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