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 struct qemu_laiocb {
  32     BlockAIOCB common;
  33     Coroutine *co;
  34     LinuxAioState *ctx;
  35     struct iocb iocb;
  36     ssize_t ret;
  37     size_t nbytes;
  38     QEMUIOVector *qiov;
  39     bool is_read;
  40     QSIMPLEQ_ENTRY(qemu_laiocb) next;
  41 };
  42
  43 typedef struct {
  44     int plugged;
  45     unsigned int in_queue;
  46     unsigned int in_flight;
  47     bool blocked;
  48     QSIMPLEQ_HEAD(, qemu_laiocb) pending;
  49 } LaioQueue;
  50
  51 struct LinuxAioState {
  52     AioContext *aio_context;
  53
  54     io_context_t ctx;
  55     EventNotifier e;
  56
  57     /* io queue for submit at batch */
  58     LaioQueue io_q;
  59
  60     /* I/O completion processing */
  61     QEMUBH *completion_bh;
  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 = -ENOSPC;
  91             }
  92         }
  93     }
  94
  95     laiocb->ret = ret;
  96     if (laiocb->co) {
  97         /* Jump and continue completion for foreign requests, don't do
  98          * anything for current request, it will be completed shortly. */
  99         if (laiocb->co != qemu_coroutine_self()) {
 100             qemu_coroutine_enter(laiocb->co);
 101         }
 102     } else {
 103         laiocb->common.cb(laiocb->common.opaque, ret);
 104         qemu_aio_unref(laiocb);
 105     }
 106 }
 107
 108 /**
 109  * aio_ring buffer which is shared between userspace and kernel.
 110  *
 111  * This copied from linux/fs/aio.c, common header does not exist
 112  * but AIO exists for ages so we assume ABI is stable.
 113  */
 114 struct aio_ring {
 115     unsigned    id;    /* kernel internal index number */
 116     unsigned    nr;    /* number of io_events */
 117     unsigned    head;  /* Written to by userland or by kernel. */
 118     unsigned    tail;
 119
 120     unsigned    magic;
 121     unsigned    compat_features;
 122     unsigned    incompat_features;
 123     unsigned    header_length;  /* size of aio_ring */
 124
 125     struct io_event io_events[0];
 126 };
 127
 128 /**
 129  * io_getevents_peek:
 130  * @ctx: AIO context
 131  * @events: pointer on events array, output value
 132
 133  * Returns the number of completed events and sets a pointer
 134  * on events array.  This function does not update the internal
 135  * ring buffer, only reads head and tail.  When @events has been
 136  * processed io_getevents_commit() must be called.
 137  */
 138 static inline unsigned int io_getevents_peek(io_context_t ctx,
 139                                              struct io_event **events)
 140 {
 141     struct aio_ring *ring = (struct aio_ring *)ctx;
 142     unsigned int head = ring->head, tail = ring->tail;
 143     unsigned int nr;
 144
 145     nr = tail >= head ? tail - head : ring->nr - head;
 146     *events = ring->io_events + head;
 147     /* To avoid speculative loads of s->events[i] before observing tail.
 148        Paired with smp_wmb() inside linux/fs/aio.c: aio_complete(). */
 149     smp_rmb();
 150
 151     return nr;
 152 }
 153
 154 /**
 155  * io_getevents_commit:
 156  * @ctx: AIO context
 157  * @nr: the number of events on which head should be advanced
 158  *
 159  * Advances head of a ring buffer.
 160  */
 161 static inline void io_getevents_commit(io_context_t ctx, unsigned int nr)
 162 {
 163     struct aio_ring *ring = (struct aio_ring *)ctx;
 164
 165     if (nr) {
 166         ring->head = (ring->head + nr) % ring->nr;
 167     }
 168 }
 169
 170 /**
 171  * io_getevents_advance_and_peek:
 172  * @ctx: AIO context
 173  * @events: pointer on events array, output value
 174  * @nr: the number of events on which head should be advanced
 175  *
 176  * Advances head of a ring buffer and returns number of elements left.
 177  */
 178 static inline unsigned int
 179 io_getevents_advance_and_peek(io_context_t ctx,
 180                               struct io_event **events,
 181                               unsigned int nr)
 182 {
 183     io_getevents_commit(ctx, nr);
 184     return io_getevents_peek(ctx, events);
 185 }
 186
 187 /**
 188  * qemu_laio_process_completions:
 189  * @s: AIO state
 190  *
 191  * Fetches completed I/O requests and invokes their callbacks.
 192  *
 193  * The function is somewhat tricky because it supports nested event loops, for
 194  * example when a request callback invokes aio_poll().  In order to do this,
 195  * indices are kept in LinuxAioState.  Function schedules BH completion so it
 196  * can be called again in a nested event loop.  When there are no events left
 197  * to complete the BH is being canceled.
 198  */
 199 static void qemu_laio_process_completions(LinuxAioState *s)
 200 {
 201     struct io_event *events;
 202
 203     /* Reschedule so nested event loops see currently pending completions */
 204     qemu_bh_schedule(s->completion_bh);
 205
 206     while ((s->event_max = io_getevents_advance_and_peek(s->ctx, &events,
 207                                                          s->event_idx))) {
 208         for (s->event_idx = 0; s->event_idx < s->event_max; ) {
 209             struct iocb *iocb = events[s->event_idx].obj;
 210             struct qemu_laiocb *laiocb =
 211                 container_of(iocb, struct qemu_laiocb, iocb);
 212
 213             laiocb->ret = io_event_ret(&events[s->event_idx]);
 214
 215             /* Change counters one-by-one because we can be nested. */
 216             s->io_q.in_flight--;
 217             s->event_idx++;
 218             qemu_laio_process_completion(laiocb);
 219         }
 220     }
 221
 222     qemu_bh_cancel(s->completion_bh);
 223
 224     /* If we are nested we have to notify the level above that we are done
 225      * by setting event_max to zero, upper level will then jump out of it's
 226      * own `for` loop.  If we are the last all counters droped to zero. */
 227     s->event_max = 0;
 228     s->event_idx = 0;
 229 }
 230
 231 static void qemu_laio_process_completions_and_submit(LinuxAioState *s)
 232 {
 233     qemu_laio_process_completions(s);
 234     if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 235         ioq_submit(s);
 236     }
 237 }
 238
 239 static void qemu_laio_completion_bh(void *opaque)
 240 {
 241     LinuxAioState *s = opaque;
 242
 243     qemu_laio_process_completions_and_submit(s);
 244 }
 245
 246 static void qemu_laio_completion_cb(EventNotifier *e)
 247 {
 248     LinuxAioState *s = container_of(e, LinuxAioState, e);
 249
 250     if (event_notifier_test_and_clear(&s->e)) {
 251         qemu_laio_process_completions_and_submit(s);
 252     }
 253 }
 254
 255 static void laio_cancel(BlockAIOCB *blockacb)
 256 {
 257     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
 258     struct io_event event;
 259     int ret;
 260
 261     if (laiocb->ret != -EINPROGRESS) {
 262         return;
 263     }
 264     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
 265     laiocb->ret = -ECANCELED;
 266     if (ret != 0) {
 267         /* iocb is not cancelled, cb will be called by the event loop later */
 268         return;
 269     }
 270
 271     laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
 272 }
 273
 274 static const AIOCBInfo laio_aiocb_info = {
 275     .aiocb_size         = sizeof(struct qemu_laiocb),
 276     .cancel_async       = laio_cancel,
 277 };
 278
 279 static void ioq_init(LaioQueue *io_q)
 280 {
 281     QSIMPLEQ_INIT(&io_q->pending);
 282     io_q->plugged = 0;
 283     io_q->in_queue = 0;
 284     io_q->in_flight = 0;
 285     io_q->blocked = false;
 286 }
 287
 288 static void ioq_submit(LinuxAioState *s)
 289 {
 290     int ret, len;
 291     struct qemu_laiocb *aiocb;
 292     struct iocb *iocbs[MAX_EVENTS];
 293     QSIMPLEQ_HEAD(, qemu_laiocb) completed;
 294
 295     do {
 296         if (s->io_q.in_flight >= MAX_EVENTS) {
 297             break;
 298         }
 299         len = 0;
 300         QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
 301             iocbs[len++] = &aiocb->iocb;
 302             if (s->io_q.in_flight + len >= MAX_EVENTS) {
 303                 break;
 304             }
 305         }
 306
 307         ret = io_submit(s->ctx, len, iocbs);
 308         if (ret == -EAGAIN) {
 309             break;
 310         }
 311         if (ret < 0) {
 312             /* Fail the first request, retry the rest */
 313             aiocb = QSIMPLEQ_FIRST(&s->io_q.pending);
 314             QSIMPLEQ_REMOVE_HEAD(&s->io_q.pending, next);
 315             s->io_q.in_queue--;
 316             aiocb->ret = ret;
 317             qemu_laio_process_completion(aiocb);
 318             continue;
 319         }
 320
 321         s->io_q.in_flight += ret;
 322         s->io_q.in_queue  -= ret;
 323         aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
 324         QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
 325     } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
 326     s->io_q.blocked = (s->io_q.in_queue > 0);
 327
 328     if (s->io_q.in_flight) {
 329         /* We can try to complete something just right away if there are
 330          * still requests in-flight. */
 331         qemu_laio_process_completions(s);
 332         /*
 333          * Even we have completed everything (in_flight == 0), the queue can
 334          * have still pended requests (in_queue > 0).  We do not attempt to
 335          * repeat submission to avoid IO hang.  The reason is simple: s->e is
 336          * still set and completion callback will be called shortly and all
 337          * pended requests will be submitted from there.
 338          */
 339     }
 340 }
 341
 342 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
 343 {
 344     s->io_q.plugged++;
 345 }
 346
 347 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s)
 348 {
 349     assert(s->io_q.plugged);
 350     if (--s->io_q.plugged == 0 &&
 351         !s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
 352         ioq_submit(s);
 353     }
 354 }
 355
 356 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
 357                           int type)
 358 {
 359     LinuxAioState *s = laiocb->ctx;
 360     struct iocb *iocbs = &laiocb->iocb;
 361     QEMUIOVector *qiov = laiocb->qiov;
 362
 363     switch (type) {
 364     case QEMU_AIO_WRITE:
 365         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
 366         break;
 367     case QEMU_AIO_READ:
 368         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
 369         break;
 370     /* Currently Linux kernel does not support other operations */
 371     default:
 372         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
 373                         __func__, type);
 374         return -EIO;
 375     }
 376     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
 377
 378     QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
 379     s->io_q.in_queue++;
 380     if (!s->io_q.blocked &&
 381         (!s->io_q.plugged ||
 382          s->io_q.in_flight + s->io_q.in_queue >= MAX_EVENTS)) {
 383         ioq_submit(s);
 384     }
 385
 386     return 0;
 387 }
 388
 389 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
 390                                 uint64_t offset, QEMUIOVector *qiov, int type)
 391 {
 392     int ret;
 393     struct qemu_laiocb laiocb = {
 394         .co         = qemu_coroutine_self(),
 395         .nbytes     = qiov->size,
 396         .ctx        = s,
 397         .ret        = -EINPROGRESS,
 398         .is_read    = (type == QEMU_AIO_READ),
 399         .qiov       = qiov,
 400     };
 401
 402     ret = laio_do_submit(fd, &laiocb, offset, type);
 403     if (ret < 0) {
 404         return ret;
 405     }
 406
 407     if (laiocb.ret == -EINPROGRESS) {
 408         qemu_coroutine_yield();
 409     }
 410     return laiocb.ret;
 411 }
 412
 413 BlockAIOCB *laio_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
 414         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 415         BlockCompletionFunc *cb, void *opaque, int type)
 416 {
 417     struct qemu_laiocb *laiocb;
 418     off_t offset = sector_num * BDRV_SECTOR_SIZE;
 419     int ret;
 420
 421     laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
 422     laiocb->nbytes = nb_sectors * BDRV_SECTOR_SIZE;
 423     laiocb->ctx = s;
 424     laiocb->ret = -EINPROGRESS;
 425     laiocb->is_read = (type == QEMU_AIO_READ);
 426     laiocb->qiov = qiov;
 427
 428     ret = laio_do_submit(fd, laiocb, offset, type);
 429     if (ret < 0) {
 430         qemu_aio_unref(laiocb);
 431         return NULL;
 432     }
 433
 434     return &laiocb->common;
 435 }
 436
 437 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
 438 {
 439     aio_set_event_notifier(old_context, &s->e, false, NULL);
 440     qemu_bh_delete(s->completion_bh);
 441 }
 442
 443 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
 444 {
 445     s->aio_context = new_context;
 446     s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
 447     aio_set_event_notifier(new_context, &s->e, false,
 448                            qemu_laio_completion_cb);
 449 }
 450
 451 LinuxAioState *laio_init(void)
 452 {
 453     LinuxAioState *s;
 454
 455     s = g_malloc0(sizeof(*s));
 456     if (event_notifier_init(&s->e, false) < 0) {
 457         goto out_free_state;
 458     }
 459
 460     if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
 461         goto out_close_efd;
 462     }
 463
 464     ioq_init(&s->io_q);
 465
 466     return s;
 467
 468 out_close_efd:
 469     event_notifier_cleanup(&s->e);
 470 out_free_state:
 471     g_free(s);
 472     return NULL;
 473 }
 474
 475 void laio_cleanup(LinuxAioState *s)
 476 {
 477     event_notifier_cleanup(&s->e);
 478
 479     if (io_destroy(s->ctx) != 0) {
 480         fprintf(stderr, "%s: destroy AIO context %p failed\n",
 481                         __func__, &s->ctx);
 482     }
 483     g_free(s);
 484 }