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 "block/aio.h"
  12 #include "qemu/queue.h"
  13 #include "block/raw-aio.h"
  14 #include "qemu/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 #define MAX_QUEUED_IO  128
  29
  30 struct qemu_laiocb {
  31     BlockDriverAIOCB common;
  32     struct qemu_laio_state *ctx;
  33     struct iocb iocb;
  34     ssize_t ret;
  35     size_t nbytes;
  36     QEMUIOVector *qiov;
  37     bool is_read;
  38     QLIST_ENTRY(qemu_laiocb) node;
  39 };
  40
  41 typedef struct {
  42     struct iocb *iocbs[MAX_QUEUED_IO];
  43     int plugged;
  44     unsigned int size;
  45     unsigned int idx;
  46 } LaioQueue;
  47
  48 struct qemu_laio_state {
  49     io_context_t ctx;
  50     EventNotifier e;
  51
  52     /* io queue for submit at batch */
  53     LaioQueue io_q;
  54
  55     /* I/O completion processing */
  56     QEMUBH *completion_bh;
  57     struct io_event events[MAX_EVENTS];
  58     int event_idx;
  59     int event_max;
  60 };
  61
  62 static inline ssize_t io_event_ret(struct io_event *ev)
  63 {
  64     return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
  65 }
  66
  67 /*
  68  * Completes an AIO request (calls the callback and frees the ACB).
  69  */
  70 static void qemu_laio_process_completion(struct qemu_laio_state *s,
  71     struct qemu_laiocb *laiocb)
  72 {
  73     int ret;
  74
  75     ret = laiocb->ret;
  76     if (ret != -ECANCELED) {
  77         if (ret == laiocb->nbytes) {
  78             ret = 0;
  79         } else if (ret >= 0) {
  80             /* Short reads mean EOF, pad with zeros. */
  81             if (laiocb->is_read) {
  82                 qemu_iovec_memset(laiocb->qiov, ret, 0,
  83                     laiocb->qiov->size - ret);
  84             } else {
  85                 ret = -EINVAL;
  86             }
  87         }
  88
  89         laiocb->common.cb(laiocb->common.opaque, ret);
  90     }
  91
  92     qemu_aio_release(laiocb);
  93 }
  94
  95 /* The completion BH fetches completed I/O requests and invokes their
  96  * callbacks.
  97  *
  98  * The function is somewhat tricky because it supports nested event loops, for
  99  * example when a request callback invokes aio_poll().  In order to do this,
 100  * the completion events array and index are kept in qemu_laio_state.  The BH
 101  * reschedules itself as long as there are completions pending so it will
 102  * either be called again in a nested event loop or will be called after all
 103  * events have been completed.  When there are no events left to complete, the
 104  * BH returns without rescheduling.
 105  */
 106 static void qemu_laio_completion_bh(void *opaque)
 107 {
 108     struct qemu_laio_state *s = opaque;
 109
 110     /* Fetch more completion events when empty */
 111     if (s->event_idx == s->event_max) {
 112         do {
 113             struct timespec ts = { 0 };
 114             s->event_max = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS,
 115                                         s->events, &ts);
 116         } while (s->event_max == -EINTR);
 117
 118         s->event_idx = 0;
 119         if (s->event_max <= 0) {
 120             s->event_max = 0;
 121             return; /* no more events */
 122         }
 123     }
 124
 125     /* Reschedule so nested event loops see currently pending completions */
 126     qemu_bh_schedule(s->completion_bh);
 127
 128     /* Process completion events */
 129     while (s->event_idx < s->event_max) {
 130         struct iocb *iocb = s->events[s->event_idx].obj;
 131         struct qemu_laiocb *laiocb =
 132                 container_of(iocb, struct qemu_laiocb, iocb);
 133
 134         laiocb->ret = io_event_ret(&s->events[s->event_idx]);
 135         s->event_idx++;
 136
 137         qemu_laio_process_completion(s, laiocb);
 138     }
 139 }
 140
 141 static void qemu_laio_completion_cb(EventNotifier *e)
 142 {
 143     struct qemu_laio_state *s = container_of(e, struct qemu_laio_state, e);
 144
 145     if (event_notifier_test_and_clear(&s->e)) {
 146         qemu_bh_schedule(s->completion_bh);
 147     }
 148 }
 149
 150 static void laio_cancel(BlockDriverAIOCB *blockacb)
 151 {
 152     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 153     struct io_event event;
 154     int ret;
 155
 156     if (laiocb->ret != -EINPROGRESS)
 157         return;
 158
 159     /*
 160      * Note that as of Linux 2.6.31 neither the block device code nor any
 161      * filesystem implements cancellation of AIO request.
 162      * Thus the polling loop below is the normal code path.
 163      */
 164     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 165     if (ret == 0) {
 166         laiocb->ret = -ECANCELED;
 167         return;
 168     }
 169
 170     /*
 171      * We have to wait for the iocb to finish.
 172      *
 173      * The only way to get the iocb status update is by polling the io context.
 174      * We might be able to do this slightly more optimal by removing the
 175      * O_NONBLOCK flag.
 176      */
 177     while (laiocb->ret == -EINPROGRESS) {
 178         qemu_laio_completion_cb(&laiocb->ctx->e);
 179     }
 180 }
 181
 182 static const AIOCBInfo laio_aiocb_info = {
 183     .aiocb_size         = sizeof(struct qemu_laiocb),
 184     .cancel             = laio_cancel,
 185 };
 186
 187 static void ioq_init(LaioQueue *io_q)
 188 {
 189     io_q->size = MAX_QUEUED_IO;
 190     io_q->idx = 0;
 191     io_q->plugged = 0;
 192 }
 193
 194 static int ioq_submit(struct qemu_laio_state *s)
 195 {
 196     int ret, i = 0;
 197     int len = s->io_q.idx;
 198
 199     do {
 200         ret = io_submit(s->ctx, len, s->io_q.iocbs);
 201     } while (i++ < 3 && ret == -EAGAIN);
 202
 203     /* empty io queue */
 204     s->io_q.idx = 0;
 205
 206     if (ret < 0) {
 207         i = 0;
 208     } else {
 209         i = ret;
 210     }
 211
 212     for (; i < len; i++) {
 213         struct qemu_laiocb *laiocb =
 214             container_of(s->io_q.iocbs[i], struct qemu_laiocb, iocb);
 215
 216         laiocb->ret = (ret < 0) ? ret : -EIO;
 217         qemu_laio_process_completion(s, laiocb);
 218     }
 219     return ret;
 220 }
 221
 222 static void ioq_enqueue(struct qemu_laio_state *s, struct iocb *iocb)
 223 {
 224     unsigned int idx = s->io_q.idx;
 225
 226     s->io_q.iocbs[idx++] = iocb;
 227     s->io_q.idx = idx;
 228
 229     /* submit immediately if queue is full */
 230     if (idx == s->io_q.size) {
 231         ioq_submit(s);
 232     }
 233 }
 234
 235 void laio_io_plug(BlockDriverState *bs, void *aio_ctx)
 236 {
 237     struct qemu_laio_state *s = aio_ctx;
 238
 239     s->io_q.plugged++;
 240 }
 241
 242 int laio_io_unplug(BlockDriverState *bs, void *aio_ctx, bool unplug)
 243 {
 244     struct qemu_laio_state *s = aio_ctx;
 245     int ret = 0;
 246
 247     assert(s->io_q.plugged > 0 || !unplug);
 248
 249     if (unplug && --s->io_q.plugged > 0) {
 250         return 0;
 251     }
 252
 253     if (s->io_q.idx > 0) {
 254         ret = ioq_submit(s);
 255     }
 256
 257     return ret;
 258 }
 259
 260 BlockDriverAIOCB *laio_submit(BlockDriverState *bs, void *aio_ctx, int fd,
 261         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 262         BlockDriverCompletionFunc *cb, void *opaque, int type)
 263 {
 264     struct qemu_laio_state *s = aio_ctx;
 265     struct qemu_laiocb *laiocb;
 266     struct iocb *iocbs;
 267     off_t offset = sector_num * 512;
 268
 269     laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
 270     laiocb->nbytes = nb_sectors * 512;
 271     laiocb->ctx = s;
 272     laiocb->ret = -EINPROGRESS;
 273     laiocb->is_read = (type == QEMU_AIO_READ);
 274     laiocb->qiov = qiov;
 275
 276     iocbs = &laiocb->iocb;
 277
 278     switch (type) {
 279     case QEMU_AIO_WRITE:
 280         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 281         break;
 282     case QEMU_AIO_READ:
 283         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 284         break;
 285     /* Currently Linux kernel does not support other operations */
 286     default:
 287         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 288                         __func__, type);
 289         goto out_free_aiocb;
 290     }
 291     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
 292
 293     if (!s->io_q.plugged) {
 294         if (io_submit(s->ctx, 1, &iocbs) < 0) {
 295             goto out_free_aiocb;
 296         }
 297     } else {
 298         ioq_enqueue(s, iocbs);
 299     }
 300     return &laiocb->common;
 301
 302 out_free_aiocb:
 303     qemu_aio_release(laiocb);
 304     return NULL;
 305 }
 306
 307 void laio_detach_aio_context(void *s_, AioContext *old_context)
 308 {
 309     struct qemu_laio_state *s = s_;
 310
 311     aio_set_event_notifier(old_context, &s->e, NULL);
 312     qemu_bh_delete(s->completion_bh);
 313 }
 314
 315 void laio_attach_aio_context(void *s_, AioContext *new_context)
 316 {
 317     struct qemu_laio_state *s = s_;
 318
 319     s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
 320     aio_set_event_notifier(new_context, &s->e, qemu_laio_completion_cb);
 321 }
 322
 323 void *laio_init(void)
 324 {
 325     struct qemu_laio_state *s;
 326
 327     s = g_malloc0(sizeof(*s));
 328     if (event_notifier_init(&s->e, false) < 0) {
 329         goto out_free_state;
 330     }
 331
 332     if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
 333         goto out_close_efd;
 334     }
 335
 336     ioq_init(&s->io_q);
 337
 338     return s;
 339
 340 out_close_efd:
 341     event_notifier_cleanup(&s->e);
 342 out_free_state:
 343     g_free(s);
 344     return NULL;
 345 }
 346
 347 void laio_cleanup(void *s_)
 348 {
 349     struct qemu_laio_state *s = s_;
 350
 351     event_notifier_cleanup(&s->e);
 352
 353     if (io_destroy(s->ctx) != 0) {
 354         fprintf(stderr, "%s: destroy AIO context %p failed\n",
 355                         __func__, &s->ctx);
 356     }
 357     g_free(s);
 358 }