| blob | a14d6cd9eeda2670251c0ec3377518912cb465a2 |
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) */
40 /* The = 0 happens when we abort creating a new inode for some reason like lack of space.. */
41 if (!(inode->i_state & I_NEW) && INODE_PKEY(inode)->k_objectid != 0) { /* also handles bad_inode case */
52 /* Do quota update inside a transaction for journaled quotas. We must do that
53 * after delete_object so that quota updates go into the same transaction as
54 * stat data deletion */
61 /* check return value from reiserfs_delete_object after
62 * ending the transaction
63 */
67 /* all items of file are deleted, so we can remove "save" link */
69 * about an error here */
71 /* no object items are in the tree */
72 ;
73 }
74 out:
78 }
82 {
90 }
92 /* take base of inode_key (it comes from inode always) (dirid, objectid) and version from an inode, set
93 offset and type of key */
96 {
100 length);
101 }
103 //
104 // when key is 0, do not set version and short key
105 //
110 {
115 }
120 /* set_ih_free_space (ih, 0); */
121 // for directory items it is entry count, for directs and stat
122 // datas - 0xffff, for indirects - 0
124 }
126 //
127 // FIXME: we might cache recently accessed indirect item
129 // Ugh. Not too eager for that....
130 // I cut the code until such time as I see a convincing argument (benchmark).
131 // I don't want a bloated inode struct..., and I don't like code complexity....
133 /* cutting the code is fine, since it really isn't in use yet and is easy
134 ** to add back in. But, Vladimir has a really good idea here. Think
135 ** about what happens for reading a file. For each page,
136 ** The VFS layer calls reiserfs_readpage, who searches the tree to find
137 ** an indirect item. This indirect item has X number of pointers, where
138 ** X is a big number if we've done the block allocation right. But,
139 ** we only use one or two of these pointers during each call to readpage,
140 ** needlessly researching again later on.
141 **
142 ** The size of the cache could be dynamic based on the size of the file.
143 **
144 ** I'd also like to see us cache the location the stat data item, since
145 ** we are needlessly researching for that frequently.
146 **
147 ** --chris
148 */
150 /* If this page has a file tail in it, and
151 ** it was read in by get_block_create_0, the page data is valid,
152 ** but tail is still sitting in a direct item, and we can't write to
153 ** it. So, look through this page, and check all the mapped buffers
154 ** to make sure they have valid block numbers. Any that don't need
155 ** to be unmapped, so that block_prepare_write will correctly call
156 ** reiserfs_get_block to convert the tail into an unformatted node
157 */
159 {
169 }
172 }
173 }
175 /* reiserfs_get_block does not need to allocate a block only if it has been
176 done already or non-hole position has been found in the indirect item */
180 {
187 }
190 {
192 }
196 {
198 }
200 //
201 // files which were created in the earlier version can not be longer,
202 // than 2 gb
203 //
205 {
211 }
215 {
225 /* we cannot restart while nested */
228 }
235 }
237 }
239 // it is called by get_block when create == 0. Returns block number
240 // for 'block'-th logical block of file. When it hits direct item it
241 // returns 0 (being called from bmap) or read direct item into piece
242 // of page (bh_result)
244 // Please improve the english/clarity in the comment above, as it is
245 // hard to understand.
249 {
255 b_blocknr_t blocknr;
263 // prepare the key to look for the 'block'-th block of file
268 research:
276 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
277 // That there is some MMAPED data associated with it that is yet to be written to disk.
281 }
283 }
284 //
290 /* FIXME: here we could cache indirect item or part of it in
291 the inode to avoid search_by_key in case of subsequent
292 access to file */
300 }
302 // We do not return -ENOENT if there is a hole but page is uptodate, because it means
303 // That there is some MMAPED data associated with it that is yet to be written to disk.
307 }
313 }
314 // requested data are in direct item(s)
316 // we are called by bmap. FIXME: we can not map block of file
317 // when it is stored in direct item(s)
322 }
324 /* if we've got a direct item, and the buffer or page was uptodate,
325 ** we don't want to pull data off disk again. skip to the
326 ** end, where we map the buffer and return
327 */
331 /*
332 ** grab_tail_page can trigger calls to reiserfs_get_block on up to date
333 ** pages without any buffers. If the page is up to date, we don't want
334 ** read old data off disk. Set the up to date bit on the buffer instead
335 ** and jump to the end
336 */
340 }
341 // read file tail into part of page
346 /* we only want to kmap if we are reading the tail into the page.
347 ** this is not the common case, so we don't kmap until we are
348 ** sure we need to. But, this means the item might move if
349 ** kmap schedules
350 */
356 }
357 }
363 }
364 /* make sure we don't read more bytes than actually exist in
365 ** the file. This can happen in odd cases where i_size isn't
366 ** correct, and when direct item padding results in a few
367 ** extra bytes at the end of the direct item
368 */
372 chars =
378 }
387 // we done, if read direct item is not the last item of
388 // node FIXME: we could try to check right delimiting key
389 // to see whether direct item continues in the right
390 // neighbor or rely on i_size
393 // update key to look for the next piece
397 // i/o error most likely
406 finished:
412 /* this buffer has valid data, but isn't valid for io. mapping it to
413 * block #0 tells the rest of reiserfs it just has a tail in it
414 */
418 }
420 // this is called to create file map. So, _get_block_create_0 will not
421 // read direct item
424 {
429 /* do not read the direct item */
433 }
435 /* special version of get_block that is only used by grab_tail_page right
436 ** now. It is sent to block_prepare_write, and when you try to get a
437 ** block past the end of the file (or a block from a hole) it returns
438 ** -ENOENT instead of a valid buffer. block_prepare_write expects to
439 ** be able to do i/o on the buffers returned, unless an error value
440 ** is also returned.
441 **
442 ** So, this allows block_prepare_write to be used for reading a single block
443 ** in a page. Where it does not produce a valid page for holes, or past the
444 ** end of the file. This turns out to be exactly what we need for reading
445 ** tails for conversion.
446 **
447 ** The point of the wrapper is forcing a certain value for create, even
448 ** though the VFS layer is calling this function with create==1. If you
449 ** don't want to send create == GET_BLOCK_NO_HOLE to reiserfs_get_block,
450 ** don't use this function.
451 */
455 {
457 }
459 /* This is special helper for reiserfs_get_block in case we are executing
460 direct_IO request. */
462 sector_t iblock,
465 {
470 /* We set the b_size before reiserfs_get_block call since it is
471 referenced in convert_tail_for_hole() that may be called from
472 reiserfs_get_block() */
480 /* don't allow direct io onto tail pages */
482 /* make sure future calls to the direct io funcs for this offset
483 ** in the file fail by unmapping the buffer
484 */
487 }
488 /* Possible unpacked tail. Flush the data before pages have
489 disappeared */
498 }
499 out:
501 }
503 /*
504 ** helper function for when reiserfs_get_block is called for a hole
505 ** but the file tail is still in a direct item
506 ** bh_result is the buffer head for the hole
507 ** tail_offset is the offset of the start of the tail in the file
508 **
509 ** This calls prepare_write, which will start a new transaction
510 ** you should not be in a transaction, or have any paths held when you
511 ** call this.
512 */
515 loff_t tail_offset)
516 {
527 /* always try to read until the end of the block */
532 /* hole_page can be zero in case of direct_io, we are sure
533 that we cannot get here if we write with O_DIRECT into
534 tail page */
540 }
543 }
545 /* we don't have to make sure the conversion did not happen while
546 ** we were locking the page because anyone that could convert
547 ** must first take i_mutex.
548 **
549 ** We must fix the tail page for writing because it might have buffers
550 ** that are mapped, but have a block number of 0. This indicates tail
551 ** data that has been read directly into the page, and block_prepare_write
552 ** won't trigger a get_block in this case.
553 */
559 /* tail conversion might change the data in the page */
564 unlock:
568 }
569 out:
571 }
574 sector_t block,
578 {
581 #ifdef REISERFS_PREALLOCATE
585 }
586 #endif
588 block);
589 }
593 {
605 /* space reserved in transaction batch:
606 . 3 balancings in direct->indirect conversion
607 . 1 block involved into reiserfs_update_sd()
608 XXX in practically impossible worst case direct2indirect()
609 can incur (much) more than 3 balancings.
610 quota update for user, group */
616 loff_t new_offset =
619 /* bad.... */
626 }
628 /* if !create, we aren't changing the FS, so we don't need to
629 ** log anything, so we don't need to start a transaction
630 */
633 /* find number of block-th logical block of the file */
638 }
639 /*
640 * if we're already in a transaction, make sure to close
641 * any new transactions we start in this func
642 */
647 /* If file is of such a size, that it might have a tail and tails are enabled
648 ** we should mark it as possibly needing tail packing on close
649 */
656 /* set the key of the first byte in the 'block'-th block of file */
659 start_trans:
664 }
666 }
667 research:
673 }
685 /* we have to allocate block for the unformatted node */
689 }
691 repeat =
696 /* restart the transaction to give the journal a chance to free
697 ** some blocks. releases the path, so we have to go back to
698 ** research if we succeed on the second try
699 */
704 repeat =
710 }
713 else
716 }
721 }
722 }
725 b_blocknr_t unfm_ptr;
726 /* 'block'-th block is in the file already (there is
727 corresponding cell in some indirect item). But it may be
728 zero unformatted node pointer (hole) */
731 /* use allocated block to plug the hole */
736 bh);
738 }
747 }
756 /* the item was found, so new blocks were not added to the file
757 ** there is no need to make sure the inode is updated with this
758 ** transaction
759 */
761 }
766 }
768 /* desired position is not found or is in the direct item. We have
769 to append file with holes up to 'block'-th block converting
770 direct items to indirect one if necessary */
777 /* indirect item has to be inserted */
783 /* we are going to add 'block'-th block to the file. Use
784 allocated block for that */
790 }
795 retval =
802 }
803 //mark_tail_converted (inode);
805 /* direct item has to be converted */
806 loff_t tail_offset;
808 tail_offset =
812 /* direct item we just found fits into block we have
813 to map. Convert it into unformatted node: use
814 bh_result for the conversion */
820 /* we have to padd file tail stored in direct item(s)
821 up to block size and convert it to unformatted
822 node. FIXME: this should also get into page cache */
825 /*
826 * ugly, but we can only end the transaction if
827 * we aren't nested
828 */
831 retval =
832 reiserfs_end_persistent_transaction
837 }
839 retval =
841 tail_offset);
846 "convert tail failed "
849 retval);
851 /* the bitmap, the super, and the stat data == 3 */
856 inode,
857 allocated_block_nr,
859 }
861 }
863 }
864 retval =
866 tail_offset);
872 }
873 /* it is important the set_buffer_uptodate is done after
874 ** the direct2indirect. The buffer might contain valid
875 ** data newer than the data on disk (read by readpage, changed,
876 ** and then sent here by writepage). direct2indirect needs
877 ** to know if unbh was already up to date, so it can decide
878 ** if the data in unbh needs to be replaced with data from
879 ** the disk
880 */
883 /* unbh->b_page == NULL in case of DIRECT_IO request, this means
884 buffer will disappear shortly, so it should not be added to
885 */
887 /* we've converted the tail, so we must
888 ** flush unbh before the transaction commits
889 */
892 /* mark it dirty now to prevent commit_write from adding
893 ** this buffer to the inode's dirty buffer list
894 */
895 /*
896 * AKPM: changed __mark_buffer_dirty to mark_buffer_dirty().
897 * It's still atomic, but it sets the page dirty too,
898 * which makes it eligible for writeback at any time by the
899 * VM (which was also the case with __mark_buffer_dirty())
900 */
902 }
904 /* append indirect item with holes if needed, when appending
905 pointer to 'block'-th block use block, which is already
906 allocated */
909 // one block which is a fastpath
911 __u64 max_to_insert =
913 UNFM_P_SIZE;
914 __u64 blocks_needed;
918 /* indirect item has to be appended, set up key of that position */
924 //pos_in_item * inode->i_sb->s_blocksize,
929 blocks_needed =
933 s_blocksize_bits);
938 un = kzalloc(min(blocks_needed, max_to_insert) * UNFM_P_SIZE, GFP_ATOMIC); // We need to avoid scheduling.
943 }
944 }
946 /* we are going to add target block to the file. Use allocated
947 block for that */
955 /* paste hole to the indirect item */
956 /* If kmalloc failed, max_to_insert becomes zero and it means we
957 only have space for one block */
958 blocks_needed =
960 }
961 retval =
964 UNFM_P_SIZE *
965 blocks_needed);
974 }
976 /* We need to mark new file size in case this function will be
977 interrupted/aborted later on. And we may do this only for
978 holes. */
981 }
982 }
987 /* this loop could log more blocks than we had originally asked
988 ** for. So, we have to allow the transaction to end if it is
989 ** too big or too full. Update the inode so things are
990 ** consistent if we crash before the function returns
991 **
992 ** release the path so that anybody waiting on the path before
993 ** ending their transaction will be able to continue.
994 */
999 }
1000 /* inserting indirect pointers for a hole can take a
1001 ** long time. reschedule if needed
1002 */
1009 }
1019 }
1028 failure:
1036 }
1041 }
1043 static int
1046 {
1048 }
1050 /* Compute real number of used bytes by file
1051 * Following three functions can go away when we'll have enough space in stat item
1052 */
1054 {
1061 /* End of file is also in full block with indirect reference, so round
1062 ** up to the next block.
1063 **
1064 ** there is just no way to know if the tail is actually packed
1065 ** on the file, so we have to assume it isn't. When we pack the
1066 ** tail, we add 4 bytes to pretend there really is an unformatted
1067 ** node pointer
1068 */
1069 bytes =
1072 sd_size;
1074 }
1078 {
1082 }
1085 }
1087 /* Compute number of blocks used by file in ReiserFS counting */
1089 {
1093 /* keeps fsck and non-quota versions of reiserfs happy */
1096 }
1098 /* files from before the quota patch might i_blocks such that
1099 ** bytes < real_space. Deal with that here to prevent it from
1100 ** going negative.
1101 */
1105 }
1107 //
1108 // BAD: new directories have stat data of new type and all other items
1109 // of old type. Version stored in the inode says about body items, so
1110 // in update_stat_data we can not rely on inode, but have to check
1111 // item version directly
1112 //
1114 // called by read_locked_inode
1116 {
1119 __u32 rdev;
1120 //int version = ITEM_VERSION_1;
1160 // there was a bug in <=3.5.23 when i_blocks could take negative
1161 // values. Starting from 3.5.17 this value could even be stored in
1162 // stat data. For such files we set i_blocks based on file
1163 // size. Just 2 notes: this can be wrong for sparce files. On-disk value will be
1164 // only updated if file's inode will ever change
1166 }
1171 /* an early bug in the quota code can give us an odd number for the
1172 ** block count. This is incorrect, fix it here.
1173 */
1176 }
1179 SD_V1_SIZE));
1180 /* nopack is initially zero for v1 objects. For v2 objects,
1181 nopack is initialised from sd_attrs */
1184 // new stat data found, but object may have old items
1185 // (directories and symlinks)
1204 else
1209 else
1215 SD_V2_SIZE));
1216 /* read persistent inode attributes from sd and initalise
1217 generic inode flags from them */
1220 }
1237 }
1238 }
1240 // update new stat data with inode fields
1242 {
1244 __u16 flags;
1257 else
1262 }
1264 // used to copy inode's fields to old stat data
1266 {
1280 else
1283 // Sigh. i_first_direct_byte is back
1286 }
1288 /* NOTE, you must prepare the buffer head before sending it here,
1289 ** and then log it after the call
1290 */
1292 loff_t size)
1293 {
1305 // path points to old stat data
1309 }
1312 }
1316 {
1330 /* look for the object's stat data */
1334 "i/o failure occurred trying to "
1337 }
1342 /*reiserfs_warning (inode->i_sb, "vs-13050: reiserfs_update_sd: i_nlink == 0, stat data not found"); */
1344 }
1346 "stat data of object %k (nlink == %d) "
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 "i/o failure occurred trying to find "
1429 }
1431 /* a stale NFS handle can trigger this without it being an error */
1436 }
1440 /* It is possible that knfsd is trying to access inode of a file
1441 that is being removed from the disk by some other thread. As we
1442 update sd on unlink all that is required is to check for nlink
1443 here. This bug was first found by Sizif when debugging
1444 SquidNG/Butterfly, forgotten, and found again after Philippe
1445 Gramoulle <philippe.gramoulle@mmania.com> reproduced it.
1447 More logical fix would require changes in fs/inode.c:iput() to
1448 remove inode from hash-table _after_ fs cleaned disk stuff up and
1449 in iget() to return NULL if I_FREEING inode is found in
1450 hash-table. */
1451 /* Currently there is one place where it's ok to meet inode with
1452 nlink==0: processing of open-unlinked and half-truncated files
1453 during mount (fs/reiserfs/super.c:finish_unfinished()). */
1457 "dead inode read from disk %K. "
1461 }
1465 }
1467 /**
1468 * reiserfs_find_actor() - "find actor" reiserfs supplies to iget5_locked().
1469 *
1470 * @inode: inode from hash table to check
1471 * @opaque: "cookie" passed to iget5_locked(). This is &reiserfs_iget_args.
1472 *
1473 * This function is called by iget5_locked() to distinguish reiserfs inodes
1474 * having the same inode numbers. Such inodes can only exist due to some
1475 * error condition. One of them should be bad. Inodes with identical
1476 * inode numbers (objectids) are distinguished by parent directory ids.
1477 *
1478 */
1480 {
1484 /* args is already in CPU order */
1487 }
1490 {
1505 }
1508 /* either due to i/o error or a stale NFS handle */
1511 }
1513 }
1518 {
1530 }
1534 }
1538 {
1539 /* fhtype happens to reflect the number of u32s encoded.
1540 * due to a bug in earlier code, fhtype might indicate there
1541 * are more u32s then actually fitted.
1542 * so if fhtype seems to be more than len, reduce fhtype.
1543 * Valid types are:
1544 * 2 - objectid + dir_id - legacy support
1545 * 3 - objectid + dir_id + generation
1546 * 4 - objectid + dir_id + objectid and dirid of parent - legacy
1547 * 5 - objectid + dir_id + generation + objectid and dirid of parent
1548 * 6 - as above plus generation of directory
1549 * 6 does not fit in NFSv2 handles
1550 */
1557 }
1561 }
1565 {
1573 }
1577 {
1588 /* no room for directory info? return what we've stored so far */
1600 }
1603 }
1605 /* looks for stat data, then copies fields to it, marks the buffer
1606 containing stat data as dirty */
1607 /* reiserfs inodes are never really dirty, since the dirty inode call
1608 ** always logs them. This call allows the VFS inode marking routines
1609 ** to properly mark inodes for datasync and such, but only actually
1610 ** does something when called for a synchronous update.
1611 */
1613 {
1619 /* memory pressure can sometimes initiate write_inode calls with sync == 1,
1620 ** these cases are just when the system needs ram, not when the
1621 ** inode needs to reach disk for safety, and they can safely be
1622 ** ignored because the altered inode has already been logged.
1623 */
1629 }
1631 }
1633 }
1635 /* stat data of new object is inserted already, this inserts the item
1636 containing "." and ".." entries */
1641 {
1654 /* compose item head for new item. Directories consist of items of
1655 old type (ITEM_VERSION_1). Do not set key (second arg is 0), it
1656 is done by reiserfs_new_inode */
1673 }
1675 /* look for place in the tree for new item */
1681 }
1688 }
1690 /* insert item, that is empty directory item */
1692 }
1694 /* stat data of object has been inserted, this inserts the item
1695 containing the body of symlink */
1696 static int reiserfs_new_symlink(struct reiserfs_transaction_handle *th, struct inode *inode, /* Inode of symlink */
1700 {
1715 /* look for place in the tree for new item */
1721 }
1728 }
1730 /* insert item, that is body of symlink */
1732 }
1734 /* inserts the stat data into the tree, and then calls
1735 reiserfs_new_directory (to insert ".", ".." item if new object is
1736 directory) or reiserfs_new_symlink (to insert symlink body if new
1737 object is symlink) or nothing (if new object is regular file)
1739 NOTE! uid and gid must already be set in the inode. If we return
1740 non-zero due to an error, we have to drop the quota previously allocated
1741 for the fresh inode. This can only be done outside a transaction, so
1742 if we return non-zero, we also end the transaction. */
1745 /* 0 for regular, EMTRY_DIR_SIZE for dirs,
1746 strlen (symname) for symlinks) */
1750 {
1765 }
1769 }
1773 /* item head of new item */
1779 }
1784 else
1793 }
1795 /* not a perfect generation count, as object ids can be reused, but
1796 ** this is as good as reiserfs can do right now.
1797 ** note that the private part of inode isn't filled in yet, we have
1798 ** to use the directory.
1799 */
1801 else
1802 #if defined( USE_INODE_GENERATION_COUNTER )
1805 #else
1807 #endif
1809 /* fill stat data */
1812 /* uid and gid must already be set by the caller for quota init */
1814 /* symlink cannot be immutable or append only, right? */
1837 /* key to search for correct place for new stat data */
1842 /* find proper place for inserting of stat data */
1847 }
1852 }
1856 /* i_uid or i_gid is too big to be stored in stat data v3.5 */
1859 }
1863 }
1864 // store in in-core inode the key of stat data and version all
1865 // object items will have (directory items will have old offset
1866 // format, other new objects will consist of new items)
1869 else
1873 else
1876 /* insert the stat data into the tree */
1877 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1880 #endif
1881 retval =
1888 }
1889 #ifdef DISPLACE_NEW_PACKING_LOCALITIES
1892 #endif
1894 /* insert item with "." and ".." */
1895 retval =
1897 }
1900 /* insert body of symlink */
1903 retval =
1905 i_size);
1906 }
1912 }
1921 }
1924 "ACLs aren't enabled in the fs, "
1939 }
1940 }
1947 /* it looks like you can easily compress these two goto targets into
1948 * one. Keeping it like this doesn't actually hurt anything, and they
1949 * are place holders for what the quota code actually needs.
1950 */
1951 out_bad_inode:
1952 /* Invalidate the object, nothing was inserted yet */
1955 /* Quota change must be inside a transaction for journaling */
1958 out_end_trans:
1960 /* Drop can be outside and it needs more credits so it's better to have it outside */
1965 out_inserted_sd:
1971 }
1973 /*
1974 ** finds the tail page in the page cache,
1975 ** reads the last block in.
1976 **
1977 ** On success, page_result is set to a locked, pinned page, and bh_result
1978 ** is set to an up to date buffer for the last block in the file. returns 0.
1979 **
1980 ** tail conversion is not done, so bh_result might not be valid for writing
1981 ** check buffer_mapped(bh_result) and bh_result->b_blocknr != 0 before
1982 ** trying to write the block.
1983 **
1984 ** on failure, nonzero is returned, page_result and bh_result are untouched.
1985 */
1989 {
1991 /* we want the page with the last byte in the file,
1992 ** not the page that will hold the next byte for appending
1993 */
2004 /* we know that we are only called with inode->i_size > 0.
2005 ** we also know that a file tail can never be as big as a block
2006 ** If i_size % blocksize == 0, our file is currently block aligned
2007 ** and it won't need converting or zeroing after a truncate.
2008 */
2011 }
2016 }
2017 /* start within the page of the last block in the file */
2021 reiserfs_get_block_create_0);
2030 }
2036 /* note, this should never happen, prepare_write should
2037 ** be taking care of this for us. If the buffer isn't up to date,
2038 ** I've screwed up the code to find the buffer, or the code to
2039 ** call prepare_write
2040 */
2045 }
2049 out:
2052 unlock:
2056 }
2058 /*
2059 ** vfs version of truncate file. Must NOT be called with
2060 ** a transaction already started.
2061 **
2062 ** some code taken from block_truncate_page
2063 */
2065 {
2067 /* we want the offset for the first byte after the end of the file */
2081 // -ENOENT means we truncated past the end of the file,
2082 // and get_block_create_0 could not find a block to read in,
2083 // which is ok.
2087 error);
2090 }
2091 }
2093 /* so, if page != NULL, we have a buffer head for the offset at
2094 ** the end of the file. if the bh is mapped, and bh->b_blocknr != 0,
2095 ** then we have an unformatted node. Otherwise, we have a direct item,
2096 ** and no zeroing is required on disk. We zero after the truncate,
2097 ** because the truncate might pack the item anyway
2098 ** (it will unmap bh if it packs).
2099 */
2100 /* it is enough to reserve space in transaction for 2 balancings:
2101 one for "save" link adding and another for the first
2102 cut_from_item. 1 is for update_sd */
2109 /* we are doing real truncate: if the system crashes before the last
2110 transaction of truncating gets committed - on reboot the file
2111 either appears truncated properly or not truncated at all */
2114 error =
2119 /* check reiserfs_do_truncate after ending the transaction */
2123 }
2129 }
2133 /* if we are not on a block boundary */
2139 }
2140 }
2143 }
2147 out:
2151 }
2154 }
2159 {
2177 /* catch places below that try to log something without starting a trans */
2182 }
2185 start_over:
2189 research:
2194 }
2201 /* we've found an unformatted node */
2206 }
2208 /* crap, we are writing to a hole */
2211 }
2224 /* vs-3050 is gone, no need to drop the path */
2233 bh);
2235 }
2236 }
2244 }
2247 copy_size);
2253 /* are there still bytes left? */
2258 copy_size);
2260 }
2266 }
2269 out:
2276 }
2279 /* this is where we fill in holes in the file. */
2282 GET_BLOCK_CREATE | GET_BLOCK_NO_IMUX
2287 /* get_block failed to find a mapped unformatted node. */
2290 }
2291 }
2292 }
2296 /* we've copied data from the page into the direct item, so the
2297 * buffer in the page is now clean, mark it to reflect that.
2298 */
2302 }
2304 }
2306 /*
2307 * mason@suse.com: updated in 2.5.54 to follow the same general io
2308 * start/recovery path as __block_write_full_page, along with special
2309 * code to handle reiserfs tails.
2310 */
2313 {
2318 sector_t last_block;
2328 /* no logging allowed when nonblocking or from PF_MEMALLOC */
2333 }
2335 /* The page dirty bit is cleared before writepage is called, which
2336 * means we have to tell create_empty_buffers to make dirty buffers
2337 * The page really should be up to date at this point, so tossing
2338 * in the BH_Uptodate is just a sanity check.
2339 */
2343 }
2346 /* last page in the file, zero out any contents past the
2347 ** last byte in the file
2348 */
2353 /* no file contents in this page */
2357 }
2359 }
2363 /* first map all the buffers, logging any direct items we find */
2366 /*
2367 * This can happen when the block size is less than
2368 * the page size. The corresponding bytes in the page
2369 * were zero filled above
2370 */
2377 /* not mapped yet, or it points to a direct item, search
2378 * the btree for the mapping info, and log any direct
2379 * items found
2380 */
2383 }
2384 }
2386 block++;
2389 /*
2390 * we start the transaction after map_block_for_writepage,
2391 * because it can create holes in the file (an unbounded operation).
2392 * starting it here, we can make a reliable estimate for how many
2393 * blocks we're going to log
2394 */
2402 }
2404 }
2405 /* now go through and lock any dirty buffers on the page */
2417 }
2418 /* from this point on, we know the buffer is mapped to a
2419 * real block and not a direct item
2420 */
2427 }
2428 }
2433 }
2441 }
2446 /*
2447 * since any buffer might be the only dirty buffer on the page,
2448 * the first submit_bh can bring the page out of writeback.
2449 * be careful with the buffers.
2450 */
2455 nr++;
2456 }
2462 done:
2464 /*
2465 * if this page only had a direct item, it is very possible for
2466 * no io to be required without there being an error. Or,
2467 * someone else could have locked them and sent them down the
2468 * pipe without locking the page
2469 */
2475 }
2481 }
2484 fail:
2485 /* catches various errors, we need to make sure any valid dirty blocks
2486 * get to the media. The page is currently locked and not marked for
2487 * writeback
2488 */
2497 /*
2498 * clear any dirty bits that might have come from getting
2499 * attached to a dirty page
2500 */
2502 }
2514 nr++;
2515 }
2520 }
2523 {
2525 }
2528 {
2532 }
2538 {
2541 pgoff_t index;
2549 pos ++;
2551 }
2564 journal_info;
2569 }
2571 reiserfs_get_block);
2574 /* this gets a little ugly. If reiserfs_get_block returned an
2575 * error and left a transacstion running, we've got to close it,
2576 * and we've got to free handle if it was a persistent transaction.
2577 *
2578 * But, if we had nested into an existing transaction, we need
2579 * to just drop the ref count on the handle.
2580 *
2581 * If old_ref == 0, the transaction is from reiserfs_get_block,
2582 * and it was a persistent trans. Otherwise, it was nested above.
2583 */
2594 }
2595 }
2596 }
2600 }
2602 }
2606 {
2616 journal_info;
2621 }
2626 /* this gets a little ugly. If reiserfs_get_block returned an
2627 * error and left a transacstion running, we've got to close it,
2628 * and we've got to free handle if it was a persistent transaction.
2629 *
2630 * But, if we had nested into an existing transaction, we need
2631 * to just drop the ref count on the handle.
2632 *
2633 * If old_ref == 0, the transaction is from reiserfs_get_block,
2634 * and it was a persistent trans. Otherwise, it was nested above.
2635 */
2646 }
2647 }
2648 }
2651 }
2654 {
2656 }
2661 {
2669 pos ++;
2674 else
2683 }
2688 /* generic_commit_write does this for us, but does not update the
2689 ** transaction tracking stuff when the size changes. So, we have
2690 ** to do the i_size updates here.
2691 */
2696 /* If the file have grown beyond the border where it
2697 can have a tail, unmark it as needing a tail
2698 packing */
2709 }
2712 /*
2713 * this will just nest into our transaction. It's important
2714 * to use mark_inode_dirty so the inode gets pushed around on the
2715 * dirty lists, and so that O_SYNC works as expected
2716 */
2724 }
2733 }
2735 out:
2740 journal_error:
2747 }
2750 }
2754 {
2764 }
2767 /* generic_commit_write does this for us, but does not update the
2768 ** transaction tracking stuff when the size changes. So, we have
2769 ** to do the i_size updates here.
2770 */
2774 /* If the file have grown beyond the border where it
2775 can have a tail, unmark it as needing a tail
2776 packing */
2787 }
2790 /*
2791 * this will just nest into our transaction. It's important
2792 * to use mark_inode_dirty so the inode gets pushed around on the
2793 * dirty lists, and so that O_SYNC works as expected
2794 */
2802 }
2811 }
2813 out:
2816 journal_error:
2823 }
2826 }
2829 {
2833 else
2837 else
2841 else
2845 else
2849 else
2851 }
2852 }
2855 {
2859 else
2863 else
2867 else
2871 else
2873 }
2874 }
2876 /* decide if this buffer needs to stay around for data logging or ordered
2877 ** write purposes
2878 */
2880 {
2888 }
2889 /* the page is locked, and the only places that log a data buffer
2890 * also lock the page.
2891 */
2893 /*
2894 * very conservative, leave the buffer pinned if
2895 * anyone might need it.
2896 */
2899 }
2904 /* why is this safe?
2905 * reiserfs_setattr updates i_size in the on disk
2906 * stat data before allowing vmtruncate to be called.
2907 *
2908 * If buffer was put onto the ordered list for this
2909 * transaction, we know for sure either this transaction
2910 * or an older one already has updated i_size on disk,
2911 * and this ordered data won't be referenced in the file
2912 * if we crash.
2913 *
2914 * if the buffer was put onto the ordered list for an older
2915 * transaction, we need to leave it around
2916 */
2920 }
2921 free_jh:
2924 }
2928 }
2930 /* clm -- taken from fs/buffer.c:block_invalidate_page */
2932 {
2952 /*
2953 * is this block fully invalidated?
2954 */
2958 else
2960 }
2965 /*
2966 * We release buffers only if the entire page is being invalidated.
2967 * The get_block cached value has been unconditionally invalidated,
2968 * so real IO is not possible anymore.
2969 */
2972 /* maybe should BUG_ON(!ret); - neilb */
2973 }
2974 out:
2976 }
2979 {
2984 }
2986 }
2988 /*
2989 * Returns 1 if the page's buffers were dropped. The page is locked.
2990 *
2991 * Takes j_dirty_buffers_lock to protect the b_assoc_buffers list_heads
2992 * in the buffers at page_buffers(page).
2993 *
2994 * even in -o notail mode, we can't be sure an old mount without -o notail
2995 * didn't create files with tails.
2996 */
2998 {
3016 }
3017 }
3024 }
3026 /* We thank Mingming Cao for helping us understand in great detail what
3027 to do in this section of the code. */
3031 {
3038 }
3041 {
3046 /* must be turned off for recursive notify_change calls */
3051 /* version 2 items will be caught by the s_maxbytes check
3052 ** done for us in vmtruncate
3053 */
3058 }
3059 /* fill in hole pointers in the expanding truncate case. */
3065 /* we're changing at most 2 bitmaps, inode + super */
3070 }
3073 }
3076 /*
3077 * file size is changed, ctime and mtime are
3078 * to be updated
3079 */
3081 }
3082 }
3087 /* stat data of format v3.5 has 16 bit uid and gid */
3090 }
3106 /* (user+group)*(old+new) structure - we count quota info and , inode write (sb, inode) */
3107 error =
3109 jbegin_count);
3112 error =
3116 jbegin_count);
3118 }
3119 /* Update corresponding info in inode so that everything is in
3120 * one transaction */
3126 error =
3128 }
3129 }
3132 }
3137 }
3139 out:
3142 }
3156 };