| blob | 52eb5d293a343dc26a544c7f5d7b3fbec1de4b04 |
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Write ahead logging implementation copyright Chris Mason 2000
4 *
5 * The background commits make this code very interrelated, and
6 * overly complex. I need to rethink things a bit....The major players:
7 *
8 * journal_begin -- call with the number of blocks you expect to log.
9 * If the current transaction is too
10 * old, it will block until the current transaction is
11 * finished, and then start a new one.
12 * Usually, your transaction will get joined in with
13 * previous ones for speed.
14 *
15 * journal_join -- same as journal_begin, but won't block on the current
16 * transaction regardless of age. Don't ever call
17 * this. Ever. There are only two places it should be
18 * called from, and they are both inside this file.
19 *
20 * journal_mark_dirty -- adds blocks into this transaction. clears any flags
21 * that might make them get sent to disk
22 * and then marks them BH_JDirty. Puts the buffer head
23 * into the current transaction hash.
24 *
25 * journal_end -- if the current transaction is batchable, it does nothing
26 * otherwise, it could do an async/synchronous commit, or
27 * a full flush of all log and real blocks in the
28 * transaction.
29 *
30 * flush_old_commits -- if the current transaction is too old, it is ended and
31 * commit blocks are sent to disk. Forces commit blocks
32 * to disk for all backgrounded commits that have been
33 * around too long.
34 * -- Note, if you call this as an immediate flush from
35 * from within kupdate, it will ignore the immediate flag
36 */
38 #include <linux/time.h>
39 #include <linux/semaphore.h>
40 #include <linux/vmalloc.h>
42 #include <linux/kernel.h>
43 #include <linux/errno.h>
44 #include <linux/fcntl.h>
45 #include <linux/stat.h>
46 #include <linux/string.h>
47 #include <linux/buffer_head.h>
48 #include <linux/workqueue.h>
49 #include <linux/writeback.h>
50 #include <linux/blkdev.h>
51 #include <linux/backing-dev.h>
52 #include <linux/uaccess.h>
53 #include <linux/slab.h>
56 /* gets a struct reiserfs_journal_list * from a list head */
57 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
58 j_list))
59 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
60 j_working_list))
62 /* must be correct to keep the desc and commit structs at 4k */
63 #define JOURNAL_TRANS_HALF 1018
66 /* cnode stat bits. Move these into reiserfs_fs.h */
68 /* this block was freed, and can't be written. */
69 #define BLOCK_FREED 2
70 /* this block was freed during this transaction, and can't be written */
71 #define BLOCK_FREED_HOLDER 3
73 /* used in flush_journal_list */
74 #define BLOCK_NEEDS_FLUSH 4
75 #define BLOCK_DIRTIED 5
77 /* journal list state bits */
78 #define LIST_TOUCHED 1
79 #define LIST_DIRTY 2
82 /* flags for do_journal_end */
102 /* values for join in do_journal_begin_r */
105 /* join the running transaction if at all possible */
107 /* called from cleanup code, ignores aborted flag */
109 };
116 {
120 }
122 /*
123 * clears BH_Dirty and sticks the buffer on the clean list. Called because
124 * I can't allow refile_buffer to make schedule happen after I've freed a
125 * block. Look at remove_from_transaction and journal_mark_freed for
126 * more details.
127 */
129 {
133 }
135 }
139 {
146 }
151 }
155 }
158 {
164 repeat:
172 }
177 }
179 }
182 {
191 }
192 }
195 {
205 /* this is ok, we'll try again when more are needed */
207 }
208 }
209 }
212 b_blocknr_t block,
214 {
220 }
223 }
227 {
236 }
237 }
238 }
240 /*
241 * only call this on FS unmount.
242 */
245 {
254 }
256 }
259 {
271 }
274 }
276 /*
277 * get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
278 * jb_array is the array to be filled in.
279 */
283 {
298 }
299 }
303 }
305 }
307 /*
308 * find an available list bitmap. If you can't find one, flush a commit list
309 * and try again
310 */
312 struct reiserfs_journal_list
314 {
329 }
332 }
333 }
334 /* double check to make sure if flushed correctly */
339 }
341 /*
342 * allocates a new chunk of X nodes, and links them all together as a list.
343 * Uses the cnode->next and cnode->prev pointers
344 * returns NULL on failure
345 */
347 {
352 }
357 }
363 }
366 }
368 /* pulls a cnode off the free list, or returns NULL on failure */
370 {
378 }
384 }
387 }
391 }
393 /*
394 * returns a cnode to the free list
395 */
398 {
405 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
409 }
410 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
412 }
415 {
418 }
420 /*
421 * return a cnode with same dev, block number and size in table,
422 * or null if not found
423 */
425 super_block
427 struct
428 reiserfs_journal_cnode
431 {
438 }
440 }
442 /*
443 * this actually means 'can this block be reallocated yet?'. If you set
444 * search_all, a block can only be allocated if it is not in the current
445 * transaction, was not freed by the current transaction, and has no chance
446 * of ever being overwritten by a replay after crashing.
447 *
448 * If you don't set search_all, a block can only be allocated if it is not
449 * in the current transaction. Since deleting a block removes it from the
450 * current transaction, this case should never happen. If you don't set
451 * search_all, make sure you never write the block without logging it.
452 *
453 * next_zero_bit is a suggestion about the next block to try for find_forward.
454 * when bl is rejected because it is set in a journal list bitmap, we search
455 * for the next zero bit in the bitmap that rejected bl. Then, we return
456 * that through next_zero_bit for find_forward to try.
457 *
458 * Just because we return something in next_zero_bit does not mean we won't
459 * reject it on the next call to reiserfs_in_journal
460 */
464 {
474 /*
475 * If we aren't doing a search_all, this is a metablock, and it
476 * will be logged before use. if we crash before the transaction
477 * that freed it commits, this transaction won't have committed
478 * either, and the block will never be written
479 */
487 data)) {
491 data),
495 }
496 }
497 }
500 /* is it in any old transactions? */
505 }
507 /* is it in the current transaction. This should never happen */
511 }
514 /* safe for reuse */
516 }
518 /* insert cn into table */
521 {
529 }
531 }
533 /* lock the current transaction */
535 {
539 }
541 /* unlock the current transaction */
543 {
545 }
548 {
550 }
554 {
558 }
561 }
563 /*
564 * this used to be much more involved, and I'm keeping it just in case
565 * things get ugly again. it gets called by flush_commit_list, and
566 * cleans up any data stored about blocks freed during a transaction.
567 */
570 {
575 }
578 }
582 {
591 }
592 }
594 }
596 /*
597 * If page->mapping was null, we failed to truncate this page for
598 * some reason. Most likely because it was truncated after being
599 * logged via data=journal.
600 *
601 * This does a check to see if the buffer belongs to one of these
602 * lost pages before doing the final put_bh. If page->mapping was
603 * null, it tries to free buffers on the page, which should make the
604 * final put_page drop the page from the lru.
605 */
607 {
618 }
619 }
622 {
627 }
630 else
635 }
638 {
641 else
645 }
648 {
658 }
661 {
668 }
670 #define CHUNK_SIZE 32
674 };
677 {
681 }
683 }
686 {
690 }
692 }
696 {
708 }
709 }
711 }
715 {
722 }
724 }
725 }
727 /*
728 * we want to free the jh when the buffer has been written
729 * and waited on
730 */
732 {
745 }
746 }
750 {
758 }
762 no_jh:
766 /*
767 * buffer must be locked for __add_jh, should be able to have
768 * two adds at the same time
769 */
773 }
779 }
782 }
785 {
787 }
789 {
791 }
793 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
798 {
816 }
824 }
825 /*
826 * in theory, dirty non-uptodate buffers should never get here,
827 * but the upper layer io error paths still have a few quirks.
828 * Handle them here as gracefully as we can
829 */
833 }
840 }
841 loop_next:
844 }
849 }
860 }
863 }
864 /*
865 * ugly interaction with invalidatepage here.
866 * reiserfs_invalidate_page will pin any buffer that has a
867 * valid journal head from an older transaction. If someone
868 * else sets our buffer dirty after we write it in the first
869 * loop, and then someone truncates the page away, nobody
870 * will ever write the buffer. We're safe if we write the
871 * page one last time after freeing the journal header.
872 */
877 }
880 }
883 }
887 {
896 find_first:
897 /*
898 * first we walk backwards to find the oldest uncommitted transation
899 */
910 }
912 /* if we didn't find any older uncommitted transactions, return now */
915 }
928 /* list we were called with is gone, return */
932 /*
933 * the one we just flushed is gone, this means
934 * all older lists are also gone, so first_jl
935 * is no longer valid either. Go back to the
936 * beginning.
937 */
941 }
942 }
948 }
949 }
951 }
954 {
963 }
966 }
968 /*
969 * if this journal list still has commit blocks unflushed, send them to disk.
970 *
971 * log areas must be flushed in order (transaction 2 can't commit before
972 * transaction 1) Before the commit block can by written, every other log
973 * block must be safely on disk
974 */
977 {
979 b_blocknr_t bn;
991 }
993 /*
994 * before we can put our commit blocks on disk, we have to make
995 * sure everyone older than us is on disk too
996 */
1003 /*
1004 * list disappeared during flush_older_commits.
1005 * return
1006 */
1008 }
1009 }
1011 /* make sure nobody is trying to flush this one at the same time */
1017 }
1020 /* this commit is done, exit */
1024 }
1027 }
1032 /*
1033 * We might sleep in numerous places inside
1034 * write_ordered_buffers. Relax the write lock.
1035 */
1042 }
1044 /*
1045 * for the description block and all the log blocks, submit any buffers
1046 * that haven't already reached the disk. Try to write at least 256
1047 * log blocks. later on, we will only wait on blocks that correspond
1048 * to this transaction, but while we're unplugging we might as well
1049 * get a chunk of data on there.
1050 */
1064 }
1066 }
1067 }
1078 /*
1079 * since we're using ll_rw_blk above, it might have skipped
1080 * over a locked buffer. Double check here
1081 */
1082 /* redundant, sync_dirty_buffer() checks */
1087 }
1089 #ifdef CONFIG_REISERFS_CHECK
1092 #endif
1094 }
1095 /* once for journal_find_get_block */
1097 /* once due to original getblk in do_journal_end */
1100 }
1104 /*
1105 * If there was a write error in the journal - we can't commit
1106 * this transaction - it will be invalid and, if successful,
1107 * will just end up propagating the write error out to
1108 * the file system.
1109 */
1118 else
1121 }
1123 /*
1124 * If there was a write error in the journal - we can't commit this
1125 * transaction - it will be invalid and, if successful, will just end
1126 * up propagating the write error out to the filesystem.
1127 */
1129 #ifdef CONFIG_REISERFS_CHECK
1131 #endif
1133 }
1139 }
1142 /*
1143 * now, every commit block is on the disk. It is safe to allow
1144 * blocks freed during this transaction to be reallocated
1145 */
1150 /* mark the metadata dirty */
1157 }
1159 put_jl:
1164 __func__);
1166 }
1168 /*
1169 * flush_journal_list frequently needs to find a newer transaction for a
1170 * given block. This does that, or returns NULL if it can't find anything
1171 */
1173 reiserfs_journal_cnode
1175 {
1183 }
1185 }
1187 }
1194 /*
1195 * once all the real blocks have been flushed, it is safe to remove them
1196 * from the journal list for this transaction. Aside from freeing the
1197 * cnode, this also allows the block to be reallocated for data blocks
1198 * if it had been deleted.
1199 */
1203 {
1208 /*
1209 * which is better, to lock once around the whole loop, or
1210 * to lock for each call to remove_journal_hash?
1211 */
1219 }
1223 }
1227 }
1229 }
1231 /*
1232 * if this timestamp is greater than the timestamp we wrote last to the
1233 * header block, write it to the header block. once this is done, I can
1234 * safely say the log area for this transaction won't ever be replayed,
1235 * and I can start releasing blocks in this transaction for reuse as data
1236 * blocks. called by flush_journal_list, before it calls
1237 * remove_all_from_journal_list
1238 */
1242 {
1256 #ifdef CONFIG_REISERFS_CHECK
1259 #endif
1261 }
1262 }
1266 b_data);
1277 else
1285 }
1286 }
1288 }
1293 {
1295 }
1297 /*
1298 ** flush any and all journal lists older than you are
1299 ** can only be called from flush_journal_list
1300 */
1303 {
1309 /*
1310 * we know we are the only ones flushing things, no extra race
1311 * protection is required.
1312 */
1313 restart:
1315 /* Did we wrap? */
1321 /* do not flush all */
1324 /* other_jl is now deleted from the list */
1326 }
1328 }
1332 {
1337 }
1338 }
1340 /*
1341 * flush a journal list, both commit and real blocks
1342 *
1343 * always set flushall to 1, unless you are calling from inside
1344 * flush_journal_list
1345 *
1346 * IMPORTANT. This can only be called while there are no journal writers,
1347 * and the journal is locked. That means it can only be called from
1348 * do_journal_end, or by journal_release
1349 */
1352 {
1369 }
1371 /* if flushall == 0, the lock is already held */
1376 }
1383 }
1385 /* if all the work is already done, get out of here */
1389 }
1391 /*
1392 * start by putting the commit list on disk. This will also flush
1393 * the commit lists of any olders transactions
1394 */
1401 /* are we done now? */
1405 }
1407 /*
1408 * loop through each cnode, see if we need to write it,
1409 * or wait on a more recent transaction, or just ignore it
1410 */
1414 }
1420 /* blocknr of 0 is no longer in the hash, ignore it */
1423 }
1425 /*
1426 * This transaction failed commit.
1427 * Don't write out to the disk
1428 */
1433 /*
1434 * the order is important here. We check pjl to make sure we
1435 * don't clear BH_JDirty_wait if we aren't the one writing this
1436 * block to disk
1437 */
1441 /*
1442 * we do this to make sure nobody releases the
1443 * buffer while we are working with it
1444 */
1452 /*
1453 * everything with !pjl && jwait
1454 * should be writable
1455 */
1457 }
1458 }
1460 /*
1461 * if someone has this block in a newer transaction, just make
1462 * sure they are committed, and don't try writing it to disk
1463 */
1468 }
1470 /*
1471 * bh == NULL when the block got to disk on its own, OR,
1472 * the block got freed in a future transaction
1473 */
1476 }
1478 /*
1479 * this should never happen. kupdate_one_transaction has
1480 * this list locked while it works, so we should never see a
1481 * buffer here that is not marked JDirty_wait
1482 */
1485 "BAD! buffer %llu %cdirty %cjwait, "
1490 }
1492 /*
1493 * we inc again because saved_bh gets decremented
1494 * at free_cnode
1495 */
1502 else
1504 count++;
1510 }
1511 free_cnode:
1515 /*
1516 * we incremented this to keep others from
1517 * taking the buffer head away
1518 */
1523 }
1524 }
1525 }
1533 }
1542 }
1544 #ifdef CONFIG_REISERFS_CHECK
1547 #endif
1549 }
1550 /*
1551 * note, we must clear the JDirty_wait bit
1552 * after the up to date check, otherwise we
1553 * race against our flushpage routine
1554 */
1558 /* drop one ref for us */
1560 /* drop one ref for journal_mark_dirty */
1562 }
1564 }
1565 }
1570 __func__);
1571 flush_older_and_return:
1573 /*
1574 * before we can update the journal header block, we _must_ flush all
1575 * real blocks from all older transactions to disk. This is because
1576 * once the header block is updated, this transaction will not be
1577 * replayed after a crash
1578 */
1581 }
1584 /*
1585 * before we can remove everything from the hash tables for this
1586 * transaction, we must make sure it can never be replayed
1587 *
1588 * since we are only called from do_journal_end, we know for sure there
1589 * are no allocations going on while we are flushing journal lists. So,
1590 * we only need to update the journal header block for the last list
1591 * being flushed
1592 */
1594 err =
1602 __func__);
1603 }
1613 }
1616 /*
1617 * not strictly required since we are freeing the list, but it should
1618 * help find code using dead lists later on
1619 */
1631 }
1636 {
1644 }
1648 /*
1649 * if the blocknr == 0, this has been cleared from the hash,
1650 * skip it
1651 */
1654 }
1657 /*
1658 * we can race against journal_mark_freed when we try
1659 * to lock_buffer(cn->bh), so we have to inc the buffer
1660 * count, and recheck things after locking
1661 */
1670 ret++;
1672 /* note, cn->bh might be null now */
1674 }
1676 }
1677 next:
1680 }
1682 }
1684 /* used by flush_commit_list */
1687 {
1695 /*
1696 * look for a more recent transaction that logged this
1697 * buffer. Only the most recent transaction with a buffer in
1698 * it is allowed to send that buffer to disk
1699 */
1704 /*
1705 * if the buffer is prepared, it will either be logged
1706 * or restored. If restored, we need to make sure
1707 * it actually gets marked dirty
1708 */
1715 }
1716 }
1718 }
1720 }
1727 {
1740 }
1742 /*
1743 * we've got j_flush_mutex held, nobody is going to delete any
1744 * of these lists out from underneath us
1745 */
1754 }
1759 transactions_flushed++;
1763 /* did we wrap? */
1766 }
1769 /* don't bother with older transactions */
1772 }
1775 }
1777 done:
1780 }
1782 /*
1783 * for o_sync and fsync heavy applications, they tend to use
1784 * all the journa list slots with tiny transactions. These
1785 * trigger lots and lots of calls to update the header block, which
1786 * adds seeks and slows things down.
1787 *
1788 * This function tries to clear out a large chunk of the journal lists
1789 * at once, which makes everything faster since only the newest journal
1790 * list updates the header block
1791 */
1794 {
1807 /* in data logging mode, try harder to flush a lot of blocks */
1810 /* flush for 256 transactions or limit blocks, whichever comes first */
1815 }
1819 }
1825 }
1829 /*
1830 * try to find a group of blocks we can flush across all the
1831 * transactions, but only bother if we've actually spanned
1832 * across multiple lists
1833 */
1836 }
1841 }
1843 /*
1844 * removes any nodes in table with name block and dev as bh.
1845 * only touchs the hnext and hprev pointers.
1846 */
1851 {
1858 }
1866 }
1871 }
1875 /*
1876 * anybody who clears the cur->bh will also
1877 * dec the nonzerolen
1878 */
1883 }
1885 }
1886 }
1889 {
1899 }
1900 /*
1901 * j_header_bh is on the journal dev, make sure
1902 * not to release the journal dev until we brelse j_header_bh
1903 */
1906 }
1908 /*
1909 * call on unmount. Only set error to 1 if you haven't made your way out
1910 * of read_super() yet. Any other caller must keep error at 0.
1911 */
1914 {
1919 /*
1920 * we only want to flush out transactions if we were
1921 * called with error == 0
1922 */
1924 /* end the current trans */
1928 /*
1929 * make sure something gets logged to force
1930 * our way into the flush code
1931 */
1939 }
1940 }
1942 /* this also catches errors during the do_journal_end above */
1951 }
1952 }
1955 /*
1956 * We must release the write lock here because
1957 * the workqueue job (flush_async_commit) needs this lock
1958 */
1961 /*
1962 * Cancel flushing of old commits. Note that neither of these works
1963 * will be requeued because superblock is being shutdown and doesn't
1964 * have SB_ACTIVE set.
1965 */
1967 /* wait for all commits to finish */
1975 }
1977 /* * call on unmount. flush all journal trans, release all alloc'd ram */
1980 {
1982 }
1984 /* only call from an error condition inside reiserfs_read_super! */
1987 {
1989 }
1991 /*
1992 * compares description block with commit block.
1993 * returns 1 if they differ, 0 if they are the same
1994 */
1998 {
2004 }
2006 }
2008 /*
2009 * returns 0 if it did not find a description block
2010 * returns -1 if it found a corrupt commit block
2011 * returns 1 if both desc and commit were valid
2012 * NOTE: only called during fs mount
2013 */
2018 {
2033 "journal-986: transaction "
2034 "is valid returning because trans_id %d is greater than "
2039 }
2043 "journal-1087: transaction "
2044 "is valid returning because mount_id %d is less than "
2049 }
2052 "Bad transaction length %d "
2056 }
2059 /*
2060 * ok, we have a journal description block,
2061 * let's see if the transaction was valid
2062 */
2063 c_bh =
2073 "journal_transaction_is_valid, commit offset %ld had bad "
2084 "journal-1004: "
2085 "transaction_is_valid setting oldest invalid trans_id "
2088 }
2090 }
2093 "journal-1006: found valid "
2102 }
2103 }
2106 {
2110 }
2111 }
2113 /*
2114 * given the start, and values for the oldest acceptable transactions,
2115 * this either reads in a replays a transaction, or returns because the
2116 * transaction is invalid, or too old.
2117 * NOTE: only called during fs mount
2118 */
2124 {
2149 cur_dblock -
2153 }
2156 "journal_read_trans skipping because %d is != "
2158 newest_mount_id);
2161 }
2168 }
2172 "journal_read_transaction, "
2181 }
2189 }
2192 /*
2193 * now we know we've got a good transaction, and it was
2194 * inside the valid time ranges
2195 */
2198 GFP_NOFS);
2201 GFP_NOFS);
2210 }
2211 /* get all the buffer heads */
2228 }
2231 "REPLAY FAILURE fsck required! "
2232 "Block to replay is outside of "
2235 }
2236 /* make sure we don't try to replay onto log or reserved area */
2240 "REPLAY FAILURE fsck required! "
2242 abort_replay:
2250 }
2251 }
2252 /* read in the log blocks, memcpy to the corresponding real block */
2259 "REPLAY FAILURE fsck required! "
2269 }
2274 }
2275 /* flush out the real blocks */
2279 }
2284 "REPLAY FAILURE, fsck required! "
2293 }
2295 }
2296 cur_dblock =
2304 /*
2305 * init starting values for the first transaction, in case
2306 * this is the last transaction to be replayed.
2307 */
2311 /* check for trans_id overflow */
2319 }
2321 /*
2322 * This function reads blocks starting from block and to max_block of bufsize
2323 * size (but no more than BUFNR blocks at a time). This proved to improve
2324 * mounting speed on self-rebuilding raid5 arrays at least.
2325 * Right now it is only used from journal code. But later we might use it
2326 * from other places.
2327 * Note: Do not use journal_getblk/sb_getblk functions here!
2328 */
2331 b_blocknr_t max_block)
2332 {
2344 }
2354 }
2364 }
2366 /*
2367 * read and replay the log
2368 * on a clean unmount, the journal header's next unflushed pointer will be
2369 * to an invalid transaction. This tests that before finding all the
2370 * transactions in the log, which makes normal mount times fast.
2371 *
2372 * After a crash, this starts with the next unflushed transaction, and
2373 * replays until it finds one too old, or invalid.
2374 *
2375 * On exit, it sets things up so the first transaction will work correctly.
2376 * NOTE: only called during fs mount
2377 */
2379 {
2400 /*
2401 * step 1, read in the journal header block. Check the transaction
2402 * it says is the first unflushed, and if that transaction is not
2403 * valid, replay is done
2404 */
2410 }
2415 oldest_start =
2421 "journal-1153: found in "
2422 "header: first_unflushed_offset %d, last_flushed_trans_id "
2427 /*
2428 * now, we try to read the first unflushed offset. If it
2429 * is not valid, there is nothing more we can do, and it
2430 * makes no sense to read through the whole log.
2431 */
2432 d_bh =
2439 }
2442 }
2444 /*
2445 * ok, there are transactions that need to be replayed. start
2446 * with the first log block, find all the valid transactions, and
2447 * pick out the oldest.
2448 */
2453 /*
2454 * Note that it is required for blocksize of primary fs
2455 * device and journal device to be the same
2456 */
2457 d_bh =
2462 ret =
2473 "journal-1179: Setting "
2475 oldest_start -
2476 SB_ONDISK_JOURNAL_1st_BLOCK
2479 /* one we just read was older */
2483 "journal-1180: Resetting "
2485 oldest_start -
2486 SB_ONDISK_JOURNAL_1st_BLOCK
2488 }
2492 "journal-1299: Setting "
2495 }
2498 cur_dblock++;
2499 }
2501 }
2503 start_log_replay:
2507 "journal-1206: Starting replay "
2510 oldest_trans_id);
2512 }
2515 ret =
2522 }
2523 cur_dblock =
2525 replay_count++;
2528 }
2533 }
2534 /*
2535 * j_start does not get set correctly if we don't replay any
2536 * transactions. if we had a valid journal_header, set j_start
2537 * to the first unflushed transaction value, copy the trans_id
2538 * from the header
2539 */
2544 /* check for trans_id overflow */
2552 }
2560 }
2561 /* needed to satisfy the locking in _update_journal_header_block */
2567 /*
2568 * replay failed, caller must call free_journal_ram and abort
2569 * the mount
2570 */
2572 }
2575 }
2578 {
2590 }
2593 {
2595 }
2599 {
2603 }
2604 }
2609 {
2611 dev_t jdev;
2624 /* there is no "jdev" option and journal is on separate device */
2629 journal);
2642 }
2653 }
2660 }
2662 /*
2663 * When creating/tuning a file system user can assign some
2664 * journal params within boundaries which depend on the ratio
2665 * blocksize/standard_blocksize.
2666 *
2667 * For blocks >= standard_blocksize transaction size should
2668 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2669 * then JOURNAL_TRANS_MAX_DEFAULT.
2670 *
2671 * For blocks < standard_blocksize these boundaries should be
2672 * decreased proportionally.
2673 */
2674 #define REISERFS_STANDARD_BLKSIZE (4096)
2678 {
2680 /* Non-default journal params. Do sanity check for them. */
2688 JOURNAL_MIN_RATIO) {
2690 "bad transaction max size (%u). "
2693 }
2697 "bad transaction max batch (%u). "
2700 }
2702 /*
2703 * Default journal params.
2704 * The file system was created by old version
2705 * of mkreiserfs, so some fields contain zeros,
2706 * and we need to advise proper values for them
2707 */
2712 }
2716 }
2718 }
2720 /* must be called once on fs mount. calls journal_read for you */
2723 {
2737 }
2749 /* reserved for journal area support */
2751 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2755 REISERFS_DISK_OFFSET_IN_BYTES /
2758 /*
2759 * Sanity check to see is the standard journal fitting
2760 * within first bitmap (actual for small blocksizes)
2761 */
2766 "journal does not fit for area addressed "
2767 "by first of bitmap blocks. It starts at "
2773 }
2779 }
2783 /* read journal header */
2791 }
2794 /* make sure that journal matches to the super block */
2799 "journal header magic %x (device %pg) does "
2806 }
2821 }
2824 "journal first block %u, max trans len %u, max batch %u, "
2869 "allocation failed (%ld bytes). Journal is "
2870 "too large for available memory. Usually "
2874 }
2879 /*
2880 * get_list_bitmap() may call flush_commit_list() which
2881 * requires the lock. Calling flush_commit_list() shouldn't happen
2882 * this early but I like to be paranoid.
2883 */
2891 }
2898 }
2903 free_and_return:
2906 }
2908 /*
2909 * test for a polite end of the current transaction. Used by file_write,
2910 * and should be used by delete to make sure they don't write more than
2911 * can fit inside a single transaction
2912 */
2915 {
2918 /* cannot restart while nested */
2928 }
2933 }
2935 /* this must be called inside a transaction */
2937 {
2943 }
2945 /* this must be called without a transaction started */
2947 {
2951 }
2953 /* this must be called without a transaction started */
2955 {
2959 }
2962 {
2963 wait_queue_entry_t wait;
2967 /*
2968 * we don't want to use wait_event here because
2969 * we only want to wait once.
2970 */
2978 }
2981 }
2984 {
2988 }
2991 {
3006 }
3012 }
3013 }
3015 /*
3016 * join == true if you must join an existing transaction.
3017 * join == false if you can deal with waiting for others to finish
3018 *
3019 * this will block until the transaction is joinable. send the number of
3020 * blocks you expect to use in nblocks.
3021 */
3025 {
3038 /* set here for journal_join */
3042 relock:
3048 }
3058 }
3061 /*
3062 * if there is no room in the journal OR
3063 * if this transaction is too old, and we weren't called joinable,
3064 * wait for it to finish before beginning we don't sleep if there
3065 * aren't other writers
3066 */
3069 (!join
3079 /* allow others to finish this transaction */
3087 sched_count++;
3090 }
3091 }
3092 /*
3093 * don't mess with joining the transaction if all we
3094 * have to do is wait for someone else to do a commit
3095 */
3100 }
3102 }
3107 /* someone might have ended the transaction while we joined */
3112 }
3119 }
3120 /* we are the first writer, set trans_id */
3123 }
3133 out_fail:
3135 /*
3136 * Re-set th->t_super, so we can properly keep track of how many
3137 * persistent transactions there are. We need to do this so if this
3138 * call is part of a failed restart_transaction, we can free it later
3139 */
3142 }
3145 super_block
3148 {
3152 /*
3153 * if we're nesting into an existing transaction. It will be
3154 * persistent on its own
3155 */
3162 }
3170 }
3174 }
3177 {
3182 else
3187 }
3189 }
3193 {
3196 /*
3197 * this keeps do_journal_end from NULLing out the
3198 * current->journal_info pointer
3199 */
3203 }
3207 {
3210 /*
3211 * this keeps do_journal_end from NULLing out the
3212 * current->journal_info pointer
3213 */
3217 }
3221 {
3227 /* we are nesting into the current transaction */
3234 "BAD: refcount <= 1, but "
3238 /*
3239 * we've ended up with a handle from a different
3240 * filesystem. save it and restore on journal_end.
3241 * This should never really happen...
3242 */
3247 }
3250 }
3254 /*
3255 * I guess this boils down to being the reciprocal of clm-2100 above.
3256 * If do_journal_begin_r fails, we need to put it back, since
3257 * journal_end won't be called to do it. */
3260 else
3264 }
3266 /*
3267 * puts bh into the current transaction. If it was already there, reorders
3268 * removes the old pointers from the hash, and puts new ones in (to make
3269 * sure replay happen in the right order).
3270 *
3271 * if it was dirty, cleans and files onto the clean list. I can't let it
3272 * be dirty again until the transaction is committed.
3273 *
3274 * if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3275 */
3278 {
3291 }
3295 /* already in this transaction, we are done */
3299 }
3301 /*
3302 * this must be turned into a panic instead of a warning. We can't
3303 * allow a dirty or journal_dirty or locked buffer to be logged, as
3304 * some changes could get to disk too early. NOT GOOD.
3305 */
3308 "buffer %llu bad state "
3315 }
3322 }
3323 /*
3324 * this error means I've screwed up, and we've overflowed
3325 * the transaction. Nothing can be done here, except make the
3326 * FS readonly or panic.
3327 */
3332 }
3338 }
3342 }
3346 /* now put this guy on the end */
3351 }
3356 }
3367 }
3368 }
3378 }
3381 }
3384 {
3393 }
3400 /*
3401 * we aren't allowed to close a nested transaction on a
3402 * different filesystem from the one in the task struct
3403 */
3409 }
3413 }
3414 }
3416 /*
3417 * removes from the current transaction, relsing and descrementing any counters.
3418 * also files the removed buffer directly onto the clean list
3419 *
3420 * called by journal_mark_freed when a block has been deleted
3421 *
3422 * returns 1 if it cleaned and relsed the buffer. 0 otherwise
3423 */
3426 {
3435 }
3439 }
3442 }
3445 }
3448 }
3462 }
3464 }
3469 }
3471 /*
3472 * for any cnode in a journal list, it can only be dirtied of all the
3473 * transactions that include it are committed to disk.
3474 * this checks through each transaction, and returns 1 if you are allowed
3475 * to dirty, and 0 if you aren't
3476 *
3477 * it is called by dirty_journal_list, which is called after
3478 * flush_commit_list has gotten all the log blocks for a given
3479 * transaction on disk
3480 *
3481 */
3483 {
3489 /*
3490 * first test hprev. These are all newer than cn, so any node here
3491 * with the same block number and dev means this node can't be sent
3492 * to disk right now.
3493 */
3498 }
3500 }
3501 /*
3502 * then test hnext. These are all older than cn. As long as they
3503 * are committed to the log, it is safe to write cn to disk
3504 */
3511 }
3513 }
3515 }
3517 /*
3518 * syncs the commit blocks, but does not force the real buffers to disk
3519 * will wait until the current transaction is done/committed before returning
3520 */
3522 {
3527 /* you can sync while nested, very, very bad */
3533 }
3535 }
3537 /* writeback the pending async commits to disk */
3539 {
3548 /* last entry is the youngest, commit it and you get everything */
3552 }
3554 }
3556 /*
3557 * flushes any old transactions to disk
3558 * ends the current transaction if it is too old
3559 */
3561 {
3567 /*
3568 * safety check so we don't flush while we are replaying the log during
3569 * mount
3570 */
3574 /*
3575 * check the current transaction. If there are no writers, and it is
3576 * too old, finish it, and force the commit blocks to disk
3577 */
3588 /*
3589 * we're only being called from kreiserfsd, it makes
3590 * no sense to do an async commit so that kreiserfsd
3591 * can do it later
3592 */
3594 }
3595 }
3596 }
3598 /*
3599 * returns 0 if do_journal_end should return right away, returns 1 if
3600 * do_journal_end should finish the commit
3601 *
3602 * if the current transaction is too old, but still has writers, this will
3603 * wait on j_join_wait until all the writers are done. By the time it
3604 * wakes up, the transaction it was called has already ended, so it just
3605 * flushes the commit list and returns 0.
3606 *
3607 * Won't batch when flush or commit_now is set. Also won't batch when
3608 * others are waiting on j_join_wait.
3609 *
3610 * Note, we can't allow the journal_end to proceed while there are still
3611 * writers in the log.
3612 */
3614 {
3630 }
3633 /* <= 0 is allowed. unmounting might not call begin */
3637 /*
3638 * BUG, deal with case where j_len is 0, but people previously
3639 * freed blocks need to be released will be dealt with by next
3640 * transaction that actually writes something, but should be taken
3641 * care of in this trans
3642 */
3645 /*
3646 * if wcount > 0, and we are called to with flush or commit_now,
3647 * we wait on j_join_wait. We will wake up when the last writer has
3648 * finished the transaction, and started it on its way to the disk.
3649 * Then, we flush the commit or journal list, and just return 0
3650 * because the rest of journal end was already done for this
3651 * transaction.
3652 */
3664 }
3667 /*
3668 * sleep while the current transaction is
3669 * still j_jlocked
3670 */
3679 }
3681 }
3682 }
3689 }
3691 }
3694 }
3696 /* deal with old transactions where we are the last writers */
3701 }
3702 /* don't batch when someone is waiting on j_join_wait */
3703 /* don't batch when syncing the commit or flushing the whole trans */
3711 }
3717 }
3719 }
3721 /*
3722 * Does all the work that makes deleting blocks safe.
3723 * when deleting a block mark BH_JNew, just remove it from the current
3724 * transaction, clean it's buffer_head and move on.
3725 *
3726 * otherwise:
3727 * set a bit for the block in the journal bitmap. That will prevent it from
3728 * being allocated for unformatted nodes before this transaction has finished.
3729 *
3730 * mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers.
3731 * That will prevent any old transactions with this block from trying to flush
3732 * to the real location. Since we aren't removing the cnode from the
3733 * journal_list_hash, *the block can't be reallocated yet.
3734 *
3735 * Then remove it from the current transaction, decrementing any counters and
3736 * filing it on the clean list.
3737 */
3740 {
3752 }
3753 /* if it is journal new, we just remove it from this transaction */
3760 /*
3761 * set the bit for this block in the journal bitmap
3762 * for this transaction
3763 */
3768 }
3771 /* Note, the entire while loop is not allowed to schedule. */
3776 }
3779 /*
3780 * find all older transactions with this block,
3781 * make sure they don't try to write it out
3782 */
3784 blocknr);
3789 /*
3790 * remove_from_transaction will brelse
3791 * the buffer if it was in the current
3792 * trans
3793 */
3796 bh);
3799 bh);
3807 }
3808 }
3809 /*
3810 * since we are clearing the bh,
3811 * we MUST dec nonzerolen
3812 */
3815 j_nonzerolen);
3816 }
3818 }
3819 }
3821 }
3822 }
3827 }
3830 {
3834 }
3836 /*
3837 * returns -1 on error, 0 if no commits/barriers were done and 1
3838 * if a transaction was actually committed and the barrier was done
3839 */
3842 {
3848 /*
3849 * is it from the current transaction,
3850 * or from an unknown transaction?
3851 */
3854 /*
3855 * try to let other writers come in and
3856 * grow this transaction
3857 */
3861 }
3867 /* someone might have ended this transaction while we joined */
3874 }
3881 /*
3882 * this gets tricky, we have to make sure the journal list in
3883 * the inode still exists. We know the list is still around
3884 * if we've got a larger transaction id than the oldest list
3885 */
3886 flush_commit_only:
3888 /*
3889 * we only set ret to 1 when we know for sure
3890 * the barrier hasn't been started yet on the commit
3891 * block.
3892 */
3898 }
3899 }
3900 /* otherwise the list is gone, and long since committed */
3902 }
3905 {
3909 /*
3910 * for the whole inode, assume unset id means it was
3911 * changed in the current transaction. More conservative
3912 */
3916 /* jl will be updated in __commit_trans_jl */
3917 }
3920 }
3924 {
3929 }
3940 }
3942 }
3944 }
3947 /*
3948 * before we can change a metadata block, we have to make sure it won't
3949 * be written to disk while we are altering it. So, we must:
3950 * clean it
3951 * wait on it.
3952 */
3955 {
3962 }
3967 }
3970 }
3972 /*
3973 * long and ugly. If flush, will not return until all commit
3974 * blocks and all real buffers in the trans are on disk.
3975 * If no_async, won't return until all commit blocks are on disk.
3976 *
3977 * keep reading, there are comments as you go along
3978 *
3979 * If the journal is aborted, we just clean up. Things like flushing
3980 * journal lists, etc just won't happen.
3981 */
3983 {
4008 /*
4009 * protect flush_older_commits from doing mistakes if the
4010 * transaction ID counter gets overflowed.
4011 */
4023 }
4029 }
4033 }
4035 /*
4036 * check_journal_end locks the journal, and unlocks if it does
4037 * not return 1 it tells us if we should continue with the
4038 * journal_end, or just return
4039 */
4045 }
4047 /* check_journal_end might set these, check again */
4050 }
4052 /*
4053 * j must wait means we have to flush the log blocks, and the
4054 * real blocks for this transaction
4055 */
4058 }
4059 #ifdef REISERFS_PREALLOCATE
4060 /*
4061 * quota ops might need to nest, setup the journal_info pointer
4062 * for them and raise the refcount so that it is > 0.
4063 */
4067 /* it should not involve new blocks into the transaction */
4072 #endif
4074 /* setup description block */
4075 d_bh =
4085 /*
4086 * setup commit block. Don't write (keep it clean too) this one
4087 * until after everyone else is written
4088 */
4097 /* init this journal list */
4100 /*
4101 * we lock the commit before doing anything because
4102 * we want to make sure nobody tries to run flush_commit_list until
4103 * the new transaction is fully setup, and we've already flushed the
4104 * ordered bh list
4105 */
4108 /* save the transaction id in case we need to commit it later */
4121 /*
4122 * The ENTIRE FOR LOOP MUST not cause schedule to occur.
4123 * for each real block, add it to the journal list hash,
4124 * copy into real block index array in the commit or desc block
4125 */
4133 }
4136 }
4141 }
4143 /*
4144 * make sure the block we are trying to log
4145 * is not a block of journal or reserved area
4146 */
4150 "Trying to log block %lu, "
4153 }
4166 }
4168 i--;
4169 }
4170 }
4176 /*
4177 * special check in case all buffers in the journal
4178 * were marked for not logging
4179 */
4182 /*
4183 * we're about to dirty all the log blocks, mark the description block
4184 * dirty now too. Don't mark the commit block dirty until all the
4185 * others are on disk
4186 */
4189 /*
4190 * first data block is j_start + 1, so add one to
4191 * cur_write_start wherever you use it
4192 */
4198 /* copy all the real blocks into log area. dirty log blocks */
4203 tmp_bh =
4207 jindex) %
4217 jindex++;
4221 /*
4222 * JDirty cleared sometime during transaction.
4223 * don't log this one
4224 */
4226 "BAD, buffer in journal hash, "
4229 }
4234 }
4236 /*
4237 * we are done with both the c_bh and d_bh, but
4238 * c_bh must be written after all other commit blocks,
4239 * so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4240 */
4244 /* now it is safe to insert this transaction on the main list */
4249 /* reset journal values for the next transaction */
4260 /* check for trans_id overflow */
4270 /*
4271 * make sure reiserfs_add_jh sees the new current_jl before we
4272 * write out the tails
4273 */
4276 /*
4277 * tail conversion targets have to hit the disk before we end the
4278 * transaction. Otherwise a later transaction might repack the tail
4279 * before this transaction commits, leaving the data block unflushed
4280 * and clean, if we crash before the later transaction commits, the
4281 * data block is lost.
4282 */
4288 }
4292 /*
4293 * honor the flush wishes from the caller, simple commits can
4294 * be done outside the journal lock, they are done below
4295 *
4296 * if we don't flush the commit list right now, we put it into
4297 * the work queue so the people waiting on the async progress work
4298 * queue don't wait for this proc to flush journal lists and such.
4299 */
4304 /*
4305 * Avoid queueing work when sb is being shut down. Transaction
4306 * will be flushed on journal shutdown.
4307 */
4311 }
4313 /*
4314 * if the next transaction has any chance of wrapping, flush
4315 * transactions that might get overwritten. If any journal lists
4316 * are very old flush them as well.
4317 */
4318 first_jl:
4329 /*
4330 * if we don't cross into the next
4331 * transaction and we don't wrap, there is
4332 * no way we can overlap any later transactions
4333 * break now
4334 */
4336 }
4346 /*
4347 * we don't overlap anything from out start
4348 * to the end of the log, and our wrapped
4349 * portion doesn't overlap anything at
4350 * the start of the log. We can break
4351 */
4353 }
4354 }
4355 }
4363 }
4367 /* wake up any body waiting to join. */
4374 }
4375 out:
4379 /*
4380 * Re-set th->t_super, so we can properly keep track of how many
4381 * persistent transactions there are. We need to do this so if this
4382 * call is part of a failed restart_transaction, we can free it later
4383 */
4387 }
4389 /* Send the file system read only and refuse new transactions */
4391 {
4402 #ifdef CONFIG_REISERFS_CHECK
4404 #endif
4405 }