source3/modules/vfs_aio_fork.c

   1 /*
   2  * Simulate the Posix AIO using mmap/fork
   3  *
   4  * Copyright (C) Volker Lendecke 2008
   5  * Copyright (C) Jeremy Allison 2010
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 2 of the License, or
  10  * (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #include "includes.h"
  23
  24 struct mmap_area {
  25         size_t size;
  26         volatile void *ptr;
  27 };
  28
  29 static int mmap_area_destructor(struct mmap_area *area)
  30 {
  31         munmap((void *)area->ptr, area->size);
  32         return 0;
  33 }
  34
  35 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
  36 {
  37         struct mmap_area *result;
  38         int fd;
  39
  40         result = talloc(mem_ctx, struct mmap_area);
  41         if (result == NULL) {
  42                 DEBUG(0, ("talloc failed\n"));
  43                 goto fail;
  44         }
  45
  46         fd = open("/dev/zero", O_RDWR);
  47         if (fd == -1) {
  48                 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
  49                           strerror(errno)));
  50                 goto fail;
  51         }
  52
  53         result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
  54                            MAP_SHARED|MAP_FILE, fd, 0);
  55         if (result->ptr == MAP_FAILED) {
  56                 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
  57                 goto fail;
  58         }
  59
  60         close(fd);
  61
  62         result->size = size;
  63         talloc_set_destructor(result, mmap_area_destructor);
  64
  65         return result;
  66
  67 fail:
  68         TALLOC_FREE(result);
  69         return NULL;
  70 }
  71
  72 struct rw_cmd {
  73         size_t n;
  74         SMB_OFF_T offset;
  75         bool read_cmd;
  76 };
  77
  78 struct rw_ret {
  79         ssize_t size;
  80         int ret_errno;
  81 };
  82
  83 struct aio_child_list;
  84
  85 struct aio_child {
  86         struct aio_child *prev, *next;
  87         struct aio_child_list *list;
  88         SMB_STRUCT_AIOCB *aiocb;
  89         pid_t pid;
  90         int sockfd;
  91         struct fd_event *sock_event;
  92         struct rw_ret retval;
  93         struct mmap_area *map;  /* ==NULL means write request */
  94         bool dont_delete;       /* Marked as in use since last cleanup */
  95         bool cancelled;
  96         bool read_cmd;
  97 };
  98
  99 struct aio_child_list {
 100         struct aio_child *children;
 101         struct timed_event *cleanup_event;
 102 };
 103
 104 static void free_aio_children(void **p)
 105 {
 106         TALLOC_FREE(*p);
 107 }
 108
 109 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
 110 {
 111         struct msghdr msg;
 112         struct iovec iov[1];
 113         ssize_t n;
 114 #ifndef HAVE_MSGHDR_MSG_CONTROL
 115         int newfd;
 116 #endif
 117
 118 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 119         union {
 120           struct cmsghdr        cm;
 121           char                          control[CMSG_SPACE(sizeof(int))];
 122         } control_un;
 123         struct cmsghdr  *cmptr;
 124
 125         msg.msg_control = control_un.control;
 126         msg.msg_controllen = sizeof(control_un.control);
 127 #else
 128 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
 129         msg.msg_accrights = (caddr_t) &newfd;
 130         msg.msg_accrightslen = sizeof(int);
 131 #else
 132 #error Can not pass file descriptors
 133 #endif
 134 #endif
 135
 136         msg.msg_name = NULL;
 137         msg.msg_namelen = 0;
 138
 139         iov[0].iov_base = (void *)ptr;
 140         iov[0].iov_len = nbytes;
 141         msg.msg_iov = iov;
 142         msg.msg_iovlen = 1;
 143
 144         if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
 145                 return(n);
 146         }
 147
 148 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 149         if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
 150             && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
 151                 if (cmptr->cmsg_level != SOL_SOCKET) {
 152                         DEBUG(10, ("control level != SOL_SOCKET"));
 153                         errno = EINVAL;
 154                         return -1;
 155                 }
 156                 if (cmptr->cmsg_type != SCM_RIGHTS) {
 157                         DEBUG(10, ("control type != SCM_RIGHTS"));
 158                         errno = EINVAL;
 159                         return -1;
 160                 }
 161                 *recvfd = *((int *) CMSG_DATA(cmptr));
 162         } else {
 163                 *recvfd = -1;           /* descriptor was not passed */
 164         }
 165 #else
 166         if (msg.msg_accrightslen == sizeof(int)) {
 167                 *recvfd = newfd;
 168         }
 169         else {
 170                 *recvfd = -1;           /* descriptor was not passed */
 171         }
 172 #endif
 173
 174         return(n);
 175 }
 176
 177 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
 178 {
 179         struct msghdr   msg;
 180         struct iovec    iov[1];
 181
 182 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 183         union {
 184                 struct cmsghdr  cm;
 185                 char control[CMSG_SPACE(sizeof(int))];
 186         } control_un;
 187         struct cmsghdr  *cmptr;
 188
 189         ZERO_STRUCT(msg);
 190         ZERO_STRUCT(control_un);
 191
 192         msg.msg_control = control_un.control;
 193         msg.msg_controllen = sizeof(control_un.control);
 194
 195         cmptr = CMSG_FIRSTHDR(&msg);
 196         cmptr->cmsg_len = CMSG_LEN(sizeof(int));
 197         cmptr->cmsg_level = SOL_SOCKET;
 198         cmptr->cmsg_type = SCM_RIGHTS;
 199         *((int *) CMSG_DATA(cmptr)) = sendfd;
 200 #else
 201         ZERO_STRUCT(msg);
 202         msg.msg_accrights = (caddr_t) &sendfd;
 203         msg.msg_accrightslen = sizeof(int);
 204 #endif
 205
 206         msg.msg_name = NULL;
 207         msg.msg_namelen = 0;
 208
 209         ZERO_STRUCT(iov);
 210         iov[0].iov_base = (void *)ptr;
 211         iov[0].iov_len = nbytes;
 212         msg.msg_iov = iov;
 213         msg.msg_iovlen = 1;
 214
 215         return (sendmsg(fd, &msg, 0));
 216 }
 217
 218 static void aio_child_cleanup(struct event_context *event_ctx,
 219                               struct timed_event *te,
 220                               struct timeval now,
 221                               void *private_data)
 222 {
 223         struct aio_child_list *list = talloc_get_type_abort(
 224                 private_data, struct aio_child_list);
 225         struct aio_child *child, *next;
 226
 227         TALLOC_FREE(list->cleanup_event);
 228
 229         for (child = list->children; child != NULL; child = next) {
 230                 next = child->next;
 231
 232                 if (child->aiocb != NULL) {
 233                         DEBUG(10, ("child %d currently active\n",
 234                                    (int)child->pid));
 235                         continue;
 236                 }
 237
 238                 if (child->dont_delete) {
 239                         DEBUG(10, ("Child %d was active since last cleanup\n",
 240                                    (int)child->pid));
 241                         child->dont_delete = false;
 242                         continue;
 243                 }
 244
 245                 DEBUG(10, ("Child %d idle for more than 30 seconds, "
 246                            "deleting\n", (int)child->pid));
 247
 248                 TALLOC_FREE(child);
 249         }
 250
 251         if (list->children != NULL) {
 252                 /*
 253                  * Re-schedule the next cleanup round
 254                  */
 255                 list->cleanup_event = event_add_timed(smbd_event_context(), list,
 256                                                       timeval_add(&now, 30, 0),
 257                                                       aio_child_cleanup, list);
 258
 259         }
 260 }
 261
 262 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 263 {
 264         struct aio_child_list *data = NULL;
 265
 266         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 267                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 268                                         return NULL);
 269         }
 270
 271         if (data == NULL) {
 272                 data = TALLOC_ZERO_P(NULL, struct aio_child_list);
 273                 if (data == NULL) {
 274                         return NULL;
 275                 }
 276         }
 277
 278         /*
 279          * Regardless of whether the child_list had been around or not, make
 280          * sure that we have a cleanup timed event. This timed event will
 281          * delete itself when it finds that no children are around anymore.
 282          */
 283
 284         if (data->cleanup_event == NULL) {
 285                 data->cleanup_event = event_add_timed(smbd_event_context(), data,
 286                                                       timeval_current_ofs(30, 0),
 287                                                       aio_child_cleanup, data);
 288                 if (data->cleanup_event == NULL) {
 289                         TALLOC_FREE(data);
 290                         return NULL;
 291                 }
 292         }
 293
 294         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 295                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 296                                         struct aio_child_list, return False);
 297         }
 298
 299         return data;
 300 }
 301
 302 static void aio_child_loop(int sockfd, struct mmap_area *map)
 303 {
 304         while (true) {
 305                 int fd = -1;
 306                 ssize_t ret;
 307                 struct rw_cmd cmd_struct;
 308                 struct rw_ret ret_struct;
 309
 310                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 311                 if (ret != sizeof(cmd_struct)) {
 312                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 313                                    strerror(errno)));
 314                         exit(1);
 315                 }
 316
 317                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 318                            cmd_struct.read_cmd ? "read" : "write",
 319                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 320
 321 #ifdef ENABLE_BUILD_FARM_HACKS
 322                 {
 323                         /*
 324                          * In the build farm, we want erratic behaviour for
 325                          * async I/O times
 326                          */
 327                         uint8_t randval;
 328                         unsigned msecs;
 329                         /*
 330                          * use generate_random_buffer, we just forked from a
 331                          * common parent state
 332                          */
 333                         generate_random_buffer(&randval, sizeof(randval));
 334                         msecs = randval + 20;
 335                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 336                         smb_msleep(msecs);
 337                 }
 338 #endif
 339
 340
 341                 ZERO_STRUCT(ret_struct);
 342
 343                 if (cmd_struct.read_cmd) {
 344                         ret_struct.size = sys_pread(
 345                                 fd, (void *)map->ptr, cmd_struct.n,
 346                                 cmd_struct.offset);
 347 #if 0
 348 /* This breaks "make test" when run with aio_fork module. */
 349 #ifdef ENABLE_BUILD_FARM_HACKS
 350                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 351 #endif
 352 #endif
 353                 }
 354                 else {
 355                         ret_struct.size = sys_pwrite(
 356                                 fd, (void *)map->ptr, cmd_struct.n,
 357                                 cmd_struct.offset);
 358                 }
 359
 360                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 361                            (int)ret_struct.size));
 362
 363                 if (ret_struct.size == -1) {
 364                         ret_struct.ret_errno = errno;
 365                 }
 366
 367                 /*
 368                  * Close the fd before telling our parent we're done. The
 369                  * parent might close and re-open the file very quickly, and
 370                  * with system-level share modes (GPFS) we would get an
 371                  * unjustified SHARING_VIOLATION.
 372                  */
 373                 close(fd);
 374
 375                 ret = write_data(sockfd, (char *)&ret_struct,
 376                                  sizeof(ret_struct));
 377                 if (ret != sizeof(ret_struct)) {
 378                         DEBUG(10, ("could not write ret_struct: %s\n",
 379                                    strerror(errno)));
 380                         exit(2);
 381                 }
 382         }
 383 }
 384
 385 static void handle_aio_completion(struct event_context *event_ctx,
 386                                   struct fd_event *event, uint16 flags,
 387                                   void *p)
 388 {
 389         struct aio_extra *aio_ex = NULL;
 390         struct aio_child *child = (struct aio_child *)p;
 391
 392         DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
 393
 394         if ((flags & EVENT_FD_READ) == 0) {
 395                 return;
 396         }
 397
 398         if (!NT_STATUS_IS_OK(read_data(child->sockfd,
 399                                        (char *)&child->retval,
 400                                        sizeof(child->retval)))) {
 401                 DEBUG(0, ("aio child %d died\n", (int)child->pid));
 402                 child->retval.size = -1;
 403                 child->retval.ret_errno = EIO;
 404         }
 405
 406         if (child->cancelled) {
 407                 child->aiocb = NULL;
 408                 child->cancelled = false;
 409                 return;
 410         }
 411
 412         if (child->read_cmd && (child->retval.size > 0)) {
 413                 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
 414                 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
 415                        child->retval.size);
 416         }
 417
 418         aio_ex = (struct aio_extra *)child->aiocb->aio_sigevent.sigev_value.sival_ptr;
 419         smbd_aio_complete_aio_ex(aio_ex);
 420 }
 421
 422 static int aio_child_destructor(struct aio_child *child)
 423 {
 424         SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
 425         close(child->sockfd);
 426         DLIST_REMOVE(child->list->children, child);
 427         return 0;
 428 }
 429
 430 /*
 431  * We have to close all fd's in open files, we might incorrectly hold a system
 432  * level share mode on a file.
 433  */
 434
 435 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 436                                          void *private_data)
 437 {
 438         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 439                 close(fsp->fh->fd);
 440                 fsp->fh->fd = -1;
 441         }
 442         return NULL;
 443 }
 444
 445 static NTSTATUS create_aio_child(struct aio_child_list *children,
 446                                  size_t map_size,
 447                                  struct aio_child **presult)
 448 {
 449         struct aio_child *result;
 450         int fdpair[2];
 451         NTSTATUS status;
 452
 453         fdpair[0] = fdpair[1] = -1;
 454
 455         result = TALLOC_ZERO_P(children, struct aio_child);
 456         NT_STATUS_HAVE_NO_MEMORY(result);
 457
 458         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 459                 status = map_nt_error_from_unix(errno);
 460                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 461                 goto fail;
 462         }
 463
 464         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 465
 466         result->map = mmap_area_init(result, map_size);
 467         if (result->map == NULL) {
 468                 status = map_nt_error_from_unix(errno);
 469                 DEBUG(0, ("Could not create mmap area\n"));
 470                 goto fail;
 471         }
 472
 473         result->pid = sys_fork();
 474         if (result->pid == -1) {
 475                 status = map_nt_error_from_unix(errno);
 476                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 477                 goto fail;
 478         }
 479
 480         if (result->pid == 0) {
 481                 close(fdpair[0]);
 482                 result->sockfd = fdpair[1];
 483                 files_forall(close_fsp_fd, NULL);
 484                 aio_child_loop(result->sockfd, result->map);
 485         }
 486
 487         DEBUG(10, ("Child %d created\n", result->pid));
 488
 489         result->sockfd = fdpair[0];
 490         close(fdpair[1]);
 491
 492         result->sock_event = event_add_fd(smbd_event_context(), result,
 493                                           result->sockfd, EVENT_FD_READ,
 494                                           handle_aio_completion,
 495                                           result);
 496         if (result->sock_event == NULL) {
 497                 status = NT_STATUS_NO_MEMORY;
 498                 DEBUG(0, ("event_add_fd failed\n"));
 499                 goto fail;
 500         }
 501
 502         result->list = children;
 503         DLIST_ADD(children->children, result);
 504
 505         talloc_set_destructor(result, aio_child_destructor);
 506
 507         *presult = result;
 508
 509         return NT_STATUS_OK;
 510
 511  fail:
 512         if (fdpair[0] != -1) close(fdpair[0]);
 513         if (fdpair[1] != -1) close(fdpair[1]);
 514         TALLOC_FREE(result);
 515
 516         return status;
 517 }
 518
 519 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
 520                                struct aio_child **pchild)
 521 {
 522         struct aio_child_list *children;
 523         struct aio_child *child;
 524         NTSTATUS status;
 525
 526         children = init_aio_children(handle);
 527         if (children == NULL) {
 528                 return NT_STATUS_NO_MEMORY;
 529         }
 530
 531         for (child = children->children; child != NULL; child = child->next) {
 532                 if (child->aiocb == NULL) {
 533                         /* idle */
 534                         break;
 535                 }
 536         }
 537
 538         if (child == NULL) {
 539                 DEBUG(10, ("no idle child found, creating new one\n"));
 540
 541                 status = create_aio_child(children, 128*1024, &child);
 542                 if (!NT_STATUS_IS_OK(status)) {
 543                         DEBUG(10, ("create_aio_child failed: %s\n",
 544                                    nt_errstr(status)));
 545                         return status;
 546                 }
 547         }
 548
 549         child->dont_delete = true;
 550
 551         *pchild = child;
 552         return NT_STATUS_OK;
 553 }
 554
 555 static int aio_fork_read(struct vfs_handle_struct *handle,
 556                          struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 557 {
 558         struct aio_child *child;
 559         struct rw_cmd cmd;
 560         ssize_t ret;
 561         NTSTATUS status;
 562
 563         if (aiocb->aio_nbytes > 128*1024) {
 564                 /* TODO: support variable buffers */
 565                 errno = EINVAL;
 566                 return -1;
 567         }
 568
 569         status = get_idle_child(handle, &child);
 570         if (!NT_STATUS_IS_OK(status)) {
 571                 DEBUG(10, ("Could not get an idle child\n"));
 572                 return -1;
 573         }
 574
 575         child->read_cmd = true;
 576         child->aiocb = aiocb;
 577         child->retval.ret_errno = EINPROGRESS;
 578
 579         ZERO_STRUCT(cmd);
 580         cmd.n = aiocb->aio_nbytes;
 581         cmd.offset = aiocb->aio_offset;
 582         cmd.read_cmd = child->read_cmd;
 583
 584         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 585                    (int)child->pid));
 586
 587         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 588         if (ret == -1) {
 589                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 590                 return -1;
 591         }
 592
 593         return 0;
 594 }
 595
 596 static int aio_fork_write(struct vfs_handle_struct *handle,
 597                           struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 598 {
 599         struct aio_child *child;
 600         struct rw_cmd cmd;
 601         ssize_t ret;
 602         NTSTATUS status;
 603
 604         if (aiocb->aio_nbytes > 128*1024) {
 605                 /* TODO: support variable buffers */
 606                 errno = EINVAL;
 607                 return -1;
 608         }
 609
 610         status = get_idle_child(handle, &child);
 611         if (!NT_STATUS_IS_OK(status)) {
 612                 DEBUG(10, ("Could not get an idle child\n"));
 613                 return -1;
 614         }
 615
 616         child->read_cmd = false;
 617         child->aiocb = aiocb;
 618         child->retval.ret_errno = EINPROGRESS;
 619
 620         memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
 621                aiocb->aio_nbytes);
 622
 623         ZERO_STRUCT(cmd);
 624         cmd.n = aiocb->aio_nbytes;
 625         cmd.offset = aiocb->aio_offset;
 626         cmd.read_cmd = child->read_cmd;
 627
 628         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 629                    (int)child->pid));
 630
 631         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 632         if (ret == -1) {
 633                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 634                 return -1;
 635         }
 636
 637         return 0;
 638 }
 639
 640 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
 641                                              SMB_STRUCT_AIOCB *aiocb)
 642 {
 643         struct aio_child_list *children;
 644         struct aio_child *child;
 645
 646         children = init_aio_children(handle);
 647         if (children == NULL) {
 648                 return NULL;
 649         }
 650
 651         for (child = children->children; child != NULL; child = child->next) {
 652                 if (child->aiocb == aiocb) {
 653                         return child;
 654                 }
 655         }
 656
 657         return NULL;
 658 }
 659
 660 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
 661                                   struct files_struct *fsp,
 662                                   SMB_STRUCT_AIOCB *aiocb)
 663 {
 664         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 665
 666         if (child == NULL) {
 667                 errno = EINVAL;
 668                 DEBUG(0, ("returning EINVAL\n"));
 669                 return -1;
 670         }
 671
 672         child->aiocb = NULL;
 673
 674         if (child->retval.size == -1) {
 675                 errno = child->retval.ret_errno;
 676         }
 677
 678         return child->retval.size;
 679 }
 680
 681 static int aio_fork_cancel(struct vfs_handle_struct *handle,
 682                            struct files_struct *fsp,
 683                            SMB_STRUCT_AIOCB *aiocb)
 684 {
 685         struct aio_child_list *children;
 686         struct aio_child *child;
 687
 688         children = init_aio_children(handle);
 689         if (children == NULL) {
 690                 errno = EINVAL;
 691                 return -1;
 692         }
 693
 694         for (child = children->children; child != NULL; child = child->next) {
 695                 if (child->aiocb == NULL) {
 696                         continue;
 697                 }
 698                 if (child->aiocb->aio_fildes != fsp->fh->fd) {
 699                         continue;
 700                 }
 701                 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
 702                         continue;
 703                 }
 704
 705                 /*
 706                  * We let the child do its job, but we discard the result when
 707                  * it's finished.
 708                  */
 709
 710                 child->cancelled = true;
 711         }
 712
 713         return AIO_CANCELED;
 714 }
 715
 716 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
 717                              struct files_struct *fsp,
 718                              SMB_STRUCT_AIOCB *aiocb)
 719 {
 720         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 721
 722         if (child == NULL) {
 723                 errno = EINVAL;
 724                 return -1;
 725         }
 726
 727         return child->retval.ret_errno;
 728 }
 729
 730 static void aio_fork_suspend_timed_out(struct tevent_context *event_ctx,
 731                                         struct tevent_timer *te,
 732                                         struct timeval now,
 733                                         void *private_data)
 734 {
 735         bool *timed_out = (bool *)private_data;
 736         /* Remove this timed event handler. */
 737         TALLOC_FREE(te);
 738         *timed_out = true;
 739 }
 740
 741 static int aio_fork_suspend(struct vfs_handle_struct *handle,
 742                         struct files_struct *fsp,
 743                         const SMB_STRUCT_AIOCB * const aiocb_array[],
 744                         int n,
 745                         const struct timespec *timeout)
 746 {
 747         struct aio_child_list *children = NULL;
 748         TALLOC_CTX *frame = talloc_stackframe();
 749         struct event_context *ev = NULL;
 750         int i;
 751         int ret = -1;
 752         bool timed_out = false;
 753
 754         children = init_aio_children(handle);
 755         if (children == NULL) {
 756                 errno = EINVAL;
 757                 goto out;
 758         }
 759
 760         /* This is a blocking call, and has to use a sub-event loop. */
 761         ev = event_context_init(frame);
 762         if (ev == NULL) {
 763                 errno = ENOMEM;
 764                 goto out;
 765         }
 766
 767         if (timeout) {
 768                 struct timeval tv = convert_timespec_to_timeval(*timeout);
 769                 struct tevent_timer *te = tevent_add_timer(ev,
 770                                                 frame,
 771                                                 timeval_current_ofs(tv.tv_sec,
 772                                                                     tv.tv_usec),
 773                                                 aio_fork_suspend_timed_out,
 774                                                 &timed_out);
 775                 if (!te) {
 776                         errno = ENOMEM;
 777                         goto out;
 778                 }
 779         }
 780
 781         for (i = 0; i < n; i++) {
 782                 struct aio_child *child = NULL;
 783                 const SMB_STRUCT_AIOCB *aiocb = aiocb_array[i];
 784
 785                 if (!aiocb) {
 786                         continue;
 787                 }
 788
 789                 /*
 790                  * We're going to cheat here. We know that smbd/aio.c
 791                  * only calls this when it's waiting for every single
 792                  * outstanding call to finish on a close, so just wait
 793                  * individually for each IO to complete. We don't care
 794                  * what order they finish - only that they all do. JRA.
 795                  */
 796
 797                 for (child = children->children; child != NULL; child = child->next) {
 798                         if (child->aiocb == NULL) {
 799                                 continue;
 800                         }
 801                         if (child->aiocb->aio_fildes != fsp->fh->fd) {
 802                                 continue;
 803                         }
 804                         if (child->aiocb != aiocb) {
 805                                 continue;
 806                         }
 807
 808                         if (child->aiocb->aio_sigevent.sigev_value.sival_ptr == NULL) {
 809                                 continue;
 810                         }
 811
 812                         /* We're never using this event on the
 813                          * main event context again... */
 814                         TALLOC_FREE(child->sock_event);
 815
 816                         child->sock_event = event_add_fd(ev,
 817                                                 child,
 818                                                 child->sockfd,
 819                                                 EVENT_FD_READ,
 820                                                 handle_aio_completion,
 821                                                 child);
 822
 823                         while (1) {
 824                                 if (tevent_loop_once(ev) == -1) {
 825                                         goto out;
 826                                 }
 827
 828                                 if (timed_out) {
 829                                         errno = EAGAIN;
 830                                         goto out;
 831                                 }
 832
 833                                 /* We set child->aiocb to NULL in our hooked
 834                                  * AIO_RETURN(). */
 835                                 if (child->aiocb == NULL) {
 836                                         break;
 837                                 }
 838                         }
 839                 }
 840         }
 841
 842         ret = 0;
 843
 844   out:
 845
 846         TALLOC_FREE(frame);
 847         return ret;
 848 }
 849
 850 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 851         .aio_read = aio_fork_read,
 852         .aio_write = aio_fork_write,
 853         .aio_return_fn = aio_fork_return_fn,
 854         .aio_cancel = aio_fork_cancel,
 855         .aio_error_fn = aio_fork_error_fn,
 856         .aio_suspend = aio_fork_suspend,
 857 };
 858
 859 NTSTATUS vfs_aio_fork_init(void);
 860 NTSTATUS vfs_aio_fork_init(void)
 861 {
 862         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 863                                 "aio_fork", &vfs_aio_fork_fns);
 864 }