| blob | a7d0e903831cae648b00493e83514ba69e7f73b9 |
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/exportfs.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>
27 {
28 /* We need blocks for transaction + (user+group) quota update (possibly delete) */
39 /* The = 0 happens when we abort creating a new inode for some reason like lack of space.. */
40 if (!(inode->i_state & I_NEW) && INODE_PKEY(inode)->k_objectid != 0) { /* also handles bad_inode case */
49 /* Do quota update inside a transaction for journaled quotas. We must do that
50 * after delete_object so that quota updates go into the same transaction as
51 * stat data deletion */
58 /* check return value from reiserfs_delete_object after
59 * ending the transaction
60 */
64 /* all items of file are deleted, so we can remove "save" link */
66 * about an error here */
68 /* no object items are in the tree */
69 ;
70 }
71 out:
75 }
79 {
87 }
89 /* take base of inode_key (it comes from inode always) (dirid, objectid) and version from an inode, set
90 offset and type of key */
93 {
97 length);
98 }
100 //
101 // when key is 0, do not set version and short key
102 //
107 {
112 }
117 /* set_ih_free_space (ih, 0); */
118 // for directory items it is entry count, for directs and stat
119 // datas - 0xffff, for indirects - 0
121 }
123 //
124 // FIXME: we might cache recently accessed indirect item
126 // Ugh. Not too eager for that....
127 // I cut the code until such time as I see a convincing argument (benchmark).
128 // I don't want a bloated inode struct..., and I don't like code complexity....
130 /* cutting the code is fine, since it really isn't in use yet and is easy
131 ** to add back in. But, Vladimir has a really good idea here. Think
132 ** about what happens for reading a file. For each page,
133 ** The VFS layer calls reiserfs_readpage, who searches the tree to find
134 ** an indirect item. This indirect item has X number of pointers, where
135 ** X is a big number if we've done the block allocation right. But,
136 ** we only use one or two of these pointers during each call to readpage,
137 ** needlessly researching again later on.
138 **
139 ** The size of the cache could be dynamic based on the size of the file.
140 **
141 ** I'd also like to see us cache the location the stat data item, since
142 ** we are needlessly researching for that frequently.
143 **
144 ** --chris
145 */
147 /* If this page has a file tail in it, and
148 ** it was read in by get_block_create_0, the page data is valid,
149 ** but tail is still sitting in a direct item, and we can't write to
150 ** it. So, look through this page, and check all the mapped buffers
151 ** to make sure they have valid block numbers. Any that don't need
152 ** to be unmapped, so that block_prepare_write will correctly call
153 ** reiserfs_get_block to convert the tail into an unformatted node
154 */
156 {
166 }
169 }
170 }
172 /* reiserfs_get_block does not need to allocate a block only if it has been
173 done already or non-hole position has been found in the indirect item */
177 {
184 }
187 {
189 }
193 {
195 }
197 //
198 // files which were created in the earlier version can not be longer,
199 // than 2 gb
200 //
202 {
208 }
212 {
222 /* we cannot restart while nested */
225 }
232 }
234 }
236 // it is called by get_block when create == 0. Returns block number
237 // for 'block'-th logical block of file. When it hits direct item it
238 // returns 0 (being called from bmap) or read direct item into piece
239 // of page (bh_result)
241 // Please improve the english/clarity in the comment above, as it is
242 // hard to understand.
246 {
260 // prepare the key to look for the 'block'-th block of file
265 research:
273 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
274 // That there is some MMAPED data associated with it that is yet to be written to disk.
278 }
280 }
281 //
287 /* FIXME: here we could cache indirect item or part of it in
288 the inode to avoid search_by_key in case of subsequent
289 access to file */
297 }
299 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
300 // That there is some MMAPED data associated with it that is yet to be written to disk.
304 }
310 }
311 // requested data are in direct item(s)
313 // we are called by bmap. FIXME: we can not map block of file
314 // when it is stored in direct item(s)
319 }
321 /* if we've got a direct item, and the buffer or page was uptodate,
322 ** we don't want to pull data off disk again. skip to the
323 ** end, where we map the buffer and return
324 */
328 /*
329 ** grab_tail_page can trigger calls to reiserfs_get_block on up to date
330 ** pages without any buffers. If the page is up to date, we don't want
331 ** read old data off disk. Set the up to date bit on the buffer instead
332 ** and jump to the end
333 */
337 }
338 // read file tail into part of page
343 /* we only want to kmap if we are reading the tail into the page.
344 ** this is not the common case, so we don't kmap until we are
345 ** sure we need to. But, this means the item might move if
346 ** kmap schedules
347 */
353 }
354 }
360 }
361 /* make sure we don't read more bytes than actually exist in
362 ** the file. This can happen in odd cases where i_size isn't
363 ** correct, and when direct item padding results in a few
364 ** extra bytes at the end of the direct item
365 */
369 chars =
375 }
384 // we done, if read direct item is not the last item of
385 // node FIXME: we could try to check right delimiting key
386 // to see whether direct item continues in the right
387 // neighbor or rely on i_size
390 // update key to look for the next piece
394 // i/o error most likely
403 finished:
409 /* this buffer has valid data, but isn't valid for io. mapping it to
410 * block #0 tells the rest of reiserfs it just has a tail in it
411 */
415 }
417 // this is called to create file map. So, _get_block_create_0 will not
418 // read direct item
421 {
426 /* do not read the direct item */
430 }
432 /* special version of get_block that is only used by grab_tail_page right
433 ** now. It is sent to block_prepare_write, and when you try to get a
434 ** block past the end of the file (or a block from a hole) it returns
435 ** -ENOENT instead of a valid buffer. block_prepare_write expects to
436 ** be able to do i/o on the buffers returned, unless an error value
437 ** is also returned.
438 **
439 ** So, this allows block_prepare_write to be used for reading a single block
440 ** in a page. Where it does not produce a valid page for holes, or past the
441 ** end of the file. This turns out to be exactly what we need for reading
442 ** tails for conversion.
443 **
444 ** The point of the wrapper is forcing a certain value for create, even
445 ** though the VFS layer is calling this function with create==1. If you
446 ** don't want to send create == GET_BLOCK_NO_HOLE to reiserfs_get_block,
447 ** don't use this function.
448 */
452 {
454 }
456 /* This is special helper for reiserfs_get_block in case we are executing
457 direct_IO request. */
459 sector_t iblock,
462 {
467 /* We set the b_size before reiserfs_get_block call since it is
468 referenced in convert_tail_for_hole() that may be called from
469 reiserfs_get_block() */
477 /* don't allow direct io onto tail pages */
479 /* make sure future calls to the direct io funcs for this offset
480 ** in the file fail by unmapping the buffer
481 */
484 }
485 /* Possible unpacked tail. Flush the data before pages have
486 disappeared */
495 }
496 out:
498 }
500 /*
501 ** helper function for when reiserfs_get_block is called for a hole
502 ** but the file tail is still in a direct item
503 ** bh_result is the buffer head for the hole
504 ** tail_offset is the offset of the start of the tail in the file
505 **
506 ** This calls prepare_write, which will start a new transaction
507 ** you should not be in a transaction, or have any paths held when you
508 ** call this.
509 */
512 loff_t tail_offset)
513 {
524 /* always try to read until the end of the block */
529 /* hole_page can be zero in case of direct_io, we are sure
530 that we cannot get here if we write with O_DIRECT into
531 tail page */
537 }
540 }
542 /* we don't have to make sure the conversion did not happen while
543 ** we were locking the page because anyone that could convert
544 ** must first take i_mutex.
545 **
546 ** We must fix the tail page for writing because it might have buffers
547 ** that are mapped, but have a block number of 0. This indicates tail
548 ** data that has been read directly into the page, and block_prepare_write
549 ** won't trigger a get_block in this case.
550 */
556 /* tail conversion might change the data in the page */
561 unlock:
565 }
566 out:
568 }
575 {
578 #ifdef REISERFS_PREALLOCATE
582 }
583 #endif
585 block);
586 }
590 {
602 /* space reserved in transaction batch:
603 . 3 balancings in direct->indirect conversion
604 . 1 block involved into reiserfs_update_sd()
605 XXX in practically impossible worst case direct2indirect()
606 can incur (much) more than 3 balancings.
607 quota update for user, group */
613 loff_t new_offset =
616 /* bad.... */
623 }
625 /* if !create, we aren't changing the FS, so we don't need to
626 ** log anything, so we don't need to start a transaction
627 */
630 /* find number of block-th logical block of the file */
635 }
636 /*
637 * if we're already in a transaction, make sure to close
638 * any new transactions we start in this func
639 */
644 /* If file is of such a size, that it might have a tail and tails are enabled
645 ** we should mark it as possibly needing tail packing on close
646 */
653 /* set the key of the first byte in the 'block'-th block of file */
656 start_trans:
661 }
663 }
664 research:
670 }
682 /* we have to allocate block for the unformatted node */
686 }
688 repeat =
693 /* restart the transaction to give the journal a chance to free
694 ** some blocks. releases the path, so we have to go back to
695 ** research if we succeed on the second try
696 */
701 repeat =
707 }
710 else
713 }
718 }
719 }
722 b_blocknr_t unfm_ptr;
723 /* 'block'-th block is in the file already (there is
724 corresponding cell in some indirect item). But it may be
725 zero unformatted node pointer (hole) */
728 /* use allocated block to plug the hole */
733 bh);
735 }
744 }
753 /* the item was found, so new blocks were not added to the file
754 ** there is no need to make sure the inode is updated with this
755 ** transaction
756 */
758 }
763 }
765 /* desired position is not found or is in the direct item. We have
766 to append file with holes up to 'block'-th block converting
767 direct items to indirect one if necessary */
774 /* indirect item has to be inserted */
780 /* we are going to add 'block'-th block to the file. Use
781 allocated block for that */
787 }
792 retval =
799 }
800 //mark_tail_converted (inode);
802 /* direct item has to be converted */
803 loff_t tail_offset;
805 tail_offset =
809 /* direct item we just found fits into block we have
810 to map. Convert it into unformatted node: use
811 bh_result for the conversion */
817 /* we have to padd file tail stored in direct item(s)
818 up to block size and convert it to unformatted
819 node. FIXME: this should also get into page cache */
822 /*
823 * ugly, but we can only end the transaction if
824 * we aren't nested
825 */
828 retval =
829 reiserfs_end_persistent_transaction
834 }
836 retval =
838 tail_offset);
844 retval);
846 /* the bitmap, the super, and the stat data == 3 */
851 inode,
852 allocated_block_nr,
854 }
856 }
858 }
859 retval =
861 tail_offset);
867 }
868 /* it is important the set_buffer_uptodate is done after
869 ** the direct2indirect. The buffer might contain valid
870 ** data newer than the data on disk (read by readpage, changed,
871 ** and then sent here by writepage). direct2indirect needs
872 ** to know if unbh was already up to date, so it can decide
873 ** if the data in unbh needs to be replaced with data from
874 ** the disk
875 */
878 /* unbh->b_page == NULL in case of DIRECT_IO request, this means
879 buffer will disappear shortly, so it should not be added to
880 */
882 /* we've converted the tail, so we must
883 ** flush unbh before the transaction commits
884 */
887 /* mark it dirty now to prevent commit_write from adding
888 ** this buffer to the inode's dirty buffer list
889 */
890 /*
891 * AKPM: changed __mark_buffer_dirty to mark_buffer_dirty().
892 * It's still atomic, but it sets the page dirty too,
893 * which makes it eligible for writeback at any time by the
894 * VM (which was also the case with __mark_buffer_dirty())
895 */
897 }
899 /* append indirect item with holes if needed, when appending
900 pointer to 'block'-th block use block, which is already
901 allocated */
904 // one block which is a fastpath
906 __u64 max_to_insert =
908 UNFM_P_SIZE;
909 __u64 blocks_needed;
913 /* indirect item has to be appended, set up key of that position */
919 //pos_in_item * inode->i_sb->s_blocksize,
924 blocks_needed =
928 s_blocksize_bits);
933 un = kzalloc(min(blocks_needed, max_to_insert) * UNFM_P_SIZE, GFP_ATOMIC); // We need to avoid scheduling.
938 }
939 }
941 /* we are going to add target block to the file. Use allocated
942 block for that */
950 /* paste hole to the indirect item */
951 /* If kmalloc failed, max_to_insert becomes zero and it means we
952 only have space for one block */
953 blocks_needed =
955 }
956 retval =
959 UNFM_P_SIZE *
960 blocks_needed);
969 }
971 /* We need to mark new file size in case this function will be
972 interrupted/aborted later on. And we may do this only for
973 holes. */
976 }
977 }
982 /* this loop could log more blocks than we had originally asked
983 ** for. So, we have to allow the transaction to end if it is
984 ** too big or too full. Update the inode so things are
985 ** consistent if we crash before the function returns
986 **
987 ** release the path so that anybody waiting on the path before
988 ** ending their transaction will be able to continue.
989 */
994 }
995 /* inserting indirect pointers for a hole can take a
996 ** long time. reschedule if needed
997 */
1004 }
1007 "vs-825: reiserfs_get_block: "
1015 }
1024 failure:
1032 }
1037 }
1039 static int
1042 {
1044 }
1046 /* Compute real number of used bytes by file
1047 * Following three functions can go away when we'll have enough space in stat item
1048 */
1050 {
1057 /* End of file is also in full block with indirect reference, so round
1058 ** up to the next block.
1059 **
1060 ** there is just no way to know if the tail is actually packed
1061 ** on the file, so we have to assume it isn't. When we pack the
1062 ** tail, we add 4 bytes to pretend there really is an unformatted
1063 ** node pointer
1064 */
1065 bytes =
1068 sd_size;
1070 }
1074 {
1078 }
1081 }
1083 /* Compute number of blocks used by file in ReiserFS counting */
1085 {
1089 /* keeps fsck and non-quota versions of reiserfs happy */
1092 }
1094 /* files from before the quota patch might i_blocks such that
1095 ** bytes < real_space. Deal with that here to prevent it from
1096 ** going negative.
1097 */
1101 }
1103 //
1104 // BAD: new directories have stat data of new type and all other items
1105 // of old type. Version stored in the inode says about body items, so
1106 // in update_stat_data we can not rely on inode, but have to check
1107 // item version directly
1108 //
1110 // called by read_locked_inode
1112 {
1115 __u32 rdev;
1116 //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 }
1520 {
1533 }
1543 }
1545 }
1549 {
1550 /* fhtype happens to reflect the number of u32s encoded.
1551 * due to a bug in earlier code, fhtype might indicate there
1552 * are more u32s then actually fitted.
1553 * so if fhtype seems to be more than len, reduce fhtype.
1554 * Valid types are:
1555 * 2 - objectid + dir_id - legacy support
1556 * 3 - objectid + dir_id + generation
1557 * 4 - objectid + dir_id + objectid and dirid of parent - legacy
1558 * 5 - objectid + dir_id + generation + objectid and dirid of parent
1559 * 6 - as above plus generation of directory
1560 * 6 does not fit in NFSv2 handles
1561 */
1568 }
1572 }
1576 {
1584 }
1588 {
1599 /* no room for directory info? return what we've stored so far */
1611 }
1614 }
1616 /* looks for stat data, then copies fields to it, marks the buffer
1617 containing stat data as dirty */
1618 /* reiserfs inodes are never really dirty, since the dirty inode call
1619 ** always logs them. This call allows the VFS inode marking routines
1620 ** to properly mark inodes for datasync and such, but only actually
1621 ** does something when called for a synchronous update.
1622 */
1624 {
1630 /* memory pressure can sometimes initiate write_inode calls with sync == 1,
1631 ** these cases are just when the system needs ram, not when the
1632 ** inode needs to reach disk for safety, and they can safely be
1633 ** ignored because the altered inode has already been logged.
1634 */
1640 }
1642 }
1644 }
1646 /* stat data of new object is inserted already, this inserts the item
1647 containing "." and ".." entries */
1652 {
1665 /* compose item head for new item. Directories consist of items of
1666 old type (ITEM_VERSION_1). Do not set key (second arg is 0), it
1667 is done by reiserfs_new_inode */
1684 }
1686 /* look for place in the tree for new item */
1692 }
1699 }
1701 /* insert item, that is empty directory item */
1703 }
1705 /* stat data of object has been inserted, this inserts the item
1706 containing the body of symlink */
1707 static int reiserfs_new_symlink(struct reiserfs_transaction_handle *th, struct inode *inode, /* Inode of symlink */
1711 {
1726 /* look for place in the tree for new item */
1732 }
1739 }
1741 /* insert item, that is body of symlink */
1743 }
1745 /* inserts the stat data into the tree, and then calls
1746 reiserfs_new_directory (to insert ".", ".." item if new object is
1747 directory) or reiserfs_new_symlink (to insert symlink body if new
1748 object is symlink) or nothing (if new object is regular file)
1750 NOTE! uid and gid must already be set in the inode. If we return
1751 non-zero due to an error, we have to drop the quota previously allocated
1752 for the fresh inode. This can only be done outside a transaction, so
1753 if we return non-zero, we also end the transaction. */
1756 /* 0 for regular, EMTRY_DIR_SIZE for dirs,
1757 strlen (symname) for symlinks) */
1760 {
1774 }
1778 }
1782 /* item head of new item */
1788 }
1790 /* not a perfect generation count, as object ids can be reused, but
1791 ** this is as good as reiserfs can do right now.
1792 ** note that the private part of inode isn't filled in yet, we have
1793 ** to use the directory.
1794 */
1796 else
1797 #if defined( USE_INODE_GENERATION_COUNTER )
1800 #else
1802 #endif
1804 /* fill stat data */
1807 /* uid and gid must already be set by the caller for quota init */
1809 /* symlink cannot be immutable or append only, right? */
1837 else
1841 /* key to search for correct place for new stat data */
1846 /* find proper place for inserting of stat data */
1851 }
1856 }
1860 /* i_uid or i_gid is too big to be stored in stat data v3.5 */
1863 }
1867 }
1868 // these do not go to on-disk stat data
1871 // store in in-core inode the key of stat data and version all
1872 // object items will have (directory items will have old offset
1873 // format, other new objects will consist of new items)
1877 else
1881 else
1884 /* insert the stat data into the tree */
1885 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1888 #endif
1889 retval =
1896 }
1897 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1900 #endif
1902 /* insert item with "." and ".." */
1903 retval =
1905 }
1908 /* insert body of symlink */
1911 retval =
1913 i_size);
1914 }
1920 }
1922 /* XXX CHECK THIS */
1930 }
1936 }
1944 /* it looks like you can easily compress these two goto targets into
1945 * one. Keeping it like this doesn't actually hurt anything, and they
1946 * are place holders for what the quota code actually needs.
1947 */
1948 out_bad_inode:
1949 /* Invalidate the object, nothing was inserted yet */
1952 /* Quota change must be inside a transaction for journaling */
1955 out_end_trans:
1957 /* Drop can be outside and it needs more credits so it's better to have it outside */
1962 out_inserted_sd:
1966 /* If we were inheriting an ACL, we need to release the lock so that
1967 * iput doesn't deadlock in reiserfs_delete_xattrs. The locking
1968 * code really needs to be reworked, but this will take care of it
1969 * for now. -jeffm */
1970 #ifdef CONFIG_REISERFS_FS_POSIX_ACL
1976 #endif
1979 }
1981 /*
1982 ** finds the tail page in the page cache,
1983 ** reads the last block in.
1984 **
1985 ** On success, page_result is set to a locked, pinned page, and bh_result
1986 ** is set to an up to date buffer for the last block in the file. returns 0.
1987 **
1988 ** tail conversion is not done, so bh_result might not be valid for writing
1989 ** check buffer_mapped(bh_result) and bh_result->b_blocknr != 0 before
1990 ** trying to write the block.
1991 **
1992 ** on failure, nonzero is returned, page_result and bh_result are untouched.
1993 */
1997 {
1999 /* we want the page with the last byte in the file,
2000 ** not the page that will hold the next byte for appending
2001 */
2012 /* we know that we are only called with inode->i_size > 0.
2013 ** we also know that a file tail can never be as big as a block
2014 ** If i_size % blocksize == 0, our file is currently block aligned
2015 ** and it won't need converting or zeroing after a truncate.
2016 */
2019 }
2024 }
2025 /* start within the page of the last block in the file */
2029 reiserfs_get_block_create_0);
2038 }
2044 /* note, this should never happen, prepare_write should
2045 ** be taking care of this for us. If the buffer isn't up to date,
2046 ** I've screwed up the code to find the buffer, or the code to
2047 ** call prepare_write
2048 */
2055 }
2059 out:
2062 unlock:
2066 }
2068 /*
2069 ** vfs version of truncate file. Must NOT be called with
2070 ** a transaction already started.
2071 **
2072 ** some code taken from block_truncate_page
2073 */
2075 {
2077 /* we want the offset for the first byte after the end of the file */
2090 // -ENOENT means we truncated past the end of the file,
2091 // and get_block_create_0 could not find a block to read in,
2092 // which is ok.
2096 error);
2099 }
2100 }
2102 /* so, if page != NULL, we have a buffer head for the offset at
2103 ** the end of the file. if the bh is mapped, and bh->b_blocknr != 0,
2104 ** then we have an unformatted node. Otherwise, we have a direct item,
2105 ** and no zeroing is required on disk. We zero after the truncate,
2106 ** because the truncate might pack the item anyway
2107 ** (it will unmap bh if it packs).
2108 */
2109 /* it is enough to reserve space in transaction for 2 balancings:
2110 one for "save" link adding and another for the first
2111 cut_from_item. 1 is for update_sd */
2118 /* we are doing real truncate: if the system crashes before the last
2119 transaction of truncating gets committed - on reboot the file
2120 either appears truncated properly or not truncated at all */
2123 error =
2128 /* check reiserfs_do_truncate after ending the transaction */
2132 }
2138 }
2142 /* if we are not on a block boundary */
2148 }
2149 }
2152 }
2156 out:
2160 }
2163 }
2168 {
2186 /* catch places below that try to log something without starting a trans */
2191 }
2194 start_over:
2198 research:
2203 }
2210 /* we've found an unformatted node */
2215 bytes_copied);
2216 }
2218 /* crap, we are writing to a hole */
2221 }
2234 /* vs-3050 is gone, no need to drop the path */
2243 bh);
2245 }
2246 }
2254 }
2257 copy_size);
2263 /* are there still bytes left? */
2268 copy_size);
2270 }
2277 }
2280 out:
2287 }
2290 /* this is where we fill in holes in the file. */
2293 GET_BLOCK_CREATE | GET_BLOCK_NO_IMUX
2298 /* get_block failed to find a mapped unformatted node. */
2301 }
2302 }
2303 }
2307 /* we've copied data from the page into the direct item, so the
2308 * buffer in the page is now clean, mark it to reflect that.
2309 */
2313 }
2315 }
2317 /*
2318 * mason@suse.com: updated in 2.5.54 to follow the same general io
2319 * start/recovery path as __block_write_full_page, along with special
2320 * code to handle reiserfs tails.
2321 */
2324 {
2329 sector_t last_block;
2339 /* no logging allowed when nonblocking or from PF_MEMALLOC */
2344 }
2346 /* The page dirty bit is cleared before writepage is called, which
2347 * means we have to tell create_empty_buffers to make dirty buffers
2348 * The page really should be up to date at this point, so tossing
2349 * in the BH_Uptodate is just a sanity check.
2350 */
2354 }
2357 /* last page in the file, zero out any contents past the
2358 ** last byte in the file
2359 */
2364 /* no file contents in this page */
2368 }
2370 }
2374 /* first map all the buffers, logging any direct items we find */
2377 /*
2378 * This can happen when the block size is less than
2379 * the page size. The corresponding bytes in the page
2380 * were zero filled above
2381 */
2388 /* not mapped yet, or it points to a direct item, search
2389 * the btree for the mapping info, and log any direct
2390 * items found
2391 */
2394 }
2395 }
2397 block++;
2400 /*
2401 * we start the transaction after map_block_for_writepage,
2402 * because it can create holes in the file (an unbounded operation).
2403 * starting it here, we can make a reliable estimate for how many
2404 * blocks we're going to log
2405 */
2413 }
2415 }
2416 /* now go through and lock any dirty buffers on the page */
2428 }
2429 /* from this point on, we know the buffer is mapped to a
2430 * real block and not a direct item
2431 */
2438 }
2439 }
2444 }
2452 }
2457 /*
2458 * since any buffer might be the only dirty buffer on the page,
2459 * the first submit_bh can bring the page out of writeback.
2460 * be careful with the buffers.
2461 */
2466 nr++;
2467 }
2473 done:
2475 /*
2476 * if this page only had a direct item, it is very possible for
2477 * no io to be required without there being an error. Or,
2478 * someone else could have locked them and sent them down the
2479 * pipe without locking the page
2480 */
2486 }
2492 }
2495 fail:
2496 /* catches various errors, we need to make sure any valid dirty blocks
2497 * get to the media. The page is currently locked and not marked for
2498 * writeback
2499 */
2508 /*
2509 * clear any dirty bits that might have come from getting
2510 * attached to a dirty page
2511 */
2513 }
2525 nr++;
2526 }
2531 }
2534 {
2536 }
2539 {
2543 }
2547 {
2557 journal_info;
2562 }
2567 /* this gets a little ugly. If reiserfs_get_block returned an
2568 * error and left a transacstion running, we've got to close it,
2569 * and we've got to free handle if it was a persistent transaction.
2570 *
2571 * But, if we had nested into an existing transaction, we need
2572 * to just drop the ref count on the handle.
2573 *
2574 * If old_ref == 0, the transaction is from reiserfs_get_block,
2575 * and it was a persistent trans. Otherwise, it was nested above.
2576 */
2587 }
2588 }
2589 }
2592 }
2595 {
2597 }
2601 {
2611 }
2614 /* generic_commit_write does this for us, but does not update the
2615 ** transaction tracking stuff when the size changes. So, we have
2616 ** to do the i_size updates here.
2617 */
2621 /* If the file have grown beyond the border where it
2622 can have a tail, unmark it as needing a tail
2623 packing */
2634 }
2637 /*
2638 * this will just nest into our transaction. It's important
2639 * to use mark_inode_dirty so the inode gets pushed around on the
2640 * dirty lists, and so that O_SYNC works as expected
2641 */
2649 }
2658 }
2660 out:
2663 journal_error:
2670 }
2673 }
2676 {
2680 else
2684 else
2688 else
2692 else
2696 else
2698 }
2699 }
2702 {
2706 else
2710 else
2714 else
2718 else
2720 }
2721 }
2723 /* decide if this buffer needs to stay around for data logging or ordered
2724 ** write purposes
2725 */
2727 {
2735 }
2736 /* the page is locked, and the only places that log a data buffer
2737 * also lock the page.
2738 */
2740 /*
2741 * very conservative, leave the buffer pinned if
2742 * anyone might need it.
2743 */
2746 }
2751 /* why is this safe?
2752 * reiserfs_setattr updates i_size in the on disk
2753 * stat data before allowing vmtruncate to be called.
2754 *
2755 * If buffer was put onto the ordered list for this
2756 * transaction, we know for sure either this transaction
2757 * or an older one already has updated i_size on disk,
2758 * and this ordered data won't be referenced in the file
2759 * if we crash.
2760 *
2761 * if the buffer was put onto the ordered list for an older
2762 * transaction, we need to leave it around
2763 */
2767 }
2768 free_jh:
2771 }
2775 }
2777 /* clm -- taken from fs/buffer.c:block_invalidate_page */
2779 {
2799 /*
2800 * is this block fully invalidated?
2801 */
2805 else
2807 }
2812 /*
2813 * We release buffers only if the entire page is being invalidated.
2814 * The get_block cached value has been unconditionally invalidated,
2815 * so real IO is not possible anymore.
2816 */
2819 /* maybe should BUG_ON(!ret); - neilb */
2820 }
2821 out:
2823 }
2826 {
2831 }
2833 }
2835 /*
2836 * Returns 1 if the page's buffers were dropped. The page is locked.
2837 *
2838 * Takes j_dirty_buffers_lock to protect the b_assoc_buffers list_heads
2839 * in the buffers at page_buffers(page).
2840 *
2841 * even in -o notail mode, we can't be sure an old mount without -o notail
2842 * didn't create files with tails.
2843 */
2845 {
2863 }
2864 }
2871 }
2873 /* We thank Mingming Cao for helping us understand in great detail what
2874 to do in this section of the code. */
2878 {
2885 }
2888 {
2894 /* version 2 items will be caught by the s_maxbytes check
2895 ** done for us in vmtruncate
2896 */
2901 }
2902 /* fill in hole pointers in the expanding truncate case. */
2908 /* we're changing at most 2 bitmaps, inode + super */
2913 }
2916 }
2919 /*
2920 * file size is changed, ctime and mtime are
2921 * to be updated
2922 */
2924 }
2925 }
2930 /* stat data of format v3.5 has 16 bit uid and gid */
2933 }
2949 /* (user+group)*(old+new) structure - we count quota info and , inode write (sb, inode) */
2950 error =
2952 jbegin_count);
2955 error =
2959 jbegin_count);
2961 }
2962 /* Update corresponding info in inode so that everything is in
2963 * one transaction */
2969 error =
2971 }
2972 }
2975 }
2980 }
2982 out:
2985 }
2999 };