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((void *)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 msghdr msg;
 158         struct iovec iov[1];
 159         ssize_t n;
 160 #ifndef HAVE_STRUCT_MSGHDR_MSG_CONTROL
 161         int newfd;
 162
 163         ZERO_STRUCT(msg);
 164         msg.msg_accrights = (caddr_t) &newfd;
 165         msg.msg_accrightslen = sizeof(int);
 166 #else
 167
 168         union {
 169           struct cmsghdr        cm;
 170           char                          control[CMSG_SPACE(sizeof(int))];
 171         } control_un;
 172         struct cmsghdr  *cmptr;
 173
 174         ZERO_STRUCT(msg);
 175         ZERO_STRUCT(control_un);
 176
 177         msg.msg_control = control_un.control;
 178         msg.msg_controllen = sizeof(control_un.control);
 179 #endif
 180
 181         msg.msg_name = NULL;
 182         msg.msg_namelen = 0;
 183
 184         iov[0].iov_base = (void *)ptr;
 185         iov[0].iov_len = nbytes;
 186         msg.msg_iov = iov;
 187         msg.msg_iovlen = 1;
 188
 189         if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
 190                 return(n);
 191         }
 192
 193 #ifdef  HAVE_STRUCT_MSGHDR_MSG_CONTROL
 194         if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
 195             && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
 196                 if (cmptr->cmsg_level != SOL_SOCKET) {
 197                         DEBUG(10, ("control level != SOL_SOCKET"));
 198                         errno = EINVAL;
 199                         return -1;
 200                 }
 201                 if (cmptr->cmsg_type != SCM_RIGHTS) {
 202                         DEBUG(10, ("control type != SCM_RIGHTS"));
 203                         errno = EINVAL;
 204                         return -1;
 205                 }
 206                 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
 207         } else {
 208                 *recvfd = -1;           /* descriptor was not passed */
 209         }
 210 #else
 211         if (msg.msg_accrightslen == sizeof(int)) {
 212                 *recvfd = newfd;
 213         }
 214         else {
 215                 *recvfd = -1;           /* descriptor was not passed */
 216         }
 217 #endif
 218
 219         return(n);
 220 }
 221
 222 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
 223 {
 224         struct msghdr msg;
 225         size_t bufsize = msghdr_prep_fds(NULL, NULL, 0, &sendfd, 1);
 226         uint8_t buf[bufsize];
 227         struct iovec iov;
 228
 229         msghdr_prep_fds(&msg, buf, bufsize, &sendfd, 1);
 230         msg.msg_name = NULL;
 231         msg.msg_namelen = 0;
 232
 233         iov.iov_base = (void *)ptr;
 234         iov.iov_len = nbytes;
 235         msg.msg_iov = &iov;
 236         msg.msg_iovlen = 1;
 237
 238         return (sendmsg(fd, &msg, 0));
 239 }
 240
 241 static void aio_child_cleanup(struct tevent_context *event_ctx,
 242                               struct tevent_timer *te,
 243                               struct timeval now,
 244                               void *private_data)
 245 {
 246         struct aio_child_list *list = talloc_get_type_abort(
 247                 private_data, struct aio_child_list);
 248         struct aio_child *child, *next;
 249
 250         TALLOC_FREE(list->cleanup_event);
 251
 252         for (child = list->children; child != NULL; child = next) {
 253                 next = child->next;
 254
 255                 if (child->busy) {
 256                         DEBUG(10, ("child %d currently active\n",
 257                                    (int)child->pid));
 258                         continue;
 259                 }
 260
 261                 if (child->dont_delete) {
 262                         DEBUG(10, ("Child %d was active since last cleanup\n",
 263                                    (int)child->pid));
 264                         child->dont_delete = false;
 265                         continue;
 266                 }
 267
 268                 DEBUG(10, ("Child %d idle for more than 30 seconds, "
 269                            "deleting\n", (int)child->pid));
 270
 271                 TALLOC_FREE(child);
 272                 child = next;
 273         }
 274
 275         if (list->children != NULL) {
 276                 /*
 277                  * Re-schedule the next cleanup round
 278                  */
 279                 list->cleanup_event = tevent_add_timer(server_event_context(), list,
 280                                                       timeval_add(&now, 30, 0),
 281                                                       aio_child_cleanup, list);
 282
 283         }
 284 }
 285
 286 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 287 {
 288         struct aio_child_list *data = NULL;
 289
 290         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 291                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 292                                         return NULL);
 293         }
 294
 295         if (data == NULL) {
 296                 data = talloc_zero(NULL, struct aio_child_list);
 297                 if (data == NULL) {
 298                         return NULL;
 299                 }
 300         }
 301
 302         /*
 303          * Regardless of whether the child_list had been around or not, make
 304          * sure that we have a cleanup timed event. This timed event will
 305          * delete itself when it finds that no children are around anymore.
 306          */
 307
 308         if (data->cleanup_event == NULL) {
 309                 data->cleanup_event = tevent_add_timer(server_event_context(), data,
 310                                                       timeval_current_ofs(30, 0),
 311                                                       aio_child_cleanup, data);
 312                 if (data->cleanup_event == NULL) {
 313                         TALLOC_FREE(data);
 314                         return NULL;
 315                 }
 316         }
 317
 318         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 319                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 320                                         struct aio_child_list, return False);
 321         }
 322
 323         return data;
 324 }
 325
 326 static void aio_child_loop(int sockfd, struct mmap_area *map)
 327 {
 328         while (true) {
 329                 int fd = -1;
 330                 ssize_t ret;
 331                 struct rw_cmd cmd_struct;
 332                 struct rw_ret ret_struct;
 333
 334                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 335                 if (ret != sizeof(cmd_struct)) {
 336                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 337                                    strerror(errno)));
 338                         exit(1);
 339                 }
 340
 341                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 342                            cmd_type_str(cmd_struct.cmd),
 343                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 344
 345                 if (cmd_struct.erratic_testing_mode) {
 346                         /*
 347                          * For developer testing, we want erratic behaviour for
 348                          * async I/O times
 349                          */
 350                         uint8_t randval;
 351                         unsigned msecs;
 352                         /*
 353                          * use generate_random_buffer, we just forked from a
 354                          * common parent state
 355                          */
 356                         generate_random_buffer(&randval, sizeof(randval));
 357                         msecs = randval + 20;
 358                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 359                         smb_msleep(msecs);
 360                 }
 361
 362                 ZERO_STRUCT(ret_struct);
 363
 364                 switch (cmd_struct.cmd) {
 365                 case READ_CMD:
 366                         ret_struct.size = sys_pread(
 367                                 fd, (void *)map->ptr, cmd_struct.n,
 368                                 cmd_struct.offset);
 369 #if 0
 370 /* This breaks "make test" when run with aio_fork module. */
 371 #ifdef DEVELOPER
 372                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 373 #endif
 374 #endif
 375                         break;
 376                 case WRITE_CMD:
 377                         ret_struct.size = sys_pwrite(
 378                                 fd, (void *)map->ptr, cmd_struct.n,
 379                                 cmd_struct.offset);
 380                         break;
 381                 case FSYNC_CMD:
 382                         ret_struct.size = fsync(fd);
 383                         break;
 384                 default:
 385                         ret_struct.size = -1;
 386                         errno = EINVAL;
 387                 }
 388
 389                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 390                            (int)ret_struct.size));
 391
 392                 if (ret_struct.size == -1) {
 393                         ret_struct.ret_errno = errno;
 394                 }
 395
 396                 /*
 397                  * Close the fd before telling our parent we're done. The
 398                  * parent might close and re-open the file very quickly, and
 399                  * with system-level share modes (GPFS) we would get an
 400                  * unjustified SHARING_VIOLATION.
 401                  */
 402                 close(fd);
 403
 404                 ret = write_data(sockfd, (char *)&ret_struct,
 405                                  sizeof(ret_struct));
 406                 if (ret != sizeof(ret_struct)) {
 407                         DEBUG(10, ("could not write ret_struct: %s\n",
 408                                    strerror(errno)));
 409                         exit(2);
 410                 }
 411         }
 412 }
 413
 414 static int aio_child_destructor(struct aio_child *child)
 415 {
 416         char c=0;
 417
 418         SMB_ASSERT(!child->busy);
 419
 420         DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
 421                         child->pid, child->sockfd));
 422
 423         /*
 424          * closing the sockfd makes the child not return from recvmsg() on RHEL
 425          * 5.5 so instead force the child to exit by writing bad data to it
 426          */
 427         write(child->sockfd, &c, sizeof(c));
 428         close(child->sockfd);
 429         DLIST_REMOVE(child->list->children, child);
 430         return 0;
 431 }
 432
 433 /*
 434  * We have to close all fd's in open files, we might incorrectly hold a system
 435  * level share mode on a file.
 436  */
 437
 438 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 439                                          void *private_data)
 440 {
 441         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 442                 close(fsp->fh->fd);
 443                 fsp->fh->fd = -1;
 444         }
 445         return NULL;
 446 }
 447
 448 static int create_aio_child(struct smbd_server_connection *sconn,
 449                             struct aio_child_list *children,
 450                             size_t map_size,
 451                             struct aio_child **presult)
 452 {
 453         struct aio_child *result;
 454         int fdpair[2];
 455         int ret;
 456
 457         fdpair[0] = fdpair[1] = -1;
 458
 459         result = talloc_zero(children, struct aio_child);
 460         if (result == NULL) {
 461                 return ENOMEM;
 462         }
 463
 464         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 465                 ret = errno;
 466                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 467                 goto fail;
 468         }
 469
 470         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 471
 472         result->map = mmap_area_init(result, map_size);
 473         if (result->map == NULL) {
 474                 ret = errno;
 475                 DEBUG(0, ("Could not create mmap area\n"));
 476                 goto fail;
 477         }
 478
 479         result->pid = fork();
 480         if (result->pid == -1) {
 481                 ret = errno;
 482                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 483                 goto fail;
 484         }
 485
 486         if (result->pid == 0) {
 487                 close(fdpair[0]);
 488                 result->sockfd = fdpair[1];
 489                 files_forall(sconn, close_fsp_fd, NULL);
 490                 aio_child_loop(result->sockfd, result->map);
 491         }
 492
 493         DEBUG(10, ("Child %d created with sockfd %d\n",
 494                         result->pid, fdpair[0]));
 495
 496         result->sockfd = fdpair[0];
 497         close(fdpair[1]);
 498
 499         result->list = children;
 500         DLIST_ADD(children->children, result);
 501
 502         talloc_set_destructor(result, aio_child_destructor);
 503
 504         *presult = result;
 505
 506         return 0;
 507
 508  fail:
 509         if (fdpair[0] != -1) close(fdpair[0]);
 510         if (fdpair[1] != -1) close(fdpair[1]);
 511         TALLOC_FREE(result);
 512
 513         return ret;
 514 }
 515
 516 static int get_idle_child(struct vfs_handle_struct *handle,
 517                           struct aio_child **pchild)
 518 {
 519         struct aio_child_list *children;
 520         struct aio_child *child;
 521
 522         children = init_aio_children(handle);
 523         if (children == NULL) {
 524                 return ENOMEM;
 525         }
 526
 527         for (child = children->children; child != NULL; child = child->next) {
 528                 if (!child->busy) {
 529                         break;
 530                 }
 531         }
 532
 533         if (child == NULL) {
 534                 int ret;
 535
 536                 DEBUG(10, ("no idle child found, creating new one\n"));
 537
 538                 ret = create_aio_child(handle->conn->sconn, children,
 539                                           128*1024, &child);
 540                 if (ret != 0) {
 541                         DEBUG(10, ("create_aio_child failed: %s\n",
 542                                    strerror(errno)));
 543                         return ret;
 544                 }
 545         }
 546
 547         child->dont_delete = true;
 548         child->busy = true;
 549
 550         *pchild = child;
 551         return 0;
 552 }
 553
 554 struct aio_fork_pread_state {
 555         struct aio_child *child;
 556         ssize_t ret;
 557         int err;
 558 };
 559
 560 static void aio_fork_pread_done(struct tevent_req *subreq);
 561
 562 static struct tevent_req *aio_fork_pread_send(struct vfs_handle_struct *handle,
 563                                               TALLOC_CTX *mem_ctx,
 564                                               struct tevent_context *ev,
 565                                               struct files_struct *fsp,
 566                                               void *data,
 567                                               size_t n, off_t offset)
 568 {
 569         struct tevent_req *req, *subreq;
 570         struct aio_fork_pread_state *state;
 571         struct rw_cmd cmd;
 572         ssize_t written;
 573         int err;
 574         struct aio_fork_config *config;
 575
 576         SMB_VFS_HANDLE_GET_DATA(handle, config,
 577                                 struct aio_fork_config,
 578                                 return NULL);
 579
 580         req = tevent_req_create(mem_ctx, &state, struct aio_fork_pread_state);
 581         if (req == NULL) {
 582                 return NULL;
 583         }
 584
 585         if (n > 128*1024) {
 586                 /* TODO: support variable buffers */
 587                 tevent_req_error(req, EINVAL);
 588                 return tevent_req_post(req, ev);
 589         }
 590
 591         err = get_idle_child(handle, &state->child);
 592         if (err != 0) {
 593                 tevent_req_error(req, err);
 594                 return tevent_req_post(req, ev);
 595         }
 596
 597         ZERO_STRUCT(cmd);
 598         cmd.n = n;
 599         cmd.offset = offset;
 600         cmd.cmd = READ_CMD;
 601         cmd.erratic_testing_mode = config->erratic_testing_mode;
 602
 603         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 604                    (int)state->child->pid));
 605
 606         /*
 607          * Not making this async. We're writing into an empty unix
 608          * domain socket. This should never block.
 609          */
 610         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 611                            fsp->fh->fd);
 612         if (written == -1) {
 613                 err = errno;
 614
 615                 TALLOC_FREE(state->child);
 616
 617                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 618                 tevent_req_error(req, err);
 619                 return tevent_req_post(req, ev);
 620         }
 621
 622         subreq = read_packet_send(state, ev, state->child->sockfd,
 623                                   sizeof(struct rw_ret), NULL, NULL);
 624         if (tevent_req_nomem(subreq, req)) {
 625                 TALLOC_FREE(state->child); /* we sent sth down */
 626                 return tevent_req_post(req, ev);
 627         }
 628         tevent_req_set_callback(subreq, aio_fork_pread_done, req);
 629         return req;
 630 }
 631
 632 static void aio_fork_pread_done(struct tevent_req *subreq)
 633 {
 634         struct tevent_req *req = tevent_req_callback_data(
 635                 subreq, struct tevent_req);
 636         struct aio_fork_pread_state *state = tevent_req_data(
 637                 req, struct aio_fork_pread_state);
 638         ssize_t nread;
 639         uint8_t *buf;
 640         int err;
 641         struct rw_ret *retbuf;
 642
 643         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 644         TALLOC_FREE(subreq);
 645         if (nread == -1) {
 646                 TALLOC_FREE(state->child);
 647                 tevent_req_error(req, err);
 648                 return;
 649         }
 650
 651         state->child->busy = false;
 652
 653         retbuf = (struct rw_ret *)buf;
 654         state->ret = retbuf->size;
 655         state->err = retbuf->ret_errno;
 656         tevent_req_done(req);
 657 }
 658
 659 static ssize_t aio_fork_pread_recv(struct tevent_req *req, int *err)
 660 {
 661         struct aio_fork_pread_state *state = tevent_req_data(
 662                 req, struct aio_fork_pread_state);
 663
 664         if (tevent_req_is_unix_error(req, err)) {
 665                 return -1;
 666         }
 667         if (state->ret == -1) {
 668                 *err = state->err;
 669         }
 670         return state->ret;
 671 }
 672
 673 struct aio_fork_pwrite_state {
 674         struct aio_child *child;
 675         ssize_t ret;
 676         int err;
 677 };
 678
 679 static void aio_fork_pwrite_done(struct tevent_req *subreq);
 680
 681 static struct tevent_req *aio_fork_pwrite_send(
 682         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 683         struct tevent_context *ev, struct files_struct *fsp,
 684         const void *data, size_t n, off_t offset)
 685 {
 686         struct tevent_req *req, *subreq;
 687         struct aio_fork_pwrite_state *state;
 688         struct rw_cmd cmd;
 689         ssize_t written;
 690         int err;
 691         struct aio_fork_config *config;
 692         SMB_VFS_HANDLE_GET_DATA(handle, config,
 693                                 struct aio_fork_config,
 694                                 return NULL);
 695
 696         req = tevent_req_create(mem_ctx, &state, struct aio_fork_pwrite_state);
 697         if (req == NULL) {
 698                 return NULL;
 699         }
 700
 701         if (n > 128*1024) {
 702                 /* TODO: support variable buffers */
 703                 tevent_req_error(req, EINVAL);
 704                 return tevent_req_post(req, ev);
 705         }
 706
 707         err = get_idle_child(handle, &state->child);
 708         if (err != 0) {
 709                 tevent_req_error(req, err);
 710                 return tevent_req_post(req, ev);
 711         }
 712
 713         ZERO_STRUCT(cmd);
 714         cmd.n = n;
 715         cmd.offset = offset;
 716         cmd.cmd = WRITE_CMD;
 717         cmd.erratic_testing_mode = config->erratic_testing_mode;
 718
 719         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 720                    (int)state->child->pid));
 721
 722         /*
 723          * Not making this async. We're writing into an empty unix
 724          * domain socket. This should never block.
 725          */
 726         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 727                            fsp->fh->fd);
 728         if (written == -1) {
 729                 err = errno;
 730
 731                 TALLOC_FREE(state->child);
 732
 733                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 734                 tevent_req_error(req, err);
 735                 return tevent_req_post(req, ev);
 736         }
 737
 738         subreq = read_packet_send(state, ev, state->child->sockfd,
 739                                   sizeof(struct rw_ret), NULL, NULL);
 740         if (tevent_req_nomem(subreq, req)) {
 741                 TALLOC_FREE(state->child); /* we sent sth down */
 742                 return tevent_req_post(req, ev);
 743         }
 744         tevent_req_set_callback(subreq, aio_fork_pwrite_done, req);
 745         return req;
 746 }
 747
 748 static void aio_fork_pwrite_done(struct tevent_req *subreq)
 749 {
 750         struct tevent_req *req = tevent_req_callback_data(
 751                 subreq, struct tevent_req);
 752         struct aio_fork_pwrite_state *state = tevent_req_data(
 753                 req, struct aio_fork_pwrite_state);
 754         ssize_t nread;
 755         uint8_t *buf;
 756         int err;
 757         struct rw_ret *retbuf;
 758
 759         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 760         TALLOC_FREE(subreq);
 761         if (nread == -1) {
 762                 TALLOC_FREE(state->child);
 763                 tevent_req_error(req, err);
 764                 return;
 765         }
 766
 767         state->child->busy = false;
 768
 769         retbuf = (struct rw_ret *)buf;
 770         state->ret = retbuf->size;
 771         state->err = retbuf->ret_errno;
 772         tevent_req_done(req);
 773 }
 774
 775 static ssize_t aio_fork_pwrite_recv(struct tevent_req *req, int *err)
 776 {
 777         struct aio_fork_pwrite_state *state = tevent_req_data(
 778                 req, struct aio_fork_pwrite_state);
 779
 780         if (tevent_req_is_unix_error(req, err)) {
 781                 return -1;
 782         }
 783         if (state->ret == -1) {
 784                 *err = state->err;
 785         }
 786         return state->ret;
 787 }
 788
 789 struct aio_fork_fsync_state {
 790         struct aio_child *child;
 791         ssize_t ret;
 792         int err;
 793 };
 794
 795 static void aio_fork_fsync_done(struct tevent_req *subreq);
 796
 797 static struct tevent_req *aio_fork_fsync_send(
 798         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 799         struct tevent_context *ev, struct files_struct *fsp)
 800 {
 801         struct tevent_req *req, *subreq;
 802         struct aio_fork_fsync_state *state;
 803         struct rw_cmd cmd;
 804         ssize_t written;
 805         int err;
 806         struct aio_fork_config *config;
 807
 808         SMB_VFS_HANDLE_GET_DATA(handle, config,
 809                                 struct aio_fork_config,
 810                                 return NULL);
 811
 812         req = tevent_req_create(mem_ctx, &state, struct aio_fork_fsync_state);
 813         if (req == NULL) {
 814                 return NULL;
 815         }
 816
 817         err = get_idle_child(handle, &state->child);
 818         if (err != 0) {
 819                 tevent_req_error(req, err);
 820                 return tevent_req_post(req, ev);
 821         }
 822
 823         ZERO_STRUCT(cmd);
 824         cmd.cmd = FSYNC_CMD;
 825         cmd.erratic_testing_mode = config->erratic_testing_mode;
 826
 827         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 828                    (int)state->child->pid));
 829
 830         /*
 831          * Not making this async. We're writing into an empty unix
 832          * domain socket. This should never block.
 833          */
 834         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 835                            fsp->fh->fd);
 836         if (written == -1) {
 837                 err = errno;
 838
 839                 TALLOC_FREE(state->child);
 840
 841                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 842                 tevent_req_error(req, err);
 843                 return tevent_req_post(req, ev);
 844         }
 845
 846         subreq = read_packet_send(state, ev, state->child->sockfd,
 847                                   sizeof(struct rw_ret), NULL, NULL);
 848         if (tevent_req_nomem(subreq, req)) {
 849                 TALLOC_FREE(state->child); /* we sent sth down */
 850                 return tevent_req_post(req, ev);
 851         }
 852         tevent_req_set_callback(subreq, aio_fork_fsync_done, req);
 853         return req;
 854 }
 855
 856 static void aio_fork_fsync_done(struct tevent_req *subreq)
 857 {
 858         struct tevent_req *req = tevent_req_callback_data(
 859                 subreq, struct tevent_req);
 860         struct aio_fork_fsync_state *state = tevent_req_data(
 861                 req, struct aio_fork_fsync_state);
 862         ssize_t nread;
 863         uint8_t *buf;
 864         int err;
 865         struct rw_ret *retbuf;
 866
 867         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 868         TALLOC_FREE(subreq);
 869         if (nread == -1) {
 870                 TALLOC_FREE(state->child);
 871                 tevent_req_error(req, err);
 872                 return;
 873         }
 874
 875         state->child->busy = false;
 876
 877         retbuf = (struct rw_ret *)buf;
 878         state->ret = retbuf->size;
 879         state->err = retbuf->ret_errno;
 880         tevent_req_done(req);
 881 }
 882
 883 static int aio_fork_fsync_recv(struct tevent_req *req, int *err)
 884 {
 885         struct aio_fork_fsync_state *state = tevent_req_data(
 886                 req, struct aio_fork_fsync_state);
 887
 888         if (tevent_req_is_unix_error(req, err)) {
 889                 return -1;
 890         }
 891         if (state->ret == -1) {
 892                 *err = state->err;
 893         }
 894         return state->ret;
 895 }
 896
 897 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
 898                             const char *user)
 899 {
 900         int ret;
 901         struct aio_fork_config *config;
 902         ret = SMB_VFS_NEXT_CONNECT(handle, service, user);
 903
 904         if (ret < 0) {
 905                 return ret;
 906         }
 907
 908         config = talloc_zero(handle->conn, struct aio_fork_config);
 909         if (!config) {
 910                 SMB_VFS_NEXT_DISCONNECT(handle);
 911                 DEBUG(0, ("talloc_zero() failed\n"));
 912                 return -1;
 913         }
 914
 915         config->erratic_testing_mode = lp_parm_bool(SNUM(handle->conn), "vfs_aio_fork",
 916                                                     "erratic_testing_mode", false);
 917
 918         SMB_VFS_HANDLE_SET_DATA(handle, config,
 919                                 NULL, struct aio_fork_config,
 920                                 return -1);
 921
 922         /*********************************************************************
 923          * How many threads to initialize ?
 924          * 100 per process seems insane as a default until you realize that
 925          * (a) Threads terminate after 1 second when idle.
 926          * (b) Throttling is done in SMB2 via the crediting algorithm.
 927          * (c) SMB1 clients are limited to max_mux (50) outstanding
 928          *     requests and Windows clients don't use this anyway.
 929          * Essentially we want this to be unlimited unless smb.conf
 930          * says different.
 931          *********************************************************************/
 932         aio_pending_size = 100;
 933         return 0;
 934 }
 935
 936 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 937         .connect_fn = aio_fork_connect,
 938         .pread_send_fn = aio_fork_pread_send,
 939         .pread_recv_fn = aio_fork_pread_recv,
 940         .pwrite_send_fn = aio_fork_pwrite_send,
 941         .pwrite_recv_fn = aio_fork_pwrite_recv,
 942         .fsync_send_fn = aio_fork_fsync_send,
 943         .fsync_recv_fn = aio_fork_fsync_recv,
 944 };
 945
 946 NTSTATUS vfs_aio_fork_init(void);
 947 NTSTATUS vfs_aio_fork_init(void)
 948 {
 949         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 950                                 "aio_fork", &vfs_aio_fork_fns);
 951 }