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 #include "smbprofile.h"
  27
  28 struct write_cache {
  29         off_t file_size;
  30         off_t offset;
  31         size_t alloc_size;
  32         size_t data_size;
  33         char *data;
  34 };
  35
  36 static bool setup_write_cache(files_struct *, off_t);
  37
  38 /****************************************************************************
  39  Read from write cache if we can.
  40 ****************************************************************************/
  41
  42 static bool read_from_write_cache(files_struct *fsp,char *data,off_t pos,size_t n)
  43 {
  44         struct write_cache *wcp = fsp->wcp;
  45
  46         if(!wcp) {
  47                 return False;
  48         }
  49
  50         if( n > wcp->data_size || pos < wcp->offset || pos + n > wcp->offset + wcp->data_size) {
  51                 return False;
  52         }
  53
  54         memcpy(data, wcp->data + (pos - wcp->offset), n);
  55
  56         DO_PROFILE_INC(writecache_cached_reads);
  57
  58         return True;
  59 }
  60
  61 /****************************************************************************
  62  Read from a file.
  63 ****************************************************************************/
  64
  65 ssize_t read_file(files_struct *fsp,char *data,off_t pos,size_t n)
  66 {
  67         ssize_t ret = 0;
  68
  69         /* you can't read from print files */
  70         if (fsp->print_file) {
  71                 errno = EBADF;
  72                 return -1;
  73         }
  74
  75         /*
  76          * Serve from write cache if we can.
  77          */
  78
  79         if(read_from_write_cache(fsp, data, pos, n)) {
  80                 fsp->fh->pos = pos + n;
  81                 fsp->fh->position_information = fsp->fh->pos;
  82                 return n;
  83         }
  84
  85         flush_write_cache(fsp, SAMBA_READ_FLUSH);
  86
  87         fsp->fh->pos = pos;
  88
  89         if (n > 0) {
  90                 ret = SMB_VFS_PREAD(fsp,data,n,pos);
  91
  92                 if (ret == -1) {
  93                         return -1;
  94                 }
  95         }
  96
  97         DEBUG(10,("read_file (%s): pos = %.0f, size = %lu, returned %lu\n",
  98                   fsp_str_dbg(fsp), (double)pos, (unsigned long)n, (long)ret));
  99
 100         fsp->fh->pos += ret;
 101         fsp->fh->position_information = fsp->fh->pos;
 102
 103         return(ret);
 104 }
 105
 106 /****************************************************************************
 107  *Really* write to a file.
 108 ****************************************************************************/
 109
 110 static ssize_t real_write_file(struct smb_request *req,
 111                                 files_struct *fsp,
 112                                 const char *data,
 113                                 off_t pos,
 114                                 size_t n)
 115 {
 116         ssize_t ret;
 117
 118         fsp->fh->pos = pos;
 119         if (pos && lp_strict_allocate(SNUM(fsp->conn) &&
 120                         !fsp->is_sparse)) {
 121                 if (vfs_fill_sparse(fsp, pos) == -1) {
 122                         return -1;
 123                 }
 124         }
 125         ret = vfs_pwrite_data(req, fsp, data, n, pos);
 126
 127         DEBUG(10,("real_write_file (%s): pos = %.0f, size = %lu, returned %ld\n",
 128                   fsp_str_dbg(fsp), (double)pos, (unsigned long)n, (long)ret));
 129
 130         if (ret != -1) {
 131                 fsp->fh->pos += ret;
 132
 133 /* Yes - this is correct - writes don't update this. JRA. */
 134 /* Found by Samba4 tests. */
 135 #if 0
 136                 fsp->position_information = fsp->pos;
 137 #endif
 138         }
 139
 140         return ret;
 141 }
 142
 143 /****************************************************************************
 144  File size cache change.
 145  Updates size on disk but doesn't flush the cache.
 146 ****************************************************************************/
 147
 148 static int wcp_file_size_change(files_struct *fsp)
 149 {
 150         int ret;
 151         struct write_cache *wcp = fsp->wcp;
 152
 153         wcp->file_size = wcp->offset + wcp->data_size;
 154         ret = SMB_VFS_FTRUNCATE(fsp, wcp->file_size);
 155         if (ret == -1) {
 156                 DEBUG(0,("wcp_file_size_change (%s): ftruncate of size %.0f "
 157                          "error %s\n", fsp_str_dbg(fsp),
 158                          (double)wcp->file_size, strerror(errno)));
 159         }
 160         return ret;
 161 }
 162
 163 void fsp_flush_write_time_update(struct files_struct *fsp)
 164 {
 165         /*
 166          * Note this won't expect any impersonation!
 167          * So don't call any SMB_VFS operations here!
 168          */
 169
 170         DEBUG(5, ("Update write time on %s\n", fsp_str_dbg(fsp)));
 171
 172         /* change the write time in the open file db. */
 173         (void)set_write_time(fsp->file_id, timespec_current());
 174
 175         /* And notify. */
 176         notify_fname(fsp->conn, NOTIFY_ACTION_MODIFIED,
 177                      FILE_NOTIFY_CHANGE_LAST_WRITE, fsp->fsp_name->base_name);
 178
 179         /* Remove the timed event handler. */
 180         TALLOC_FREE(fsp->update_write_time_event);
 181 }
 182
 183 static void update_write_time_handler(struct tevent_context *ctx,
 184                                       struct tevent_timer *te,
 185                                       struct timeval now,
 186                                       void *private_data)
 187 {
 188         files_struct *fsp = (files_struct *)private_data;
 189         fsp_flush_write_time_update(fsp);
 190 }
 191
 192 /*********************************************************
 193  Schedule a write time update for WRITE_TIME_UPDATE_USEC_DELAY
 194  in the future.
 195 *********************************************************/
 196
 197 void trigger_write_time_update(struct files_struct *fsp)
 198 {
 199         int delay;
 200
 201         if (fsp->posix_flags & FSP_POSIX_FLAGS_OPEN) {
 202                 /* Don't use delayed writes on POSIX files. */
 203                 return;
 204         }
 205
 206         if (fsp->write_time_forced) {
 207                 /* No point - "sticky" write times
 208                  * in effect.
 209                  */
 210                 return;
 211         }
 212
 213         /* We need to remember someone did a write
 214          * and update to current time on close. */
 215
 216         fsp->update_write_time_on_close = true;
 217
 218         if (fsp->update_write_time_triggered) {
 219                 /*
 220                  * We only update the write time after 2 seconds
 221                  * on the first normal write. After that
 222                  * no other writes affect this until close.
 223                  */
 224                 return;
 225         }
 226         fsp->update_write_time_triggered = true;
 227
 228         delay = lp_parm_int(SNUM(fsp->conn),
 229                             "smbd", "writetimeupdatedelay",
 230                             WRITE_TIME_UPDATE_USEC_DELAY);
 231
 232         DEBUG(5, ("Update write time %d usec later on %s\n",
 233                   delay, fsp_str_dbg(fsp)));
 234
 235         /*
 236          * trigger the update 2 seconds later
 237          *
 238          * Note that update_write_time_handler()
 239          * => fsp_flush_write_time_update()
 240          * won't do any SMB_VFS calls and don't
 241          * need impersonation. So we use the
 242          * raw event context for this.
 243          */
 244         fsp->update_write_time_event =
 245                 tevent_add_timer(fsp->conn->sconn->raw_ev_ctx, NULL,
 246                                  timeval_current_ofs_usec(delay),
 247                                  update_write_time_handler, fsp);
 248 }
 249
 250 void trigger_write_time_update_immediate(struct files_struct *fsp)
 251 {
 252         struct smb_file_time ft;
 253
 254         if (fsp->posix_flags & FSP_POSIX_FLAGS_OPEN) {
 255                 /* Don't use delayed writes on POSIX files. */
 256                 return;
 257         }
 258
 259         if (fsp->write_time_forced) {
 260                 /*
 261                  * No point - "sticky" write times
 262                  * in effect.
 263                  */
 264                 return;
 265         }
 266
 267         TALLOC_FREE(fsp->update_write_time_event);
 268         DEBUG(5, ("Update write time immediate on %s\n",
 269                   fsp_str_dbg(fsp)));
 270
 271         /* After an immediate update, reset the trigger. */
 272         fsp->update_write_time_triggered = true;
 273         fsp->update_write_time_on_close = false;
 274
 275         ZERO_STRUCT(ft);
 276         ft.mtime = timespec_current();
 277
 278         /* Update the time in the open file db. */
 279         (void)set_write_time(fsp->file_id, ft.mtime);
 280
 281         /* Now set on disk - takes care of notify. */
 282         (void)smb_set_file_time(fsp->conn, fsp, fsp->fsp_name, &ft, false);
 283 }
 284
 285 void mark_file_modified(files_struct *fsp)
 286 {
 287         int dosmode;
 288
 289         if (fsp->modified) {
 290                 return;
 291         }
 292
 293         fsp->modified = true;
 294
 295         if (SMB_VFS_FSTAT(fsp, &fsp->fsp_name->st) != 0) {
 296                 return;
 297         }
 298         trigger_write_time_update(fsp);
 299
 300         if (fsp->posix_flags & FSP_POSIX_FLAGS_OPEN) {
 301                 return;
 302         }
 303         if (!(lp_store_dos_attributes(SNUM(fsp->conn)) ||
 304               MAP_ARCHIVE(fsp->conn))) {
 305                 return;
 306         }
 307
 308         dosmode = dos_mode(fsp->conn, fsp->fsp_name);
 309         if (IS_DOS_ARCHIVE(dosmode)) {
 310                 return;
 311         }
 312         file_set_dosmode(fsp->conn, fsp->fsp_name,
 313                          dosmode | FILE_ATTRIBUTE_ARCHIVE, NULL, false);
 314 }
 315
 316 /****************************************************************************
 317  Write to a file.
 318 ****************************************************************************/
 319
 320 ssize_t write_file(struct smb_request *req,
 321                         files_struct *fsp,
 322                         const char *data,
 323                         off_t pos,
 324                         size_t n)
 325 {
 326         struct write_cache *wcp = fsp->wcp;
 327         ssize_t total_written = 0;
 328         int write_path = -1;
 329
 330         if (fsp->print_file) {
 331                 uint32_t t;
 332                 int ret;
 333
 334                 ret = print_spool_write(fsp, data, n, pos, &t);
 335                 if (ret) {
 336                         errno = ret;
 337                         return -1;
 338                 }
 339                 return t;
 340         }
 341
 342         if (!fsp->can_write) {
 343                 errno = EPERM;
 344                 return -1;
 345         }
 346
 347         /*
 348          * If this is the first write and we have an exclusive oplock
 349          * then setup the write cache.
 350          */
 351
 352         if (!fsp->modified &&
 353             EXCLUSIVE_OPLOCK_TYPE(fsp->oplock_type) &&
 354             (wcp == NULL)) {
 355                 /*
 356                  * Note: no write cache with leases!
 357                  * as the handles would have to share the write cache
 358                  * that's possible but an improvement for another day...
 359                  */
 360                 setup_write_cache(fsp, fsp->fsp_name->st.st_ex_size);
 361                 wcp = fsp->wcp;
 362         }
 363
 364         mark_file_modified(fsp);
 365
 366         DO_PROFILE_INC(writecache_total_writes);
 367         if (!fsp->oplock_type) {
 368                 DO_PROFILE_INC(writecache_non_oplock_writes);
 369         }
 370
 371         /*
 372          * If this file is level II oplocked then we need
 373          * to grab the shared memory lock and inform all
 374          * other files with a level II lock that they need
 375          * to flush their read caches. We keep the lock over
 376          * the shared memory area whilst doing this.
 377          */
 378
 379         /* This should actually be improved to span the write. */
 380         contend_level2_oplocks_begin(fsp, LEVEL2_CONTEND_WRITE);
 381         contend_level2_oplocks_end(fsp, LEVEL2_CONTEND_WRITE);
 382
 383         if (wcp && req->unread_bytes) {
 384                 /* If we're using receivefile don't
 385                  * deal with a write cache.
 386                  */
 387                 flush_write_cache(fsp, SAMBA_WRITE_FLUSH);
 388                 delete_write_cache(fsp);
 389                 wcp = NULL;
 390         }
 391
 392         if(!wcp) {
 393                 DO_PROFILE_INC(writecache_direct_writes);
 394                 total_written = real_write_file(req, fsp, data, pos, n);
 395                 return total_written;
 396         }
 397
 398         DEBUG(9,("write_file (%s)(fd=%d pos=%.0f size=%u) wcp->offset=%.0f "
 399                  "wcp->data_size=%u\n", fsp_str_dbg(fsp), fsp->fh->fd,
 400                  (double)pos, (unsigned int)n, (double)wcp->offset,
 401                  (unsigned int)wcp->data_size));
 402
 403         fsp->fh->pos = pos + n;
 404
 405         if ((n == 1) && (data[0] == '\0') && (pos > wcp->file_size)) {
 406                 int ret;
 407
 408                 /*
 409                  * This is a 1-byte write of a 0 beyond the EOF and
 410                  * thus implicitly also beyond the current active
 411                  * write cache, the typical file-extending (and
 412                  * allocating, but we're using the write cache here)
 413                  * write done by Windows. We just have to ftruncate
 414                  * the file and rely on posix semantics to return
 415                  * zeros for non-written file data that is within the
 416                  * file length.
 417                  *
 418                  * We can not use wcp_file_size_change here because we
 419                  * might have an existing write cache, and
 420                  * wcp_file_size_change assumes a change to just the
 421                  * end of the current write cache.
 422                  */
 423
 424                 wcp->file_size = pos + 1;
 425                 ret = SMB_VFS_FTRUNCATE(fsp, wcp->file_size);
 426                 if (ret == -1) {
 427                         DEBUG(0, ("wcp_file_size_change (%s): ftruncate of "
 428                                   "size %.0f error %s\n", fsp_str_dbg(fsp),
 429                                   (double)wcp->file_size, strerror(errno)));
 430                         return -1;
 431                 }
 432                 return 1;
 433         }
 434
 435
 436         /*
 437          * If we have active cache and it isn't contiguous then we flush.
 438          * NOTE: There is a small problem with running out of disk ....
 439          */
 440
 441         if (wcp->data_size) {
 442                 bool cache_flush_needed = False;
 443
 444                 if ((pos >= wcp->offset) &&
 445                     (pos <= wcp->offset + wcp->data_size)) {
 446
 447                         /* ASCII art.... JRA.
 448
 449       +--------------+-----
 450       | Cached data  | Rest of allocated cache buffer....
 451       +--------------+-----
 452
 453             +-------------------+
 454             | Data to write     |
 455             +-------------------+
 456
 457                         */
 458
 459                         /*
 460                          * Start of write overlaps or abutts the existing data.
 461                          */
 462
 463                         size_t data_used;
 464
 465                         data_used = MIN((wcp->alloc_size - (pos - wcp->offset)),
 466                                         n);
 467
 468                         memcpy(wcp->data + (pos - wcp->offset), data,
 469                                data_used);
 470
 471                         /*
 472                          * Update the current buffer size with the new data.
 473                          */
 474
 475                         if(pos + data_used > wcp->offset + wcp->data_size) {
 476                                 wcp->data_size = pos + data_used - wcp->offset;
 477                         }
 478
 479                         /*
 480                          * Update the file size if changed.
 481                          */
 482
 483                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 484                                 if (wcp_file_size_change(fsp) == -1) {
 485                                         return -1;
 486                                 }
 487                         }
 488
 489                         /*
 490                          * If we used all the data then
 491                          * return here.
 492                          */
 493
 494                         if(n == data_used) {
 495                                 return n;
 496                         } else {
 497                                 cache_flush_needed = True;
 498                         }
 499                         /*
 500                          * Move the start of data forward by the amount used,
 501                          * cut down the amount left by the same amount.
 502                          */
 503
 504                         data += data_used;
 505                         pos += data_used;
 506                         n -= data_used;
 507
 508                         DO_PROFILE_INC(writecache_abutted_writes);
 509                         total_written = data_used;
 510
 511                         write_path = 1;
 512
 513                 } else if ((pos < wcp->offset) &&
 514                            (pos + n > wcp->offset) &&
 515                            (pos + n <= wcp->offset + wcp->alloc_size)) {
 516
 517                         /* ASCII art.... JRA.
 518
 519                         +---------------+
 520                         | Cache buffer  |
 521                         +---------------+
 522
 523             +-------------------+
 524             | Data to write     |
 525             +-------------------+
 526
 527                         */
 528
 529                         /*
 530                          * End of write overlaps the existing data.
 531                          */
 532
 533                         size_t data_used = pos + n - wcp->offset;
 534
 535                         memcpy(wcp->data, data + n - data_used, data_used);
 536
 537                         /*
 538                          * Update the current buffer size with the new data.
 539                          */
 540
 541                         if(pos + n > wcp->offset + wcp->data_size) {
 542                                 wcp->data_size = pos + n - wcp->offset;
 543                         }
 544
 545                         /*
 546                          * Update the file size if changed.
 547                          */
 548
 549                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 550                                 if (wcp_file_size_change(fsp) == -1) {
 551                                         return -1;
 552                                 }
 553                         }
 554
 555                         /*
 556                          * We don't need to move the start of data, but we
 557                          * cut down the amount left by the amount used.
 558                          */
 559
 560                         n -= data_used;
 561
 562                         /*
 563                          * We cannot have used all the data here.
 564                          */
 565
 566                         cache_flush_needed = True;
 567
 568                         DO_PROFILE_INC(writecache_abutted_writes);
 569                         total_written = data_used;
 570
 571                         write_path = 2;
 572
 573                 } else if ((pos >= wcp->file_size) &&
 574                            (wcp->offset + wcp->data_size == wcp->file_size) &&
 575                            (pos > wcp->offset + wcp->data_size) &&
 576                            (pos < wcp->offset + wcp->alloc_size) ) {
 577
 578                         /* ASCII art.... JRA.
 579
 580                        End of file ---->|
 581
 582                         +---------------+---------------+
 583                         | Cached data   | Cache buffer  |
 584                         +---------------+---------------+
 585
 586                                               +-------------------+
 587                                               | Data to write     |
 588                                               +-------------------+
 589
 590                         */
 591
 592                         /*
 593                          * Non-contiguous write part of which fits within
 594                          * the cache buffer and is extending the file
 595                          * and the cache contents reflect the current
 596                          * data up to the current end of the file.
 597                          */
 598
 599                         size_t data_used;
 600
 601                         if(pos + n <= wcp->offset + wcp->alloc_size) {
 602                                 data_used = n;
 603                         } else {
 604                                 data_used = wcp->offset+wcp->alloc_size-pos;
 605                         }
 606
 607                         /*
 608                          * Fill in the non-continuous area with zeros.
 609                          */
 610
 611                         memset(wcp->data + wcp->data_size, '\0',
 612                                 pos - (wcp->offset + wcp->data_size) );
 613
 614                         memcpy(wcp->data + (pos - wcp->offset), data,
 615                                data_used);
 616
 617                         /*
 618                          * Update the current buffer size with the new data.
 619                          */
 620
 621                         if(pos + data_used > wcp->offset + wcp->data_size) {
 622                                 wcp->data_size = pos + data_used - wcp->offset;
 623                         }
 624
 625                         /*
 626                          * Update the file size if changed.
 627                          */
 628
 629                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 630                                 if (wcp_file_size_change(fsp) == -1) {
 631                                         return -1;
 632                                 }
 633                         }
 634
 635                         /*
 636                          * If we used all the data then
 637                          * return here.
 638                          */
 639
 640                         if(n == data_used) {
 641                                 return n;
 642                         } else {
 643                                 cache_flush_needed = True;
 644                         }
 645
 646                         /*
 647                          * Move the start of data forward by the amount used,
 648                          * cut down the amount left by the same amount.
 649                          */
 650
 651                         data += data_used;
 652                         pos += data_used;
 653                         n -= data_used;
 654
 655                         DO_PROFILE_INC(writecache_abutted_writes);
 656                         total_written = data_used;
 657
 658                         write_path = 3;
 659
 660                 } else if ( (pos >= wcp->file_size) &&
 661                             (n == 1) &&
 662                             (wcp->file_size == wcp->offset + wcp->data_size) &&
 663                             (pos < wcp->file_size + wcp->alloc_size)) {
 664
 665                         /*
 666
 667                 End of file ---->|
 668
 669                  +---------------+---------------+
 670                  | Cached data   | Cache buffer  |
 671                  +---------------+---------------+
 672
 673                                  |<------- allocated size ---------------->|
 674
 675                                                          +--------+
 676                                                          | 1 Byte |
 677                                                          +--------+
 678
 679                         MS-Office seems to do this a lot to determine if
 680                         there's enough space on the filesystem to write a new
 681                         file.
 682
 683                         Change to :
 684
 685                 End of file ---->|
 686                                  +-----------------------+--------+
 687                                  | Zeroed Cached data    | 1 Byte |
 688                                  +-----------------------+--------+
 689                         */
 690
 691                         flush_write_cache(fsp, SAMBA_WRITE_FLUSH);
 692                         wcp->offset = wcp->file_size;
 693                         wcp->data_size = pos - wcp->file_size + 1;
 694                         memset(wcp->data, '\0', wcp->data_size);
 695                         memcpy(wcp->data + wcp->data_size-1, data, 1);
 696
 697                         /*
 698                          * Update the file size if changed.
 699                          */
 700
 701                         if (wcp->offset + wcp->data_size > wcp->file_size) {
 702                                 if (wcp_file_size_change(fsp) == -1) {
 703                                         return -1;
 704                                 }
 705                         }
 706
 707                         return n;
 708
 709                 } else {
 710
 711                         /* ASCII art..... JRA.
 712
 713    Case 1).
 714
 715                         +---------------+---------------+
 716                         | Cached data   | Cache buffer  |
 717                         +---------------+---------------+
 718
 719                                                               +---------------+
 720                                                               | Data to write |
 721                                                               +---------------+
 722
 723    Case 2).
 724
 725                            +---------------+---------------+
 726                            | Cached data   | Cache buffer  |
 727                            +---------------+---------------+
 728
 729    +-------------------+
 730    | Data to write     |
 731    +-------------------+
 732
 733     Case 3).
 734
 735                            +---------------+---------------+
 736                            | Cached data   | Cache buffer  |
 737                            +---------------+---------------+
 738
 739                   +-----------------------------------------------------+
 740                   | Data to write                                       |
 741                   +-----------------------------------------------------+
 742
 743                   */
 744
 745                         /*
 746                          * Write is bigger than buffer, or there is no
 747                          * overlap on the low or high ends.
 748                          */
 749
 750                         DEBUG(9,("write_file: non cacheable write : fd = %d, "
 751                                  "pos = %.0f, len = %u, "
 752                                  "current cache pos = %.0f len = %u\n",
 753                                  fsp->fh->fd, (double)pos, (unsigned int)n,
 754                                  (double)wcp->offset,
 755                                  (unsigned int)wcp->data_size ));
 756
 757                         /*
 758                          * If write would fit in the cache, and is
 759                          * larger than the data already in the cache,
 760                          * flush the cache and preferentially copy the
 761                          * data new data into it. Otherwise just write
 762                          * the data directly.
 763                          */
 764
 765                         if ( n <= wcp->alloc_size && n > wcp->data_size) {
 766                                 cache_flush_needed = True;
 767                         } else {
 768                                 ssize_t ret = real_write_file(NULL, fsp, data,
 769                                                               pos, n);
 770
 771                                 /*
 772                                  * If the write overlaps the entire
 773                                  * cache, then discard the current
 774                                  * contents of the cache.  Fix from
 775                                  * Rasmus Borup Hansen rbh@math.ku.dk.
 776                                  */
 777
 778                                 if ((pos <= wcp->offset) &&
 779                                     (pos + n >= wcp->offset+wcp->data_size)) {
 780                                         DEBUG(9,("write_file: discarding "
 781                                                  "overwritten write cache: "
 782                                                  "fd = %d, off=%.0f, "
 783                                                  "size=%u\n", fsp->fh->fd,
 784                                                  (double)wcp->offset,
 785                                                  (unsigned)wcp->data_size));
 786                                         wcp->data_size = 0;
 787                                 }
 788
 789                                 DO_PROFILE_INC(writecache_direct_writes);
 790                                 if (ret == -1) {
 791                                         return ret;
 792                                 }
 793
 794                                 if (pos + ret > wcp->file_size) {
 795                                         wcp->file_size = pos + ret;
 796                                 }
 797
 798                                 return ret;
 799                         }
 800
 801                         write_path = 4;
 802
 803                 }
 804
 805                 if (cache_flush_needed) {
 806                         DEBUG(3, ("SAMBA_WRITE_FLUSH:%d: due to noncontinuous "
 807                                   "write: fd = %d, size = %.0f, pos = %.0f, "
 808                                   "n = %u, wcp->offset=%.0f, "
 809                                   "wcp->data_size=%u\n",
 810                                   write_path, fsp->fh->fd,
 811                                   (double)wcp->file_size, (double)pos,
 812                                   (unsigned int)n, (double)wcp->offset,
 813                                   (unsigned int)wcp->data_size ));
 814
 815                         flush_write_cache(fsp, SAMBA_WRITE_FLUSH);
 816                 }
 817         }
 818
 819         /*
 820          * If the write request is bigger than the cache
 821          * size, write it all out.
 822          */
 823
 824         if (n > wcp->alloc_size ) {
 825                 ssize_t ret = real_write_file(NULL,fsp, data, pos, n);
 826                 if (ret == -1) {
 827                         return -1;
 828                 }
 829
 830                 if (pos + ret > wcp->file_size) {
 831                         wcp->file_size = pos + n;
 832                 }
 833
 834                 DO_PROFILE_INC(writecache_direct_writes);
 835                 return total_written + n;
 836         }
 837
 838         /*
 839          * If there's any data left, cache it.
 840          */
 841
 842         if (n) {
 843                 DO_PROFILE_INC(writecache_cached_writes);
 844                 if (wcp->data_size) {
 845                         DO_PROFILE_INC(writecache_abutted_writes);
 846                 } else {
 847                         DO_PROFILE_INC(writecache_init_writes);
 848                 }
 849
 850                 if ((wcp->data_size == 0)
 851                     && (pos > wcp->file_size)
 852                     && (pos + n <= wcp->file_size + wcp->alloc_size)) {
 853                         /*
 854                          * This is a write completely beyond the
 855                          * current EOF, but within reach of the write
 856                          * cache. We expect fill-up writes pretty
 857                          * soon, so it does not make sense to start
 858                          * the write cache at the current
 859                          * offset. These fill-up writes would trigger
 860                          * separate pwrites or even unnecessary cache
 861                          * flushes because they overlap if this is a
 862                          * one-byte allocating write.
 863                          */
 864                         wcp->offset = wcp->file_size;
 865                         wcp->data_size = pos - wcp->file_size;
 866                         memset(wcp->data, 0, wcp->data_size);
 867                 }
 868
 869                 memcpy(wcp->data+wcp->data_size, data, n);
 870                 if (wcp->data_size == 0) {
 871                         wcp->offset = pos;
 872                 }
 873                 wcp->data_size += n;
 874
 875                 /*
 876                  * Update the file size if changed.
 877                  */
 878
 879                 if (wcp->offset + wcp->data_size > wcp->file_size) {
 880                         if (wcp_file_size_change(fsp) == -1) {
 881                                 return -1;
 882                         }
 883                 }
 884                 DEBUG(9, ("wcp->offset = %.0f wcp->data_size = %u cache "
 885                           "return %u\n",
 886                           (double)wcp->offset, (unsigned int)wcp->data_size,
 887                           (unsigned int)n));
 888
 889                 total_written += n;
 890                 return total_written; /* .... that's a write :) */
 891         }
 892
 893         return total_written;
 894 }
 895
 896 /****************************************************************************
 897  Delete the write cache structure.
 898 ****************************************************************************/
 899
 900 void delete_write_cache(files_struct *fsp)
 901 {
 902         struct write_cache *wcp;
 903
 904         if(!fsp) {
 905                 return;
 906         }
 907
 908         if(!(wcp = fsp->wcp)) {
 909                 return;
 910         }
 911
 912         DO_PROFILE_INC(writecache_deallocations);
 913         allocated_write_caches--;
 914
 915         SMB_ASSERT(wcp->data_size == 0);
 916
 917         SAFE_FREE(wcp->data);
 918         SAFE_FREE(fsp->wcp);
 919
 920         DEBUG(10,("delete_write_cache: File %s deleted write cache\n",
 921                   fsp_str_dbg(fsp)));
 922 }
 923
 924 /****************************************************************************
 925  Setup the write cache structure.
 926 ****************************************************************************/
 927
 928 static bool setup_write_cache(files_struct *fsp, off_t file_size)
 929 {
 930         ssize_t alloc_size = lp_write_cache_size(SNUM(fsp->conn));
 931         struct write_cache *wcp;
 932
 933         if (allocated_write_caches >= MAX_WRITE_CACHES) {
 934                 return False;
 935         }
 936
 937         if(alloc_size == 0 || fsp->wcp) {
 938                 return False;
 939         }
 940
 941         if((wcp = SMB_MALLOC_P(struct write_cache)) == NULL) {
 942                 DEBUG(0,("setup_write_cache: malloc fail.\n"));
 943                 return False;
 944         }
 945
 946         wcp->file_size = file_size;
 947         wcp->offset = 0;
 948         wcp->alloc_size = alloc_size;
 949         wcp->data_size = 0;
 950         if((wcp->data = (char *)SMB_MALLOC(wcp->alloc_size)) == NULL) {
 951                 DEBUG(0,("setup_write_cache: malloc fail for buffer size %u.\n",
 952                         (unsigned int)wcp->alloc_size ));
 953                 SAFE_FREE(wcp);
 954                 return False;
 955         }
 956
 957         memset(wcp->data, '\0', wcp->alloc_size );
 958
 959         fsp->wcp = wcp;
 960         DO_PROFILE_INC(writecache_allocations);
 961         allocated_write_caches++;
 962
 963         DEBUG(10,("setup_write_cache: File %s allocated write cache size %lu\n",
 964                   fsp_str_dbg(fsp), (unsigned long)wcp->alloc_size));
 965
 966         return True;
 967 }
 968
 969 /****************************************************************************
 970  Cope with a size change.
 971 ****************************************************************************/
 972
 973 void set_filelen_write_cache(files_struct *fsp, off_t file_size)
 974 {
 975         if(fsp->wcp) {
 976                 /* The cache *must* have been flushed before we do this. */
 977                 if (fsp->wcp->data_size != 0) {
 978                         char *msg;
 979                         if (asprintf(&msg, "set_filelen_write_cache: size change "
 980                                  "on file %s with write cache size = %lu\n",
 981                                  fsp->fsp_name->base_name,
 982                                  (unsigned long)fsp->wcp->data_size) != -1) {
 983                                 smb_panic(msg);
 984                         } else {
 985                                 smb_panic("set_filelen_write_cache");
 986                         }
 987                 }
 988                 fsp->wcp->file_size = file_size;
 989         }
 990 }
 991
 992 /*******************************************************************
 993  Flush a write cache struct to disk.
 994 ********************************************************************/
 995
 996 ssize_t flush_write_cache(files_struct *fsp, enum flush_reason_enum reason)
 997 {
 998         struct write_cache *wcp = fsp->wcp;
 999         size_t data_size;
1000         ssize_t ret;
1001
1002         if(!wcp || !wcp->data_size) {
1003                 return 0;
1004         }
1005
1006         data_size = wcp->data_size;
1007         wcp->data_size = 0;
1008
1009         switch (reason) {
1010         case SAMBA_SEEK_FLUSH:
1011                 DO_PROFILE_INC(writecache_flush_reason_seek);
1012                 break;
1013         case SAMBA_READ_FLUSH:
1014                 DO_PROFILE_INC(writecache_flush_reason_read);
1015                 break;
1016         case SAMBA_WRITE_FLUSH:
1017                 DO_PROFILE_INC(writecache_flush_reason_write);;
1018                 break;
1019         case SAMBA_READRAW_FLUSH:
1020                 DO_PROFILE_INC(writecache_flush_reason_readraw);
1021                 break;
1022         case SAMBA_OPLOCK_RELEASE_FLUSH:
1023                 DO_PROFILE_INC(writecache_flush_reason_oplock);
1024                 break;
1025         case SAMBA_CLOSE_FLUSH:
1026                 DO_PROFILE_INC(writecache_flush_reason_close);
1027                 break;
1028         case SAMBA_SYNC_FLUSH:
1029                 DO_PROFILE_INC(writecache_flush_reason_sync);
1030                 break;
1031         case SAMBA_SIZECHANGE_FLUSH:
1032                 DO_PROFILE_INC(writecache_flush_reason_sizechange);
1033                 break;
1034         default:
1035                 break;
1036         }
1037
1038         DEBUG(9,("flushing write cache: fd = %d, off=%.0f, size=%u\n",
1039                 fsp->fh->fd, (double)wcp->offset, (unsigned int)data_size));
1040
1041         if(data_size == wcp->alloc_size) {
1042                 DO_PROFILE_INC(writecache_perfect_writes);
1043         }
1044
1045         ret = real_write_file(NULL, fsp, wcp->data, wcp->offset, data_size);
1046
1047         /*
1048          * Ensure file size if kept up to date if write extends file.
1049          */
1050
1051         if ((ret != -1) && (wcp->offset + ret > wcp->file_size)) {
1052                 wcp->file_size = wcp->offset + ret;
1053         }
1054
1055         return ret;
1056 }
1057
1058 /*******************************************************************
1059 sync a file
1060 ********************************************************************/
1061
1062 NTSTATUS sync_file(connection_struct *conn, files_struct *fsp, bool write_through)
1063 {
1064         if (fsp->fh->fd == -1)
1065                 return NT_STATUS_INVALID_HANDLE;
1066
1067         if (lp_strict_sync(SNUM(conn)) &&
1068             (lp_sync_always(SNUM(conn)) || write_through)) {
1069                 int ret = flush_write_cache(fsp, SAMBA_SYNC_FLUSH);
1070                 if (ret == -1) {
1071                         return map_nt_error_from_unix(errno);
1072                 }
1073                 ret = smb_vfs_fsync_sync(fsp);
1074                 if (ret == -1) {
1075                         return map_nt_error_from_unix(errno);
1076                 }
1077         }
1078         return NT_STATUS_OK;
1079 }
1080
1081 /************************************************************
1082  Perform a stat whether a valid fd or not.
1083 ************************************************************/
1084
1085 int fsp_stat(files_struct *fsp)
1086 {
1087         if (fsp->fh->fd == -1) {
1088                 if (fsp->posix_flags & FSP_POSIX_FLAGS_OPEN) {
1089                         return SMB_VFS_LSTAT(fsp->conn, fsp->fsp_name);
1090                 } else {
1091                         return SMB_VFS_STAT(fsp->conn, fsp->fsp_name);
1092                 }
1093         } else {
1094                 return SMB_VFS_FSTAT(fsp, &fsp->fsp_name->st);
1095         }
1096 }