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-common.h"
  11 #include "qemu-aio.h"
  12 #include "qemu-queue.h"
  13 #include "block/raw-aio.h"
  14 #include "event_notifier.h"
  15
  16 #include <libaio.h>
  17
  18 /*
  19  * Queue size (per-device).
  20  *
  21  * XXX: eventually we need to communicate this to the guest and/or make it
  22  *      tunable by the guest.  If we get more outstanding requests at a time
  23  *      than this we will get EAGAIN from io_submit which is communicated to
  24  *      the guest as an I/O error.
  25  */
  26 #define MAX_EVENTS 128
  27
  28 struct qemu_laiocb {
  29     BlockDriverAIOCB common;
  30     struct qemu_laio_state *ctx;
  31     struct iocb iocb;
  32     ssize_t ret;
  33     size_t nbytes;
  34     QEMUIOVector *qiov;
  35     bool is_read;
  36     QLIST_ENTRY(qemu_laiocb) node;
  37 };
  38
  39 struct qemu_laio_state {
  40     io_context_t ctx;
  41     EventNotifier e;
  42     int count;
  43 };
  44
  45 static inline ssize_t io_event_ret(struct io_event *ev)
  46 {
  47     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
  48 }
  49
  50 /*
  51  * Completes an AIO request (calls the callback and frees the ACB).
  52  */
  53 static void qemu_laio_process_completion(struct qemu_laio_state *s,
  54     struct qemu_laiocb *laiocb)
  55 {
  56     int ret;
  57
  58     s->count--;
  59
  60     ret = laiocb->ret;
  61     if (ret != -ECANCELED) {
  62         if (ret == laiocb->nbytes) {
  63             ret = 0;
  64         } else if (ret >= 0) {
  65             /* Short reads mean EOF, pad with zeros. */
  66             if (laiocb->is_read) {
  67                 qemu_iovec_memset(laiocb->qiov, ret, 0,
  68                     laiocb->qiov->size - ret);
  69             } else {
  70                 ret = -EINVAL;
  71             }
  72         }
  73
  74         laiocb->common.cb(laiocb->common.opaque, ret);
  75     }
  76
  77     qemu_aio_release(laiocb);
  78 }
  79
  80 static void qemu_laio_completion_cb(EventNotifier *e)
  81 {
  82     struct qemu_laio_state *s = container_of(e, struct qemu_laio_state, e);
  83
  84     while (event_notifier_test_and_clear(&s->e)) {
  85         struct io_event events[MAX_EVENTS];
  86         struct timespec ts = { 0 };
  87         int nevents, i;
  88
  89         do {
  90             nevents = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS, events, &ts);
  91         } while (nevents == -EINTR);
  92
  93         for (i = 0; i < nevents; i++) {
  94             struct iocb *iocb = events[i].obj;
  95             struct qemu_laiocb *laiocb =
  96                     container_of(iocb, struct qemu_laiocb, iocb);
  97
  98             laiocb->ret = io_event_ret(&events[i]);
  99             qemu_laio_process_completion(s, laiocb);
 100         }
 101     }
 102 }
 103
 104 static int qemu_laio_flush_cb(EventNotifier *e)
 105 {
 106     struct qemu_laio_state *s = container_of(e, struct qemu_laio_state, e);
 107
 108     return (s->count > 0) ? 1 : 0;
 109 }
 110
 111 static void laio_cancel(BlockDriverAIOCB *blockacb)
 112 {
 113     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 114     struct io_event event;
 115     int ret;
 116
 117     if (laiocb->ret != -EINPROGRESS)
 118         return;
 119
 120     /*
 121      * Note that as of Linux 2.6.31 neither the block device code nor any
 122      * filesystem implements cancellation of AIO request.
 123      * Thus the polling loop below is the normal code path.
 124      */
 125     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 126     if (ret == 0) {
 127         laiocb->ret = -ECANCELED;
 128         return;
 129     }
 130
 131     /*
 132      * We have to wait for the iocb to finish.
 133      *
 134      * The only way to get the iocb status update is by polling the io context.
 135      * We might be able to do this slightly more optimal by removing the
 136      * O_NONBLOCK flag.
 137      */
 138     while (laiocb->ret == -EINPROGRESS) {
 139         qemu_laio_completion_cb(&laiocb->ctx->e);
 140     }
 141 }
 142
 143 static const AIOCBInfo laio_aiocb_info = {
 144     .aiocb_size         = sizeof(struct qemu_laiocb),
 145     .cancel             = laio_cancel,
 146 };
 147
 148 BlockDriverAIOCB *laio_submit(BlockDriverState *bs, void *aio_ctx, int fd,
 149         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 150         BlockDriverCompletionFunc *cb, void *opaque, int type)
 151 {
 152     struct qemu_laio_state *s = aio_ctx;
 153     struct qemu_laiocb *laiocb;
 154     struct iocb *iocbs;
 155     off_t offset = sector_num * 512;
 156
 157     laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
 158     laiocb->nbytes = nb_sectors * 512;
 159     laiocb->ctx = s;
 160     laiocb->ret = -EINPROGRESS;
 161     laiocb->is_read = (type == QEMU_AIO_READ);
 162     laiocb->qiov = qiov;
 163
 164     iocbs = &laiocb->iocb;
 165
 166     switch (type) {
 167     case QEMU_AIO_WRITE:
 168         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 169         break;
 170     case QEMU_AIO_READ:
 171         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 172         break;
 173     /* Currently Linux kernel does not support other operations */
 174     default:
 175         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 176                         __func__, type);
 177         goto out_free_aiocb;
 178     }
 179     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
 180     s->count++;
 181
 182     if (io_submit(s->ctx, 1, &iocbs) < 0)
 183         goto out_dec_count;
 184     return &laiocb->common;
 185
 186 out_dec_count:
 187     s->count--;
 188 out_free_aiocb:
 189     qemu_aio_release(laiocb);
 190     return NULL;
 191 }
 192
 193 void *laio_init(void)
 194 {
 195     struct qemu_laio_state *s;
 196
 197     s = g_malloc0(sizeof(*s));
 198     if (event_notifier_init(&s->e, false) < 0) {
 199         goto out_free_state;
 200     }
 201
 202     if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
 203         goto out_close_efd;
 204     }
 205
 206     qemu_aio_set_event_notifier(&s->e, qemu_laio_completion_cb,
 207                                 qemu_laio_flush_cb);
 208
 209     return s;
 210
 211 out_close_efd:
 212     event_notifier_cleanup(&s->e);
 213 out_free_state:
 214     g_free(s);
 215     return NULL;
 216 }