2 Unix SMB/Netbios implementation.
4 read/write to a files_struct
5 Copyright (C) Andrew Tridgell 1992-1998
6 Copyright (C) Jeremy Allison 2000-2002. - write cache.
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.
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.
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/>.
24 #include "smbd/smbd.h"
25 #include "smbd/globals.h"
26 #include "smbprofile.h"
36 static bool setup_write_cache(files_struct *, off_t);
38 /****************************************************************************
39 Read from write cache if we can.
40 ****************************************************************************/
42 static bool read_from_write_cache(files_struct *fsp,char *data,off_t pos,size_t n)
44 struct write_cache *wcp = fsp->wcp;
50 if( n > wcp->data_size || pos < wcp->offset || pos + n > wcp->offset + wcp->data_size) {
54 memcpy(data, wcp->data + (pos - wcp->offset), n);
56 DO_PROFILE_INC(writecache_read_hits);
61 /****************************************************************************
63 ****************************************************************************/
65 ssize_t read_file(files_struct *fsp,char *data,off_t pos,size_t n)
69 /* you can't read from print files */
70 if (fsp->print_file) {
76 * Serve from write cache if we can.
79 if(read_from_write_cache(fsp, data, pos, n)) {
80 fsp->fh->pos = pos + n;
81 fsp->fh->position_information = fsp->fh->pos;
85 flush_write_cache(fsp, READ_FLUSH);
90 ret = SMB_VFS_PREAD(fsp,data,n,pos);
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));
101 fsp->fh->position_information = fsp->fh->pos;
106 /****************************************************************************
107 *Really* write to a file.
108 ****************************************************************************/
110 static ssize_t real_write_file(struct smb_request *req,
119 ret = vfs_write_data(req, fsp, data, n);
122 if (pos && lp_strict_allocate(SNUM(fsp->conn) &&
124 if (vfs_fill_sparse(fsp, pos) == -1) {
128 ret = vfs_pwrite_data(req, fsp, data, n, pos);
131 DEBUG(10,("real_write_file (%s): pos = %.0f, size = %lu, returned %ld\n",
132 fsp_str_dbg(fsp), (double)pos, (unsigned long)n, (long)ret));
137 /* Yes - this is correct - writes don't update this. JRA. */
138 /* Found by Samba4 tests. */
140 fsp->position_information = fsp->pos;
147 /****************************************************************************
148 File size cache change.
149 Updates size on disk but doesn't flush the cache.
150 ****************************************************************************/
152 static int wcp_file_size_change(files_struct *fsp)
155 struct write_cache *wcp = fsp->wcp;
157 wcp->file_size = wcp->offset + wcp->data_size;
158 ret = SMB_VFS_FTRUNCATE(fsp, wcp->file_size);
160 DEBUG(0,("wcp_file_size_change (%s): ftruncate of size %.0f "
161 "error %s\n", fsp_str_dbg(fsp),
162 (double)wcp->file_size, strerror(errno)));
167 void update_write_time_handler(struct event_context *ctx,
168 struct timed_event *te,
172 files_struct *fsp = (files_struct *)private_data;
174 DEBUG(5, ("Update write time on %s\n", fsp_str_dbg(fsp)));
176 /* change the write time in the open file db. */
177 (void)set_write_time(fsp->file_id, timespec_current());
180 notify_fname(fsp->conn, NOTIFY_ACTION_MODIFIED,
181 FILE_NOTIFY_CHANGE_LAST_WRITE, fsp->fsp_name->base_name);
183 /* Remove the timed event handler. */
184 TALLOC_FREE(fsp->update_write_time_event);
187 /*********************************************************
188 Schedule a write time update for WRITE_TIME_UPDATE_USEC_DELAY
190 *********************************************************/
192 void trigger_write_time_update(struct files_struct *fsp)
196 if (fsp->posix_open) {
197 /* Don't use delayed writes on POSIX files. */
201 if (fsp->write_time_forced) {
202 /* No point - "sticky" write times
208 /* We need to remember someone did a write
209 * and update to current time on close. */
211 fsp->update_write_time_on_close = true;
213 if (fsp->update_write_time_triggered) {
215 * We only update the write time after 2 seconds
216 * on the first normal write. After that
217 * no other writes affect this until close.
221 fsp->update_write_time_triggered = true;
223 delay = lp_parm_int(SNUM(fsp->conn),
224 "smbd", "writetimeupdatedelay",
225 WRITE_TIME_UPDATE_USEC_DELAY);
227 DEBUG(5, ("Update write time %d usec later on %s\n",
228 delay, fsp_str_dbg(fsp)));
230 /* trigger the update 2 seconds later */
231 fsp->update_write_time_event =
232 tevent_add_timer(fsp->conn->sconn->ev_ctx, NULL,
233 timeval_current_ofs_usec(delay),
234 update_write_time_handler, fsp);
237 void trigger_write_time_update_immediate(struct files_struct *fsp)
239 struct smb_file_time ft;
241 if (fsp->posix_open) {
242 /* Don't use delayed writes on POSIX files. */
246 if (fsp->write_time_forced) {
248 * No point - "sticky" write times
254 TALLOC_FREE(fsp->update_write_time_event);
255 DEBUG(5, ("Update write time immediate on %s\n",
258 /* After an immediate update, reset the trigger. */
259 fsp->update_write_time_triggered = true;
260 fsp->update_write_time_on_close = false;
263 ft.mtime = timespec_current();
265 /* Update the time in the open file db. */
266 (void)set_write_time(fsp->file_id, ft.mtime);
268 /* Now set on disk - takes care of notify. */
269 (void)smb_set_file_time(fsp->conn, fsp, fsp->fsp_name, &ft, false);
272 void mark_file_modified(files_struct *fsp)
280 fsp->modified = true;
282 if (SMB_VFS_FSTAT(fsp, &fsp->fsp_name->st) != 0) {
285 trigger_write_time_update(fsp);
287 if (fsp->posix_open) {
290 if (!(lp_store_dos_attributes(SNUM(fsp->conn)) ||
291 MAP_ARCHIVE(fsp->conn))) {
295 dosmode = dos_mode(fsp->conn, fsp->fsp_name);
296 if (IS_DOS_ARCHIVE(dosmode)) {
299 file_set_dosmode(fsp->conn, fsp->fsp_name,
300 dosmode | FILE_ATTRIBUTE_ARCHIVE, NULL, false);
303 /****************************************************************************
305 ****************************************************************************/
307 ssize_t write_file(struct smb_request *req,
313 struct write_cache *wcp = fsp->wcp;
314 ssize_t total_written = 0;
317 if (fsp->print_file) {
321 ret = print_spool_write(fsp, data, n, pos, &t);
329 if (!fsp->can_write) {
335 * If this is the first write and we have an exclusive oplock
336 * then setup the write cache.
339 if (!fsp->modified &&
340 EXCLUSIVE_OPLOCK_TYPE(fsp->oplock_type) &&
342 setup_write_cache(fsp, fsp->fsp_name->st.st_ex_size);
346 mark_file_modified(fsp);
349 DO_PROFILE_INC(writecache_total_writes);
350 if (!fsp->oplock_type) {
351 DO_PROFILE_INC(writecache_non_oplock_writes);
356 * If this file is level II oplocked then we need
357 * to grab the shared memory lock and inform all
358 * other files with a level II lock that they need
359 * to flush their read caches. We keep the lock over
360 * the shared memory area whilst doing this.
363 /* This should actually be improved to span the write. */
364 contend_level2_oplocks_begin(fsp, LEVEL2_CONTEND_WRITE);
365 contend_level2_oplocks_end(fsp, LEVEL2_CONTEND_WRITE);
368 if (profile_p && profile_p->writecache_total_writes % 500 == 0) {
369 DEBUG(3,("WRITECACHE: initwrites=%u abutted=%u total=%u \
370 nonop=%u allocated=%u active=%u direct=%u perfect=%u readhits=%u\n",
371 profile_p->writecache_init_writes,
372 profile_p->writecache_abutted_writes,
373 profile_p->writecache_total_writes,
374 profile_p->writecache_non_oplock_writes,
375 profile_p->writecache_allocated_write_caches,
376 profile_p->writecache_num_write_caches,
377 profile_p->writecache_direct_writes,
378 profile_p->writecache_num_perfect_writes,
379 profile_p->writecache_read_hits ));
381 DEBUG(3,("WRITECACHE: Flushes SEEK=%d, READ=%d, WRITE=%d, READRAW=%d, OPLOCK=%d, CLOSE=%d, SYNC=%d\n",
382 profile_p->writecache_flushed_writes[SEEK_FLUSH],
383 profile_p->writecache_flushed_writes[READ_FLUSH],
384 profile_p->writecache_flushed_writes[WRITE_FLUSH],
385 profile_p->writecache_flushed_writes[READRAW_FLUSH],
386 profile_p->writecache_flushed_writes[OPLOCK_RELEASE_FLUSH],
387 profile_p->writecache_flushed_writes[CLOSE_FLUSH],
388 profile_p->writecache_flushed_writes[SYNC_FLUSH] ));
392 if (wcp && req->unread_bytes) {
393 /* If we're using receivefile don't
394 * deal with a write cache.
396 flush_write_cache(fsp, WRITE_FLUSH);
397 delete_write_cache(fsp);
402 DO_PROFILE_INC(writecache_direct_writes);
403 total_written = real_write_file(req, fsp, data, pos, n);
404 return total_written;
407 DEBUG(9,("write_file (%s)(fd=%d pos=%.0f size=%u) wcp->offset=%.0f "
408 "wcp->data_size=%u\n", fsp_str_dbg(fsp), fsp->fh->fd,
409 (double)pos, (unsigned int)n, (double)wcp->offset,
410 (unsigned int)wcp->data_size));
412 fsp->fh->pos = pos + n;
414 if ((n == 1) && (data[0] == '\0') && (pos > wcp->file_size)) {
418 * This is a 1-byte write of a 0 beyond the EOF and
419 * thus implicitly also beyond the current active
420 * write cache, the typical file-extending (and
421 * allocating, but we're using the write cache here)
422 * write done by Windows. We just have to ftruncate
423 * the file and rely on posix semantics to return
424 * zeros for non-written file data that is within the
427 * We can not use wcp_file_size_change here because we
428 * might have an existing write cache, and
429 * wcp_file_size_change assumes a change to just the
430 * end of the current write cache.
433 wcp->file_size = pos + 1;
434 ret = SMB_VFS_FTRUNCATE(fsp, wcp->file_size);
436 DEBUG(0,("wcp_file_size_change (%s): ftruncate of size %.0f"
437 "error %s\n", fsp_str_dbg(fsp),
438 (double)wcp->file_size, strerror(errno)));
446 * If we have active cache and it isn't contiguous then we flush.
447 * NOTE: There is a small problem with running out of disk ....
450 if (wcp->data_size) {
451 bool cache_flush_needed = False;
453 if ((pos >= wcp->offset) && (pos <= wcp->offset + wcp->data_size)) {
455 /* ASCII art.... JRA.
457 +--------------+-----
458 | Cached data | Rest of allocated cache buffer....
459 +--------------+-----
461 +-------------------+
463 +-------------------+
468 * Start of write overlaps or abutts the existing data.
471 size_t data_used = MIN((wcp->alloc_size - (pos - wcp->offset)), n);
473 memcpy(wcp->data + (pos - wcp->offset), data, data_used);
476 * Update the current buffer size with the new data.
479 if(pos + data_used > wcp->offset + wcp->data_size) {
480 wcp->data_size = pos + data_used - wcp->offset;
484 * Update the file size if changed.
487 if (wcp->offset + wcp->data_size > wcp->file_size) {
488 if (wcp_file_size_change(fsp) == -1) {
494 * If we used all the data then
501 cache_flush_needed = True;
504 * Move the start of data forward by the amount used,
505 * cut down the amount left by the same amount.
512 DO_PROFILE_INC(writecache_abutted_writes);
513 total_written = data_used;
517 } else if ((pos < wcp->offset) && (pos + n > wcp->offset) &&
518 (pos + n <= wcp->offset + wcp->alloc_size)) {
520 /* ASCII art.... JRA.
526 +-------------------+
528 +-------------------+
533 * End of write overlaps the existing data.
536 size_t data_used = pos + n - wcp->offset;
538 memcpy(wcp->data, data + n - data_used, data_used);
541 * Update the current buffer size with the new data.
544 if(pos + n > wcp->offset + wcp->data_size) {
545 wcp->data_size = pos + n - wcp->offset;
549 * Update the file size if changed.
552 if (wcp->offset + wcp->data_size > wcp->file_size) {
553 if (wcp_file_size_change(fsp) == -1) {
559 * We don't need to move the start of data, but we
560 * cut down the amount left by the amount used.
566 * We cannot have used all the data here.
569 cache_flush_needed = True;
571 DO_PROFILE_INC(writecache_abutted_writes);
572 total_written = data_used;
576 } else if ( (pos >= wcp->file_size) &&
577 (wcp->offset + wcp->data_size == wcp->file_size) &&
578 (pos > wcp->offset + wcp->data_size) &&
579 (pos < wcp->offset + wcp->alloc_size) ) {
581 /* ASCII art.... JRA.
585 +---------------+---------------+
586 | Cached data | Cache buffer |
587 +---------------+---------------+
589 +-------------------+
591 +-------------------+
596 * Non-contiguous write part of which fits within
597 * the cache buffer and is extending the file
598 * and the cache contents reflect the current
599 * data up to the current end of the file.
604 if(pos + n <= wcp->offset + wcp->alloc_size) {
607 data_used = wcp->offset + wcp->alloc_size - pos;
611 * Fill in the non-continuous area with zeros.
614 memset(wcp->data + wcp->data_size, '\0',
615 pos - (wcp->offset + wcp->data_size) );
617 memcpy(wcp->data + (pos - wcp->offset), data, data_used);
620 * Update the current buffer size with the new data.
623 if(pos + data_used > wcp->offset + wcp->data_size) {
624 wcp->data_size = pos + data_used - wcp->offset;
628 * Update the file size if changed.
631 if (wcp->offset + wcp->data_size > wcp->file_size) {
632 if (wcp_file_size_change(fsp) == -1) {
638 * If we used all the data then
645 cache_flush_needed = True;
649 * Move the start of data forward by the amount used,
650 * cut down the amount left by the same amount.
657 DO_PROFILE_INC(writecache_abutted_writes);
658 total_written = data_used;
662 } else if ( (pos >= wcp->file_size) &&
664 (wcp->file_size == wcp->offset + wcp->data_size) &&
665 (pos < wcp->file_size + wcp->alloc_size)) {
671 +---------------+---------------+
672 | Cached data | Cache buffer |
673 +---------------+---------------+
675 |<------- allocated size ---------------->|
681 MS-Office seems to do this a lot to determine if there's enough
682 space on the filesystem to write a new file.
687 +-----------------------+--------+
688 | Zeroed Cached data | 1 Byte |
689 +-----------------------+--------+
692 flush_write_cache(fsp, WRITE_FLUSH);
693 wcp->offset = wcp->file_size;
694 wcp->data_size = pos - wcp->file_size + 1;
695 memset(wcp->data, '\0', wcp->data_size);
696 memcpy(wcp->data + wcp->data_size-1, data, 1);
699 * Update the file size if changed.
702 if (wcp->offset + wcp->data_size > wcp->file_size) {
703 if (wcp_file_size_change(fsp) == -1) {
712 /* ASCII art..... JRA.
716 +---------------+---------------+
717 | Cached data | Cache buffer |
718 +---------------+---------------+
720 +-------------------+
722 +-------------------+
726 +---------------+---------------+
727 | Cached data | Cache buffer |
728 +---------------+---------------+
730 +-------------------+
732 +-------------------+
736 +---------------+---------------+
737 | Cached data | Cache buffer |
738 +---------------+---------------+
740 +-----------------------------------------------------+
742 +-----------------------------------------------------+
747 * Write is bigger than buffer, or there is no overlap on the
751 DEBUG(9,("write_file: non cacheable write : fd = %d, pos = %.0f, len = %u, current cache pos = %.0f \
752 len = %u\n",fsp->fh->fd, (double)pos, (unsigned int)n, (double)wcp->offset, (unsigned int)wcp->data_size ));
755 * If write would fit in the cache, and is larger than
756 * the data already in the cache, flush the cache and
757 * preferentially copy the data new data into it. Otherwise
758 * just write the data directly.
761 if ( n <= wcp->alloc_size && n > wcp->data_size) {
762 cache_flush_needed = True;
764 ssize_t ret = real_write_file(NULL,fsp, data, pos, n);
767 * If the write overlaps the entire cache, then
768 * discard the current contents of the cache.
769 * Fix from Rasmus Borup Hansen rbh@math.ku.dk.
772 if ((pos <= wcp->offset) &&
773 (pos + n >= wcp->offset + wcp->data_size) ) {
774 DEBUG(9,("write_file: discarding overwritten write \
775 cache: fd = %d, off=%.0f, size=%u\n", fsp->fh->fd, (double)wcp->offset, (unsigned int)wcp->data_size ));
779 DO_PROFILE_INC(writecache_direct_writes);
784 if (pos + ret > wcp->file_size) {
785 wcp->file_size = pos + ret;
795 if (cache_flush_needed) {
796 DEBUG(3,("WRITE_FLUSH:%d: due to noncontinuous write: fd = %d, size = %.0f, pos = %.0f, \
797 n = %u, wcp->offset=%.0f, wcp->data_size=%u\n",
798 write_path, fsp->fh->fd, (double)wcp->file_size, (double)pos, (unsigned int)n,
799 (double)wcp->offset, (unsigned int)wcp->data_size ));
801 flush_write_cache(fsp, WRITE_FLUSH);
806 * If the write request is bigger than the cache
807 * size, write it all out.
810 if (n > wcp->alloc_size ) {
811 ssize_t ret = real_write_file(NULL,fsp, data, pos, n);
816 if (pos + ret > wcp->file_size) {
817 wcp->file_size = pos + n;
820 DO_PROFILE_INC(writecache_direct_writes);
821 return total_written + n;
825 * If there's any data left, cache it.
830 if (wcp->data_size) {
831 DO_PROFILE_INC(writecache_abutted_writes);
833 DO_PROFILE_INC(writecache_init_writes);
837 if ((wcp->data_size == 0)
838 && (pos > wcp->file_size)
839 && (pos + n <= wcp->file_size + wcp->alloc_size)) {
841 * This is a write completely beyond the
842 * current EOF, but within reach of the write
843 * cache. We expect fill-up writes pretty
844 * soon, so it does not make sense to start
845 * the write cache at the current
846 * offset. These fill-up writes would trigger
847 * separate pwrites or even unnecessary cache
848 * flushes because they overlap if this is a
849 * one-byte allocating write.
851 wcp->offset = wcp->file_size;
852 wcp->data_size = pos - wcp->file_size;
853 memset(wcp->data, 0, wcp->data_size);
856 memcpy(wcp->data+wcp->data_size, data, n);
857 if (wcp->data_size == 0) {
859 DO_PROFILE_INC(writecache_num_write_caches);
864 * Update the file size if changed.
867 if (wcp->offset + wcp->data_size > wcp->file_size) {
868 if (wcp_file_size_change(fsp) == -1) {
872 DEBUG(9,("wcp->offset = %.0f wcp->data_size = %u cache return %u\n",
873 (double)wcp->offset, (unsigned int)wcp->data_size, (unsigned int)n));
876 return total_written; /* .... that's a write :) */
879 return total_written;
882 /****************************************************************************
883 Delete the write cache structure.
884 ****************************************************************************/
886 void delete_write_cache(files_struct *fsp)
888 struct write_cache *wcp;
894 if(!(wcp = fsp->wcp)) {
898 DO_PROFILE_DEC(writecache_allocated_write_caches);
899 allocated_write_caches--;
901 SMB_ASSERT(wcp->data_size == 0);
903 SAFE_FREE(wcp->data);
906 DEBUG(10,("delete_write_cache: File %s deleted write cache\n",
910 /****************************************************************************
911 Setup the write cache structure.
912 ****************************************************************************/
914 static bool setup_write_cache(files_struct *fsp, off_t file_size)
916 ssize_t alloc_size = lp_write_cache_size(SNUM(fsp->conn));
917 struct write_cache *wcp;
919 if (allocated_write_caches >= MAX_WRITE_CACHES) {
923 if(alloc_size == 0 || fsp->wcp) {
927 if((wcp = SMB_MALLOC_P(struct write_cache)) == NULL) {
928 DEBUG(0,("setup_write_cache: malloc fail.\n"));
932 wcp->file_size = file_size;
934 wcp->alloc_size = alloc_size;
936 if((wcp->data = (char *)SMB_MALLOC(wcp->alloc_size)) == NULL) {
937 DEBUG(0,("setup_write_cache: malloc fail for buffer size %u.\n",
938 (unsigned int)wcp->alloc_size ));
943 memset(wcp->data, '\0', wcp->alloc_size );
946 DO_PROFILE_INC(writecache_allocated_write_caches);
947 allocated_write_caches++;
949 DEBUG(10,("setup_write_cache: File %s allocated write cache size %lu\n",
950 fsp_str_dbg(fsp), (unsigned long)wcp->alloc_size));
955 /****************************************************************************
956 Cope with a size change.
957 ****************************************************************************/
959 void set_filelen_write_cache(files_struct *fsp, off_t file_size)
962 /* The cache *must* have been flushed before we do this. */
963 if (fsp->wcp->data_size != 0) {
965 if (asprintf(&msg, "set_filelen_write_cache: size change "
966 "on file %s with write cache size = %lu\n",
967 fsp->fsp_name->base_name,
968 (unsigned long)fsp->wcp->data_size) != -1) {
971 smb_panic("set_filelen_write_cache");
974 fsp->wcp->file_size = file_size;
978 /*******************************************************************
979 Flush a write cache struct to disk.
980 ********************************************************************/
982 ssize_t flush_write_cache(files_struct *fsp, enum flush_reason_enum reason)
984 struct write_cache *wcp = fsp->wcp;
988 if(!wcp || !wcp->data_size) {
992 data_size = wcp->data_size;
995 DO_PROFILE_DEC_INC(writecache_num_write_caches,writecache_flushed_writes[reason]);
997 DEBUG(9,("flushing write cache: fd = %d, off=%.0f, size=%u\n",
998 fsp->fh->fd, (double)wcp->offset, (unsigned int)data_size));
1001 if(data_size == wcp->alloc_size) {
1002 DO_PROFILE_INC(writecache_num_perfect_writes);
1006 ret = real_write_file(NULL, fsp, wcp->data, wcp->offset, data_size);
1009 * Ensure file size if kept up to date if write extends file.
1012 if ((ret != -1) && (wcp->offset + ret > wcp->file_size)) {
1013 wcp->file_size = wcp->offset + ret;
1019 /*******************************************************************
1021 ********************************************************************/
1023 NTSTATUS sync_file(connection_struct *conn, files_struct *fsp, bool write_through)
1025 if (fsp->fh->fd == -1)
1026 return NT_STATUS_INVALID_HANDLE;
1028 if (lp_strict_sync(SNUM(conn)) &&
1029 (lp_syncalways(SNUM(conn)) || write_through)) {
1030 int ret = flush_write_cache(fsp, SYNC_FLUSH);
1032 return map_nt_error_from_unix(errno);
1034 ret = SMB_VFS_FSYNC(fsp);
1036 return map_nt_error_from_unix(errno);
1039 return NT_STATUS_OK;
1042 /************************************************************
1043 Perform a stat whether a valid fd or not.
1044 ************************************************************/
1046 int fsp_stat(files_struct *fsp)
1048 if (fsp->fh->fd == -1) {
1049 if (fsp->posix_open) {
1050 return SMB_VFS_LSTAT(fsp->conn, fsp->fsp_name);
1052 return SMB_VFS_STAT(fsp->conn, fsp->fsp_name);
1055 return SMB_VFS_FSTAT(fsp, &fsp->fsp_name->st);