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 <signal.h>
  21 #include <string.h>
  22 #include <stdlib.h>
  23 #include <stdio.h>
  24
  25 #include "qemu-queue.h"
  26 #include "osdep.h"
  27 #include "qemu-common.h"
  28 #include "block_int.h"
  29 #include "compatfd.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 fd;
  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 size_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          * successfull 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 size_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 size_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 size_t handle_aiocb_rw_linear(struct qemu_paiocb *aiocb, char *buf)
 202 {
 203     size_t offset = 0;
 204     size_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 size_t handle_aiocb_rw(struct qemu_paiocb *aiocb)
 233 {
 234     size_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_memalign(512, 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     sigset_t set;
 306
 307     pid = getpid();
 308
 309     /* block all signals */
 310     if (sigfillset(&set)) die("sigfillset");
 311     if (sigprocmask(SIG_BLOCK, &set, NULL)) die("sigprocmask");
 312
 313     while (1) {
 314         struct qemu_paiocb *aiocb;
 315         size_t ret = 0;
 316         qemu_timeval tv;
 317         struct timespec ts;
 318
 319         qemu_gettimeofday(&tv);
 320         ts.tv_sec = tv.tv_sec + 10;
 321         ts.tv_nsec = 0;
 322
 323         mutex_lock(&lock);
 324
 325         while (QTAILQ_EMPTY(&request_list) &&
 326                !(ret == ETIMEDOUT)) {
 327             ret = cond_timedwait(&cond, &lock, &ts);
 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         idle_threads--;
 337         mutex_unlock(&lock);
 338
 339         switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
 340         case QEMU_AIO_READ:
 341         case QEMU_AIO_WRITE:
 342                 ret = handle_aiocb_rw(aiocb);
 343                 break;
 344         case QEMU_AIO_FLUSH:
 345                 ret = handle_aiocb_flush(aiocb);
 346                 break;
 347         case QEMU_AIO_IOCTL:
 348                 ret = handle_aiocb_ioctl(aiocb);
 349                 break;
 350         default:
 351                 fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
 352                 ret = -EINVAL;
 353                 break;
 354         }
 355
 356         mutex_lock(&lock);
 357         aiocb->ret = ret;
 358         idle_threads++;
 359         mutex_unlock(&lock);
 360
 361         if (kill(pid, aiocb->ev_signo)) die("kill failed");
 362     }
 363
 364     idle_threads--;
 365     cur_threads--;
 366     mutex_unlock(&lock);
 367
 368     return NULL;
 369 }
 370
 371 static void spawn_thread(void)
 372 {
 373     cur_threads++;
 374     idle_threads++;
 375     thread_create(&thread_id, &attr, aio_thread, NULL);
 376 }
 377
 378 static void qemu_paio_submit(struct qemu_paiocb *aiocb)
 379 {
 380     aiocb->ret = -EINPROGRESS;
 381     aiocb->active = 0;
 382     mutex_lock(&lock);
 383     if (idle_threads == 0 && cur_threads < max_threads)
 384         spawn_thread();
 385     QTAILQ_INSERT_TAIL(&request_list, aiocb, node);
 386     mutex_unlock(&lock);
 387     cond_signal(&cond);
 388 }
 389
 390 static ssize_t qemu_paio_return(struct qemu_paiocb *aiocb)
 391 {
 392     ssize_t ret;
 393
 394     mutex_lock(&lock);
 395     ret = aiocb->ret;
 396     mutex_unlock(&lock);
 397
 398     return ret;
 399 }
 400
 401 static int qemu_paio_error(struct qemu_paiocb *aiocb)
 402 {
 403     ssize_t ret = qemu_paio_return(aiocb);
 404
 405     if (ret < 0)
 406         ret = -ret;
 407     else
 408         ret = 0;
 409
 410     return ret;
 411 }
 412
 413 static void posix_aio_read(void *opaque)
 414 {
 415     PosixAioState *s = opaque;
 416     struct qemu_paiocb *acb, **pacb;
 417     int ret;
 418     union {
 419         struct qemu_signalfd_siginfo siginfo;
 420         char buf[128];
 421     } sig;
 422     size_t offset;
 423
 424     /* try to read from signalfd, don't freak out if we can't read anything */
 425     offset = 0;
 426     while (offset < 128) {
 427         ssize_t len;
 428
 429         len = read(s->fd, sig.buf + offset, 128 - offset);
 430         if (len == -1 && errno == EINTR)
 431             continue;
 432         if (len == -1 && errno == EAGAIN) {
 433             /* there is no natural reason for this to happen,
 434              * so we'll spin hard until we get everything just
 435              * to be on the safe side. */
 436             if (offset > 0)
 437                 continue;
 438         }
 439
 440         offset += len;
 441     }
 442
 443     for(;;) {
 444         pacb = &s->first_aio;
 445         for(;;) {
 446             acb = *pacb;
 447             if (!acb)
 448                 goto the_end;
 449             ret = qemu_paio_error(acb);
 450             if (ret == ECANCELED) {
 451                 /* remove the request */
 452                 *pacb = acb->next;
 453                 qemu_aio_release(acb);
 454             } else if (ret != EINPROGRESS) {
 455                 /* end of aio */
 456                 if (ret == 0) {
 457                     ret = qemu_paio_return(acb);
 458                     if (ret == acb->aio_nbytes)
 459                         ret = 0;
 460                     else
 461                         ret = -EINVAL;
 462                 } else {
 463                     ret = -ret;
 464                 }
 465                 /* remove the request */
 466                 *pacb = acb->next;
 467                 /* call the callback */
 468                 acb->common.cb(acb->common.opaque, ret);
 469                 qemu_aio_release(acb);
 470                 break;
 471             } else {
 472                 pacb = &acb->next;
 473             }
 474         }
 475     }
 476  the_end: ;
 477 }
 478
 479 static int posix_aio_flush(void *opaque)
 480 {
 481     PosixAioState *s = opaque;
 482     return !!s->first_aio;
 483 }
 484
 485 static PosixAioState *posix_aio_state;
 486
 487 static void paio_remove(struct qemu_paiocb *acb)
 488 {
 489     struct qemu_paiocb **pacb;
 490
 491     /* remove the callback from the queue */
 492     pacb = &posix_aio_state->first_aio;
 493     for(;;) {
 494         if (*pacb == NULL) {
 495             fprintf(stderr, "paio_remove: aio request not found!\n");
 496             break;
 497         } else if (*pacb == acb) {
 498             *pacb = acb->next;
 499             qemu_aio_release(acb);
 500             break;
 501         }
 502         pacb = &(*pacb)->next;
 503     }
 504 }
 505
 506 static void paio_cancel(BlockDriverAIOCB *blockacb)
 507 {
 508     struct qemu_paiocb *acb = (struct qemu_paiocb *)blockacb;
 509     int active = 0;
 510
 511     mutex_lock(&lock);
 512     if (!acb->active) {
 513         QTAILQ_REMOVE(&request_list, acb, node);
 514         acb->ret = -ECANCELED;
 515     } else if (acb->ret == -EINPROGRESS) {
 516         active = 1;
 517     }
 518     mutex_unlock(&lock);
 519
 520     if (active) {
 521         /* fail safe: if the aio could not be canceled, we wait for
 522            it */
 523         while (qemu_paio_error(acb) == EINPROGRESS)
 524             ;
 525     }
 526
 527     paio_remove(acb);
 528 }
 529
 530 static AIOPool raw_aio_pool = {
 531     .aiocb_size         = sizeof(struct qemu_paiocb),
 532     .cancel             = paio_cancel,
 533 };
 534
 535 BlockDriverAIOCB *paio_submit(BlockDriverState *bs, void *aio_ctx, int fd,
 536         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 537         BlockDriverCompletionFunc *cb, void *opaque, int type)
 538 {
 539     struct qemu_paiocb *acb;
 540
 541     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 542     if (!acb)
 543         return NULL;
 544     acb->aio_type = type;
 545     acb->aio_fildes = fd;
 546     acb->ev_signo = SIGUSR2;
 547     if (qiov) {
 548         acb->aio_iov = qiov->iov;
 549         acb->aio_niov = qiov->niov;
 550     }
 551     acb->aio_nbytes = nb_sectors * 512;
 552     acb->aio_offset = sector_num * 512;
 553
 554     acb->next = posix_aio_state->first_aio;
 555     posix_aio_state->first_aio = acb;
 556
 557     qemu_paio_submit(acb);
 558     return &acb->common;
 559 }
 560
 561 BlockDriverAIOCB *paio_ioctl(BlockDriverState *bs, int fd,
 562         unsigned long int req, void *buf,
 563         BlockDriverCompletionFunc *cb, void *opaque)
 564 {
 565     struct qemu_paiocb *acb;
 566
 567     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 568     if (!acb)
 569         return NULL;
 570     acb->aio_type = QEMU_AIO_IOCTL;
 571     acb->aio_fildes = fd;
 572     acb->ev_signo = SIGUSR2;
 573     acb->aio_offset = 0;
 574     acb->aio_ioctl_buf = buf;
 575     acb->aio_ioctl_cmd = req;
 576
 577     acb->next = posix_aio_state->first_aio;
 578     posix_aio_state->first_aio = acb;
 579
 580     qemu_paio_submit(acb);
 581     return &acb->common;
 582 }
 583
 584 void *paio_init(void)
 585 {
 586     sigset_t mask;
 587     PosixAioState *s;
 588     int ret;
 589
 590     if (posix_aio_state)
 591         return posix_aio_state;
 592
 593     s = qemu_malloc(sizeof(PosixAioState));
 594
 595     /* Make sure to block AIO signal */
 596     sigemptyset(&mask);
 597     sigaddset(&mask, SIGUSR2);
 598     sigprocmask(SIG_BLOCK, &mask, NULL);
 599
 600     s->first_aio = NULL;
 601     s->fd = qemu_signalfd(&mask);
 602     if (s->fd == -1) {
 603         fprintf(stderr, "failed to create signalfd\n");
 604         return NULL;
 605     }
 606
 607     fcntl(s->fd, F_SETFL, O_NONBLOCK);
 608
 609     qemu_aio_set_fd_handler(s->fd, posix_aio_read, NULL, posix_aio_flush, s);
 610
 611     ret = pthread_attr_init(&attr);
 612     if (ret)
 613         die2(ret, "pthread_attr_init");
 614
 615     ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
 616     if (ret)
 617         die2(ret, "pthread_attr_setdetachstate");
 618
 619     QTAILQ_INIT(&request_list);
 620
 621     posix_aio_state = s;
 622
 623     return posix_aio_state;
 624 }