source3/modules/vfs_aio_fork.c

   1 /*
   2  * Simulate the Posix AIO using mmap/fork
   3  *
   4  * Copyright (C) Volker Lendecke 2008
   5  * Copyright (C) Jeremy Allison 2010
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 2 of the License, or
  10  * (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #include "includes.h"
  23 #include "system/filesys.h"
  24 #include "system/shmem.h"
  25 #include "smbd/smbd.h"
  26 #include "smbd/globals.h"
  27 #include "lib/async_req/async_sock.h"
  28 #include "lib/util/tevent_unix.h"
  29
  30 #undef recvmsg
  31
  32 #ifndef MAP_FILE
  33 #define MAP_FILE 0
  34 #endif
  35
  36 struct mmap_area {
  37         size_t size;
  38         volatile void *ptr;
  39 };
  40
  41 static int mmap_area_destructor(struct mmap_area *area)
  42 {
  43         munmap((void *)area->ptr, area->size);
  44         return 0;
  45 }
  46
  47 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
  48 {
  49         struct mmap_area *result;
  50         int fd;
  51
  52         result = talloc(mem_ctx, struct mmap_area);
  53         if (result == NULL) {
  54                 DEBUG(0, ("talloc failed\n"));
  55                 goto fail;
  56         }
  57
  58         fd = open("/dev/zero", O_RDWR);
  59         if (fd == -1) {
  60                 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
  61                           strerror(errno)));
  62                 goto fail;
  63         }
  64
  65         result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
  66                            MAP_SHARED|MAP_FILE, fd, 0);
  67         if (result->ptr == MAP_FAILED) {
  68                 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
  69                 goto fail;
  70         }
  71
  72         close(fd);
  73
  74         result->size = size;
  75         talloc_set_destructor(result, mmap_area_destructor);
  76
  77         return result;
  78
  79 fail:
  80         TALLOC_FREE(result);
  81         return NULL;
  82 }
  83
  84 enum cmd_type {
  85         READ_CMD,
  86         WRITE_CMD,
  87         FSYNC_CMD
  88 };
  89
  90 static const char *cmd_type_str(enum cmd_type cmd)
  91 {
  92         const char *result;
  93
  94         switch (cmd) {
  95         case READ_CMD:
  96                 result = "READ";
  97                 break;
  98         case WRITE_CMD:
  99                 result = "WRITE";
 100                 break;
 101         case FSYNC_CMD:
 102                 result = "FSYNC";
 103                 break;
 104         default:
 105                 result = "<UNKNOWN>";
 106                 break;
 107         }
 108         return result;
 109 }
 110
 111 struct rw_cmd {
 112         size_t n;
 113         off_t offset;
 114         enum cmd_type cmd;
 115 };
 116
 117 struct rw_ret {
 118         ssize_t size;
 119         int ret_errno;
 120 };
 121
 122 struct aio_child_list;
 123
 124 struct aio_child {
 125         struct aio_child *prev, *next;
 126         struct aio_child_list *list;
 127         pid_t pid;
 128         int sockfd;
 129         struct mmap_area *map;
 130         bool dont_delete;       /* Marked as in use since last cleanup */
 131         bool busy;
 132 };
 133
 134 struct aio_child_list {
 135         struct aio_child *children;
 136         struct timed_event *cleanup_event;
 137 };
 138
 139 static void free_aio_children(void **p)
 140 {
 141         TALLOC_FREE(*p);
 142 }
 143
 144 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
 145 {
 146         struct msghdr msg;
 147         struct iovec iov[1];
 148         ssize_t n;
 149 #ifndef HAVE_MSGHDR_MSG_CONTROL
 150         int newfd;
 151 #endif
 152
 153 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 154         union {
 155           struct cmsghdr        cm;
 156           char                          control[CMSG_SPACE(sizeof(int))];
 157         } control_un;
 158         struct cmsghdr  *cmptr;
 159
 160         msg.msg_control = control_un.control;
 161         msg.msg_controllen = sizeof(control_un.control);
 162 #else
 163 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
 164         msg.msg_accrights = (caddr_t) &newfd;
 165         msg.msg_accrightslen = sizeof(int);
 166 #else
 167 #error Can not pass file descriptors
 168 #endif
 169 #endif
 170
 171         msg.msg_name = NULL;
 172         msg.msg_namelen = 0;
 173         msg.msg_flags = 0;
 174
 175         iov[0].iov_base = (void *)ptr;
 176         iov[0].iov_len = nbytes;
 177         msg.msg_iov = iov;
 178         msg.msg_iovlen = 1;
 179
 180         if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
 181                 return(n);
 182         }
 183
 184 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 185         if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
 186             && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
 187                 if (cmptr->cmsg_level != SOL_SOCKET) {
 188                         DEBUG(10, ("control level != SOL_SOCKET"));
 189                         errno = EINVAL;
 190                         return -1;
 191                 }
 192                 if (cmptr->cmsg_type != SCM_RIGHTS) {
 193                         DEBUG(10, ("control type != SCM_RIGHTS"));
 194                         errno = EINVAL;
 195                         return -1;
 196                 }
 197                 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
 198         } else {
 199                 *recvfd = -1;           /* descriptor was not passed */
 200         }
 201 #else
 202         if (msg.msg_accrightslen == sizeof(int)) {
 203                 *recvfd = newfd;
 204         }
 205         else {
 206                 *recvfd = -1;           /* descriptor was not passed */
 207         }
 208 #endif
 209
 210         return(n);
 211 }
 212
 213 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
 214 {
 215         struct msghdr   msg;
 216         struct iovec    iov[1];
 217
 218 #ifdef  HAVE_MSGHDR_MSG_CONTROL
 219         union {
 220                 struct cmsghdr  cm;
 221                 char control[CMSG_SPACE(sizeof(int))];
 222         } control_un;
 223         struct cmsghdr  *cmptr;
 224
 225         ZERO_STRUCT(msg);
 226         ZERO_STRUCT(control_un);
 227
 228         msg.msg_control = control_un.control;
 229         msg.msg_controllen = sizeof(control_un.control);
 230
 231         cmptr = CMSG_FIRSTHDR(&msg);
 232         cmptr->cmsg_len = CMSG_LEN(sizeof(int));
 233         cmptr->cmsg_level = SOL_SOCKET;
 234         cmptr->cmsg_type = SCM_RIGHTS;
 235         memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
 236 #else
 237         ZERO_STRUCT(msg);
 238         msg.msg_accrights = (caddr_t) &sendfd;
 239         msg.msg_accrightslen = sizeof(int);
 240 #endif
 241
 242         msg.msg_name = NULL;
 243         msg.msg_namelen = 0;
 244
 245         ZERO_STRUCT(iov);
 246         iov[0].iov_base = (void *)ptr;
 247         iov[0].iov_len = nbytes;
 248         msg.msg_iov = iov;
 249         msg.msg_iovlen = 1;
 250
 251         return (sendmsg(fd, &msg, 0));
 252 }
 253
 254 static void aio_child_cleanup(struct event_context *event_ctx,
 255                               struct timed_event *te,
 256                               struct timeval now,
 257                               void *private_data)
 258 {
 259         struct aio_child_list *list = talloc_get_type_abort(
 260                 private_data, struct aio_child_list);
 261         struct aio_child *child, *next;
 262
 263         TALLOC_FREE(list->cleanup_event);
 264
 265         for (child = list->children; child != NULL; child = next) {
 266                 next = child->next;
 267
 268                 if (child->busy) {
 269                         DEBUG(10, ("child %d currently active\n",
 270                                    (int)child->pid));
 271                         continue;
 272                 }
 273
 274                 if (child->dont_delete) {
 275                         DEBUG(10, ("Child %d was active since last cleanup\n",
 276                                    (int)child->pid));
 277                         child->dont_delete = false;
 278                         continue;
 279                 }
 280
 281                 DEBUG(10, ("Child %d idle for more than 30 seconds, "
 282                            "deleting\n", (int)child->pid));
 283
 284                 TALLOC_FREE(child);
 285                 child = next;
 286         }
 287
 288         if (list->children != NULL) {
 289                 /*
 290                  * Re-schedule the next cleanup round
 291                  */
 292                 list->cleanup_event = event_add_timed(server_event_context(), list,
 293                                                       timeval_add(&now, 30, 0),
 294                                                       aio_child_cleanup, list);
 295
 296         }
 297 }
 298
 299 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
 300 {
 301         struct aio_child_list *data = NULL;
 302
 303         if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
 304                 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
 305                                         return NULL);
 306         }
 307
 308         if (data == NULL) {
 309                 data = talloc_zero(NULL, struct aio_child_list);
 310                 if (data == NULL) {
 311                         return NULL;
 312                 }
 313         }
 314
 315         /*
 316          * Regardless of whether the child_list had been around or not, make
 317          * sure that we have a cleanup timed event. This timed event will
 318          * delete itself when it finds that no children are around anymore.
 319          */
 320
 321         if (data->cleanup_event == NULL) {
 322                 data->cleanup_event = event_add_timed(server_event_context(), data,
 323                                                       timeval_current_ofs(30, 0),
 324                                                       aio_child_cleanup, data);
 325                 if (data->cleanup_event == NULL) {
 326                         TALLOC_FREE(data);
 327                         return NULL;
 328                 }
 329         }
 330
 331         if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
 332                 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
 333                                         struct aio_child_list, return False);
 334         }
 335
 336         return data;
 337 }
 338
 339 static void aio_child_loop(int sockfd, struct mmap_area *map)
 340 {
 341         while (true) {
 342                 int fd = -1;
 343                 ssize_t ret;
 344                 struct rw_cmd cmd_struct;
 345                 struct rw_ret ret_struct;
 346
 347                 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
 348                 if (ret != sizeof(cmd_struct)) {
 349                         DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
 350                                    strerror(errno)));
 351                         exit(1);
 352                 }
 353
 354                 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
 355                            cmd_type_str(cmd_struct.cmd),
 356                            (int)cmd_struct.n, (int)cmd_struct.offset, fd));
 357
 358 #ifdef ENABLE_BUILD_FARM_HACKS
 359                 {
 360                         /*
 361                          * In the build farm, we want erratic behaviour for
 362                          * async I/O times
 363                          */
 364                         uint8_t randval;
 365                         unsigned msecs;
 366                         /*
 367                          * use generate_random_buffer, we just forked from a
 368                          * common parent state
 369                          */
 370                         generate_random_buffer(&randval, sizeof(randval));
 371                         msecs = randval + 20;
 372                         DEBUG(10, ("delaying for %u msecs\n", msecs));
 373                         smb_msleep(msecs);
 374                 }
 375 #endif
 376
 377
 378                 ZERO_STRUCT(ret_struct);
 379
 380                 switch (cmd_struct.cmd) {
 381                 case READ_CMD:
 382                         ret_struct.size = sys_pread(
 383                                 fd, (void *)map->ptr, cmd_struct.n,
 384                                 cmd_struct.offset);
 385 #if 0
 386 /* This breaks "make test" when run with aio_fork module. */
 387 #ifdef ENABLE_BUILD_FARM_HACKS
 388                         ret_struct.size = MAX(1, ret_struct.size * 0.9);
 389 #endif
 390 #endif
 391                         break;
 392                 case WRITE_CMD:
 393                         ret_struct.size = sys_pwrite(
 394                                 fd, (void *)map->ptr, cmd_struct.n,
 395                                 cmd_struct.offset);
 396                         break;
 397                 case FSYNC_CMD:
 398                         ret_struct.size = fsync(fd);
 399                         break;
 400                 default:
 401                         ret_struct.size = -1;
 402                         errno = EINVAL;
 403                 }
 404
 405                 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
 406                            (int)ret_struct.size));
 407
 408                 if (ret_struct.size == -1) {
 409                         ret_struct.ret_errno = errno;
 410                 }
 411
 412                 /*
 413                  * Close the fd before telling our parent we're done. The
 414                  * parent might close and re-open the file very quickly, and
 415                  * with system-level share modes (GPFS) we would get an
 416                  * unjustified SHARING_VIOLATION.
 417                  */
 418                 close(fd);
 419
 420                 ret = write_data(sockfd, (char *)&ret_struct,
 421                                  sizeof(ret_struct));
 422                 if (ret != sizeof(ret_struct)) {
 423                         DEBUG(10, ("could not write ret_struct: %s\n",
 424                                    strerror(errno)));
 425                         exit(2);
 426                 }
 427         }
 428 }
 429
 430 static int aio_child_destructor(struct aio_child *child)
 431 {
 432         char c=0;
 433
 434         SMB_ASSERT(!child->busy);
 435
 436         DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
 437                         child->pid, child->sockfd));
 438
 439         /*
 440          * closing the sockfd makes the child not return from recvmsg() on RHEL
 441          * 5.5 so instead force the child to exit by writing bad data to it
 442          */
 443         write(child->sockfd, &c, sizeof(c));
 444         close(child->sockfd);
 445         DLIST_REMOVE(child->list->children, child);
 446         return 0;
 447 }
 448
 449 /*
 450  * We have to close all fd's in open files, we might incorrectly hold a system
 451  * level share mode on a file.
 452  */
 453
 454 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
 455                                          void *private_data)
 456 {
 457         if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
 458                 close(fsp->fh->fd);
 459                 fsp->fh->fd = -1;
 460         }
 461         return NULL;
 462 }
 463
 464 static int create_aio_child(struct smbd_server_connection *sconn,
 465                             struct aio_child_list *children,
 466                             size_t map_size,
 467                             struct aio_child **presult)
 468 {
 469         struct aio_child *result;
 470         int fdpair[2];
 471         int ret;
 472
 473         fdpair[0] = fdpair[1] = -1;
 474
 475         result = talloc_zero(children, struct aio_child);
 476         if (result == NULL) {
 477                 return ENOMEM;
 478         }
 479
 480         if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
 481                 ret = errno;
 482                 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
 483                 goto fail;
 484         }
 485
 486         DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
 487
 488         result->map = mmap_area_init(result, map_size);
 489         if (result->map == NULL) {
 490                 ret = errno;
 491                 DEBUG(0, ("Could not create mmap area\n"));
 492                 goto fail;
 493         }
 494
 495         result->pid = fork();
 496         if (result->pid == -1) {
 497                 ret = errno;
 498                 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
 499                 goto fail;
 500         }
 501
 502         if (result->pid == 0) {
 503                 close(fdpair[0]);
 504                 result->sockfd = fdpair[1];
 505                 files_forall(sconn, close_fsp_fd, NULL);
 506                 aio_child_loop(result->sockfd, result->map);
 507         }
 508
 509         DEBUG(10, ("Child %d created with sockfd %d\n",
 510                         result->pid, fdpair[0]));
 511
 512         result->sockfd = fdpair[0];
 513         close(fdpair[1]);
 514
 515         result->list = children;
 516         DLIST_ADD(children->children, result);
 517
 518         talloc_set_destructor(result, aio_child_destructor);
 519
 520         *presult = result;
 521
 522         return 0;
 523
 524  fail:
 525         if (fdpair[0] != -1) close(fdpair[0]);
 526         if (fdpair[1] != -1) close(fdpair[1]);
 527         TALLOC_FREE(result);
 528
 529         return ret;
 530 }
 531
 532 static int get_idle_child(struct vfs_handle_struct *handle,
 533                           struct aio_child **pchild)
 534 {
 535         struct aio_child_list *children;
 536         struct aio_child *child;
 537
 538         children = init_aio_children(handle);
 539         if (children == NULL) {
 540                 return ENOMEM;
 541         }
 542
 543         for (child = children->children; child != NULL; child = child->next) {
 544                 if (!child->busy) {
 545                         break;
 546                 }
 547         }
 548
 549         if (child == NULL) {
 550                 int ret;
 551
 552                 DEBUG(10, ("no idle child found, creating new one\n"));
 553
 554                 ret = create_aio_child(handle->conn->sconn, children,
 555                                           128*1024, &child);
 556                 if (ret != 0) {
 557                         DEBUG(10, ("create_aio_child failed: %s\n",
 558                                    strerror(errno)));
 559                         return ret;
 560                 }
 561         }
 562
 563         child->dont_delete = true;
 564         child->busy = true;
 565
 566         *pchild = child;
 567         return 0;
 568 }
 569
 570 struct aio_fork_pread_state {
 571         struct aio_child *child;
 572         ssize_t ret;
 573         int err;
 574 };
 575
 576 static void aio_fork_pread_done(struct tevent_req *subreq);
 577
 578 static struct tevent_req *aio_fork_pread_send(struct vfs_handle_struct *handle,
 579                                               TALLOC_CTX *mem_ctx,
 580                                               struct tevent_context *ev,
 581                                               struct files_struct *fsp,
 582                                               void *data,
 583                                               size_t n, off_t offset)
 584 {
 585         struct tevent_req *req, *subreq;
 586         struct aio_fork_pread_state *state;
 587         struct rw_cmd cmd;
 588         ssize_t written;
 589         int err;
 590
 591         req = tevent_req_create(mem_ctx, &state, struct aio_fork_pread_state);
 592         if (req == NULL) {
 593                 return NULL;
 594         }
 595
 596         if (n > 128*1024) {
 597                 /* TODO: support variable buffers */
 598                 tevent_req_error(req, EINVAL);
 599                 return tevent_req_post(req, ev);
 600         }
 601
 602         err = get_idle_child(handle, &state->child);
 603         if (err != 0) {
 604                 tevent_req_error(req, err);
 605                 return tevent_req_post(req, ev);
 606         }
 607
 608         ZERO_STRUCT(cmd);
 609         cmd.n = n;
 610         cmd.offset = offset;
 611         cmd.cmd = READ_CMD;
 612
 613         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 614                    (int)state->child->pid));
 615
 616         /*
 617          * Not making this async. We're writing into an empty unix
 618          * domain socket. This should never block.
 619          */
 620         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 621                            fsp->fh->fd);
 622         if (written == -1) {
 623                 err = errno;
 624
 625                 TALLOC_FREE(state->child);
 626
 627                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 628                 tevent_req_error(req, err);
 629                 return tevent_req_post(req, ev);
 630         }
 631
 632         subreq = read_packet_send(state, ev, state->child->sockfd,
 633                                   sizeof(struct rw_ret), NULL, NULL);
 634         if (tevent_req_nomem(subreq, req)) {
 635                 TALLOC_FREE(state->child); /* we sent sth down */
 636                 return tevent_req_post(req, ev);
 637         }
 638         tevent_req_set_callback(subreq, aio_fork_pread_done, req);
 639         return req;
 640 }
 641
 642 static void aio_fork_pread_done(struct tevent_req *subreq)
 643 {
 644         struct tevent_req *req = tevent_req_callback_data(
 645                 subreq, struct tevent_req);
 646         struct aio_fork_pread_state *state = tevent_req_data(
 647                 req, struct aio_fork_pread_state);
 648         ssize_t nread;
 649         uint8_t *buf;
 650         int err;
 651         struct rw_ret *retbuf;
 652
 653         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 654         TALLOC_FREE(subreq);
 655         if (nread == -1) {
 656                 TALLOC_FREE(state->child);
 657                 tevent_req_error(req, err);
 658                 return;
 659         }
 660
 661         state->child->busy = false;
 662
 663         retbuf = (struct rw_ret *)buf;
 664         state->ret = retbuf->size;
 665         state->err = retbuf->ret_errno;
 666         tevent_req_done(req);
 667 }
 668
 669 static ssize_t aio_fork_pread_recv(struct tevent_req *req, int *err)
 670 {
 671         struct aio_fork_pread_state *state = tevent_req_data(
 672                 req, struct aio_fork_pread_state);
 673
 674         if (tevent_req_is_unix_error(req, err)) {
 675                 return -1;
 676         }
 677         if (state->ret == -1) {
 678                 *err = state->err;
 679         }
 680         return state->ret;
 681 }
 682
 683 struct aio_fork_pwrite_state {
 684         struct aio_child *child;
 685         ssize_t ret;
 686         int err;
 687 };
 688
 689 static void aio_fork_pwrite_done(struct tevent_req *subreq);
 690
 691 static struct tevent_req *aio_fork_pwrite_send(
 692         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 693         struct tevent_context *ev, struct files_struct *fsp,
 694         const void *data, size_t n, off_t offset)
 695 {
 696         struct tevent_req *req, *subreq;
 697         struct aio_fork_pwrite_state *state;
 698         struct rw_cmd cmd;
 699         ssize_t written;
 700         int err;
 701
 702         req = tevent_req_create(mem_ctx, &state, struct aio_fork_pwrite_state);
 703         if (req == NULL) {
 704                 return NULL;
 705         }
 706
 707         if (n > 128*1024) {
 708                 /* TODO: support variable buffers */
 709                 tevent_req_error(req, EINVAL);
 710                 return tevent_req_post(req, ev);
 711         }
 712
 713         err = get_idle_child(handle, &state->child);
 714         if (err != 0) {
 715                 tevent_req_error(req, err);
 716                 return tevent_req_post(req, ev);
 717         }
 718
 719         ZERO_STRUCT(cmd);
 720         cmd.n = n;
 721         cmd.offset = offset;
 722         cmd.cmd = WRITE_CMD;
 723
 724         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 725                    (int)state->child->pid));
 726
 727         /*
 728          * Not making this async. We're writing into an empty unix
 729          * domain socket. This should never block.
 730          */
 731         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 732                            fsp->fh->fd);
 733         if (written == -1) {
 734                 err = errno;
 735
 736                 TALLOC_FREE(state->child);
 737
 738                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 739                 tevent_req_error(req, err);
 740                 return tevent_req_post(req, ev);
 741         }
 742
 743         subreq = read_packet_send(state, ev, state->child->sockfd,
 744                                   sizeof(struct rw_ret), NULL, NULL);
 745         if (tevent_req_nomem(subreq, req)) {
 746                 TALLOC_FREE(state->child); /* we sent sth down */
 747                 return tevent_req_post(req, ev);
 748         }
 749         tevent_req_set_callback(subreq, aio_fork_pwrite_done, req);
 750         return req;
 751 }
 752
 753 static void aio_fork_pwrite_done(struct tevent_req *subreq)
 754 {
 755         struct tevent_req *req = tevent_req_callback_data(
 756                 subreq, struct tevent_req);
 757         struct aio_fork_pwrite_state *state = tevent_req_data(
 758                 req, struct aio_fork_pwrite_state);
 759         ssize_t nread;
 760         uint8_t *buf;
 761         int err;
 762         struct rw_ret *retbuf;
 763
 764         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 765         TALLOC_FREE(subreq);
 766         if (nread == -1) {
 767                 TALLOC_FREE(state->child);
 768                 tevent_req_error(req, err);
 769                 return;
 770         }
 771
 772         state->child->busy = false;
 773
 774         retbuf = (struct rw_ret *)buf;
 775         state->ret = retbuf->size;
 776         state->err = retbuf->ret_errno;
 777         tevent_req_done(req);
 778 }
 779
 780 static ssize_t aio_fork_pwrite_recv(struct tevent_req *req, int *err)
 781 {
 782         struct aio_fork_pwrite_state *state = tevent_req_data(
 783                 req, struct aio_fork_pwrite_state);
 784
 785         if (tevent_req_is_unix_error(req, err)) {
 786                 return -1;
 787         }
 788         if (state->ret == -1) {
 789                 *err = state->err;
 790         }
 791         return state->ret;
 792 }
 793
 794 struct aio_fork_fsync_state {
 795         struct aio_child *child;
 796         ssize_t ret;
 797         int err;
 798 };
 799
 800 static void aio_fork_fsync_done(struct tevent_req *subreq);
 801
 802 static struct tevent_req *aio_fork_fsync_send(
 803         struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
 804         struct tevent_context *ev, struct files_struct *fsp)
 805 {
 806         struct tevent_req *req, *subreq;
 807         struct aio_fork_fsync_state *state;
 808         struct rw_cmd cmd;
 809         ssize_t written;
 810         int err;
 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
 826         DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
 827                    (int)state->child->pid));
 828
 829         /*
 830          * Not making this async. We're writing into an empty unix
 831          * domain socket. This should never block.
 832          */
 833         written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
 834                            fsp->fh->fd);
 835         if (written == -1) {
 836                 err = errno;
 837
 838                 TALLOC_FREE(state->child);
 839
 840                 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
 841                 tevent_req_error(req, err);
 842                 return tevent_req_post(req, ev);
 843         }
 844
 845         subreq = read_packet_send(state, ev, state->child->sockfd,
 846                                   sizeof(struct rw_ret), NULL, NULL);
 847         if (tevent_req_nomem(subreq, req)) {
 848                 TALLOC_FREE(state->child); /* we sent sth down */
 849                 return tevent_req_post(req, ev);
 850         }
 851         tevent_req_set_callback(subreq, aio_fork_fsync_done, req);
 852         return req;
 853 }
 854
 855 static void aio_fork_fsync_done(struct tevent_req *subreq)
 856 {
 857         struct tevent_req *req = tevent_req_callback_data(
 858                 subreq, struct tevent_req);
 859         struct aio_fork_fsync_state *state = tevent_req_data(
 860                 req, struct aio_fork_fsync_state);
 861         ssize_t nread;
 862         uint8_t *buf;
 863         int err;
 864         struct rw_ret *retbuf;
 865
 866         nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
 867         TALLOC_FREE(subreq);
 868         if (nread == -1) {
 869                 TALLOC_FREE(state->child);
 870                 tevent_req_error(req, err);
 871                 return;
 872         }
 873
 874         state->child->busy = false;
 875
 876         retbuf = (struct rw_ret *)buf;
 877         state->ret = retbuf->size;
 878         state->err = retbuf->ret_errno;
 879         tevent_req_done(req);
 880 }
 881
 882 static int aio_fork_fsync_recv(struct tevent_req *req, int *err)
 883 {
 884         struct aio_fork_fsync_state *state = tevent_req_data(
 885                 req, struct aio_fork_fsync_state);
 886
 887         if (tevent_req_is_unix_error(req, err)) {
 888                 return -1;
 889         }
 890         if (state->ret == -1) {
 891                 *err = state->err;
 892         }
 893         return state->ret;
 894 }
 895
 896 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
 897                             const char *user)
 898 {
 899         /*********************************************************************
 900          * How many threads to initialize ?
 901          * 100 per process seems insane as a default until you realize that
 902          * (a) Threads terminate after 1 second when idle.
 903          * (b) Throttling is done in SMB2 via the crediting algorithm.
 904          * (c) SMB1 clients are limited to max_mux (50) outstanding
 905          *     requests and Windows clients don't use this anyway.
 906          * Essentially we want this to be unlimited unless smb.conf
 907          * says different.
 908          *********************************************************************/
 909         aio_pending_size = 100;
 910         return SMB_VFS_NEXT_CONNECT(handle, service, user);
 911 }
 912
 913 static struct vfs_fn_pointers vfs_aio_fork_fns = {
 914         .connect_fn = aio_fork_connect,
 915         .pread_send_fn = aio_fork_pread_send,
 916         .pread_recv_fn = aio_fork_pread_recv,
 917         .pwrite_send_fn = aio_fork_pwrite_send,
 918         .pwrite_recv_fn = aio_fork_pwrite_recv,
 919         .fsync_send_fn = aio_fork_fsync_send,
 920         .fsync_recv_fn = aio_fork_fsync_recv,
 921 };
 922
 923 NTSTATUS vfs_aio_fork_init(void);
 924 NTSTATUS vfs_aio_fork_init(void)
 925 {
 926         return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
 927                                 "aio_fork", &vfs_aio_fork_fns);
 928 }