| blob | 9087b10209e634c4f694aceaad2e14bde7e01ba2 |
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>
20 #include <linux/swap.h>
28 {
29 /* 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 */
53 /* Do quota update inside a transaction for journaled quotas. We must do that
54 * after delete_object so that quota updates go into the same transaction as
55 * stat data deletion */
62 /* check return value from reiserfs_delete_object after
63 * ending the transaction
64 */
68 /* all items of file are deleted, so we can remove "save" link */
70 * about an error here */
72 /* no object items are in the tree */
73 ;
74 }
75 out:
79 }
83 {
91 }
93 /* take base of inode_key (it comes from inode always) (dirid, objectid) and version from an inode, set
94 offset and type of key */
97 {
101 length);
102 }
104 //
105 // when key is 0, do not set version and short key
106 //
111 {
116 }
121 /* set_ih_free_space (ih, 0); */
122 // for directory items it is entry count, for directs and stat
123 // datas - 0xffff, for indirects - 0
125 }
127 //
128 // FIXME: we might cache recently accessed indirect item
130 // Ugh. Not too eager for that....
131 // I cut the code until such time as I see a convincing argument (benchmark).
132 // I don't want a bloated inode struct..., and I don't like code complexity....
134 /* cutting the code is fine, since it really isn't in use yet and is easy
135 ** to add back in. But, Vladimir has a really good idea here. Think
136 ** about what happens for reading a file. For each page,
137 ** The VFS layer calls reiserfs_readpage, who searches the tree to find
138 ** an indirect item. This indirect item has X number of pointers, where
139 ** X is a big number if we've done the block allocation right. But,
140 ** we only use one or two of these pointers during each call to readpage,
141 ** needlessly researching again later on.
142 **
143 ** The size of the cache could be dynamic based on the size of the file.
144 **
145 ** I'd also like to see us cache the location the stat data item, since
146 ** we are needlessly researching for that frequently.
147 **
148 ** --chris
149 */
151 /* If this page has a file tail in it, and
152 ** it was read in by get_block_create_0, the page data is valid,
153 ** but tail is still sitting in a direct item, and we can't write to
154 ** it. So, look through this page, and check all the mapped buffers
155 ** to make sure they have valid block numbers. Any that don't need
156 ** to be unmapped, so that block_prepare_write will correctly call
157 ** reiserfs_get_block to convert the tail into an unformatted node
158 */
160 {
170 }
173 }
174 }
176 /* reiserfs_get_block does not need to allocate a block only if it has been
177 done already or non-hole position has been found in the indirect item */
181 {
188 }
191 {
193 }
197 {
199 }
201 //
202 // files which were created in the earlier version can not be longer,
203 // than 2 gb
204 //
206 {
212 }
216 {
226 /* we cannot restart while nested */
229 }
236 }
238 }
240 // it is called by get_block when create == 0. Returns block number
241 // for 'block'-th logical block of file. When it hits direct item it
242 // returns 0 (being called from bmap) or read direct item into piece
243 // of page (bh_result)
245 // Please improve the english/clarity in the comment above, as it is
246 // hard to understand.
250 {
255 b_blocknr_t blocknr;
263 // prepare the key to look for the 'block'-th block of file
275 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
276 // That there is some MMAPED data associated with it that is yet to be written to disk.
280 }
282 }
283 //
289 /* FIXME: here we could cache indirect item or part of it in
290 the inode to avoid search_by_key in case of subsequent
291 access to file */
299 }
301 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
302 // That there is some MMAPED data associated with it that is yet to be written to disk.
306 }
312 }
313 // requested data are in direct item(s)
315 // we are called by bmap. FIXME: we can not map block of file
316 // when it is stored in direct item(s)
321 }
323 /* if we've got a direct item, and the buffer or page was uptodate,
324 ** we don't want to pull data off disk again. skip to the
325 ** end, where we map the buffer and return
326 */
330 /*
331 ** grab_tail_page can trigger calls to reiserfs_get_block on up to date
332 ** pages without any buffers. If the page is up to date, we don't want
333 ** read old data off disk. Set the up to date bit on the buffer instead
334 ** and jump to the end
335 */
339 }
340 // read file tail into part of page
344 /* we only want to kmap if we are reading the tail into the page.
345 ** this is not the common case, so we don't kmap until we are
346 ** sure we need to. But, this means the item might move if
347 ** kmap schedules
348 */
357 }
358 /* make sure we don't read more bytes than actually exist in
359 ** the file. This can happen in odd cases where i_size isn't
360 ** correct, and when direct item padding results in a few
361 ** extra bytes at the end of the direct item
362 */
366 chars =
372 }
381 // we done, if read direct item is not the last item of
382 // node FIXME: we could try to check right delimiting key
383 // to see whether direct item continues in the right
384 // neighbor or rely on i_size
387 // update key to look for the next piece
391 // i/o error most likely
400 finished:
406 /* this buffer has valid data, but isn't valid for io. mapping it to
407 * block #0 tells the rest of reiserfs it just has a tail in it
408 */
412 }
414 // this is called to create file map. So, _get_block_create_0 will not
415 // read direct item
418 {
423 /* do not read the direct item */
427 }
429 /* special version of get_block that is only used by grab_tail_page right
430 ** now. It is sent to block_prepare_write, and when you try to get a
431 ** block past the end of the file (or a block from a hole) it returns
432 ** -ENOENT instead of a valid buffer. block_prepare_write expects to
433 ** be able to do i/o on the buffers returned, unless an error value
434 ** is also returned.
435 **
436 ** So, this allows block_prepare_write to be used for reading a single block
437 ** in a page. Where it does not produce a valid page for holes, or past the
438 ** end of the file. This turns out to be exactly what we need for reading
439 ** tails for conversion.
440 **
441 ** The point of the wrapper is forcing a certain value for create, even
442 ** though the VFS layer is calling this function with create==1. If you
443 ** don't want to send create == GET_BLOCK_NO_HOLE to reiserfs_get_block,
444 ** don't use this function.
445 */
449 {
451 }
453 /* This is special helper for reiserfs_get_block in case we are executing
454 direct_IO request. */
456 sector_t iblock,
459 {
464 /* We set the b_size before reiserfs_get_block call since it is
465 referenced in convert_tail_for_hole() that may be called from
466 reiserfs_get_block() */
474 /* don't allow direct io onto tail pages */
476 /* make sure future calls to the direct io funcs for this offset
477 ** in the file fail by unmapping the buffer
478 */
481 }
482 /* Possible unpacked tail. Flush the data before pages have
483 disappeared */
496 }
497 out:
499 }
501 /*
502 ** helper function for when reiserfs_get_block is called for a hole
503 ** but the file tail is still in a direct item
504 ** bh_result is the buffer head for the hole
505 ** tail_offset is the offset of the start of the tail in the file
506 **
507 ** This calls prepare_write, which will start a new transaction
508 ** you should not be in a transaction, or have any paths held when you
509 ** call this.
510 */
513 loff_t tail_offset)
514 {
525 /* always try to read until the end of the block */
530 /* hole_page can be zero in case of direct_io, we are sure
531 that we cannot get here if we write with O_DIRECT into
532 tail page */
538 }
541 }
543 /* we don't have to make sure the conversion did not happen while
544 ** we were locking the page because anyone that could convert
545 ** must first take i_mutex.
546 **
547 ** We must fix the tail page for writing because it might have buffers
548 ** that are mapped, but have a block number of 0. This indicates tail
549 ** data that has been read directly into the page, and block_prepare_write
550 ** won't trigger a get_block in this case.
551 */
557 /* tail conversion might change the data in the page */
562 unlock:
566 }
567 out:
569 }
572 sector_t block,
576 {
579 #ifdef REISERFS_PREALLOCATE
583 }
584 #endif
586 block);
587 }
591 {
604 /* space reserved in transaction batch:
605 . 3 balancings in direct->indirect conversion
606 . 1 block involved into reiserfs_update_sd()
607 XXX in practically impossible worst case direct2indirect()
608 can incur (much) more than 3 balancings.
609 quota update for user, group */
615 loff_t new_offset =
624 }
626 /* if !create, we aren't changing the FS, so we don't need to
627 ** log anything, so we don't need to start a transaction
628 */
631 /* find number of block-th logical block of the file */
636 }
637 /*
638 * if we're already in a transaction, make sure to close
639 * any new transactions we start in this func
640 */
645 /* If file is of such a size, that it might have a tail and tails are enabled
646 ** we should mark it as possibly needing tail packing on close
647 */
654 /* set the key of the first byte in the 'block'-th block of file */
657 start_trans:
662 }
664 }
665 research:
671 }
683 /* we have to allocate block for the unformatted node */
687 }
689 repeat =
694 /* restart the transaction to give the journal a chance to free
695 ** some blocks. releases the path, so we have to go back to
696 ** research if we succeed on the second try
697 */
702 repeat =
708 }
711 else
714 }
719 }
720 }
723 b_blocknr_t unfm_ptr;
724 /* 'block'-th block is in the file already (there is
725 corresponding cell in some indirect item). But it may be
726 zero unformatted node pointer (hole) */
729 /* use allocated block to plug the hole */
734 bh);
736 }
745 }
754 /* the item was found, so new blocks were not added to the file
755 ** there is no need to make sure the inode is updated with this
756 ** transaction
757 */
759 }
764 }
766 /* desired position is not found or is in the direct item. We have
767 to append file with holes up to 'block'-th block converting
768 direct items to indirect one if necessary */
775 /* indirect item has to be inserted */
781 /* we are going to add 'block'-th block to the file. Use
782 allocated block for that */
788 }
793 retval =
800 }
801 //mark_tail_converted (inode);
803 /* direct item has to be converted */
804 loff_t tail_offset;
806 tail_offset =
810 /* direct item we just found fits into block we have
811 to map. Convert it into unformatted node: use
812 bh_result for the conversion */
818 /* we have to padd file tail stored in direct item(s)
819 up to block size and convert it to unformatted
820 node. FIXME: this should also get into page cache */
823 /*
824 * ugly, but we can only end the transaction if
825 * we aren't nested
826 */
829 retval =
830 reiserfs_end_persistent_transaction
835 }
837 retval =
839 tail_offset);
844 "convert tail failed "
847 retval);
849 /* the bitmap, the super, and the stat data == 3 */
854 inode,
855 allocated_block_nr,
857 }
859 }
861 }
862 retval =
864 tail_offset);
870 }
871 /* it is important the set_buffer_uptodate is done after
872 ** the direct2indirect. The buffer might contain valid
873 ** data newer than the data on disk (read by readpage, changed,
874 ** and then sent here by writepage). direct2indirect needs
875 ** to know if unbh was already up to date, so it can decide
876 ** if the data in unbh needs to be replaced with data from
877 ** the disk
878 */
881 /* unbh->b_page == NULL in case of DIRECT_IO request, this means
882 buffer will disappear shortly, so it should not be added to
883 */
885 /* we've converted the tail, so we must
886 ** flush unbh before the transaction commits
887 */
890 /* mark it dirty now to prevent commit_write from adding
891 ** this buffer to the inode's dirty buffer list
892 */
893 /*
894 * AKPM: changed __mark_buffer_dirty to mark_buffer_dirty().
895 * It's still atomic, but it sets the page dirty too,
896 * which makes it eligible for writeback at any time by the
897 * VM (which was also the case with __mark_buffer_dirty())
898 */
900 }
902 /* append indirect item with holes if needed, when appending
903 pointer to 'block'-th block use block, which is already
904 allocated */
907 // one block which is a fastpath
909 __u64 max_to_insert =
911 UNFM_P_SIZE;
912 __u64 blocks_needed;
916 /* indirect item has to be appended, set up key of that position */
922 //pos_in_item * inode->i_sb->s_blocksize,
927 blocks_needed =
931 s_blocksize_bits);
941 }
942 }
944 /* we are going to add target block to the file. Use allocated
945 block for that */
953 /* paste hole to the indirect item */
954 /* If kmalloc failed, max_to_insert becomes zero and it means we
955 only have space for one block */
956 blocks_needed =
958 }
959 retval =
962 UNFM_P_SIZE *
963 blocks_needed);
972 }
974 /* We need to mark new file size in case this function will be
975 interrupted/aborted later on. And we may do this only for
976 holes. */
979 }
980 }
985 /* this loop could log more blocks than we had originally asked
986 ** for. So, we have to allow the transaction to end if it is
987 ** too big or too full. Update the inode so things are
988 ** consistent if we crash before the function returns
989 **
990 ** release the path so that anybody waiting on the path before
991 ** ending their transaction will be able to continue.
992 */
997 }
998 /*
999 * inserting indirect pointers for a hole can take a
1000 * long time. reschedule if needed and also release the write
1001 * lock for others.
1002 */
1007 }
1013 }
1023 }
1032 failure:
1040 }
1045 }
1047 static int
1050 {
1052 }
1054 /* Compute real number of used bytes by file
1055 * Following three functions can go away when we'll have enough space in stat item
1056 */
1058 {
1065 /* End of file is also in full block with indirect reference, so round
1066 ** up to the next block.
1067 **
1068 ** there is just no way to know if the tail is actually packed
1069 ** on the file, so we have to assume it isn't. When we pack the
1070 ** tail, we add 4 bytes to pretend there really is an unformatted
1071 ** node pointer
1072 */
1073 bytes =
1076 sd_size;
1078 }
1082 {
1086 }
1089 }
1091 /* Compute number of blocks used by file in ReiserFS counting */
1093 {
1097 /* keeps fsck and non-quota versions of reiserfs happy */
1100 }
1102 /* files from before the quota patch might i_blocks such that
1103 ** bytes < real_space. Deal with that here to prevent it from
1104 ** going negative.
1105 */
1109 }
1111 //
1112 // BAD: new directories have stat data of new type and all other items
1113 // of old type. Version stored in the inode says about body items, so
1114 // in update_stat_data we can not rely on inode, but have to check
1115 // item version directly
1116 //
1118 // called by read_locked_inode
1120 {
1123 __u32 rdev;
1124 //int version = ITEM_VERSION_1;
1164 // there was a bug in <=3.5.23 when i_blocks could take negative
1165 // values. Starting from 3.5.17 this value could even be stored in
1166 // stat data. For such files we set i_blocks based on file
1167 // size. Just 2 notes: this can be wrong for sparce files. On-disk value will be
1168 // only updated if file's inode will ever change
1170 }
1175 /* an early bug in the quota code can give us an odd number for the
1176 ** block count. This is incorrect, fix it here.
1177 */
1180 }
1183 SD_V1_SIZE));
1184 /* nopack is initially zero for v1 objects. For v2 objects,
1185 nopack is initialised from sd_attrs */
1188 // new stat data found, but object may have old items
1189 // (directories and symlinks)
1208 else
1213 else
1219 SD_V2_SIZE));
1220 /* read persistent inode attributes from sd and initalise
1221 generic inode flags from them */
1224 }
1241 }
1242 }
1244 // update new stat data with inode fields
1246 {
1248 __u16 flags;
1261 else
1266 }
1268 // used to copy inode's fields to old stat data
1270 {
1284 else
1287 // Sigh. i_first_direct_byte is back
1290 }
1292 /* NOTE, you must prepare the buffer head before sending it here,
1293 ** and then log it after the call
1294 */
1296 loff_t size)
1297 {
1309 // path points to old stat data
1313 }
1316 }
1320 {
1334 /* look for the object's stat data */
1338 "i/o failure occurred trying to "
1341 }
1346 /*reiserfs_warning (inode->i_sb, "vs-13050: reiserfs_update_sd: i_nlink == 0, stat data not found"); */
1348 }
1350 "stat data of object %k (nlink == %d) "
1353 pos);
1356 }
1358 /* sigh, prepare_for_journal might schedule. When it schedules the
1359 ** FS might change. We have to detect that, and loop back to the
1360 ** search if the stat data item has moved
1361 */
1371 }
1373 }
1378 }
1380 /* reiserfs_read_locked_inode is called to read the inode off disk, and it
1381 ** does a make_bad_inode when things go wrong. But, we need to make sure
1382 ** and clear the key in the private portion of the inode, otherwise a
1383 ** corresponding iput might try to delete whatever object the inode last
1384 ** represented.
1385 */
1387 {
1390 }
1392 //
1393 // initially this function was derived from minix or ext2's analog and
1394 // evolved as the prototype did
1395 //
1398 {
1403 }
1405 /* looks for stat data in the tree, and fills up the fields of in-core
1406 inode stat data fields */
1409 {
1417 /* set version 1, version 2 could be used too, because stat data
1418 key is the same in both versions */
1425 /* look for the object's stat data */
1429 "i/o failure occurred trying to find "
1433 }
1435 /* a stale NFS handle can trigger this without it being an error */
1440 }
1444 /* It is possible that knfsd is trying to access inode of a file
1445 that is being removed from the disk by some other thread. As we
1446 update sd on unlink all that is required is to check for nlink
1447 here. This bug was first found by Sizif when debugging
1448 SquidNG/Butterfly, forgotten, and found again after Philippe
1449 Gramoulle <philippe.gramoulle@mmania.com> reproduced it.
1451 More logical fix would require changes in fs/inode.c:iput() to
1452 remove inode from hash-table _after_ fs cleaned disk stuff up and
1453 in iget() to return NULL if I_FREEING inode is found in
1454 hash-table. */
1455 /* Currently there is one place where it's ok to meet inode with
1456 nlink==0: processing of open-unlinked and half-truncated files
1457 during mount (fs/reiserfs/super.c:finish_unfinished()). */
1461 "dead inode read from disk %K. "
1465 }
1469 }
1471 /**
1472 * reiserfs_find_actor() - "find actor" reiserfs supplies to iget5_locked().
1473 *
1474 * @inode: inode from hash table to check
1475 * @opaque: "cookie" passed to iget5_locked(). This is &reiserfs_iget_args.
1476 *
1477 * This function is called by iget5_locked() to distinguish reiserfs inodes
1478 * having the same inode numbers. Such inodes can only exist due to some
1479 * error condition. One of them should be bad. Inodes with identical
1480 * inode numbers (objectids) are distinguished by parent directory ids.
1481 *
1482 */
1484 {
1488 /* args is already in CPU order */
1491 }
1494 {
1509 }
1512 /* either due to i/o error or a stale NFS handle */
1515 }
1517 }
1522 {
1534 }
1538 }
1542 {
1543 /* fhtype happens to reflect the number of u32s encoded.
1544 * due to a bug in earlier code, fhtype might indicate there
1545 * are more u32s then actually fitted.
1546 * so if fhtype seems to be more than len, reduce fhtype.
1547 * Valid types are:
1548 * 2 - objectid + dir_id - legacy support
1549 * 3 - objectid + dir_id + generation
1550 * 4 - objectid + dir_id + objectid and dirid of parent - legacy
1551 * 5 - objectid + dir_id + generation + objectid and dirid of parent
1552 * 6 - as above plus generation of directory
1553 * 6 does not fit in NFSv2 handles
1554 */
1561 }
1565 }
1569 {
1577 }
1581 {
1592 /* no room for directory info? return what we've stored so far */
1604 }
1607 }
1609 /* looks for stat data, then copies fields to it, marks the buffer
1610 containing stat data as dirty */
1611 /* reiserfs inodes are never really dirty, since the dirty inode call
1612 ** always logs them. This call allows the VFS inode marking routines
1613 ** to properly mark inodes for datasync and such, but only actually
1614 ** does something when called for a synchronous update.
1615 */
1617 {
1623 /* memory pressure can sometimes initiate write_inode calls with sync == 1,
1624 ** these cases are just when the system needs ram, not when the
1625 ** inode needs to reach disk for safety, and they can safely be
1626 ** ignored because the altered inode has already been logged.
1627 */
1633 }
1635 }
1637 }
1639 /* stat data of new object is inserted already, this inserts the item
1640 containing "." and ".." entries */
1645 {
1658 /* compose item head for new item. Directories consist of items of
1659 old type (ITEM_VERSION_1). Do not set key (second arg is 0), it
1660 is done by reiserfs_new_inode */
1677 }
1679 /* look for place in the tree for new item */
1685 }
1692 }
1694 /* insert item, that is empty directory item */
1696 }
1698 /* stat data of object has been inserted, this inserts the item
1699 containing the body of symlink */
1700 static int reiserfs_new_symlink(struct reiserfs_transaction_handle *th, struct inode *inode, /* Inode of symlink */
1704 {
1719 /* look for place in the tree for new item */
1725 }
1732 }
1734 /* insert item, that is body of symlink */
1736 }
1738 /* inserts the stat data into the tree, and then calls
1739 reiserfs_new_directory (to insert ".", ".." item if new object is
1740 directory) or reiserfs_new_symlink (to insert symlink body if new
1741 object is symlink) or nothing (if new object is regular file)
1743 NOTE! uid and gid must already be set in the inode. If we return
1744 non-zero due to an error, we have to drop the quota previously allocated
1745 for the fresh inode. This can only be done outside a transaction, so
1746 if we return non-zero, we also end the transaction. */
1749 /* 0 for regular, EMTRY_DIR_SIZE for dirs,
1750 strlen (symname) for symlinks) */
1754 {
1769 }
1773 }
1777 /* item head of new item */
1783 }
1788 else
1797 }
1799 /* not a perfect generation count, as object ids can be reused, but
1800 ** this is as good as reiserfs can do right now.
1801 ** note that the private part of inode isn't filled in yet, we have
1802 ** to use the directory.
1803 */
1805 else
1806 #if defined( USE_INODE_GENERATION_COUNTER )
1809 #else
1811 #endif
1813 /* fill stat data */
1816 /* uid and gid must already be set by the caller for quota init */
1818 /* symlink cannot be immutable or append only, right? */
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 // store in in-core inode the key of stat data and version all
1869 // object items will have (directory items will have old offset
1870 // format, other new objects will consist of new items)
1873 else
1877 else
1880 /* insert the stat data into the tree */
1881 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1884 #endif
1885 retval =
1892 }
1893 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1896 #endif
1898 /* insert item with "." and ".." */
1899 retval =
1901 }
1904 /* insert body of symlink */
1907 retval =
1909 i_size);
1910 }
1916 }
1925 }
1928 "ACLs aren't enabled in the fs, "
1943 }
1944 }
1951 /* it looks like you can easily compress these two goto targets into
1952 * one. Keeping it like this doesn't actually hurt anything, and they
1953 * are place holders for what the quota code actually needs.
1954 */
1955 out_bad_inode:
1956 /* Invalidate the object, nothing was inserted yet */
1959 /* Quota change must be inside a transaction for journaling */
1962 out_end_trans:
1964 /* Drop can be outside and it needs more credits so it's better to have it outside */
1969 out_inserted_sd:
1975 }
1977 /*
1978 ** finds the tail page in the page cache,
1979 ** reads the last block in.
1980 **
1981 ** On success, page_result is set to a locked, pinned page, and bh_result
1982 ** is set to an up to date buffer for the last block in the file. returns 0.
1983 **
1984 ** tail conversion is not done, so bh_result might not be valid for writing
1985 ** check buffer_mapped(bh_result) and bh_result->b_blocknr != 0 before
1986 ** trying to write the block.
1987 **
1988 ** on failure, nonzero is returned, page_result and bh_result are untouched.
1989 */
1993 {
1995 /* we want the page with the last byte in the file,
1996 ** not the page that will hold the next byte for appending
1997 */
2008 /* we know that we are only called with inode->i_size > 0.
2009 ** we also know that a file tail can never be as big as a block
2010 ** If i_size % blocksize == 0, our file is currently block aligned
2011 ** and it won't need converting or zeroing after a truncate.
2012 */
2015 }
2020 }
2021 /* start within the page of the last block in the file */
2025 reiserfs_get_block_create_0);
2034 }
2040 /* note, this should never happen, prepare_write should
2041 ** be taking care of this for us. If the buffer isn't up to date,
2042 ** I've screwed up the code to find the buffer, or the code to
2043 ** call prepare_write
2044 */
2049 }
2053 out:
2056 unlock:
2060 }
2062 /*
2063 ** vfs version of truncate file. Must NOT be called with
2064 ** a transaction already started.
2065 **
2066 ** some code taken from block_truncate_page
2067 */
2069 {
2071 /* we want the offset for the first byte after the end of the file */
2086 // -ENOENT means we truncated past the end of the file,
2087 // and get_block_create_0 could not find a block to read in,
2088 // which is ok.
2092 error);
2095 }
2096 }
2098 /* so, if page != NULL, we have a buffer head for the offset at
2099 ** the end of the file. if the bh is mapped, and bh->b_blocknr != 0,
2100 ** then we have an unformatted node. Otherwise, we have a direct item,
2101 ** and no zeroing is required on disk. We zero after the truncate,
2102 ** because the truncate might pack the item anyway
2103 ** (it will unmap bh if it packs).
2104 */
2105 /* it is enough to reserve space in transaction for 2 balancings:
2106 one for "save" link adding and another for the first
2107 cut_from_item. 1 is for update_sd */
2114 /* we are doing real truncate: if the system crashes before the last
2115 transaction of truncating gets committed - on reboot the file
2116 either appears truncated properly or not truncated at all */
2119 error =
2124 /* check reiserfs_do_truncate after ending the transaction */
2128 }
2134 }
2138 /* if we are not on a block boundary */
2144 }
2145 }
2148 }
2153 out:
2157 }
2162 }
2167 {
2185 /* catch places below that try to log something without starting a trans */
2190 }
2193 start_over:
2197 research:
2202 }
2209 /* we've found an unformatted node */
2214 }
2216 /* crap, we are writing to a hole */
2219 }
2232 /* vs-3050 is gone, no need to drop the path */
2241 bh);
2243 }
2244 }
2252 }
2255 copy_size);
2261 /* are there still bytes left? */
2266 copy_size);
2268 }
2274 }
2277 out:
2284 }
2287 /* this is where we fill in holes in the file. */
2290 GET_BLOCK_CREATE | GET_BLOCK_NO_IMUX
2295 /* get_block failed to find a mapped unformatted node. */
2298 }
2299 }
2300 }
2304 /* we've copied data from the page into the direct item, so the
2305 * buffer in the page is now clean, mark it to reflect that.
2306 */
2310 }
2312 }
2314 /*
2315 * mason@suse.com: updated in 2.5.54 to follow the same general io
2316 * start/recovery path as __block_write_full_page, along with special
2317 * code to handle reiserfs tails.
2318 */
2321 {
2326 sector_t last_block;
2336 /* no logging allowed when nonblocking or from PF_MEMALLOC */
2341 }
2343 /* The page dirty bit is cleared before writepage is called, which
2344 * means we have to tell create_empty_buffers to make dirty buffers
2345 * The page really should be up to date at this point, so tossing
2346 * in the BH_Uptodate is just a sanity check.
2347 */
2351 }
2354 /* last page in the file, zero out any contents past the
2355 ** last byte in the file
2356 */
2361 /* no file contents in this page */
2365 }
2367 }
2371 /* first map all the buffers, logging any direct items we find */
2374 /*
2375 * This can happen when the block size is less than
2376 * the page size. The corresponding bytes in the page
2377 * were zero filled above
2378 */
2385 /* not mapped yet, or it points to a direct item, search
2386 * the btree for the mapping info, and log any direct
2387 * items found
2388 */
2391 }
2392 }
2394 block++;
2397 /*
2398 * we start the transaction after map_block_for_writepage,
2399 * because it can create holes in the file (an unbounded operation).
2400 * starting it here, we can make a reliable estimate for how many
2401 * blocks we're going to log
2402 */
2410 }
2412 }
2413 /* now go through and lock any dirty buffers on the page */
2425 }
2426 /* from this point on, we know the buffer is mapped to a
2427 * real block and not a direct item
2428 */
2435 }
2436 }
2441 }
2449 }
2454 /*
2455 * since any buffer might be the only dirty buffer on the page,
2456 * the first submit_bh can bring the page out of writeback.
2457 * be careful with the buffers.
2458 */
2463 nr++;
2464 }
2470 done:
2472 /*
2473 * if this page only had a direct item, it is very possible for
2474 * no io to be required without there being an error. Or,
2475 * someone else could have locked them and sent them down the
2476 * pipe without locking the page
2477 */
2483 }
2489 }
2492 fail:
2493 /* catches various errors, we need to make sure any valid dirty blocks
2494 * get to the media. The page is currently locked and not marked for
2495 * writeback
2496 */
2505 /*
2506 * clear any dirty bits that might have come from getting
2507 * attached to a dirty page
2508 */
2510 }
2522 nr++;
2523 }
2528 }
2531 {
2533 }
2536 {
2540 }
2543 {
2546 }
2552 {
2555 pgoff_t index;
2563 pos ++;
2565 }
2578 journal_info;
2583 }
2585 reiserfs_get_block);
2588 /* this gets a little ugly. If reiserfs_get_block returned an
2589 * error and left a transacstion running, we've got to close it,
2590 * and we've got to free handle if it was a persistent transaction.
2591 *
2592 * But, if we had nested into an existing transaction, we need
2593 * to just drop the ref count on the handle.
2594 *
2595 * If old_ref == 0, the transaction is from reiserfs_get_block,
2596 * and it was a persistent trans. Otherwise, it was nested above.
2597 */
2608 }
2609 }
2610 }
2614 /* Truncate allocated blocks */
2616 }
2618 }
2622 {
2635 journal_info;
2640 }
2645 /* this gets a little ugly. If reiserfs_get_block returned an
2646 * error and left a transacstion running, we've got to close it,
2647 * and we've got to free handle if it was a persistent transaction.
2648 *
2649 * But, if we had nested into an existing transaction, we need
2650 * to just drop the ref count on the handle.
2651 *
2652 * If old_ref == 0, the transaction is from reiserfs_get_block,
2653 * and it was a persistent trans. Otherwise, it was nested above.
2654 */
2665 }
2666 }
2667 }
2670 }
2673 {
2675 }
2680 {
2690 pos ++;
2695 else
2704 }
2709 /* generic_commit_write does this for us, but does not update the
2710 ** transaction tracking stuff when the size changes. So, we have
2711 ** to do the i_size updates here.
2712 */
2717 /* If the file have grown beyond the border where it
2718 can have a tail, unmark it as needing a tail
2719 packing */
2732 /*
2733 * this will just nest into our transaction. It's important
2734 * to use mark_inode_dirty so the inode gets pushed around on the
2735 * dirty lists, and so that O_SYNC works as expected
2736 */
2743 }
2748 }
2754 }
2756 out:
2767 journal_error:
2774 }
2776 }
2780 {
2793 }
2796 /* generic_commit_write does this for us, but does not update the
2797 ** transaction tracking stuff when the size changes. So, we have
2798 ** to do the i_size updates here.
2799 */
2802 /* If the file have grown beyond the border where it
2803 can have a tail, unmark it as needing a tail
2804 packing */
2817 /*
2818 * this will just nest into our transaction. It's important
2819 * to use mark_inode_dirty so the inode gets pushed around on the
2820 * dirty lists, and so that O_SYNC works as expected
2821 */
2828 }
2835 }
2837 out:
2840 journal_error:
2845 }
2848 }
2851 {
2855 else
2859 else
2863 else
2867 else
2871 else
2873 }
2874 }
2877 {
2881 else
2885 else
2889 else
2893 else
2895 }
2896 }
2898 /* decide if this buffer needs to stay around for data logging or ordered
2899 ** write purposes
2900 */
2902 {
2910 }
2911 /* the page is locked, and the only places that log a data buffer
2912 * also lock the page.
2913 */
2915 /*
2916 * very conservative, leave the buffer pinned if
2917 * anyone might need it.
2918 */
2921 }
2926 /* why is this safe?
2927 * reiserfs_setattr updates i_size in the on disk
2928 * stat data before allowing vmtruncate to be called.
2929 *
2930 * If buffer was put onto the ordered list for this
2931 * transaction, we know for sure either this transaction
2932 * or an older one already has updated i_size on disk,
2933 * and this ordered data won't be referenced in the file
2934 * if we crash.
2935 *
2936 * if the buffer was put onto the ordered list for an older
2937 * transaction, we need to leave it around
2938 */
2942 }
2943 free_jh:
2946 }
2950 }
2952 /* clm -- taken from fs/buffer.c:block_invalidate_page */
2954 {
2974 /*
2975 * is this block fully invalidated?
2976 */
2980 else
2982 }
2987 /*
2988 * We release buffers only if the entire page is being invalidated.
2989 * The get_block cached value has been unconditionally invalidated,
2990 * so real IO is not possible anymore.
2991 */
2994 /* maybe should BUG_ON(!ret); - neilb */
2995 }
2996 out:
2998 }
3001 {
3006 }
3008 }
3010 /*
3011 * Returns 1 if the page's buffers were dropped. The page is locked.
3012 *
3013 * Takes j_dirty_buffers_lock to protect the b_assoc_buffers list_heads
3014 * in the buffers at page_buffers(page).
3015 *
3016 * even in -o notail mode, we can't be sure an old mount without -o notail
3017 * didn't create files with tails.
3018 */
3020 {
3038 }
3039 }
3046 }
3048 /* We thank Mingming Cao for helping us understand in great detail what
3049 to do in this section of the code. */
3053 {
3060 }
3063 {
3069 /* must be turned off for recursive notify_change calls */
3074 /* version 2 items will be caught by the s_maxbytes check
3075 ** done for us in vmtruncate
3076 */
3081 }
3082 /* fill in hole pointers in the expanding truncate case. */
3088 /* we're changing at most 2 bitmaps, inode + super */
3093 }
3096 }
3099 /*
3100 * file size is changed, ctime and mtime are
3101 * to be updated
3102 */
3104 }
3105 }
3110 /* stat data of format v3.5 has 16 bit uid and gid */
3113 }
3129 /* (user+group)*(old+new) structure - we count quota info and , inode write (sb, inode) */
3130 error =
3132 jbegin_count);
3135 error =
3139 jbegin_count);
3141 }
3142 /* Update corresponding info in inode so that everything is in
3143 * one transaction */
3149 error =
3151 }
3152 }
3154 /*
3155 * Relax the lock here, as it might truncate the
3156 * inode pages and wait for inode pages locks.
3157 * To release such page lock, the owner needs the
3158 * reiserfs lock
3159 */
3163 }
3164 }
3169 }
3171 out:
3175 }
3189 };