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