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         msg.msg_flags = 0;
 149
 150         iov[0].iov_base = (void *)ptr;
 151         iov[0].iov_len = nbytes;
 152         msg.msg_iov = iov;
 153         msg.msg_iovlen = 1;
 154
 155         if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
 156                 return(n);
 157         }
 158
 159 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 160         if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
 161             && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
 162                 if (cmptr->cmsg_level != SOL_SOCKET) {
 163                         DEBUG(10, ("control level != SOL_SOCKET"));
 164                         errno = EINVAL;
 165                         return -1;
 166                 }
 167                 if (cmptr->cmsg_type != SCM_RIGHTS) {
 168                         DEBUG(10, ("control type != SCM_RIGHTS"));
 169                         errno = EINVAL;
 170                         return -1;
 171                 }
 172                 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
 173         } else {
 174                 *recvfd = -1;           /* descriptor was not passed */
 175         }
 176 #else
 177         if (msg.msg_accrightslen == sizeof(int)) {
 178                 *recvfd = newfd;
 179         }
 180         else {
 181                 *recvfd = -1;           /* descriptor was not passed */
 182         }
 183 #endif
 184
 185         return(n);
 186 }
 187
 188 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
 189 {
 190         struct msghdr   msg;
 191         struct iovec    iov[1];
 192
 193 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 194         union {
 195                 struct cmsghdr  cm;
 196                 char control[CMSG_SPACE(sizeof(int))];
 197         } control_un;
 198         struct cmsghdr  *cmptr;
 199
 200         ZERO_STRUCT(msg);
 201         ZERO_STRUCT(control_un);
 202
 203         msg.msg_control = control_un.control;
 204         msg.msg_controllen = sizeof(control_un.control);
 205
 206         cmptr = CMSG_FIRSTHDR(&msg);
 207         cmptr->cmsg_len = CMSG_LEN(sizeof(int));
 208         cmptr->cmsg_level = SOL_SOCKET;
 209         cmptr->cmsg_type = SCM_RIGHTS;
 210         memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
 211 #else
 212         ZERO_STRUCT(msg);
 213         msg.msg_accrights = (caddr_t) &sendfd;
 214         msg.msg_accrightslen = sizeof(int);
 215 #endif
 216
 217         msg.msg_name = NULL;
 218         msg.msg_namelen = 0;
 219
 220         ZERO_STRUCT(iov);
 221         iov[0].iov_base = (void *)ptr;
 222         iov[0].iov_len = nbytes;
 223         msg.msg_iov = iov;
 224         msg.msg_iovlen = 1;
 225
 226         return (sendmsg(fd, &msg, 0));
 227 }
 228
 229 static void aio_child_cleanup(struct event_context *event_ctx,
 230                               struct timed_event *te,
 231                               struct timeval now,
 232                               void *private_data)
 233 {
 234         struct aio_child_list *list = talloc_get_type_abort(
 235                 private_data, struct aio_child_list);
 236         struct aio_child *child, *next;
 237
 238         TALLOC_FREE(list->cleanup_event);
 239
 240         for (child = list->children; child != NULL; child = next) {
 241                 next = child->next;
 242
 243                 if (child->aiocb != NULL) {
 244                         DEBUG(10, ("child %d currently active\n",
 245                                    (int)child->pid));
 246                         continue;
 247                 }
 248
 249                 if (child->dont_delete) {
 250                         DEBUG(10, ("Child %d was active since last cleanup\n",
 251                                    (int)child->pid));
 252                         child->dont_delete = false;
 253                         continue;
 254                 }
 255
 256                 DEBUG(10, ("Child %d idle for more than 30 seconds, "
 257                            "deleting\n", (int)child->pid));
 258
 259                 TALLOC_FREE(child);
 260                 child = next;
 261         }
 262
 263         if (list->children != NULL) {
 264                 /*
 265                  * Re-schedule the next cleanup round
 266                  */
 267                 list->cleanup_event = event_add_timed(server_event_context(), list,
 268                                                       timeval_add(&now, 30, 0),
 269                                                       aio_child_cleanup, list);
 270
 271         }
 272 }
 273
 274 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 275 {
 276         struct aio_child_list *data = NULL;
 277
 278         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 279                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 280                                         return NULL);
 281         }
 282
 283         if (data == NULL) {
 284                 data = talloc_zero(NULL, struct aio_child_list);
 285                 if (data == NULL) {
 286                         return NULL;
 287                 }
 288         }
 289
 290         /*
 291          * Regardless of whether the child_list had been around or not, make
 292          * sure that we have a cleanup timed event. This timed event will
 293          * delete itself when it finds that no children are around anymore.
 294          */
 295
 296         if (data->cleanup_event == NULL) {
 297                 data->cleanup_event = event_add_timed(server_event_context(), data,
 298                                                       timeval_current_ofs(30, 0),
 299                                                       aio_child_cleanup, data);
 300                 if (data->cleanup_event == NULL) {
 301                         TALLOC_FREE(data);
 302                         return NULL;
 303                 }
 304         }
 305
 306         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 307                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 308                                         struct aio_child_list, return False);
 309         }
 310
 311         return data;
 312 }
 313
 314 static void aio_child_loop(int sockfd, struct mmap_area *map)
 315 {
 316         while (true) {
 317                 int fd = -1;
 318                 ssize_t ret;
 319                 struct rw_cmd cmd_struct;
 320                 struct rw_ret ret_struct;
 321
 322                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 323                 if (ret != sizeof(cmd_struct)) {
 324                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 325                                    strerror(errno)));
 326                         exit(1);
 327                 }
 328
 329                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 330                            cmd_struct.read_cmd ? "read" : "write",
 331                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 332
 333 #ifdef ENABLE_BUILD_FARM_HACKS
 334                 {
 335                         /*
 336                          * In the build farm, we want erratic behaviour for
 337                          * async I/O times
 338                          */
 339                         uint8_t randval;
 340                         unsigned msecs;
 341                         /*
 342                          * use generate_random_buffer, we just forked from a
 343                          * common parent state
 344                          */
 345                         generate_random_buffer(&randval, sizeof(randval));
 346                         msecs = randval + 20;
 347                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 348                         smb_msleep(msecs);
 349                 }
 350 #endif
 351
 352
 353                 ZERO_STRUCT(ret_struct);
 354
 355                 if (cmd_struct.read_cmd) {
 356                         ret_struct.size = sys_pread(
 357                                 fd, (void *)map->ptr, cmd_struct.n,
 358                                 cmd_struct.offset);
 359 #if 0
 360 /* This breaks "make test" when run with aio_fork module. */
 361 #ifdef ENABLE_BUILD_FARM_HACKS
 362                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 363 #endif
 364 #endif
 365                 }
 366                 else {
 367                         ret_struct.size = sys_pwrite(
 368                                 fd, (void *)map->ptr, cmd_struct.n,
 369                                 cmd_struct.offset);
 370                 }
 371
 372                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 373                            (int)ret_struct.size));
 374
 375                 if (ret_struct.size == -1) {
 376                         ret_struct.ret_errno = errno;
 377                 }
 378
 379                 /*
 380                  * Close the fd before telling our parent we're done. The
 381                  * parent might close and re-open the file very quickly, and
 382                  * with system-level share modes (GPFS) we would get an
 383                  * unjustified SHARING_VIOLATION.
 384                  */
 385                 close(fd);
 386
 387                 ret = write_data(sockfd, (char *)&ret_struct,
 388                                  sizeof(ret_struct));
 389                 if (ret != sizeof(ret_struct)) {
 390                         DEBUG(10, ("could not write ret_struct: %s\n",
 391                                    strerror(errno)));
 392                         exit(2);
 393                 }
 394         }
 395 }
 396
 397 static void handle_aio_completion(struct event_context *event_ctx,
 398                                   struct fd_event *event, uint16 flags,
 399                                   void *p)
 400 {
 401         struct aio_extra *aio_ex = NULL;
 402         struct aio_child *child = (struct aio_child *)p;
 403         NTSTATUS status;
 404
 405         DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
 406
 407         if ((flags & EVENT_FD_READ) == 0) {
 408                 return;
 409         }
 410
 411         status = read_data(child->sockfd, (char *)&child->retval,
 412                            sizeof(child->retval));
 413
 414         if (!NT_STATUS_IS_OK(status)) {
 415                 DEBUG(1, ("aio child %d died: %s\n", (int)child->pid,
 416                           nt_errstr(status)));
 417                 child->retval.size = -1;
 418                 child->retval.ret_errno = EIO;
 419         }
 420
 421         if (child->aiocb == NULL) {
 422                 DEBUG(1, ("Inactive child died\n"));
 423                 TALLOC_FREE(child);
 424                 return;
 425         }
 426
 427         if (child->cancelled) {
 428                 child->aiocb = NULL;
 429                 child->cancelled = false;
 430                 return;
 431         }
 432
 433         if (child->read_cmd && (child->retval.size > 0)) {
 434                 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
 435                 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
 436                        child->retval.size);
 437         }
 438
 439         if (child->called_from_suspend) {
 440                 child->completion_done = true;
 441                 return;
 442         }
 443         aio_ex = (struct aio_extra *)child->aiocb->aio_sigevent.sigev_value.sival_ptr;
 444         smbd_aio_complete_aio_ex(aio_ex);
 445         TALLOC_FREE(aio_ex);
 446 }
 447
 448 static int aio_child_destructor(struct aio_child *child)
 449 {
 450         char c=0;
 451
 452         SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
 453
 454         DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
 455                         child->pid, child->sockfd));
 456
 457         /*
 458          * closing the sockfd makes the child not return from recvmsg() on RHEL
 459          * 5.5 so instead force the child to exit by writing bad data to it
 460          */
 461         write(child->sockfd, &c, sizeof(c));
 462         close(child->sockfd);
 463         DLIST_REMOVE(child->list->children, child);
 464         return 0;
 465 }
 466
 467 /*
 468  * We have to close all fd's in open files, we might incorrectly hold a system
 469  * level share mode on a file.
 470  */
 471
 472 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 473                                          void *private_data)
 474 {
 475         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 476                 close(fsp->fh->fd);
 477                 fsp->fh->fd = -1;
 478         }
 479         return NULL;
 480 }
 481
 482 static NTSTATUS create_aio_child(struct smbd_server_connection *sconn,
 483                                  struct aio_child_list *children,
 484                                  size_t map_size,
 485                                  struct aio_child **presult)
 486 {
 487         struct aio_child *result;
 488         int fdpair[2];
 489         NTSTATUS status;
 490
 491         fdpair[0] = fdpair[1] = -1;
 492
 493         result = talloc_zero(children, struct aio_child);
 494         NT_STATUS_HAVE_NO_MEMORY(result);
 495
 496         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 497                 status = map_nt_error_from_unix(errno);
 498                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 499                 goto fail;
 500         }
 501
 502         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 503
 504         result->map = mmap_area_init(result, map_size);
 505         if (result->map == NULL) {
 506                 status = map_nt_error_from_unix(errno);
 507                 DEBUG(0, ("Could not create mmap area\n"));
 508                 goto fail;
 509         }
 510
 511         result->pid = fork();
 512         if (result->pid == -1) {
 513                 status = map_nt_error_from_unix(errno);
 514                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 515                 goto fail;
 516         }
 517
 518         if (result->pid == 0) {
 519                 close(fdpair[0]);
 520                 result->sockfd = fdpair[1];
 521                 files_forall(sconn, close_fsp_fd, NULL);
 522                 aio_child_loop(result->sockfd, result->map);
 523         }
 524
 525         DEBUG(10, ("Child %d created with sockfd %d\n",
 526                         result->pid, fdpair[0]));
 527
 528         result->sockfd = fdpair[0];
 529         close(fdpair[1]);
 530
 531         result->sock_event = event_add_fd(server_event_context(), result,
 532                                           result->sockfd, EVENT_FD_READ,
 533                                           handle_aio_completion,
 534                                           result);
 535         if (result->sock_event == NULL) {
 536                 status = NT_STATUS_NO_MEMORY;
 537                 DEBUG(0, ("event_add_fd failed\n"));
 538                 goto fail;
 539         }
 540
 541         result->list = children;
 542         DLIST_ADD(children->children, result);
 543
 544         talloc_set_destructor(result, aio_child_destructor);
 545
 546         *presult = result;
 547
 548         return NT_STATUS_OK;
 549
 550  fail:
 551         if (fdpair[0] != -1) close(fdpair[0]);
 552         if (fdpair[1] != -1) close(fdpair[1]);
 553         TALLOC_FREE(result);
 554
 555         return status;
 556 }
 557
 558 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
 559                                struct aio_child **pchild)
 560 {
 561         struct aio_child_list *children;
 562         struct aio_child *child;
 563         NTSTATUS status;
 564
 565         children = init_aio_children(handle);
 566         if (children == NULL) {
 567                 return NT_STATUS_NO_MEMORY;
 568         }
 569
 570         for (child = children->children; child != NULL; child = child->next) {
 571                 if (child->aiocb == NULL) {
 572                         /* idle */
 573                         break;
 574                 }
 575         }
 576
 577         if (child == NULL) {
 578                 DEBUG(10, ("no idle child found, creating new one\n"));
 579
 580                 status = create_aio_child(handle->conn->sconn, children,
 581                                           128*1024, &child);
 582                 if (!NT_STATUS_IS_OK(status)) {
 583                         DEBUG(10, ("create_aio_child failed: %s\n",
 584                                    nt_errstr(status)));
 585                         return status;
 586                 }
 587         }
 588
 589         child->dont_delete = true;
 590
 591         *pchild = child;
 592         return NT_STATUS_OK;
 593 }
 594
 595 static int aio_fork_read(struct vfs_handle_struct *handle,
 596                          struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 597 {
 598         struct aio_child *child;
 599         struct rw_cmd cmd;
 600         ssize_t ret;
 601         NTSTATUS status;
 602
 603         if (aiocb->aio_nbytes > 128*1024) {
 604                 /* TODO: support variable buffers */
 605                 errno = EINVAL;
 606                 return -1;
 607         }
 608
 609         status = get_idle_child(handle, &child);
 610         if (!NT_STATUS_IS_OK(status)) {
 611                 DEBUG(10, ("Could not get an idle child\n"));
 612                 return -1;
 613         }
 614
 615         child->read_cmd = true;
 616         child->aiocb = aiocb;
 617         child->retval.ret_errno = EINPROGRESS;
 618
 619         ZERO_STRUCT(cmd);
 620         cmd.n = aiocb->aio_nbytes;
 621         cmd.offset = aiocb->aio_offset;
 622         cmd.read_cmd = child->read_cmd;
 623
 624         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 625                    (int)child->pid));
 626
 627         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 628         if (ret == -1) {
 629                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 630                 return -1;
 631         }
 632
 633         return 0;
 634 }
 635
 636 static int aio_fork_write(struct vfs_handle_struct *handle,
 637                           struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
 638 {
 639         struct aio_child *child;
 640         struct rw_cmd cmd;
 641         ssize_t ret;
 642         NTSTATUS status;
 643
 644         if (aiocb->aio_nbytes > 128*1024) {
 645                 /* TODO: support variable buffers */
 646                 errno = EINVAL;
 647                 return -1;
 648         }
 649
 650         status = get_idle_child(handle, &child);
 651         if (!NT_STATUS_IS_OK(status)) {
 652                 DEBUG(10, ("Could not get an idle child\n"));
 653                 return -1;
 654         }
 655
 656         child->read_cmd = false;
 657         child->aiocb = aiocb;
 658         child->retval.ret_errno = EINPROGRESS;
 659
 660         memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
 661                aiocb->aio_nbytes);
 662
 663         ZERO_STRUCT(cmd);
 664         cmd.n = aiocb->aio_nbytes;
 665         cmd.offset = aiocb->aio_offset;
 666         cmd.read_cmd = child->read_cmd;
 667
 668         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 669                    (int)child->pid));
 670
 671         ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
 672         if (ret == -1) {
 673                 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
 674                 return -1;
 675         }
 676
 677         return 0;
 678 }
 679
 680 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
 681                                              SMB_STRUCT_AIOCB *aiocb)
 682 {
 683         struct aio_child_list *children;
 684         struct aio_child *child;
 685
 686         children = init_aio_children(handle);
 687         if (children == NULL) {
 688                 return NULL;
 689         }
 690
 691         for (child = children->children; child != NULL; child = child->next) {
 692                 if (child->aiocb == aiocb) {
 693                         return child;
 694                 }
 695         }
 696
 697         return NULL;
 698 }
 699
 700 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
 701                                   struct files_struct *fsp,
 702                                   SMB_STRUCT_AIOCB *aiocb)
 703 {
 704         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 705
 706         if (child == NULL) {
 707                 errno = EINVAL;
 708                 DEBUG(0, ("returning EINVAL\n"));
 709                 return -1;
 710         }
 711
 712         child->aiocb = NULL;
 713
 714         if (child->cancelled) {
 715                 errno = ECANCELED;
 716                 return -1;
 717         }
 718
 719         if (child->retval.size == -1) {
 720                 errno = child->retval.ret_errno;
 721         }
 722
 723         return child->retval.size;
 724 }
 725
 726 static int aio_fork_cancel(struct vfs_handle_struct *handle,
 727                            struct files_struct *fsp,
 728                            SMB_STRUCT_AIOCB *aiocb)
 729 {
 730         struct aio_child_list *children;
 731         struct aio_child *child;
 732
 733         children = init_aio_children(handle);
 734         if (children == NULL) {
 735                 errno = EINVAL;
 736                 return -1;
 737         }
 738
 739         for (child = children->children; child != NULL; child = child->next) {
 740                 if (child->aiocb == NULL) {
 741                         continue;
 742                 }
 743                 if (child->aiocb->aio_fildes != fsp->fh->fd) {
 744                         continue;
 745                 }
 746                 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
 747                         continue;
 748                 }
 749
 750                 /*
 751                  * We let the child do its job, but we discard the result when
 752                  * it's finished.
 753                  */
 754
 755                 child->cancelled = true;
 756         }
 757
 758         return AIO_CANCELED;
 759 }
 760
 761 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
 762                              struct files_struct *fsp,
 763                              SMB_STRUCT_AIOCB *aiocb)
 764 {
 765         struct aio_child *child = aio_fork_find_child(handle, aiocb);
 766
 767         if (child == NULL) {
 768                 errno = EINVAL;
 769                 return -1;
 770         }
 771
 772         return child->retval.ret_errno;
 773 }
 774
 775 static void aio_fork_suspend_timed_out(struct tevent_context *event_ctx,
 776                                         struct tevent_timer *te,
 777                                         struct timeval now,
 778                                         void *private_data)
 779 {
 780         bool *timed_out = (bool *)private_data;
 781         /* Remove this timed event handler. */
 782         TALLOC_FREE(te);
 783         *timed_out = true;
 784 }
 785
 786 static int aio_fork_suspend(struct vfs_handle_struct *handle,
 787                         struct files_struct *fsp,
 788                         const SMB_STRUCT_AIOCB * const aiocb_array[],
 789                         int n,
 790                         const struct timespec *timeout)
 791 {
 792         struct aio_child_list *children = NULL;
 793         TALLOC_CTX *frame = talloc_stackframe();
 794         struct event_context *ev = NULL;
 795         int i;
 796         int ret = -1;
 797         bool timed_out = false;
 798         int err;
 799
 800         children = init_aio_children(handle);
 801         if (children == NULL) {
 802                 errno = EINVAL;
 803                 goto out;
 804         }
 805
 806         /* This is a blocking call, and has to use a sub-event loop. */
 807         ev = event_context_init(frame);
 808         if (ev == NULL) {
 809                 errno = ENOMEM;
 810                 goto out;
 811         }
 812
 813         if (timeout) {
 814                 struct timeval tv = convert_timespec_to_timeval(*timeout);
 815                 struct tevent_timer *te = tevent_add_timer(ev,
 816                                                 frame,
 817                                                 timeval_current_ofs(tv.tv_sec,
 818                                                                     tv.tv_usec),
 819                                                 aio_fork_suspend_timed_out,
 820                                                 &timed_out);
 821                 if (!te) {
 822                         errno = ENOMEM;
 823                         goto out;
 824                 }
 825         }
 826
 827         for (i = 0; i < n; i++) {
 828                 struct aio_child *child = NULL;
 829                 const SMB_STRUCT_AIOCB *aiocb = aiocb_array[i];
 830
 831                 if (!aiocb) {
 832                         continue;
 833                 }
 834
 835                 /*
 836                  * We're going to cheat here. We know that smbd/aio.c
 837                  * only calls this when it's waiting for every single
 838                  * outstanding call to finish on a close, so just wait
 839                  * individually for each IO to complete. We don't care
 840                  * what order they finish - only that they all do. JRA.
 841                  */
 842
 843                 for (child = children->children; child != NULL; child = child->next) {
 844                         struct tevent_fd *event;
 845
 846                         if (child->aiocb == NULL) {
 847                                 continue;
 848                         }
 849                         if (child->aiocb->aio_fildes != fsp->fh->fd) {
 850                                 continue;
 851                         }
 852                         if (child->aiocb != aiocb) {
 853                                 continue;
 854                         }
 855
 856                         if (child->aiocb->aio_sigevent.sigev_value.sival_ptr == NULL) {
 857                                 continue;
 858                         }
 859
 860                         event = event_add_fd(ev,
 861                                              frame,
 862                                              child->sockfd,
 863                                              EVENT_FD_READ,
 864                                              handle_aio_completion,
 865                                              child);
 866                         if (event == NULL) {
 867                                 errno = ENOMEM;
 868                                 goto out;
 869                         }
 870
 871                         child->called_from_suspend = true;
 872
 873                         while (!child->completion_done) {
 874                                 if (tevent_loop_once(ev) == -1) {
 875                                         goto out;
 876                                 }
 877
 878                                 if (timed_out) {
 879                                         errno = EAGAIN;
 880                                         goto out;
 881                                 }
 882                         }
 883                 }
 884         }
 885
 886         ret = 0;
 887
 888   out:
 889
 890         err = errno;
 891         TALLOC_FREE(frame);
 892         errno = err;
 893         return ret;
 894 }
 895
 896 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
 897                             const char *user)
 898 {
 899         /*********************************************************************
 900          * How many threads to initialize ?
 901          * 100 per process seems insane as a default until you realize that
 902          * (a) Threads terminate after 1 second when idle.
 903          * (b) Throttling is done in SMB2 via the crediting algorithm.
 904          * (c) SMB1 clients are limited to max_mux (50) outstanding
 905          *     requests and Windows clients don't use this anyway.
 906          * Essentially we want this to be unlimited unless smb.conf
 907          * says different.
 908          *********************************************************************/
 909         aio_pending_size = 100;
 910         return SMB_VFS_NEXT_CONNECT(handle, service, user);
 911 }
 912
 913 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 914         .connect_fn = aio_fork_connect,
 915         .aio_read_fn = aio_fork_read,
 916         .aio_write_fn = aio_fork_write,
 917         .aio_return_fn = aio_fork_return_fn,
 918         .aio_cancel_fn = aio_fork_cancel,
 919         .aio_error_fn = aio_fork_error_fn,
 920         .aio_suspend_fn = aio_fork_suspend,
 921 };
 922
 923 NTSTATUS vfs_aio_fork_init(void);
 924 NTSTATUS vfs_aio_fork_init(void)
 925 {
 926         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 927                                 "aio_fork", &vfs_aio_fork_fns);
 928 }