| blob | 9c69bcacad2286c1e7a926bf2a5000ae9d84da64 |
1 /*
2 * Copyright 2000 by Hans Reiser, licensing governed by reiserfs/README
3 */
5 #include <linux/time.h>
6 #include <linux/fs.h>
7 #include <linux/reiserfs_fs.h>
8 #include <linux/reiserfs_acl.h>
9 #include <linux/reiserfs_xattr.h>
10 #include <linux/smp_lock.h>
11 #include <linux/pagemap.h>
12 #include <linux/highmem.h>
13 #include <asm/uaccess.h>
14 #include <asm/unaligned.h>
15 #include <linux/buffer_head.h>
16 #include <linux/mpage.h>
17 #include <linux/writeback.h>
18 #include <linux/quotaops.h>
26 {
27 /* We need blocks for transaction + (user+group) quota update (possibly delete) */
38 /* The = 0 happens when we abort creating a new inode for some reason like lack of space.. */
39 if (!(inode->i_state & I_NEW) && INODE_PKEY(inode)->k_objectid != 0) { /* also handles bad_inode case */
48 /* Do quota update inside a transaction for journaled quotas. We must do that
49 * after delete_object so that quota updates go into the same transaction as
50 * stat data deletion */
57 /* check return value from reiserfs_delete_object after
58 * ending the transaction
59 */
63 /* all items of file are deleted, so we can remove "save" link */
65 * about an error here */
67 /* no object items are in the tree */
68 ;
69 }
70 out:
74 }
78 {
86 }
88 /* take base of inode_key (it comes from inode always) (dirid, objectid) and version from an inode, set
89 offset and type of key */
92 {
96 length);
97 }
99 //
100 // when key is 0, do not set version and short key
101 //
106 {
111 }
116 /* set_ih_free_space (ih, 0); */
117 // for directory items it is entry count, for directs and stat
118 // datas - 0xffff, for indirects - 0
120 }
122 //
123 // FIXME: we might cache recently accessed indirect item
125 // Ugh. Not too eager for that....
126 // I cut the code until such time as I see a convincing argument (benchmark).
127 // I don't want a bloated inode struct..., and I don't like code complexity....
129 /* cutting the code is fine, since it really isn't in use yet and is easy
130 ** to add back in. But, Vladimir has a really good idea here. Think
131 ** about what happens for reading a file. For each page,
132 ** The VFS layer calls reiserfs_readpage, who searches the tree to find
133 ** an indirect item. This indirect item has X number of pointers, where
134 ** X is a big number if we've done the block allocation right. But,
135 ** we only use one or two of these pointers during each call to readpage,
136 ** needlessly researching again later on.
137 **
138 ** The size of the cache could be dynamic based on the size of the file.
139 **
140 ** I'd also like to see us cache the location the stat data item, since
141 ** we are needlessly researching for that frequently.
142 **
143 ** --chris
144 */
146 /* If this page has a file tail in it, and
147 ** it was read in by get_block_create_0, the page data is valid,
148 ** but tail is still sitting in a direct item, and we can't write to
149 ** it. So, look through this page, and check all the mapped buffers
150 ** to make sure they have valid block numbers. Any that don't need
151 ** to be unmapped, so that block_prepare_write will correctly call
152 ** reiserfs_get_block to convert the tail into an unformatted node
153 */
155 {
165 }
168 }
169 }
171 /* reiserfs_get_block does not need to allocate a block only if it has been
172 done already or non-hole position has been found in the indirect item */
176 {
183 }
186 {
188 }
192 {
194 }
196 //
197 // files which were created in the earlier version can not be longer,
198 // than 2 gb
199 //
201 {
207 }
211 {
219 /* we cannot restart while nested */
222 }
230 }
232 }
234 // it is called by get_block when create == 0. Returns block number
235 // for 'block'-th logical block of file. When it hits direct item it
236 // returns 0 (being called from bmap) or read direct item into piece
237 // of page (bh_result)
239 // Please improve the english/clarity in the comment above, as it is
240 // hard to understand.
244 {
258 // prepare the key to look for the 'block'-th block of file
263 research:
271 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
272 // That there is some MMAPED data associated with it that is yet to be written to disk.
276 }
278 }
279 //
285 /* FIXME: here we could cache indirect item or part of it in
286 the inode to avoid search_by_key in case of subsequent
287 access to file */
295 }
297 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
298 // That there is some MMAPED data associated with it that is yet to be written to disk.
302 }
308 }
309 // requested data are in direct item(s)
311 // we are called by bmap. FIXME: we can not map block of file
312 // when it is stored in direct item(s)
317 }
319 /* if we've got a direct item, and the buffer or page was uptodate,
320 ** we don't want to pull data off disk again. skip to the
321 ** end, where we map the buffer and return
322 */
326 /*
327 ** grab_tail_page can trigger calls to reiserfs_get_block on up to date
328 ** pages without any buffers. If the page is up to date, we don't want
329 ** read old data off disk. Set the up to date bit on the buffer instead
330 ** and jump to the end
331 */
335 }
336 // read file tail into part of page
341 /* we only want to kmap if we are reading the tail into the page.
342 ** this is not the common case, so we don't kmap until we are
343 ** sure we need to. But, this means the item might move if
344 ** kmap schedules
345 */
351 }
352 }
358 }
359 /* make sure we don't read more bytes than actually exist in
360 ** the file. This can happen in odd cases where i_size isn't
361 ** correct, and when direct item padding results in a few
362 ** extra bytes at the end of the direct item
363 */
367 chars =
373 }
382 // we done, if read direct item is not the last item of
383 // node FIXME: we could try to check right delimiting key
384 // to see whether direct item continues in the right
385 // neighbor or rely on i_size
388 // update key to look for the next piece
392 // i/o error most likely
401 finished:
407 /* this buffer has valid data, but isn't valid for io. mapping it to
408 * block #0 tells the rest of reiserfs it just has a tail in it
409 */
413 }
415 // this is called to create file map. So, _get_block_create_0 will not
416 // read direct item
419 {
424 /* do not read the direct item */
428 }
430 /* special version of get_block that is only used by grab_tail_page right
431 ** now. It is sent to block_prepare_write, and when you try to get a
432 ** block past the end of the file (or a block from a hole) it returns
433 ** -ENOENT instead of a valid buffer. block_prepare_write expects to
434 ** be able to do i/o on the buffers returned, unless an error value
435 ** is also returned.
436 **
437 ** So, this allows block_prepare_write to be used for reading a single block
438 ** in a page. Where it does not produce a valid page for holes, or past the
439 ** end of the file. This turns out to be exactly what we need for reading
440 ** tails for conversion.
441 **
442 ** The point of the wrapper is forcing a certain value for create, even
443 ** though the VFS layer is calling this function with create==1. If you
444 ** don't want to send create == GET_BLOCK_NO_HOLE to reiserfs_get_block,
445 ** don't use this function.
446 */
450 {
452 }
454 /* This is special helper for reiserfs_get_block in case we are executing
455 direct_IO request. */
457 sector_t iblock,
460 {
465 /* We set the b_size before reiserfs_get_block call since it is
466 referenced in convert_tail_for_hole() that may be called from
467 reiserfs_get_block() */
475 /* don't allow direct io onto tail pages */
477 /* make sure future calls to the direct io funcs for this offset
478 ** in the file fail by unmapping the buffer
479 */
482 }
483 /* Possible unpacked tail. Flush the data before pages have
484 disappeared */
493 }
494 out:
496 }
498 /*
499 ** helper function for when reiserfs_get_block is called for a hole
500 ** but the file tail is still in a direct item
501 ** bh_result is the buffer head for the hole
502 ** tail_offset is the offset of the start of the tail in the file
503 **
504 ** This calls prepare_write, which will start a new transaction
505 ** you should not be in a transaction, or have any paths held when you
506 ** call this.
507 */
510 loff_t tail_offset)
511 {
522 /* always try to read until the end of the block */
527 /* hole_page can be zero in case of direct_io, we are sure
528 that we cannot get here if we write with O_DIRECT into
529 tail page */
535 }
538 }
540 /* we don't have to make sure the conversion did not happen while
541 ** we were locking the page because anyone that could convert
542 ** must first take i_mutex.
543 **
544 ** We must fix the tail page for writing because it might have buffers
545 ** that are mapped, but have a block number of 0. This indicates tail
546 ** data that has been read directly into the page, and block_prepare_write
547 ** won't trigger a get_block in this case.
548 */
554 /* tail conversion might change the data in the page */
559 unlock:
563 }
564 out:
566 }
573 {
576 #ifdef REISERFS_PREALLOCATE
580 }
581 #endif
583 block);
584 }
588 {
600 /* space reserved in transaction batch:
601 . 3 balancings in direct->indirect conversion
602 . 1 block involved into reiserfs_update_sd()
603 XXX in practically impossible worst case direct2indirect()
604 can incur (much) more than 3 balancings.
605 quota update for user, group */
611 loff_t new_offset =
614 /* bad.... */
621 }
623 /* if !create, we aren't changing the FS, so we don't need to
624 ** log anything, so we don't need to start a transaction
625 */
628 /* find number of block-th logical block of the file */
633 }
634 /*
635 * if we're already in a transaction, make sure to close
636 * any new transactions we start in this func
637 */
642 /* If file is of such a size, that it might have a tail and tails are enabled
643 ** we should mark it as possibly needing tail packing on close
644 */
651 /* set the key of the first byte in the 'block'-th block of file */
654 start_trans:
659 }
661 }
662 research:
668 }
680 /* we have to allocate block for the unformatted node */
684 }
686 repeat =
691 /* restart the transaction to give the journal a chance to free
692 ** some blocks. releases the path, so we have to go back to
693 ** research if we succeed on the second try
694 */
699 repeat =
705 }
708 else
711 }
716 }
717 }
720 b_blocknr_t unfm_ptr;
721 /* 'block'-th block is in the file already (there is
722 corresponding cell in some indirect item). But it may be
723 zero unformatted node pointer (hole) */
726 /* use allocated block to plug the hole */
731 bh);
733 }
742 }
751 /* the item was found, so new blocks were not added to the file
752 ** there is no need to make sure the inode is updated with this
753 ** transaction
754 */
756 }
761 }
763 /* desired position is not found or is in the direct item. We have
764 to append file with holes up to 'block'-th block converting
765 direct items to indirect one if necessary */
772 /* indirect item has to be inserted */
778 /* we are going to add 'block'-th block to the file. Use
779 allocated block for that */
785 }
790 retval =
797 }
798 //mark_tail_converted (inode);
800 /* direct item has to be converted */
801 loff_t tail_offset;
803 tail_offset =
807 /* direct item we just found fits into block we have
808 to map. Convert it into unformatted node: use
809 bh_result for the conversion */
815 /* we have to padd file tail stored in direct item(s)
816 up to block size and convert it to unformatted
817 node. FIXME: this should also get into page cache */
820 /*
821 * ugly, but we can only end the transaction if
822 * we aren't nested
823 */
826 retval =
827 reiserfs_end_persistent_transaction
832 }
834 retval =
836 tail_offset);
842 retval);
844 /* the bitmap, the super, and the stat data == 3 */
849 inode,
850 allocated_block_nr,
852 }
854 }
856 }
857 retval =
859 tail_offset);
865 }
866 /* it is important the set_buffer_uptodate is done after
867 ** the direct2indirect. The buffer might contain valid
868 ** data newer than the data on disk (read by readpage, changed,
869 ** and then sent here by writepage). direct2indirect needs
870 ** to know if unbh was already up to date, so it can decide
871 ** if the data in unbh needs to be replaced with data from
872 ** the disk
873 */
876 /* unbh->b_page == NULL in case of DIRECT_IO request, this means
877 buffer will disappear shortly, so it should not be added to
878 */
880 /* we've converted the tail, so we must
881 ** flush unbh before the transaction commits
882 */
885 /* mark it dirty now to prevent commit_write from adding
886 ** this buffer to the inode's dirty buffer list
887 */
888 /*
889 * AKPM: changed __mark_buffer_dirty to mark_buffer_dirty().
890 * It's still atomic, but it sets the page dirty too,
891 * which makes it eligible for writeback at any time by the
892 * VM (which was also the case with __mark_buffer_dirty())
893 */
895 }
897 /* append indirect item with holes if needed, when appending
898 pointer to 'block'-th block use block, which is already
899 allocated */
902 // one block which is a fastpath
904 __u64 max_to_insert =
906 UNFM_P_SIZE;
907 __u64 blocks_needed;
911 /* indirect item has to be appended, set up key of that position */
917 //pos_in_item * inode->i_sb->s_blocksize,
922 blocks_needed =
926 s_blocksize_bits);
931 un = kmalloc(min(blocks_needed, max_to_insert) * UNFM_P_SIZE, GFP_ATOMIC); // We need to avoid scheduling.
939 max_to_insert));
940 }
942 /* we are going to add target block to the file. Use allocated
943 block for that */
951 /* paste hole to the indirect item */
952 /* If kmalloc failed, max_to_insert becomes zero and it means we
953 only have space for one block */
954 blocks_needed =
956 }
957 retval =
960 UNFM_P_SIZE *
961 blocks_needed);
970 }
972 /* We need to mark new file size in case this function will be
973 interrupted/aborted later on. And we may do this only for
974 holes. */
977 }
978 }
983 /* this loop could log more blocks than we had originally asked
984 ** for. So, we have to allow the transaction to end if it is
985 ** too big or too full. Update the inode so things are
986 ** consistent if we crash before the function returns
987 **
988 ** release the path so that anybody waiting on the path before
989 ** ending their transaction will be able to continue.
990 */
995 }
996 /* inserting indirect pointers for a hole can take a
997 ** long time. reschedule if needed
998 */
1005 }
1008 "vs-825: reiserfs_get_block: "
1016 }
1025 failure:
1033 }
1038 }
1040 static int
1043 {
1045 }
1047 /* Compute real number of used bytes by file
1048 * Following three functions can go away when we'll have enough space in stat item
1049 */
1051 {
1058 /* End of file is also in full block with indirect reference, so round
1059 ** up to the next block.
1060 **
1061 ** there is just no way to know if the tail is actually packed
1062 ** on the file, so we have to assume it isn't. When we pack the
1063 ** tail, we add 4 bytes to pretend there really is an unformatted
1064 ** node pointer
1065 */
1066 bytes =
1069 sd_size;
1071 }
1075 {
1079 }
1082 }
1084 /* Compute number of blocks used by file in ReiserFS counting */
1086 {
1090 /* keeps fsck and non-quota versions of reiserfs happy */
1093 }
1095 /* files from before the quota patch might i_blocks such that
1096 ** bytes < real_space. Deal with that here to prevent it from
1097 ** going negative.
1098 */
1102 }
1104 //
1105 // BAD: new directories have stat data of new type and all other items
1106 // of old type. Version stored in the inode says about body items, so
1107 // in update_stat_data we can not rely on inode, but have to check
1108 // item version directly
1109 //
1111 // called by read_locked_inode
1113 {
1116 __u32 rdev;
1117 //int version = ITEM_VERSION_1;
1158 // there was a bug in <=3.5.23 when i_blocks could take negative
1159 // values. Starting from 3.5.17 this value could even be stored in
1160 // stat data. For such files we set i_blocks based on file
1161 // size. Just 2 notes: this can be wrong for sparce files. On-disk value will be
1162 // only updated if file's inode will ever change
1164 }
1169 /* an early bug in the quota code can give us an odd number for the
1170 ** block count. This is incorrect, fix it here.
1171 */
1174 }
1177 SD_V1_SIZE));
1178 /* nopack is initially zero for v1 objects. For v2 objects,
1179 nopack is initialised from sd_attrs */
1182 // new stat data found, but object may have old items
1183 // (directories and symlinks)
1202 else
1207 else
1213 SD_V2_SIZE));
1214 /* read persistent inode attributes from sd and initalise
1215 generic inode flags from them */
1218 }
1235 }
1236 }
1238 // update new stat data with inode fields
1240 {
1242 __u16 flags;
1255 else
1260 }
1262 // used to copy inode's fields to old stat data
1264 {
1278 else
1281 // Sigh. i_first_direct_byte is back
1284 }
1286 /* NOTE, you must prepare the buffer head before sending it here,
1287 ** and then log it after the call
1288 */
1290 loff_t size)
1291 {
1304 // path points to old stat data
1308 }
1311 }
1315 {
1329 /* look for the object's stat data */
1333 "vs-13050: reiserfs_update_sd: "
1337 }
1342 /*reiserfs_warning (inode->i_sb, "vs-13050: reiserfs_update_sd: i_nlink == 0, stat data not found"); */
1344 }
1346 "vs-13060: reiserfs_update_sd: "
1349 pos);
1352 }
1354 /* sigh, prepare_for_journal might schedule. When it schedules the
1355 ** FS might change. We have to detect that, and loop back to the
1356 ** search if the stat data item has moved
1357 */
1367 }
1369 }
1374 }
1376 /* reiserfs_read_locked_inode is called to read the inode off disk, and it
1377 ** does a make_bad_inode when things go wrong. But, we need to make sure
1378 ** and clear the key in the private portion of the inode, otherwise a
1379 ** corresponding iput might try to delete whatever object the inode last
1380 ** represented.
1381 */
1383 {
1386 }
1388 //
1389 // initially this function was derived from minix or ext2's analog and
1390 // evolved as the prototype did
1391 //
1394 {
1399 }
1401 /* looks for stat data in the tree, and fills up the fields of in-core
1402 inode stat data fields */
1405 {
1413 /* set version 1, version 2 could be used too, because stat data
1414 key is the same in both versions */
1421 /* look for the object's stat data */
1425 "vs-13070: reiserfs_read_locked_inode: "
1430 }
1432 /* a stale NFS handle can trigger this without it being an error */
1437 }
1441 /* It is possible that knfsd is trying to access inode of a file
1442 that is being removed from the disk by some other thread. As we
1443 update sd on unlink all that is required is to check for nlink
1444 here. This bug was first found by Sizif when debugging
1445 SquidNG/Butterfly, forgotten, and found again after Philippe
1446 Gramoulle <philippe.gramoulle@mmania.com> reproduced it.
1448 More logical fix would require changes in fs/inode.c:iput() to
1449 remove inode from hash-table _after_ fs cleaned disk stuff up and
1450 in iget() to return NULL if I_FREEING inode is found in
1451 hash-table. */
1452 /* Currently there is one place where it's ok to meet inode with
1453 nlink==0: processing of open-unlinked and half-truncated files
1454 during mount (fs/reiserfs/super.c:finish_unfinished()). */
1458 "vs-13075: reiserfs_read_locked_inode: "
1459 "dead inode read from disk %K. "
1463 }
1467 }
1469 /**
1470 * reiserfs_find_actor() - "find actor" reiserfs supplies to iget5_locked().
1471 *
1472 * @inode: inode from hash table to check
1473 * @opaque: "cookie" passed to iget5_locked(). This is &reiserfs_iget_args.
1474 *
1475 * This function is called by iget5_locked() to distinguish reiserfs inodes
1476 * having the same inode numbers. Such inodes can only exist due to some
1477 * error condition. One of them should be bad. Inodes with identical
1478 * inode numbers (objectids) are distinguished by parent directory ids.
1479 *
1480 */
1482 {
1486 /* args is already in CPU order */
1489 }
1492 {
1507 }
1510 /* either due to i/o error or a stale NFS handle */
1513 }
1515 }
1518 {
1532 }
1542 }
1544 }
1551 {
1554 /* fhtype happens to reflect the number of u32s encoded.
1555 * due to a bug in earlier code, fhtype might indicate there
1556 * are more u32s then actually fitted.
1557 * so if fhtype seems to be more than len, reduce fhtype.
1558 * Valid types are:
1559 * 2 - objectid + dir_id - legacy support
1560 * 3 - objectid + dir_id + generation
1561 * 4 - objectid + dir_id + objectid and dirid of parent - legacy
1562 * 5 - objectid + dir_id + generation + objectid and dirid of parent
1563 * 6 - as above plus generation of directory
1564 * 6 does not fit in NFSv2 handles
1565 */
1572 }
1578 else
1586 else
1588 }
1592 }
1596 {
1607 /* no room for directory info? return what we've stored so far */
1619 }
1622 }
1624 /* looks for stat data, then copies fields to it, marks the buffer
1625 containing stat data as dirty */
1626 /* reiserfs inodes are never really dirty, since the dirty inode call
1627 ** always logs them. This call allows the VFS inode marking routines
1628 ** to properly mark inodes for datasync and such, but only actually
1629 ** does something when called for a synchronous update.
1630 */
1632 {
1638 /* memory pressure can sometimes initiate write_inode calls with sync == 1,
1639 ** these cases are just when the system needs ram, not when the
1640 ** inode needs to reach disk for safety, and they can safely be
1641 ** ignored because the altered inode has already been logged.
1642 */
1648 }
1650 }
1652 }
1654 /* stat data of new object is inserted already, this inserts the item
1655 containing "." and ".." entries */
1660 {
1673 /* compose item head for new item. Directories consist of items of
1674 old type (ITEM_VERSION_1). Do not set key (second arg is 0), it
1675 is done by reiserfs_new_inode */
1692 }
1694 /* look for place in the tree for new item */
1700 }
1707 }
1709 /* insert item, that is empty directory item */
1711 }
1713 /* stat data of object has been inserted, this inserts the item
1714 containing the body of symlink */
1715 static int reiserfs_new_symlink(struct reiserfs_transaction_handle *th, struct inode *inode, /* Inode of symlink */
1719 {
1734 /* look for place in the tree for new item */
1740 }
1747 }
1749 /* insert item, that is body of symlink */
1751 }
1753 /* inserts the stat data into the tree, and then calls
1754 reiserfs_new_directory (to insert ".", ".." item if new object is
1755 directory) or reiserfs_new_symlink (to insert symlink body if new
1756 object is symlink) or nothing (if new object is regular file)
1758 NOTE! uid and gid must already be set in the inode. If we return
1759 non-zero due to an error, we have to drop the quota previously allocated
1760 for the fresh inode. This can only be done outside a transaction, so
1761 if we return non-zero, we also end the transaction. */
1764 /* 0 for regular, EMTRY_DIR_SIZE for dirs,
1765 strlen (symname) for symlinks) */
1768 {
1782 }
1786 }
1790 /* item head of new item */
1796 }
1798 /* not a perfect generation count, as object ids can be reused, but
1799 ** this is as good as reiserfs can do right now.
1800 ** note that the private part of inode isn't filled in yet, we have
1801 ** to use the directory.
1802 */
1804 else
1805 #if defined( USE_INODE_GENERATION_COUNTER )
1808 #else
1810 #endif
1812 /* fill stat data */
1815 /* uid and gid must already be set by the caller for quota init */
1817 /* symlink cannot be immutable or append only, right? */
1844 else
1848 /* key to search for correct place for new stat data */
1853 /* find proper place for inserting of stat data */
1858 }
1863 }
1867 /* i_uid or i_gid is too big to be stored in stat data v3.5 */
1870 }
1874 }
1875 // these do not go to on-disk stat data
1878 // store in in-core inode the key of stat data and version all
1879 // object items will have (directory items will have old offset
1880 // format, other new objects will consist of new items)
1884 else
1888 else
1891 /* insert the stat data into the tree */
1892 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1895 #endif
1896 retval =
1903 }
1904 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1907 #endif
1909 /* insert item with "." and ".." */
1910 retval =
1912 }
1915 /* insert body of symlink */
1918 retval =
1920 i_size);
1921 }
1927 }
1929 /* XXX CHECK THIS */
1937 }
1943 }
1951 /* it looks like you can easily compress these two goto targets into
1952 * one. Keeping it like this doesn't actually hurt anything, and they
1953 * are place holders for what the quota code actually needs.
1954 */
1955 out_bad_inode:
1956 /* Invalidate the object, nothing was inserted yet */
1959 /* Quota change must be inside a transaction for journaling */
1962 out_end_trans:
1964 /* Drop can be outside and it needs more credits so it's better to have it outside */
1969 out_inserted_sd:
1973 /* If we were inheriting an ACL, we need to release the lock so that
1974 * iput doesn't deadlock in reiserfs_delete_xattrs. The locking
1975 * code really needs to be reworked, but this will take care of it
1976 * for now. -jeffm */
1977 #ifdef CONFIG_REISERFS_FS_POSIX_ACL
1983 #endif
1986 }
1988 /*
1989 ** finds the tail page in the page cache,
1990 ** reads the last block in.
1991 **
1992 ** On success, page_result is set to a locked, pinned page, and bh_result
1993 ** is set to an up to date buffer for the last block in the file. returns 0.
1994 **
1995 ** tail conversion is not done, so bh_result might not be valid for writing
1996 ** check buffer_mapped(bh_result) and bh_result->b_blocknr != 0 before
1997 ** trying to write the block.
1998 **
1999 ** on failure, nonzero is returned, page_result and bh_result are untouched.
2000 */
2004 {
2006 /* we want the page with the last byte in the file,
2007 ** not the page that will hold the next byte for appending
2008 */
2019 /* we know that we are only called with inode->i_size > 0.
2020 ** we also know that a file tail can never be as big as a block
2021 ** If i_size % blocksize == 0, our file is currently block aligned
2022 ** and it won't need converting or zeroing after a truncate.
2023 */
2026 }
2031 }
2032 /* start within the page of the last block in the file */
2036 reiserfs_get_block_create_0);
2045 }
2051 /* note, this should never happen, prepare_write should
2052 ** be taking care of this for us. If the buffer isn't up to date,
2053 ** I've screwed up the code to find the buffer, or the code to
2054 ** call prepare_write
2055 */
2062 }
2066 out:
2069 unlock:
2073 }
2075 /*
2076 ** vfs version of truncate file. Must NOT be called with
2077 ** a transaction already started.
2078 **
2079 ** some code taken from block_truncate_page
2080 */
2082 {
2084 /* we want the offset for the first byte after the end of the file */
2097 // -ENOENT means we truncated past the end of the file,
2098 // and get_block_create_0 could not find a block to read in,
2099 // which is ok.
2103 error);
2106 }
2107 }
2109 /* so, if page != NULL, we have a buffer head for the offset at
2110 ** the end of the file. if the bh is mapped, and bh->b_blocknr != 0,
2111 ** then we have an unformatted node. Otherwise, we have a direct item,
2112 ** and no zeroing is required on disk. We zero after the truncate,
2113 ** because the truncate might pack the item anyway
2114 ** (it will unmap bh if it packs).
2115 */
2116 /* it is enough to reserve space in transaction for 2 balancings:
2117 one for "save" link adding and another for the first
2118 cut_from_item. 1 is for update_sd */
2125 /* we are doing real truncate: if the system crashes before the last
2126 transaction of truncating gets committed - on reboot the file
2127 either appears truncated properly or not truncated at all */
2130 error =
2135 /* check reiserfs_do_truncate after ending the transaction */
2139 }
2145 }
2149 /* if we are not on a block boundary */
2160 }
2161 }
2164 }
2168 out:
2172 }
2175 }
2180 {
2198 /* catch places below that try to log something without starting a trans */
2203 }
2206 start_over:
2210 research:
2215 }
2222 /* we've found an unformatted node */
2227 bytes_copied);
2228 }
2230 /* crap, we are writing to a hole */
2233 }
2246 /* vs-3050 is gone, no need to drop the path */
2255 bh);
2257 }
2258 }
2266 }
2269 copy_size);
2275 /* are there still bytes left? */
2280 copy_size);
2282 }
2289 }
2292 out:
2299 }
2302 /* this is where we fill in holes in the file. */
2305 GET_BLOCK_CREATE | GET_BLOCK_NO_IMUX
2310 /* get_block failed to find a mapped unformatted node. */
2313 }
2314 }
2315 }
2319 /* we've copied data from the page into the direct item, so the
2320 * buffer in the page is now clean, mark it to reflect that.
2321 */
2325 }
2327 }
2329 /*
2330 * mason@suse.com: updated in 2.5.54 to follow the same general io
2331 * start/recovery path as __block_write_full_page, along with special
2332 * code to handle reiserfs tails.
2333 */
2336 {
2341 sector_t last_block;
2351 /* no logging allowed when nonblocking or from PF_MEMALLOC */
2356 }
2358 /* The page dirty bit is cleared before writepage is called, which
2359 * means we have to tell create_empty_buffers to make dirty buffers
2360 * The page really should be up to date at this point, so tossing
2361 * in the BH_Uptodate is just a sanity check.
2362 */
2366 }
2369 /* last page in the file, zero out any contents past the
2370 ** last byte in the file
2371 */
2377 /* no file contents in this page */
2381 }
2386 }
2390 /* first map all the buffers, logging any direct items we find */
2393 /*
2394 * This can happen when the block size is less than
2395 * the page size. The corresponding bytes in the page
2396 * were zero filled above
2397 */
2404 /* not mapped yet, or it points to a direct item, search
2405 * the btree for the mapping info, and log any direct
2406 * items found
2407 */
2410 }
2411 }
2413 block++;
2416 /*
2417 * we start the transaction after map_block_for_writepage,
2418 * because it can create holes in the file (an unbounded operation).
2419 * starting it here, we can make a reliable estimate for how many
2420 * blocks we're going to log
2421 */
2429 }
2431 }
2432 /* now go through and lock any dirty buffers on the page */
2444 }
2445 /* from this point on, we know the buffer is mapped to a
2446 * real block and not a direct item
2447 */
2454 }
2455 }
2460 }
2468 }
2473 /*
2474 * since any buffer might be the only dirty buffer on the page,
2475 * the first submit_bh can bring the page out of writeback.
2476 * be careful with the buffers.
2477 */
2482 nr++;
2483 }
2489 done:
2491 /*
2492 * if this page only had a direct item, it is very possible for
2493 * no io to be required without there being an error. Or,
2494 * someone else could have locked them and sent them down the
2495 * pipe without locking the page
2496 */
2502 }
2508 }
2511 fail:
2512 /* catches various errors, we need to make sure any valid dirty blocks
2513 * get to the media. The page is currently locked and not marked for
2514 * writeback
2515 */
2524 /*
2525 * clear any dirty bits that might have come from getting
2526 * attached to a dirty page
2527 */
2529 }
2541 nr++;
2542 }
2547 }
2550 {
2552 }
2555 {
2559 }
2563 {
2573 journal_info;
2578 }
2583 /* this gets a little ugly. If reiserfs_get_block returned an
2584 * error and left a transacstion running, we've got to close it,
2585 * and we've got to free handle if it was a persistent transaction.
2586 *
2587 * But, if we had nested into an existing transaction, we need
2588 * to just drop the ref count on the handle.
2589 *
2590 * If old_ref == 0, the transaction is from reiserfs_get_block,
2591 * and it was a persistent trans. Otherwise, it was nested above.
2592 */
2603 }
2604 }
2605 }
2608 }
2611 {
2613 }
2617 {
2627 }
2630 /* generic_commit_write does this for us, but does not update the
2631 ** transaction tracking stuff when the size changes. So, we have
2632 ** to do the i_size updates here.
2633 */
2637 /* If the file have grown beyond the border where it
2638 can have a tail, unmark it as needing a tail
2639 packing */
2650 }
2653 /*
2654 * this will just nest into our transaction. It's important
2655 * to use mark_inode_dirty so the inode gets pushed around on the
2656 * dirty lists, and so that O_SYNC works as expected
2657 */
2665 }
2674 }
2676 out:
2679 journal_error:
2686 }
2689 }
2692 {
2696 else
2700 else
2704 else
2708 else
2712 else
2714 }
2715 }
2718 {
2722 else
2726 else
2730 else
2734 else
2736 }
2737 }
2739 /* decide if this buffer needs to stay around for data logging or ordered
2740 ** write purposes
2741 */
2743 {
2751 }
2752 /* the page is locked, and the only places that log a data buffer
2753 * also lock the page.
2754 */
2756 /*
2757 * very conservative, leave the buffer pinned if
2758 * anyone might need it.
2759 */
2762 }
2767 /* why is this safe?
2768 * reiserfs_setattr updates i_size in the on disk
2769 * stat data before allowing vmtruncate to be called.
2770 *
2771 * If buffer was put onto the ordered list for this
2772 * transaction, we know for sure either this transaction
2773 * or an older one already has updated i_size on disk,
2774 * and this ordered data won't be referenced in the file
2775 * if we crash.
2776 *
2777 * if the buffer was put onto the ordered list for an older
2778 * transaction, we need to leave it around
2779 */
2783 }
2784 free_jh:
2787 }
2791 }
2793 /* clm -- taken from fs/buffer.c:block_invalidate_page */
2795 {
2815 /*
2816 * is this block fully invalidated?
2817 */
2821 else
2823 }
2828 /*
2829 * We release buffers only if the entire page is being invalidated.
2830 * The get_block cached value has been unconditionally invalidated,
2831 * so real IO is not possible anymore.
2832 */
2835 /* maybe should BUG_ON(!ret); - neilb */
2836 }
2837 out:
2839 }
2842 {
2847 }
2849 }
2851 /*
2852 * Returns 1 if the page's buffers were dropped. The page is locked.
2853 *
2854 * Takes j_dirty_buffers_lock to protect the b_assoc_buffers list_heads
2855 * in the buffers at page_buffers(page).
2856 *
2857 * even in -o notail mode, we can't be sure an old mount without -o notail
2858 * didn't create files with tails.
2859 */
2861 {
2879 }
2880 }
2887 }
2889 /* We thank Mingming Cao for helping us understand in great detail what
2890 to do in this section of the code. */
2894 {
2901 }
2904 {
2910 /* version 2 items will be caught by the s_maxbytes check
2911 ** done for us in vmtruncate
2912 */
2917 }
2918 /* fill in hole pointers in the expanding truncate case. */
2924 /* we're changing at most 2 bitmaps, inode + super */
2929 }
2932 }
2935 /*
2936 * file size is changed, ctime and mtime are
2937 * to be updated
2938 */
2940 }
2941 }
2946 /* stat data of format v3.5 has 16 bit uid and gid */
2949 }
2965 /* (user+group)*(old+new) structure - we count quota info and , inode write (sb, inode) */
2966 error =
2968 jbegin_count);
2971 error =
2975 jbegin_count);
2977 }
2978 /* Update corresponding info in inode so that everything is in
2979 * one transaction */
2985 error =
2987 }
2988 }
2991 }
2996 }
2998 out:
3001 }
3015 };