posix-aio-compat.c

   1 /*
   2  * QEMU posix-aio emulation
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *
   6  * Authors:
   7  *  Anthony Liguori   <aliguori@us.ibm.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  */
  13
  14 #include <sys/ioctl.h>
  15 #include <sys/types.h>
  16 #include <pthread.h>
  17 #include <unistd.h>
  18 #include <errno.h>
  19 #include <time.h>
  20 #include <string.h>
  21 #include <stdlib.h>
  22 #include <stdio.h>
  23
  24 #include "qemu-queue.h"
  25 #include "osdep.h"
  26 #include "sysemu.h"
  27 #include "qemu-common.h"
  28 #include "trace.h"
  29 #include "block_int.h"
  30
  31 #include "block/raw-posix-aio.h"
  32
  33
  34 struct qemu_paiocb {
  35     BlockDriverAIOCB common;
  36     int aio_fildes;
  37     union {
  38         struct iovec *aio_iov;
  39         void *aio_ioctl_buf;
  40     };
  41     int aio_niov;
  42     size_t aio_nbytes;
  43 #define aio_ioctl_cmd   aio_nbytes /* for QEMU_AIO_IOCTL */
  44     int ev_signo;
  45     off_t aio_offset;
  46
  47     QTAILQ_ENTRY(qemu_paiocb) node;
  48     int aio_type;
  49     ssize_t ret;
  50     int active;
  51     struct qemu_paiocb *next;
  52 };
  53
  54 typedef struct PosixAioState {
  55     int rfd, wfd;
  56     struct qemu_paiocb *first_aio;
  57 } PosixAioState;
  58
  59
  60 static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
  61 static pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
  62 static pthread_t thread_id;
  63 static pthread_attr_t attr;
  64 static int max_threads = 64;
  65 static int cur_threads = 0;
  66 static int idle_threads = 0;
  67 static QTAILQ_HEAD(, qemu_paiocb) request_list;
  68
  69 #ifdef CONFIG_PREADV
  70 static int preadv_present = 1;
  71 #else
  72 static int preadv_present = 0;
  73 #endif
  74
  75 static void die2(int err, const char *what)
  76 {
  77     fprintf(stderr, "%s failed: %s\n", what, strerror(err));
  78     abort();
  79 }
  80
  81 static void die(const char *what)
  82 {
  83     die2(errno, what);
  84 }
  85
  86 static void mutex_lock(pthread_mutex_t *mutex)
  87 {
  88     int ret = pthread_mutex_lock(mutex);
  89     if (ret) die2(ret, "pthread_mutex_lock");
  90 }
  91
  92 static void mutex_unlock(pthread_mutex_t *mutex)
  93 {
  94     int ret = pthread_mutex_unlock(mutex);
  95     if (ret) die2(ret, "pthread_mutex_unlock");
  96 }
  97
  98 static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex,
  99                            struct timespec *ts)
 100 {
 101     int ret = pthread_cond_timedwait(cond, mutex, ts);
 102     if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait");
 103     return ret;
 104 }
 105
 106 static void cond_signal(pthread_cond_t *cond)
 107 {
 108     int ret = pthread_cond_signal(cond);
 109     if (ret) die2(ret, "pthread_cond_signal");
 110 }
 111
 112 static void thread_create(pthread_t *thread, pthread_attr_t *attr,
 113                           void *(*start_routine)(void*), void *arg)
 114 {
 115     int ret = pthread_create(thread, attr, start_routine, arg);
 116     if (ret) die2(ret, "pthread_create");
 117 }
 118
 119 static ssize_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb)
 120 {
 121     int ret;
 122
 123     ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf);
 124     if (ret == -1)
 125         return -errno;
 126
 127     /*
 128      * This looks weird, but the aio code only consideres a request
 129      * successful if it has written the number full number of bytes.
 130      *
 131      * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
 132      * so in fact we return the ioctl command here to make posix_aio_read()
 133      * happy..
 134      */
 135     return aiocb->aio_nbytes;
 136 }
 137
 138 static ssize_t handle_aiocb_flush(struct qemu_paiocb *aiocb)
 139 {
 140     int ret;
 141
 142     ret = qemu_fdatasync(aiocb->aio_fildes);
 143     if (ret == -1)
 144         return -errno;
 145     return 0;
 146 }
 147
 148 #ifdef CONFIG_PREADV
 149
 150 static ssize_t
 151 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 152 {
 153     return preadv(fd, iov, nr_iov, offset);
 154 }
 155
 156 static ssize_t
 157 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 158 {
 159     return pwritev(fd, iov, nr_iov, offset);
 160 }
 161
 162 #else
 163
 164 static ssize_t
 165 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 166 {
 167     return -ENOSYS;
 168 }
 169
 170 static ssize_t
 171 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 172 {
 173     return -ENOSYS;
 174 }
 175
 176 #endif
 177
 178 static ssize_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb)
 179 {
 180     size_t offset = 0;
 181     ssize_t len;
 182
 183     do {
 184         if (aiocb->aio_type & QEMU_AIO_WRITE)
 185             len = qemu_pwritev(aiocb->aio_fildes,
 186                                aiocb->aio_iov,
 187                                aiocb->aio_niov,
 188                                aiocb->aio_offset + offset);
 189          else
 190             len = qemu_preadv(aiocb->aio_fildes,
 191                               aiocb->aio_iov,
 192                               aiocb->aio_niov,
 193                               aiocb->aio_offset + offset);
 194     } while (len == -1 && errno == EINTR);
 195
 196     if (len == -1)
 197         return -errno;
 198     return len;
 199 }
 200
 201 static ssize_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf)
 202 {
 203     ssize_t offset = 0;
 204     ssize_t len;
 205
 206     while (offset < aiocb->aio_nbytes) {
 207          if (aiocb->aio_type & QEMU_AIO_WRITE)
 208              len = pwrite(aiocb->aio_fildes,
 209                           (const char *)buf + offset,
 210                           aiocb->aio_nbytes - offset,
 211                           aiocb->aio_offset + offset);
 212          else
 213              len = pread(aiocb->aio_fildes,
 214                          buf + offset,
 215                          aiocb->aio_nbytes - offset,
 216                          aiocb->aio_offset + offset);
 217
 218          if (len == -1 && errno == EINTR)
 219              continue;
 220          else if (len == -1) {
 221              offset = -errno;
 222              break;
 223          } else if (len == 0)
 224              break;
 225
 226          offset += len;
 227     }
 228
 229     return offset;
 230 }
 231
 232 static ssize_t handle_aiocb_rw(struct qemu_paiocb *aiocb)
 233 {
 234     ssize_t nbytes;
 235     char *buf;
 236
 237     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
 238         /*
 239          * If there is just a single buffer, and it is properly aligned
 240          * we can just use plain pread/pwrite without any problems.
 241          */
 242         if (aiocb->aio_niov == 1)
 243              return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
 244
 245         /*
 246          * We have more than one iovec, and all are properly aligned.
 247          *
 248          * Try preadv/pwritev first and fall back to linearizing the
 249          * buffer if it's not supported.
 250          */
 251         if (preadv_present) {
 252             nbytes = handle_aiocb_rw_vector(aiocb);
 253             if (nbytes == aiocb->aio_nbytes)
 254                 return nbytes;
 255             if (nbytes < 0 && nbytes != -ENOSYS)
 256                 return nbytes;
 257             preadv_present = 0;
 258         }
 259
 260         /*
 261          * XXX(hch): short read/write.  no easy way to handle the reminder
 262          * using these interfaces.  For now retry using plain
 263          * pread/pwrite?
 264          */
 265     }
 266
 267     /*
 268      * Ok, we have to do it the hard way, copy all segments into
 269      * a single aligned buffer.
 270      */
 271     buf = qemu_blockalign(aiocb->common.bs, aiocb->aio_nbytes);
 272     if (aiocb->aio_type & QEMU_AIO_WRITE) {
 273         char *p = buf;
 274         int i;
 275
 276         for (i = 0; i < aiocb->aio_niov; ++i) {
 277             memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len);
 278             p += aiocb->aio_iov[i].iov_len;
 279         }
 280     }
 281
 282     nbytes = handle_aiocb_rw_linear(aiocb, buf);
 283     if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
 284         char *p = buf;
 285         size_t count = aiocb->aio_nbytes, copy;
 286         int i;
 287
 288         for (i = 0; i < aiocb->aio_niov && count; ++i) {
 289             copy = count;
 290             if (copy > aiocb->aio_iov[i].iov_len)
 291                 copy = aiocb->aio_iov[i].iov_len;
 292             memcpy(aiocb->aio_iov[i].iov_base, p, copy);
 293             p     += copy;
 294             count -= copy;
 295         }
 296     }
 297     qemu_vfree(buf);
 298
 299     return nbytes;
 300 }
 301
 302 static void *aio_thread(void *unused)
 303 {
 304     pid_t pid;
 305
 306     pid = getpid();
 307
 308     while (1) {
 309         struct qemu_paiocb *aiocb;
 310         ssize_t ret = 0;
 311         qemu_timeval tv;
 312         struct timespec ts;
 313
 314         qemu_gettimeofday(&tv);
 315         ts.tv_sec = tv.tv_sec + 10;
 316         ts.tv_nsec = 0;
 317
 318         mutex_lock(&lock);
 319
 320         while (QTAILQ_EMPTY(&request_list) &&
 321                !(ret == ETIMEDOUT)) {
 322             idle_threads++;
 323             ret = cond_timedwait(&cond, &lock, &ts);
 324             idle_threads--;
 325         }
 326
 327         if (QTAILQ_EMPTY(&request_list))
 328             break;
 329
 330         aiocb = QTAILQ_FIRST(&request_list);
 331         QTAILQ_REMOVE(&request_list, aiocb, node);
 332         aiocb->active = 1;
 333         mutex_unlock(&lock);
 334
 335         switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
 336         case QEMU_AIO_READ:
 337         case QEMU_AIO_WRITE:
 338             ret = handle_aiocb_rw(aiocb);
 339             break;
 340         case QEMU_AIO_FLUSH:
 341             ret = handle_aiocb_flush(aiocb);
 342             break;
 343         case QEMU_AIO_IOCTL:
 344             ret = handle_aiocb_ioctl(aiocb);
 345             break;
 346         default:
 347             fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
 348             ret = -EINVAL;
 349             break;
 350         }
 351
 352         mutex_lock(&lock);
 353         aiocb->ret = ret;
 354         mutex_unlock(&lock);
 355
 356         if (kill(pid, aiocb->ev_signo)) die("kill failed");
 357     }
 358
 359     cur_threads--;
 360     mutex_unlock(&lock);
 361
 362     return NULL;
 363 }
 364
 365 static void spawn_thread(void)
 366 {
 367     sigset_t set, oldset;
 368
 369     cur_threads++;
 370
 371     /* block all signals */
 372     if (sigfillset(&set)) die("sigfillset");
 373     if (sigprocmask(SIG_SETMASK, &set, &oldset)) die("sigprocmask");
 374
 375     thread_create(&thread_id, &attr, aio_thread, NULL);
 376
 377     if (sigprocmask(SIG_SETMASK, &oldset, NULL)) die("sigprocmask restore");
 378 }
 379
 380 static void qemu_paio_submit(struct qemu_paiocb *aiocb)
 381 {
 382     aiocb->ret = -EINPROGRESS;
 383     aiocb->active = 0;
 384     mutex_lock(&lock);
 385     if (idle_threads == 0 && cur_threads < max_threads)
 386         spawn_thread();
 387     QTAILQ_INSERT_TAIL(&request_list, aiocb, node);
 388     mutex_unlock(&lock);
 389     cond_signal(&cond);
 390 }
 391
 392 static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
 393 {
 394     ssize_t ret;
 395
 396     mutex_lock(&lock);
 397     ret = aiocb->ret;
 398     mutex_unlock(&lock);
 399
 400     return ret;
 401 }
 402
 403 static int qemu_paio_error(struct qemu_paiocb *aiocb)
 404 {
 405     ssize_t ret = qemu_paio_return(aiocb);
 406
 407     if (ret < 0)
 408         ret = -ret;
 409     else
 410         ret = 0;
 411
 412     return ret;
 413 }
 414
 415 static int posix_aio_process_queue(void *opaque)
 416 {
 417     PosixAioState *s = opaque;
 418     struct qemu_paiocb *acb, **pacb;
 419     int ret;
 420     int result = 0;
 421
 422     for(;;) {
 423         pacb = &s->first_aio;
 424         for(;;) {
 425             acb = *pacb;
 426             if (!acb)
 427                 return result;
 428
 429             ret = qemu_paio_error(acb);
 430             if (ret == ECANCELED) {
 431                 /* remove the request */
 432                 *pacb = acb->next;
 433                 qemu_aio_release(acb);
 434                 result = 1;
 435             } else if (ret != EINPROGRESS) {
 436                 /* end of aio */
 437                 if (ret == 0) {
 438                     ret = qemu_paio_return(acb);
 439                     if (ret == acb->aio_nbytes)
 440                         ret = 0;
 441                     else
 442                         ret = -EINVAL;
 443                 } else {
 444                     ret = -ret;
 445                 }
 446
 447                 trace_paio_complete(acb, acb->common.opaque, ret);
 448
 449                 /* remove the request */
 450                 *pacb = acb->next;
 451                 /* call the callback */
 452                 acb->common.cb(acb->common.opaque, ret);
 453                 qemu_aio_release(acb);
 454                 result = 1;
 455                 break;
 456             } else {
 457                 pacb = &acb->next;
 458             }
 459         }
 460     }
 461
 462     return result;
 463 }
 464
 465 static void posix_aio_read(void *opaque)
 466 {
 467     PosixAioState *s = opaque;
 468     ssize_t len;
 469
 470     /* read all bytes from signal pipe */
 471     for (;;) {
 472         char bytes[16];
 473
 474         len = read(s->rfd, bytes, sizeof(bytes));
 475         if (len == -1 && errno == EINTR)
 476             continue; /* try again */
 477         if (len == sizeof(bytes))
 478             continue; /* more to read */
 479         break;
 480     }
 481
 482     posix_aio_process_queue(s);
 483 }
 484
 485 static int posix_aio_flush(void *opaque)
 486 {
 487     PosixAioState *s = opaque;
 488     return !!s->first_aio;
 489 }
 490
 491 static PosixAioState *posix_aio_state;
 492
 493 static void aio_signal_handler(int signum)
 494 {
 495     if (posix_aio_state) {
 496         char byte = 0;
 497         ssize_t ret;
 498
 499         ret = write(posix_aio_state->wfd, &byte, sizeof(byte));
 500         if (ret < 0 && errno != EAGAIN)
 501             die("write()");
 502     }
 503
 504     qemu_service_io();
 505 }
 506
 507 static void paio_remove(struct qemu_paiocb *acb)
 508 {
 509     struct qemu_paiocb **pacb;
 510
 511     /* remove the callback from the queue */
 512     pacb = &posix_aio_state->first_aio;
 513     for(;;) {
 514         if (*pacb == NULL) {
 515             fprintf(stderr, "paio_remove: aio request not found!\n");
 516             break;
 517         } else if (*pacb == acb) {
 518             *pacb = acb->next;
 519             qemu_aio_release(acb);
 520             break;
 521         }
 522         pacb = &(*pacb)->next;
 523     }
 524 }
 525
 526 static void paio_cancel(BlockDriverAIOCB *blockacb)
 527 {
 528     struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb;
 529     int active = 0;
 530
 531     trace_paio_cancel(acb, acb->common.opaque);
 532
 533     mutex_lock(&lock);
 534     if (!acb->active) {
 535         QTAILQ_REMOVE(&request_list, acb, node);
 536         acb->ret = -ECANCELED;
 537     } else if (acb->ret == -EINPROGRESS) {
 538         active = 1;
 539     }
 540     mutex_unlock(&lock);
 541
 542     if (active) {
 543         /* fail safe: if the aio could not be canceled, we wait for
 544            it */
 545         while (qemu_paio_error(acb) == EINPROGRESS)
 546             ;
 547     }
 548
 549     paio_remove(acb);
 550 }
 551
 552 static AIOPool raw_aio_pool = {
 553     .aiocb_size         = sizeof(struct qemu_paiocb),
 554     .cancel             = paio_cancel,
 555 };
 556
 557 BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
 558         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 559         BlockDriverCompletionFunc *cb, void *opaque, int type)
 560 {
 561     struct qemu_paiocb *acb;
 562
 563     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 564     if (!acb)
 565         return NULL;
 566     acb->aio_type = type;
 567     acb->aio_fildes = fd;
 568     acb->ev_signo = SIGUSR2;
 569
 570     if (qiov) {
 571         acb->aio_iov = qiov->iov;
 572         acb->aio_niov = qiov->niov;
 573     }
 574     acb->aio_nbytes = nb_sectors * 512;
 575     acb->aio_offset = sector_num * 512;
 576
 577     acb->next = posix_aio_state->first_aio;
 578     posix_aio_state->first_aio = acb;
 579
 580     trace_paio_submit(acb, opaque, sector_num, nb_sectors, type);
 581     qemu_paio_submit(acb);
 582     return &acb->common;
 583 }
 584
 585 BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
 586         unsigned long int req, void *buf,
 587         BlockDriverCompletionFunc *cb, void *opaque)
 588 {
 589     struct qemu_paiocb *acb;
 590
 591     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 592     if (!acb)
 593         return NULL;
 594     acb->aio_type = QEMU_AIO_IOCTL;
 595     acb->aio_fildes = fd;
 596     acb->ev_signo = SIGUSR2;
 597     acb->aio_offset = 0;
 598     acb->aio_ioctl_buf = buf;
 599     acb->aio_ioctl_cmd = req;
 600
 601     acb->next = posix_aio_state->first_aio;
 602     posix_aio_state->first_aio = acb;
 603
 604     qemu_paio_submit(acb);
 605     return &acb->common;
 606 }
 607
 608 int paio_init(void)
 609 {
 610     struct sigaction act;
 611     PosixAioState *s;
 612     int fds[2];
 613     int ret;
 614
 615     if (posix_aio_state)
 616         return 0;
 617
 618     s = qemu_malloc(sizeof(PosixAioState));
 619
 620     sigfillset(&act.sa_mask);
 621     act.sa_flags = 0; /* do not restart syscalls to interrupt select() */
 622     act.sa_handler = aio_signal_handler;
 623     sigaction(SIGUSR2, &act, NULL);
 624
 625     s->first_aio = NULL;
 626     if (qemu_pipe(fds) == -1) {
 627         fprintf(stderr, "failed to create pipe\n");
 628         return -1;
 629     }
 630
 631     s->rfd = fds[0];
 632     s->wfd = fds[1];
 633
 634     fcntl(s->rfd, F_SETFL, O_NONBLOCK);
 635     fcntl(s->wfd, F_SETFL, O_NONBLOCK);
 636
 637     qemu_aio_set_fd_handler(s->rfd, posix_aio_read, NULL, posix_aio_flush,
 638         posix_aio_process_queue, s);
 639
 640     ret = pthread_attr_init(&attr);
 641     if (ret)
 642         die2(ret, "pthread_attr_init");
 643
 644     ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
 645     if (ret)
 646         die2(ret, "pthread_attr_setdetachstate");
 647
 648     QTAILQ_INIT(&request_list);
 649
 650     posix_aio_state = s;
 651     return 0;
 652 }