| blob | 70057359fbaf3b7d98c8109dd0bcba1fe112b5cb |
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 }
356 }
362 }
365 }
367 /* pulls a cnode off the free list, or returns NULL on failure */
369 {
377 }
383 }
386 }
390 }
392 /*
393 * returns a cnode to the free list
394 */
397 {
404 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
408 }
409 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
411 }
414 {
417 }
419 /*
420 * return a cnode with same dev, block number and size in table,
421 * or null if not found
422 */
424 super_block
426 struct
427 reiserfs_journal_cnode
430 {
437 }
439 }
441 /*
442 * this actually means 'can this block be reallocated yet?'. If you set
443 * search_all, a block can only be allocated if it is not in the current
444 * transaction, was not freed by the current transaction, and has no chance
445 * of ever being overwritten by a replay after crashing.
446 *
447 * If you don't set search_all, a block can only be allocated if it is not
448 * in the current transaction. Since deleting a block removes it from the
449 * current transaction, this case should never happen. If you don't set
450 * search_all, make sure you never write the block without logging it.
451 *
452 * next_zero_bit is a suggestion about the next block to try for find_forward.
453 * when bl is rejected because it is set in a journal list bitmap, we search
454 * for the next zero bit in the bitmap that rejected bl. Then, we return
455 * that through next_zero_bit for find_forward to try.
456 *
457 * Just because we return something in next_zero_bit does not mean we won't
458 * reject it on the next call to reiserfs_in_journal
459 */
463 {
473 /*
474 * If we aren't doing a search_all, this is a metablock, and it
475 * will be logged before use. if we crash before the transaction
476 * that freed it commits, this transaction won't have committed
477 * either, and the block will never be written
478 */
486 data)) {
490 data),
494 }
495 }
496 }
499 /* is it in any old transactions? */
504 }
506 /* is it in the current transaction. This should never happen */
510 }
513 /* safe for reuse */
515 }
517 /* insert cn into table */
520 {
528 }
530 }
532 /* lock the current transaction */
534 {
538 }
540 /* unlock the current transaction */
542 {
544 }
547 {
549 }
553 {
557 }
560 }
562 /*
563 * this used to be much more involved, and I'm keeping it just in case
564 * things get ugly again. it gets called by flush_commit_list, and
565 * cleans up any data stored about blocks freed during a transaction.
566 */
569 {
574 }
577 }
581 {
590 }
591 }
593 }
595 /*
596 * If page->mapping was null, we failed to truncate this page for
597 * some reason. Most likely because it was truncated after being
598 * logged via data=journal.
599 *
600 * This does a check to see if the buffer belongs to one of these
601 * lost pages before doing the final put_bh. If page->mapping was
602 * null, it tries to free buffers on the page, which should make the
603 * final put_page drop the page from the lru.
604 */
606 {
617 }
618 }
621 {
626 }
629 else
634 }
637 {
640 else
644 }
647 {
657 }
660 {
667 }
669 #define CHUNK_SIZE 32
673 };
676 {
680 }
682 }
685 {
689 }
691 }
695 {
707 }
708 }
710 }
714 {
721 }
723 }
724 }
726 /*
727 * we want to free the jh when the buffer has been written
728 * and waited on
729 */
731 {
744 }
745 }
749 {
757 }
761 no_jh:
765 /*
766 * buffer must be locked for __add_jh, should be able to have
767 * two adds at the same time
768 */
772 }
778 }
781 }
784 {
786 }
788 {
790 }
792 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
797 {
815 }
823 }
824 /*
825 * in theory, dirty non-uptodate buffers should never get here,
826 * but the upper layer io error paths still have a few quirks.
827 * Handle them here as gracefully as we can
828 */
832 }
839 }
840 loop_next:
843 }
848 }
859 }
862 }
863 /*
864 * ugly interaction with invalidatepage here.
865 * reiserfs_invalidate_page will pin any buffer that has a
866 * valid journal head from an older transaction. If someone
867 * else sets our buffer dirty after we write it in the first
868 * loop, and then someone truncates the page away, nobody
869 * will ever write the buffer. We're safe if we write the
870 * page one last time after freeing the journal header.
871 */
876 }
879 }
882 }
886 {
895 find_first:
896 /*
897 * first we walk backwards to find the oldest uncommitted transation
898 */
909 }
911 /* if we didn't find any older uncommitted transactions, return now */
914 }
927 /* list we were called with is gone, return */
931 /*
932 * the one we just flushed is gone, this means
933 * all older lists are also gone, so first_jl
934 * is no longer valid either. Go back to the
935 * beginning.
936 */
940 }
941 }
947 }
948 }
950 }
953 {
962 }
965 }
967 /*
968 * if this journal list still has commit blocks unflushed, send them to disk.
969 *
970 * log areas must be flushed in order (transaction 2 can't commit before
971 * transaction 1) Before the commit block can by written, every other log
972 * block must be safely on disk
973 */
976 {
978 b_blocknr_t bn;
990 }
992 /*
993 * before we can put our commit blocks on disk, we have to make
994 * sure everyone older than us is on disk too
995 */
1002 /*
1003 * list disappeared during flush_older_commits.
1004 * return
1005 */
1007 }
1008 }
1010 /* make sure nobody is trying to flush this one at the same time */
1016 }
1019 /* this commit is done, exit */
1023 }
1026 }
1031 /*
1032 * We might sleep in numerous places inside
1033 * write_ordered_buffers. Relax the write lock.
1034 */
1041 }
1043 /*
1044 * for the description block and all the log blocks, submit any buffers
1045 * that haven't already reached the disk. Try to write at least 256
1046 * log blocks. later on, we will only wait on blocks that correspond
1047 * to this transaction, but while we're unplugging we might as well
1048 * get a chunk of data on there.
1049 */
1063 }
1065 }
1066 }
1077 /*
1078 * since we're using ll_rw_blk above, it might have skipped
1079 * over a locked buffer. Double check here
1080 */
1081 /* redundant, sync_dirty_buffer() checks */
1086 }
1088 #ifdef CONFIG_REISERFS_CHECK
1091 #endif
1093 }
1094 /* once for journal_find_get_block */
1096 /* once due to original getblk in do_journal_end */
1099 }
1103 /*
1104 * If there was a write error in the journal - we can't commit
1105 * this transaction - it will be invalid and, if successful,
1106 * will just end up propagating the write error out to
1107 * the file system.
1108 */
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);
1276 else
1284 }
1285 }
1287 }
1292 {
1294 }
1296 /*
1297 ** flush any and all journal lists older than you are
1298 ** can only be called from flush_journal_list
1299 */
1302 {
1308 /*
1309 * we know we are the only ones flushing things, no extra race
1310 * protection is required.
1311 */
1312 restart:
1314 /* Did we wrap? */
1320 /* do not flush all */
1323 /* other_jl is now deleted from the list */
1325 }
1327 }
1331 {
1336 }
1337 }
1339 /*
1340 * flush a journal list, both commit and real blocks
1341 *
1342 * always set flushall to 1, unless you are calling from inside
1343 * flush_journal_list
1344 *
1345 * IMPORTANT. This can only be called while there are no journal writers,
1346 * and the journal is locked. That means it can only be called from
1347 * do_journal_end, or by journal_release
1348 */
1351 {
1368 }
1370 /* if flushall == 0, the lock is already held */
1375 }
1382 }
1384 /* if all the work is already done, get out of here */
1388 }
1390 /*
1391 * start by putting the commit list on disk. This will also flush
1392 * the commit lists of any olders transactions
1393 */
1400 /* are we done now? */
1404 }
1406 /*
1407 * loop through each cnode, see if we need to write it,
1408 * or wait on a more recent transaction, or just ignore it
1409 */
1413 }
1419 /* blocknr of 0 is no longer in the hash, ignore it */
1422 }
1424 /*
1425 * This transaction failed commit.
1426 * Don't write out to the disk
1427 */
1432 /*
1433 * the order is important here. We check pjl to make sure we
1434 * don't clear BH_JDirty_wait if we aren't the one writing this
1435 * block to disk
1436 */
1440 /*
1441 * we do this to make sure nobody releases the
1442 * buffer while we are working with it
1443 */
1451 /*
1452 * everything with !pjl && jwait
1453 * should be writable
1454 */
1456 }
1457 }
1459 /*
1460 * if someone has this block in a newer transaction, just make
1461 * sure they are committed, and don't try writing it to disk
1462 */
1467 }
1469 /*
1470 * bh == NULL when the block got to disk on its own, OR,
1471 * the block got freed in a future transaction
1472 */
1475 }
1477 /*
1478 * this should never happen. kupdate_one_transaction has
1479 * this list locked while it works, so we should never see a
1480 * buffer here that is not marked JDirty_wait
1481 */
1484 "BAD! buffer %llu %cdirty %cjwait, "
1489 }
1491 /*
1492 * we inc again because saved_bh gets decremented
1493 * at free_cnode
1494 */
1501 else
1503 count++;
1509 }
1510 free_cnode:
1514 /*
1515 * we incremented this to keep others from
1516 * taking the buffer head away
1517 */
1522 }
1523 }
1524 }
1532 }
1541 }
1543 #ifdef CONFIG_REISERFS_CHECK
1546 #endif
1548 }
1549 /*
1550 * note, we must clear the JDirty_wait bit
1551 * after the up to date check, otherwise we
1552 * race against our flushpage routine
1553 */
1557 /* drop one ref for us */
1559 /* drop one ref for journal_mark_dirty */
1561 }
1563 }
1564 }
1569 __func__);
1570 flush_older_and_return:
1572 /*
1573 * before we can update the journal header block, we _must_ flush all
1574 * real blocks from all older transactions to disk. This is because
1575 * once the header block is updated, this transaction will not be
1576 * replayed after a crash
1577 */
1580 }
1583 /*
1584 * before we can remove everything from the hash tables for this
1585 * transaction, we must make sure it can never be replayed
1586 *
1587 * since we are only called from do_journal_end, we know for sure there
1588 * are no allocations going on while we are flushing journal lists. So,
1589 * we only need to update the journal header block for the last list
1590 * being flushed
1591 */
1593 err =
1601 __func__);
1602 }
1612 }
1615 /*
1616 * not strictly required since we are freeing the list, but it should
1617 * help find code using dead lists later on
1618 */
1630 }
1635 {
1643 }
1647 /*
1648 * if the blocknr == 0, this has been cleared from the hash,
1649 * skip it
1650 */
1653 }
1656 /*
1657 * we can race against journal_mark_freed when we try
1658 * to lock_buffer(cn->bh), so we have to inc the buffer
1659 * count, and recheck things after locking
1660 */
1669 ret++;
1671 /* note, cn->bh might be null now */
1673 }
1675 }
1676 next:
1679 }
1681 }
1683 /* used by flush_commit_list */
1686 {
1694 /*
1695 * look for a more recent transaction that logged this
1696 * buffer. Only the most recent transaction with a buffer in
1697 * it is allowed to send that buffer to disk
1698 */
1703 /*
1704 * if the buffer is prepared, it will either be logged
1705 * or restored. If restored, we need to make sure
1706 * it actually gets marked dirty
1707 */
1714 }
1715 }
1717 }
1719 }
1726 {
1739 }
1741 /*
1742 * we've got j_flush_mutex held, nobody is going to delete any
1743 * of these lists out from underneath us
1744 */
1753 }
1758 transactions_flushed++;
1762 /* did we wrap? */
1765 }
1768 /* don't bother with older transactions */
1771 }
1774 }
1776 done:
1779 }
1781 /*
1782 * for o_sync and fsync heavy applications, they tend to use
1783 * all the journa list slots with tiny transactions. These
1784 * trigger lots and lots of calls to update the header block, which
1785 * adds seeks and slows things down.
1786 *
1787 * This function tries to clear out a large chunk of the journal lists
1788 * at once, which makes everything faster since only the newest journal
1789 * list updates the header block
1790 */
1793 {
1806 /* in data logging mode, try harder to flush a lot of blocks */
1809 /* flush for 256 transactions or limit blocks, whichever comes first */
1814 }
1818 }
1824 }
1828 /*
1829 * try to find a group of blocks we can flush across all the
1830 * transactions, but only bother if we've actually spanned
1831 * across multiple lists
1832 */
1835 }
1840 }
1842 /*
1843 * removes any nodes in table with name block and dev as bh.
1844 * only touchs the hnext and hprev pointers.
1845 */
1850 {
1857 }
1865 }
1870 }
1874 /*
1875 * anybody who clears the cur->bh will also
1876 * dec the nonzerolen
1877 */
1882 }
1884 }
1885 }
1888 {
1898 }
1899 /*
1900 * j_header_bh is on the journal dev, make sure
1901 * not to release the journal dev until we brelse j_header_bh
1902 */
1905 }
1907 /*
1908 * call on unmount. Only set error to 1 if you haven't made your way out
1909 * of read_super() yet. Any other caller must keep error at 0.
1910 */
1913 {
1918 /*
1919 * we only want to flush out transactions if we were
1920 * called with error == 0
1921 */
1923 /* end the current trans */
1927 /*
1928 * make sure something gets logged to force
1929 * our way into the flush code
1930 */
1938 }
1939 }
1941 /* this also catches errors during the do_journal_end above */
1950 }
1951 }
1954 /*
1955 * We must release the write lock here because
1956 * the workqueue job (flush_async_commit) needs this lock
1957 */
1960 /*
1961 * Cancel flushing of old commits. Note that neither of these works
1962 * will be requeued because superblock is being shutdown and doesn't
1963 * have SB_ACTIVE set.
1964 */
1966 /* wait for all commits to finish */
1974 }
1976 /* * call on unmount. flush all journal trans, release all alloc'd ram */
1979 {
1981 }
1983 /* only call from an error condition inside reiserfs_read_super! */
1986 {
1988 }
1990 /*
1991 * compares description block with commit block.
1992 * returns 1 if they differ, 0 if they are the same
1993 */
1997 {
2003 }
2005 }
2007 /*
2008 * returns 0 if it did not find a description block
2009 * returns -1 if it found a corrupt commit block
2010 * returns 1 if both desc and commit were valid
2011 * NOTE: only called during fs mount
2012 */
2017 {
2032 "journal-986: transaction "
2033 "is valid returning because trans_id %d is greater than "
2038 }
2042 "journal-1087: transaction "
2043 "is valid returning because mount_id %d is less than "
2048 }
2051 "Bad transaction length %d "
2055 }
2058 /*
2059 * ok, we have a journal description block,
2060 * let's see if the transaction was valid
2061 */
2062 c_bh =
2072 "journal_transaction_is_valid, commit offset %ld had bad "
2083 "journal-1004: "
2084 "transaction_is_valid setting oldest invalid trans_id "
2087 }
2089 }
2092 "journal-1006: found valid "
2101 }
2102 }
2105 {
2109 }
2110 }
2112 /*
2113 * given the start, and values for the oldest acceptable transactions,
2114 * this either reads in a replays a transaction, or returns because the
2115 * transaction is invalid, or too old.
2116 * NOTE: only called during fs mount
2117 */
2123 {
2148 cur_dblock -
2152 }
2155 "journal_read_trans skipping because %d is != "
2157 newest_mount_id);
2160 }
2167 }
2171 "journal_read_transaction, "
2180 }
2188 }
2191 /*
2192 * now we know we've got a good transaction, and it was
2193 * inside the valid time ranges
2194 */
2207 }
2208 /* get all the buffer heads */
2225 }
2228 "REPLAY FAILURE fsck required! "
2229 "Block to replay is outside of "
2232 }
2233 /* make sure we don't try to replay onto log or reserved area */
2237 "REPLAY FAILURE fsck required! "
2239 abort_replay:
2247 }
2248 }
2249 /* read in the log blocks, memcpy to the corresponding real block */
2256 "REPLAY FAILURE fsck required! "
2266 }
2271 }
2272 /* flush out the real blocks */
2276 }
2281 "REPLAY FAILURE, fsck required! "
2290 }
2292 }
2293 cur_dblock =
2301 /*
2302 * init starting values for the first transaction, in case
2303 * this is the last transaction to be replayed.
2304 */
2308 /* check for trans_id overflow */
2316 }
2318 /*
2319 * This function reads blocks starting from block and to max_block of bufsize
2320 * size (but no more than BUFNR blocks at a time). This proved to improve
2321 * mounting speed on self-rebuilding raid5 arrays at least.
2322 * Right now it is only used from journal code. But later we might use it
2323 * from other places.
2324 * Note: Do not use journal_getblk/sb_getblk functions here!
2325 */
2328 b_blocknr_t max_block)
2329 {
2341 }
2351 }
2361 }
2363 /*
2364 * read and replay the log
2365 * on a clean unmount, the journal header's next unflushed pointer will be
2366 * to an invalid transaction. This tests that before finding all the
2367 * transactions in the log, which makes normal mount times fast.
2368 *
2369 * After a crash, this starts with the next unflushed transaction, and
2370 * replays until it finds one too old, or invalid.
2371 *
2372 * On exit, it sets things up so the first transaction will work correctly.
2373 * NOTE: only called during fs mount
2374 */
2376 {
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 {
2734 }
2746 /* reserved for journal area support */
2748 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2752 REISERFS_DISK_OFFSET_IN_BYTES /
2755 /*
2756 * Sanity check to see is the standard journal fitting
2757 * within first bitmap (actual for small blocksizes)
2758 */
2763 "journal does not fit for area addressed "
2764 "by first of bitmap blocks. It starts at "
2770 }
2776 }
2780 /* read journal header */
2788 }
2791 /* make sure that journal matches to the super block */
2796 "journal header magic %x (device %pg) does "
2803 }
2818 }
2821 "journal first block %u, max trans len %u, max batch %u, "
2866 "allocation failed (%ld bytes). Journal is "
2867 "too large for available memory. Usually "
2871 }
2876 /*
2877 * get_list_bitmap() may call flush_commit_list() which
2878 * requires the lock. Calling flush_commit_list() shouldn't happen
2879 * this early but I like to be paranoid.
2880 */
2888 }
2895 }
2900 free_and_return:
2903 }
2905 /*
2906 * test for a polite end of the current transaction. Used by file_write,
2907 * and should be used by delete to make sure they don't write more than
2908 * can fit inside a single transaction
2909 */
2912 {
2915 /* cannot restart while nested */
2925 }
2930 }
2932 /* this must be called inside a transaction */
2934 {
2940 }
2942 /* this must be called without a transaction started */
2944 {
2948 }
2950 /* this must be called without a transaction started */
2952 {
2956 }
2959 {
2960 wait_queue_entry_t wait;
2964 /*
2965 * we don't want to use wait_event here because
2966 * we only want to wait once.
2967 */
2975 }
2978 }
2981 {
2985 }
2988 {
3003 }
3009 }
3010 }
3012 /*
3013 * join == true if you must join an existing transaction.
3014 * join == false if you can deal with waiting for others to finish
3015 *
3016 * this will block until the transaction is joinable. send the number of
3017 * blocks you expect to use in nblocks.
3018 */
3022 {
3035 /* set here for journal_join */
3039 relock:
3045 }
3055 }
3058 /*
3059 * if there is no room in the journal OR
3060 * if this transaction is too old, and we weren't called joinable,
3061 * wait for it to finish before beginning we don't sleep if there
3062 * aren't other writers
3063 */
3066 (!join
3076 /* allow others to finish this transaction */
3084 sched_count++;
3087 }
3088 }
3089 /*
3090 * don't mess with joining the transaction if all we
3091 * have to do is wait for someone else to do a commit
3092 */
3097 }
3099 }
3104 /* someone might have ended the transaction while we joined */
3109 }
3116 }
3117 /* we are the first writer, set trans_id */
3120 }
3130 out_fail:
3132 /*
3133 * Re-set th->t_super, so we can properly keep track of how many
3134 * persistent transactions there are. We need to do this so if this
3135 * call is part of a failed restart_transaction, we can free it later
3136 */
3139 }
3142 super_block
3145 {
3149 /*
3150 * if we're nesting into an existing transaction. It will be
3151 * persistent on its own
3152 */
3159 }
3167 }
3171 }
3174 {
3179 else
3184 }
3186 }
3190 {
3193 /*
3194 * this keeps do_journal_end from NULLing out the
3195 * current->journal_info pointer
3196 */
3200 }
3204 {
3207 /*
3208 * this keeps do_journal_end from NULLing out the
3209 * current->journal_info pointer
3210 */
3214 }
3218 {
3224 /* we are nesting into the current transaction */
3231 "BAD: refcount <= 1, but "
3235 /*
3236 * we've ended up with a handle from a different
3237 * filesystem. save it and restore on journal_end.
3238 * This should never really happen...
3239 */
3244 }
3247 }
3251 /*
3252 * I guess this boils down to being the reciprocal of clm-2100 above.
3253 * If do_journal_begin_r fails, we need to put it back, since
3254 * journal_end won't be called to do it. */
3257 else
3261 }
3263 /*
3264 * puts bh into the current transaction. If it was already there, reorders
3265 * removes the old pointers from the hash, and puts new ones in (to make
3266 * sure replay happen in the right order).
3267 *
3268 * if it was dirty, cleans and files onto the clean list. I can't let it
3269 * be dirty again until the transaction is committed.
3270 *
3271 * if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3272 */
3275 {
3288 }
3292 /* already in this transaction, we are done */
3296 }
3298 /*
3299 * this must be turned into a panic instead of a warning. We can't
3300 * allow a dirty or journal_dirty or locked buffer to be logged, as
3301 * some changes could get to disk too early. NOT GOOD.
3302 */
3305 "buffer %llu bad state "
3312 }
3319 }
3320 /*
3321 * this error means I've screwed up, and we've overflowed
3322 * the transaction. Nothing can be done here, except make the
3323 * FS readonly or panic.
3324 */
3329 }
3335 }
3339 }
3343 /* now put this guy on the end */
3348 }
3353 }
3364 }
3365 }
3375 }
3378 }
3381 {
3390 }
3397 /*
3398 * we aren't allowed to close a nested transaction on a
3399 * different filesystem from the one in the task struct
3400 */
3406 }
3410 }
3411 }
3413 /*
3414 * removes from the current transaction, relsing and descrementing any counters.
3415 * also files the removed buffer directly onto the clean list
3416 *
3417 * called by journal_mark_freed when a block has been deleted
3418 *
3419 * returns 1 if it cleaned and relsed the buffer. 0 otherwise
3420 */
3423 {
3432 }
3436 }
3439 }
3442 }
3445 }
3459 }
3461 }
3466 }
3468 /*
3469 * for any cnode in a journal list, it can only be dirtied of all the
3470 * transactions that include it are committed to disk.
3471 * this checks through each transaction, and returns 1 if you are allowed
3472 * to dirty, and 0 if you aren't
3473 *
3474 * it is called by dirty_journal_list, which is called after
3475 * flush_commit_list has gotten all the log blocks for a given
3476 * transaction on disk
3477 *
3478 */
3480 {
3486 /*
3487 * first test hprev. These are all newer than cn, so any node here
3488 * with the same block number and dev means this node can't be sent
3489 * to disk right now.
3490 */
3495 }
3497 }
3498 /*
3499 * then test hnext. These are all older than cn. As long as they
3500 * are committed to the log, it is safe to write cn to disk
3501 */
3508 }
3510 }
3512 }
3514 /*
3515 * syncs the commit blocks, but does not force the real buffers to disk
3516 * will wait until the current transaction is done/committed before returning
3517 */
3519 {
3524 /* you can sync while nested, very, very bad */
3530 }
3532 }
3534 /* writeback the pending async commits to disk */
3536 {
3545 /* last entry is the youngest, commit it and you get everything */
3549 }
3551 }
3553 /*
3554 * flushes any old transactions to disk
3555 * ends the current transaction if it is too old
3556 */
3558 {
3564 /*
3565 * safety check so we don't flush while we are replaying the log during
3566 * mount
3567 */
3571 /*
3572 * check the current transaction. If there are no writers, and it is
3573 * too old, finish it, and force the commit blocks to disk
3574 */
3585 /*
3586 * we're only being called from kreiserfsd, it makes
3587 * no sense to do an async commit so that kreiserfsd
3588 * can do it later
3589 */
3591 }
3592 }
3593 }
3595 /*
3596 * returns 0 if do_journal_end should return right away, returns 1 if
3597 * do_journal_end should finish the commit
3598 *
3599 * if the current transaction is too old, but still has writers, this will
3600 * wait on j_join_wait until all the writers are done. By the time it
3601 * wakes up, the transaction it was called has already ended, so it just
3602 * flushes the commit list and returns 0.
3603 *
3604 * Won't batch when flush or commit_now is set. Also won't batch when
3605 * others are waiting on j_join_wait.
3606 *
3607 * Note, we can't allow the journal_end to proceed while there are still
3608 * writers in the log.
3609 */
3611 {
3627 }
3630 /* <= 0 is allowed. unmounting might not call begin */
3634 /*
3635 * BUG, deal with case where j_len is 0, but people previously
3636 * freed blocks need to be released will be dealt with by next
3637 * transaction that actually writes something, but should be taken
3638 * care of in this trans
3639 */
3642 /*
3643 * if wcount > 0, and we are called to with flush or commit_now,
3644 * we wait on j_join_wait. We will wake up when the last writer has
3645 * finished the transaction, and started it on its way to the disk.
3646 * Then, we flush the commit or journal list, and just return 0
3647 * because the rest of journal end was already done for this
3648 * transaction.
3649 */
3661 }
3664 /*
3665 * sleep while the current transaction is
3666 * still j_jlocked
3667 */
3676 }
3678 }
3679 }
3686 }
3688 }
3691 }
3693 /* deal with old transactions where we are the last writers */
3698 }
3699 /* don't batch when someone is waiting on j_join_wait */
3700 /* don't batch when syncing the commit or flushing the whole trans */
3708 }
3714 }
3716 }
3718 /*
3719 * Does all the work that makes deleting blocks safe.
3720 * when deleting a block mark BH_JNew, just remove it from the current
3721 * transaction, clean it's buffer_head and move on.
3722 *
3723 * otherwise:
3724 * set a bit for the block in the journal bitmap. That will prevent it from
3725 * being allocated for unformatted nodes before this transaction has finished.
3726 *
3727 * mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers.
3728 * That will prevent any old transactions with this block from trying to flush
3729 * to the real location. Since we aren't removing the cnode from the
3730 * journal_list_hash, *the block can't be reallocated yet.
3731 *
3732 * Then remove it from the current transaction, decrementing any counters and
3733 * filing it on the clean list.
3734 */
3737 {
3749 }
3750 /* if it is journal new, we just remove it from this transaction */
3757 /*
3758 * set the bit for this block in the journal bitmap
3759 * for this transaction
3760 */
3765 }
3768 /* Note, the entire while loop is not allowed to schedule. */
3773 }
3776 /*
3777 * find all older transactions with this block,
3778 * make sure they don't try to write it out
3779 */
3781 blocknr);
3786 /*
3787 * remove_from_transaction will brelse
3788 * the buffer if it was in the current
3789 * trans
3790 */
3793 bh);
3796 bh);
3804 }
3805 }
3806 /*
3807 * since we are clearing the bh,
3808 * we MUST dec nonzerolen
3809 */
3812 j_nonzerolen);
3813 }
3815 }
3816 }
3818 }
3819 }
3824 }
3827 {
3831 }
3833 /*
3834 * returns -1 on error, 0 if no commits/barriers were done and 1
3835 * if a transaction was actually committed and the barrier was done
3836 */
3839 {
3845 /*
3846 * is it from the current transaction,
3847 * or from an unknown transaction?
3848 */
3851 /*
3852 * try to let other writers come in and
3853 * grow this transaction
3854 */
3858 }
3864 /* someone might have ended this transaction while we joined */
3871 }
3878 /*
3879 * this gets tricky, we have to make sure the journal list in
3880 * the inode still exists. We know the list is still around
3881 * if we've got a larger transaction id than the oldest list
3882 */
3883 flush_commit_only:
3885 /*
3886 * we only set ret to 1 when we know for sure
3887 * the barrier hasn't been started yet on the commit
3888 * block.
3889 */
3895 }
3896 }
3897 /* otherwise the list is gone, and long since committed */
3899 }
3902 {
3906 /*
3907 * for the whole inode, assume unset id means it was
3908 * changed in the current transaction. More conservative
3909 */
3913 /* jl will be updated in __commit_trans_jl */
3914 }
3917 }
3921 {
3926 }
3937 }
3939 }
3941 }
3944 /*
3945 * before we can change a metadata block, we have to make sure it won't
3946 * be written to disk while we are altering it. So, we must:
3947 * clean it
3948 * wait on it.
3949 */
3952 {
3959 }
3964 }
3967 }
3969 /*
3970 * long and ugly. If flush, will not return until all commit
3971 * blocks and all real buffers in the trans are on disk.
3972 * If no_async, won't return until all commit blocks are on disk.
3973 *
3974 * keep reading, there are comments as you go along
3975 *
3976 * If the journal is aborted, we just clean up. Things like flushing
3977 * journal lists, etc just won't happen.
3978 */
3980 {
4005 /*
4006 * protect flush_older_commits from doing mistakes if the
4007 * transaction ID counter gets overflowed.
4008 */
4020 }
4026 }
4030 }
4032 /*
4033 * check_journal_end locks the journal, and unlocks if it does
4034 * not return 1 it tells us if we should continue with the
4035 * journal_end, or just return
4036 */
4042 }
4044 /* check_journal_end might set these, check again */
4047 }
4049 /*
4050 * j must wait means we have to flush the log blocks, and the
4051 * real blocks for this transaction
4052 */
4055 }
4056 #ifdef REISERFS_PREALLOCATE
4057 /*
4058 * quota ops might need to nest, setup the journal_info pointer
4059 * for them and raise the refcount so that it is > 0.
4060 */
4064 /* it should not involve new blocks into the transaction */
4069 #endif
4071 /* setup description block */
4072 d_bh =
4082 /*
4083 * setup commit block. Don't write (keep it clean too) this one
4084 * until after everyone else is written
4085 */
4094 /* init this journal list */
4097 /*
4098 * we lock the commit before doing anything because
4099 * we want to make sure nobody tries to run flush_commit_list until
4100 * the new transaction is fully setup, and we've already flushed the
4101 * ordered bh list
4102 */
4105 /* save the transaction id in case we need to commit it later */
4118 /*
4119 * The ENTIRE FOR LOOP MUST not cause schedule to occur.
4120 * for each real block, add it to the journal list hash,
4121 * copy into real block index array in the commit or desc block
4122 */
4130 }
4133 }
4138 }
4140 /*
4141 * make sure the block we are trying to log
4142 * is not a block of journal or reserved area
4143 */
4147 "Trying to log block %lu, "
4150 }
4163 }
4165 i--;
4166 }
4167 }
4173 /*
4174 * special check in case all buffers in the journal
4175 * were marked for not logging
4176 */
4179 /*
4180 * we're about to dirty all the log blocks, mark the description block
4181 * dirty now too. Don't mark the commit block dirty until all the
4182 * others are on disk
4183 */
4186 /*
4187 * first data block is j_start + 1, so add one to
4188 * cur_write_start wherever you use it
4189 */
4195 /* copy all the real blocks into log area. dirty log blocks */
4200 tmp_bh =
4204 jindex) %
4214 jindex++;
4218 /*
4219 * JDirty cleared sometime during transaction.
4220 * don't log this one
4221 */
4223 "BAD, buffer in journal hash, "
4226 }
4231 }
4233 /*
4234 * we are done with both the c_bh and d_bh, but
4235 * c_bh must be written after all other commit blocks,
4236 * so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4237 */
4241 /* now it is safe to insert this transaction on the main list */
4246 /* reset journal values for the next transaction */
4257 /* check for trans_id overflow */
4267 /*
4268 * make sure reiserfs_add_jh sees the new current_jl before we
4269 * write out the tails
4270 */
4273 /*
4274 * tail conversion targets have to hit the disk before we end the
4275 * transaction. Otherwise a later transaction might repack the tail
4276 * before this transaction commits, leaving the data block unflushed
4277 * and clean, if we crash before the later transaction commits, the
4278 * data block is lost.
4279 */
4285 }
4289 /*
4290 * honor the flush wishes from the caller, simple commits can
4291 * be done outside the journal lock, they are done below
4292 *
4293 * if we don't flush the commit list right now, we put it into
4294 * the work queue so the people waiting on the async progress work
4295 * queue don't wait for this proc to flush journal lists and such.
4296 */
4301 /*
4302 * Avoid queueing work when sb is being shut down. Transaction
4303 * will be flushed on journal shutdown.
4304 */
4308 }
4310 /*
4311 * if the next transaction has any chance of wrapping, flush
4312 * transactions that might get overwritten. If any journal lists
4313 * are very old flush them as well.
4314 */
4315 first_jl:
4326 /*
4327 * if we don't cross into the next
4328 * transaction and we don't wrap, there is
4329 * no way we can overlap any later transactions
4330 * break now
4331 */
4333 }
4343 /*
4344 * we don't overlap anything from out start
4345 * to the end of the log, and our wrapped
4346 * portion doesn't overlap anything at
4347 * the start of the log. We can break
4348 */
4350 }
4351 }
4352 }
4360 }
4364 /* wake up any body waiting to join. */
4371 }
4372 out:
4376 /*
4377 * Re-set th->t_super, so we can properly keep track of how many
4378 * persistent transactions there are. We need to do this so if this
4379 * call is part of a failed restart_transaction, we can free it later
4380 */
4384 }
4386 /* Send the file system read only and refuse new transactions */
4388 {
4399 #ifdef CONFIG_REISERFS_CHECK
4401 #endif
4402 }