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 #include "lib/sys_rw.h"
  30 #include "lib/sys_rw_data.h"
  31 #include "lib/msghdr.h"
  32
  33 #if !defined(HAVE_STRUCT_MSGHDR_MSG_CONTROL) && !defined(HAVE_STRUCT_MSGHDR_MSG_ACCRIGHTS)
  34 # error Can not pass file descriptors
  35 #endif
  36
  37 #undef recvmsg
  38
  39 #ifndef MAP_FILE
  40 #define MAP_FILE 0
  41 #endif
  42
  43 struct aio_fork_config {
  44         bool erratic_testing_mode;
  45 };
  46
  47 struct mmap_area {
  48         size_t size;
  49         volatile void *ptr;
  50 };
  51
  52 static int mmap_area_destructor(struct mmap_area *area)
  53 {
  54         munmap(discard_const(area->ptr), area->size);
  55         return 0;
  56 }
  57
  58 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
  59 {
  60         struct mmap_area *result;
  61         int fd;
  62
  63         result = talloc(mem_ctx, struct mmap_area);
  64         if (result == NULL) {
  65                 DEBUG(0, ("talloc failed\n"));
  66                 goto fail;
  67         }
  68
  69         fd = open("/dev/zero", O_RDWR);
  70         if (fd == -1) {
  71                 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
  72                           strerror(errno)));
  73                 goto fail;
  74         }
  75
  76         result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
  77                            MAP_SHARED|MAP_FILE, fd, 0);
  78         close(fd);
  79         if (result->ptr == MAP_FAILED) {
  80                 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
  81                 goto fail;
  82         }
  83
  84         result->size = size;
  85         talloc_set_destructor(result, mmap_area_destructor);
  86
  87         return result;
  88
  89 fail:
  90         TALLOC_FREE(result);
  91         return NULL;
  92 }
  93
  94 enum cmd_type {
  95         READ_CMD,
  96         WRITE_CMD,
  97         FSYNC_CMD
  98 };
  99
 100 static const char *cmd_type_str(enum cmd_type cmd)
 101 {
 102         const char *result;
 103
 104         switch (cmd) {
 105         case READ_CMD:
 106                 result = "READ";
 107                 break;
 108         case WRITE_CMD:
 109                 result = "WRITE";
 110                 break;
 111         case FSYNC_CMD:
 112                 result = "FSYNC";
 113                 break;
 114         default:
 115                 result = "<UNKNOWN>";
 116                 break;
 117         }
 118         return result;
 119 }
 120
 121 struct rw_cmd {
 122         size_t n;
 123         off_t offset;
 124         enum cmd_type cmd;
 125         bool erratic_testing_mode;
 126 };
 127
 128 struct rw_ret {
 129         ssize_t size;
 130         int ret_errno;
 131 };
 132
 133 struct aio_child_list;
 134
 135 struct aio_child {
 136         struct aio_child *prev, *next;
 137         struct aio_child_list *list;
 138         pid_t pid;
 139         int sockfd;
 140         struct mmap_area *map;
 141         bool dont_delete;       /* Marked as in use since last cleanup */
 142         bool busy;
 143 };
 144
 145 struct aio_child_list {
 146         struct aio_child *children;
 147         struct tevent_timer *cleanup_event;
 148 };
 149
 150 static void free_aio_children(void **p)
 151 {
 152         TALLOC_FREE(*p);
 153 }
 154
 155 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
 156 {
 157         struct iovec iov[1];
 158         struct msghdr msg = { .msg_iov = iov, .msg_iovlen = 1 };
 159         ssize_t n;
 160         size_t bufsize = msghdr_prep_recv_fds(NULL, NULL, 0, 1);
 161         uint8_t buf[bufsize];
 162
 163         msghdr_prep_recv_fds(&msg, buf, bufsize, 1);
 164
 165         iov[0].iov_base = (void *)ptr;
 166         iov[0].iov_len = nbytes;
 167
 168         do {
 169                 n = recvmsg(fd, &msg, 0);
 170         } while ((n == -1) && (errno == EINTR));
 171
 172         if (n <= 0) {
 173                 return n;
 174         }
 175
 176         {
 177                 size_t num_fds = msghdr_extract_fds(&msg, NULL, 0);
 178                 int fds[num_fds];
 179
 180                 msghdr_extract_fds(&msg, fds, num_fds);
 181
 182                 if (num_fds != 1) {
 183                         size_t i;
 184
 185                         for (i=0; i<num_fds; i++) {
 186                                 close(fds[i]);
 187                         }
 188
 189                         *recvfd = -1;
 190                         return n;
 191                 }
 192
 193                 *recvfd = fds[0];
 194         }
 195
 196         return(n);
 197 }
 198
 199 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
 200 {
 201         struct msghdr msg = {0};
 202         size_t bufsize = msghdr_prep_fds(NULL, NULL, 0, &sendfd, 1);
 203         uint8_t buf[bufsize];
 204         struct iovec iov;
 205         ssize_t sent;
 206
 207         msghdr_prep_fds(&msg, buf, bufsize, &sendfd, 1);
 208
 209         iov.iov_base = (void *)ptr;
 210         iov.iov_len = nbytes;
 211         msg.msg_iov = &iov;
 212         msg.msg_iovlen = 1;
 213
 214         do {
 215                 sent = sendmsg(fd, &msg, 0);
 216         } while ((sent == -1) && (errno == EINTR));
 217
 218         return sent;
 219 }
 220
 221 static void aio_child_cleanup(struct tevent_context *event_ctx,
 222                               struct tevent_timer *te,
 223                               struct timeval now,
 224                               void *private_data)
 225 {
 226         struct aio_child_list *list = talloc_get_type_abort(
 227                 private_data, struct aio_child_list);
 228         struct aio_child *child, *next;
 229
 230         TALLOC_FREE(list->cleanup_event);
 231
 232         for (child = list->children; child != NULL; child = next) {
 233                 next = child->next;
 234
 235                 if (child->busy) {
 236                         DEBUG(10, ("child %d currently active\n",
 237                                    (int)child->pid));
 238                         continue;
 239                 }
 240
 241                 if (child->dont_delete) {
 242                         DEBUG(10, ("Child %d was active since last cleanup\n",
 243                                    (int)child->pid));
 244                         child->dont_delete = false;
 245                         continue;
 246                 }
 247
 248                 DEBUG(10, ("Child %d idle for more than 30 seconds, "
 249                            "deleting\n", (int)child->pid));
 250
 251                 TALLOC_FREE(child);
 252                 child = next;
 253         }
 254
 255         if (list->children != NULL) {
 256                 /*
 257                  * Re-schedule the next cleanup round
 258                  */
 259                 list->cleanup_event = tevent_add_timer(server_event_context(), list,
 260                                                       timeval_add(&now, 30, 0),
 261                                                       aio_child_cleanup, list);
 262
 263         }
 264 }
 265
 266 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 267 {
 268         struct aio_child_list *data = NULL;
 269
 270         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 271                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 272                                         return NULL);
 273         }
 274
 275         if (data == NULL) {
 276                 data = talloc_zero(NULL, struct aio_child_list);
 277                 if (data == NULL) {
 278                         return NULL;
 279                 }
 280         }
 281
 282         /*
 283          * Regardless of whether the child_list had been around or not, make
 284          * sure that we have a cleanup timed event. This timed event will
 285          * delete itself when it finds that no children are around anymore.
 286          */
 287
 288         if (data->cleanup_event == NULL) {
 289                 data->cleanup_event = tevent_add_timer(server_event_context(), data,
 290                                                       timeval_current_ofs(30, 0),
 291                                                       aio_child_cleanup, data);
 292                 if (data->cleanup_event == NULL) {
 293                         TALLOC_FREE(data);
 294                         return NULL;
 295                 }
 296         }
 297
 298         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 299                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 300                                         struct aio_child_list, return False);
 301         }
 302
 303         return data;
 304 }
 305
 306 static void aio_child_loop(int sockfd, struct mmap_area *map)
 307 {
 308         while (true) {
 309                 int fd = -1;
 310                 ssize_t ret;
 311                 struct rw_cmd cmd_struct;
 312                 struct rw_ret ret_struct;
 313
 314                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 315                 if (ret != sizeof(cmd_struct)) {
 316                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 317                                    strerror(errno)));
 318                         exit(1);
 319                 }
 320
 321                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 322                            cmd_type_str(cmd_struct.cmd),
 323                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 324
 325                 if (cmd_struct.erratic_testing_mode) {
 326                         /*
 327                          * For developer testing, we want erratic behaviour for
 328                          * async I/O times
 329                          */
 330                         uint8_t randval;
 331                         unsigned msecs;
 332                         /*
 333                          * use generate_random_buffer, we just forked from a
 334                          * common parent state
 335                          */
 336                         generate_random_buffer(&randval, sizeof(randval));
 337                         msecs = randval + 20;
 338                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 339                         smb_msleep(msecs);
 340                 }
 341
 342                 ZERO_STRUCT(ret_struct);
 343
 344                 switch (cmd_struct.cmd) {
 345                 case READ_CMD:
 346                         ret_struct.size = sys_pread(
 347                                 fd, discard_const(map->ptr), cmd_struct.n,
 348                                 cmd_struct.offset);
 349 #if 0
 350 /* This breaks "make test" when run with aio_fork module. */
 351 #ifdef DEVELOPER
 352                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 353 #endif
 354 #endif
 355                         break;
 356                 case WRITE_CMD:
 357                         ret_struct.size = sys_pwrite(
 358                                 fd, discard_const(map->ptr), cmd_struct.n,
 359                                 cmd_struct.offset);
 360                         break;
 361                 case FSYNC_CMD:
 362                         ret_struct.size = fsync(fd);
 363                         break;
 364                 default:
 365                         ret_struct.size = -1;
 366                         errno = EINVAL;
 367                 }
 368
 369                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 370                            (int)ret_struct.size));
 371
 372                 if (ret_struct.size == -1) {
 373                         ret_struct.ret_errno = errno;
 374                 }
 375
 376                 /*
 377                  * Close the fd before telling our parent we're done. The
 378                  * parent might close and re-open the file very quickly, and
 379                  * with system-level share modes (GPFS) we would get an
 380                  * unjustified SHARING_VIOLATION.
 381                  */
 382                 close(fd);
 383
 384                 ret = write_data(sockfd, (char *)&ret_struct,
 385                                  sizeof(ret_struct));
 386                 if (ret != sizeof(ret_struct)) {
 387                         DEBUG(10, ("could not write ret_struct: %s\n",
 388                                    strerror(errno)));
 389                         exit(2);
 390                 }
 391         }
 392 }
 393
 394 static int aio_child_destructor(struct aio_child *child)
 395 {
 396         char c=0;
 397
 398         SMB_ASSERT(!child->busy);
 399
 400         DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
 401                    (int)child->pid, child->sockfd));
 402
 403         /*
 404          * closing the sockfd makes the child not return from recvmsg() on RHEL
 405          * 5.5 so instead force the child to exit by writing bad data to it
 406          */
 407         write(child->sockfd, &c, sizeof(c));
 408         close(child->sockfd);
 409         DLIST_REMOVE(child->list->children, child);
 410         return 0;
 411 }
 412
 413 /*
 414  * We have to close all fd's in open files, we might incorrectly hold a system
 415  * level share mode on a file.
 416  */
 417
 418 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 419                                          void *private_data)
 420 {
 421         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 422                 close(fsp->fh->fd);
 423                 fsp->fh->fd = -1;
 424         }
 425         return NULL;
 426 }
 427
 428 static int create_aio_child(struct smbd_server_connection *sconn,
 429                             struct aio_child_list *children,
 430                             size_t map_size,
 431                             struct aio_child **presult)
 432 {
 433         struct aio_child *result;
 434         int fdpair[2];
 435         int ret;
 436
 437         fdpair[0] = fdpair[1] = -1;
 438
 439         result = talloc_zero(children, struct aio_child);
 440         if (result == NULL) {
 441                 return ENOMEM;
 442         }
 443
 444         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 445                 ret = errno;
 446                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 447                 goto fail;
 448         }
 449
 450         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 451
 452         result->map = mmap_area_init(result, map_size);
 453         if (result->map == NULL) {
 454                 ret = errno;
 455                 DEBUG(0, ("Could not create mmap area\n"));
 456                 goto fail;
 457         }
 458
 459         result->pid = fork();
 460         if (result->pid == -1) {
 461                 ret = errno;
 462                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 463                 goto fail;
 464         }
 465
 466         if (result->pid == 0) {
 467                 close(fdpair[0]);
 468                 result->sockfd = fdpair[1];
 469                 files_forall(sconn, close_fsp_fd, NULL);
 470                 aio_child_loop(result->sockfd, result->map);
 471         }
 472
 473         DEBUG(10, ("Child %d created with sockfd %d\n",
 474                    (int)result->pid, fdpair[0]));
 475
 476         result->sockfd = fdpair[0];
 477         close(fdpair[1]);
 478
 479         result->list = children;
 480         DLIST_ADD(children->children, result);
 481
 482         talloc_set_destructor(result, aio_child_destructor);
 483
 484         *presult = result;
 485
 486         return 0;
 487
 488  fail:
 489         if (fdpair[0] != -1) close(fdpair[0]);
 490         if (fdpair[1] != -1) close(fdpair[1]);
 491         TALLOC_FREE(result);
 492
 493         return ret;
 494 }
 495
 496 static int get_idle_child(struct vfs_handle_struct *handle,
 497                           struct aio_child **pchild)
 498 {
 499         struct aio_child_list *children;
 500         struct aio_child *child;
 501
 502         children = init_aio_children(handle);
 503         if (children == NULL) {
 504                 return ENOMEM;
 505         }
 506
 507         for (child = children->children; child != NULL; child = child->next) {
 508                 if (!child->busy) {
 509                         break;
 510                 }
 511         }
 512
 513         if (child == NULL) {
 514                 int ret;
 515
 516                 DEBUG(10, ("no idle child found, creating new one\n"));
 517
 518                 ret = create_aio_child(handle->conn->sconn, children,
 519                                           128*1024, &child);
 520                 if (ret != 0) {
 521                         DEBUG(10, ("create_aio_child failed: %s\n",
 522                                    strerror(errno)));
 523                         return ret;
 524                 }
 525         }
 526
 527         child->dont_delete = true;
 528         child->busy = true;
 529
 530         *pchild = child;
 531         return 0;
 532 }
 533
 534 struct aio_fork_pread_state {
 535         struct aio_child *child;
 536         ssize_t ret;
 537         int err;
 538 };
 539
 540 static void aio_fork_pread_done(struct tevent_req *subreq);
 541
 542 static struct tevent_req *aio_fork_pread_send(struct vfs_handle_struct *handle,
 543                                               TALLOC_CTX *mem_ctx,
 544                                               struct tevent_context *ev,
 545                                               struct files_struct *fsp,
 546                                               void *data,
 547                                               size_t n, off_t offset)
 548 {
 549         struct tevent_req *req, *subreq;
 550         struct aio_fork_pread_state *state;
 551         struct rw_cmd cmd;
 552         ssize_t written;
 553         int err;
 554         struct aio_fork_config *config;
 555
 556         SMB_VFS_HANDLE_GET_DATA(handle, config,
 557                                 struct aio_fork_config,
 558                                 return NULL);
 559
 560         req = tevent_req_create(mem_ctx, &state, struct aio_fork_pread_state);
 561         if (req == NULL) {
 562                 return NULL;
 563         }
 564
 565         if (n > 128*1024) {
 566                 /* TODO: support variable buffers */
 567                 tevent_req_error(req, EINVAL);
 568                 return tevent_req_post(req, ev);
 569         }
 570
 571         err = get_idle_child(handle, &state->child);
 572         if (err != 0) {
 573                 tevent_req_error(req, err);
 574                 return tevent_req_post(req, ev);
 575         }
 576
 577         ZERO_STRUCT(cmd);
 578         cmd.n = n;
 579         cmd.offset = offset;
 580         cmd.cmd = READ_CMD;
 581         cmd.erratic_testing_mode = config->erratic_testing_mode;
 582
 583         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 584                    (int)state->child->pid));
 585
 586         /*
 587          * Not making this async. We're writing into an empty unix
 588          * domain socket. This should never block.
 589          */
 590         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 591                            fsp->fh->fd);
 592         if (written == -1) {
 593                 err = errno;
 594
 595                 TALLOC_FREE(state->child);
 596
 597                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 598                 tevent_req_error(req, err);
 599                 return tevent_req_post(req, ev);
 600         }
 601
 602         subreq = read_packet_send(state, ev, state->child->sockfd,
 603                                   sizeof(struct rw_ret), NULL, NULL);
 604         if (tevent_req_nomem(subreq, req)) {
 605                 TALLOC_FREE(state->child); /* we sent sth down */
 606                 return tevent_req_post(req, ev);
 607         }
 608         tevent_req_set_callback(subreq, aio_fork_pread_done, req);
 609         return req;
 610 }
 611
 612 static void aio_fork_pread_done(struct tevent_req *subreq)
 613 {
 614         struct tevent_req *req = tevent_req_callback_data(
 615                 subreq, struct tevent_req);
 616         struct aio_fork_pread_state *state = tevent_req_data(
 617                 req, struct aio_fork_pread_state);
 618         ssize_t nread;
 619         uint8_t *buf;
 620         int err;
 621         struct rw_ret *retbuf;
 622
 623         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 624         TALLOC_FREE(subreq);
 625         if (nread == -1) {
 626                 TALLOC_FREE(state->child);
 627                 tevent_req_error(req, err);
 628                 return;
 629         }
 630
 631         state->child->busy = false;
 632
 633         retbuf = (struct rw_ret *)buf;
 634         state->ret = retbuf->size;
 635         state->err = retbuf->ret_errno;
 636         tevent_req_done(req);
 637 }
 638
 639 static ssize_t aio_fork_pread_recv(struct tevent_req *req, int *err)
 640 {
 641         struct aio_fork_pread_state *state = tevent_req_data(
 642                 req, struct aio_fork_pread_state);
 643
 644         if (tevent_req_is_unix_error(req, err)) {
 645                 return -1;
 646         }
 647         if (state->ret == -1) {
 648                 *err = state->err;
 649         }
 650         return state->ret;
 651 }
 652
 653 struct aio_fork_pwrite_state {
 654         struct aio_child *child;
 655         ssize_t ret;
 656         int err;
 657 };
 658
 659 static void aio_fork_pwrite_done(struct tevent_req *subreq);
 660
 661 static struct tevent_req *aio_fork_pwrite_send(
 662         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 663         struct tevent_context *ev, struct files_struct *fsp,
 664         const void *data, size_t n, off_t offset)
 665 {
 666         struct tevent_req *req, *subreq;
 667         struct aio_fork_pwrite_state *state;
 668         struct rw_cmd cmd;
 669         ssize_t written;
 670         int err;
 671         struct aio_fork_config *config;
 672         SMB_VFS_HANDLE_GET_DATA(handle, config,
 673                                 struct aio_fork_config,
 674                                 return NULL);
 675
 676         req = tevent_req_create(mem_ctx, &state, struct aio_fork_pwrite_state);
 677         if (req == NULL) {
 678                 return NULL;
 679         }
 680
 681         if (n > 128*1024) {
 682                 /* TODO: support variable buffers */
 683                 tevent_req_error(req, EINVAL);
 684                 return tevent_req_post(req, ev);
 685         }
 686
 687         err = get_idle_child(handle, &state->child);
 688         if (err != 0) {
 689                 tevent_req_error(req, err);
 690                 return tevent_req_post(req, ev);
 691         }
 692
 693         ZERO_STRUCT(cmd);
 694         cmd.n = n;
 695         cmd.offset = offset;
 696         cmd.cmd = WRITE_CMD;
 697         cmd.erratic_testing_mode = config->erratic_testing_mode;
 698
 699         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 700                    (int)state->child->pid));
 701
 702         /*
 703          * Not making this async. We're writing into an empty unix
 704          * domain socket. This should never block.
 705          */
 706         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 707                            fsp->fh->fd);
 708         if (written == -1) {
 709                 err = errno;
 710
 711                 TALLOC_FREE(state->child);
 712
 713                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 714                 tevent_req_error(req, err);
 715                 return tevent_req_post(req, ev);
 716         }
 717
 718         subreq = read_packet_send(state, ev, state->child->sockfd,
 719                                   sizeof(struct rw_ret), NULL, NULL);
 720         if (tevent_req_nomem(subreq, req)) {
 721                 TALLOC_FREE(state->child); /* we sent sth down */
 722                 return tevent_req_post(req, ev);
 723         }
 724         tevent_req_set_callback(subreq, aio_fork_pwrite_done, req);
 725         return req;
 726 }
 727
 728 static void aio_fork_pwrite_done(struct tevent_req *subreq)
 729 {
 730         struct tevent_req *req = tevent_req_callback_data(
 731                 subreq, struct tevent_req);
 732         struct aio_fork_pwrite_state *state = tevent_req_data(
 733                 req, struct aio_fork_pwrite_state);
 734         ssize_t nread;
 735         uint8_t *buf;
 736         int err;
 737         struct rw_ret *retbuf;
 738
 739         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 740         TALLOC_FREE(subreq);
 741         if (nread == -1) {
 742                 TALLOC_FREE(state->child);
 743                 tevent_req_error(req, err);
 744                 return;
 745         }
 746
 747         state->child->busy = false;
 748
 749         retbuf = (struct rw_ret *)buf;
 750         state->ret = retbuf->size;
 751         state->err = retbuf->ret_errno;
 752         tevent_req_done(req);
 753 }
 754
 755 static ssize_t aio_fork_pwrite_recv(struct tevent_req *req, int *err)
 756 {
 757         struct aio_fork_pwrite_state *state = tevent_req_data(
 758                 req, struct aio_fork_pwrite_state);
 759
 760         if (tevent_req_is_unix_error(req, err)) {
 761                 return -1;
 762         }
 763         if (state->ret == -1) {
 764                 *err = state->err;
 765         }
 766         return state->ret;
 767 }
 768
 769 struct aio_fork_fsync_state {
 770         struct aio_child *child;
 771         ssize_t ret;
 772         int err;
 773 };
 774
 775 static void aio_fork_fsync_done(struct tevent_req *subreq);
 776
 777 static struct tevent_req *aio_fork_fsync_send(
 778         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 779         struct tevent_context *ev, struct files_struct *fsp)
 780 {
 781         struct tevent_req *req, *subreq;
 782         struct aio_fork_fsync_state *state;
 783         struct rw_cmd cmd;
 784         ssize_t written;
 785         int err;
 786         struct aio_fork_config *config;
 787
 788         SMB_VFS_HANDLE_GET_DATA(handle, config,
 789                                 struct aio_fork_config,
 790                                 return NULL);
 791
 792         req = tevent_req_create(mem_ctx, &state, struct aio_fork_fsync_state);
 793         if (req == NULL) {
 794                 return NULL;
 795         }
 796
 797         err = get_idle_child(handle, &state->child);
 798         if (err != 0) {
 799                 tevent_req_error(req, err);
 800                 return tevent_req_post(req, ev);
 801         }
 802
 803         ZERO_STRUCT(cmd);
 804         cmd.cmd = FSYNC_CMD;
 805         cmd.erratic_testing_mode = config->erratic_testing_mode;
 806
 807         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 808                    (int)state->child->pid));
 809
 810         /*
 811          * Not making this async. We're writing into an empty unix
 812          * domain socket. This should never block.
 813          */
 814         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 815                            fsp->fh->fd);
 816         if (written == -1) {
 817                 err = errno;
 818
 819                 TALLOC_FREE(state->child);
 820
 821                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 822                 tevent_req_error(req, err);
 823                 return tevent_req_post(req, ev);
 824         }
 825
 826         subreq = read_packet_send(state, ev, state->child->sockfd,
 827                                   sizeof(struct rw_ret), NULL, NULL);
 828         if (tevent_req_nomem(subreq, req)) {
 829                 TALLOC_FREE(state->child); /* we sent sth down */
 830                 return tevent_req_post(req, ev);
 831         }
 832         tevent_req_set_callback(subreq, aio_fork_fsync_done, req);
 833         return req;
 834 }
 835
 836 static void aio_fork_fsync_done(struct tevent_req *subreq)
 837 {
 838         struct tevent_req *req = tevent_req_callback_data(
 839                 subreq, struct tevent_req);
 840         struct aio_fork_fsync_state *state = tevent_req_data(
 841                 req, struct aio_fork_fsync_state);
 842         ssize_t nread;
 843         uint8_t *buf;
 844         int err;
 845         struct rw_ret *retbuf;
 846
 847         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 848         TALLOC_FREE(subreq);
 849         if (nread == -1) {
 850                 TALLOC_FREE(state->child);
 851                 tevent_req_error(req, err);
 852                 return;
 853         }
 854
 855         state->child->busy = false;
 856
 857         retbuf = (struct rw_ret *)buf;
 858         state->ret = retbuf->size;
 859         state->err = retbuf->ret_errno;
 860         tevent_req_done(req);
 861 }
 862
 863 static int aio_fork_fsync_recv(struct tevent_req *req, int *err)
 864 {
 865         struct aio_fork_fsync_state *state = tevent_req_data(
 866                 req, struct aio_fork_fsync_state);
 867
 868         if (tevent_req_is_unix_error(req, err)) {
 869                 return -1;
 870         }
 871         if (state->ret == -1) {
 872                 *err = state->err;
 873         }
 874         return state->ret;
 875 }
 876
 877 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
 878                             const char *user)
 879 {
 880         int ret;
 881         struct aio_fork_config *config;
 882         ret = SMB_VFS_NEXT_CONNECT(handle, service, user);
 883
 884         if (ret < 0) {
 885                 return ret;
 886         }
 887
 888         config = talloc_zero(handle->conn, struct aio_fork_config);
 889         if (!config) {
 890                 SMB_VFS_NEXT_DISCONNECT(handle);
 891                 DEBUG(0, ("talloc_zero() failed\n"));
 892                 return -1;
 893         }
 894
 895         config->erratic_testing_mode = lp_parm_bool(SNUM(handle->conn), "vfs_aio_fork",
 896                                                     "erratic_testing_mode", false);
 897
 898         SMB_VFS_HANDLE_SET_DATA(handle, config,
 899                                 NULL, struct aio_fork_config,
 900                                 return -1);
 901
 902         /*********************************************************************
 903          * How many threads to initialize ?
 904          * 100 per process seems insane as a default until you realize that
 905          * (a) Threads terminate after 1 second when idle.
 906          * (b) Throttling is done in SMB2 via the crediting algorithm.
 907          * (c) SMB1 clients are limited to max_mux (50) outstanding
 908          *     requests and Windows clients don't use this anyway.
 909          * Essentially we want this to be unlimited unless smb.conf
 910          * says different.
 911          *********************************************************************/
 912         aio_pending_size = 100;
 913         return 0;
 914 }
 915
 916 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 917         .connect_fn = aio_fork_connect,
 918         .pread_send_fn = aio_fork_pread_send,
 919         .pread_recv_fn = aio_fork_pread_recv,
 920         .pwrite_send_fn = aio_fork_pwrite_send,
 921         .pwrite_recv_fn = aio_fork_pwrite_recv,
 922         .fsync_send_fn = aio_fork_fsync_send,
 923         .fsync_recv_fn = aio_fork_fsync_recv,
 924 };
 925
 926 NTSTATUS vfs_aio_fork_init(void);
 927 NTSTATUS vfs_aio_fork_init(void)
 928 {
 929         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 930                                 "aio_fork", &vfs_aio_fork_fns);
 931 }