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 #include "smbd/globals.h"
  27
  28 #ifndef MAP_FILE
  29 #define MAP_FILE 0
  30 #endif
  31
  32 struct mmap_area {
  33         size_t size;
  34         volatile void *ptr;
  35 };
  36
  37 static int mmap_area_destructor(struct mmap_area *area)
  38 {
  39         munmap((void *)area->ptr, area->size);
  40         return 0;
  41 }
  42
  43 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
  44 {
  45         struct mmap_area *result;
  46         int fd;
  47
  48         result = talloc(mem_ctx, struct mmap_area);
  49         if (result == NULL) {
  50                 DEBUG(0, ("talloc failed\n"));
  51                 goto fail;
  52         }
  53
  54         fd = open("/dev/zero", O_RDWR);
  55         if (fd == -1) {
  56                 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
  57                           strerror(errno)));
  58                 goto fail;
  59         }
  60
  61         result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
  62                            MAP_SHARED|MAP_FILE, fd, 0);
  63         if (result->ptr == MAP_FAILED) {
  64                 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
  65                 goto fail;
  66         }
  67
  68         close(fd);
  69
  70         result->size = size;
  71         talloc_set_destructor(result, mmap_area_destructor);
  72
  73         return result;
  74
  75 fail:
  76         TALLOC_FREE(result);
  77         return NULL;
  78 }
  79
  80 struct rw_cmd {
  81         size_t n;
  82         off_t offset;
  83         bool read_cmd;
  84 };
  85
  86 struct rw_ret {
  87         ssize_t size;
  88         int ret_errno;
  89 };
  90
  91 struct aio_child_list;
  92
  93 struct aio_child {
  94         struct aio_child *prev, *next;
  95         struct aio_child_list *list;
  96         SMB_STRUCT_AIOCB *aiocb;
  97         pid_t pid;
  98         int sockfd;
  99         struct fd_event *sock_event;
 100         struct rw_ret retval;
 101         struct mmap_area *map;  /* ==NULL means write request */
 102         bool dont_delete;       /* Marked as in use since last cleanup */
 103         bool cancelled;
 104         bool read_cmd;
 105         bool called_from_suspend;
 106         bool completion_done;
 107 };
 108
 109 struct aio_child_list {
 110         struct aio_child *children;
 111         struct timed_event *cleanup_event;
 112 };
 113
 114 static void free_aio_children(void **p)
 115 {
 116         TALLOC_FREE(*p);
 117 }
 118
 119 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
 120 {
 121         struct msghdr msg;
 122         struct iovec iov[1];
 123         ssize_t n;
 124 #ifndef HAVE_MSGHDR_MSG_CONTROL
 125         int newfd;
 126 #endif
 127
 128 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 129         union {
 130           struct cmsghdr        cm;
 131           char                          control[CMSG_SPACE(sizeof(int))];
 132         } control_un;
 133         struct cmsghdr  *cmptr;
 134
 135         msg.msg_control = control_un.control;
 136         msg.msg_controllen = sizeof(control_un.control);
 137 #else
 138 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
 139         msg.msg_accrights = (caddr_t) &newfd;
 140         msg.msg_accrightslen = sizeof(int);
 141 #else
 142 #error Can not pass file descriptors
 143 #endif
 144 #endif
 145
 146         msg.msg_name = NULL;
 147         msg.msg_namelen = 0;
 148
 149         iov[0].iov_base = (void *)ptr;
 150         iov[0].iov_len = nbytes;
 151         msg.msg_iov = iov;
 152         msg.msg_iovlen = 1;
 153
 154         if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
 155                 return(n);
 156         }
 157
 158 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 159         if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
 160             && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
 161                 if (cmptr->cmsg_level != SOL_SOCKET) {
 162                         DEBUG(10, ("control level != SOL_SOCKET"));
 163                         errno = EINVAL;
 164                         return -1;
 165                 }
 166                 if (cmptr->cmsg_type != SCM_RIGHTS) {
 167                         DEBUG(10, ("control type != SCM_RIGHTS"));
 168                         errno = EINVAL;
 169                         return -1;
 170                 }
 171                 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
 172         } else {
 173                 *recvfd = -1;           /* descriptor was not passed */
 174         }
 175 #else
 176         if (msg.msg_accrightslen == sizeof(int)) {
 177                 *recvfd = newfd;
 178         }
 179         else {
 180                 *recvfd = -1;           /* descriptor was not passed */
 181         }
 182 #endif
 183
 184         return(n);
 185 }
 186
 187 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
 188 {
 189         struct msghdr   msg;
 190         struct iovec    iov[1];
 191
 192 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 193         union {
 194                 struct cmsghdr  cm;
 195                 char control[CMSG_SPACE(sizeof(int))];
 196         } control_un;
 197         struct cmsghdr  *cmptr;
 198
 199         ZERO_STRUCT(msg);
 200         ZERO_STRUCT(control_un);
 201
 202         msg.msg_control = control_un.control;
 203         msg.msg_controllen = sizeof(control_un.control);
 204
 205         cmptr = CMSG_FIRSTHDR(&msg);
 206         cmptr->cmsg_len = CMSG_LEN(sizeof(int));
 207         cmptr->cmsg_level = SOL_SOCKET;
 208         cmptr->cmsg_type = SCM_RIGHTS;
 209         memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
 210 #else
 211         ZERO_STRUCT(msg);
 212         msg.msg_accrights = (caddr_t) &sendfd;
 213         msg.msg_accrightslen = sizeof(int);
 214 #endif
 215
 216         msg.msg_name = NULL;
 217         msg.msg_namelen = 0;
 218
 219         ZERO_STRUCT(iov);
 220         iov[0].iov_base = (void *)ptr;
 221         iov[0].iov_len = nbytes;
 222         msg.msg_iov = iov;
 223         msg.msg_iovlen = 1;
 224
 225         return (sendmsg(fd, &msg, 0));
 226 }
 227
 228 static void aio_child_cleanup(struct event_context *event_ctx,
 229                               struct timed_event *te,
 230                               struct timeval now,
 231                               void *private_data)
 232 {
 233         struct aio_child_list *list = talloc_get_type_abort(
 234                 private_data, struct aio_child_list);
 235         struct aio_child *child, *next;
 236
 237         TALLOC_FREE(list->cleanup_event);
 238
 239         for (child = list->children; child != NULL; child = next) {
 240                 next = child->next;
 241
 242                 if (child->aiocb != NULL) {
 243                         DEBUG(10, ("child %d currently active\n",
 244                                    (int)child->pid));
 245                         continue;
 246                 }
 247
 248                 if (child->dont_delete) {
 249                         DEBUG(10, ("Child %d was active since last cleanup\n",
 250                                    (int)child->pid));
 251                         child->dont_delete = false;
 252                         continue;
 253                 }
 254
 255                 DEBUG(10, ("Child %d idle for more than 30 seconds, "
 256                            "deleting\n", (int)child->pid));
 257
 258                 TALLOC_FREE(child);
 259                 child = next;
 260         }
 261
 262         if (list->children != NULL) {
 263                 /*
 264                  * Re-schedule the next cleanup round
 265                  */
 266                 list->cleanup_event = event_add_timed(server_event_context(), list,
 267                                                       timeval_add(&now, 30, 0),
 268                                                       aio_child_cleanup, list);
 269
 270         }
 271 }
 272
 273 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 274 {
 275         struct aio_child_list *data = NULL;
 276
 277         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 278                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 279                                         return NULL);
 280         }
 281
 282         if (data == NULL) {
 283                 data = talloc_zero(NULL, struct aio_child_list);
 284                 if (data == NULL) {
 285                         return NULL;
 286                 }
 287         }
 288
 289         /*
 290          * Regardless of whether the child_list had been around or not, make
 291          * sure that we have a cleanup timed event. This timed event will
 292          * delete itself when it finds that no children are around anymore.
 293          */
 294
 295         if (data->cleanup_event == NULL) {
 296                 data->cleanup_event = event_add_timed(server_event_context(), data,
 297                                                       timeval_current_ofs(30, 0),
 298                                                       aio_child_cleanup, data);
 299                 if (data->cleanup_event == NULL) {
 300                         TALLOC_FREE(data);
 301                         return NULL;
 302                 }
 303         }
 304
 305         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 306                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 307                                         struct aio_child_list, return False);
 308         }
 309
 310         return data;
 311 }
 312
 313 static void aio_child_loop(int sockfd, struct mmap_area *map)
 314 {
 315         while (true) {
 316                 int fd = -1;
 317                 ssize_t ret;
 318                 struct rw_cmd cmd_struct;
 319                 struct rw_ret ret_struct;
 320
 321                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 322                 if (ret != sizeof(cmd_struct)) {
 323                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 324                                    strerror(errno)));
 325                         exit(1);
 326                 }
 327
 328                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 329                            cmd_struct.read_cmd ? "read" : "write",
 330                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 331
 332 #ifdef ENABLE_BUILD_FARM_HACKS
 333                 {
 334                         /*
 335                          * In the build farm, we want erratic behaviour for
 336                          * async I/O times
 337                          */
 338                         uint8_t randval;
 339                         unsigned msecs;
 340                         /*
 341                          * use generate_random_buffer, we just forked from a
 342                          * common parent state
 343                          */
 344                         generate_random_buffer(&randval, sizeof(randval));
 345                         msecs = randval + 20;
 346                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 347                         smb_msleep(msecs);
 348                 }
 349 #endif
 350
 351
 352                 ZERO_STRUCT(ret_struct);
 353
 354                 if (cmd_struct.read_cmd) {
 355                         ret_struct.size = sys_pread(
 356                                 fd, (void *)map->ptr, cmd_struct.n,
 357                                 cmd_struct.offset);
 358 #if 0
 359 /* This breaks "make test" when run with aio_fork module. */
 360 #ifdef ENABLE_BUILD_FARM_HACKS
 361                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 362 #endif
 363 #endif
 364                 }
 365                 else {
 366                         ret_struct.size = sys_pwrite(
 367                                 fd, (void *)map->ptr, cmd_struct.n,
 368                                 cmd_struct.offset);
 369                 }
 370
 371                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 372                            (int)ret_struct.size));
 373
 374                 if (ret_struct.size == -1) {
 375                         ret_struct.ret_errno = errno;
 376                 }
 377
 378                 /*
 379                  * Close the fd before telling our parent we're done. The
 380                  * parent might close and re-open the file very quickly, and
 381                  * with system-level share modes (GPFS) we would get an
 382                  * unjustified SHARING_VIOLATION.
 383                  */
 384                 close(fd);
 385
 386                 ret = write_data(sockfd, (char *)&ret_struct,
 387                                  sizeof(ret_struct));
 388                 if (ret != sizeof(ret_struct)) {
 389                         DEBUG(10, ("could not write ret_struct: %s\n",
 390                                    strerror(errno)));
 391                         exit(2);
 392                 }
 393         }
 394 }
 395
 396 static void handle_aio_completion(struct event_context *event_ctx,
 397                                   struct fd_event *event, uint16 flags,
 398                                   void *p)
 399 {
 400         struct aio_extra *aio_ex = NULL;
 401         struct aio_child *child = (struct aio_child *)p;
 402         NTSTATUS status;
 403
 404         DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
 405
 406         if ((flags & EVENT_FD_READ) == 0) {
 407                 return;
 408         }
 409
 410         status = read_data(child->sockfd, (char *)&child->retval,
 411                            sizeof(child->retval));
 412
 413         if (!NT_STATUS_IS_OK(status)) {
 414                 DEBUG(1, ("aio child %d died: %s\n", (int)child->pid,
 415                           nt_errstr(status)));
 416                 child->retval.size = -1;
 417                 child->retval.ret_errno = EIO;
 418         }
 419
 420         if (child->aiocb == NULL) {
 421                 DEBUG(1, ("Inactive child died\n"));
 422                 TALLOC_FREE(child);
 423                 return;
 424         }
 425
 426         if (child->cancelled) {
 427                 child->aiocb = NULL;
 428                 child->cancelled = false;
 429                 return;
 430         }
 431
 432         if (child->read_cmd && (child->retval.size > 0)) {
 433                 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
 434                 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
 435                        child->retval.size);
 436         }
 437
 438         if (child->called_from_suspend) {
 439                 child->completion_done = true;
 440                 return;
 441         }
 442         aio_ex = (struct aio_extra *)child->aiocb->aio_sigevent.sigev_value.sival_ptr;
 443         smbd_aio_complete_aio_ex(aio_ex);
 444         TALLOC_FREE(aio_ex);
 445 }
 446
 447 static int aio_child_destructor(struct aio_child *child)
 448 {
 449         char c=0;
 450
 451         SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
 452
 453         DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
 454                         child->pid, child->sockfd));
 455
 456         /*
 457          * closing the sockfd makes the child not return from recvmsg() on RHEL
 458          * 5.5 so instead force the child to exit by writing bad data to it
 459          */
 460         write(child->sockfd, &c, sizeof(c));
 461         close(child->sockfd);
 462         DLIST_REMOVE(child->list->children, child);
 463         return 0;
 464 }
 465
 466 /*
 467  * We have to close all fd's in open files, we might incorrectly hold a system
 468  * level share mode on a file.
 469  */
 470
 471 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 472                                          void *private_data)
 473 {
 474         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 475                 close(fsp->fh->fd);
 476                 fsp->fh->fd = -1;
 477         }
 478         return NULL;
 479 }
 480
 481 static NTSTATUS create_aio_child(struct smbd_server_connection *sconn,
 482                                  struct aio_child_list *children,
 483                                  size_t map_size,
 484                                  struct aio_child **presult)
 485 {
 486         struct aio_child *result;
 487         int fdpair[2];
 488         NTSTATUS status;
 489
 490         fdpair[0] = fdpair[1] = -1;
 491
 492         result = talloc_zero(children, struct aio_child);
 493         NT_STATUS_HAVE_NO_MEMORY(result);
 494
 495         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 496                 status = map_nt_error_from_unix(errno);
 497                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 498                 goto fail;
 499         }
 500
 501         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 502
 503         result->map = mmap_area_init(result, map_size);
 504         if (result->map == NULL) {
 505                 status = map_nt_error_from_unix(errno);
 506                 DEBUG(0, ("Could not create mmap area\n"));
 507                 goto fail;
 508         }
 509
 510         result->pid = fork();
 511         if (result->pid == -1) {
 512                 status = map_nt_error_from_unix(errno);
 513                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 514                 goto fail;
 515         }
 516
 517         if (result->pid == 0) {
 518                 close(fdpair[0]);
 519                 result->sockfd = fdpair[1];
 520                 files_forall(sconn, close_fsp_fd, NULL);
 521                 aio_child_loop(result->sockfd, result->map);
 522         }
 523
 524         DEBUG(10, ("Child %d created with sockfd %d\n",
 525                         result->pid, fdpair[0]));
 526
 527         result->sockfd = fdpair[0];
 528         close(fdpair[1]);
 529
 530         result->sock_event = event_add_fd(server_event_context(), result,
 531                                           result->sockfd, EVENT_FD_READ,
 532                                           handle_aio_completion,
 533                                           result);
 534         if (result->sock_event == NULL) {
 535                 status = NT_STATUS_NO_MEMORY;
 536                 DEBUG(0, ("event_add_fd failed\n"));
 537                 goto fail;
 538         }
 539
 540         result->list = children;
 541         DLIST_ADD(children->children, result);
 542
 543         talloc_set_destructor(result, aio_child_destructor);
 544
 545         *presult = result;
 546
 547         return NT_STATUS_OK;
 548
 549  fail:
 550         if (fdpair[0] != -1) close(fdpair[0]);
 551         if (fdpair[1] != -1) close(fdpair[1]);
 552         TALLOC_FREE(result);
 553
 554         return status;
 555 }
 556
 557 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
 558                                struct aio_child **pchild)
 559 {
 560         struct aio_child_list *children;
 561         struct aio_child *child;
 562         NTSTATUS status;
 563
 564         children = init_aio_children(handle);
 565         if (children == NULL) {
 566                 return NT_STATUS_NO_MEMORY;
 567         }
 568
 569         for (child = children->children; child != NULL; child = child->next) {
 570                 if (child->aiocb == NULL) {
 571                         /* idle */
 572                         break;
 573                 }
 574         }
 575
 576         if (child == NULL) {
 577                 DEBUG(10, ("no idle child found, creating new one\n"));
 578
 579                 status = create_aio_child(handle->conn->sconn, children,
 580                                           128*1024, &child);
 581                 if (!NT_STATUS_IS_OK(status)) {
 582                         DEBUG(10, ("create_aio_child failed: %s\n",
 583                                    nt_errstr(status)));
 584                         return status;
 585                 }
 586         }
 587
 588         child->dont_delete = true;
 589
 590         *pchild = child;
 591         return NT_STATUS_OK;
 592 }
 593
 594 static int aio_fork_read(struct vfs_handle_struct *handle,
 595                          struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 596 {
 597         struct aio_child *child;
 598         struct rw_cmd cmd;
 599         ssize_t ret;
 600         NTSTATUS status;
 601
 602         if (aiocb->aio_nbytes > 128*1024) {
 603                 /* TODO: support variable buffers */
 604                 errno = EINVAL;
 605                 return -1;
 606         }
 607
 608         status = get_idle_child(handle, &child);
 609         if (!NT_STATUS_IS_OK(status)) {
 610                 DEBUG(10, ("Could not get an idle child\n"));
 611                 return -1;
 612         }
 613
 614         child->read_cmd = true;
 615         child->aiocb = aiocb;
 616         child->retval.ret_errno = EINPROGRESS;
 617
 618         ZERO_STRUCT(cmd);
 619         cmd.n = aiocb->aio_nbytes;
 620         cmd.offset = aiocb->aio_offset;
 621         cmd.read_cmd = child->read_cmd;
 622
 623         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 624                    (int)child->pid));
 625
 626         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 627         if (ret == -1) {
 628                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 629                 return -1;
 630         }
 631
 632         return 0;
 633 }
 634
 635 static int aio_fork_write(struct vfs_handle_struct *handle,
 636                           struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 637 {
 638         struct aio_child *child;
 639         struct rw_cmd cmd;
 640         ssize_t ret;
 641         NTSTATUS status;
 642
 643         if (aiocb->aio_nbytes > 128*1024) {
 644                 /* TODO: support variable buffers */
 645                 errno = EINVAL;
 646                 return -1;
 647         }
 648
 649         status = get_idle_child(handle, &child);
 650         if (!NT_STATUS_IS_OK(status)) {
 651                 DEBUG(10, ("Could not get an idle child\n"));
 652                 return -1;
 653         }
 654
 655         child->read_cmd = false;
 656         child->aiocb = aiocb;
 657         child->retval.ret_errno = EINPROGRESS;
 658
 659         memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
 660                aiocb->aio_nbytes);
 661
 662         ZERO_STRUCT(cmd);
 663         cmd.n = aiocb->aio_nbytes;
 664         cmd.offset = aiocb->aio_offset;
 665         cmd.read_cmd = child->read_cmd;
 666
 667         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 668                    (int)child->pid));
 669
 670         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 671         if (ret == -1) {
 672                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 673                 return -1;
 674         }
 675
 676         return 0;
 677 }
 678
 679 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
 680                                              SMB_STRUCT_AIOCB *aiocb)
 681 {
 682         struct aio_child_list *children;
 683         struct aio_child *child;
 684
 685         children = init_aio_children(handle);
 686         if (children == NULL) {
 687                 return NULL;
 688         }
 689
 690         for (child = children->children; child != NULL; child = child->next) {
 691                 if (child->aiocb == aiocb) {
 692                         return child;
 693                 }
 694         }
 695
 696         return NULL;
 697 }
 698
 699 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
 700                                   struct files_struct *fsp,
 701                                   SMB_STRUCT_AIOCB *aiocb)
 702 {
 703         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 704
 705         if (child == NULL) {
 706                 errno = EINVAL;
 707                 DEBUG(0, ("returning EINVAL\n"));
 708                 return -1;
 709         }
 710
 711         child->aiocb = NULL;
 712
 713         if (child->cancelled) {
 714                 errno = ECANCELED;
 715                 return -1;
 716         }
 717
 718         if (child->retval.size == -1) {
 719                 errno = child->retval.ret_errno;
 720         }
 721
 722         return child->retval.size;
 723 }
 724
 725 static int aio_fork_cancel(struct vfs_handle_struct *handle,
 726                            struct files_struct *fsp,
 727                            SMB_STRUCT_AIOCB *aiocb)
 728 {
 729         struct aio_child_list *children;
 730         struct aio_child *child;
 731
 732         children = init_aio_children(handle);
 733         if (children == NULL) {
 734                 errno = EINVAL;
 735                 return -1;
 736         }
 737
 738         for (child = children->children; child != NULL; child = child->next) {
 739                 if (child->aiocb == NULL) {
 740                         continue;
 741                 }
 742                 if (child->aiocb->aio_fildes != fsp->fh->fd) {
 743                         continue;
 744                 }
 745                 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
 746                         continue;
 747                 }
 748
 749                 /*
 750                  * We let the child do its job, but we discard the result when
 751                  * it's finished.
 752                  */
 753
 754                 child->cancelled = true;
 755         }
 756
 757         return AIO_CANCELED;
 758 }
 759
 760 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
 761                              struct files_struct *fsp,
 762                              SMB_STRUCT_AIOCB *aiocb)
 763 {
 764         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 765
 766         if (child == NULL) {
 767                 errno = EINVAL;
 768                 return -1;
 769         }
 770
 771         return child->retval.ret_errno;
 772 }
 773
 774 static void aio_fork_suspend_timed_out(struct tevent_context *event_ctx,
 775                                         struct tevent_timer *te,
 776                                         struct timeval now,
 777                                         void *private_data)
 778 {
 779         bool *timed_out = (bool *)private_data;
 780         /* Remove this timed event handler. */
 781         TALLOC_FREE(te);
 782         *timed_out = true;
 783 }
 784
 785 static int aio_fork_suspend(struct vfs_handle_struct *handle,
 786                         struct files_struct *fsp,
 787                         const SMB_STRUCT_AIOCB * const aiocb_array[],
 788                         int n,
 789                         const struct timespec *timeout)
 790 {
 791         struct aio_child_list *children = NULL;
 792         TALLOC_CTX *frame = talloc_stackframe();
 793         struct event_context *ev = NULL;
 794         int i;
 795         int ret = -1;
 796         bool timed_out = false;
 797
 798         children = init_aio_children(handle);
 799         if (children == NULL) {
 800                 errno = EINVAL;
 801                 goto out;
 802         }
 803
 804         /* This is a blocking call, and has to use a sub-event loop. */
 805         ev = event_context_init(frame);
 806         if (ev == NULL) {
 807                 errno = ENOMEM;
 808                 goto out;
 809         }
 810
 811         if (timeout) {
 812                 struct timeval tv = convert_timespec_to_timeval(*timeout);
 813                 struct tevent_timer *te = tevent_add_timer(ev,
 814                                                 frame,
 815                                                 timeval_current_ofs(tv.tv_sec,
 816                                                                     tv.tv_usec),
 817                                                 aio_fork_suspend_timed_out,
 818                                                 &timed_out);
 819                 if (!te) {
 820                         errno = ENOMEM;
 821                         goto out;
 822                 }
 823         }
 824
 825         for (i = 0; i < n; i++) {
 826                 struct aio_child *child = NULL;
 827                 const SMB_STRUCT_AIOCB *aiocb = aiocb_array[i];
 828
 829                 if (!aiocb) {
 830                         continue;
 831                 }
 832
 833                 /*
 834                  * We're going to cheat here. We know that smbd/aio.c
 835                  * only calls this when it's waiting for every single
 836                  * outstanding call to finish on a close, so just wait
 837                  * individually for each IO to complete. We don't care
 838                  * what order they finish - only that they all do. JRA.
 839                  */
 840
 841                 for (child = children->children; child != NULL; child = child->next) {
 842                         struct tevent_fd *event;
 843
 844                         if (child->aiocb == NULL) {
 845                                 continue;
 846                         }
 847                         if (child->aiocb->aio_fildes != fsp->fh->fd) {
 848                                 continue;
 849                         }
 850                         if (child->aiocb != aiocb) {
 851                                 continue;
 852                         }
 853
 854                         if (child->aiocb->aio_sigevent.sigev_value.sival_ptr == NULL) {
 855                                 continue;
 856                         }
 857
 858                         event = event_add_fd(ev,
 859                                              frame,
 860                                              child->sockfd,
 861                                              EVENT_FD_READ,
 862                                              handle_aio_completion,
 863                                              child);
 864
 865                         child->called_from_suspend = true;
 866
 867                         while (!child->completion_done) {
 868                                 if (tevent_loop_once(ev) == -1) {
 869                                         goto out;
 870                                 }
 871
 872                                 if (timed_out) {
 873                                         errno = EAGAIN;
 874                                         goto out;
 875                                 }
 876                         }
 877                 }
 878         }
 879
 880         ret = 0;
 881
 882   out:
 883
 884         TALLOC_FREE(frame);
 885         return ret;
 886 }
 887
 888 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
 889                             const char *user)
 890 {
 891         /*********************************************************************
 892          * How many threads to initialize ?
 893          * 100 per process seems insane as a default until you realize that
 894          * (a) Threads terminate after 1 second when idle.
 895          * (b) Throttling is done in SMB2 via the crediting algorithm.
 896          * (c) SMB1 clients are limited to max_mux (50) outstanding
 897          *     requests and Windows clients don't use this anyway.
 898          * Essentially we want this to be unlimited unless smb.conf
 899          * says different.
 900          *********************************************************************/
 901         aio_pending_size = 100;
 902         return SMB_VFS_NEXT_CONNECT(handle, service, user);
 903 }
 904
 905 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 906         .connect_fn = aio_fork_connect,
 907         .aio_read_fn = aio_fork_read,
 908         .aio_write_fn = aio_fork_write,
 909         .aio_return_fn = aio_fork_return_fn,
 910         .aio_cancel_fn = aio_fork_cancel,
 911         .aio_error_fn = aio_fork_error_fn,
 912         .aio_suspend_fn = aio_fork_suspend,
 913 };
 914
 915 NTSTATUS vfs_aio_fork_init(void);
 916 NTSTATUS vfs_aio_fork_init(void)
 917 {
 918         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 919                                 "aio_fork", &vfs_aio_fork_fns);
 920 }