| blob | d60f6238c66a48184fa7c35ec749ba8e6f110456 |
1 /*
2 * Copyright 2000 by Hans Reiser, licensing governed by reiserfs/README
3 */
5 #include <linux/config.h>
6 #include <linux/time.h>
7 #include <linux/fs.h>
8 #include <linux/reiserfs_fs.h>
9 #include <linux/reiserfs_acl.h>
10 #include <linux/reiserfs_xattr.h>
11 #include <linux/smp_lock.h>
12 #include <linux/pagemap.h>
13 #include <linux/highmem.h>
14 #include <asm/uaccess.h>
15 #include <asm/unaligned.h>
16 #include <linux/buffer_head.h>
17 #include <linux/mpage.h>
18 #include <linux/writeback.h>
19 #include <linux/quotaops.h>
29 {
30 /* We need blocks for transaction + (user+group) quota update (possibly delete) */
41 /* The = 0 happens when we abort creating a new inode for some reason like lack of space.. */
42 if (!(inode->i_state & I_NEW) && INODE_PKEY(inode)->k_objectid != 0) { /* also handles bad_inode case */
50 }
55 /* Do quota update inside a transaction for journaled quotas. We must do that
56 * after delete_object so that quota updates go into the same transaction as
57 * stat data deletion */
64 }
68 /* check return value from reiserfs_delete_object after
69 * ending the transaction
70 */
74 /* all items of file are deleted, so we can remove "save" link */
76 * about an error here */
78 /* no object items are in the tree */
79 ;
80 }
81 out:
85 }
89 {
97 }
99 /* take base of inode_key (it comes from inode always) (dirid, objectid) and version from an inode, set
100 offset and type of key */
103 {
107 length);
108 }
110 //
111 // when key is 0, do not set version and short key
112 //
117 {
122 }
127 /* set_ih_free_space (ih, 0); */
128 // for directory items it is entry count, for directs and stat
129 // datas - 0xffff, for indirects - 0
131 }
133 //
134 // FIXME: we might cache recently accessed indirect item
136 // Ugh. Not too eager for that....
137 // I cut the code until such time as I see a convincing argument (benchmark).
138 // I don't want a bloated inode struct..., and I don't like code complexity....
140 /* cutting the code is fine, since it really isn't in use yet and is easy
141 ** to add back in. But, Vladimir has a really good idea here. Think
142 ** about what happens for reading a file. For each page,
143 ** The VFS layer calls reiserfs_readpage, who searches the tree to find
144 ** an indirect item. This indirect item has X number of pointers, where
145 ** X is a big number if we've done the block allocation right. But,
146 ** we only use one or two of these pointers during each call to readpage,
147 ** needlessly researching again later on.
148 **
149 ** The size of the cache could be dynamic based on the size of the file.
150 **
151 ** I'd also like to see us cache the location the stat data item, since
152 ** we are needlessly researching for that frequently.
153 **
154 ** --chris
155 */
157 /* If this page has a file tail in it, and
158 ** it was read in by get_block_create_0, the page data is valid,
159 ** but tail is still sitting in a direct item, and we can't write to
160 ** it. So, look through this page, and check all the mapped buffers
161 ** to make sure they have valid block numbers. Any that don't need
162 ** to be unmapped, so that block_prepare_write will correctly call
163 ** reiserfs_get_block to convert the tail into an unformatted node
164 */
166 {
176 }
179 }
180 }
182 /* reiserfs_get_block does not need to allocate a block only if it has been
183 done already or non-hole position has been found in the indirect item */
187 {
194 }
197 {
199 }
203 {
205 }
207 //
208 // files which were created in the earlier version can not be longer,
209 // than 2 gb
210 //
212 {
218 }
222 {
230 /* we cannot restart while nested */
233 }
241 }
243 }
245 // it is called by get_block when create == 0. Returns block number
246 // for 'block'-th logical block of file. When it hits direct item it
247 // returns 0 (being called from bmap) or read direct item into piece
248 // of page (bh_result)
250 // Please improve the english/clarity in the comment above, as it is
251 // hard to understand.
255 {
269 // prepare the key to look for the 'block'-th block of file
274 research:
282 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
283 // That there is some MMAPED data associated with it that is yet to be written to disk.
287 }
289 }
290 //
296 /* FIXME: here we could cache indirect item or part of it in
297 the inode to avoid search_by_key in case of subsequent
298 access to file */
306 }
308 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
309 // That there is some MMAPED data associated with it that is yet to be written to disk.
313 }
319 }
320 // requested data are in direct item(s)
322 // we are called by bmap. FIXME: we can not map block of file
323 // when it is stored in direct item(s)
328 }
330 /* if we've got a direct item, and the buffer or page was uptodate,
331 ** we don't want to pull data off disk again. skip to the
332 ** end, where we map the buffer and return
333 */
337 /*
338 ** grab_tail_page can trigger calls to reiserfs_get_block on up to date
339 ** pages without any buffers. If the page is up to date, we don't want
340 ** read old data off disk. Set the up to date bit on the buffer instead
341 ** and jump to the end
342 */
346 }
347 // read file tail into part of page
352 /* we only want to kmap if we are reading the tail into the page.
353 ** this is not the common case, so we don't kmap until we are
354 ** sure we need to. But, this means the item might move if
355 ** kmap schedules
356 */
362 }
363 }
369 }
370 /* make sure we don't read more bytes than actually exist in
371 ** the file. This can happen in odd cases where i_size isn't
372 ** correct, and when direct item padding results in a few
373 ** extra bytes at the end of the direct item
374 */
378 chars =
384 }
393 // we done, if read direct item is not the last item of
394 // node FIXME: we could try to check right delimiting key
395 // to see whether direct item continues in the right
396 // neighbor or rely on i_size
399 // update key to look for the next piece
403 // i/o error most likely
412 finished:
418 /* this buffer has valid data, but isn't valid for io. mapping it to
419 * block #0 tells the rest of reiserfs it just has a tail in it
420 */
424 }
426 // this is called to create file map. So, _get_block_create_0 will not
427 // read direct item
430 {
435 /* do not read the direct item */
439 }
441 /* special version of get_block that is only used by grab_tail_page right
442 ** now. It is sent to block_prepare_write, and when you try to get a
443 ** block past the end of the file (or a block from a hole) it returns
444 ** -ENOENT instead of a valid buffer. block_prepare_write expects to
445 ** be able to do i/o on the buffers returned, unless an error value
446 ** is also returned.
447 **
448 ** So, this allows block_prepare_write to be used for reading a single block
449 ** in a page. Where it does not produce a valid page for holes, or past the
450 ** end of the file. This turns out to be exactly what we need for reading
451 ** tails for conversion.
452 **
453 ** The point of the wrapper is forcing a certain value for create, even
454 ** though the VFS layer is calling this function with create==1. If you
455 ** don't want to send create == GET_BLOCK_NO_HOLE to reiserfs_get_block,
456 ** don't use this function.
457 */
461 {
463 }
465 /* This is special helper for reiserfs_get_block in case we are executing
466 direct_IO request. */
468 sector_t iblock,
472 {
477 /* We set the b_size before reiserfs_get_block call since it is
478 referenced in convert_tail_for_hole() that may be called from
479 reiserfs_get_block() */
487 /* don't allow direct io onto tail pages */
489 /* make sure future calls to the direct io funcs for this offset
490 ** in the file fail by unmapping the buffer
491 */
494 }
495 /* Possible unpacked tail. Flush the data before pages have
496 disappeared */
505 }
506 out:
508 }
510 /*
511 ** helper function for when reiserfs_get_block is called for a hole
512 ** but the file tail is still in a direct item
513 ** bh_result is the buffer head for the hole
514 ** tail_offset is the offset of the start of the tail in the file
515 **
516 ** This calls prepare_write, which will start a new transaction
517 ** you should not be in a transaction, or have any paths held when you
518 ** call this.
519 */
522 loff_t tail_offset)
523 {
534 /* always try to read until the end of the block */
539 /* hole_page can be zero in case of direct_io, we are sure
540 that we cannot get here if we write with O_DIRECT into
541 tail page */
547 }
550 }
552 /* we don't have to make sure the conversion did not happen while
553 ** we were locking the page because anyone that could convert
554 ** must first take i_mutex.
555 **
556 ** We must fix the tail page for writing because it might have buffers
557 ** that are mapped, but have a block number of 0. This indicates tail
558 ** data that has been read directly into the page, and block_prepare_write
559 ** won't trigger a get_block in this case.
560 */
566 /* tail conversion might change the data in the page */
571 unlock:
575 }
576 out:
578 }
585 {
588 #ifdef REISERFS_PREALLOCATE
592 }
593 #endif
595 block);
596 }
600 {
612 /* space reserved in transaction batch:
613 . 3 balancings in direct->indirect conversion
614 . 1 block involved into reiserfs_update_sd()
615 XXX in practically impossible worst case direct2indirect()
616 can incur (much) more than 3 balancings.
617 quota update for user, group */
623 loff_t new_offset =
626 /* bad.... */
633 }
635 /* if !create, we aren't changing the FS, so we don't need to
636 ** log anything, so we don't need to start a transaction
637 */
640 /* find number of block-th logical block of the file */
645 }
646 /*
647 * if we're already in a transaction, make sure to close
648 * any new transactions we start in this func
649 */
654 /* If file is of such a size, that it might have a tail and tails are enabled
655 ** we should mark it as possibly needing tail packing on close
656 */
663 /* set the key of the first byte in the 'block'-th block of file */
666 start_trans:
671 }
673 }
674 research:
680 }
692 /* we have to allocate block for the unformatted node */
696 }
698 repeat =
703 /* restart the transaction to give the journal a chance to free
704 ** some blocks. releases the path, so we have to go back to
705 ** research if we succeed on the second try
706 */
711 repeat =
717 }
720 else
723 }
728 }
729 }
732 b_blocknr_t unfm_ptr;
733 /* 'block'-th block is in the file already (there is
734 corresponding cell in some indirect item). But it may be
735 zero unformatted node pointer (hole) */
738 /* use allocated block to plug the hole */
743 bh);
745 }
754 }
763 /* the item was found, so new blocks were not added to the file
764 ** there is no need to make sure the inode is updated with this
765 ** transaction
766 */
768 }
773 }
775 /* desired position is not found or is in the direct item. We have
776 to append file with holes up to 'block'-th block converting
777 direct items to indirect one if necessary */
784 /* indirect item has to be inserted */
790 /* we are going to add 'block'-th block to the file. Use
791 allocated block for that */
797 }
802 retval =
809 }
810 //mark_tail_converted (inode);
812 /* direct item has to be converted */
813 loff_t tail_offset;
815 tail_offset =
819 /* direct item we just found fits into block we have
820 to map. Convert it into unformatted node: use
821 bh_result for the conversion */
827 /* we have to padd file tail stored in direct item(s)
828 up to block size and convert it to unformatted
829 node. FIXME: this should also get into page cache */
832 /*
833 * ugly, but we can only end the transaction if
834 * we aren't nested
835 */
838 retval =
839 reiserfs_end_persistent_transaction
844 }
846 retval =
848 tail_offset);
854 retval);
856 /* the bitmap, the super, and the stat data == 3 */
861 inode,
862 allocated_block_nr,
864 }
866 }
868 }
869 retval =
871 tail_offset);
877 }
878 /* it is important the set_buffer_uptodate is done after
879 ** the direct2indirect. The buffer might contain valid
880 ** data newer than the data on disk (read by readpage, changed,
881 ** and then sent here by writepage). direct2indirect needs
882 ** to know if unbh was already up to date, so it can decide
883 ** if the data in unbh needs to be replaced with data from
884 ** the disk
885 */
888 /* unbh->b_page == NULL in case of DIRECT_IO request, this means
889 buffer will disappear shortly, so it should not be added to
890 */
892 /* we've converted the tail, so we must
893 ** flush unbh before the transaction commits
894 */
897 /* mark it dirty now to prevent commit_write from adding
898 ** this buffer to the inode's dirty buffer list
899 */
900 /*
901 * AKPM: changed __mark_buffer_dirty to mark_buffer_dirty().
902 * It's still atomic, but it sets the page dirty too,
903 * which makes it eligible for writeback at any time by the
904 * VM (which was also the case with __mark_buffer_dirty())
905 */
907 }
909 /* append indirect item with holes if needed, when appending
910 pointer to 'block'-th block use block, which is already
911 allocated */
914 // one block which is a fastpath
916 __u64 max_to_insert =
918 UNFM_P_SIZE;
919 __u64 blocks_needed;
923 /* indirect item has to be appended, set up key of that position */
929 //pos_in_item * inode->i_sb->s_blocksize,
934 blocks_needed =
938 s_blocksize_bits);
943 un = kmalloc(min(blocks_needed, max_to_insert) * UNFM_P_SIZE, GFP_ATOMIC); // We need to avoid scheduling.
951 max_to_insert));
952 }
954 /* we are going to add target block to the file. Use allocated
955 block for that */
963 /* paste hole to the indirect item */
964 /* If kmalloc failed, max_to_insert becomes zero and it means we
965 only have space for one block */
966 blocks_needed =
968 }
969 retval =
972 UNFM_P_SIZE *
973 blocks_needed);
982 }
984 /* We need to mark new file size in case this function will be
985 interrupted/aborted later on. And we may do this only for
986 holes. */
989 }
990 }
995 /* this loop could log more blocks than we had originally asked
996 ** for. So, we have to allow the transaction to end if it is
997 ** too big or too full. Update the inode so things are
998 ** consistent if we crash before the function returns
999 **
1000 ** release the path so that anybody waiting on the path before
1001 ** ending their transaction will be able to continue.
1002 */
1007 }
1008 /* inserting indirect pointers for a hole can take a
1009 ** long time. reschedule if needed
1010 */
1017 }
1020 "vs-825: reiserfs_get_block: "
1028 }
1037 failure:
1045 }
1050 }
1052 static int
1055 {
1057 }
1059 /* Compute real number of used bytes by file
1060 * Following three functions can go away when we'll have enough space in stat item
1061 */
1063 {
1070 /* End of file is also in full block with indirect reference, so round
1071 ** up to the next block.
1072 **
1073 ** there is just no way to know if the tail is actually packed
1074 ** on the file, so we have to assume it isn't. When we pack the
1075 ** tail, we add 4 bytes to pretend there really is an unformatted
1076 ** node pointer
1077 */
1078 bytes =
1081 sd_size;
1083 }
1087 {
1091 }
1094 }
1096 /* Compute number of blocks used by file in ReiserFS counting */
1098 {
1102 /* keeps fsck and non-quota versions of reiserfs happy */
1105 }
1107 /* files from before the quota patch might i_blocks such that
1108 ** bytes < real_space. Deal with that here to prevent it from
1109 ** going negative.
1110 */
1114 }
1116 //
1117 // BAD: new directories have stat data of new type and all other items
1118 // of old type. Version stored in the inode says about body items, so
1119 // in update_stat_data we can not rely on inode, but have to check
1120 // item version directly
1121 //
1123 // called by read_locked_inode
1125 {
1128 __u32 rdev;
1129 //int version = ITEM_VERSION_1;
1171 // there was a bug in <=3.5.23 when i_blocks could take negative
1172 // values. Starting from 3.5.17 this value could even be stored in
1173 // stat data. For such files we set i_blocks based on file
1174 // size. Just 2 notes: this can be wrong for sparce files. On-disk value will be
1175 // only updated if file's inode will ever change
1177 }
1182 /* an early bug in the quota code can give us an odd number for the
1183 ** block count. This is incorrect, fix it here.
1184 */
1187 }
1190 SD_V1_SIZE));
1191 /* nopack is initially zero for v1 objects. For v2 objects,
1192 nopack is initialised from sd_attrs */
1195 // new stat data found, but object may have old items
1196 // (directories and symlinks)
1215 else
1220 else
1226 SD_V2_SIZE));
1227 /* read persistent inode attributes from sd and initalise
1228 generic inode flags from them */
1231 }
1248 }
1249 }
1251 // update new stat data with inode fields
1253 {
1255 __u16 flags;
1268 else
1273 }
1275 // used to copy inode's fields to old stat data
1277 {
1291 else
1294 // Sigh. i_first_direct_byte is back
1297 }
1299 /* NOTE, you must prepare the buffer head before sending it here,
1300 ** and then log it after the call
1301 */
1303 loff_t size)
1304 {
1317 // path points to old stat data
1321 }
1324 }
1328 {
1342 /* look for the object's stat data */
1346 "vs-13050: reiserfs_update_sd: "
1350 }
1355 /*reiserfs_warning (inode->i_sb, "vs-13050: reiserfs_update_sd: i_nlink == 0, stat data not found"); */
1357 }
1359 "vs-13060: reiserfs_update_sd: "
1362 pos);
1365 }
1367 /* sigh, prepare_for_journal might schedule. When it schedules the
1368 ** FS might change. We have to detect that, and loop back to the
1369 ** search if the stat data item has moved
1370 */
1380 }
1382 }
1387 }
1389 /* reiserfs_read_locked_inode is called to read the inode off disk, and it
1390 ** does a make_bad_inode when things go wrong. But, we need to make sure
1391 ** and clear the key in the private portion of the inode, otherwise a
1392 ** corresponding iput might try to delete whatever object the inode last
1393 ** represented.
1394 */
1396 {
1399 }
1401 //
1402 // initially this function was derived from minix or ext2's analog and
1403 // evolved as the prototype did
1404 //
1407 {
1412 }
1414 /* looks for stat data in the tree, and fills up the fields of in-core
1415 inode stat data fields */
1418 {
1426 /* set version 1, version 2 could be used too, because stat data
1427 key is the same in both versions */
1434 /* look for the object's stat data */
1438 "vs-13070: reiserfs_read_locked_inode: "
1443 }
1445 /* a stale NFS handle can trigger this without it being an error */
1450 }
1454 /* It is possible that knfsd is trying to access inode of a file
1455 that is being removed from the disk by some other thread. As we
1456 update sd on unlink all that is required is to check for nlink
1457 here. This bug was first found by Sizif when debugging
1458 SquidNG/Butterfly, forgotten, and found again after Philippe
1459 Gramoulle <philippe.gramoulle@mmania.com> reproduced it.
1461 More logical fix would require changes in fs/inode.c:iput() to
1462 remove inode from hash-table _after_ fs cleaned disk stuff up and
1463 in iget() to return NULL if I_FREEING inode is found in
1464 hash-table. */
1465 /* Currently there is one place where it's ok to meet inode with
1466 nlink==0: processing of open-unlinked and half-truncated files
1467 during mount (fs/reiserfs/super.c:finish_unfinished()). */
1471 "vs-13075: reiserfs_read_locked_inode: "
1472 "dead inode read from disk %K. "
1476 }
1480 }
1482 /**
1483 * reiserfs_find_actor() - "find actor" reiserfs supplies to iget5_locked().
1484 *
1485 * @inode: inode from hash table to check
1486 * @opaque: "cookie" passed to iget5_locked(). This is &reiserfs_iget_args.
1487 *
1488 * This function is called by iget5_locked() to distinguish reiserfs inodes
1489 * having the same inode numbers. Such inodes can only exist due to some
1490 * error condition. One of them should be bad. Inodes with identical
1491 * inode numbers (objectids) are distinguished by parent directory ids.
1492 *
1493 */
1495 {
1499 /* args is already in CPU order */
1502 }
1505 {
1520 }
1523 /* either due to i/o error or a stale NFS handle */
1526 }
1528 }
1531 {
1545 }
1555 }
1557 }
1564 {
1567 /* fhtype happens to reflect the number of u32s encoded.
1568 * due to a bug in earlier code, fhtype might indicate there
1569 * are more u32s then actually fitted.
1570 * so if fhtype seems to be more than len, reduce fhtype.
1571 * Valid types are:
1572 * 2 - objectid + dir_id - legacy support
1573 * 3 - objectid + dir_id + generation
1574 * 4 - objectid + dir_id + objectid and dirid of parent - legacy
1575 * 5 - objectid + dir_id + generation + objectid and dirid of parent
1576 * 6 - as above plus generation of directory
1577 * 6 does not fit in NFSv2 handles
1578 */
1585 }
1591 else
1599 else
1601 }
1605 }
1609 {
1620 /* no room for directory info? return what we've stored so far */
1632 }
1635 }
1637 /* looks for stat data, then copies fields to it, marks the buffer
1638 containing stat data as dirty */
1639 /* reiserfs inodes are never really dirty, since the dirty inode call
1640 ** always logs them. This call allows the VFS inode marking routines
1641 ** to properly mark inodes for datasync and such, but only actually
1642 ** does something when called for a synchronous update.
1643 */
1645 {
1651 /* memory pressure can sometimes initiate write_inode calls with sync == 1,
1652 ** these cases are just when the system needs ram, not when the
1653 ** inode needs to reach disk for safety, and they can safely be
1654 ** ignored because the altered inode has already been logged.
1655 */
1661 }
1663 }
1665 }
1667 /* stat data of new object is inserted already, this inserts the item
1668 containing "." and ".." entries */
1673 {
1686 /* compose item head for new item. Directories consist of items of
1687 old type (ITEM_VERSION_1). Do not set key (second arg is 0), it
1688 is done by reiserfs_new_inode */
1705 }
1707 /* look for place in the tree for new item */
1713 }
1720 }
1722 /* insert item, that is empty directory item */
1724 }
1726 /* stat data of object has been inserted, this inserts the item
1727 containing the body of symlink */
1728 static int reiserfs_new_symlink(struct reiserfs_transaction_handle *th, struct inode *inode, /* Inode of symlink */
1732 {
1747 /* look for place in the tree for new item */
1753 }
1760 }
1762 /* insert item, that is body of symlink */
1764 }
1766 /* inserts the stat data into the tree, and then calls
1767 reiserfs_new_directory (to insert ".", ".." item if new object is
1768 directory) or reiserfs_new_symlink (to insert symlink body if new
1769 object is symlink) or nothing (if new object is regular file)
1771 NOTE! uid and gid must already be set in the inode. If we return
1772 non-zero due to an error, we have to drop the quota previously allocated
1773 for the fresh inode. This can only be done outside a transaction, so
1774 if we return non-zero, we also end the transaction. */
1777 /* 0 for regular, EMTRY_DIR_SIZE for dirs,
1778 strlen (symname) for symlinks) */
1781 {
1795 }
1799 }
1803 /* item head of new item */
1809 }
1811 /* not a perfect generation count, as object ids can be reused, but
1812 ** this is as good as reiserfs can do right now.
1813 ** note that the private part of inode isn't filled in yet, we have
1814 ** to use the directory.
1815 */
1817 else
1818 #if defined( USE_INODE_GENERATION_COUNTER )
1821 #else
1823 #endif
1825 /* fill stat data */
1828 /* uid and gid must already be set by the caller for quota init */
1830 /* symlink cannot be immutable or append only, right? */
1857 else
1861 /* key to search for correct place for new stat data */
1866 /* find proper place for inserting of stat data */
1871 }
1876 }
1880 /* i_uid or i_gid is too big to be stored in stat data v3.5 */
1883 }
1887 }
1888 // these do not go to on-disk stat data
1892 // store in in-core inode the key of stat data and version all
1893 // object items will have (directory items will have old offset
1894 // format, other new objects will consist of new items)
1898 else
1902 else
1905 /* insert the stat data into the tree */
1906 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1909 #endif
1910 retval =
1917 }
1918 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1921 #endif
1923 /* insert item with "." and ".." */
1924 retval =
1926 }
1929 /* insert body of symlink */
1932 retval =
1934 i_size);
1935 }
1941 }
1943 /* XXX CHECK THIS */
1951 }
1957 }
1965 /* it looks like you can easily compress these two goto targets into
1966 * one. Keeping it like this doesn't actually hurt anything, and they
1967 * are place holders for what the quota code actually needs.
1968 */
1969 out_bad_inode:
1970 /* Invalidate the object, nothing was inserted yet */
1973 /* Quota change must be inside a transaction for journaling */
1976 out_end_trans:
1978 /* Drop can be outside and it needs more credits so it's better to have it outside */
1983 out_inserted_sd:
1987 /* If we were inheriting an ACL, we need to release the lock so that
1988 * iput doesn't deadlock in reiserfs_delete_xattrs. The locking
1989 * code really needs to be reworked, but this will take care of it
1990 * for now. -jeffm */
1998 }
2000 /*
2001 ** finds the tail page in the page cache,
2002 ** reads the last block in.
2003 **
2004 ** On success, page_result is set to a locked, pinned page, and bh_result
2005 ** is set to an up to date buffer for the last block in the file. returns 0.
2006 **
2007 ** tail conversion is not done, so bh_result might not be valid for writing
2008 ** check buffer_mapped(bh_result) and bh_result->b_blocknr != 0 before
2009 ** trying to write the block.
2010 **
2011 ** on failure, nonzero is returned, page_result and bh_result are untouched.
2012 */
2016 {
2018 /* we want the page with the last byte in the file,
2019 ** not the page that will hold the next byte for appending
2020 */
2031 /* we know that we are only called with inode->i_size > 0.
2032 ** we also know that a file tail can never be as big as a block
2033 ** If i_size % blocksize == 0, our file is currently block aligned
2034 ** and it won't need converting or zeroing after a truncate.
2035 */
2038 }
2043 }
2044 /* start within the page of the last block in the file */
2048 reiserfs_get_block_create_0);
2057 }
2063 /* note, this should never happen, prepare_write should
2064 ** be taking care of this for us. If the buffer isn't up to date,
2065 ** I've screwed up the code to find the buffer, or the code to
2066 ** call prepare_write
2067 */
2074 }
2078 out:
2081 unlock:
2085 }
2087 /*
2088 ** vfs version of truncate file. Must NOT be called with
2089 ** a transaction already started.
2090 **
2091 ** some code taken from block_truncate_page
2092 */
2094 {
2096 /* we want the offset for the first byte after the end of the file */
2109 // -ENOENT means we truncated past the end of the file,
2110 // and get_block_create_0 could not find a block to read in,
2111 // which is ok.
2115 error);
2118 }
2119 }
2121 /* so, if page != NULL, we have a buffer head for the offset at
2122 ** the end of the file. if the bh is mapped, and bh->b_blocknr != 0,
2123 ** then we have an unformatted node. Otherwise, we have a direct item,
2124 ** and no zeroing is required on disk. We zero after the truncate,
2125 ** because the truncate might pack the item anyway
2126 ** (it will unmap bh if it packs).
2127 */
2128 /* it is enough to reserve space in transaction for 2 balancings:
2129 one for "save" link adding and another for the first
2130 cut_from_item. 1 is for update_sd */
2137 /* we are doing real truncate: if the system crashes before the last
2138 transaction of truncating gets committed - on reboot the file
2139 either appears truncated properly or not truncated at all */
2142 error =
2147 /* check reiserfs_do_truncate after ending the transaction */
2151 }
2157 }
2161 /* if we are not on a block boundary */
2172 }
2173 }
2176 }
2180 out:
2184 }
2187 }
2192 {
2210 /* catch places below that try to log something without starting a trans */
2215 }
2218 start_over:
2222 research:
2227 }
2234 /* we've found an unformatted node */
2239 bytes_copied);
2240 }
2242 /* crap, we are writing to a hole */
2245 }
2258 /* vs-3050 is gone, no need to drop the path */
2267 bh);
2269 }
2270 }
2278 }
2281 copy_size);
2287 /* are there still bytes left? */
2292 copy_size);
2294 }
2301 }
2304 out:
2311 }
2314 /* this is where we fill in holes in the file. */
2317 GET_BLOCK_CREATE | GET_BLOCK_NO_IMUX
2322 /* get_block failed to find a mapped unformatted node. */
2325 }
2326 }
2327 }
2331 /* we've copied data from the page into the direct item, so the
2332 * buffer in the page is now clean, mark it to reflect that.
2333 */
2337 }
2339 }
2341 /*
2342 * mason@suse.com: updated in 2.5.54 to follow the same general io
2343 * start/recovery path as __block_write_full_page, along with special
2344 * code to handle reiserfs tails.
2345 */
2348 {
2362 /* no logging allowed when nonblocking or from PF_MEMALLOC */
2367 }
2369 /* The page dirty bit is cleared before writepage is called, which
2370 * means we have to tell create_empty_buffers to make dirty buffers
2371 * The page really should be up to date at this point, so tossing
2372 * in the BH_Uptodate is just a sanity check.
2373 */
2377 }
2380 /* last page in the file, zero out any contents past the
2381 ** last byte in the file
2382 */
2388 /* no file contents in this page */
2392 }
2397 }
2400 /* first map all the buffers, logging any direct items we find */
2406 /* not mapped yet, or it points to a direct item, search
2407 * the btree for the mapping info, and log any direct
2408 * items found
2409 */
2412 }
2413 }
2415 block++;
2418 /*
2419 * we start the transaction after map_block_for_writepage,
2420 * because it can create holes in the file (an unbounded operation).
2421 * starting it here, we can make a reliable estimate for how many
2422 * blocks we're going to log
2423 */
2431 }
2433 }
2434 /* now go through and lock any dirty buffers on the page */
2446 }
2447 /* from this point on, we know the buffer is mapped to a
2448 * real block and not a direct item
2449 */
2456 }
2457 }
2462 }
2470 }
2475 /*
2476 * since any buffer might be the only dirty buffer on the page,
2477 * the first submit_bh can bring the page out of writeback.
2478 * be careful with the buffers.
2479 */
2484 nr++;
2485 }
2491 done:
2493 /*
2494 * if this page only had a direct item, it is very possible for
2495 * no io to be required without there being an error. Or,
2496 * someone else could have locked them and sent them down the
2497 * pipe without locking the page
2498 */
2504 }
2510 }
2513 fail:
2514 /* catches various errors, we need to make sure any valid dirty blocks
2515 * get to the media. The page is currently locked and not marked for
2516 * writeback
2517 */
2526 /*
2527 * clear any dirty bits that might have come from getting
2528 * attached to a dirty page
2529 */
2531 }
2543 nr++;
2544 }
2549 }
2552 {
2554 }
2557 {
2561 }
2565 {
2575 journal_info;
2580 }
2585 /* this gets a little ugly. If reiserfs_get_block returned an
2586 * error and left a transacstion running, we've got to close it,
2587 * and we've got to free handle if it was a persistent transaction.
2588 *
2589 * But, if we had nested into an existing transaction, we need
2590 * to just drop the ref count on the handle.
2591 *
2592 * If old_ref == 0, the transaction is from reiserfs_get_block,
2593 * and it was a persistent trans. Otherwise, it was nested above.
2594 */
2605 }
2606 }
2607 }
2610 }
2613 {
2615 }
2619 {
2629 }
2632 /* generic_commit_write does this for us, but does not update the
2633 ** transaction tracking stuff when the size changes. So, we have
2634 ** to do the i_size updates here.
2635 */
2639 /* If the file have grown beyond the border where it
2640 can have a tail, unmark it as needing a tail
2641 packing */
2652 }
2655 /*
2656 * this will just nest into our transaction. It's important
2657 * to use mark_inode_dirty so the inode gets pushed around on the
2658 * dirty lists, and so that O_SYNC works as expected
2659 */
2667 }
2676 }
2678 out:
2681 journal_error:
2688 }
2691 }
2694 {
2698 else
2702 else
2706 else
2710 else
2714 else
2716 }
2717 }
2720 {
2724 else
2728 else
2732 else
2736 else
2738 }
2739 }
2741 /* decide if this buffer needs to stay around for data logging or ordered
2742 ** write purposes
2743 */
2745 {
2753 }
2754 /* the page is locked, and the only places that log a data buffer
2755 * also lock the page.
2756 */
2758 /*
2759 * very conservative, leave the buffer pinned if
2760 * anyone might need it.
2761 */
2764 }
2769 /* why is this safe?
2770 * reiserfs_setattr updates i_size in the on disk
2771 * stat data before allowing vmtruncate to be called.
2772 *
2773 * If buffer was put onto the ordered list for this
2774 * transaction, we know for sure either this transaction
2775 * or an older one already has updated i_size on disk,
2776 * and this ordered data won't be referenced in the file
2777 * if we crash.
2778 *
2779 * if the buffer was put onto the ordered list for an older
2780 * transaction, we need to leave it around
2781 */
2785 }
2786 free_jh:
2789 }
2793 }
2795 /* clm -- taken from fs/buffer.c:block_invalidate_page */
2797 {
2817 /*
2818 * is this block fully invalidated?
2819 */
2823 else
2825 }
2830 /*
2831 * We release buffers only if the entire page is being invalidated.
2832 * The get_block cached value has been unconditionally invalidated,
2833 * so real IO is not possible anymore.
2834 */
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 }
2941 /* stat data of format v3.5 has 16 bit uid and gid */
2944 }
2960 /* (user+group)*(old+new) structure - we count quota info and , inode write (sb, inode) */
2961 error =
2963 jbegin_count);
2966 error =
2970 jbegin_count);
2972 }
2973 /* Update corresponding info in inode so that everything is in
2974 * one transaction */
2980 error =
2982 }
2983 }
2986 }
2991 }
2993 out:
2996 }
3010 };