;
  43 };
  44
  45 typedef struct {
  46     int plugged;
  47     unsigned int n;
  48     bool blocked;
  49     QSIMPLEQ_HEAD(, qemu_laiocb) pending;
  50 } LaioQueue;
  51
  52 struct LinuxAioState {
  53     AioContext *aio_context;
  54
  55     io_context_t ctx;
  56     EventNotifier e;
  57
  58     /* io queue for submit at batch */
  59     LaioQueue io_q;
  60
  61     /* I/O completion processing */
  62     QEMUBH *completion_bh;
  63     struct io_event events[MAX_EVENTS];
  64     int event_idx;
  65     int event_max;
  66 };
  67
  68 static void ioq_submit(LinuxAioState *s);
  69
  70 static inline ssize_t io_event_ret(struct io_event *ev)
  71 {
  72     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
  73 }
  74
  75 /*
  76  * Completes an AIO request (calls the callback and frees the ACB).
  77  */
  78 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
  79 {
  80     int ret;
  81
  82     ret = laiocb->ret;
  83     if (ret != -ECANCELED) {
  84         if (ret == laiocb->nbytes) {
  85             ret = 0;
  86         } else if (ret >= 0) {
  87             /* Short reads mean EOF, pad with zeros. */
  88             if (laiocb->is_read) {
  89                 qemu_iovec_memset(laiocb->qiov, ret, 0,
  90                     laiocb->qiov->size - ret);
  91             } else {
  92                 ret = -ENOSPC;
  93             }
  94         }
  95     }
  96
  97     laiocb->ret = ret;
  98     if (laiocb->co) {
  99         qemu_coroutine_enter(laiocb->co);
 100     } else {
 101         laiocb->common.cb(laiocb->common.opaque, ret);
 102         qemu_aio_unref(laiocb);
 103     }
 104 }
 105
 106 /* The completion BH fetches completed I/O requests and invokes their
 107  * callbacks.
 108  *
 109  * The function is somewhat tricky because it supports nested event loops, for
 110  * example when a request callback invokes aio_poll().  In order to do this,
 111  * the completion events array and index are kept in LinuxAioState.  The BH
 112  * reschedules itself as long as there are completions pending so it will
 113  * either be called again in a nested event loop or will be called after all
 114  * events have been completed.  When there are no events left to complete, the
 115  * BH returns without rescheduling.
 116  */
 117 static void qemu_laio_completion_bh(void *opaque)
 118 {
 119     LinuxAioState *s = opaque;
 120
 121     /* Fetch more completion events when empty */
 122     if (s->event_idx == s->event_max) {
 123         do {
 124             struct timespec ts = { 0 };
 125             s->event_max = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS,
 126                                         s->events, &ts);
 127         } while (s->event_max == -EINTR);
 128
 129         s->event_idx = 0;
 130         if (s->event_max <= 0) {
 131             s->event_max = 0;
 132             return; /* no more events */
 133         }
 134     }
 135
 136     /* Reschedule so nested event loops see currently pending completions */
 137     qemu_bh_schedule(s->completion_bh);
 138
 139     /* Process completion events */
 140     while (s->event_idx < s->event_max) {
 141         struct iocb *iocb = s->events[s->event_idx].obj;
 142         struct qemu_laiocb *laiocb =
 143                 container_of(iocb, struct qemu_laiocb, iocb);
 144
 145         laiocb->ret = io_event_ret(&s->events[s->event_idx]);
 146         s->event_idx++;
 147
 148         qemu_laio_process_completion(laiocb);
 149     }
 150
 151     if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 152         ioq_submit(s);
 153     }
 154
 155     qemu_bh_cancel(s->completion_bh);
 156 }
 157
 158 static void qemu_laio_completion_cb(EventNotifier *e)
 159 {
 160     LinuxAioState *s = container_of(e, LinuxAioState, e);
 161
 162     if (event_notifier_test_and_clear(&s->e)) {
 163         qemu_laio_completion_bh(s);
 164     }
 165 }
 166
 167 static void laio_cancel(BlockAIOCB *blockacb)
 168 {
 169     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 170     struct io_event event;
 171     int ret;
 172
 173     if (laiocb->ret != -EINPROGRESS) {
 174         return;
 175     }
 176     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 177     laiocb->ret = -ECANCELED;
 178     if (ret != 0) {
 179         /* iocb is not cancelled, cb will be called by the event loop later */
 180         return;
 181     }
 182
 183     laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
 184 }
 185
 186 static const AIOCBInfo laio_aiocb_info = {
 187     .aiocb_size         = sizeof(struct qemu_laiocb),
 188     .cancel_async       = laio_cancel,
 189 };
 190
 191 static void ioq_init(LaioQueue *io_q)
 192 {
 193     QSIMPLEQ_INIT(&io_q->pending);
 194     io_q->plugged = 0;
 195     io_q->n = 0;
 196     io_q->blocked = false;
 197 }
 198
 199 static void ioq_submit(LinuxAioState *s)
 200 {
 201     int ret, len;
 202     struct qemu_laiocb *aiocb;
 203     struct iocb *iocbs[MAX_QUEUED_IO];
 204     QSIMPLEQ_HEAD(, qemu_laiocb) completed;
 205
 206     do {
 207         len = 0;
 208         QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
 209             iocbs[len++] = &aiocb->iocb;
 210             if (len == MAX_QUEUED_IO) {
 211                 break;
 212             }
 213         }
 214
 215         ret = io_submit(s->ctx, len, iocbs);
 216         if (ret == -EAGAIN) {
 217             break;
 218         }
 219         if (ret < 0) {
 220             abort();
 221         }
 222
 223         s->io_q.n -= ret;
 224         aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
 225         QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
 226     } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
 227     s->io_q.blocked = (s->io_q.n > 0);
 228 }
 229
 230 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
 231 {
 232     s->io_q.plugged++;
 233 }
 234
 235 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s)
 236 {
 237     assert(s->io_q.plugged);
 238     if (--s->io_q.plugged == 0 &&
 239         !s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 240         ioq_submit(s);
 241     }
 242 }
 243
 244 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
 245                           int type)
 246 {
 247     LinuxAioState *s = laiocb->ctx;
 248     struct iocb *iocbs = &laiocb->iocb;
 249     QEMUIOVector *qiov = laiocb->qiov;
 250
 251     switch (type) {
 252     case QEMU_AIO_WRITE:
 253         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 254         break;
 255     case QEMU_AIO_READ:
 256         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 257         break;
 258     /* Currently Linux kernel does not support other operations */
 259     default:
 260         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 261                         __func__, type);
 262         return -EIO;
 263     }
 264     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
 265
 266     QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
 267     s->io_q.n++;
 268     if (!s->io_q.blocked &&
 269         (!s->io_q.plugged || s->io_q.n >= MAX_QUEUED_IO)) {
 270         ioq_submit(s);
 271     }
 272
 273     return 0;
 274 }
 275
 276 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
 277                                 uint64_t offset, QEMUIOVector *qiov, int type)
 278 {
 279     int ret;
 280     struct qemu_laiocb laiocb = {
 281         .co         = qemu_coroutine_self(),
 282         .nbytes     = qiov->size,
 283         .ctx        = s,
 284         .is_read    = (type == QEMU_AIO_READ),
 285         .qiov       = qiov,
 286     };
 287
 288     ret = laio_do_submit(fd, &laiocb, offset, type);
 289     if (ret < 0) {
 290         return ret;
 291     }
 292
 293     qemu_coroutine_yield();
 294     return laiocb.ret;
 295 }
 296
 297 BlockAIOCB *laio_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
 298         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 299         BlockCompletionFunc *cb, void *opaque, int type)
 300 {
 301     struct qemu_laiocb *laiocb;
 302     off_t offset = sector_num * BDRV_SECTOR_SIZE;
 303     int ret;
 304
 305     laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
 306     laiocb->nbytes = nb_sectors * BDRV_SECTOR_SIZE;
 307     laiocb->ctx = s;
 308     laiocb->ret = -EINPROGRESS;
 309     laiocb->is_read = (type == QEMU_AIO_READ);
 310     laiocb->qiov = qiov;
 311
 312     ret = laio_do_submit(fd, laiocb, offset, type);
 313     if (ret < 0) {
 314         qemu_aio_unref(laiocb);
 315         return NULL;
 316     }
 317
 318     return &laiocb->common;
 319 }
 320
 321 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
 322 {
 323     aio_set_event_notifier(old_context, &s->e, false, NULL);
 324     qemu_bh_delete(s->completion_bh);
 325 }
 326
 327 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
 328 {
 329     s->aio_context = new_context;
 330     s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
 331     aio_set_event_notifier(new_context, &s->e, false,
 332                            qemu_laio_completion_cb);
 333 }
 334
 335 LinuxAioState *laio_init(void)
 336 {
 337     LinuxAioState *s;
 338
 339     s = g_malloc0(sizeof(*s));
 340     if (event_notifier_init(&s->e, false) < 0) {
 341         goto out_free_state;
 342     }
 343
 344     if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
 345         goto out_close_efd;
 346     }
 347
 348     ioq_init(&s->io_q);
 349
 350     return s;
 351
 352 out_close_efd:
 353     event_notifier_cleanup(&s->e);
 354 out_free_state:
 355     g_free(s);
 356     return NULL;
 357 }
 358
 359 void laio_cleanup(LinuxAioState *s)
 360 {
 361     event_notifier_cleanup(&s->e);
 362
 363     if (io_destroy(s->ctx) != 0) {
 364         fprintf(stderr, "%s: destroy AIO context %p failed\n",
 365                         __func__, &s->ctx);
 366     }
 367     g_free(s);
 368 }