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 #include "system/filesys.h"
  24 #include "system/shmem.h"
  25 #include "smbd/smbd.h"
  26
  27 #ifndef MAP_FILE
  28 #define MAP_FILE 0
  29 #endif
  30
  31 struct mmap_area {
  32         size_t size;
  33         volatile void *ptr;
  34 };
  35
  36 static int mmap_area_destructor(struct mmap_area *area)
  37 {
  38         munmap((void *)area->ptr, area->size);
  39         return 0;
  40 }
  41
  42 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
  43 {
  44         struct mmap_area *result;
  45         int fd;
  46
  47         result = talloc(mem_ctx, struct mmap_area);
  48         if (result == NULL) {
  49                 DEBUG(0, ("talloc failed\n"));
  50                 goto fail;
  51         }
  52
  53         fd = open("/dev/zero", O_RDWR);
  54         if (fd == -1) {
  55                 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
  56                           strerror(errno)));
  57                 goto fail;
  58         }
  59
  60         result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
  61                            MAP_SHARED|MAP_FILE, fd, 0);
  62         if (result->ptr == MAP_FAILED) {
  63                 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
  64                 goto fail;
  65         }
  66
  67         close(fd);
  68
  69         result->size = size;
  70         talloc_set_destructor(result, mmap_area_destructor);
  71
  72         return result;
  73
  74 fail:
  75         TALLOC_FREE(result);
  76         return NULL;
  77 }
  78
  79 struct rw_cmd {
  80         size_t n;
  81         SMB_OFF_T offset;
  82         bool read_cmd;
  83 };
  84
  85 struct rw_ret {
  86         ssize_t size;
  87         int ret_errno;
  88 };
  89
  90 struct aio_child_list;
  91
  92 struct aio_child {
  93         struct aio_child *prev, *next;
  94         struct aio_child_list *list;
  95         SMB_STRUCT_AIOCB *aiocb;
  96         pid_t pid;
  97         int sockfd;
  98         struct fd_event *sock_event;
  99         struct rw_ret retval;
 100         struct mmap_area *map;  /* ==NULL means write request */
 101         bool dont_delete;       /* Marked as in use since last cleanup */
 102         bool cancelled;
 103         bool read_cmd;
 104 };
 105
 106 struct aio_child_list {
 107         struct aio_child *children;
 108         struct timed_event *cleanup_event;
 109 };
 110
 111 static void free_aio_children(void **p)
 112 {
 113         TALLOC_FREE(*p);
 114 }
 115
 116 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
 117 {
 118         struct msghdr msg;
 119         struct iovec iov[1];
 120         ssize_t n;
 121 #ifndef HAVE_MSGHDR_MSG_CONTROL
 122         int newfd;
 123 #endif
 124
 125 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 126         union {
 127           struct cmsghdr        cm;
 128           char                          control[CMSG_SPACE(sizeof(int))];
 129         } control_un;
 130         struct cmsghdr  *cmptr;
 131
 132         msg.msg_control = control_un.control;
 133         msg.msg_controllen = sizeof(control_un.control);
 134 #else
 135 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
 136         msg.msg_accrights = (caddr_t) &newfd;
 137         msg.msg_accrightslen = sizeof(int);
 138 #else
 139 #error Can not pass file descriptors
 140 #endif
 141 #endif
 142
 143         msg.msg_name = NULL;
 144         msg.msg_namelen = 0;
 145
 146         iov[0].iov_base = (void *)ptr;
 147         iov[0].iov_len = nbytes;
 148         msg.msg_iov = iov;
 149         msg.msg_iovlen = 1;
 150
 151         if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
 152                 return(n);
 153         }
 154
 155 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 156         if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
 157             && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
 158                 if (cmptr->cmsg_level != SOL_SOCKET) {
 159                         DEBUG(10, ("control level != SOL_SOCKET"));
 160                         errno = EINVAL;
 161                         return -1;
 162                 }
 163                 if (cmptr->cmsg_type != SCM_RIGHTS) {
 164                         DEBUG(10, ("control type != SCM_RIGHTS"));
 165                         errno = EINVAL;
 166                         return -1;
 167                 }
 168                 *recvfd = *((int *) CMSG_DATA(cmptr));
 169         } else {
 170                 *recvfd = -1;           /* descriptor was not passed */
 171         }
 172 #else
 173         if (msg.msg_accrightslen == sizeof(int)) {
 174                 *recvfd = newfd;
 175         }
 176         else {
 177                 *recvfd = -1;           /* descriptor was not passed */
 178         }
 179 #endif
 180
 181         return(n);
 182 }
 183
 184 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
 185 {
 186         struct msghdr   msg;
 187         struct iovec    iov[1];
 188
 189 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 190         union {
 191                 struct cmsghdr  cm;
 192                 char control[CMSG_SPACE(sizeof(int))];
 193         } control_un;
 194         struct cmsghdr  *cmptr;
 195
 196         ZERO_STRUCT(msg);
 197         ZERO_STRUCT(control_un);
 198
 199         msg.msg_control = control_un.control;
 200         msg.msg_controllen = sizeof(control_un.control);
 201
 202         cmptr = CMSG_FIRSTHDR(&msg);
 203         cmptr->cmsg_len = CMSG_LEN(sizeof(int));
 204         cmptr->cmsg_level = SOL_SOCKET;
 205         cmptr->cmsg_type = SCM_RIGHTS;
 206         *((int *) CMSG_DATA(cmptr)) = sendfd;
 207 #else
 208         ZERO_STRUCT(msg);
 209         msg.msg_accrights = (caddr_t) &sendfd;
 210         msg.msg_accrightslen = sizeof(int);
 211 #endif
 212
 213         msg.msg_name = NULL;
 214         msg.msg_namelen = 0;
 215
 216         ZERO_STRUCT(iov);
 217         iov[0].iov_base = (void *)ptr;
 218         iov[0].iov_len = nbytes;
 219         msg.msg_iov = iov;
 220         msg.msg_iovlen = 1;
 221
 222         return (sendmsg(fd, &msg, 0));
 223 }
 224
 225 static void aio_child_cleanup(struct event_context *event_ctx,
 226                               struct timed_event *te,
 227                               struct timeval now,
 228                               void *private_data)
 229 {
 230         struct aio_child_list *list = talloc_get_type_abort(
 231                 private_data, struct aio_child_list);
 232         struct aio_child *child, *next;
 233
 234         TALLOC_FREE(list->cleanup_event);
 235
 236         for (child = list->children; child != NULL; child = next) {
 237                 next = child->next;
 238
 239                 if (child->aiocb != NULL) {
 240                         DEBUG(10, ("child %d currently active\n",
 241                                    (int)child->pid));
 242                         continue;
 243                 }
 244
 245                 if (child->dont_delete) {
 246                         DEBUG(10, ("Child %d was active since last cleanup\n",
 247                                    (int)child->pid));
 248                         child->dont_delete = false;
 249                         continue;
 250                 }
 251
 252                 DEBUG(10, ("Child %d idle for more than 30 seconds, "
 253                            "deleting\n", (int)child->pid));
 254
 255                 TALLOC_FREE(child);
 256                 child = next;
 257         }
 258
 259         if (list->children != NULL) {
 260                 /*
 261                  * Re-schedule the next cleanup round
 262                  */
 263                 list->cleanup_event = event_add_timed(server_event_context(), list,
 264                                                       timeval_add(&now, 30, 0),
 265                                                       aio_child_cleanup, list);
 266
 267         }
 268 }
 269
 270 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 271 {
 272         struct aio_child_list *data = NULL;
 273
 274         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 275                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 276                                         return NULL);
 277         }
 278
 279         if (data == NULL) {
 280                 data = talloc_zero(NULL, struct aio_child_list);
 281                 if (data == NULL) {
 282                         return NULL;
 283                 }
 284         }
 285
 286         /*
 287          * Regardless of whether the child_list had been around or not, make
 288          * sure that we have a cleanup timed event. This timed event will
 289          * delete itself when it finds that no children are around anymore.
 290          */
 291
 292         if (data->cleanup_event == NULL) {
 293                 data->cleanup_event = event_add_timed(server_event_context(), data,
 294                                                       timeval_current_ofs(30, 0),
 295                                                       aio_child_cleanup, data);
 296                 if (data->cleanup_event == NULL) {
 297                         TALLOC_FREE(data);
 298                         return NULL;
 299                 }
 300         }
 301
 302         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 303                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 304                                         struct aio_child_list, return False);
 305         }
 306
 307         return data;
 308 }
 309
 310 static void aio_child_loop(int sockfd, struct mmap_area *map)
 311 {
 312         while (true) {
 313                 int fd = -1;
 314                 ssize_t ret;
 315                 struct rw_cmd cmd_struct;
 316                 struct rw_ret ret_struct;
 317
 318                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 319                 if (ret != sizeof(cmd_struct)) {
 320                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 321                                    strerror(errno)));
 322                         exit(1);
 323                 }
 324
 325                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 326                            cmd_struct.read_cmd ? "read" : "write",
 327                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 328
 329 #ifdef ENABLE_BUILD_FARM_HACKS
 330                 {
 331                         /*
 332                          * In the build farm, we want erratic behaviour for
 333                          * async I/O times
 334                          */
 335                         uint8_t randval;
 336                         unsigned msecs;
 337                         /*
 338                          * use generate_random_buffer, we just forked from a
 339                          * common parent state
 340                          */
 341                         generate_random_buffer(&randval, sizeof(randval));
 342                         msecs = randval + 20;
 343                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 344                         smb_msleep(msecs);
 345                 }
 346 #endif
 347
 348
 349                 ZERO_STRUCT(ret_struct);
 350
 351                 if (cmd_struct.read_cmd) {
 352                         ret_struct.size = sys_pread(
 353                                 fd, (void *)map->ptr, cmd_struct.n,
 354                                 cmd_struct.offset);
 355 #if 0
 356 /* This breaks "make test" when run with aio_fork module. */
 357 #ifdef ENABLE_BUILD_FARM_HACKS
 358                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 359 #endif
 360 #endif
 361                 }
 362                 else {
 363                         ret_struct.size = sys_pwrite(
 364                                 fd, (void *)map->ptr, cmd_struct.n,
 365                                 cmd_struct.offset);
 366                 }
 367
 368                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 369                            (int)ret_struct.size));
 370
 371                 if (ret_struct.size == -1) {
 372                         ret_struct.ret_errno = errno;
 373                 }
 374
 375                 /*
 376                  * Close the fd before telling our parent we're done. The
 377                  * parent might close and re-open the file very quickly, and
 378                  * with system-level share modes (GPFS) we would get an
 379                  * unjustified SHARING_VIOLATION.
 380                  */
 381                 close(fd);
 382
 383                 ret = write_data(sockfd, (char *)&ret_struct,
 384                                  sizeof(ret_struct));
 385                 if (ret != sizeof(ret_struct)) {
 386                         DEBUG(10, ("could not write ret_struct: %s\n",
 387                                    strerror(errno)));
 388                         exit(2);
 389                 }
 390         }
 391 }
 392
 393 static void handle_aio_completion(struct event_context *event_ctx,
 394                                   struct fd_event *event, uint16 flags,
 395                                   void *p)
 396 {
 397         struct aio_extra *aio_ex = NULL;
 398         struct aio_child *child = (struct aio_child *)p;
 399         NTSTATUS status;
 400
 401         DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
 402
 403         if ((flags & EVENT_FD_READ) == 0) {
 404                 return;
 405         }
 406
 407         status = read_data(child->sockfd, (char *)&child->retval,
 408                            sizeof(child->retval));
 409
 410         if (!NT_STATUS_IS_OK(status)) {
 411                 DEBUG(1, ("aio child %d died: %s\n", (int)child->pid,
 412                           nt_errstr(status)));
 413                 child->retval.size = -1;
 414                 child->retval.ret_errno = EIO;
 415         }
 416
 417         if (child->aiocb == NULL) {
 418                 DEBUG(1, ("Inactive child died\n"));
 419                 TALLOC_FREE(child);
 420                 return;
 421         }
 422
 423         if (child->cancelled) {
 424                 child->aiocb = NULL;
 425                 child->cancelled = false;
 426                 return;
 427         }
 428
 429         if (child->read_cmd && (child->retval.size > 0)) {
 430                 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
 431                 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
 432                        child->retval.size);
 433         }
 434
 435         aio_ex = (struct aio_extra *)child->aiocb->aio_sigevent.sigev_value.sival_ptr;
 436         smbd_aio_complete_aio_ex(aio_ex);
 437         TALLOC_FREE(aio_ex);
 438 }
 439
 440 static int aio_child_destructor(struct aio_child *child)
 441 {
 442         char c=0;
 443
 444         SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
 445
 446         DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
 447                         child->pid, child->sockfd));
 448
 449         /*
 450          * closing the sockfd makes the child not return from recvmsg() on RHEL
 451          * 5.5 so instead force the child to exit by writing bad data to it
 452          */
 453         write(child->sockfd, &c, sizeof(c));
 454         close(child->sockfd);
 455         DLIST_REMOVE(child->list->children, child);
 456         return 0;
 457 }
 458
 459 /*
 460  * We have to close all fd's in open files, we might incorrectly hold a system
 461  * level share mode on a file.
 462  */
 463
 464 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 465                                          void *private_data)
 466 {
 467         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 468                 close(fsp->fh->fd);
 469                 fsp->fh->fd = -1;
 470         }
 471         return NULL;
 472 }
 473
 474 static NTSTATUS create_aio_child(struct smbd_server_connection *sconn,
 475                                  struct aio_child_list *children,
 476                                  size_t map_size,
 477                                  struct aio_child **presult)
 478 {
 479         struct aio_child *result;
 480         int fdpair[2];
 481         NTSTATUS status;
 482
 483         fdpair[0] = fdpair[1] = -1;
 484
 485         result = talloc_zero(children, struct aio_child);
 486         NT_STATUS_HAVE_NO_MEMORY(result);
 487
 488         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 489                 status = map_nt_error_from_unix(errno);
 490                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 491                 goto fail;
 492         }
 493
 494         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 495
 496         result->map = mmap_area_init(result, map_size);
 497         if (result->map == NULL) {
 498                 status = map_nt_error_from_unix(errno);
 499                 DEBUG(0, ("Could not create mmap area\n"));
 500                 goto fail;
 501         }
 502
 503         result->pid = sys_fork();
 504         if (result->pid == -1) {
 505                 status = map_nt_error_from_unix(errno);
 506                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 507                 goto fail;
 508         }
 509
 510         if (result->pid == 0) {
 511                 close(fdpair[0]);
 512                 result->sockfd = fdpair[1];
 513                 files_forall(sconn, close_fsp_fd, NULL);
 514                 aio_child_loop(result->sockfd, result->map);
 515         }
 516
 517         DEBUG(10, ("Child %d created with sockfd %d\n",
 518                         result->pid, fdpair[0]));
 519
 520         result->sockfd = fdpair[0];
 521         close(fdpair[1]);
 522
 523         result->sock_event = event_add_fd(server_event_context(), result,
 524                                           result->sockfd, EVENT_FD_READ,
 525                                           handle_aio_completion,
 526                                           result);
 527         if (result->sock_event == NULL) {
 528                 status = NT_STATUS_NO_MEMORY;
 529                 DEBUG(0, ("event_add_fd failed\n"));
 530                 goto fail;
 531         }
 532
 533         result->list = children;
 534         DLIST_ADD(children->children, result);
 535
 536         talloc_set_destructor(result, aio_child_destructor);
 537
 538         *presult = result;
 539
 540         return NT_STATUS_OK;
 541
 542  fail:
 543         if (fdpair[0] != -1) close(fdpair[0]);
 544         if (fdpair[1] != -1) close(fdpair[1]);
 545         TALLOC_FREE(result);
 546
 547         return status;
 548 }
 549
 550 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
 551                                struct aio_child **pchild)
 552 {
 553         struct aio_child_list *children;
 554         struct aio_child *child;
 555         NTSTATUS status;
 556
 557         children = init_aio_children(handle);
 558         if (children == NULL) {
 559                 return NT_STATUS_NO_MEMORY;
 560         }
 561
 562         for (child = children->children; child != NULL; child = child->next) {
 563                 if (child->aiocb == NULL) {
 564                         /* idle */
 565                         break;
 566                 }
 567         }
 568
 569         if (child == NULL) {
 570                 DEBUG(10, ("no idle child found, creating new one\n"));
 571
 572                 status = create_aio_child(handle->conn->sconn, children,
 573                                           128*1024, &child);
 574                 if (!NT_STATUS_IS_OK(status)) {
 575                         DEBUG(10, ("create_aio_child failed: %s\n",
 576                                    nt_errstr(status)));
 577                         return status;
 578                 }
 579         }
 580
 581         child->dont_delete = true;
 582
 583         *pchild = child;
 584         return NT_STATUS_OK;
 585 }
 586
 587 static int aio_fork_read(struct vfs_handle_struct *handle,
 588                          struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 589 {
 590         struct aio_child *child;
 591         struct rw_cmd cmd;
 592         ssize_t ret;
 593         NTSTATUS status;
 594
 595         if (aiocb->aio_nbytes > 128*1024) {
 596                 /* TODO: support variable buffers */
 597                 errno = EINVAL;
 598                 return -1;
 599         }
 600
 601         status = get_idle_child(handle, &child);
 602         if (!NT_STATUS_IS_OK(status)) {
 603                 DEBUG(10, ("Could not get an idle child\n"));
 604                 return -1;
 605         }
 606
 607         child->read_cmd = true;
 608         child->aiocb = aiocb;
 609         child->retval.ret_errno = EINPROGRESS;
 610
 611         ZERO_STRUCT(cmd);
 612         cmd.n = aiocb->aio_nbytes;
 613         cmd.offset = aiocb->aio_offset;
 614         cmd.read_cmd = child->read_cmd;
 615
 616         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 617                    (int)child->pid));
 618
 619         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 620         if (ret == -1) {
 621                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 622                 return -1;
 623         }
 624
 625         return 0;
 626 }
 627
 628 static int aio_fork_write(struct vfs_handle_struct *handle,
 629                           struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 630 {
 631         struct aio_child *child;
 632         struct rw_cmd cmd;
 633         ssize_t ret;
 634         NTSTATUS status;
 635
 636         if (aiocb->aio_nbytes > 128*1024) {
 637                 /* TODO: support variable buffers */
 638                 errno = EINVAL;
 639                 return -1;
 640         }
 641
 642         status = get_idle_child(handle, &child);
 643         if (!NT_STATUS_IS_OK(status)) {
 644                 DEBUG(10, ("Could not get an idle child\n"));
 645                 return -1;
 646         }
 647
 648         child->read_cmd = false;
 649         child->aiocb = aiocb;
 650         child->retval.ret_errno = EINPROGRESS;
 651
 652         memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
 653                aiocb->aio_nbytes);
 654
 655         ZERO_STRUCT(cmd);
 656         cmd.n = aiocb->aio_nbytes;
 657         cmd.offset = aiocb->aio_offset;
 658         cmd.read_cmd = child->read_cmd;
 659
 660         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 661                    (int)child->pid));
 662
 663         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 664         if (ret == -1) {
 665                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 666                 return -1;
 667         }
 668
 669         return 0;
 670 }
 671
 672 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
 673                                              SMB_STRUCT_AIOCB *aiocb)
 674 {
 675         struct aio_child_list *children;
 676         struct aio_child *child;
 677
 678         children = init_aio_children(handle);
 679         if (children == NULL) {
 680                 return NULL;
 681         }
 682
 683         for (child = children->children; child != NULL; child = child->next) {
 684                 if (child->aiocb == aiocb) {
 685                         return child;
 686                 }
 687         }
 688
 689         return NULL;
 690 }
 691
 692 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
 693                                   struct files_struct *fsp,
 694                                   SMB_STRUCT_AIOCB *aiocb)
 695 {
 696         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 697
 698         if (child == NULL) {
 699                 errno = EINVAL;
 700                 DEBUG(0, ("returning EINVAL\n"));
 701                 return -1;
 702         }
 703
 704         child->aiocb = NULL;
 705
 706         if (child->retval.size == -1) {
 707                 errno = child->retval.ret_errno;
 708         }
 709
 710         return child->retval.size;
 711 }
 712
 713 static int aio_fork_cancel(struct vfs_handle_struct *handle,
 714                            struct files_struct *fsp,
 715                            SMB_STRUCT_AIOCB *aiocb)
 716 {
 717         struct aio_child_list *children;
 718         struct aio_child *child;
 719
 720         children = init_aio_children(handle);
 721         if (children == NULL) {
 722                 errno = EINVAL;
 723                 return -1;
 724         }
 725
 726         for (child = children->children; child != NULL; child = child->next) {
 727                 if (child->aiocb == NULL) {
 728                         continue;
 729                 }
 730                 if (child->aiocb->aio_fildes != fsp->fh->fd) {
 731                         continue;
 732                 }
 733                 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
 734                         continue;
 735                 }
 736
 737                 /*
 738                  * We let the child do its job, but we discard the result when
 739                  * it's finished.
 740                  */
 741
 742                 child->cancelled = true;
 743         }
 744
 745         return AIO_CANCELED;
 746 }
 747
 748 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
 749                              struct files_struct *fsp,
 750                              SMB_STRUCT_AIOCB *aiocb)
 751 {
 752         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 753
 754         if (child == NULL) {
 755                 errno = EINVAL;
 756                 return -1;
 757         }
 758
 759         return child->retval.ret_errno;
 760 }
 761
 762 static void aio_fork_suspend_timed_out(struct tevent_context *event_ctx,
 763                                         struct tevent_timer *te,
 764                                         struct timeval now,
 765                                         void *private_data)
 766 {
 767         bool *timed_out = (bool *)private_data;
 768         /* Remove this timed event handler. */
 769         TALLOC_FREE(te);
 770         *timed_out = true;
 771 }
 772
 773 static int aio_fork_suspend(struct vfs_handle_struct *handle,
 774                         struct files_struct *fsp,
 775                         const SMB_STRUCT_AIOCB * const aiocb_array[],
 776                         int n,
 777                         const struct timespec *timeout)
 778 {
 779         struct aio_child_list *children = NULL;
 780         TALLOC_CTX *frame = talloc_stackframe();
 781         struct event_context *ev = NULL;
 782         int i;
 783         int ret = -1;
 784         bool timed_out = false;
 785
 786         children = init_aio_children(handle);
 787         if (children == NULL) {
 788                 errno = EINVAL;
 789                 goto out;
 790         }
 791
 792         /* This is a blocking call, and has to use a sub-event loop. */
 793         ev = event_context_init(frame);
 794         if (ev == NULL) {
 795                 errno = ENOMEM;
 796                 goto out;
 797         }
 798
 799         if (timeout) {
 800                 struct timeval tv = convert_timespec_to_timeval(*timeout);
 801                 struct tevent_timer *te = tevent_add_timer(ev,
 802                                                 frame,
 803                                                 timeval_current_ofs(tv.tv_sec,
 804                                                                     tv.tv_usec),
 805                                                 aio_fork_suspend_timed_out,
 806                                                 &timed_out);
 807                 if (!te) {
 808                         errno = ENOMEM;
 809                         goto out;
 810                 }
 811         }
 812
 813         for (i = 0; i < n; i++) {
 814                 struct aio_child *child = NULL;
 815                 const SMB_STRUCT_AIOCB *aiocb = aiocb_array[i];
 816
 817                 if (!aiocb) {
 818                         continue;
 819                 }
 820
 821                 /*
 822                  * We're going to cheat here. We know that smbd/aio.c
 823                  * only calls this when it's waiting for every single
 824                  * outstanding call to finish on a close, so just wait
 825                  * individually for each IO to complete. We don't care
 826                  * what order they finish - only that they all do. JRA.
 827                  */
 828
 829                 for (child = children->children; child != NULL; child = child->next) {
 830                         if (child->aiocb == NULL) {
 831                                 continue;
 832                         }
 833                         if (child->aiocb->aio_fildes != fsp->fh->fd) {
 834                                 continue;
 835                         }
 836                         if (child->aiocb != aiocb) {
 837                                 continue;
 838                         }
 839
 840                         if (child->aiocb->aio_sigevent.sigev_value.sival_ptr == NULL) {
 841                                 continue;
 842                         }
 843
 844                         /* We're never using this event on the
 845                          * main event context again... */
 846                         TALLOC_FREE(child->sock_event);
 847
 848                         child->sock_event = event_add_fd(ev,
 849                                                 child,
 850                                                 child->sockfd,
 851                                                 EVENT_FD_READ,
 852                                                 handle_aio_completion,
 853                                                 child);
 854
 855                         while (1) {
 856                                 if (tevent_loop_once(ev) == -1) {
 857                                         goto out;
 858                                 }
 859
 860                                 if (timed_out) {
 861                                         errno = EAGAIN;
 862                                         goto out;
 863                                 }
 864
 865                                 /* We set child->aiocb to NULL in our hooked
 866                                  * AIO_RETURN(). */
 867                                 if (child->aiocb == NULL) {
 868                                         break;
 869                                 }
 870                         }
 871                 }
 872         }
 873
 874         ret = 0;
 875
 876   out:
 877
 878         TALLOC_FREE(frame);
 879         return ret;
 880 }
 881
 882 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 883         .aio_read_fn = aio_fork_read,
 884         .aio_write_fn = aio_fork_write,
 885         .aio_return_fn = aio_fork_return_fn,
 886         .aio_cancel_fn = aio_fork_cancel,
 887         .aio_error_fn = aio_fork_error_fn,
 888         .aio_suspend_fn = aio_fork_suspend,
 889 };
 890
 891 NTSTATUS vfs_aio_fork_init(void);
 892 NTSTATUS vfs_aio_fork_init(void)
 893 {
 894         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 895                                 "aio_fork", &vfs_aio_fork_fns);
 896 }