| blob | 132ec4406ed00733947e9ba8c2fc770de08ec749 |
1 /*
2 * Copyright 2000 by Hans Reiser, licensing governed by reiserfs/README
3 */
5 #include <linux/time.h>
6 #include <linux/fs.h>
10 #include <linux/exportfs.h>
11 #include <linux/pagemap.h>
12 #include <linux/highmem.h>
13 #include <linux/slab.h>
14 #include <linux/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>
20 #include <linux/swap.h>
21 #include <linux/uio.h>
22 #include <linux/bio.h>
28 {
29 /*
30 * We need blocks for transaction + (user+group) quota
31 * update (possibly delete)
32 */
46 /*
47 * The = 0 happens when we abort creating a new inode
48 * for some reason like lack of space..
49 * also handles bad_inode case
50 */
65 /*
66 * Do quota update inside a transaction for journaled quotas.
67 * We must do that after delete_object so that quota updates
68 * go into the same transaction as stat data deletion
69 */
74 }
79 /*
80 * check return value from reiserfs_delete_object after
81 * ending the transaction
82 */
86 /*
87 * all items of file are deleted, so we can remove
88 * "save" link
89 * we can't do anything about an error here
90 */
92 out:
95 /* no object items are in the tree */
96 ;
97 }
99 /* note this must go after the journal_end to prevent deadlock */
106 no_delete:
109 }
113 {
121 }
123 /*
124 * take base of inode_key (it comes from inode always) (dirid, objectid)
125 * and version from an inode, set offset and type of key
126 */
129 {
133 length);
134 }
136 /* when key is 0, do not set version and short key */
141 {
146 }
151 /* set_ih_free_space (ih, 0); */
152 /*
153 * for directory items it is entry count, for directs and stat
154 * datas - 0xffff, for indirects - 0
155 */
157 }
159 /*
160 * FIXME: we might cache recently accessed indirect item
161 * Ugh. Not too eager for that....
162 * I cut the code until such time as I see a convincing argument (benchmark).
163 * I don't want a bloated inode struct..., and I don't like code complexity....
164 */
166 /*
167 * cutting the code is fine, since it really isn't in use yet and is easy
168 * to add back in. But, Vladimir has a really good idea here. Think
169 * about what happens for reading a file. For each page,
170 * The VFS layer calls reiserfs_readpage, who searches the tree to find
171 * an indirect item. This indirect item has X number of pointers, where
172 * X is a big number if we've done the block allocation right. But,
173 * we only use one or two of these pointers during each call to readpage,
174 * needlessly researching again later on.
175 *
176 * The size of the cache could be dynamic based on the size of the file.
177 *
178 * I'd also like to see us cache the location the stat data item, since
179 * we are needlessly researching for that frequently.
180 *
181 * --chris
182 */
184 /*
185 * If this page has a file tail in it, and
186 * it was read in by get_block_create_0, the page data is valid,
187 * but tail is still sitting in a direct item, and we can't write to
188 * it. So, look through this page, and check all the mapped buffers
189 * to make sure they have valid block numbers. Any that don't need
190 * to be unmapped, so that __block_write_begin will correctly call
191 * reiserfs_get_block to convert the tail into an unformatted node
192 */
194 {
204 }
207 }
208 }
210 /*
211 * reiserfs_get_block does not need to allocate a block only if it has been
212 * done already or non-hole position has been found in the indirect item
213 */
217 {
224 }
227 {
229 }
233 {
235 }
237 /*
238 * files which were created in the earlier version can not be longer,
239 * than 2 gb
240 */
242 {
243 /* it is new file. */
245 /* old file, but 'block' is inside of 2gb */
250 }
254 {
263 /* we cannot restart while nested */
266 }
273 }
275 }
277 /*
278 * it is called by get_block when create == 0. Returns block number
279 * for 'block'-th logical block of file. When it hits direct item it
280 * returns 0 (being called from bmap) or read direct item into piece
281 * of page (bh_result)
282 * Please improve the english/clarity in the comment above, as it is
283 * hard to understand.
284 */
287 {
292 b_blocknr_t blocknr;
300 /* prepare the key to look for the 'block'-th block of file */
312 /*
313 * We do not return -ENOENT if there is a hole but page is
314 * uptodate, because it means that there is some MMAPED data
315 * associated with it that is yet to be written to disk.
316 */
320 }
322 }
329 /*
330 * FIXME: here we could cache indirect item or part of it in
331 * the inode to avoid search_by_key in case of subsequent
332 * access to file
333 */
341 }
343 /*
344 * We do not return -ENOENT if there is a hole but
345 * page is uptodate, because it means that there is
346 * some MMAPED data associated with it that is
347 * yet to be written to disk.
348 */
352 }
358 }
359 /* requested data are in direct item(s) */
361 /*
362 * we are called by bmap. FIXME: we can not map block of file
363 * when it is stored in direct item(s)
364 */
369 }
371 /*
372 * if we've got a direct item, and the buffer or page was uptodate,
373 * we don't want to pull data off disk again. skip to the
374 * end, where we map the buffer and return
375 */
379 /*
380 * grab_tail_page can trigger calls to reiserfs_get_block on
381 * up to date pages without any buffers. If the page is up
382 * to date, we don't want read old data off disk. Set the up
383 * to date bit on the buffer instead and jump to the end
384 */
388 }
389 /* read file tail into part of page */
393 /*
394 * we only want to kmap if we are reading the tail into the page.
395 * this is not the common case, so we don't kmap until we are
396 * sure we need to. But, this means the item might move if
397 * kmap schedules
398 */
407 }
408 /*
409 * make sure we don't read more bytes than actually exist in
410 * the file. This can happen in odd cases where i_size isn't
411 * correct, and when direct item padding results in a few
412 * extra bytes at the end of the direct item
413 */
417 chars =
423 }
431 /*
432 * we done, if read direct item is not the last item of
433 * node FIXME: we could try to check right delimiting key
434 * to see whether direct item continues in the right
435 * neighbor or rely on i_size
436 */
440 /* update key to look for the next piece */
444 /* i/o error most likely */
453 finished:
459 /*
460 * this buffer has valid data, but isn't valid for io. mapping it to
461 * block #0 tells the rest of reiserfs it just has a tail in it
462 */
466 }
468 /*
469 * this is called to create file map. So, _get_block_create_0 will not
470 * read direct item
471 */
474 {
479 /* do not read the direct item */
483 }
485 /*
486 * special version of get_block that is only used by grab_tail_page right
487 * now. It is sent to __block_write_begin, and when you try to get a
488 * block past the end of the file (or a block from a hole) it returns
489 * -ENOENT instead of a valid buffer. __block_write_begin expects to
490 * be able to do i/o on the buffers returned, unless an error value
491 * is also returned.
492 *
493 * So, this allows __block_write_begin to be used for reading a single block
494 * in a page. Where it does not produce a valid page for holes, or past the
495 * end of the file. This turns out to be exactly what we need for reading
496 * tails for conversion.
497 *
498 * The point of the wrapper is forcing a certain value for create, even
499 * though the VFS layer is calling this function with create==1. If you
500 * don't want to send create == GET_BLOCK_NO_HOLE to reiserfs_get_block,
501 * don't use this function.
502 */
506 {
508 }
510 /*
511 * This is special helper for reiserfs_get_block in case we are executing
512 * direct_IO request.
513 */
515 sector_t iblock,
518 {
523 /*
524 * We set the b_size before reiserfs_get_block call since it is
525 * referenced in convert_tail_for_hole() that may be called from
526 * reiserfs_get_block()
527 */
535 /* don't allow direct io onto tail pages */
537 /*
538 * make sure future calls to the direct io funcs for this
539 * offset in the file fail by unmapping the buffer
540 */
543 }
545 /*
546 * Possible unpacked tail. Flush the data before pages have
547 * disappeared
548 */
561 }
562 out:
564 }
566 /*
567 * helper function for when reiserfs_get_block is called for a hole
568 * but the file tail is still in a direct item
569 * bh_result is the buffer head for the hole
570 * tail_offset is the offset of the start of the tail in the file
571 *
572 * This calls prepare_write, which will start a new transaction
573 * you should not be in a transaction, or have any paths held when you
574 * call this.
575 */
578 loff_t tail_offset)
579 {
590 /* always try to read until the end of the block */
595 /*
596 * hole_page can be zero in case of direct_io, we are sure
597 * that we cannot get here if we write with O_DIRECT into tail page
598 */
604 }
607 }
609 /*
610 * we don't have to make sure the conversion did not happen while
611 * we were locking the page because anyone that could convert
612 * must first take i_mutex.
613 *
614 * We must fix the tail page for writing because it might have buffers
615 * that are mapped, but have a block number of 0. This indicates tail
616 * data that has been read directly into the page, and
617 * __block_write_begin won't trigger a get_block in this case.
618 */
625 /* tail conversion might change the data in the page */
630 unlock:
634 }
635 out:
637 }
640 sector_t block,
644 {
647 #ifdef REISERFS_PREALLOCATE
651 }
652 #endif
654 block);
655 }
659 {
661 /* b_blocknr_t is (unsigned) 32 bit int*/
672 /*
673 * space reserved in transaction batch:
674 * . 3 balancings in direct->indirect conversion
675 * . 1 block involved into reiserfs_update_sd()
676 * XXX in practically impossible worst case direct2indirect()
677 * can incur (much) more than 3 balancings.
678 * quota update for user, group
679 */
685 loff_t new_offset =
694 }
696 /*
697 * if !create, we aren't changing the FS, so we don't need to
698 * log anything, so we don't need to start a transaction
699 */
702 /* find number of block-th logical block of the file */
707 }
709 /*
710 * if we're already in a transaction, make sure to close
711 * any new transactions we start in this func
712 */
717 /*
718 * If file is of such a size, that it might have a tail and
719 * tails are enabled we should mark it as possibly needing
720 * tail packing on close
721 */
728 /* set the key of the first byte in the 'block'-th block of file */
731 start_trans:
736 }
738 }
739 research:
745 }
757 /* we have to allocate block for the unformatted node */
761 }
763 repeat =
767 /*
768 * restart the transaction to give the journal a chance to free
769 * some blocks. releases the path, so we have to go back to
770 * research if we succeed on the second try
771 */
777 repeat =
783 }
786 else
789 }
794 }
795 }
798 b_blocknr_t unfm_ptr;
799 /*
800 * 'block'-th block is in the file already (there is
801 * corresponding cell in some indirect item). But it may be
802 * zero unformatted node pointer (hole)
803 */
806 /* use allocated block to plug the hole */
811 bh);
813 }
822 }
831 /*
832 * the item was found, so new blocks were not added to the file
833 * there is no need to make sure the inode is updated with this
834 * transaction
835 */
837 }
842 }
844 /*
845 * desired position is not found or is in the direct item. We have
846 * to append file with holes up to 'block'-th block converting
847 * direct items to indirect one if necessary
848 */
855 /* indirect item has to be inserted */
860 /*
861 * we are going to add 'block'-th block to the file.
862 * Use allocated block for that
863 */
870 }
875 retval =
881 /*
882 * retval == -ENOSPC, -EDQUOT or -EIO
883 * or -EEXIST
884 */
886 }
888 /* direct item has to be converted */
889 loff_t tail_offset;
891 tail_offset =
895 /*
896 * direct item we just found fits into block we have
897 * to map. Convert it into unformatted node: use
898 * bh_result for the conversion
899 */
906 /*
907 * we have to pad file tail stored in direct
908 * item(s) up to block size and convert it
909 * to unformatted node. FIXME: this should
910 * also get into page cache
911 */
914 /*
915 * ugly, but we can only end the transaction if
916 * we aren't nested
917 */
920 retval =
921 reiserfs_end_persistent_transaction
926 }
928 retval =
930 tail_offset);
935 "convert tail failed "
938 retval);
940 /*
941 * the bitmap, the super,
942 * and the stat data == 3
943 */
948 inode,
949 allocated_block_nr,
951 }
953 }
955 }
956 retval =
958 tail_offset);
964 }
965 /*
966 * it is important the set_buffer_uptodate is done
967 * after the direct2indirect. The buffer might
968 * contain valid data newer than the data on disk
969 * (read by readpage, changed, and then sent here by
970 * writepage). direct2indirect needs to know if unbh
971 * was already up to date, so it can decide if the
972 * data in unbh needs to be replaced with data from
973 * the disk
974 */
977 /*
978 * unbh->b_page == NULL in case of DIRECT_IO request,
979 * this means buffer will disappear shortly, so it
980 * should not be added to
981 */
983 /*
984 * we've converted the tail, so we must
985 * flush unbh before the transaction commits
986 */
989 /*
990 * mark it dirty now to prevent commit_write
991 * from adding this buffer to the inode's
992 * dirty buffer list
993 */
994 /*
995 * AKPM: changed __mark_buffer_dirty to
996 * mark_buffer_dirty(). It's still atomic,
997 * but it sets the page dirty too, which makes
998 * it eligible for writeback at any time by the
999 * VM (which was also the case with
1000 * __mark_buffer_dirty())
1001 */
1003 }
1005 /*
1006 * append indirect item with holes if needed, when
1007 * appending pointer to 'block'-th block use block,
1008 * which is already allocated
1009 */
1011 /*
1012 * We use this in case we need to allocate
1013 * only one block which is a fastpath
1014 */
1017 __u64 max_to_insert =
1019 UNFM_P_SIZE;
1020 __u64 blocks_needed;
1024 /*
1025 * indirect item has to be appended,
1026 * set up key of that position
1027 * (key type is unimportant)
1028 */
1038 blocks_needed =
1042 s_blocksize_bits);
1053 }
1054 }
1056 /*
1057 * we are going to add target block to
1058 * the file. Use allocated block for that
1059 */
1067 /* paste hole to the indirect item */
1068 /*
1069 * If kmalloc failed, max_to_insert becomes
1070 * zero and it means we only have space for
1071 * one block
1072 */
1073 blocks_needed =
1075 }
1076 retval =
1079 UNFM_P_SIZE *
1080 blocks_needed);
1089 }
1091 /*
1092 * We need to mark new file size in case
1093 * this function will be interrupted/aborted
1094 * later on. And we may do this only for
1095 * holes.
1096 */
1099 }
1100 }
1105 /*
1106 * this loop could log more blocks than we had originally
1107 * asked for. So, we have to allow the transaction to end
1108 * if it is too big or too full. Update the inode so things
1109 * are consistent if we crash before the function returns
1110 * release the path so that anybody waiting on the path before
1111 * ending their transaction will be able to continue.
1112 */
1117 }
1118 /*
1119 * inserting indirect pointers for a hole can take a
1120 * long time. reschedule if needed and also release the write
1121 * lock for others.
1122 */
1129 }
1139 }
1148 failure:
1156 }
1161 }
1163 static int
1166 {
1168 }
1170 /*
1171 * Compute real number of used bytes by file
1172 * Following three functions can go away when we'll have enough space in
1173 * stat item
1174 */
1176 {
1183 /*
1184 * End of file is also in full block with indirect reference, so round
1185 * up to the next block.
1186 *
1187 * there is just no way to know if the tail is actually packed
1188 * on the file, so we have to assume it isn't. When we pack the
1189 * tail, we add 4 bytes to pretend there really is an unformatted
1190 * node pointer
1191 */
1192 bytes =
1195 sd_size;
1197 }
1201 {
1205 }
1208 }
1210 /* Compute number of blocks used by file in ReiserFS counting */
1212 {
1216 /* keeps fsck and non-quota versions of reiserfs happy */
1219 }
1221 /*
1222 * files from before the quota patch might i_blocks such that
1223 * bytes < real_space. Deal with that here to prevent it from
1224 * going negative.
1225 */
1229 }
1231 /*
1232 * BAD: new directories have stat data of new type and all other items
1233 * of old type. Version stored in the inode says about body items, so
1234 * in update_stat_data we can not rely on inode, but have to check
1235 * item version directly
1236 */
1238 /* called by read_locked_inode */
1240 {
1243 __u32 rdev;
1282 /*
1283 * there was a bug in <=3.5.23 when i_blocks could take
1284 * negative values. Starting from 3.5.17 this value could
1285 * even be stored in stat data. For such files we set
1286 * i_blocks based on file size. Just 2 notes: this can be
1287 * wrong for sparse files. On-disk value will be only
1288 * updated if file's inode will ever change
1289 */
1292 }
1298 /*
1299 * an early bug in the quota code can give us an odd
1300 * number for the block count. This is incorrect, fix it here.
1301 */
1304 }
1307 SD_V1_SIZE));
1308 /*
1309 * nopack is initially zero for v1 objects. For v2 objects,
1310 * nopack is initialised from sd_attrs
1311 */
1314 /*
1315 * new stat data found, but object may have old items
1316 * (directories and symlinks)
1317 */
1336 else
1341 else
1347 SD_V2_SIZE));
1348 /*
1349 * read persistent inode attributes from sd and initialise
1350 * generic inode flags from them
1351 */
1354 }
1372 }
1373 }
1375 /* update new stat data with inode fields */
1377 {
1391 else
1394 }
1396 /* used to copy inode's fields to old stat data */
1398 {
1412 else
1415 /* Sigh. i_first_direct_byte is back */
1418 }
1420 /*
1421 * NOTE, you must prepare the buffer head before sending it here,
1422 * and then log it after the call
1423 */
1425 loff_t size)
1426 {
1437 /* path points to old stat data */
1442 }
1445 }
1449 {
1459 /* key type is unimportant */
1464 /* look for the object's stat data */
1468 "i/o failure occurred trying to "
1471 }
1476 /*reiserfs_warning (inode->i_sb, "vs-13050: reiserfs_update_sd: i_nlink == 0, stat data not found"); */
1478 }
1480 "stat data of object %k (nlink == %d) "
1483 pos);
1486 }
1488 /*
1489 * sigh, prepare_for_journal might schedule. When it
1490 * schedules the FS might change. We have to detect that,
1491 * and loop back to the search if the stat data item has moved
1492 */
1499 /* Stat_data item has been moved after scheduling. */
1504 }
1506 }
1511 }
1513 /*
1514 * reiserfs_read_locked_inode is called to read the inode off disk, and it
1515 * does a make_bad_inode when things go wrong. But, we need to make sure
1516 * and clear the key in the private portion of the inode, otherwise a
1517 * corresponding iput might try to delete whatever object the inode last
1518 * represented.
1519 */
1521 {
1524 }
1526 /*
1527 * initially this function was derived from minix or ext2's analog and
1528 * evolved as the prototype did
1529 */
1531 {
1536 }
1538 /*
1539 * looks for stat data in the tree, and fills up the fields of in-core
1540 * inode stat data fields
1541 */
1544 {
1552 /*
1553 * set version 1, version 2 could be used too, because stat data
1554 * key is the same in both versions
1555 */
1562 /* look for the object's stat data */
1566 "i/o failure occurred trying to find "
1570 }
1572 /* a stale NFS handle can trigger this without it being an error */
1578 }
1582 /*
1583 * It is possible that knfsd is trying to access inode of a file
1584 * that is being removed from the disk by some other thread. As we
1585 * update sd on unlink all that is required is to check for nlink
1586 * here. This bug was first found by Sizif when debugging
1587 * SquidNG/Butterfly, forgotten, and found again after Philippe
1588 * Gramoulle <philippe.gramoulle@mmania.com> reproduced it.
1590 * More logical fix would require changes in fs/inode.c:iput() to
1591 * remove inode from hash-table _after_ fs cleaned disk stuff up and
1592 * in iget() to return NULL if I_FREEING inode is found in
1593 * hash-table.
1594 */
1596 /*
1597 * Currently there is one place where it's ok to meet inode with
1598 * nlink==0: processing of open-unlinked and half-truncated files
1599 * during mount (fs/reiserfs/super.c:finish_unfinished()).
1600 */
1604 "dead inode read from disk %K. "
1608 }
1610 /* init inode should be relsing */
1613 /*
1614 * Stat data v1 doesn't support ACLs.
1615 */
1618 }
1620 /*
1621 * reiserfs_find_actor() - "find actor" reiserfs supplies to iget5_locked().
1622 *
1623 * @inode: inode from hash table to check
1624 * @opaque: "cookie" passed to iget5_locked(). This is &reiserfs_iget_args.
1625 *
1626 * This function is called by iget5_locked() to distinguish reiserfs inodes
1627 * having the same inode numbers. Such inodes can only exist due to some
1628 * error condition. One of them should be bad. Inodes with identical
1629 * inode numbers (objectids) are distinguished by parent directory ids.
1630 *
1631 */
1633 {
1637 /* args is already in CPU order */
1640 }
1643 {
1661 }
1664 /* either due to i/o error or a stale NFS handle */
1667 }
1669 }
1674 {
1686 }
1690 }
1694 {
1695 /*
1696 * fhtype happens to reflect the number of u32s encoded.
1697 * due to a bug in earlier code, fhtype might indicate there
1698 * are more u32s then actually fitted.
1699 * so if fhtype seems to be more than len, reduce fhtype.
1700 * Valid types are:
1701 * 2 - objectid + dir_id - legacy support
1702 * 3 - objectid + dir_id + generation
1703 * 4 - objectid + dir_id + objectid and dirid of parent - legacy
1704 * 5 - objectid + dir_id + generation + objectid and dirid of parent
1705 * 6 - as above plus generation of directory
1706 * 6 does not fit in NFSv2 handles
1707 */
1714 }
1720 }
1724 {
1734 }
1738 {
1747 }
1760 }
1761 }
1763 }
1765 /*
1766 * looks for stat data, then copies fields to it, marks the buffer
1767 * containing stat data as dirty
1768 */
1769 /*
1770 * reiserfs inodes are never really dirty, since the dirty inode call
1771 * always logs them. This call allows the VFS inode marking routines
1772 * to properly mark inodes for datasync and such, but only actually
1773 * does something when called for a synchronous update.
1774 */
1776 {
1782 /*
1783 * memory pressure can sometimes initiate write_inode calls with
1784 * sync == 1,
1785 * these cases are just when the system needs ram, not when the
1786 * inode needs to reach disk for safety, and they can safely be
1787 * ignored because the altered inode has already been logged.
1788 */
1794 }
1796 }
1798 }
1800 /*
1801 * stat data of new object is inserted already, this inserts the item
1802 * containing "." and ".." entries
1803 */
1808 {
1821 /*
1822 * compose item head for new item. Directories consist of items of
1823 * old type (ITEM_VERSION_1). Do not set key (second arg is 0), it
1824 * is done by reiserfs_new_inode
1825 */
1842 }
1844 /* look for place in the tree for new item */
1850 }
1857 }
1859 /* insert item, that is empty directory item */
1861 }
1863 /*
1864 * stat data of object has been inserted, this inserts the item
1865 * containing the body of symlink
1866 */
1872 {
1887 /* look for place in the tree for new item */
1893 }
1900 }
1902 /* insert item, that is body of symlink */
1904 }
1906 /*
1907 * inserts the stat data into the tree, and then calls
1908 * reiserfs_new_directory (to insert ".", ".." item if new object is
1909 * directory) or reiserfs_new_symlink (to insert symlink body if new
1910 * object is symlink) or nothing (if new object is regular file)
1912 * NOTE! uid and gid must already be set in the inode. If we return
1913 * non-zero due to an error, we have to drop the quota previously allocated
1914 * for the fresh inode. This can only be done outside a transaction, so
1915 * if we return non-zero, we also end the transaction.
1916 *
1917 * @th: active transaction handle
1918 * @dir: parent directory for new inode
1919 * @mode: mode of new inode
1920 * @symname: symlink contents if inode is symlink
1921 * @isize: 0 for regular file, EMPTY_DIR_SIZE for dirs, strlen(symname) for
1922 * symlinks
1923 * @inode: inode to be filled
1924 * @security: optional security context to associate with this inode
1925 */
1928 /* 0 for regular, EMTRY_DIR_SIZE for dirs,
1929 strlen (symname) for symlinks) */
1933 {
1954 }
1956 /* item head of new item */
1962 }
1967 else
1980 }
1983 /*
1984 * not a perfect generation count, as object ids can be reused,
1985 * but this is as good as reiserfs can do right now.
1986 * note that the private part of inode isn't filled in yet,
1987 * we have to use the directory.
1988 */
1990 else
1991 #if defined( USE_INODE_GENERATION_COUNTER )
1994 #else
1996 #endif
1998 /* fill stat data */
2001 /* uid and gid must already be set by the caller for quota init */
2021 /* key to search for correct place for new stat data */
2026 /* find proper place for inserting of stat data */
2031 }
2036 }
2038 /* i_uid or i_gid is too big to be stored in stat data v3.5 */
2043 }
2047 }
2048 /*
2049 * store in in-core inode the key of stat data and version all
2050 * object items will have (directory items will have old offset
2051 * format, other new objects will consist of new items)
2052 */
2055 else
2059 else
2062 /* insert the stat data into the tree */
2063 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
2066 #endif
2067 retval =
2074 }
2075 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
2078 #endif
2080 /* insert item with "." and ".." */
2081 retval =
2083 }
2086 /* insert body of symlink */
2089 retval =
2091 i_size);
2092 }
2098 }
2109 }
2112 "ACLs aren't enabled in the fs, "
2128 }
2129 }
2136 out_bad_inode:
2137 /* Invalidate the object, nothing was inserted yet */
2140 /* Quota change must be inside a transaction for journaling */
2145 out_end_trans:
2147 /*
2148 * Drop can be outside and it needs more credits so it's better
2149 * to have it outside
2150 */
2157 out_inserted_sd:
2163 }
2165 /*
2166 * finds the tail page in the page cache,
2167 * reads the last block in.
2168 *
2169 * On success, page_result is set to a locked, pinned page, and bh_result
2170 * is set to an up to date buffer for the last block in the file. returns 0.
2171 *
2172 * tail conversion is not done, so bh_result might not be valid for writing
2173 * check buffer_mapped(bh_result) and bh_result->b_blocknr != 0 before
2174 * trying to write the block.
2175 *
2176 * on failure, nonzero is returned, page_result and bh_result are untouched.
2177 */
2181 {
2183 /*
2184 * we want the page with the last byte in the file,
2185 * not the page that will hold the next byte for appending
2186 */
2197 /*
2198 * we know that we are only called with inode->i_size > 0.
2199 * we also know that a file tail can never be as big as a block
2200 * If i_size % blocksize == 0, our file is currently block aligned
2201 * and it won't need converting or zeroing after a truncate.
2202 */
2205 }
2210 }
2211 /* start within the page of the last block in the file */
2215 reiserfs_get_block_create_0);
2224 }
2230 /*
2231 * note, this should never happen, prepare_write should be
2232 * taking care of this for us. If the buffer isn't up to
2233 * date, I've screwed up the code to find the buffer, or the
2234 * code to call prepare_write
2235 */
2240 }
2244 out:
2247 unlock:
2251 }
2253 /*
2254 * vfs version of truncate file. Must NOT be called with
2255 * a transaction already started.
2256 *
2257 * some code taken from block_truncate_page
2258 */
2260 {
2262 /* we want the offset for the first byte after the end of the file */
2276 /*
2277 * -ENOENT means we truncated past the end of the
2278 * file, and get_block_create_0 could not find a
2279 * block to read in, which is ok.
2280 */
2284 error);
2287 }
2288 }
2290 /*
2291 * so, if page != NULL, we have a buffer head for the offset at
2292 * the end of the file. if the bh is mapped, and bh->b_blocknr != 0,
2293 * then we have an unformatted node. Otherwise, we have a direct item,
2294 * and no zeroing is required on disk. We zero after the truncate,
2295 * because the truncate might pack the item anyway
2296 * (it will unmap bh if it packs).
2297 *
2298 * it is enough to reserve space in transaction for 2 balancings:
2299 * one for "save" link adding and another for the first
2300 * cut_from_item. 1 is for update_sd
2301 */
2308 /*
2309 * we are doing real truncate: if the system crashes
2310 * before the last transaction of truncating gets committed
2311 * - on reboot the file either appears truncated properly
2312 * or not truncated at all
2313 */
2320 /* check reiserfs_do_truncate after ending the transaction */
2324 }
2330 }
2334 /* if we are not on a block boundary */
2340 }
2341 }
2344 }
2349 out:
2353 }
2358 }
2363 {
2381 /*
2382 * catch places below that try to log something without
2383 * starting a trans
2384 */
2389 }
2392 start_over:
2396 research:
2401 }
2408 /* we've found an unformatted node */
2413 }
2415 /* crap, we are writing to a hole */
2418 }
2431 /* vs-3050 is gone, no need to drop the path */
2440 bh);
2442 }
2443 }
2451 }
2454 copy_size);
2460 /* are there still bytes left? */
2465 copy_size);
2467 }
2473 }
2476 out:
2483 }
2486 /* this is where we fill in holes in the file. */
2489 GET_BLOCK_CREATE | GET_BLOCK_NO_IMUX
2494 /* get_block failed to find a mapped unformatted node. */
2497 }
2498 }
2499 }
2503 /*
2504 * we've copied data from the page into the direct item, so the
2505 * buffer in the page is now clean, mark it to reflect that.
2506 */
2510 }
2512 }
2514 /*
2515 * mason@suse.com: updated in 2.5.54 to follow the same general io
2516 * start/recovery path as __block_write_full_page, along with special
2517 * code to handle reiserfs tails.
2518 */
2521 {
2526 sector_t last_block;
2536 /* no logging allowed when nonblocking or from PF_MEMALLOC */
2541 }
2543 /*
2544 * The page dirty bit is cleared before writepage is called, which
2545 * means we have to tell create_empty_buffers to make dirty buffers
2546 * The page really should be up to date at this point, so tossing
2547 * in the BH_Uptodate is just a sanity check.
2548 */
2552 }
2555 /*
2556 * last page in the file, zero out any contents past the
2557 * last byte in the file
2558 */
2563 /* no file contents in this page */
2567 }
2569 }
2573 /* first map all the buffers, logging any direct items we find */
2576 /*
2577 * This can happen when the block size is less than
2578 * the page size. The corresponding bytes in the page
2579 * were zero filled above
2580 */
2587 /*
2588 * not mapped yet, or it points to a direct item, search
2589 * the btree for the mapping info, and log any direct
2590 * items found
2591 */
2594 }
2595 }
2597 block++;
2600 /*
2601 * we start the transaction after map_block_for_writepage,
2602 * because it can create holes in the file (an unbounded operation).
2603 * starting it here, we can make a reliable estimate for how many
2604 * blocks we're going to log
2605 */
2613 }
2615 }
2616 /* now go through and lock any dirty buffers on the page */
2628 }
2629 /*
2630 * from this point on, we know the buffer is mapped to a
2631 * real block and not a direct item
2632 */
2639 }
2640 }
2645 }
2653 }
2658 /*
2659 * since any buffer might be the only dirty buffer on the page,
2660 * the first submit_bh can bring the page out of writeback.
2661 * be careful with the buffers.
2662 */
2667 nr++;
2668 }
2674 done:
2676 /*
2677 * if this page only had a direct item, it is very possible for
2678 * no io to be required without there being an error. Or,
2679 * someone else could have locked them and sent them down the
2680 * pipe without locking the page
2681 */
2687 }
2693 }
2696 fail:
2697 /*
2698 * catches various errors, we need to make sure any valid dirty blocks
2699 * get to the media. The page is currently locked and not marked for
2700 * writeback
2701 */
2710 /*
2711 * clear any dirty bits that might have come from
2712 * getting attached to a dirty page
2713 */
2715 }
2727 nr++;
2728 }
2733 }
2736 {
2738 }
2741 {
2745 }
2748 {
2751 }
2757 {
2760 pgoff_t index;
2768 pos ++;
2770 }
2783 journal_info;
2788 }
2792 /*
2793 * this gets a little ugly. If reiserfs_get_block returned an
2794 * error and left a transacstion running, we've got to close
2795 * it, and we've got to free handle if it was a persistent
2796 * transaction.
2797 *
2798 * But, if we had nested into an existing transaction, we need
2799 * to just drop the ref count on the handle.
2800 *
2801 * If old_ref == 0, the transaction is from reiserfs_get_block,
2802 * and it was a persistent trans. Otherwise, it was nested
2803 * above.
2804 */
2815 }
2816 }
2817 }
2821 /* Truncate allocated blocks */
2823 }
2825 }
2828 {
2842 journal_info;
2847 }
2852 /*
2853 * this gets a little ugly. If reiserfs_get_block returned an
2854 * error and left a transacstion running, we've got to close
2855 * it, and we've got to free handle if it was a persistent
2856 * transaction.
2857 *
2858 * But, if we had nested into an existing transaction, we need
2859 * to just drop the ref count on the handle.
2860 *
2861 * If old_ref == 0, the transaction is from reiserfs_get_block,
2862 * and it was a persistent trans. Otherwise, it was nested
2863 * above.
2864 */
2875 }
2876 }
2877 }
2880 }
2883 {
2885 }
2890 {
2899 pos ++;
2904 else
2913 }
2918 /*
2919 * generic_commit_write does this for us, but does not update the
2920 * transaction tracking stuff when the size changes. So, we have
2921 * to do the i_size updates here.
2922 */
2927 /*
2928 * If the file have grown beyond the border where it
2929 * can have a tail, unmark it as needing a tail
2930 * packing
2931 */
2944 /*
2945 * this will just nest into our transaction. It's important
2946 * to use mark_inode_dirty so the inode gets pushed around on
2947 * the dirty lists, and so that O_SYNC works as expected
2948 */
2955 }
2960 }
2966 }
2968 out:
2979 journal_error:
2986 }
2988 }
2992 {
3006 }
3009 /*
3010 * generic_commit_write does this for us, but does not update the
3011 * transaction tracking stuff when the size changes. So, we have
3012 * to do the i_size updates here.
3013 */
3016 /*
3017 * If the file have grown beyond the border where it
3018 * can have a tail, unmark it as needing a tail
3019 * packing
3020 */
3033 /*
3034 * this will just nest into our transaction. It's important
3035 * to use mark_inode_dirty so the inode gets pushed around
3036 * on the dirty lists, and so that O_SYNC works as expected
3037 */
3044 }
3051 }
3053 out:
3056 journal_error:
3061 }
3064 }
3067 {
3071 else
3075 else
3079 else
3083 else
3087 else
3089 }
3090 }
3092 /*
3093 * decide if this buffer needs to stay around for data logging or ordered
3094 * write purposes
3095 */
3097 {
3105 }
3106 /*
3107 * the page is locked, and the only places that log a data buffer
3108 * also lock the page.
3109 */
3111 /*
3112 * very conservative, leave the buffer pinned if
3113 * anyone might need it.
3114 */
3117 }
3122 /*
3123 * why is this safe?
3124 * reiserfs_setattr updates i_size in the on disk
3125 * stat data before allowing vmtruncate to be called.
3126 *
3127 * If buffer was put onto the ordered list for this
3128 * transaction, we know for sure either this transaction
3129 * or an older one already has updated i_size on disk,
3130 * and this ordered data won't be referenced in the file
3131 * if we crash.
3132 *
3133 * if the buffer was put onto the ordered list for an older
3134 * transaction, we need to leave it around
3135 */
3139 }
3140 free_jh:
3143 }
3147 }
3149 /* clm -- taken from fs/buffer.c:block_invalidate_page */
3152 {
3177 /*
3178 * is this block fully invalidated?
3179 */
3183 else
3185 }
3190 /*
3191 * We release buffers only if the entire page is being invalidated.
3192 * The get_block cached value has been unconditionally invalidated,
3193 * so real IO is not possible anymore.
3194 */
3197 /* maybe should BUG_ON(!ret); - neilb */
3198 }
3199 out:
3201 }
3204 {
3209 }
3211 }
3213 /*
3214 * Returns 1 if the page's buffers were dropped. The page is locked.
3215 *
3216 * Takes j_dirty_buffers_lock to protect the b_assoc_buffers list_heads
3217 * in the buffers at page_buffers(page).
3218 *
3219 * even in -o notail mode, we can't be sure an old mount without -o notail
3220 * didn't create files with tails.
3221 */
3223 {
3241 }
3242 }
3249 }
3251 /*
3252 * We thank Mingming Cao for helping us understand in great detail what
3253 * to do in this section of the code.
3254 */
3256 {
3260 ssize_t ret;
3263 reiserfs_get_blocks_direct_io);
3265 /*
3266 * In case of error extending write may have instantiated a few
3267 * blocks outside i_size. Trim these off again.
3268 */
3276 }
3277 }
3280 }
3283 {
3292 /* must be turned off for recursive notify_change calls */
3299 }
3302 /*
3303 * version 2 items will be caught by the s_maxbytes check
3304 * done for us in vmtruncate
3305 */
3311 }
3315 /* fill in hole pointers in the expanding truncate case. */
3321 /* we're changing at most 2 bitmaps, inode + super */
3326 }
3329 }
3333 }
3334 /*
3335 * file size is changed, ctime and mtime are
3336 * to be updated
3337 */
3339 }
3340 }
3346 /* stat data of format v3.5 has 16 bit uid and gid */
3349 }
3365 /*
3366 * (user+group)*(old+new) structure - we count quota
3367 * info and , inode write (sb, inode)
3368 */
3380 }
3382 /*
3383 * Update corresponding info in inode so that everything
3384 * is in one transaction
3385 */
3395 }
3401 /*
3402 * Could race against reiserfs_file_release
3403 * if called from NFS, so take tailpack mutex.
3404 */
3409 }
3410 }
3415 }
3420 }
3422 out:
3424 }
3437 };