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