source3/smbd/fileio.c

   1 /*
   2    Unix SMB/Netbios implementation.
   3    Version 1.9.
   4    read/write to a files_struct
   5    Copyright (C) Andrew Tridgell 1992-1998
   6    Copyright (C) Jeremy Allison 2000-2002. - write cache.
   7
   8    This program is free software; you can redistribute it and/or modify
   9    it under the terms of the GNU General Public License as published by
  10    the Free Software Foundation; either version 3 of the License, or
  11    (at your option) any later version.
  12
  13    This program is distributed in the hope that it will be useful,
  14    but WITHOUT ANY WARRANTY; without even the implied warranty of
  15    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16    GNU General Public License for more details.
  17
  18    You should have received a copy of the GNU General Public License
  19    along with this program.  If not, see <http://www.gnu.org/licenses/>.
  20 */
  21
  22 #include "includes.h"
  23 #include "printing.h"
  24 #include "smbd/smbd.h"
  25 #include "smbd/globals.h"
  26
  27 static bool setup_write_cache(files_struct *, SMB_OFF_T);
  28
  29 /****************************************************************************
  30  Read from write cache if we can.
  31 ****************************************************************************/
  32
  33 static bool read_from_write_cache(files_struct *fsp,char *data,SMB_OFF_T pos,size_t n)
  34 {
  35         write_cache *wcp = fsp->wcp;
  36
  37         if(!wcp) {
  38                 return False;
  39         }
  40
  41         if( n > wcp->data_size || pos < wcp->offset || pos + n > wcp->offset + wcp->data_size) {
  42                 return False;
  43         }
  44
  45         memcpy(data, wcp->data + (pos - wcp->offset), n);
  46
  47         DO_PROFILE_INC(writecache_read_hits);
  48
  49         return True;
  50 }
  51
  52 /****************************************************************************
  53  Read from a file.
  54 ****************************************************************************/
  55
  56 ssize_t read_file(files_struct *fsp,char *data,SMB_OFF_T pos,size_t n)
  57 {
  58         ssize_t ret=0,readret;
  59
  60         /* you can't read from print files */
  61         if (fsp->print_file) {
  62                 errno = EBADF;
  63                 return -1;
  64         }
  65
  66         /*
  67          * Serve from write cache if we can.
  68          */
  69
  70         if(read_from_write_cache(fsp, data, pos, n)) {
  71                 fsp->fh->pos = pos + n;
  72                 fsp->fh->position_information = fsp->fh->pos;
  73                 return n;
  74         }
  75
  76         flush_write_cache(fsp, READ_FLUSH);
  77
  78         fsp->fh->pos = pos;
  79
  80         if (n > 0) {
  81 #ifdef DMF_FIX
  82                 int numretries = 3;
  83 tryagain:
  84                 readret = SMB_VFS_PREAD(fsp,data,n,pos);
  85
  86                 if (readret == -1) {
  87                         if ((errno == EAGAIN) && numretries) {
  88                                 DEBUG(3,("read_file EAGAIN retry in 10 seconds\n"));
  89                                 (void)sleep(10);
  90                                 --numretries;
  91                                 goto tryagain;
  92                         }
  93                         return -1;
  94                 }
  95 #else /* NO DMF fix. */
  96                 readret = SMB_VFS_PREAD(fsp,data,n,pos);
  97
  98                 if (readret == -1) {
  99                         return -1;
 100                 }
 101 #endif
 102                 if (readret > 0) {
 103                         ret += readret;
 104                 }
 105         }
 106
 107         DEBUG(10,("read_file (%s): pos = %.0f, size = %lu, returned %lu\n",
 108                   fsp_str_dbg(fsp), (double)pos, (unsigned long)n, (long)ret));
 109
 110         fsp->fh->pos += ret;
 111         fsp->fh->position_information = fsp->fh->pos;
 112
 113         return(ret);
 114 }
 115
 116 /****************************************************************************
 117  *Really* write to a file.
 118 ****************************************************************************/
 119
 120 static ssize_t real_write_file(struct smb_request *req,
 121                                 files_struct *fsp,
 122                                 const char *data,
 123                                 SMB_OFF_T pos,
 124                                 size_t n)
 125 {
 126         ssize_t ret;
 127
 128         if (pos == -1) {
 129                 ret = vfs_write_data(req, fsp, data, n);
 130         } else {
 131                 fsp->fh->pos = pos;
 132                 if (pos && lp_strict_allocate(SNUM(fsp->conn) &&
 133                                 !fsp->is_sparse)) {
 134                         if (vfs_fill_sparse(fsp, pos) == -1) {
 135                                 return -1;
 136                         }
 137                 }
 138                 ret = vfs_pwrite_data(req, fsp, data, n, pos);
 139         }
 140
 141         DEBUG(10,("real_write_file (%s): pos = %.0f, size = %lu, returned %ld\n",
 142                   fsp_str_dbg(fsp), (double)pos, (unsigned long)n, (long)ret));
 143
 144         if (ret != -1) {
 145                 fsp->fh->pos += ret;
 146
 147 /* Yes - this is correct - writes don't update this. JRA. */
 148 /* Found by Samba4 tests. */
 149 #if 0
 150                 fsp->position_information = fsp->pos;
 151 #endif
 152         }
 153
 154         return ret;
 155 }
 156
 157 /****************************************************************************
 158  File size cache change.
 159  Updates size on disk but doesn't flush the cache.
 160 ****************************************************************************/
 161
 162 static int wcp_file_size_change(files_struct *fsp)
 163 {
 164         int ret;
 165         write_cache *wcp = fsp->wcp;
 166
 167         wcp->file_size = wcp->offset + wcp->data_size;
 168         ret = SMB_VFS_FTRUNCATE(fsp, wcp->file_size);
 169         if (ret == -1) {
 170                 DEBUG(0,("wcp_file_size_change (%s): ftruncate of size %.0f "
 171                          "error %s\n", fsp_str_dbg(fsp),
 172                          (double)wcp->file_size, strerror(errno)));
 173         }
 174         return ret;
 175 }
 176
 177 void update_write_time_handler(struct event_context *ctx,
 178                                       struct timed_event *te,
 179                                       struct timeval now,
 180                                       void *private_data)
 181 {
 182         files_struct *fsp = (files_struct *)private_data;
 183
 184         DEBUG(5, ("Update write time on %s\n", fsp_str_dbg(fsp)));
 185
 186         /* change the write time in the open file db. */
 187         (void)set_write_time(fsp->file_id, timespec_current());
 188
 189         /* And notify. */
 190         notify_fname(fsp->conn, NOTIFY_ACTION_MODIFIED,
 191                      FILE_NOTIFY_CHANGE_LAST_WRITE, fsp->fsp_name->base_name);
 192
 193         /* Remove the timed event handler. */
 194         TALLOC_FREE(fsp->update_write_time_event);
 195 }
 196
 197 /*********************************************************
 198  Schedule a write time update for WRITE_TIME_UPDATE_USEC_DELAY
 199  in the future.
 200 *********************************************************/
 201
 202 void trigger_write_time_update(struct files_struct *fsp)
 203 {
 204         int delay;
 205
 206         if (fsp->posix_open) {
 207                 /* Don't use delayed writes on POSIX files. */
 208                 return;
 209         }
 210
 211         if (fsp->write_time_forced) {
 212                 /* No point - "sticky" write times
 213                  * in effect.
 214                  */
 215                 return;
 216         }
 217
 218         /* We need to remember someone did a write
 219          * and update to current time on close. */
 220
 221         fsp->update_write_time_on_close = true;
 222
 223         if (fsp->update_write_time_triggered) {
 224                 /*
 225                  * We only update the write time after 2 seconds
 226                  * on the first normal write. After that
 227                  * no other writes affect this until close.
 228                  */
 229                 return;
 230         }
 231         fsp->update_write_time_triggered = true;
 232
 233         delay = lp_parm_int(SNUM(fsp->conn),
 234                             "smbd", "writetimeupdatedelay",
 235                             WRITE_TIME_UPDATE_USEC_DELAY);
 236
 237         DEBUG(5, ("Update write time %d usec later on %s\n",
 238                   delay, fsp_str_dbg(fsp)));
 239
 240         /* trigger the update 2 seconds later */
 241         fsp->update_write_time_event =
 242                 event_add_timed(smbd_event_context(), NULL,
 243                                 timeval_current_ofs(0, delay),
 244                                 update_write_time_handler, fsp);
 245 }
 246
 247 void trigger_write_time_update_immediate(struct files_struct *fsp)
 248 {
 249         struct smb_file_time ft;
 250
 251         if (fsp->posix_open) {
 252                 /* Don't use delayed writes on POSIX files. */
 253                 return;
 254         }
 255
 256         if (fsp->write_time_forced) {
 257                 /*
 258                  * No point - "sticky" write times
 259                  * in effect.
 260                  */
 261                 return;
 262         }
 263
 264         TALLOC_FREE(fsp->update_write_time_event);
 265         DEBUG(5, ("Update write time immediate on %s\n",
 266                   fsp_str_dbg(fsp)));
 267
 268         /* After an immediate update, reset the trigger. */
 269         fsp->update_write_time_triggered = true;
 270         fsp->update_write_time_on_close = false;
 271
 272         ZERO_STRUCT(ft);
 273         ft.mtime = timespec_current();
 274
 275         /* Update the time in the open file db. */
 276         (void)set_write_time(fsp->file_id, ft.mtime);
 277
 278         /* Now set on disk - takes care of notify. */
 279         (void)smb_set_file_time(fsp->conn, fsp, fsp->fsp_name, &ft, false);
 280 }
 281
 282 /****************************************************************************
 283  Write to a file.
 284 ****************************************************************************/
 285
 286 ssize_t write_file(struct smb_request *req,
 287                         files_struct *fsp,
 288                         const char *data,
 289                         SMB_OFF_T pos,
 290                         size_t n)
 291 {
 292         write_cache *wcp = fsp->wcp;
 293         ssize_t total_written = 0;
 294         int write_path = -1;
 295
 296         if (fsp->print_file) {
 297                 uint32_t t;
 298                 int ret;
 299
 300                 ret = print_spool_write(fsp, data, n, pos, &t);
 301                 if (ret) {
 302                         errno = ret;
 303                         return -1;
 304                 }
 305                 return t;
 306         }
 307
 308         if (!fsp->can_write) {
 309                 errno = EPERM;
 310                 return -1;
 311         }
 312
 313         if (!fsp->modified) {
 314                 fsp->modified = True;
 315
 316                 if (SMB_VFS_FSTAT(fsp, &fsp->fsp_name->st) == 0) {
 317                         trigger_write_time_update(fsp);
 318                         if (!fsp->posix_open &&
 319                                         (lp_store_dos_attributes(SNUM(fsp->conn)) ||
 320                                         MAP_ARCHIVE(fsp->conn))) {
 321                                 int dosmode = dos_mode(fsp->conn, fsp->fsp_name);
 322                                 if (!IS_DOS_ARCHIVE(dosmode)) {
 323                                         file_set_dosmode(fsp->conn, fsp->fsp_name,
 324                                                  dosmode | aARCH, NULL, false);
 325                                 }
 326                         }
 327
 328                         /*
 329                          * If this is the first write and we have an exclusive oplock then setup
 330                          * the write cache.
 331                          */
 332
 333                         if (EXCLUSIVE_OPLOCK_TYPE(fsp->oplock_type) && !wcp) {
 334                                 setup_write_cache(fsp,
 335                                                  fsp->fsp_name->st.st_ex_size);
 336                                 wcp = fsp->wcp;
 337                         }
 338                 }
 339         }
 340
 341 #ifdef WITH_PROFILE
 342         DO_PROFILE_INC(writecache_total_writes);
 343         if (!fsp->oplock_type) {
 344                 DO_PROFILE_INC(writecache_non_oplock_writes);
 345         }
 346 #endif
 347
 348         /*
 349          * If this file is level II oplocked then we need
 350          * to grab the shared memory lock and inform all
 351          * other files with a level II lock that they need
 352          * to flush their read caches. We keep the lock over
 353          * the shared memory area whilst doing this.
 354          */
 355
 356         /* This should actually be improved to span the write. */
 357         contend_level2_oplocks_begin(fsp, LEVEL2_CONTEND_WRITE);
 358         contend_level2_oplocks_end(fsp, LEVEL2_CONTEND_WRITE);
 359
 360 #ifdef WITH_PROFILE
 361         if (profile_p && profile_p->writecache_total_writes % 500 == 0) {
 362                 DEBUG(3,("WRITECACHE: initwrites=%u abutted=%u total=%u \
 363 nonop=%u allocated=%u active=%u direct=%u perfect=%u readhits=%u\n",
 364                         profile_p->writecache_init_writes,
 365                         profile_p->writecache_abutted_writes,
 366                         profile_p->writecache_total_writes,
 367                         profile_p->writecache_non_oplock_writes,
 368                         profile_p->writecache_allocated_write_caches,
 369                         profile_p->writecache_num_write_caches,
 370                         profile_p->writecache_direct_writes,
 371                         profile_p->writecache_num_perfect_writes,
 372                         profile_p->writecache_read_hits ));
 373
 374                 DEBUG(3,("WRITECACHE: Flushes SEEK=%d, READ=%d, WRITE=%d, READRAW=%d, OPLOCK=%d, CLOSE=%d, SYNC=%d\n",
 375                         profile_p->writecache_flushed_writes[SEEK_FLUSH],
 376                         profile_p->writecache_flushed_writes[READ_FLUSH],
 377                         profile_p->writecache_flushed_writes[WRITE_FLUSH],
 378                         profile_p->writecache_flushed_writes[READRAW_FLUSH],
 379                         profile_p->writecache_flushed_writes[OPLOCK_RELEASE_FLUSH],
 380                         profile_p->writecache_flushed_writes[CLOSE_FLUSH],
 381                         profile_p->writecache_flushed_writes[SYNC_FLUSH] ));
 382         }
 383 #endif
 384
 385         if (wcp && req->unread_bytes) {
 386                 /* If we're using receivefile don't
 387                  * deal with a write cache.
 388                  */
 389                 flush_write_cache(fsp, WRITE_FLUSH);
 390                 delete_write_cache(fsp);
 391                 wcp = NULL;
 392         }
 393
 394         if(!wcp) {
 395                 DO_PROFILE_INC(writecache_direct_writes);
 396                 total_written = real_write_file(req, fsp, data, pos, n);
 397                 return total_written;
 398         }
 399
 400         DEBUG(9,("write_file (%s)(fd=%d pos=%.0f size=%u) wcp->offset=%.0f "
 401                  "wcp->data_size=%u\n", fsp_str_dbg(fsp), fsp->fh->fd,
 402                  (double)pos, (unsigned int)n, (double)wcp->offset,
 403                  (unsigned int)wcp->data_size));
 404
 405         fsp->fh->pos = pos + n;
 406
 407         if ((n == 1) && (data[0] == '\0') && (pos > wcp->file_size)) {
 408                 int ret;
 409
 410                 /*
 411                  * This is a 1-byte write of a 0 beyond the EOF and
 412                  * thus implicitly also beyond the current active
 413                  * write cache, the typical file-extending (and
 414                  * allocating, but we're using the write cache here)
 415                  * write done by Windows. We just have to ftruncate
 416                  * the file and rely on posix semantics to return
 417                  * zeros for non-written file data that is within the
 418                  * file length.
 419                  *
 420                  * We can not use wcp_file_size_change here because we
 421                  * might have an existing write cache, and
 422                  * wcp_file_size_change assumes a change to just the
 423                  * end of the current write cache.
 424                  */
 425
 426                 wcp->file_size = pos + 1;
 427                 ret = SMB_VFS_FTRUNCATE(fsp, wcp->file_size);
 428                 if (ret == -1) {
 429                         DEBUG(0,("wcp_file_size_change (%s): ftruncate of size %.0f"
 430                                  "error %s\n", fsp_str_dbg(fsp),
 431                                  (double)wcp->file_size, strerror(errno)));
 432                         return -1;
 433                 }
 434                 return 1;
 435         }
 436
 437
 438         /*
 439          * If we have active cache and it isn't contiguous then we flush.
 440          * NOTE: There is a small problem with running out of disk ....
 441          */
 442
 443         if (wcp->data_size) {
 444                 bool cache_flush_needed = False;
 445
 446                 if ((pos >= wcp->offset) && (pos <= wcp->offset + wcp->data_size)) {
 447
 448                         /* ASCII art.... JRA.
 449
 450       +--------------+-----
 451       | Cached data  | Rest of allocated cache buffer....
 452       +--------------+-----
 453
 454             +-------------------+
 455             | Data to write     |
 456             +-------------------+
 457
 458                         */
 459
 460                         /*
 461                          * Start of write overlaps or abutts the existing data.
 462                          */
 463
 464                         size_t data_used = MIN((wcp->alloc_size - (pos - wcp->offset)), n);
 465
 466                         memcpy(wcp->data + (pos - wcp->offset), data, data_used);
 467
 468                         /*
 469                          * Update the current buffer size with the new data.
 470                          */
 471
 472                         if(pos + data_used > wcp->offset + wcp->data_size) {
 473                                 wcp->data_size = pos + data_used - wcp->offset;
 474                         }
 475
 476                         /*
 477                          * Update the file size if changed.
 478                          */
 479
 480                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 481                                 if (wcp_file_size_change(fsp) == -1) {
 482                                         return -1;
 483                                 }
 484                         }
 485
 486                         /*
 487                          * If we used all the data then
 488                          * return here.
 489                          */
 490
 491                         if(n == data_used) {
 492                                 return n;
 493                         } else {
 494                                 cache_flush_needed = True;
 495                         }
 496                         /*
 497                          * Move the start of data forward by the amount used,
 498                          * cut down the amount left by the same amount.
 499                          */
 500
 501                         data += data_used;
 502                         pos += data_used;
 503                         n -= data_used;
 504
 505                         DO_PROFILE_INC(writecache_abutted_writes);
 506                         total_written = data_used;
 507
 508                         write_path = 1;
 509
 510                 } else if ((pos < wcp->offset) && (pos + n > wcp->offset) &&
 511                                         (pos + n <= wcp->offset + wcp->alloc_size)) {
 512
 513                         /* ASCII art.... JRA.
 514
 515                         +---------------+
 516                         | Cache buffer  |
 517                         +---------------+
 518
 519             +-------------------+
 520             | Data to write     |
 521             +-------------------+
 522
 523                         */
 524
 525                         /*
 526                          * End of write overlaps the existing data.
 527                          */
 528
 529                         size_t data_used = pos + n - wcp->offset;
 530
 531                         memcpy(wcp->data, data + n - data_used, data_used);
 532
 533                         /*
 534                          * Update the current buffer size with the new data.
 535                          */
 536
 537                         if(pos + n > wcp->offset + wcp->data_size) {
 538                                 wcp->data_size = pos + n - wcp->offset;
 539                         }
 540
 541                         /*
 542                          * Update the file size if changed.
 543                          */
 544
 545                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 546                                 if (wcp_file_size_change(fsp) == -1) {
 547                                         return -1;
 548                                 }
 549                         }
 550
 551                         /*
 552                          * We don't need to move the start of data, but we
 553                          * cut down the amount left by the amount used.
 554                          */
 555
 556                         n -= data_used;
 557
 558                         /*
 559                          * We cannot have used all the data here.
 560                          */
 561
 562                         cache_flush_needed = True;
 563
 564                         DO_PROFILE_INC(writecache_abutted_writes);
 565                         total_written = data_used;
 566
 567                         write_path = 2;
 568
 569                 } else if ( (pos >= wcp->file_size) &&
 570                                         (wcp->offset + wcp->data_size == wcp->file_size) &&
 571                                         (pos > wcp->offset + wcp->data_size) &&
 572                                         (pos < wcp->offset + wcp->alloc_size) ) {
 573
 574                         /* ASCII art.... JRA.
 575
 576                        End of file ---->|
 577
 578                         +---------------+---------------+
 579                         | Cached data   | Cache buffer  |
 580                         +---------------+---------------+
 581
 582                                               +-------------------+
 583                                               | Data to write     |
 584                                               +-------------------+
 585
 586                         */
 587
 588                         /*
 589                          * Non-contiguous write part of which fits within
 590                          * the cache buffer and is extending the file
 591                          * and the cache contents reflect the current
 592                          * data up to the current end of the file.
 593                          */
 594
 595                         size_t data_used;
 596
 597                         if(pos + n <= wcp->offset + wcp->alloc_size) {
 598                                 data_used = n;
 599                         } else {
 600                                 data_used = wcp->offset + wcp->alloc_size - pos;
 601                         }
 602
 603                         /*
 604                          * Fill in the non-continuous area with zeros.
 605                          */
 606
 607                         memset(wcp->data + wcp->data_size, '\0',
 608                                 pos - (wcp->offset + wcp->data_size) );
 609
 610                         memcpy(wcp->data + (pos - wcp->offset), data, data_used);
 611
 612                         /*
 613                          * Update the current buffer size with the new data.
 614                          */
 615
 616                         if(pos + data_used > wcp->offset + wcp->data_size) {
 617                                 wcp->data_size = pos + data_used - wcp->offset;
 618                         }
 619
 620                         /*
 621                          * Update the file size if changed.
 622                          */
 623
 624                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 625                                 if (wcp_file_size_change(fsp) == -1) {
 626                                         return -1;
 627                                 }
 628                         }
 629
 630                         /*
 631                          * If we used all the data then
 632                          * return here.
 633                          */
 634
 635                         if(n == data_used) {
 636                                 return n;
 637                         } else {
 638                                 cache_flush_needed = True;
 639                         }
 640
 641                         /*
 642                          * Move the start of data forward by the amount used,
 643                          * cut down the amount left by the same amount.
 644                          */
 645
 646                         data += data_used;
 647                         pos += data_used;
 648                         n -= data_used;
 649
 650                         DO_PROFILE_INC(writecache_abutted_writes);
 651                         total_written = data_used;
 652
 653                         write_path = 3;
 654
 655                 } else if ( (pos >= wcp->file_size) &&
 656                             (n == 1) &&
 657                             (wcp->file_size == wcp->offset + wcp->data_size) &&
 658                             (pos < wcp->file_size + wcp->alloc_size)) {
 659
 660                         /*
 661
 662                 End of file ---->|
 663
 664                  +---------------+---------------+
 665                  | Cached data   | Cache buffer  |
 666                  +---------------+---------------+
 667
 668                                  |<------- allocated size ---------------->|
 669
 670                                                          +--------+
 671                                                          | 1 Byte |
 672                                                          +--------+
 673
 674                         MS-Office seems to do this a lot to determine if there's enough
 675                         space on the filesystem to write a new file.
 676
 677                         Change to :
 678
 679                 End of file ---->|
 680                                  +-----------------------+--------+
 681                                  | Zeroed Cached data    | 1 Byte |
 682                                  +-----------------------+--------+
 683                         */
 684
 685                         flush_write_cache(fsp, WRITE_FLUSH);
 686                         wcp->offset = wcp->file_size;
 687                         wcp->data_size = pos - wcp->file_size + 1;
 688                         memset(wcp->data, '\0', wcp->data_size);
 689                         memcpy(wcp->data + wcp->data_size-1, data, 1);
 690
 691                         /*
 692                          * Update the file size if changed.
 693                          */
 694
 695                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 696                                 if (wcp_file_size_change(fsp) == -1) {
 697                                         return -1;
 698                                 }
 699                         }
 700
 701                         return n;
 702
 703                 } else {
 704
 705                         /* ASCII art..... JRA.
 706
 707    Case 1).
 708
 709                         +---------------+---------------+
 710                         | Cached data   | Cache buffer  |
 711                         +---------------+---------------+
 712
 713                                                               +-------------------+
 714                                                               | Data to write     |
 715                                                               +-------------------+
 716
 717    Case 2).
 718
 719                            +---------------+---------------+
 720                            | Cached data   | Cache buffer  |
 721                            +---------------+---------------+
 722
 723    +-------------------+
 724    | Data to write     |
 725    +-------------------+
 726
 727     Case 3).
 728
 729                            +---------------+---------------+
 730                            | Cached data   | Cache buffer  |
 731                            +---------------+---------------+
 732
 733                   +-----------------------------------------------------+
 734                   | Data to write                                       |
 735                   +-----------------------------------------------------+
 736
 737                   */
 738
 739                         /*
 740                          * Write is bigger than buffer, or there is no overlap on the
 741                          * low or high ends.
 742                          */
 743
 744                         DEBUG(9,("write_file: non cacheable write : fd = %d, pos = %.0f, len = %u, current cache pos = %.0f \
 745 len = %u\n",fsp->fh->fd, (double)pos, (unsigned int)n, (double)wcp->offset, (unsigned int)wcp->data_size ));
 746
 747                         /*
 748                          * If write would fit in the cache, and is larger than
 749                          * the data already in the cache, flush the cache and
 750                          * preferentially copy the data new data into it. Otherwise
 751                          * just write the data directly.
 752                          */
 753
 754                         if ( n <= wcp->alloc_size && n > wcp->data_size) {
 755                                 cache_flush_needed = True;
 756                         } else {
 757                                 ssize_t ret = real_write_file(NULL,fsp, data, pos, n);
 758
 759                                 /*
 760                                  * If the write overlaps the entire cache, then
 761                                  * discard the current contents of the cache.
 762                                  * Fix from Rasmus Borup Hansen rbh@math.ku.dk.
 763                                  */
 764
 765                                 if ((pos <= wcp->offset) &&
 766                                                 (pos + n >= wcp->offset + wcp->data_size) ) {
 767                                         DEBUG(9,("write_file: discarding overwritten write \
 768 cache: fd = %d, off=%.0f, size=%u\n", fsp->fh->fd, (double)wcp->offset, (unsigned int)wcp->data_size ));
 769                                         wcp->data_size = 0;
 770                                 }
 771
 772                                 DO_PROFILE_INC(writecache_direct_writes);
 773                                 if (ret == -1) {
 774                                         return ret;
 775                                 }
 776
 777                                 if (pos + ret > wcp->file_size) {
 778                                         wcp->file_size = pos + ret;
 779                                 }
 780
 781                                 return ret;
 782                         }
 783
 784                         write_path = 4;
 785
 786                 }
 787
 788                 if (cache_flush_needed) {
 789                         DEBUG(3,("WRITE_FLUSH:%d: due to noncontinuous write: fd = %d, size = %.0f, pos = %.0f, \
 790 n = %u, wcp->offset=%.0f, wcp->data_size=%u\n",
 791                                 write_path, fsp->fh->fd, (double)wcp->file_size, (double)pos, (unsigned int)n,
 792                                 (double)wcp->offset, (unsigned int)wcp->data_size ));
 793
 794                         flush_write_cache(fsp, WRITE_FLUSH);
 795                 }
 796         }
 797
 798         /*
 799          * If the write request is bigger than the cache
 800          * size, write it all out.
 801          */
 802
 803         if (n > wcp->alloc_size ) {
 804                 ssize_t ret = real_write_file(NULL,fsp, data, pos, n);
 805                 if (ret == -1) {
 806                         return -1;
 807                 }
 808
 809                 if (pos + ret > wcp->file_size) {
 810                         wcp->file_size = pos + n;
 811                 }
 812
 813                 DO_PROFILE_INC(writecache_direct_writes);
 814                 return total_written + n;
 815         }
 816
 817         /*
 818          * If there's any data left, cache it.
 819          */
 820
 821         if (n) {
 822 #ifdef WITH_PROFILE
 823                 if (wcp->data_size) {
 824                         DO_PROFILE_INC(writecache_abutted_writes);
 825                 } else {
 826                         DO_PROFILE_INC(writecache_init_writes);
 827                 }
 828 #endif
 829
 830                 if ((wcp->data_size == 0)
 831                     && (pos > wcp->file_size)
 832                     && (pos + n <= wcp->file_size + wcp->alloc_size)) {
 833                         /*
 834                          * This is a write completely beyond the
 835                          * current EOF, but within reach of the write
 836                          * cache. We expect fill-up writes pretty
 837                          * soon, so it does not make sense to start
 838                          * the write cache at the current
 839                          * offset. These fill-up writes would trigger
 840                          * separate pwrites or even unnecessary cache
 841                          * flushes because they overlap if this is a
 842                          * one-byte allocating write.
 843                          */
 844                         wcp->offset = wcp->file_size;
 845                         wcp->data_size = pos - wcp->file_size;
 846                         memset(wcp->data, 0, wcp->data_size);
 847                 }
 848
 849                 memcpy(wcp->data+wcp->data_size, data, n);
 850                 if (wcp->data_size == 0) {
 851                         wcp->offset = pos;
 852                         DO_PROFILE_INC(writecache_num_write_caches);
 853                 }
 854                 wcp->data_size += n;
 855
 856                 /*
 857                  * Update the file size if changed.
 858                  */
 859
 860                 if (wcp->offset + wcp->data_size > wcp->file_size) {
 861                         if (wcp_file_size_change(fsp) == -1) {
 862                                 return -1;
 863                         }
 864                 }
 865                 DEBUG(9,("wcp->offset = %.0f wcp->data_size = %u cache return %u\n",
 866                         (double)wcp->offset, (unsigned int)wcp->data_size, (unsigned int)n));
 867
 868                 total_written += n;
 869                 return total_written; /* .... that's a write :) */
 870         }
 871
 872         return total_written;
 873 }
 874
 875 /****************************************************************************
 876  Delete the write cache structure.
 877 ****************************************************************************/
 878
 879 void delete_write_cache(files_struct *fsp)
 880 {
 881         write_cache *wcp;
 882
 883         if(!fsp) {
 884                 return;
 885         }
 886
 887         if(!(wcp = fsp->wcp)) {
 888                 return;
 889         }
 890
 891         DO_PROFILE_DEC(writecache_allocated_write_caches);
 892         allocated_write_caches--;
 893
 894         SMB_ASSERT(wcp->data_size == 0);
 895
 896         SAFE_FREE(wcp->data);
 897         SAFE_FREE(fsp->wcp);
 898
 899         DEBUG(10,("delete_write_cache: File %s deleted write cache\n",
 900                   fsp_str_dbg(fsp)));
 901 }
 902
 903 /****************************************************************************
 904  Setup the write cache structure.
 905 ****************************************************************************/
 906
 907 static bool setup_write_cache(files_struct *fsp, SMB_OFF_T file_size)
 908 {
 909         ssize_t alloc_size = lp_write_cache_size(SNUM(fsp->conn));
 910         write_cache *wcp;
 911
 912         if (allocated_write_caches >= MAX_WRITE_CACHES) {
 913                 return False;
 914         }
 915
 916         if(alloc_size == 0 || fsp->wcp) {
 917                 return False;
 918         }
 919
 920         if((wcp = SMB_MALLOC_P(write_cache)) == NULL) {
 921                 DEBUG(0,("setup_write_cache: malloc fail.\n"));
 922                 return False;
 923         }
 924
 925         wcp->file_size = file_size;
 926         wcp->offset = 0;
 927         wcp->alloc_size = alloc_size;
 928         wcp->data_size = 0;
 929         if((wcp->data = (char *)SMB_MALLOC(wcp->alloc_size)) == NULL) {
 930                 DEBUG(0,("setup_write_cache: malloc fail for buffer size %u.\n",
 931                         (unsigned int)wcp->alloc_size ));
 932                 SAFE_FREE(wcp);
 933                 return False;
 934         }
 935
 936         memset(wcp->data, '\0', wcp->alloc_size );
 937
 938         fsp->wcp = wcp;
 939         DO_PROFILE_INC(writecache_allocated_write_caches);
 940         allocated_write_caches++;
 941
 942         DEBUG(10,("setup_write_cache: File %s allocated write cache size %lu\n",
 943                   fsp_str_dbg(fsp), (unsigned long)wcp->alloc_size));
 944
 945         return True;
 946 }
 947
 948 /****************************************************************************
 949  Cope with a size change.
 950 ****************************************************************************/
 951
 952 void set_filelen_write_cache(files_struct *fsp, SMB_OFF_T file_size)
 953 {
 954         if(fsp->wcp) {
 955                 /* The cache *must* have been flushed before we do this. */
 956                 if (fsp->wcp->data_size != 0) {
 957                         char *msg;
 958                         if (asprintf(&msg, "set_filelen_write_cache: size change "
 959                                  "on file %s with write cache size = %lu\n",
 960                                  fsp->fsp_name->base_name,
 961                                  (unsigned long)fsp->wcp->data_size) != -1) {
 962                                 smb_panic(msg);
 963                         } else {
 964                                 smb_panic("set_filelen_write_cache");
 965                         }
 966                 }
 967                 fsp->wcp->file_size = file_size;
 968         }
 969 }
 970
 971 /*******************************************************************
 972  Flush a write cache struct to disk.
 973 ********************************************************************/
 974
 975 ssize_t flush_write_cache(files_struct *fsp, enum flush_reason_enum reason)
 976 {
 977         write_cache *wcp = fsp->wcp;
 978         size_t data_size;
 979         ssize_t ret;
 980
 981         if(!wcp || !wcp->data_size) {
 982                 return 0;
 983         }
 984
 985         data_size = wcp->data_size;
 986         wcp->data_size = 0;
 987
 988         DO_PROFILE_DEC_INC(writecache_num_write_caches,writecache_flushed_writes[reason]);
 989
 990         DEBUG(9,("flushing write cache: fd = %d, off=%.0f, size=%u\n",
 991                 fsp->fh->fd, (double)wcp->offset, (unsigned int)data_size));
 992
 993 #ifdef WITH_PROFILE
 994         if(data_size == wcp->alloc_size) {
 995                 DO_PROFILE_INC(writecache_num_perfect_writes);
 996         }
 997 #endif
 998
 999         ret = real_write_file(NULL, fsp, wcp->data, wcp->offset, data_size);
1000
1001         /*
1002          * Ensure file size if kept up to date if write extends file.
1003          */
1004
1005         if ((ret != -1) && (wcp->offset + ret > wcp->file_size)) {
1006                 wcp->file_size = wcp->offset + ret;
1007         }
1008
1009         return ret;
1010 }
1011
1012 /*******************************************************************
1013 sync a file
1014 ********************************************************************/
1015
1016 NTSTATUS sync_file(connection_struct *conn, files_struct *fsp, bool write_through)
1017 {
1018         if (fsp->fh->fd == -1)
1019                 return NT_STATUS_INVALID_HANDLE;
1020
1021         if (lp_strict_sync(SNUM(conn)) &&
1022             (lp_syncalways(SNUM(conn)) || write_through)) {
1023                 int ret = flush_write_cache(fsp, SYNC_FLUSH);
1024                 if (ret == -1) {
1025                         return map_nt_error_from_unix(errno);
1026                 }
1027                 ret = SMB_VFS_FSYNC(fsp);
1028                 if (ret == -1) {
1029                         return map_nt_error_from_unix(errno);
1030                 }
1031         }
1032         return NT_STATUS_OK;
1033 }
1034
1035 /************************************************************
1036  Perform a stat whether a valid fd or not.
1037 ************************************************************/
1038
1039 int fsp_stat(files_struct *fsp)
1040 {
1041         if (fsp->fh->fd == -1) {
1042                 if (fsp->posix_open) {
1043                         return SMB_VFS_LSTAT(fsp->conn, fsp->fsp_name);
1044                 } else {
1045                         return SMB_VFS_STAT(fsp->conn, fsp->fsp_name);
1046                 }
1047         } else {
1048                 return SMB_VFS_FSTAT(fsp, &fsp->fsp_name->st);
1049         }
1050 }