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 static void do_spawn_thread(void);
  34
  35 struct qemu_paiocb {
  36     BlockDriverAIOCB common;
  37     int aio_fildes;
  38     union {
  39         struct iovec *aio_iov;
  40         void *aio_ioctl_buf;
  41     };
  42     int aio_niov;
  43     size_t aio_nbytes;
  44 #define aio_ioctl_cmd   aio_nbytes /* for QEMU_AIO_IOCTL */
  45     int ev_signo;
  46     off_t aio_offset;
  47
  48     QTAILQ_ENTRY(qemu_paiocb) node;
  49     int aio_type;
  50     ssize_t ret;
  51     int active;
  52     struct qemu_paiocb *next;
  53 };
  54
  55 typedef struct PosixAioState {
  56     int rfd, wfd;
  57     struct qemu_paiocb *first_aio;
  58 } PosixAioState;
  59
  60
  61 static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
  62 static pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
  63 static pthread_t thread_id;
  64 static pthread_attr_t attr;
  65 static int max_threads = 64;
  66 static int cur_threads = 0;
  67 static int idle_threads = 0;
  68 static int new_threads = 0;     /* backlog of threads we need to create */
  69 static int pending_threads = 0; /* threads created but not running yet */
  70 static QEMUBH *new_thread_bh;
  71 static QTAILQ_HEAD(, qemu_paiocb) request_list;
  72
  73 #ifdef CONFIG_PREADV
  74 static int preadv_present = 1;
  75 #else
  76 static int preadv_present = 0;
  77 #endif
  78
  79 static void die2(int err, const char *what)
  80 {
  81     fprintf(stderr, "%s failed: %s\n", what, strerror(err));
  82     abort();
  83 }
  84
  85 static void die(const char *what)
  86 {
  87     die2(errno, what);
  88 }
  89
  90 static void mutex_lock(pthread_mutex_t *mutex)
  91 {
  92     int ret = pthread_mutex_lock(mutex);
  93     if (ret) die2(ret, "pthread_mutex_lock");
  94 }
  95
  96 static void mutex_unlock(pthread_mutex_t *mutex)
  97 {
  98     int ret = pthread_mutex_unlock(mutex);
  99     if (ret) die2(ret, "pthread_mutex_unlock");
 100 }
 101
 102 static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex,
 103                            struct timespec *ts)
 104 {
 105     int ret = pthread_cond_timedwait(cond, mutex, ts);
 106     if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait");
 107     return ret;
 108 }
 109
 110 static void cond_signal(pthread_cond_t *cond)
 111 {
 112     int ret = pthread_cond_signal(cond);
 113     if (ret) die2(ret, "pthread_cond_signal");
 114 }
 115
 116 static void thread_create(pthread_t *thread, pthread_attr_t *attr,
 117                           void *(*start_routine)(void*), void *arg)
 118 {
 119     int ret = pthread_create(thread, attr, start_routine, arg);
 120     if (ret) die2(ret, "pthread_create");
 121 }
 122
 123 static ssize_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb)
 124 {
 125     int ret;
 126
 127     ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf);
 128     if (ret == -1)
 129         return -errno;
 130
 131     /*
 132      * This looks weird, but the aio code only consideres a request
 133      * successful if it has written the number full number of bytes.
 134      *
 135      * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
 136      * so in fact we return the ioctl command here to make posix_aio_read()
 137      * happy..
 138      */
 139     return aiocb->aio_nbytes;
 140 }
 141
 142 static ssize_t handle_aiocb_flush(struct qemu_paiocb *aiocb)
 143 {
 144     int ret;
 145
 146     ret = qemu_fdatasync(aiocb->aio_fildes);
 147     if (ret == -1)
 148         return -errno;
 149     return 0;
 150 }
 151
 152 #ifdef CONFIG_PREADV
 153
 154 static ssize_t
 155 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 156 {
 157     return preadv(fd, iov, nr_iov, offset);
 158 }
 159
 160 static ssize_t
 161 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 162 {
 163     return pwritev(fd, iov, nr_iov, offset);
 164 }
 165
 166 #else
 167
 168 static ssize_t
 169 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 170 {
 171     return -ENOSYS;
 172 }
 173
 174 static ssize_t
 175 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 176 {
 177     return -ENOSYS;
 178 }
 179
 180 #endif
 181
 182 static ssize_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb)
 183 {
 184     size_t offset = 0;
 185     ssize_t len;
 186
 187     do {
 188         if (aiocb->aio_type & QEMU_AIO_WRITE)
 189             len = qemu_pwritev(aiocb->aio_fildes,
 190                                aiocb->aio_iov,
 191                                aiocb->aio_niov,
 192                                aiocb->aio_offset + offset);
 193          else
 194             len = qemu_preadv(aiocb->aio_fildes,
 195                               aiocb->aio_iov,
 196                               aiocb->aio_niov,
 197                               aiocb->aio_offset + offset);
 198     } while (len == -1 && errno == EINTR);
 199
 200     if (len == -1)
 201         return -errno;
 202     return len;
 203 }
 204
 205 /*
 206  * Read/writes the data to/from a given linear buffer.
 207  *
 208  * Returns the number of bytes handles or -errno in case of an error. Short
 209  * reads are only returned if the end of the file is reached.
 210  */
 211 static ssize_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf)
 212 {
 213     ssize_t offset = 0;
 214     ssize_t len;
 215
 216     while (offset < aiocb->aio_nbytes) {
 217          if (aiocb->aio_type & QEMU_AIO_WRITE)
 218              len = pwrite(aiocb->aio_fildes,
 219                           (const char *)buf + offset,
 220                           aiocb->aio_nbytes - offset,
 221                           aiocb->aio_offset + offset);
 222          else
 223              len = pread(aiocb->aio_fildes,
 224                          buf + offset,
 225                          aiocb->aio_nbytes - offset,
 226                          aiocb->aio_offset + offset);
 227
 228          if (len == -1 && errno == EINTR)
 229              continue;
 230          else if (len == -1) {
 231              offset = -errno;
 232              break;
 233          } else if (len == 0)
 234              break;
 235
 236          offset += len;
 237     }
 238
 239     return offset;
 240 }
 241
 242 static ssize_t handle_aiocb_rw(struct qemu_paiocb *aiocb)
 243 {
 244     ssize_t nbytes;
 245     char *buf;
 246
 247     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
 248         /*
 249          * If there is just a single buffer, and it is properly aligned
 250          * we can just use plain pread/pwrite without any problems.
 251          */
 252         if (aiocb->aio_niov == 1)
 253              return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
 254
 255         /*
 256          * We have more than one iovec, and all are properly aligned.
 257          *
 258          * Try preadv/pwritev first and fall back to linearizing the
 259          * buffer if it's not supported.
 260          */
 261         if (preadv_present) {
 262             nbytes = handle_aiocb_rw_vector(aiocb);
 263             if (nbytes == aiocb->aio_nbytes)
 264                 return nbytes;
 265             if (nbytes < 0 && nbytes != -ENOSYS)
 266                 return nbytes;
 267             preadv_present = 0;
 268         }
 269
 270         /*
 271          * XXX(hch): short read/write.  no easy way to handle the reminder
 272          * using these interfaces.  For now retry using plain
 273          * pread/pwrite?
 274          */
 275     }
 276
 277     /*
 278      * Ok, we have to do it the hard way, copy all segments into
 279      * a single aligned buffer.
 280      */
 281     buf = qemu_blockalign(aiocb->common.bs, aiocb->aio_nbytes);
 282     if (aiocb->aio_type & QEMU_AIO_WRITE) {
 283         char *p = buf;
 284         int i;
 285
 286         for (i = 0; i < aiocb->aio_niov; ++i) {
 287             memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len);
 288             p += aiocb->aio_iov[i].iov_len;
 289         }
 290     }
 291
 292     nbytes = handle_aiocb_rw_linear(aiocb, buf);
 293     if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
 294         char *p = buf;
 295         size_t count = aiocb->aio_nbytes, copy;
 296         int i;
 297
 298         for (i = 0; i < aiocb->aio_niov && count; ++i) {
 299             copy = count;
 300             if (copy > aiocb->aio_iov[i].iov_len)
 301                 copy = aiocb->aio_iov[i].iov_len;
 302             memcpy(aiocb->aio_iov[i].iov_base, p, copy);
 303             p     += copy;
 304             count -= copy;
 305         }
 306     }
 307     qemu_vfree(buf);
 308
 309     return nbytes;
 310 }
 311
 312 static void *aio_thread(void *unused)
 313 {
 314     pid_t pid;
 315
 316     pid = getpid();
 317
 318     mutex_lock(&lock);
 319     pending_threads--;
 320     mutex_unlock(&lock);
 321     do_spawn_thread();
 322
 323     while (1) {
 324         struct qemu_paiocb *aiocb;
 325         ssize_t ret = 0;
 326         qemu_timeval tv;
 327         struct timespec ts;
 328
 329         qemu_gettimeofday(&tv);
 330         ts.tv_sec = tv.tv_sec + 10;
 331         ts.tv_nsec = 0;
 332
 333         mutex_lock(&lock);
 334
 335         while (QTAILQ_EMPTY(&request_list) &&
 336                !(ret == ETIMEDOUT)) {
 337             idle_threads++;
 338             ret = cond_timedwait(&cond, &lock, &ts);
 339             idle_threads--;
 340         }
 341
 342         if (QTAILQ_EMPTY(&request_list))
 343             break;
 344
 345         aiocb = QTAILQ_FIRST(&request_list);
 346         QTAILQ_REMOVE(&request_list, aiocb, node);
 347         aiocb->active = 1;
 348         mutex_unlock(&lock);
 349
 350         switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
 351         case QEMU_AIO_READ:
 352             ret = handle_aiocb_rw(aiocb);
 353             if (ret >= 0 && ret < aiocb->aio_nbytes && aiocb->common.bs->growable) {
 354                 /* A short read means that we have reached EOF. Pad the buffer
 355                  * with zeros for bytes after EOF. */
 356                 QEMUIOVector qiov;
 357
 358                 qemu_iovec_init_external(&qiov, aiocb->aio_iov,
 359                                          aiocb->aio_niov);
 360                 qemu_iovec_memset_skip(&qiov, 0, aiocb->aio_nbytes - ret, ret);
 361
 362                 ret = aiocb->aio_nbytes;
 363             }
 364             break;
 365         case QEMU_AIO_WRITE:
 366             ret = handle_aiocb_rw(aiocb);
 367             break;
 368         case QEMU_AIO_FLUSH:
 369             ret = handle_aiocb_flush(aiocb);
 370             break;
 371         case QEMU_AIO_IOCTL:
 372             ret = handle_aiocb_ioctl(aiocb);
 373             break;
 374         default:
 375             fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
 376             ret = -EINVAL;
 377             break;
 378         }
 379
 380         mutex_lock(&lock);
 381         aiocb->ret = ret;
 382         mutex_unlock(&lock);
 383
 384         if (kill(pid, aiocb->ev_signo)) die("kill failed");
 385     }
 386
 387     cur_threads--;
 388     mutex_unlock(&lock);
 389
 390     return NULL;
 391 }
 392
 393 static void do_spawn_thread(void)
 394 {
 395     sigset_t set, oldset;
 396
 397     mutex_lock(&lock);
 398     if (!new_threads) {
 399         mutex_unlock(&lock);
 400         return;
 401     }
 402
 403     new_threads--;
 404     pending_threads++;
 405
 406     mutex_unlock(&lock);
 407
 408     /* block all signals */
 409     if (sigfillset(&set)) die("sigfillset");
 410     if (sigprocmask(SIG_SETMASK, &set, &oldset)) die("sigprocmask");
 411
 412     thread_create(&thread_id, &attr, aio_thread, NULL);
 413
 414     if (sigprocmask(SIG_SETMASK, &oldset, NULL)) die("sigprocmask restore");
 415 }
 416
 417 static void spawn_thread_bh_fn(void *opaque)
 418 {
 419     do_spawn_thread();
 420 }
 421
 422 static void spawn_thread(void)
 423 {
 424     cur_threads++;
 425     new_threads++;
 426     /* If there are threads being created, they will spawn new workers, so
 427      * we don't spend time creating many threads in a loop holding a mutex or
 428      * starving the current vcpu.
 429      *
 430      * If there are no idle threads, ask the main thread to create one, so we
 431      * inherit the correct affinity instead of the vcpu affinity.
 432      */
 433     if (!pending_threads) {
 434         qemu_bh_schedule(new_thread_bh);
 435     }
 436 }
 437
 438 static void qemu_paio_submit(struct qemu_paiocb *aiocb)
 439 {
 440     aiocb->ret = -EINPROGRESS;
 441     aiocb->active = 0;
 442     mutex_lock(&lock);
 443     if (idle_threads == 0 && cur_threads < max_threads)
 444         spawn_thread();
 445     QTAILQ_INSERT_TAIL(&request_list, aiocb, node);
 446     mutex_unlock(&lock);
 447     cond_signal(&cond);
 448 }
 449
 450 static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
 451 {
 452     ssize_t ret;
 453
 454     mutex_lock(&lock);
 455     ret = aiocb->ret;
 456     mutex_unlock(&lock);
 457
 458     return ret;
 459 }
 460
 461 static int qemu_paio_error(struct qemu_paiocb *aiocb)
 462 {
 463     ssize_t ret = qemu_paio_return(aiocb);
 464
 465     if (ret < 0)
 466         ret = -ret;
 467     else
 468         ret = 0;
 469
 470     return ret;
 471 }
 472
 473 static int posix_aio_process_queue(void *opaque)
 474 {
 475     PosixAioState *s = opaque;
 476     struct qemu_paiocb *acb, **pacb;
 477     int ret;
 478     int result = 0;
 479
 480     for(;;) {
 481         pacb = &s->first_aio;
 482         for(;;) {
 483             acb = *pacb;
 484             if (!acb)
 485                 return result;
 486
 487             ret = qemu_paio_error(acb);
 488             if (ret == ECANCELED) {
 489                 /* remove the request */
 490                 *pacb = acb->next;
 491                 qemu_aio_release(acb);
 492                 result = 1;
 493             } else if (ret != EINPROGRESS) {
 494                 /* end of aio */
 495                 if (ret == 0) {
 496                     ret = qemu_paio_return(acb);
 497                     if (ret == acb->aio_nbytes)
 498                         ret = 0;
 499                     else
 500                         ret = -EINVAL;
 501                 } else {
 502                     ret = -ret;
 503                 }
 504
 505                 trace_paio_complete(acb, acb->common.opaque, ret);
 506
 507                 /* remove the request */
 508                 *pacb = acb->next;
 509                 /* call the callback */
 510                 acb->common.cb(acb->common.opaque, ret);
 511                 qemu_aio_release(acb);
 512                 result = 1;
 513                 break;
 514             } else {
 515                 pacb = &acb->next;
 516             }
 517         }
 518     }
 519
 520     return result;
 521 }
 522
 523 static void posix_aio_read(void *opaque)
 524 {
 525     PosixAioState *s = opaque;
 526     ssize_t len;
 527
 528     /* read all bytes from signal pipe */
 529     for (;;) {
 530         char bytes[16];
 531
 532         len = read(s->rfd, bytes, sizeof(bytes));
 533         if (len == -1 && errno == EINTR)
 534             continue; /* try again */
 535         if (len == sizeof(bytes))
 536             continue; /* more to read */
 537         break;
 538     }
 539
 540     posix_aio_process_queue(s);
 541 }
 542
 543 static int posix_aio_flush(void *opaque)
 544 {
 545     PosixAioState *s = opaque;
 546     return !!s->first_aio;
 547 }
 548
 549 static PosixAioState *posix_aio_state;
 550
 551 static void aio_signal_handler(int signum)
 552 {
 553     if (posix_aio_state) {
 554         char byte = 0;
 555         ssize_t ret;
 556
 557         ret = write(posix_aio_state->wfd, &byte, sizeof(byte));
 558         if (ret < 0 && errno != EAGAIN)
 559             die("write()");
 560     }
 561
 562     qemu_service_io();
 563 }
 564
 565 static void paio_remove(struct qemu_paiocb *acb)
 566 {
 567     struct qemu_paiocb **pacb;
 568
 569     /* remove the callback from the queue */
 570     pacb = &posix_aio_state->first_aio;
 571     for(;;) {
 572         if (*pacb == NULL) {
 573             fprintf(stderr, "paio_remove: aio request not found!\n");
 574             break;
 575         } else if (*pacb == acb) {
 576             *pacb = acb->next;
 577             qemu_aio_release(acb);
 578             break;
 579         }
 580         pacb = &(*pacb)->next;
 581     }
 582 }
 583
 584 static void paio_cancel(BlockDriverAIOCB *blockacb)
 585 {
 586     struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb;
 587     int active = 0;
 588
 589     trace_paio_cancel(acb, acb->common.opaque);
 590
 591     mutex_lock(&lock);
 592     if (!acb->active) {
 593         QTAILQ_REMOVE(&request_list, acb, node);
 594         acb->ret = -ECANCELED;
 595     } else if (acb->ret == -EINPROGRESS) {
 596         active = 1;
 597     }
 598     mutex_unlock(&lock);
 599
 600     if (active) {
 601         /* fail safe: if the aio could not be canceled, we wait for
 602            it */
 603         while (qemu_paio_error(acb) == EINPROGRESS)
 604             ;
 605     }
 606
 607     paio_remove(acb);
 608 }
 609
 610 static AIOPool raw_aio_pool = {
 611     .aiocb_size         = sizeof(struct qemu_paiocb),
 612     .cancel             = paio_cancel,
 613 };
 614
 615 BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
 616         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 617         BlockDriverCompletionFunc *cb, void *opaque, int type)
 618 {
 619     struct qemu_paiocb *acb;
 620
 621     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 622     if (!acb)
 623         return NULL;
 624     acb->aio_type = type;
 625     acb->aio_fildes = fd;
 626     acb->ev_signo = SIGUSR2;
 627
 628     if (qiov) {
 629         acb->aio_iov = qiov->iov;
 630         acb->aio_niov = qiov->niov;
 631     }
 632     acb->aio_nbytes = nb_sectors * 512;
 633     acb->aio_offset = sector_num * 512;
 634
 635     acb->next = posix_aio_state->first_aio;
 636     posix_aio_state->first_aio = acb;
 637
 638     trace_paio_submit(acb, opaque, sector_num, nb_sectors, type);
 639     qemu_paio_submit(acb);
 640     return &acb->common;
 641 }
 642
 643 BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
 644         unsigned long int req, void *buf,
 645         BlockDriverCompletionFunc *cb, void *opaque)
 646 {
 647     struct qemu_paiocb *acb;
 648
 649     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 650     if (!acb)
 651         return NULL;
 652     acb->aio_type = QEMU_AIO_IOCTL;
 653     acb->aio_fildes = fd;
 654     acb->ev_signo = SIGUSR2;
 655     acb->aio_offset = 0;
 656     acb->aio_ioctl_buf = buf;
 657     acb->aio_ioctl_cmd = req;
 658
 659     acb->next = posix_aio_state->first_aio;
 660     posix_aio_state->first_aio = acb;
 661
 662     qemu_paio_submit(acb);
 663     return &acb->common;
 664 }
 665
 666 int paio_init(void)
 667 {
 668     struct sigaction act;
 669     PosixAioState *s;
 670     int fds[2];
 671     int ret;
 672
 673     if (posix_aio_state)
 674         return 0;
 675
 676     s = g_malloc(sizeof(PosixAioState));
 677
 678     sigfillset(&act.sa_mask);
 679     act.sa_flags = 0; /* do not restart syscalls to interrupt select() */
 680     act.sa_handler = aio_signal_handler;
 681     sigaction(SIGUSR2, &act, NULL);
 682
 683     s->first_aio = NULL;
 684     if (qemu_pipe(fds) == -1) {
 685         fprintf(stderr, "failed to create pipe\n");
 686         return -1;
 687     }
 688
 689     s->rfd = fds[0];
 690     s->wfd = fds[1];
 691
 692     fcntl(s->rfd, F_SETFL, O_NONBLOCK);
 693     fcntl(s->wfd, F_SETFL, O_NONBLOCK);
 694
 695     qemu_aio_set_fd_handler(s->rfd, posix_aio_read, NULL, posix_aio_flush,
 696         posix_aio_process_queue, s);
 697
 698     ret = pthread_attr_init(&attr);
 699     if (ret)
 700         die2(ret, "pthread_attr_init");
 701
 702     ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
 703     if (ret)
 704         die2(ret, "pthread_attr_setdetachstate");
 705
 706     QTAILQ_INIT(&request_list);
 707     new_thread_bh = qemu_bh_new(spawn_thread_bh_fn, NULL);
 708
 709     posix_aio_state = s;
 710     return 0;
 711 }