| blob | 9d6486d416a33308047e1f21a92cd9cf21f7c144 |
1 /*
2 * Write ahead logging implementation copyright Chris Mason 2000
3 *
4 * The background commits make this code very interrelated, and
5 * overly complex. I need to rethink things a bit....The major players:
6 *
7 * journal_begin -- call with the number of blocks you expect to log.
8 * If the current transaction is too
9 * old, it will block until the current transaction is
10 * finished, and then start a new one.
11 * Usually, your transaction will get joined in with
12 * previous ones for speed.
13 *
14 * journal_join -- same as journal_begin, but won't block on the current
15 * transaction regardless of age. Don't ever call
16 * this. Ever. There are only two places it should be
17 * called from, and they are both inside this file.
18 *
19 * journal_mark_dirty -- adds blocks into this transaction. clears any flags
20 * that might make them get sent to disk
21 * and then marks them BH_JDirty. Puts the buffer head
22 * into the current transaction hash.
23 *
24 * journal_end -- if the current transaction is batchable, it does nothing
25 * otherwise, it could do an async/synchronous commit, or
26 * a full flush of all log and real blocks in the
27 * transaction.
28 *
29 * flush_old_commits -- if the current transaction is too old, it is ended and
30 * commit blocks are sent to disk. Forces commit blocks
31 * to disk for all backgrounded commits that have been
32 * around too long.
33 * -- Note, if you call this as an immediate flush from
34 * from within kupdate, it will ignore the immediate flag
35 */
37 #include <linux/time.h>
38 #include <linux/semaphore.h>
39 #include <linux/vmalloc.h>
41 #include <linux/kernel.h>
42 #include <linux/errno.h>
43 #include <linux/fcntl.h>
44 #include <linux/stat.h>
45 #include <linux/string.h>
46 #include <linux/buffer_head.h>
47 #include <linux/workqueue.h>
48 #include <linux/writeback.h>
49 #include <linux/blkdev.h>
50 #include <linux/backing-dev.h>
51 #include <linux/uaccess.h>
52 #include <linux/slab.h>
55 /* gets a struct reiserfs_journal_list * from a list head */
56 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
57 j_list))
58 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
59 j_working_list))
61 /* must be correct to keep the desc and commit structs at 4k */
62 #define JOURNAL_TRANS_HALF 1018
65 /* cnode stat bits. Move these into reiserfs_fs.h */
67 /* this block was freed, and can't be written. */
68 #define BLOCK_FREED 2
69 /* this block was freed during this transaction, and can't be written */
70 #define BLOCK_FREED_HOLDER 3
72 /* used in flush_journal_list */
73 #define BLOCK_NEEDS_FLUSH 4
74 #define BLOCK_DIRTIED 5
76 /* journal list state bits */
77 #define LIST_TOUCHED 1
78 #define LIST_DIRTY 2
81 /* flags for do_journal_end */
101 /* values for join in do_journal_begin_r */
104 /* join the running transaction if at all possible */
106 /* called from cleanup code, ignores aborted flag */
108 };
115 {
119 }
121 /*
122 * clears BH_Dirty and sticks the buffer on the clean list. Called because
123 * I can't allow refile_buffer to make schedule happen after I've freed a
124 * block. Look at remove_from_transaction and journal_mark_freed for
125 * more details.
126 */
128 {
132 }
134 }
138 {
145 }
150 }
154 }
157 {
163 repeat:
171 }
176 }
178 }
181 {
190 }
191 }
194 {
204 /* this is ok, we'll try again when more are needed */
206 }
207 }
208 }
211 b_blocknr_t block,
213 {
219 }
222 }
226 {
235 }
236 }
237 }
239 /*
240 * only call this on FS unmount.
241 */
244 {
253 }
255 }
258 {
270 }
273 }
275 /*
276 * get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
277 * jb_array is the array to be filled in.
278 */
282 {
297 }
298 }
302 }
304 }
306 /*
307 * find an available list bitmap. If you can't find one, flush a commit list
308 * and try again
309 */
311 struct reiserfs_journal_list
313 {
328 }
331 }
332 }
333 /* double check to make sure if flushed correctly */
338 }
340 /*
341 * allocates a new chunk of X nodes, and links them all together as a list.
342 * Uses the cnode->next and cnode->prev pointers
343 * returns NULL on failure
344 */
346 {
351 }
355 }
361 }
364 }
366 /* pulls a cnode off the free list, or returns NULL on failure */
368 {
376 }
382 }
385 }
389 }
391 /*
392 * returns a cnode to the free list
393 */
396 {
403 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
407 }
408 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
410 }
413 {
416 }
418 /*
419 * return a cnode with same dev, block number and size in table,
420 * or null if not found
421 */
423 super_block
425 struct
426 reiserfs_journal_cnode
429 {
436 }
438 }
440 /*
441 * this actually means 'can this block be reallocated yet?'. If you set
442 * search_all, a block can only be allocated if it is not in the current
443 * transaction, was not freed by the current transaction, and has no chance
444 * of ever being overwritten by a replay after crashing.
445 *
446 * If you don't set search_all, a block can only be allocated if it is not
447 * in the current transaction. Since deleting a block removes it from the
448 * current transaction, this case should never happen. If you don't set
449 * search_all, make sure you never write the block without logging it.
450 *
451 * next_zero_bit is a suggestion about the next block to try for find_forward.
452 * when bl is rejected because it is set in a journal list bitmap, we search
453 * for the next zero bit in the bitmap that rejected bl. Then, we return
454 * that through next_zero_bit for find_forward to try.
455 *
456 * Just because we return something in next_zero_bit does not mean we won't
457 * reject it on the next call to reiserfs_in_journal
458 */
462 {
472 /*
473 * If we aren't doing a search_all, this is a metablock, and it
474 * will be logged before use. if we crash before the transaction
475 * that freed it commits, this transaction won't have committed
476 * either, and the block will never be written
477 */
485 data)) {
489 data),
493 }
494 }
495 }
498 /* is it in any old transactions? */
503 }
505 /* is it in the current transaction. This should never happen */
509 }
512 /* safe for reuse */
514 }
516 /* insert cn into table */
519 {
527 }
529 }
531 /* lock the current transaction */
533 {
537 }
539 /* unlock the current transaction */
541 {
543 }
546 {
548 }
552 {
556 }
559 }
561 /*
562 * this used to be much more involved, and I'm keeping it just in case
563 * things get ugly again. it gets called by flush_commit_list, and
564 * cleans up any data stored about blocks freed during a transaction.
565 */
568 {
573 }
576 }
580 {
589 }
590 }
592 }
594 /*
595 * If page->mapping was null, we failed to truncate this page for
596 * some reason. Most likely because it was truncated after being
597 * logged via data=journal.
598 *
599 * This does a check to see if the buffer belongs to one of these
600 * lost pages before doing the final put_bh. If page->mapping was
601 * null, it tries to free buffers on the page, which should make the
602 * final page_cache_release drop the page from the lru.
603 */
605 {
616 }
617 }
620 {
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 */
1117 else
1120 }
1122 /*
1123 * If there was a write error in the journal - we can't commit this
1124 * transaction - it will be invalid and, if successful, will just end
1125 * up propagating the write error out to the filesystem.
1126 */
1128 #ifdef CONFIG_REISERFS_CHECK
1130 #endif
1132 }
1138 }
1141 /*
1142 * now, every commit block is on the disk. It is safe to allow
1143 * blocks freed during this transaction to be reallocated
1144 */
1149 /* mark the metadata dirty */
1156 }
1158 put_jl:
1163 __func__);
1165 }
1167 /*
1168 * flush_journal_list frequently needs to find a newer transaction for a
1169 * given block. This does that, or returns NULL if it can't find anything
1170 */
1172 reiserfs_journal_cnode
1174 {
1182 }
1184 }
1186 }
1193 /*
1194 * once all the real blocks have been flushed, it is safe to remove them
1195 * from the journal list for this transaction. Aside from freeing the
1196 * cnode, this also allows the block to be reallocated for data blocks
1197 * if it had been deleted.
1198 */
1202 {
1207 /*
1208 * which is better, to lock once around the whole loop, or
1209 * to lock for each call to remove_journal_hash?
1210 */
1218 }
1222 }
1226 }
1228 }
1230 /*
1231 * if this timestamp is greater than the timestamp we wrote last to the
1232 * header block, write it to the header block. once this is done, I can
1233 * safely say the log area for this transaction won't ever be replayed,
1234 * and I can start releasing blocks in this transaction for reuse as data
1235 * blocks. called by flush_journal_list, before it calls
1236 * remove_all_from_journal_list
1237 */
1241 {
1255 #ifdef CONFIG_REISERFS_CHECK
1258 #endif
1260 }
1261 }
1265 b_data);
1275 else
1283 }
1284 }
1286 }
1291 {
1293 }
1295 /*
1296 ** flush any and all journal lists older than you are
1297 ** can only be called from flush_journal_list
1298 */
1301 {
1307 /*
1308 * we know we are the only ones flushing things, no extra race
1309 * protection is required.
1310 */
1311 restart:
1313 /* Did we wrap? */
1319 /* do not flush all */
1322 /* other_jl is now deleted from the list */
1324 }
1326 }
1330 {
1335 }
1336 }
1338 /*
1339 * flush a journal list, both commit and real blocks
1340 *
1341 * always set flushall to 1, unless you are calling from inside
1342 * flush_journal_list
1343 *
1344 * IMPORTANT. This can only be called while there are no journal writers,
1345 * and the journal is locked. That means it can only be called from
1346 * do_journal_end, or by journal_release
1347 */
1350 {
1367 }
1369 /* if flushall == 0, the lock is already held */
1374 }
1381 }
1383 /* if all the work is already done, get out of here */
1387 }
1389 /*
1390 * start by putting the commit list on disk. This will also flush
1391 * the commit lists of any olders transactions
1392 */
1399 /* are we done now? */
1403 }
1405 /*
1406 * loop through each cnode, see if we need to write it,
1407 * or wait on a more recent transaction, or just ignore it
1408 */
1412 }
1418 /* blocknr of 0 is no longer in the hash, ignore it */
1421 }
1423 /*
1424 * This transaction failed commit.
1425 * Don't write out to the disk
1426 */
1431 /*
1432 * the order is important here. We check pjl to make sure we
1433 * don't clear BH_JDirty_wait if we aren't the one writing this
1434 * block to disk
1435 */
1439 /*
1440 * we do this to make sure nobody releases the
1441 * buffer while we are working with it
1442 */
1450 /*
1451 * everything with !pjl && jwait
1452 * should be writable
1453 */
1455 }
1456 }
1458 /*
1459 * if someone has this block in a newer transaction, just make
1460 * sure they are committed, and don't try writing it to disk
1461 */
1466 }
1468 /*
1469 * bh == NULL when the block got to disk on its own, OR,
1470 * the block got freed in a future transaction
1471 */
1474 }
1476 /*
1477 * this should never happen. kupdate_one_transaction has
1478 * this list locked while it works, so we should never see a
1479 * buffer here that is not marked JDirty_wait
1480 */
1483 "BAD! buffer %llu %cdirty %cjwait, "
1488 }
1490 /*
1491 * we inc again because saved_bh gets decremented
1492 * at free_cnode
1493 */
1500 else
1502 count++;
1508 }
1509 free_cnode:
1513 /*
1514 * we incremented this to keep others from
1515 * taking the buffer head away
1516 */
1521 }
1522 }
1523 }
1531 }
1540 }
1542 #ifdef CONFIG_REISERFS_CHECK
1545 #endif
1547 }
1548 /*
1549 * note, we must clear the JDirty_wait bit
1550 * after the up to date check, otherwise we
1551 * race against our flushpage routine
1552 */
1556 /* drop one ref for us */
1558 /* drop one ref for journal_mark_dirty */
1560 }
1562 }
1563 }
1568 __func__);
1569 flush_older_and_return:
1571 /*
1572 * before we can update the journal header block, we _must_ flush all
1573 * real blocks from all older transactions to disk. This is because
1574 * once the header block is updated, this transaction will not be
1575 * replayed after a crash
1576 */
1579 }
1582 /*
1583 * before we can remove everything from the hash tables for this
1584 * transaction, we must make sure it can never be replayed
1585 *
1586 * since we are only called from do_journal_end, we know for sure there
1587 * are no allocations going on while we are flushing journal lists. So,
1588 * we only need to update the journal header block for the last list
1589 * being flushed
1590 */
1592 err =
1600 __func__);
1601 }
1611 }
1614 /*
1615 * not strictly required since we are freeing the list, but it should
1616 * help find code using dead lists later on
1617 */
1629 }
1634 {
1642 }
1646 /*
1647 * if the blocknr == 0, this has been cleared from the hash,
1648 * skip it
1649 */
1652 }
1655 /*
1656 * we can race against journal_mark_freed when we try
1657 * to lock_buffer(cn->bh), so we have to inc the buffer
1658 * count, and recheck things after locking
1659 */
1668 ret++;
1670 /* note, cn->bh might be null now */
1672 }
1674 }
1675 next:
1678 }
1680 }
1682 /* used by flush_commit_list */
1685 {
1693 /*
1694 * look for a more recent transaction that logged this
1695 * buffer. Only the most recent transaction with a buffer in
1696 * it is allowed to send that buffer to disk
1697 */
1702 /*
1703 * if the buffer is prepared, it will either be logged
1704 * or restored. If restored, we need to make sure
1705 * it actually gets marked dirty
1706 */
1713 }
1714 }
1716 }
1718 }
1725 {
1738 }
1740 /*
1741 * we've got j_flush_mutex held, nobody is going to delete any
1742 * of these lists out from underneath us
1743 */
1752 }
1757 transactions_flushed++;
1761 /* did we wrap? */
1764 }
1767 /* don't bother with older transactions */
1770 }
1773 }
1775 done:
1778 }
1780 /*
1781 * for o_sync and fsync heavy applications, they tend to use
1782 * all the journa list slots with tiny transactions. These
1783 * trigger lots and lots of calls to update the header block, which
1784 * adds seeks and slows things down.
1785 *
1786 * This function tries to clear out a large chunk of the journal lists
1787 * at once, which makes everything faster since only the newest journal
1788 * list updates the header block
1789 */
1792 {
1805 /* in data logging mode, try harder to flush a lot of blocks */
1808 /* flush for 256 transactions or limit blocks, whichever comes first */
1813 }
1817 }
1823 }
1827 /*
1828 * try to find a group of blocks we can flush across all the
1829 * transactions, but only bother if we've actually spanned
1830 * across multiple lists
1831 */
1834 }
1839 }
1841 /*
1842 * removes any nodes in table with name block and dev as bh.
1843 * only touchs the hnext and hprev pointers.
1844 */
1849 {
1856 }
1864 }
1869 }
1873 /*
1874 * anybody who clears the cur->bh will also
1875 * dec the nonzerolen
1876 */
1881 }
1883 }
1884 }
1887 {
1897 }
1898 /*
1899 * j_header_bh is on the journal dev, make sure
1900 * not to release the journal dev until we brelse j_header_bh
1901 */
1904 }
1906 /*
1907 * call on unmount. Only set error to 1 if you haven't made your way out
1908 * of read_super() yet. Any other caller must keep error at 0.
1909 */
1912 {
1917 /*
1918 * we only want to flush out transactions if we were
1919 * called with error == 0
1920 */
1922 /* end the current trans */
1926 /*
1927 * make sure something gets logged to force
1928 * our way into the flush code
1929 */
1937 }
1938 }
1940 /* this also catches errors during the do_journal_end above */
1949 }
1950 }
1953 /*
1954 * We must release the write lock here because
1955 * the workqueue job (flush_async_commit) needs this lock
1956 */
1959 /*
1960 * Cancel flushing of old commits. Note that neither of these works
1961 * will be requeued because superblock is being shutdown and doesn't
1962 * have MS_ACTIVE set.
1963 */
1965 /* wait for all commits to finish */
1973 }
1975 /* * call on unmount. flush all journal trans, release all alloc'd ram */
1978 {
1980 }
1982 /* only call from an error condition inside reiserfs_read_super! */
1985 {
1987 }
1989 /*
1990 * compares description block with commit block.
1991 * returns 1 if they differ, 0 if they are the same
1992 */
1996 {
2002 }
2004 }
2006 /*
2007 * returns 0 if it did not find a description block
2008 * returns -1 if it found a corrupt commit block
2009 * returns 1 if both desc and commit were valid
2010 * NOTE: only called during fs mount
2011 */
2016 {
2031 "journal-986: transaction "
2032 "is valid returning because trans_id %d is greater than "
2037 }
2041 "journal-1087: transaction "
2042 "is valid returning because mount_id %d is less than "
2047 }
2050 "Bad transaction length %d "
2054 }
2057 /*
2058 * ok, we have a journal description block,
2059 * let's see if the transaction was valid
2060 */
2061 c_bh =
2071 "journal_transaction_is_valid, commit offset %ld had bad "
2082 "journal-1004: "
2083 "transaction_is_valid setting oldest invalid trans_id "
2086 }
2088 }
2091 "journal-1006: found valid "
2100 }
2101 }
2104 {
2108 }
2109 }
2111 /*
2112 * given the start, and values for the oldest acceptable transactions,
2113 * this either reads in a replays a transaction, or returns because the
2114 * transaction is invalid, or too old.
2115 * NOTE: only called during fs mount
2116 */
2122 {
2147 cur_dblock -
2151 }
2154 "journal_read_trans skipping because %d is != "
2156 newest_mount_id);
2159 }
2166 }
2170 "journal_read_transaction, "
2179 }
2187 }
2190 /*
2191 * now we know we've got a good transaction, and it was
2192 * inside the valid time ranges
2193 */
2206 }
2207 /* get all the buffer heads */
2224 }
2227 "REPLAY FAILURE fsck required! "
2228 "Block to replay is outside of "
2231 }
2232 /* make sure we don't try to replay onto log or reserved area */
2236 "REPLAY FAILURE fsck required! "
2238 abort_replay:
2246 }
2247 }
2248 /* read in the log blocks, memcpy to the corresponding real block */
2255 "REPLAY FAILURE fsck required! "
2265 }
2270 }
2271 /* flush out the real blocks */
2275 }
2280 "REPLAY FAILURE, fsck required! "
2289 }
2291 }
2292 cur_dblock =
2300 /*
2301 * init starting values for the first transaction, in case
2302 * this is the last transaction to be replayed.
2303 */
2307 /* check for trans_id overflow */
2315 }
2317 /*
2318 * This function reads blocks starting from block and to max_block of bufsize
2319 * size (but no more than BUFNR blocks at a time). This proved to improve
2320 * mounting speed on self-rebuilding raid5 arrays at least.
2321 * Right now it is only used from journal code. But later we might use it
2322 * from other places.
2323 * Note: Do not use journal_getblk/sb_getblk functions here!
2324 */
2327 b_blocknr_t max_block)
2328 {
2340 }
2350 }
2360 }
2362 /*
2363 * read and replay the log
2364 * on a clean unmount, the journal header's next unflushed pointer will be
2365 * to an invalid transaction. This tests that before finding all the
2366 * transactions in the log, which makes normal mount times fast.
2367 *
2368 * After a crash, this starts with the next unflushed transaction, and
2369 * replays until it finds one too old, or invalid.
2370 *
2371 * On exit, it sets things up so the first transaction will work correctly.
2372 * NOTE: only called during fs mount
2373 */
2375 {
2397 /*
2398 * step 1, read in the journal header block. Check the transaction
2399 * it says is the first unflushed, and if that transaction is not
2400 * valid, replay is done
2401 */
2407 }
2412 oldest_start =
2418 "journal-1153: found in "
2419 "header: first_unflushed_offset %d, last_flushed_trans_id "
2424 /*
2425 * now, we try to read the first unflushed offset. If it
2426 * is not valid, there is nothing more we can do, and it
2427 * makes no sense to read through the whole log.
2428 */
2429 d_bh =
2436 }
2439 }
2441 /*
2442 * ok, there are transactions that need to be replayed. start
2443 * with the first log block, find all the valid transactions, and
2444 * pick out the oldest.
2445 */
2450 /*
2451 * Note that it is required for blocksize of primary fs
2452 * device and journal device to be the same
2453 */
2454 d_bh =
2459 ret =
2470 "journal-1179: Setting "
2472 oldest_start -
2473 SB_ONDISK_JOURNAL_1st_BLOCK
2476 /* one we just read was older */
2480 "journal-1180: Resetting "
2482 oldest_start -
2483 SB_ONDISK_JOURNAL_1st_BLOCK
2485 }
2489 "journal-1299: Setting "
2492 }
2495 cur_dblock++;
2496 }
2498 }
2500 start_log_replay:
2504 "journal-1206: Starting replay "
2507 oldest_trans_id);
2509 }
2512 ret =
2519 }
2520 cur_dblock =
2522 replay_count++;
2525 }
2530 }
2531 /*
2532 * j_start does not get set correctly if we don't replay any
2533 * transactions. if we had a valid journal_header, set j_start
2534 * to the first unflushed transaction value, copy the trans_id
2535 * from the header
2536 */
2541 /* check for trans_id overflow */
2549 }
2557 }
2558 /* needed to satisfy the locking in _update_journal_header_block */
2564 /*
2565 * replay failed, caller must call free_journal_ram and abort
2566 * the mount
2567 */
2569 }
2572 }
2575 {
2587 }
2590 {
2592 }
2596 {
2600 }
2601 }
2606 {
2608 dev_t jdev;
2621 /* there is no "jdev" option and journal is on separate device */
2626 journal);
2639 }
2650 }
2657 }
2659 /*
2660 * When creating/tuning a file system user can assign some
2661 * journal params within boundaries which depend on the ratio
2662 * blocksize/standard_blocksize.
2663 *
2664 * For blocks >= standard_blocksize transaction size should
2665 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2666 * then JOURNAL_TRANS_MAX_DEFAULT.
2667 *
2668 * For blocks < standard_blocksize these boundaries should be
2669 * decreased proportionally.
2670 */
2671 #define REISERFS_STANDARD_BLKSIZE (4096)
2675 {
2677 /* Non-default journal params. Do sanity check for them. */
2685 JOURNAL_MIN_RATIO) {
2687 "bad transaction max size (%u). "
2690 }
2694 "bad transaction max batch (%u). "
2697 }
2699 /*
2700 * Default journal params.
2701 * The file system was created by old version
2702 * of mkreiserfs, so some fields contain zeros,
2703 * and we need to advise proper values for them
2704 */
2709 }
2713 }
2715 }
2717 /* must be called once on fs mount. calls journal_read for you */
2720 {
2735 }
2747 /* reserved for journal area support */
2749 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2753 REISERFS_DISK_OFFSET_IN_BYTES /
2756 /*
2757 * Sanity check to see is the standard journal fitting
2758 * within first bitmap (actual for small blocksizes)
2759 */
2764 "journal does not fit for area addressed "
2765 "by first of bitmap blocks. It starts at "
2771 }
2777 }
2781 /* read journal header */
2789 }
2792 /* make sure that journal matches to the super block */
2797 "journal header magic %x (device %s) does "
2804 }
2819 }
2822 "journal first block %u, max trans len %u, max batch %u, "
2867 "allocation failed (%ld bytes). Journal is "
2868 "too large for available memory. Usually "
2872 }
2877 /*
2878 * get_list_bitmap() may call flush_commit_list() which
2879 * requires the lock. Calling flush_commit_list() shouldn't happen
2880 * this early but I like to be paranoid.
2881 */
2889 }
2896 }
2901 free_and_return:
2904 }
2906 /*
2907 * test for a polite end of the current transaction. Used by file_write,
2908 * and should be used by delete to make sure they don't write more than
2909 * can fit inside a single transaction
2910 */
2913 {
2916 /* cannot restart while nested */
2926 }
2931 }
2933 /* this must be called inside a transaction */
2935 {
2941 }
2943 /* this must be called without a transaction started */
2945 {
2949 }
2951 /* this must be called without a transaction started */
2953 {
2957 }
2960 {
2961 wait_queue_t wait;
2965 /*
2966 * we don't want to use wait_event here because
2967 * we only want to wait once.
2968 */
2976 }
2979 }
2982 {
2986 }
2989 {
3004 }
3010 }
3011 }
3013 /*
3014 * join == true if you must join an existing transaction.
3015 * join == false if you can deal with waiting for others to finish
3016 *
3017 * this will block until the transaction is joinable. send the number of
3018 * blocks you expect to use in nblocks.
3019 */
3023 {
3036 /* set here for journal_join */
3040 relock:
3046 }
3056 }
3059 /*
3060 * if there is no room in the journal OR
3061 * if this transaction is too old, and we weren't called joinable,
3062 * wait for it to finish before beginning we don't sleep if there
3063 * aren't other writers
3064 */
3067 (!join
3077 /* allow others to finish this transaction */
3085 sched_count++;
3088 }
3089 }
3090 /*
3091 * don't mess with joining the transaction if all we
3092 * have to do is wait for someone else to do a commit
3093 */
3098 }
3100 }
3105 /* someone might have ended the transaction while we joined */
3110 }
3117 }
3118 /* we are the first writer, set trans_id */
3121 }
3131 out_fail:
3133 /*
3134 * Re-set th->t_super, so we can properly keep track of how many
3135 * persistent transactions there are. We need to do this so if this
3136 * call is part of a failed restart_transaction, we can free it later
3137 */
3140 }
3143 super_block
3146 {
3150 /*
3151 * if we're nesting into an existing transaction. It will be
3152 * persistent on its own
3153 */
3160 }
3168 }
3172 }
3175 {
3180 else
3185 }
3187 }
3191 {
3194 /*
3195 * this keeps do_journal_end from NULLing out the
3196 * current->journal_info pointer
3197 */
3201 }
3205 {
3208 /*
3209 * this keeps do_journal_end from NULLing out the
3210 * current->journal_info pointer
3211 */
3215 }
3219 {
3225 /* we are nesting into the current transaction */
3232 "BAD: refcount <= 1, but "
3236 /*
3237 * we've ended up with a handle from a different
3238 * filesystem. save it and restore on journal_end.
3239 * This should never really happen...
3240 */
3245 }
3248 }
3252 /*
3253 * I guess this boils down to being the reciprocal of clm-2100 above.
3254 * If do_journal_begin_r fails, we need to put it back, since
3255 * journal_end won't be called to do it. */
3258 else
3262 }
3264 /*
3265 * puts bh into the current transaction. If it was already there, reorders
3266 * removes the old pointers from the hash, and puts new ones in (to make
3267 * sure replay happen in the right order).
3268 *
3269 * if it was dirty, cleans and files onto the clean list. I can't let it
3270 * be dirty again until the transaction is committed.
3271 *
3272 * if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3273 */
3276 {
3289 }
3293 /* already in this transaction, we are done */
3297 }
3299 /*
3300 * this must be turned into a panic instead of a warning. We can't
3301 * allow a dirty or journal_dirty or locked buffer to be logged, as
3302 * some changes could get to disk too early. NOT GOOD.
3303 */
3306 "buffer %llu bad state "
3313 }
3320 }
3321 /*
3322 * this error means I've screwed up, and we've overflowed
3323 * the transaction. Nothing can be done here, except make the
3324 * FS readonly or panic.
3325 */
3330 }
3336 }
3340 }
3344 /* now put this guy on the end */
3349 }
3354 }
3365 }
3366 }
3376 }
3379 }
3382 {
3391 }
3398 /*
3399 * we aren't allowed to close a nested transaction on a
3400 * different filesystem from the one in the task struct
3401 */
3407 }
3411 }
3412 }
3414 /*
3415 * removes from the current transaction, relsing and descrementing any counters.
3416 * also files the removed buffer directly onto the clean list
3417 *
3418 * called by journal_mark_freed when a block has been deleted
3419 *
3420 * returns 1 if it cleaned and relsed the buffer. 0 otherwise
3421 */
3424 {
3433 }
3437 }
3440 }
3443 }
3446 }
3460 }
3462 }
3467 }
3469 /*
3470 * for any cnode in a journal list, it can only be dirtied of all the
3471 * transactions that include it are committed to disk.
3472 * this checks through each transaction, and returns 1 if you are allowed
3473 * to dirty, and 0 if you aren't
3474 *
3475 * it is called by dirty_journal_list, which is called after
3476 * flush_commit_list has gotten all the log blocks for a given
3477 * transaction on disk
3478 *
3479 */
3481 {
3487 /*
3488 * first test hprev. These are all newer than cn, so any node here
3489 * with the same block number and dev means this node can't be sent
3490 * to disk right now.
3491 */
3496 }
3498 }
3499 /*
3500 * then test hnext. These are all older than cn. As long as they
3501 * are committed to the log, it is safe to write cn to disk
3502 */
3509 }
3511 }
3513 }
3515 /*
3516 * syncs the commit blocks, but does not force the real buffers to disk
3517 * will wait until the current transaction is done/committed before returning
3518 */
3520 {
3525 /* you can sync while nested, very, very bad */
3531 }
3533 }
3535 /* writeback the pending async commits to disk */
3537 {
3546 /* last entry is the youngest, commit it and you get everything */
3550 }
3552 }
3554 /*
3555 * flushes any old transactions to disk
3556 * ends the current transaction if it is too old
3557 */
3559 {
3565 /*
3566 * safety check so we don't flush while we are replaying the log during
3567 * mount
3568 */
3572 /*
3573 * check the current transaction. If there are no writers, and it is
3574 * too old, finish it, and force the commit blocks to disk
3575 */
3586 /*
3587 * we're only being called from kreiserfsd, it makes
3588 * no sense to do an async commit so that kreiserfsd
3589 * can do it later
3590 */
3592 }
3593 }
3594 }
3596 /*
3597 * returns 0 if do_journal_end should return right away, returns 1 if
3598 * do_journal_end should finish the commit
3599 *
3600 * if the current transaction is too old, but still has writers, this will
3601 * wait on j_join_wait until all the writers are done. By the time it
3602 * wakes up, the transaction it was called has already ended, so it just
3603 * flushes the commit list and returns 0.
3604 *
3605 * Won't batch when flush or commit_now is set. Also won't batch when
3606 * others are waiting on j_join_wait.
3607 *
3608 * Note, we can't allow the journal_end to proceed while there are still
3609 * writers in the log.
3610 */
3612 {
3628 }
3631 /* <= 0 is allowed. unmounting might not call begin */
3635 /*
3636 * BUG, deal with case where j_len is 0, but people previously
3637 * freed blocks need to be released will be dealt with by next
3638 * transaction that actually writes something, but should be taken
3639 * care of in this trans
3640 */
3643 /*
3644 * if wcount > 0, and we are called to with flush or commit_now,
3645 * we wait on j_join_wait. We will wake up when the last writer has
3646 * finished the transaction, and started it on its way to the disk.
3647 * Then, we flush the commit or journal list, and just return 0
3648 * because the rest of journal end was already done for this
3649 * transaction.
3650 */
3662 }
3665 /*
3666 * sleep while the current transaction is
3667 * still j_jlocked
3668 */
3677 }
3679 }
3680 }
3687 }
3689 }
3692 }
3694 /* deal with old transactions where we are the last writers */
3699 }
3700 /* don't batch when someone is waiting on j_join_wait */
3701 /* don't batch when syncing the commit or flushing the whole trans */
3709 }
3715 }
3717 }
3719 /*
3720 * Does all the work that makes deleting blocks safe.
3721 * when deleting a block mark BH_JNew, just remove it from the current
3722 * transaction, clean it's buffer_head and move on.
3723 *
3724 * otherwise:
3725 * set a bit for the block in the journal bitmap. That will prevent it from
3726 * being allocated for unformatted nodes before this transaction has finished.
3727 *
3728 * mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers.
3729 * That will prevent any old transactions with this block from trying to flush
3730 * to the real location. Since we aren't removing the cnode from the
3731 * journal_list_hash, *the block can't be reallocated yet.
3732 *
3733 * Then remove it from the current transaction, decrementing any counters and
3734 * filing it on the clean list.
3735 */
3738 {
3750 }
3751 /* if it is journal new, we just remove it from this transaction */
3758 /*
3759 * set the bit for this block in the journal bitmap
3760 * for this transaction
3761 */
3766 }
3769 /* Note, the entire while loop is not allowed to schedule. */
3774 }
3777 /*
3778 * find all older transactions with this block,
3779 * make sure they don't try to write it out
3780 */
3782 blocknr);
3787 /*
3788 * remove_from_transaction will brelse
3789 * the buffer if it was in the current
3790 * trans
3791 */
3794 bh);
3797 bh);
3805 }
3806 }
3807 /*
3808 * since we are clearing the bh,
3809 * we MUST dec nonzerolen
3810 */
3813 j_nonzerolen);
3814 }
3816 }
3817 }
3819 }
3820 }
3825 }
3828 {
3832 }
3834 /*
3835 * returns -1 on error, 0 if no commits/barriers were done and 1
3836 * if a transaction was actually committed and the barrier was done
3837 */
3840 {
3846 /*
3847 * is it from the current transaction,
3848 * or from an unknown transaction?
3849 */
3852 /*
3853 * try to let other writers come in and
3854 * grow this transaction
3855 */
3859 }
3865 /* someone might have ended this transaction while we joined */
3872 }
3879 /*
3880 * this gets tricky, we have to make sure the journal list in
3881 * the inode still exists. We know the list is still around
3882 * if we've got a larger transaction id than the oldest list
3883 */
3884 flush_commit_only:
3886 /*
3887 * we only set ret to 1 when we know for sure
3888 * the barrier hasn't been started yet on the commit
3889 * block.
3890 */
3896 }
3897 }
3898 /* otherwise the list is gone, and long since committed */
3900 }
3903 {
3907 /*
3908 * for the whole inode, assume unset id means it was
3909 * changed in the current transaction. More conservative
3910 */
3914 /* jl will be updated in __commit_trans_jl */
3915 }
3918 }
3922 {
3927 }
3938 }
3940 }
3942 }
3945 /*
3946 * before we can change a metadata block, we have to make sure it won't
3947 * be written to disk while we are altering it. So, we must:
3948 * clean it
3949 * wait on it.
3950 */
3953 {
3960 }
3965 }
3968 }
3970 /*
3971 * long and ugly. If flush, will not return until all commit
3972 * blocks and all real buffers in the trans are on disk.
3973 * If no_async, won't return until all commit blocks are on disk.
3974 *
3975 * keep reading, there are comments as you go along
3976 *
3977 * If the journal is aborted, we just clean up. Things like flushing
3978 * journal lists, etc just won't happen.
3979 */
3981 {
4006 /*
4007 * protect flush_older_commits from doing mistakes if the
4008 * transaction ID counter gets overflowed.
4009 */
4021 }
4027 }
4031 }
4033 /*
4034 * check_journal_end locks the journal, and unlocks if it does
4035 * not return 1 it tells us if we should continue with the
4036 * journal_end, or just return
4037 */
4043 }
4045 /* check_journal_end might set these, check again */
4048 }
4050 /*
4051 * j must wait means we have to flush the log blocks, and the
4052 * real blocks for this transaction
4053 */
4056 }
4057 #ifdef REISERFS_PREALLOCATE
4058 /*
4059 * quota ops might need to nest, setup the journal_info pointer
4060 * for them and raise the refcount so that it is > 0.
4061 */
4065 /* it should not involve new blocks into the transaction */
4070 #endif
4072 /* setup description block */
4073 d_bh =
4083 /*
4084 * setup commit block. Don't write (keep it clean too) this one
4085 * until after everyone else is written
4086 */
4095 /* init this journal list */
4098 /*
4099 * we lock the commit before doing anything because
4100 * we want to make sure nobody tries to run flush_commit_list until
4101 * the new transaction is fully setup, and we've already flushed the
4102 * ordered bh list
4103 */
4106 /* save the transaction id in case we need to commit it later */
4119 /*
4120 * The ENTIRE FOR LOOP MUST not cause schedule to occur.
4121 * for each real block, add it to the journal list hash,
4122 * copy into real block index array in the commit or desc block
4123 */
4131 }
4134 }
4139 }
4141 /*
4142 * make sure the block we are trying to log
4143 * is not a block of journal or reserved area
4144 */
4148 "Trying to log block %lu, "
4151 }
4164 }
4166 i--;
4167 }
4168 }
4174 /*
4175 * special check in case all buffers in the journal
4176 * were marked for not logging
4177 */
4180 /*
4181 * we're about to dirty all the log blocks, mark the description block
4182 * dirty now too. Don't mark the commit block dirty until all the
4183 * others are on disk
4184 */
4187 /*
4188 * first data block is j_start + 1, so add one to
4189 * cur_write_start wherever you use it
4190 */
4196 /* copy all the real blocks into log area. dirty log blocks */
4201 tmp_bh =
4205 jindex) %
4215 jindex++;
4219 /*
4220 * JDirty cleared sometime during transaction.
4221 * don't log this one
4222 */
4224 "BAD, buffer in journal hash, "
4227 }
4232 }
4234 /*
4235 * we are done with both the c_bh and d_bh, but
4236 * c_bh must be written after all other commit blocks,
4237 * so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4238 */
4242 /* now it is safe to insert this transaction on the main list */
4247 /* reset journal values for the next transaction */
4258 /* check for trans_id overflow */
4268 /*
4269 * make sure reiserfs_add_jh sees the new current_jl before we
4270 * write out the tails
4271 */
4274 /*
4275 * tail conversion targets have to hit the disk before we end the
4276 * transaction. Otherwise a later transaction might repack the tail
4277 * before this transaction commits, leaving the data block unflushed
4278 * and clean, if we crash before the later transaction commits, the
4279 * data block is lost.
4280 */
4286 }
4290 /*
4291 * honor the flush wishes from the caller, simple commits can
4292 * be done outside the journal lock, they are done below
4293 *
4294 * if we don't flush the commit list right now, we put it into
4295 * the work queue so the people waiting on the async progress work
4296 * queue don't wait for this proc to flush journal lists and such.
4297 */
4302 /*
4303 * Avoid queueing work when sb is being shut down. Transaction
4304 * will be flushed on journal shutdown.
4305 */
4309 }
4311 /*
4312 * if the next transaction has any chance of wrapping, flush
4313 * transactions that might get overwritten. If any journal lists
4314 * are very old flush them as well.
4315 */
4316 first_jl:
4327 /*
4328 * if we don't cross into the next
4329 * transaction and we don't wrap, there is
4330 * no way we can overlap any later transactions
4331 * break now
4332 */
4334 }
4344 /*
4345 * we don't overlap anything from out start
4346 * to the end of the log, and our wrapped
4347 * portion doesn't overlap anything at
4348 * the start of the log. We can break
4349 */
4351 }
4352 }
4353 }
4361 }
4365 /* wake up any body waiting to join. */
4372 }
4373 out:
4377 /*
4378 * Re-set th->t_super, so we can properly keep track of how many
4379 * persistent transactions there are. We need to do this so if this
4380 * call is part of a failed restart_transaction, we can free it later
4381 */
4385 }
4387 /* Send the file system read only and refuse new transactions */
4389 {
4400 #ifdef CONFIG_REISERFS_CHECK
4402 #endif
4403 }