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 "block_int.h"
  13 #include "block/raw-posix-aio.h"
  14
  15 #include <sys/eventfd.h>
  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     QLIST_ENTRY(qemu_laiocb) node;
  35 };
  36
  37 struct qemu_laio_state {
  38     io_context_t ctx;
  39     int efd;
  40     int count;
  41 };
  42
  43 static inline ssize_t io_event_ret(struct io_event *ev)
  44 {
  45     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
  46 }
  47
  48 /*
  49  * Completes an AIO request (calls the callback and frees the ACB).
  50  */
  51 static void qemu_laio_process_completion(struct qemu_laio_state *s,
  52     struct qemu_laiocb *laiocb)
  53 {
  54     int ret;
  55
  56     s->count--;
  57
  58     ret = laiocb->ret;
  59     if (ret != -ECANCELED) {
  60         if (ret == laiocb->nbytes)
  61             ret = 0;
  62         else if (ret >= 0)
  63             ret = -EINVAL;
  64
  65         laiocb->common.cb(laiocb->common.opaque, ret);
  66     }
  67
  68     qemu_aio_release(laiocb);
  69 }
  70
  71 static void qemu_laio_completion_cb(void *opaque)
  72 {
  73     struct qemu_laio_state *s = opaque;
  74
  75     while (1) {
  76         struct io_event events[MAX_EVENTS];
  77         uint64_t val;
  78         ssize_t ret;
  79         struct timespec ts = { 0 };
  80         int nevents, i;
  81
  82         do {
  83             ret = read(s->efd, &val, sizeof(val));
  84         } while (ret == -1 && errno == EINTR);
  85
  86         if (ret == -1 && errno == EAGAIN)
  87             break;
  88
  89         if (ret != 8)
  90             break;
  91
  92         do {
  93             nevents = io_getevents(s->ctx, val, MAX_EVENTS, events, &ts);
  94         } while (nevents == -EINTR);
  95
  96         for (i = 0; i < nevents; i++) {
  97             struct iocb *iocb = events[i].obj;
  98             struct qemu_laiocb *laiocb =
  99                     container_of(iocb, struct qemu_laiocb, iocb);
 100
 101             laiocb->ret = io_event_ret(&events[i]);
 102             qemu_laio_process_completion(s, laiocb);
 103         }
 104     }
 105 }
 106
 107 static int qemu_laio_flush_cb(void *opaque)
 108 {
 109     struct qemu_laio_state *s = opaque;
 110
 111     return (s->count > 0) ? 1 : 0;
 112 }
 113
 114 static void laio_cancel(BlockDriverAIOCB *blockacb)
 115 {
 116     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 117     struct io_event event;
 118     int ret;
 119
 120     if (laiocb->ret != -EINPROGRESS)
 121         return;
 122
 123     /*
 124      * Note that as of Linux 2.6.31 neither the block device code nor any
 125      * filesystem implements cancellation of AIO request.
 126      * Thus the polling loop below is the normal code path.
 127      */
 128     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 129     if (ret == 0) {
 130         laiocb->ret = -ECANCELED;
 131         return;
 132     }
 133
 134     /*
 135      * We have to wait for the iocb to finish.
 136      *
 137      * The only way to get the iocb status update is by polling the io context.
 138      * We might be able to do this slightly more optimal by removing the
 139      * O_NONBLOCK flag.
 140      */
 141     while (laiocb->ret == -EINPROGRESS)
 142         qemu_laio_completion_cb(laiocb->ctx);
 143 }
 144
 145 static AIOPool laio_pool = {
 146     .aiocb_size         = sizeof(struct qemu_laiocb),
 147     .cancel             = laio_cancel,
 148 };
 149
 150 BlockDriverAIOCB *laio_submit(BlockDriverState *bs, void *aio_ctx, int fd,
 151         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 152         BlockDriverCompletionFunc *cb, void *opaque, int type)
 153 {
 154     struct qemu_laio_state *s = aio_ctx;
 155     struct qemu_laiocb *laiocb;
 156     struct iocb *iocbs;
 157     off_t offset = sector_num * 512;
 158
 159     laiocb = qemu_aio_get(&laio_pool, bs, cb, opaque);
 160     if (!laiocb)
 161         return NULL;
 162     laiocb->nbytes = nb_sectors * 512;
 163     laiocb->ctx = s;
 164     laiocb->ret = -EINPROGRESS;
 165
 166     iocbs = &laiocb->iocb;
 167
 168     switch (type) {
 169     case QEMU_AIO_WRITE:
 170         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 171         break;
 172     case QEMU_AIO_READ:
 173         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 174         break;
 175     /* Currently Linux kernel does not support other operations */
 176     default:
 177         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 178                         __func__, type);
 179         goto out_free_aiocb;
 180     }
 181     io_set_eventfd(&laiocb->iocb, s->efd);
 182     s->count++;
 183
 184     if (io_submit(s->ctx, 1, &iocbs) < 0)
 185         goto out_dec_count;
 186     return &laiocb->common;
 187
 188 out_free_aiocb:
 189     qemu_aio_release(laiocb);
 190 out_dec_count:
 191     s->count--;
 192     return NULL;
 193 }
 194
 195 void *laio_init(void)
 196 {
 197     struct qemu_laio_state *s;
 198
 199     s = g_malloc0(sizeof(*s));
 200     s->efd = eventfd(0, 0);
 201     if (s->efd == -1)
 202         goto out_free_state;
 203     fcntl(s->efd, F_SETFL, O_NONBLOCK);
 204
 205     if (io_setup(MAX_EVENTS, &s->ctx) != 0)
 206         goto out_close_efd;
 207
 208     qemu_aio_set_fd_handler(s->efd, qemu_laio_completion_cb, NULL,
 209         qemu_laio_flush_cb, NULL, s);
 210
 211     return s;
 212
 213 out_close_efd:
 214     close(s->efd);
 215 out_free_state:
 216     g_free(s);
 217     return NULL;
 218 }