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 /*
  72  * All requests are directly processed when they complete, so there's nothing
  73  * left to do during qemu_aio_wait().
  74  */
  75 static int qemu_laio_process_requests(void *opaque)
  76 {
  77     return 0;
  78 }
  79
  80 static void qemu_laio_completion_cb(void *opaque)
  81 {
  82     struct qemu_laio_state *s = opaque;
  83
  84     while (1) {
  85         struct io_event events[MAX_EVENTS];
  86         uint64_t val;
  87         ssize_t ret;
  88         struct timespec ts = { 0 };
  89         int nevents, i;
  90
  91         do {
  92             ret = read(s->efd, &val, sizeof(val));
  93         } while (ret == -1 && errno == EINTR);
  94
  95         if (ret == -1 && errno == EAGAIN)
  96             break;
  97
  98         if (ret != 8)
  99             break;
 100
 101         do {
 102             nevents = io_getevents(s->ctx, val, MAX_EVENTS, events, &ts);
 103         } while (nevents == -EINTR);
 104
 105         for (i = 0; i < nevents; i++) {
 106             struct iocb *iocb = events[i].obj;
 107             struct qemu_laiocb *laiocb =
 108                     container_of(iocb, struct qemu_laiocb, iocb);
 109
 110             laiocb->ret = io_event_ret(&events[i]);
 111             qemu_laio_process_completion(s, laiocb);
 112         }
 113     }
 114 }
 115
 116 static int qemu_laio_flush_cb(void *opaque)
 117 {
 118     struct qemu_laio_state *s = opaque;
 119
 120     return (s->count > 0) ? 1 : 0;
 121 }
 122
 123 static void laio_cancel(BlockDriverAIOCB *blockacb)
 124 {
 125     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 126     struct io_event event;
 127     int ret;
 128
 129     if (laiocb->ret != -EINPROGRESS)
 130         return;
 131
 132     /*
 133      * Note that as of Linux 2.6.31 neither the block device code nor any
 134      * filesystem implements cancellation of AIO request.
 135      * Thus the polling loop below is the normal code path.
 136      */
 137     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 138     if (ret == 0) {
 139         laiocb->ret = -ECANCELED;
 140         return;
 141     }
 142
 143     /*
 144      * We have to wait for the iocb to finish.
 145      *
 146      * The only way to get the iocb status update is by polling the io context.
 147      * We might be able to do this slightly more optimal by removing the
 148      * O_NONBLOCK flag.
 149      */
 150     while (laiocb->ret == -EINPROGRESS)
 151         qemu_laio_completion_cb(laiocb->ctx);
 152 }
 153
 154 static AIOPool laio_pool = {
 155     .aiocb_size         = sizeof(struct qemu_laiocb),
 156     .cancel             = laio_cancel,
 157 };
 158
 159 BlockDriverAIOCB *laio_submit(BlockDriverState *bs, void *aio_ctx, int fd,
 160         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 161         BlockDriverCompletionFunc *cb, void *opaque, int type)
 162 {
 163     struct qemu_laio_state *s = aio_ctx;
 164     struct qemu_laiocb *laiocb;
 165     struct iocb *iocbs;
 166     off_t offset = sector_num * 512;
 167
 168     laiocb = qemu_aio_get(&laio_pool, bs, cb, opaque);
 169     if (!laiocb)
 170         return NULL;
 171     laiocb->nbytes = nb_sectors * 512;
 172     laiocb->ctx = s;
 173     laiocb->ret = -EINPROGRESS;
 174
 175     iocbs = &laiocb->iocb;
 176
 177     switch (type) {
 178     case QEMU_AIO_WRITE:
 179         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 180         break;
 181     case QEMU_AIO_READ:
 182         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 183         break;
 184     default:
 185         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 186                         __func__, type);
 187         goto out_free_aiocb;
 188     }
 189     io_set_eventfd(&laiocb->iocb, s->efd);
 190     s->count++;
 191
 192     if (io_submit(s->ctx, 1, &iocbs) < 0)
 193         goto out_dec_count;
 194     return &laiocb->common;
 195
 196 out_free_aiocb:
 197     qemu_aio_release(laiocb);
 198 out_dec_count:
 199     s->count--;
 200     return NULL;
 201 }
 202
 203 void *laio_init(void)
 204 {
 205     struct qemu_laio_state *s;
 206
 207     s = qemu_mallocz(sizeof(*s));
 208     s->efd = eventfd(0, 0);
 209     if (s->efd == -1)
 210         goto out_free_state;
 211     fcntl(s->efd, F_SETFL, O_NONBLOCK);
 212
 213     if (io_setup(MAX_EVENTS, &s->ctx) != 0)
 214         goto out_close_efd;
 215
 216     qemu_aio_set_fd_handler(s->efd, qemu_laio_completion_cb, NULL,
 217         qemu_laio_flush_cb, qemu_laio_process_requests, s);
 218
 219     return s;
 220
 221 out_close_efd:
 222     close(s->efd);
 223 out_free_state:
 224     qemu_free(s);
 225     return NULL;
 226 }