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/raw-aio.h"
  15 #include "qemu/event_notifier.h"
  16
  17 #include <libaio.h>
  18
  19 /*
  20  * Queue size (per-device).
  21  *
  22  * XXX: eventually we need to communicate this to the guest and/or make it
  23  *      tunable by the guest.  If we get more outstanding requests at a time
  24  *      than this we will get EAGAIN from io_submit which is communicated to
  25  *      the guest as an I/O error.
  26  */
  27 #define MAX_EVENTS 128
  28
  29 #define MAX_QUEUED_IO  128
  30
  31 struct qemu_laiocb {
  32     BlockAIOCB common;
  33     LinuxAioState *ctx;
  34     struct iocb iocb;
  35     ssize_t ret;
  36     size_t nbytes;
  37     QEMUIOVector *qiov;
  38     bool is_read;
  39     QSIMPLEQ_ENTRY(qemu_laiocb) next;
  40 };
  41
  42 typedef struct {
  43     int plugged;
  44     unsigned int n;
  45     bool blocked;
  46     QSIMPLEQ_HEAD(, qemu_laiocb) pending;
  47 } LaioQueue;
  48
  49 struct LinuxAioState {
  50     io_context_t ctx;
  51     EventNotifier e;
  52
  53     /* io queue for submit at batch */
  54     LaioQueue io_q;
  55
  56     /* I/O completion processing */
  57     QEMUBH *completion_bh;
  58     struct io_event events[MAX_EVENTS];
  59     int event_idx;
  60     int event_max;
  61 };
  62
  63 static void ioq_submit(LinuxAioState *s);
  64
  65 static inline ssize_t io_event_ret(struct io_event *ev)
  66 {
  67     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
  68 }
  69
  70 /*
  71  * Completes an AIO request (calls the callback and frees the ACB).
  72  */
  73 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
  74 {
  75     int ret;
  76
  77     ret = laiocb->ret;
  78     if (ret != -ECANCELED) {
  79         if (ret == laiocb->nbytes) {
  80             ret = 0;
  81         } else if (ret >= 0) {
  82             /* Short reads mean EOF, pad with zeros. */
  83             if (laiocb->is_read) {
  84                 qemu_iovec_memset(laiocb->qiov, ret, 0,
  85                     laiocb->qiov->size - ret);
  86             } else {
  87                 ret = -EINVAL;
  88             }
  89         }
  90     }
  91     laiocb->common.cb(laiocb->common.opaque, ret);
  92
  93     qemu_aio_unref(laiocb);
  94 }
  95
  96 /* The completion BH fetches completed I/O requests and invokes their
  97  * callbacks.
  98  *
  99  * The function is somewhat tricky because it supports nested event loops, for
 100  * example when a request callback invokes aio_poll().  In order to do this,
 101  * the completion events array and index are kept in LinuxAioState.  The BH
 102  * reschedules itself as long as there are completions pending so it will
 103  * either be called again in a nested event loop or will be called after all
 104  * events have been completed.  When there are no events left to complete, the
 105  * BH returns without rescheduling.
 106  */
 107 static void qemu_laio_completion_bh(void *opaque)
 108 {
 109     LinuxAioState *s = opaque;
 110
 111     /* Fetch more completion events when empty */
 112     if (s->event_idx == s->event_max) {
 113         do {
 114             struct timespec ts = { 0 };
 115             s->event_max = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS,
 116                                         s->events, &ts);
 117         } while (s->event_max == -EINTR);
 118
 119         s->event_idx = 0;
 120         if (s->event_max <= 0) {
 121             s->event_max = 0;
 122             return; /* no more events */
 123         }
 124     }
 125
 126     /* Reschedule so nested event loops see currently pending completions */
 127     qemu_bh_schedule(s->completion_bh);
 128
 129     /* Process completion events */
 130     while (s->event_idx < s->event_max) {
 131         struct iocb *iocb = s->events[s->event_idx].obj;
 132         struct qemu_laiocb *laiocb =
 133                 container_of(iocb, struct qemu_laiocb, iocb);
 134
 135         laiocb->ret = io_event_ret(&s->events[s->event_idx]);
 136         s->event_idx++;
 137
 138         qemu_laio_process_completion(laiocb);
 139     }
 140
 141     if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 142         ioq_submit(s);
 143     }
 144 }
 145
 146 static void qemu_laio_completion_cb(EventNotifier *e)
 147 {
 148     LinuxAioState *s = container_of(e, LinuxAioState, e);
 149
 150     if (event_notifier_test_and_clear(&s->e)) {
 151         qemu_bh_schedule(s->completion_bh);
 152     }
 153 }
 154
 155 static void laio_cancel(BlockAIOCB *blockacb)
 156 {
 157     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 158     struct io_event event;
 159     int ret;
 160
 161     if (laiocb->ret != -EINPROGRESS) {
 162         return;
 163     }
 164     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 165     laiocb->ret = -ECANCELED;
 166     if (ret != 0) {
 167         /* iocb is not cancelled, cb will be called by the event loop later */
 168         return;
 169     }
 170
 171     laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
 172 }
 173
 174 static const AIOCBInfo laio_aiocb_info = {
 175     .aiocb_size         = sizeof(struct qemu_laiocb),
 176     .cancel_async       = laio_cancel,
 177 };
 178
 179 static void ioq_init(LaioQueue *io_q)
 180 {
 181     QSIMPLEQ_INIT(&io_q->pending);
 182     io_q->plugged = 0;
 183     io_q->n = 0;
 184     io_q->blocked = false;
 185 }
 186
 187 static void ioq_submit(LinuxAioState *s)
 188 {
 189     int ret, len;
 190     struct qemu_laiocb *aiocb;
 191     struct iocb *iocbs[MAX_QUEUED_IO];
 192     QSIMPLEQ_HEAD(, qemu_laiocb) completed;
 193
 194     do {
 195         len = 0;
 196         QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
 197             iocbs[len++] = &aiocb->iocb;
 198             if (len == MAX_QUEUED_IO) {
 199                 break;
 200             }
 201         }
 202
 203         ret = io_submit(s->ctx, len, iocbs);
 204         if (ret == -EAGAIN) {
 205             break;
 206         }
 207         if (ret < 0) {
 208             abort();
 209         }
 210
 211         s->io_q.n -= ret;
 212         aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
 213         QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
 214     } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
 215     s->io_q.blocked = (s->io_q.n > 0);
 216 }
 217
 218 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
 219 {
 220     assert(!s->io_q.plugged);
 221     s->io_q.plugged = 1;
 222 }
 223
 224 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s)
 225 {
 226     assert(s->io_q.plugged);
 227     s->io_q.plugged = 0;
 228     if (!s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 229         ioq_submit(s);
 230     }
 231 }
 232
 233 BlockAIOCB *laio_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
 234         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 235         BlockCompletionFunc *cb, void *opaque, int type)
 236 {
 237     struct qemu_laiocb *laiocb;
 238     struct iocb *iocbs;
 239     off_t offset = sector_num * 512;
 240
 241     laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
 242     laiocb->nbytes = nb_sectors * 512;
 243     laiocb->ctx = s;
 244     laiocb->ret = -EINPROGRESS;
 245     laiocb->is_read = (type == QEMU_AIO_READ);
 246     laiocb->qiov = qiov;
 247
 248     iocbs = &laiocb->iocb;
 249
 250     switch (type) {
 251     case QEMU_AIO_WRITE:
 252         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 253         break;
 254     case QEMU_AIO_READ:
 255         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 256         break;
 257     /* Currently Linux kernel does not support other operations */
 258     default:
 259         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 260                         __func__, type);
 261         goto out_free_aiocb;
 262     }
 263     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
 264
 265     QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
 266     s->io_q.n++;
 267     if (!s->io_q.blocked &&
 268         (!s->io_q.plugged || s->io_q.n >= MAX_QUEUED_IO)) {
 269         ioq_submit(s);
 270     }
 271     return &laiocb->common;
 272
 273 out_free_aiocb:
 274     qemu_aio_unref(laiocb);
 275     return NULL;
 276 }
 277
 278 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
 279 {
 280     aio_set_event_notifier(old_context, &s->e, false, NULL);
 281     qemu_bh_delete(s->completion_bh);
 282 }
 283
 284 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
 285 {
 286     s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
 287     aio_set_event_notifier(new_context, &s->e, false,
 288                            qemu_laio_completion_cb);
 289 }
 290
 291 LinuxAioState *laio_init(void)
 292 {
 293     LinuxAioState *s;
 294
 295     s = g_malloc0(sizeof(*s));
 296     if (event_notifier_init(&s->e, false) < 0) {
 297         goto out_free_state;
 298     }
 299
 300     if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
 301         goto out_close_efd;
 302     }
 303
 304     ioq_init(&s->io_q);
 305
 306     return s;
 307
 308 out_close_efd:
 309     event_notifier_cleanup(&s->e);
 310 out_free_state:
 311     g_free(s);
 312     return NULL;
 313 }
 314
 315 void laio_cleanup(LinuxAioState *s)
 316 {
 317     event_notifier_cleanup(&s->e);
 318
 319     if (io_destroy(s->ctx) != 0) {
 320         fprintf(stderr, "%s: destroy AIO context %p failed\n",
 321                         __func__, &s->ctx);
 322     }
 323     g_free(s);
 324 }