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                 child = next;
 250         }
 251
 252         if (list->children != NULL) {
 253                 /*
 254                  * Re-schedule the next cleanup round
 255                  */
 256                 list->cleanup_event = event_add_timed(smbd_event_context(), list,
 257                                                       timeval_add(&now, 30, 0),
 258                                                       aio_child_cleanup, list);
 259
 260         }
 261 }
 262
 263 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 264 {
 265         struct aio_child_list *data = NULL;
 266
 267         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 268                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 269                                         return NULL);
 270         }
 271
 272         if (data == NULL) {
 273                 data = TALLOC_ZERO_P(NULL, struct aio_child_list);
 274                 if (data == NULL) {
 275                         return NULL;
 276                 }
 277         }
 278
 279         /*
 280          * Regardless of whether the child_list had been around or not, make
 281          * sure that we have a cleanup timed event. This timed event will
 282          * delete itself when it finds that no children are around anymore.
 283          */
 284
 285         if (data->cleanup_event == NULL) {
 286                 data->cleanup_event = event_add_timed(smbd_event_context(), data,
 287                                                       timeval_current_ofs(30, 0),
 288                                                       aio_child_cleanup, data);
 289                 if (data->cleanup_event == NULL) {
 290                         TALLOC_FREE(data);
 291                         return NULL;
 292                 }
 293         }
 294
 295         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 296                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 297                                         struct aio_child_list, return False);
 298         }
 299
 300         return data;
 301 }
 302
 303 static void aio_child_loop(int sockfd, struct mmap_area *map)
 304 {
 305         while (true) {
 306                 int fd = -1;
 307                 ssize_t ret;
 308                 struct rw_cmd cmd_struct;
 309                 struct rw_ret ret_struct;
 310
 311                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 312                 if (ret != sizeof(cmd_struct)) {
 313                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 314                                    strerror(errno)));
 315                         exit(1);
 316                 }
 317
 318                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 319                            cmd_struct.read_cmd ? "read" : "write",
 320                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 321
 322 #ifdef ENABLE_BUILD_FARM_HACKS
 323                 {
 324                         /*
 325                          * In the build farm, we want erratic behaviour for
 326                          * async I/O times
 327                          */
 328                         uint8_t randval;
 329                         unsigned msecs;
 330                         /*
 331                          * use generate_random_buffer, we just forked from a
 332                          * common parent state
 333                          */
 334                         generate_random_buffer(&randval, sizeof(randval));
 335                         msecs = randval + 20;
 336                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 337                         smb_msleep(msecs);
 338                 }
 339 #endif
 340
 341
 342                 ZERO_STRUCT(ret_struct);
 343
 344                 if (cmd_struct.read_cmd) {
 345                         ret_struct.size = sys_pread(
 346                                 fd, (void *)map->ptr, cmd_struct.n,
 347                                 cmd_struct.offset);
 348 #if 0
 349 /* This breaks "make test" when run with aio_fork module. */
 350 #ifdef ENABLE_BUILD_FARM_HACKS
 351                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 352 #endif
 353 #endif
 354                 }
 355                 else {
 356                         ret_struct.size = sys_pwrite(
 357                                 fd, (void *)map->ptr, cmd_struct.n,
 358                                 cmd_struct.offset);
 359                 }
 360
 361                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 362                            (int)ret_struct.size));
 363
 364                 if (ret_struct.size == -1) {
 365                         ret_struct.ret_errno = errno;
 366                 }
 367
 368                 /*
 369                  * Close the fd before telling our parent we're done. The
 370                  * parent might close and re-open the file very quickly, and
 371                  * with system-level share modes (GPFS) we would get an
 372                  * unjustified SHARING_VIOLATION.
 373                  */
 374                 close(fd);
 375
 376                 ret = write_data(sockfd, (char *)&ret_struct,
 377                                  sizeof(ret_struct));
 378                 if (ret != sizeof(ret_struct)) {
 379                         DEBUG(10, ("could not write ret_struct: %s\n",
 380                                    strerror(errno)));
 381                         exit(2);
 382                 }
 383         }
 384 }
 385
 386 static void handle_aio_completion(struct event_context *event_ctx,
 387                                   struct fd_event *event, uint16 flags,
 388                                   void *p)
 389 {
 390         struct aio_extra *aio_ex = NULL;
 391         struct aio_child *child = (struct aio_child *)p;
 392         NTSTATUS status;
 393
 394         DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
 395
 396         if ((flags & EVENT_FD_READ) == 0) {
 397                 return;
 398         }
 399
 400         status = read_data(child->sockfd, (char *)&child->retval,
 401                            sizeof(child->retval));
 402
 403         if (!NT_STATUS_IS_OK(status)) {
 404                 DEBUG(1, ("aio child %d died: %s\n", (int)child->pid,
 405                           nt_errstr(status)));
 406                 child->retval.size = -1;
 407                 child->retval.ret_errno = EIO;
 408         }
 409
 410         if (child->aiocb == NULL) {
 411                 DEBUG(1, ("Inactive child died\n"));
 412                 TALLOC_FREE(child);
 413                 return;
 414         }
 415
 416         if (child->cancelled) {
 417                 child->aiocb = NULL;
 418                 child->cancelled = false;
 419                 return;
 420         }
 421
 422         if (child->read_cmd && (child->retval.size > 0)) {
 423                 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
 424                 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
 425                        child->retval.size);
 426         }
 427
 428         aio_ex = (struct aio_extra *)child->aiocb->aio_sigevent.sigev_value.sival_ptr;
 429         smbd_aio_complete_aio_ex(aio_ex);
 430 }
 431
 432 static int aio_child_destructor(struct aio_child *child)
 433 {
 434         char c=0;
 435
 436         SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
 437
 438         DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
 439                         child->pid, child->sockfd));
 440
 441         /*
 442          * closing the sockfd makes the child not return from recvmsg() on RHEL
 443          * 5.5 so instead force the child to exit by writing bad data to it
 444          */
 445         write(child->sockfd, &c, sizeof(c));
 446         close(child->sockfd);
 447         DLIST_REMOVE(child->list->children, child);
 448         return 0;
 449 }
 450
 451 /*
 452  * We have to close all fd's in open files, we might incorrectly hold a system
 453  * level share mode on a file.
 454  */
 455
 456 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 457                                          void *private_data)
 458 {
 459         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 460                 close(fsp->fh->fd);
 461                 fsp->fh->fd = -1;
 462         }
 463         return NULL;
 464 }
 465
 466 static NTSTATUS create_aio_child(struct smbd_server_connection *sconn,
 467                                  struct aio_child_list *children,
 468                                  size_t map_size,
 469                                  struct aio_child **presult)
 470 {
 471         struct aio_child *result;
 472         int fdpair[2];
 473         NTSTATUS status;
 474
 475         fdpair[0] = fdpair[1] = -1;
 476
 477         result = TALLOC_ZERO_P(children, struct aio_child);
 478         NT_STATUS_HAVE_NO_MEMORY(result);
 479
 480         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 481                 status = map_nt_error_from_unix(errno);
 482                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 483                 goto fail;
 484         }
 485
 486         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 487
 488         result->map = mmap_area_init(result, map_size);
 489         if (result->map == NULL) {
 490                 status = map_nt_error_from_unix(errno);
 491                 DEBUG(0, ("Could not create mmap area\n"));
 492                 goto fail;
 493         }
 494
 495         result->pid = sys_fork();
 496         if (result->pid == -1) {
 497                 status = map_nt_error_from_unix(errno);
 498                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 499                 goto fail;
 500         }
 501
 502         if (result->pid == 0) {
 503                 close(fdpair[0]);
 504                 result->sockfd = fdpair[1];
 505                 files_forall(sconn, close_fsp_fd, NULL);
 506                 aio_child_loop(result->sockfd, result->map);
 507         }
 508
 509         DEBUG(10, ("Child %d created with sockfd %d\n",
 510                         result->pid, fdpair[0]));
 511
 512         result->sockfd = fdpair[0];
 513         close(fdpair[1]);
 514
 515         result->sock_event = event_add_fd(smbd_event_context(), result,
 516                                           result->sockfd, EVENT_FD_READ,
 517                                           handle_aio_completion,
 518                                           result);
 519         if (result->sock_event == NULL) {
 520                 status = NT_STATUS_NO_MEMORY;
 521                 DEBUG(0, ("event_add_fd failed\n"));
 522                 goto fail;
 523         }
 524
 525         result->list = children;
 526         DLIST_ADD(children->children, result);
 527
 528         talloc_set_destructor(result, aio_child_destructor);
 529
 530         *presult = result;
 531
 532         return NT_STATUS_OK;
 533
 534  fail:
 535         if (fdpair[0] != -1) close(fdpair[0]);
 536         if (fdpair[1] != -1) close(fdpair[1]);
 537         TALLOC_FREE(result);
 538
 539         return status;
 540 }
 541
 542 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
 543                                struct aio_child **pchild)
 544 {
 545         struct aio_child_list *children;
 546         struct aio_child *child;
 547         NTSTATUS status;
 548
 549         children = init_aio_children(handle);
 550         if (children == NULL) {
 551                 return NT_STATUS_NO_MEMORY;
 552         }
 553
 554         for (child = children->children; child != NULL; child = child->next) {
 555                 if (child->aiocb == NULL) {
 556                         /* idle */
 557                         break;
 558                 }
 559         }
 560
 561         if (child == NULL) {
 562                 DEBUG(10, ("no idle child found, creating new one\n"));
 563
 564                 status = create_aio_child(handle->conn->sconn, children,
 565                                           128*1024, &child);
 566                 if (!NT_STATUS_IS_OK(status)) {
 567                         DEBUG(10, ("create_aio_child failed: %s\n",
 568                                    nt_errstr(status)));
 569                         return status;
 570                 }
 571         }
 572
 573         child->dont_delete = true;
 574
 575         *pchild = child;
 576         return NT_STATUS_OK;
 577 }
 578
 579 static int aio_fork_read(struct vfs_handle_struct *handle,
 580                          struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 581 {
 582         struct aio_child *child;
 583         struct rw_cmd cmd;
 584         ssize_t ret;
 585         NTSTATUS status;
 586
 587         if (aiocb->aio_nbytes > 128*1024) {
 588                 /* TODO: support variable buffers */
 589                 errno = EINVAL;
 590                 return -1;
 591         }
 592
 593         status = get_idle_child(handle, &child);
 594         if (!NT_STATUS_IS_OK(status)) {
 595                 DEBUG(10, ("Could not get an idle child\n"));
 596                 return -1;
 597         }
 598
 599         child->read_cmd = true;
 600         child->aiocb = aiocb;
 601         child->retval.ret_errno = EINPROGRESS;
 602
 603         ZERO_STRUCT(cmd);
 604         cmd.n = aiocb->aio_nbytes;
 605         cmd.offset = aiocb->aio_offset;
 606         cmd.read_cmd = child->read_cmd;
 607
 608         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 609                    (int)child->pid));
 610
 611         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 612         if (ret == -1) {
 613                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 614                 return -1;
 615         }
 616
 617         return 0;
 618 }
 619
 620 static int aio_fork_write(struct vfs_handle_struct *handle,
 621                           struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 622 {
 623         struct aio_child *child;
 624         struct rw_cmd cmd;
 625         ssize_t ret;
 626         NTSTATUS status;
 627
 628         if (aiocb->aio_nbytes > 128*1024) {
 629                 /* TODO: support variable buffers */
 630                 errno = EINVAL;
 631                 return -1;
 632         }
 633
 634         status = get_idle_child(handle, &child);
 635         if (!NT_STATUS_IS_OK(status)) {
 636                 DEBUG(10, ("Could not get an idle child\n"));
 637                 return -1;
 638         }
 639
 640         child->read_cmd = false;
 641         child->aiocb = aiocb;
 642         child->retval.ret_errno = EINPROGRESS;
 643
 644         memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
 645                aiocb->aio_nbytes);
 646
 647         ZERO_STRUCT(cmd);
 648         cmd.n = aiocb->aio_nbytes;
 649         cmd.offset = aiocb->aio_offset;
 650         cmd.read_cmd = child->read_cmd;
 651
 652         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 653                    (int)child->pid));
 654
 655         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 656         if (ret == -1) {
 657                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 658                 return -1;
 659         }
 660
 661         return 0;
 662 }
 663
 664 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
 665                                              SMB_STRUCT_AIOCB *aiocb)
 666 {
 667         struct aio_child_list *children;
 668         struct aio_child *child;
 669
 670         children = init_aio_children(handle);
 671         if (children == NULL) {
 672                 return NULL;
 673         }
 674
 675         for (child = children->children; child != NULL; child = child->next) {
 676                 if (child->aiocb == aiocb) {
 677                         return child;
 678                 }
 679         }
 680
 681         return NULL;
 682 }
 683
 684 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
 685                                   struct files_struct *fsp,
 686                                   SMB_STRUCT_AIOCB *aiocb)
 687 {
 688         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 689
 690         if (child == NULL) {
 691                 errno = EINVAL;
 692                 DEBUG(0, ("returning EINVAL\n"));
 693                 return -1;
 694         }
 695
 696         child->aiocb = NULL;
 697
 698         if (child->retval.size == -1) {
 699                 errno = child->retval.ret_errno;
 700         }
 701
 702         return child->retval.size;
 703 }
 704
 705 static int aio_fork_cancel(struct vfs_handle_struct *handle,
 706                            struct files_struct *fsp,
 707                            SMB_STRUCT_AIOCB *aiocb)
 708 {
 709         struct aio_child_list *children;
 710         struct aio_child *child;
 711
 712         children = init_aio_children(handle);
 713         if (children == NULL) {
 714                 errno = EINVAL;
 715                 return -1;
 716         }
 717
 718         for (child = children->children; child != NULL; child = child->next) {
 719                 if (child->aiocb == NULL) {
 720                         continue;
 721                 }
 722                 if (child->aiocb->aio_fildes != fsp->fh->fd) {
 723                         continue;
 724                 }
 725                 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
 726                         continue;
 727                 }
 728
 729                 /*
 730                  * We let the child do its job, but we discard the result when
 731                  * it's finished.
 732                  */
 733
 734                 child->cancelled = true;
 735         }
 736
 737         return AIO_CANCELED;
 738 }
 739
 740 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
 741                              struct files_struct *fsp,
 742                              SMB_STRUCT_AIOCB *aiocb)
 743 {
 744         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 745
 746         if (child == NULL) {
 747                 errno = EINVAL;
 748                 return -1;
 749         }
 750
 751         return child->retval.ret_errno;
 752 }
 753
 754 static void aio_fork_suspend_timed_out(struct tevent_context *event_ctx,
 755                                         struct tevent_timer *te,
 756                                         struct timeval now,
 757                                         void *private_data)
 758 {
 759         bool *timed_out = (bool *)private_data;
 760         /* Remove this timed event handler. */
 761         TALLOC_FREE(te);
 762         *timed_out = true;
 763 }
 764
 765 static int aio_fork_suspend(struct vfs_handle_struct *handle,
 766                         struct files_struct *fsp,
 767                         const SMB_STRUCT_AIOCB * const aiocb_array[],
 768                         int n,
 769                         const struct timespec *timeout)
 770 {
 771         struct aio_child_list *children = NULL;
 772         TALLOC_CTX *frame = talloc_stackframe();
 773         struct event_context *ev = NULL;
 774         int i;
 775         int ret = -1;
 776         bool timed_out = false;
 777
 778         children = init_aio_children(handle);
 779         if (children == NULL) {
 780                 errno = EINVAL;
 781                 goto out;
 782         }
 783
 784         /* This is a blocking call, and has to use a sub-event loop. */
 785         ev = event_context_init(frame);
 786         if (ev == NULL) {
 787                 errno = ENOMEM;
 788                 goto out;
 789         }
 790
 791         if (timeout) {
 792                 struct timeval tv = convert_timespec_to_timeval(*timeout);
 793                 struct tevent_timer *te = tevent_add_timer(ev,
 794                                                 frame,
 795                                                 timeval_current_ofs(tv.tv_sec,
 796                                                                     tv.tv_usec),
 797                                                 aio_fork_suspend_timed_out,
 798                                                 &timed_out);
 799                 if (!te) {
 800                         errno = ENOMEM;
 801                         goto out;
 802                 }
 803         }
 804
 805         for (i = 0; i < n; i++) {
 806                 struct aio_child *child = NULL;
 807                 const SMB_STRUCT_AIOCB *aiocb = aiocb_array[i];
 808
 809                 if (!aiocb) {
 810                         continue;
 811                 }
 812
 813                 /*
 814                  * We're going to cheat here. We know that smbd/aio.c
 815                  * only calls this when it's waiting for every single
 816                  * outstanding call to finish on a close, so just wait
 817                  * individually for each IO to complete. We don't care
 818                  * what order they finish - only that they all do. JRA.
 819                  */
 820
 821                 for (child = children->children; child != NULL; child = child->next) {
 822                         if (child->aiocb == NULL) {
 823                                 continue;
 824                         }
 825                         if (child->aiocb->aio_fildes != fsp->fh->fd) {
 826                                 continue;
 827                         }
 828                         if (child->aiocb != aiocb) {
 829                                 continue;
 830                         }
 831
 832                         if (child->aiocb->aio_sigevent.sigev_value.sival_ptr == NULL) {
 833                                 continue;
 834                         }
 835
 836                         /* We're never using this event on the
 837                          * main event context again... */
 838                         TALLOC_FREE(child->sock_event);
 839
 840                         child->sock_event = event_add_fd(ev,
 841                                                 child,
 842                                                 child->sockfd,
 843                                                 EVENT_FD_READ,
 844                                                 handle_aio_completion,
 845                                                 child);
 846
 847                         while (1) {
 848                                 if (tevent_loop_once(ev) == -1) {
 849                                         goto out;
 850                                 }
 851
 852                                 if (timed_out) {
 853                                         errno = EAGAIN;
 854                                         goto out;
 855                                 }
 856
 857                                 /* We set child->aiocb to NULL in our hooked
 858                                  * AIO_RETURN(). */
 859                                 if (child->aiocb == NULL) {
 860                                         break;
 861                                 }
 862                         }
 863                 }
 864         }
 865
 866         ret = 0;
 867
 868   out:
 869
 870         TALLOC_FREE(frame);
 871         return ret;
 872 }
 873
 874 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 875         .aio_read = aio_fork_read,
 876         .aio_write = aio_fork_write,
 877         .aio_return_fn = aio_fork_return_fn,
 878         .aio_cancel = aio_fork_cancel,
 879         .aio_error_fn = aio_fork_error_fn,
 880         .aio_suspend = aio_fork_suspend,
 881 };
 882
 883 NTSTATUS vfs_aio_fork_init(void);
 884 NTSTATUS vfs_aio_fork_init(void)
 885 {
 886         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 887                                 "aio_fork", &vfs_aio_fork_fns);
 888 }