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