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 #include "compatfd.h"
  31
  32 #include "block/raw-posix-aio.h"
  33
  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     int async_context_id;
  55 };
  56
  57 typedef struct PosixAioState {
  58     int fd;
  59     struct qemu_paiocb *first_aio;
  60 } PosixAioState;
  61
  62
  63 static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
  64 static pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
  65 static pthread_t thread_id;
  66 static pthread_attr_t attr;
  67 static int max_threads = 64;
  68 static int cur_threads = 0;
  69 static int idle_threads = 0;
  70 static QTAILQ_HEAD(, qemu_paiocb) request_list;
  71
  72 #ifdef CONFIG_PREADV
  73 static int preadv_present = 1;
  74 #else
  75 static int preadv_present = 0;
  76 #endif
  77
  78 static void die2(int err, const char *what)
  79 {
  80     fprintf(stderr, "%s failed: %s\n", what, strerror(err));
  81     abort();
  82 }
  83
  84 static void die(const char *what)
  85 {
  86     die2(errno, what);
  87 }
  88
  89 static void mutex_lock(pthread_mutex_t *mutex)
  90 {
  91     int ret = pthread_mutex_lock(mutex);
  92     if (ret) die2(ret, "pthread_mutex_lock");
  93 }
  94
  95 static void mutex_unlock(pthread_mutex_t *mutex)
  96 {
  97     int ret = pthread_mutex_unlock(mutex);
  98     if (ret) die2(ret, "pthread_mutex_unlock");
  99 }
 100
 101 static int cond_timedwait(pthread_cond_t *cond, pthread_mutex_t *mutex,
 102                            struct timespec *ts)
 103 {
 104     int ret = pthread_cond_timedwait(cond, mutex, ts);
 105     if (ret && ret != ETIMEDOUT) die2(ret, "pthread_cond_timedwait");
 106     return ret;
 107 }
 108
 109 static void cond_signal(pthread_cond_t *cond)
 110 {
 111     int ret = pthread_cond_signal(cond);
 112     if (ret) die2(ret, "pthread_cond_signal");
 113 }
 114
 115 static void thread_create(pthread_t *thread, pthread_attr_t *attr,
 116                           void *(*start_routine)(void*), void *arg)
 117 {
 118     int ret = pthread_create(thread, attr, start_routine, arg);
 119     if (ret) die2(ret, "pthread_create");
 120 }
 121
 122 static ssize_t handle_aiocb_ioctl(struct qemu_paiocb *aiocb)
 123 {
 124     int ret;
 125
 126     ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf);
 127     if (ret == -1)
 128         return -errno;
 129
 130     /*
 131      * This looks weird, but the aio code only consideres a request
 132      * successful if it has written the number full number of bytes.
 133      *
 134      * Now we overload aio_nbytes as aio_ioctl_cmd for the ioctl command,
 135      * so in fact we return the ioctl command here to make posix_aio_read()
 136      * happy..
 137      */
 138     return aiocb->aio_nbytes;
 139 }
 140
 141 static ssize_t handle_aiocb_flush(struct qemu_paiocb *aiocb)
 142 {
 143     int ret;
 144
 145     ret = qemu_fdatasync(aiocb->aio_fildes);
 146     if (ret == -1)
 147         return -errno;
 148     return 0;
 149 }
 150
 151 #ifdef CONFIG_PREADV
 152
 153 static ssize_t
 154 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 155 {
 156     return preadv(fd, iov, nr_iov, offset);
 157 }
 158
 159 static ssize_t
 160 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 161 {
 162     return pwritev(fd, iov, nr_iov, offset);
 163 }
 164
 165 #else
 166
 167 static ssize_t
 168 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 169 {
 170     return -ENOSYS;
 171 }
 172
 173 static ssize_t
 174 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
 175 {
 176     return -ENOSYS;
 177 }
 178
 179 #endif
 180
 181 static ssize_t handle_aiocb_rw_vector(struct qemu_paiocb *aiocb)
 182 {
 183     size_t offset = 0;
 184     ssize_t len;
 185
 186     do {
 187         if (aiocb->aio_type & QEMU_AIO_WRITE)
 188             len = qemu_pwritev(aiocb->aio_fildes,
 189                                aiocb->aio_iov,
 190                                aiocb->aio_niov,
 191                                aiocb->aio_offset + offset);
 192          else
 193             len = qemu_preadv(aiocb->aio_fildes,
 194                               aiocb->aio_iov,
 195                               aiocb->aio_niov,
 196                               aiocb->aio_offset + offset);
 197     } while (len == -1 && errno == EINTR);
 198
 199     if (len == -1)
 200         return -errno;
 201     return len;
 202 }
 203
 204 static ssize_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf)
 205 {
 206     ssize_t offset = 0;
 207     ssize_t len;
 208
 209     while (offset < aiocb->aio_nbytes) {
 210          if (aiocb->aio_type & QEMU_AIO_WRITE)
 211              len = pwrite(aiocb->aio_fildes,
 212                           (const char *)buf + offset,
 213                           aiocb->aio_nbytes - offset,
 214                           aiocb->aio_offset + offset);
 215          else
 216              len = pread(aiocb->aio_fildes,
 217                          buf + offset,
 218                          aiocb->aio_nbytes - offset,
 219                          aiocb->aio_offset + offset);
 220
 221          if (len == -1 && errno == EINTR)
 222              continue;
 223          else if (len == -1) {
 224              offset = -errno;
 225              break;
 226          } else if (len == 0)
 227              break;
 228
 229          offset += len;
 230     }
 231
 232     return offset;
 233 }
 234
 235 static ssize_t handle_aiocb_rw(struct qemu_paiocb *aiocb)
 236 {
 237     ssize_t nbytes;
 238     char *buf;
 239
 240     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
 241         /*
 242          * If there is just a single buffer, and it is properly aligned
 243          * we can just use plain pread/pwrite without any problems.
 244          */
 245         if (aiocb->aio_niov == 1)
 246              return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
 247
 248         /*
 249          * We have more than one iovec, and all are properly aligned.
 250          *
 251          * Try preadv/pwritev first and fall back to linearizing the
 252          * buffer if it's not supported.
 253          */
 254         if (preadv_present) {
 255             nbytes = handle_aiocb_rw_vector(aiocb);
 256             if (nbytes == aiocb->aio_nbytes)
 257                 return nbytes;
 258             if (nbytes < 0 && nbytes != -ENOSYS)
 259                 return nbytes;
 260             preadv_present = 0;
 261         }
 262
 263         /*
 264          * XXX(hch): short read/write.  no easy way to handle the reminder
 265          * using these interfaces.  For now retry using plain
 266          * pread/pwrite?
 267          */
 268     }
 269
 270     /*
 271      * Ok, we have to do it the hard way, copy all segments into
 272      * a single aligned buffer.
 273      */
 274     buf = qemu_blockalign(aiocb->common.bs, aiocb->aio_nbytes);
 275     if (aiocb->aio_type & QEMU_AIO_WRITE) {
 276         char *p = buf;
 277         int i;
 278
 279         for (i = 0; i < aiocb->aio_niov; ++i) {
 280             memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len);
 281             p += aiocb->aio_iov[i].iov_len;
 282         }
 283     }
 284
 285     nbytes = handle_aiocb_rw_linear(aiocb, buf);
 286     if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
 287         char *p = buf;
 288         size_t count = aiocb->aio_nbytes, copy;
 289         int i;
 290
 291         for (i = 0; i < aiocb->aio_niov && count; ++i) {
 292             copy = count;
 293             if (copy > aiocb->aio_iov[i].iov_len)
 294                 copy = aiocb->aio_iov[i].iov_len;
 295             memcpy(aiocb->aio_iov[i].iov_base, p, copy);
 296             p     += copy;
 297             count -= copy;
 298         }
 299     }
 300     qemu_vfree(buf);
 301
 302     return nbytes;
 303 }
 304
 305 static void *aio_thread(void *unused)
 306 {
 307     pid_t pid;
 308
 309     pid = getpid();
 310
 311     while (1) {
 312         struct qemu_paiocb *aiocb;
 313         ssize_t ret = 0;
 314         qemu_timeval tv;
 315         struct timespec ts;
 316
 317         qemu_gettimeofday(&tv);
 318         ts.tv_sec = tv.tv_sec + 10;
 319         ts.tv_nsec = 0;
 320
 321         mutex_lock(&lock);
 322
 323         while (QTAILQ_EMPTY(&request_list) &&
 324                !(ret == ETIMEDOUT)) {
 325             idle_threads++;
 326             ret = cond_timedwait(&cond, &lock, &ts);
 327             idle_threads--;
 328         }
 329
 330         if (QTAILQ_EMPTY(&request_list))
 331             break;
 332
 333         aiocb = QTAILQ_FIRST(&request_list);
 334         QTAILQ_REMOVE(&request_list, aiocb, node);
 335         aiocb->active = 1;
 336         mutex_unlock(&lock);
 337
 338         switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
 339         case QEMU_AIO_READ:
 340         case QEMU_AIO_WRITE:
 341             ret = handle_aiocb_rw(aiocb);
 342             break;
 343         case QEMU_AIO_FLUSH:
 344             ret = handle_aiocb_flush(aiocb);
 345             break;
 346         case QEMU_AIO_IOCTL:
 347             ret = handle_aiocb_ioctl(aiocb);
 348             break;
 349         default:
 350             fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
 351             ret = -EINVAL;
 352             break;
 353         }
 354
 355         mutex_lock(&lock);
 356         aiocb->ret = ret;
 357         mutex_unlock(&lock);
 358
 359         if (kill(pid, aiocb->ev_signo)) die("kill failed");
 360     }
 361
 362     cur_threads--;
 363     mutex_unlock(&lock);
 364
 365     return NULL;
 366 }
 367
 368 static void spawn_thread(void)
 369 {
 370     sigset_t set, oldset;
 371
 372     cur_threads++;
 373
 374     /* block all signals */
 375     if (sigfillset(&set)) die("sigfillset");
 376     if (sigprocmask(SIG_SETMASK, &set, &oldset)) die("sigprocmask");
 377
 378     thread_create(&thread_id, &attr, aio_thread, NULL);
 379
 380     if (sigprocmask(SIG_SETMASK, &oldset, NULL)) die("sigprocmask restore");
 381 }
 382
 383 static void qemu_paio_submit(struct qemu_paiocb *aiocb)
 384 {
 385     aiocb->ret = -EINPROGRESS;
 386     aiocb->active = 0;
 387     mutex_lock(&lock);
 388     if (idle_threads == 0 && cur_threads < max_threads)
 389         spawn_thread();
 390     QTAILQ_INSERT_TAIL(&request_list, aiocb, node);
 391     mutex_unlock(&lock);
 392     cond_signal(&cond);
 393 }
 394
 395 static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
 396 {
 397     ssize_t ret;
 398
 399     mutex_lock(&lock);
 400     ret = aiocb->ret;
 401     mutex_unlock(&lock);
 402
 403     return ret;
 404 }
 405
 406 static int qemu_paio_error(struct qemu_paiocb *aiocb)
 407 {
 408     ssize_t ret = qemu_paio_return(aiocb);
 409
 410     if (ret < 0)
 411         ret = -ret;
 412     else
 413         ret = 0;
 414
 415     return ret;
 416 }
 417
 418 static int posix_aio_process_queue(void *opaque)
 419 {
 420     PosixAioState *s = opaque;
 421     struct qemu_paiocb *acb, **pacb;
 422     int ret;
 423     int result = 0;
 424     int async_context_id = get_async_context_id();
 425
 426     for(;;) {
 427         pacb = &s->first_aio;
 428         for(;;) {
 429             acb = *pacb;
 430             if (!acb)
 431                 return result;
 432
 433             /* we're only interested in requests in the right context */
 434             if (acb->async_context_id != async_context_id) {
 435                 pacb = &acb->next;
 436                 continue;
 437             }
 438
 439             ret = qemu_paio_error(acb);
 440             if (ret == ECANCELED) {
 441                 /* remove the request */
 442                 *pacb = acb->next;
 443                 qemu_aio_release(acb);
 444                 result = 1;
 445             } else if (ret != EINPROGRESS) {
 446                 /* end of aio */
 447                 if (ret == 0) {
 448                     ret = qemu_paio_return(acb);
 449                     if (ret == acb->aio_nbytes)
 450                         ret = 0;
 451                     else
 452                         ret = -EINVAL;
 453                 } else {
 454                     ret = -ret;
 455                 }
 456
 457                 trace_paio_complete(acb, acb->common.opaque, ret);
 458
 459                 /* remove the request */
 460                 *pacb = acb->next;
 461                 /* call the callback */
 462                 acb->common.cb(acb->common.opaque, ret);
 463                 qemu_aio_release(acb);
 464                 result = 1;
 465                 break;
 466             } else {
 467                 pacb = &acb->next;
 468             }
 469         }
 470     }
 471
 472     return result;
 473 }
 474
 475 static void posix_aio_read(void *opaque)
 476 {
 477     PosixAioState *s = opaque;
 478     union {
 479         struct qemu_signalfd_siginfo siginfo;
 480         char buf[128];
 481     } sig;
 482     size_t offset;
 483
 484     /* try to read from signalfd, don't freak out if we can't read anything */
 485     offset = 0;
 486     while (offset < 128) {
 487         ssize_t len;
 488
 489         len = read(s->fd, sig.buf + offset, 128 - offset);
 490         if (len == -1 && errno == EINTR)
 491             continue;
 492         if (len == -1 && errno == EAGAIN) {
 493             /* there is no natural reason for this to happen,
 494              * so we'll spin hard until we get everything just
 495              * to be on the safe side. */
 496             if (offset > 0)
 497                 continue;
 498         }
 499
 500         offset += len;
 501     }
 502
 503     posix_aio_process_queue(s);
 504 }
 505
 506 static int posix_aio_flush(void *opaque)
 507 {
 508     PosixAioState *s = opaque;
 509     return !!s->first_aio;
 510 }
 511
 512 static PosixAioState *posix_aio_state;
 513
 514 static void paio_remove(struct qemu_paiocb *acb)
 515 {
 516     struct qemu_paiocb **pacb;
 517
 518     /* remove the callback from the queue */
 519     pacb = &posix_aio_state->first_aio;
 520     for(;;) {
 521         if (*pacb == NULL) {
 522             fprintf(stderr, "paio_remove: aio request not found!\n");
 523             break;
 524         } else if (*pacb == acb) {
 525             *pacb = acb->next;
 526             qemu_aio_release(acb);
 527             break;
 528         }
 529         pacb = &(*pacb)->next;
 530     }
 531 }
 532
 533 static void paio_cancel(BlockDriverAIOCB *blockacb)
 534 {
 535     struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb;
 536     int active = 0;
 537
 538     trace_paio_cancel(acb, acb->common.opaque);
 539
 540     mutex_lock(&lock);
 541     if (!acb->active) {
 542         QTAILQ_REMOVE(&request_list, acb, node);
 543         acb->ret = -ECANCELED;
 544     } else if (acb->ret == -EINPROGRESS) {
 545         active = 1;
 546     }
 547     mutex_unlock(&lock);
 548
 549     if (active) {
 550         /* fail safe: if the aio could not be canceled, we wait for
 551            it */
 552         while (qemu_paio_error(acb) == EINPROGRESS)
 553             ;
 554     }
 555
 556     paio_remove(acb);
 557 }
 558
 559 static AIOPool raw_aio_pool = {
 560     .aiocb_size         = sizeof(struct qemu_paiocb),
 561     .cancel             = paio_cancel,
 562 };
 563
 564 BlockDriverAIOCB *paio_submit(BlockDriverState *bs, int fd,
 565         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 566         BlockDriverCompletionFunc *cb, void *opaque, int type)
 567 {
 568     struct qemu_paiocb *acb;
 569
 570     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 571     if (!acb)
 572         return NULL;
 573     acb->aio_type = type;
 574     acb->aio_fildes = fd;
 575     acb->ev_signo = SIGUSR2;
 576     acb->async_context_id = get_async_context_id();
 577
 578     if (qiov) {
 579         acb->aio_iov = qiov->iov;
 580         acb->aio_niov = qiov->niov;
 581     }
 582     acb->aio_nbytes = nb_sectors * 512;
 583     acb->aio_offset = sector_num * 512;
 584
 585     acb->next = posix_aio_state->first_aio;
 586     posix_aio_state->first_aio = acb;
 587
 588     trace_paio_submit(acb, opaque, sector_num, nb_sectors, type);
 589     qemu_paio_submit(acb);
 590     return &acb->common;
 591 }
 592
 593 BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
 594         unsigned long int req, void *buf,
 595         BlockDriverCompletionFunc *cb, void *opaque)
 596 {
 597     struct qemu_paiocb *acb;
 598
 599     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 600     if (!acb)
 601         return NULL;
 602     acb->aio_type = QEMU_AIO_IOCTL;
 603     acb->aio_fildes = fd;
 604     acb->ev_signo = SIGUSR2;
 605     acb->async_context_id = get_async_context_id();
 606     acb->aio_offset = 0;
 607     acb->aio_ioctl_buf = buf;
 608     acb->aio_ioctl_cmd = req;
 609
 610     acb->next = posix_aio_state->first_aio;
 611     posix_aio_state->first_aio = acb;
 612
 613     qemu_paio_submit(acb);
 614     return &acb->common;
 615 }
 616
 617 int paio_init(void)
 618 {
 619     sigset_t mask;
 620     PosixAioState *s;
 621     int ret;
 622
 623     if (posix_aio_state)
 624         return 0;
 625
 626     s = qemu_malloc(sizeof(PosixAioState));
 627
 628     /* Make sure to block AIO signal */
 629     sigemptyset(&mask);
 630     sigaddset(&mask, SIGUSR2);
 631     sigprocmask(SIG_BLOCK, &mask, NULL);
 632
 633     s->first_aio = NULL;
 634     s->fd = qemu_signalfd(&mask);
 635     if (s->fd == -1) {
 636         fprintf(stderr, "failed to create signalfd\n");
 637         return -1;
 638     }
 639
 640     fcntl(s->fd, F_SETFL, O_NONBLOCK);
 641
 642     qemu_aio_set_fd_handler(s->fd, posix_aio_read, NULL, posix_aio_flush,
 643         posix_aio_process_queue, s);
 644
 645     ret = pthread_attr_init(&attr);
 646     if (ret)
 647         die2(ret, "pthread_attr_init");
 648
 649     ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
 650     if (ret)
 651         die2(ret, "pthread_attr_setdetachstate");
 652
 653     QTAILQ_INIT(&request_list);
 654
 655     posix_aio_state = s;
 656     return 0;
 657 }