| blob | 2ace90e981f07a51c69c38cc16b415e221ca35d0 |
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 put_page drop the page from the lru.
603 */
605 {
616 }
617 }
620 {
625 }
628 else
633 }
636 {
639 else
643 }
646 {
656 }
659 {
666 }
668 #define CHUNK_SIZE 32
672 };
675 {
679 }
681 }
684 {
688 }
690 }
694 {
706 }
707 }
709 }
713 {
720 }
722 }
723 }
725 /*
726 * we want to free the jh when the buffer has been written
727 * and waited on
728 */
730 {
743 }
744 }
748 {
756 }
760 no_jh:
764 /*
765 * buffer must be locked for __add_jh, should be able to have
766 * two adds at the same time
767 */
771 }
777 }
780 }
783 {
785 }
787 {
789 }
791 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
796 {
814 }
822 }
823 /*
824 * in theory, dirty non-uptodate buffers should never get here,
825 * but the upper layer io error paths still have a few quirks.
826 * Handle them here as gracefully as we can
827 */
831 }
838 }
839 loop_next:
842 }
847 }
858 }
861 }
862 /*
863 * ugly interaction with invalidatepage here.
864 * reiserfs_invalidate_page will pin any buffer that has a
865 * valid journal head from an older transaction. If someone
866 * else sets our buffer dirty after we write it in the first
867 * loop, and then someone truncates the page away, nobody
868 * will ever write the buffer. We're safe if we write the
869 * page one last time after freeing the journal header.
870 */
875 }
878 }
881 }
885 {
894 find_first:
895 /*
896 * first we walk backwards to find the oldest uncommitted transation
897 */
908 }
910 /* if we didn't find any older uncommitted transactions, return now */
913 }
926 /* list we were called with is gone, return */
930 /*
931 * the one we just flushed is gone, this means
932 * all older lists are also gone, so first_jl
933 * is no longer valid either. Go back to the
934 * beginning.
935 */
939 }
940 }
946 }
947 }
949 }
952 {
961 }
964 }
966 /*
967 * if this journal list still has commit blocks unflushed, send them to disk.
968 *
969 * log areas must be flushed in order (transaction 2 can't commit before
970 * transaction 1) Before the commit block can by written, every other log
971 * block must be safely on disk
972 */
975 {
977 b_blocknr_t bn;
989 }
991 /*
992 * before we can put our commit blocks on disk, we have to make
993 * sure everyone older than us is on disk too
994 */
1001 /*
1002 * list disappeared during flush_older_commits.
1003 * return
1004 */
1006 }
1007 }
1009 /* make sure nobody is trying to flush this one at the same time */
1015 }
1018 /* this commit is done, exit */
1022 }
1025 }
1030 /*
1031 * We might sleep in numerous places inside
1032 * write_ordered_buffers. Relax the write lock.
1033 */
1040 }
1042 /*
1043 * for the description block and all the log blocks, submit any buffers
1044 * that haven't already reached the disk. Try to write at least 256
1045 * log blocks. later on, we will only wait on blocks that correspond
1046 * to this transaction, but while we're unplugging we might as well
1047 * get a chunk of data on there.
1048 */
1062 }
1064 }
1065 }
1076 /*
1077 * since we're using ll_rw_blk above, it might have skipped
1078 * over a locked buffer. Double check here
1079 */
1080 /* redundant, sync_dirty_buffer() checks */
1085 }
1087 #ifdef CONFIG_REISERFS_CHECK
1090 #endif
1092 }
1093 /* once for journal_find_get_block */
1095 /* once due to original getblk in do_journal_end */
1098 }
1102 /*
1103 * If there was a write error in the journal - we can't commit
1104 * this transaction - it will be invalid and, if successful,
1105 * will just end up propagating the write error out to
1106 * the file system.
1107 */
1115 else
1118 }
1120 /*
1121 * If there was a write error in the journal - we can't commit this
1122 * transaction - it will be invalid and, if successful, will just end
1123 * up propagating the write error out to the filesystem.
1124 */
1126 #ifdef CONFIG_REISERFS_CHECK
1128 #endif
1130 }
1136 }
1139 /*
1140 * now, every commit block is on the disk. It is safe to allow
1141 * blocks freed during this transaction to be reallocated
1142 */
1147 /* mark the metadata dirty */
1154 }
1156 put_jl:
1161 __func__);
1163 }
1165 /*
1166 * flush_journal_list frequently needs to find a newer transaction for a
1167 * given block. This does that, or returns NULL if it can't find anything
1168 */
1170 reiserfs_journal_cnode
1172 {
1180 }
1182 }
1184 }
1191 /*
1192 * once all the real blocks have been flushed, it is safe to remove them
1193 * from the journal list for this transaction. Aside from freeing the
1194 * cnode, this also allows the block to be reallocated for data blocks
1195 * if it had been deleted.
1196 */
1200 {
1205 /*
1206 * which is better, to lock once around the whole loop, or
1207 * to lock for each call to remove_journal_hash?
1208 */
1216 }
1220 }
1224 }
1226 }
1228 /*
1229 * if this timestamp is greater than the timestamp we wrote last to the
1230 * header block, write it to the header block. once this is done, I can
1231 * safely say the log area for this transaction won't ever be replayed,
1232 * and I can start releasing blocks in this transaction for reuse as data
1233 * blocks. called by flush_journal_list, before it calls
1234 * remove_all_from_journal_list
1235 */
1239 {
1253 #ifdef CONFIG_REISERFS_CHECK
1256 #endif
1258 }
1259 }
1263 b_data);
1273 else
1281 }
1282 }
1284 }
1289 {
1291 }
1293 /*
1294 ** flush any and all journal lists older than you are
1295 ** can only be called from flush_journal_list
1296 */
1299 {
1305 /*
1306 * we know we are the only ones flushing things, no extra race
1307 * protection is required.
1308 */
1309 restart:
1311 /* Did we wrap? */
1317 /* do not flush all */
1320 /* other_jl is now deleted from the list */
1322 }
1324 }
1328 {
1333 }
1334 }
1336 /*
1337 * flush a journal list, both commit and real blocks
1338 *
1339 * always set flushall to 1, unless you are calling from inside
1340 * flush_journal_list
1341 *
1342 * IMPORTANT. This can only be called while there are no journal writers,
1343 * and the journal is locked. That means it can only be called from
1344 * do_journal_end, or by journal_release
1345 */
1348 {
1365 }
1367 /* if flushall == 0, the lock is already held */
1372 }
1379 }
1381 /* if all the work is already done, get out of here */
1385 }
1387 /*
1388 * start by putting the commit list on disk. This will also flush
1389 * the commit lists of any olders transactions
1390 */
1397 /* are we done now? */
1401 }
1403 /*
1404 * loop through each cnode, see if we need to write it,
1405 * or wait on a more recent transaction, or just ignore it
1406 */
1410 }
1416 /* blocknr of 0 is no longer in the hash, ignore it */
1419 }
1421 /*
1422 * This transaction failed commit.
1423 * Don't write out to the disk
1424 */
1429 /*
1430 * the order is important here. We check pjl to make sure we
1431 * don't clear BH_JDirty_wait if we aren't the one writing this
1432 * block to disk
1433 */
1437 /*
1438 * we do this to make sure nobody releases the
1439 * buffer while we are working with it
1440 */
1448 /*
1449 * everything with !pjl && jwait
1450 * should be writable
1451 */
1453 }
1454 }
1456 /*
1457 * if someone has this block in a newer transaction, just make
1458 * sure they are committed, and don't try writing it to disk
1459 */
1464 }
1466 /*
1467 * bh == NULL when the block got to disk on its own, OR,
1468 * the block got freed in a future transaction
1469 */
1472 }
1474 /*
1475 * this should never happen. kupdate_one_transaction has
1476 * this list locked while it works, so we should never see a
1477 * buffer here that is not marked JDirty_wait
1478 */
1481 "BAD! buffer %llu %cdirty %cjwait, "
1486 }
1488 /*
1489 * we inc again because saved_bh gets decremented
1490 * at free_cnode
1491 */
1498 else
1500 count++;
1506 }
1507 free_cnode:
1511 /*
1512 * we incremented this to keep others from
1513 * taking the buffer head away
1514 */
1519 }
1520 }
1521 }
1529 }
1538 }
1540 #ifdef CONFIG_REISERFS_CHECK
1543 #endif
1545 }
1546 /*
1547 * note, we must clear the JDirty_wait bit
1548 * after the up to date check, otherwise we
1549 * race against our flushpage routine
1550 */
1554 /* drop one ref for us */
1556 /* drop one ref for journal_mark_dirty */
1558 }
1560 }
1561 }
1566 __func__);
1567 flush_older_and_return:
1569 /*
1570 * before we can update the journal header block, we _must_ flush all
1571 * real blocks from all older transactions to disk. This is because
1572 * once the header block is updated, this transaction will not be
1573 * replayed after a crash
1574 */
1577 }
1580 /*
1581 * before we can remove everything from the hash tables for this
1582 * transaction, we must make sure it can never be replayed
1583 *
1584 * since we are only called from do_journal_end, we know for sure there
1585 * are no allocations going on while we are flushing journal lists. So,
1586 * we only need to update the journal header block for the last list
1587 * being flushed
1588 */
1590 err =
1598 __func__);
1599 }
1609 }
1612 /*
1613 * not strictly required since we are freeing the list, but it should
1614 * help find code using dead lists later on
1615 */
1627 }
1632 {
1640 }
1644 /*
1645 * if the blocknr == 0, this has been cleared from the hash,
1646 * skip it
1647 */
1650 }
1653 /*
1654 * we can race against journal_mark_freed when we try
1655 * to lock_buffer(cn->bh), so we have to inc the buffer
1656 * count, and recheck things after locking
1657 */
1666 ret++;
1668 /* note, cn->bh might be null now */
1670 }
1672 }
1673 next:
1676 }
1678 }
1680 /* used by flush_commit_list */
1683 {
1691 /*
1692 * look for a more recent transaction that logged this
1693 * buffer. Only the most recent transaction with a buffer in
1694 * it is allowed to send that buffer to disk
1695 */
1700 /*
1701 * if the buffer is prepared, it will either be logged
1702 * or restored. If restored, we need to make sure
1703 * it actually gets marked dirty
1704 */
1711 }
1712 }
1714 }
1716 }
1723 {
1736 }
1738 /*
1739 * we've got j_flush_mutex held, nobody is going to delete any
1740 * of these lists out from underneath us
1741 */
1750 }
1755 transactions_flushed++;
1759 /* did we wrap? */
1762 }
1765 /* don't bother with older transactions */
1768 }
1771 }
1773 done:
1776 }
1778 /*
1779 * for o_sync and fsync heavy applications, they tend to use
1780 * all the journa list slots with tiny transactions. These
1781 * trigger lots and lots of calls to update the header block, which
1782 * adds seeks and slows things down.
1783 *
1784 * This function tries to clear out a large chunk of the journal lists
1785 * at once, which makes everything faster since only the newest journal
1786 * list updates the header block
1787 */
1790 {
1803 /* in data logging mode, try harder to flush a lot of blocks */
1806 /* flush for 256 transactions or limit blocks, whichever comes first */
1811 }
1815 }
1821 }
1825 /*
1826 * try to find a group of blocks we can flush across all the
1827 * transactions, but only bother if we've actually spanned
1828 * across multiple lists
1829 */
1832 }
1837 }
1839 /*
1840 * removes any nodes in table with name block and dev as bh.
1841 * only touchs the hnext and hprev pointers.
1842 */
1847 {
1854 }
1862 }
1867 }
1871 /*
1872 * anybody who clears the cur->bh will also
1873 * dec the nonzerolen
1874 */
1879 }
1881 }
1882 }
1885 {
1895 }
1896 /*
1897 * j_header_bh is on the journal dev, make sure
1898 * not to release the journal dev until we brelse j_header_bh
1899 */
1902 }
1904 /*
1905 * call on unmount. Only set error to 1 if you haven't made your way out
1906 * of read_super() yet. Any other caller must keep error at 0.
1907 */
1910 {
1915 /*
1916 * we only want to flush out transactions if we were
1917 * called with error == 0
1918 */
1920 /* end the current trans */
1924 /*
1925 * make sure something gets logged to force
1926 * our way into the flush code
1927 */
1935 }
1936 }
1938 /* this also catches errors during the do_journal_end above */
1947 }
1948 }
1951 /*
1952 * We must release the write lock here because
1953 * the workqueue job (flush_async_commit) needs this lock
1954 */
1957 /*
1958 * Cancel flushing of old commits. Note that neither of these works
1959 * will be requeued because superblock is being shutdown and doesn't
1960 * have MS_ACTIVE set.
1961 */
1963 /* wait for all commits to finish */
1971 }
1973 /* * call on unmount. flush all journal trans, release all alloc'd ram */
1976 {
1978 }
1980 /* only call from an error condition inside reiserfs_read_super! */
1983 {
1985 }
1987 /*
1988 * compares description block with commit block.
1989 * returns 1 if they differ, 0 if they are the same
1990 */
1994 {
2000 }
2002 }
2004 /*
2005 * returns 0 if it did not find a description block
2006 * returns -1 if it found a corrupt commit block
2007 * returns 1 if both desc and commit were valid
2008 * NOTE: only called during fs mount
2009 */
2014 {
2029 "journal-986: transaction "
2030 "is valid returning because trans_id %d is greater than "
2035 }
2039 "journal-1087: transaction "
2040 "is valid returning because mount_id %d is less than "
2045 }
2048 "Bad transaction length %d "
2052 }
2055 /*
2056 * ok, we have a journal description block,
2057 * let's see if the transaction was valid
2058 */
2059 c_bh =
2069 "journal_transaction_is_valid, commit offset %ld had bad "
2080 "journal-1004: "
2081 "transaction_is_valid setting oldest invalid trans_id "
2084 }
2086 }
2089 "journal-1006: found valid "
2098 }
2099 }
2102 {
2106 }
2107 }
2109 /*
2110 * given the start, and values for the oldest acceptable transactions,
2111 * this either reads in a replays a transaction, or returns because the
2112 * transaction is invalid, or too old.
2113 * NOTE: only called during fs mount
2114 */
2120 {
2145 cur_dblock -
2149 }
2152 "journal_read_trans skipping because %d is != "
2154 newest_mount_id);
2157 }
2164 }
2168 "journal_read_transaction, "
2177 }
2185 }
2188 /*
2189 * now we know we've got a good transaction, and it was
2190 * inside the valid time ranges
2191 */
2204 }
2205 /* get all the buffer heads */
2222 }
2225 "REPLAY FAILURE fsck required! "
2226 "Block to replay is outside of "
2229 }
2230 /* make sure we don't try to replay onto log or reserved area */
2234 "REPLAY FAILURE fsck required! "
2236 abort_replay:
2244 }
2245 }
2246 /* read in the log blocks, memcpy to the corresponding real block */
2253 "REPLAY FAILURE fsck required! "
2263 }
2268 }
2269 /* flush out the real blocks */
2273 }
2278 "REPLAY FAILURE, fsck required! "
2287 }
2289 }
2290 cur_dblock =
2298 /*
2299 * init starting values for the first transaction, in case
2300 * this is the last transaction to be replayed.
2301 */
2305 /* check for trans_id overflow */
2313 }
2315 /*
2316 * This function reads blocks starting from block and to max_block of bufsize
2317 * size (but no more than BUFNR blocks at a time). This proved to improve
2318 * mounting speed on self-rebuilding raid5 arrays at least.
2319 * Right now it is only used from journal code. But later we might use it
2320 * from other places.
2321 * Note: Do not use journal_getblk/sb_getblk functions here!
2322 */
2325 b_blocknr_t max_block)
2326 {
2338 }
2348 }
2358 }
2360 /*
2361 * read and replay the log
2362 * on a clean unmount, the journal header's next unflushed pointer will be
2363 * to an invalid transaction. This tests that before finding all the
2364 * transactions in the log, which makes normal mount times fast.
2365 *
2366 * After a crash, this starts with the next unflushed transaction, and
2367 * replays until it finds one too old, or invalid.
2368 *
2369 * On exit, it sets things up so the first transaction will work correctly.
2370 * NOTE: only called during fs mount
2371 */
2373 {
2394 /*
2395 * step 1, read in the journal header block. Check the transaction
2396 * it says is the first unflushed, and if that transaction is not
2397 * valid, replay is done
2398 */
2404 }
2409 oldest_start =
2415 "journal-1153: found in "
2416 "header: first_unflushed_offset %d, last_flushed_trans_id "
2421 /*
2422 * now, we try to read the first unflushed offset. If it
2423 * is not valid, there is nothing more we can do, and it
2424 * makes no sense to read through the whole log.
2425 */
2426 d_bh =
2433 }
2436 }
2438 /*
2439 * ok, there are transactions that need to be replayed. start
2440 * with the first log block, find all the valid transactions, and
2441 * pick out the oldest.
2442 */
2447 /*
2448 * Note that it is required for blocksize of primary fs
2449 * device and journal device to be the same
2450 */
2451 d_bh =
2456 ret =
2467 "journal-1179: Setting "
2469 oldest_start -
2470 SB_ONDISK_JOURNAL_1st_BLOCK
2473 /* one we just read was older */
2477 "journal-1180: Resetting "
2479 oldest_start -
2480 SB_ONDISK_JOURNAL_1st_BLOCK
2482 }
2486 "journal-1299: Setting "
2489 }
2492 cur_dblock++;
2493 }
2495 }
2497 start_log_replay:
2501 "journal-1206: Starting replay "
2504 oldest_trans_id);
2506 }
2509 ret =
2516 }
2517 cur_dblock =
2519 replay_count++;
2522 }
2527 }
2528 /*
2529 * j_start does not get set correctly if we don't replay any
2530 * transactions. if we had a valid journal_header, set j_start
2531 * to the first unflushed transaction value, copy the trans_id
2532 * from the header
2533 */
2538 /* check for trans_id overflow */
2546 }
2554 }
2555 /* needed to satisfy the locking in _update_journal_header_block */
2561 /*
2562 * replay failed, caller must call free_journal_ram and abort
2563 * the mount
2564 */
2566 }
2569 }
2572 {
2584 }
2587 {
2589 }
2593 {
2597 }
2598 }
2603 {
2605 dev_t jdev;
2618 /* there is no "jdev" option and journal is on separate device */
2623 journal);
2636 }
2647 }
2654 }
2656 /*
2657 * When creating/tuning a file system user can assign some
2658 * journal params within boundaries which depend on the ratio
2659 * blocksize/standard_blocksize.
2660 *
2661 * For blocks >= standard_blocksize transaction size should
2662 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2663 * then JOURNAL_TRANS_MAX_DEFAULT.
2664 *
2665 * For blocks < standard_blocksize these boundaries should be
2666 * decreased proportionally.
2667 */
2668 #define REISERFS_STANDARD_BLKSIZE (4096)
2672 {
2674 /* Non-default journal params. Do sanity check for them. */
2682 JOURNAL_MIN_RATIO) {
2684 "bad transaction max size (%u). "
2687 }
2691 "bad transaction max batch (%u). "
2694 }
2696 /*
2697 * Default journal params.
2698 * The file system was created by old version
2699 * of mkreiserfs, so some fields contain zeros,
2700 * and we need to advise proper values for them
2701 */
2706 }
2710 }
2712 }
2714 /* must be called once on fs mount. calls journal_read for you */
2717 {
2731 }
2743 /* reserved for journal area support */
2745 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2749 REISERFS_DISK_OFFSET_IN_BYTES /
2752 /*
2753 * Sanity check to see is the standard journal fitting
2754 * within first bitmap (actual for small blocksizes)
2755 */
2760 "journal does not fit for area addressed "
2761 "by first of bitmap blocks. It starts at "
2767 }
2773 }
2777 /* read journal header */
2785 }
2788 /* make sure that journal matches to the super block */
2793 "journal header magic %x (device %pg) does "
2800 }
2815 }
2818 "journal first block %u, max trans len %u, max batch %u, "
2863 "allocation failed (%ld bytes). Journal is "
2864 "too large for available memory. Usually "
2868 }
2873 /*
2874 * get_list_bitmap() may call flush_commit_list() which
2875 * requires the lock. Calling flush_commit_list() shouldn't happen
2876 * this early but I like to be paranoid.
2877 */
2885 }
2892 }
2897 free_and_return:
2900 }
2902 /*
2903 * test for a polite end of the current transaction. Used by file_write,
2904 * and should be used by delete to make sure they don't write more than
2905 * can fit inside a single transaction
2906 */
2909 {
2912 /* cannot restart while nested */
2922 }
2927 }
2929 /* this must be called inside a transaction */
2931 {
2937 }
2939 /* this must be called without a transaction started */
2941 {
2945 }
2947 /* this must be called without a transaction started */
2949 {
2953 }
2956 {
2957 wait_queue_t wait;
2961 /*
2962 * we don't want to use wait_event here because
2963 * we only want to wait once.
2964 */
2972 }
2975 }
2978 {
2982 }
2985 {
3000 }
3006 }
3007 }
3009 /*
3010 * join == true if you must join an existing transaction.
3011 * join == false if you can deal with waiting for others to finish
3012 *
3013 * this will block until the transaction is joinable. send the number of
3014 * blocks you expect to use in nblocks.
3015 */
3019 {
3032 /* set here for journal_join */
3036 relock:
3042 }
3052 }
3055 /*
3056 * if there is no room in the journal OR
3057 * if this transaction is too old, and we weren't called joinable,
3058 * wait for it to finish before beginning we don't sleep if there
3059 * aren't other writers
3060 */
3063 (!join
3073 /* allow others to finish this transaction */
3081 sched_count++;
3084 }
3085 }
3086 /*
3087 * don't mess with joining the transaction if all we
3088 * have to do is wait for someone else to do a commit
3089 */
3094 }
3096 }
3101 /* someone might have ended the transaction while we joined */
3106 }
3113 }
3114 /* we are the first writer, set trans_id */
3117 }
3127 out_fail:
3129 /*
3130 * Re-set th->t_super, so we can properly keep track of how many
3131 * persistent transactions there are. We need to do this so if this
3132 * call is part of a failed restart_transaction, we can free it later
3133 */
3136 }
3139 super_block
3142 {
3146 /*
3147 * if we're nesting into an existing transaction. It will be
3148 * persistent on its own
3149 */
3156 }
3164 }
3168 }
3171 {
3176 else
3181 }
3183 }
3187 {
3190 /*
3191 * this keeps do_journal_end from NULLing out the
3192 * current->journal_info pointer
3193 */
3197 }
3201 {
3204 /*
3205 * this keeps do_journal_end from NULLing out the
3206 * current->journal_info pointer
3207 */
3211 }
3215 {
3221 /* we are nesting into the current transaction */
3228 "BAD: refcount <= 1, but "
3232 /*
3233 * we've ended up with a handle from a different
3234 * filesystem. save it and restore on journal_end.
3235 * This should never really happen...
3236 */
3241 }
3244 }
3248 /*
3249 * I guess this boils down to being the reciprocal of clm-2100 above.
3250 * If do_journal_begin_r fails, we need to put it back, since
3251 * journal_end won't be called to do it. */
3254 else
3258 }
3260 /*
3261 * puts bh into the current transaction. If it was already there, reorders
3262 * removes the old pointers from the hash, and puts new ones in (to make
3263 * sure replay happen in the right order).
3264 *
3265 * if it was dirty, cleans and files onto the clean list. I can't let it
3266 * be dirty again until the transaction is committed.
3267 *
3268 * if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3269 */
3272 {
3285 }
3289 /* already in this transaction, we are done */
3293 }
3295 /*
3296 * this must be turned into a panic instead of a warning. We can't
3297 * allow a dirty or journal_dirty or locked buffer to be logged, as
3298 * some changes could get to disk too early. NOT GOOD.
3299 */
3302 "buffer %llu bad state "
3309 }
3316 }
3317 /*
3318 * this error means I've screwed up, and we've overflowed
3319 * the transaction. Nothing can be done here, except make the
3320 * FS readonly or panic.
3321 */
3326 }
3332 }
3336 }
3340 /* now put this guy on the end */
3345 }
3350 }
3361 }
3362 }
3372 }
3375 }
3378 {
3387 }
3394 /*
3395 * we aren't allowed to close a nested transaction on a
3396 * different filesystem from the one in the task struct
3397 */
3403 }
3407 }
3408 }
3410 /*
3411 * removes from the current transaction, relsing and descrementing any counters.
3412 * also files the removed buffer directly onto the clean list
3413 *
3414 * called by journal_mark_freed when a block has been deleted
3415 *
3416 * returns 1 if it cleaned and relsed the buffer. 0 otherwise
3417 */
3420 {
3429 }
3433 }
3436 }
3439 }
3442 }
3456 }
3458 }
3463 }
3465 /*
3466 * for any cnode in a journal list, it can only be dirtied of all the
3467 * transactions that include it are committed to disk.
3468 * this checks through each transaction, and returns 1 if you are allowed
3469 * to dirty, and 0 if you aren't
3470 *
3471 * it is called by dirty_journal_list, which is called after
3472 * flush_commit_list has gotten all the log blocks for a given
3473 * transaction on disk
3474 *
3475 */
3477 {
3483 /*
3484 * first test hprev. These are all newer than cn, so any node here
3485 * with the same block number and dev means this node can't be sent
3486 * to disk right now.
3487 */
3492 }
3494 }
3495 /*
3496 * then test hnext. These are all older than cn. As long as they
3497 * are committed to the log, it is safe to write cn to disk
3498 */
3505 }
3507 }
3509 }
3511 /*
3512 * syncs the commit blocks, but does not force the real buffers to disk
3513 * will wait until the current transaction is done/committed before returning
3514 */
3516 {
3521 /* you can sync while nested, very, very bad */
3527 }
3529 }
3531 /* writeback the pending async commits to disk */
3533 {
3542 /* last entry is the youngest, commit it and you get everything */
3546 }
3548 }
3550 /*
3551 * flushes any old transactions to disk
3552 * ends the current transaction if it is too old
3553 */
3555 {
3561 /*
3562 * safety check so we don't flush while we are replaying the log during
3563 * mount
3564 */
3568 /*
3569 * check the current transaction. If there are no writers, and it is
3570 * too old, finish it, and force the commit blocks to disk
3571 */
3582 /*
3583 * we're only being called from kreiserfsd, it makes
3584 * no sense to do an async commit so that kreiserfsd
3585 * can do it later
3586 */
3588 }
3589 }
3590 }
3592 /*
3593 * returns 0 if do_journal_end should return right away, returns 1 if
3594 * do_journal_end should finish the commit
3595 *
3596 * if the current transaction is too old, but still has writers, this will
3597 * wait on j_join_wait until all the writers are done. By the time it
3598 * wakes up, the transaction it was called has already ended, so it just
3599 * flushes the commit list and returns 0.
3600 *
3601 * Won't batch when flush or commit_now is set. Also won't batch when
3602 * others are waiting on j_join_wait.
3603 *
3604 * Note, we can't allow the journal_end to proceed while there are still
3605 * writers in the log.
3606 */
3608 {
3624 }
3627 /* <= 0 is allowed. unmounting might not call begin */
3631 /*
3632 * BUG, deal with case where j_len is 0, but people previously
3633 * freed blocks need to be released will be dealt with by next
3634 * transaction that actually writes something, but should be taken
3635 * care of in this trans
3636 */
3639 /*
3640 * if wcount > 0, and we are called to with flush or commit_now,
3641 * we wait on j_join_wait. We will wake up when the last writer has
3642 * finished the transaction, and started it on its way to the disk.
3643 * Then, we flush the commit or journal list, and just return 0
3644 * because the rest of journal end was already done for this
3645 * transaction.
3646 */
3658 }
3661 /*
3662 * sleep while the current transaction is
3663 * still j_jlocked
3664 */
3673 }
3675 }
3676 }
3683 }
3685 }
3688 }
3690 /* deal with old transactions where we are the last writers */
3695 }
3696 /* don't batch when someone is waiting on j_join_wait */
3697 /* don't batch when syncing the commit or flushing the whole trans */
3705 }
3711 }
3713 }
3715 /*
3716 * Does all the work that makes deleting blocks safe.
3717 * when deleting a block mark BH_JNew, just remove it from the current
3718 * transaction, clean it's buffer_head and move on.
3719 *
3720 * otherwise:
3721 * set a bit for the block in the journal bitmap. That will prevent it from
3722 * being allocated for unformatted nodes before this transaction has finished.
3723 *
3724 * mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers.
3725 * That will prevent any old transactions with this block from trying to flush
3726 * to the real location. Since we aren't removing the cnode from the
3727 * journal_list_hash, *the block can't be reallocated yet.
3728 *
3729 * Then remove it from the current transaction, decrementing any counters and
3730 * filing it on the clean list.
3731 */
3734 {
3746 }
3747 /* if it is journal new, we just remove it from this transaction */
3754 /*
3755 * set the bit for this block in the journal bitmap
3756 * for this transaction
3757 */
3762 }
3765 /* Note, the entire while loop is not allowed to schedule. */
3770 }
3773 /*
3774 * find all older transactions with this block,
3775 * make sure they don't try to write it out
3776 */
3778 blocknr);
3783 /*
3784 * remove_from_transaction will brelse
3785 * the buffer if it was in the current
3786 * trans
3787 */
3790 bh);
3793 bh);
3801 }
3802 }
3803 /*
3804 * since we are clearing the bh,
3805 * we MUST dec nonzerolen
3806 */
3809 j_nonzerolen);
3810 }
3812 }
3813 }
3815 }
3816 }
3821 }
3824 {
3828 }
3830 /*
3831 * returns -1 on error, 0 if no commits/barriers were done and 1
3832 * if a transaction was actually committed and the barrier was done
3833 */
3836 {
3842 /*
3843 * is it from the current transaction,
3844 * or from an unknown transaction?
3845 */
3848 /*
3849 * try to let other writers come in and
3850 * grow this transaction
3851 */
3855 }
3861 /* someone might have ended this transaction while we joined */
3868 }
3875 /*
3876 * this gets tricky, we have to make sure the journal list in
3877 * the inode still exists. We know the list is still around
3878 * if we've got a larger transaction id than the oldest list
3879 */
3880 flush_commit_only:
3882 /*
3883 * we only set ret to 1 when we know for sure
3884 * the barrier hasn't been started yet on the commit
3885 * block.
3886 */
3892 }
3893 }
3894 /* otherwise the list is gone, and long since committed */
3896 }
3899 {
3903 /*
3904 * for the whole inode, assume unset id means it was
3905 * changed in the current transaction. More conservative
3906 */
3910 /* jl will be updated in __commit_trans_jl */
3911 }
3914 }
3918 {
3923 }
3934 }
3936 }
3938 }
3941 /*
3942 * before we can change a metadata block, we have to make sure it won't
3943 * be written to disk while we are altering it. So, we must:
3944 * clean it
3945 * wait on it.
3946 */
3949 {
3956 }
3961 }
3964 }
3966 /*
3967 * long and ugly. If flush, will not return until all commit
3968 * blocks and all real buffers in the trans are on disk.
3969 * If no_async, won't return until all commit blocks are on disk.
3970 *
3971 * keep reading, there are comments as you go along
3972 *
3973 * If the journal is aborted, we just clean up. Things like flushing
3974 * journal lists, etc just won't happen.
3975 */
3977 {
4002 /*
4003 * protect flush_older_commits from doing mistakes if the
4004 * transaction ID counter gets overflowed.
4005 */
4017 }
4023 }
4027 }
4029 /*
4030 * check_journal_end locks the journal, and unlocks if it does
4031 * not return 1 it tells us if we should continue with the
4032 * journal_end, or just return
4033 */
4039 }
4041 /* check_journal_end might set these, check again */
4044 }
4046 /*
4047 * j must wait means we have to flush the log blocks, and the
4048 * real blocks for this transaction
4049 */
4052 }
4053 #ifdef REISERFS_PREALLOCATE
4054 /*
4055 * quota ops might need to nest, setup the journal_info pointer
4056 * for them and raise the refcount so that it is > 0.
4057 */
4061 /* it should not involve new blocks into the transaction */
4066 #endif
4068 /* setup description block */
4069 d_bh =
4079 /*
4080 * setup commit block. Don't write (keep it clean too) this one
4081 * until after everyone else is written
4082 */
4091 /* init this journal list */
4094 /*
4095 * we lock the commit before doing anything because
4096 * we want to make sure nobody tries to run flush_commit_list until
4097 * the new transaction is fully setup, and we've already flushed the
4098 * ordered bh list
4099 */
4102 /* save the transaction id in case we need to commit it later */
4115 /*
4116 * The ENTIRE FOR LOOP MUST not cause schedule to occur.
4117 * for each real block, add it to the journal list hash,
4118 * copy into real block index array in the commit or desc block
4119 */
4127 }
4130 }
4135 }
4137 /*
4138 * make sure the block we are trying to log
4139 * is not a block of journal or reserved area
4140 */
4144 "Trying to log block %lu, "
4147 }
4160 }
4162 i--;
4163 }
4164 }
4170 /*
4171 * special check in case all buffers in the journal
4172 * were marked for not logging
4173 */
4176 /*
4177 * we're about to dirty all the log blocks, mark the description block
4178 * dirty now too. Don't mark the commit block dirty until all the
4179 * others are on disk
4180 */
4183 /*
4184 * first data block is j_start + 1, so add one to
4185 * cur_write_start wherever you use it
4186 */
4192 /* copy all the real blocks into log area. dirty log blocks */
4197 tmp_bh =
4201 jindex) %
4211 jindex++;
4215 /*
4216 * JDirty cleared sometime during transaction.
4217 * don't log this one
4218 */
4220 "BAD, buffer in journal hash, "
4223 }
4228 }
4230 /*
4231 * we are done with both the c_bh and d_bh, but
4232 * c_bh must be written after all other commit blocks,
4233 * so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4234 */
4238 /* now it is safe to insert this transaction on the main list */
4243 /* reset journal values for the next transaction */
4254 /* check for trans_id overflow */
4264 /*
4265 * make sure reiserfs_add_jh sees the new current_jl before we
4266 * write out the tails
4267 */
4270 /*
4271 * tail conversion targets have to hit the disk before we end the
4272 * transaction. Otherwise a later transaction might repack the tail
4273 * before this transaction commits, leaving the data block unflushed
4274 * and clean, if we crash before the later transaction commits, the
4275 * data block is lost.
4276 */
4282 }
4286 /*
4287 * honor the flush wishes from the caller, simple commits can
4288 * be done outside the journal lock, they are done below
4289 *
4290 * if we don't flush the commit list right now, we put it into
4291 * the work queue so the people waiting on the async progress work
4292 * queue don't wait for this proc to flush journal lists and such.
4293 */
4298 /*
4299 * Avoid queueing work when sb is being shut down. Transaction
4300 * will be flushed on journal shutdown.
4301 */
4305 }
4307 /*
4308 * if the next transaction has any chance of wrapping, flush
4309 * transactions that might get overwritten. If any journal lists
4310 * are very old flush them as well.
4311 */
4312 first_jl:
4323 /*
4324 * if we don't cross into the next
4325 * transaction and we don't wrap, there is
4326 * no way we can overlap any later transactions
4327 * break now
4328 */
4330 }
4340 /*
4341 * we don't overlap anything from out start
4342 * to the end of the log, and our wrapped
4343 * portion doesn't overlap anything at
4344 * the start of the log. We can break
4345 */
4347 }
4348 }
4349 }
4357 }
4361 /* wake up any body waiting to join. */
4368 }
4369 out:
4373 /*
4374 * Re-set th->t_super, so we can properly keep track of how many
4375 * persistent transactions there are. We need to do this so if this
4376 * call is part of a failed restart_transaction, we can free it later
4377 */
4381 }
4383 /* Send the file system read only and refuse new transactions */
4385 {
4396 #ifdef CONFIG_REISERFS_CHECK
4398 #endif
4399 }