| blob | ba98546fabbd1921beedc54074f7399115f14f54 |
1 /*
2 ** Write ahead logging implementation copyright Chris Mason 2000
3 **
4 ** The background commits make this code very interelated, 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>
40 #include <linux/reiserfs_fs.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/smp_lock.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>
54 #include <asm/system.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 /* the number of mounted filesystems. This is used to decide when to
63 ** start and kill the commit workqueue
64 */
70 structs at 4k */
73 /* cnode stat bits. Move these into reiserfs_fs.h */
76 #define BLOCK_FREED_HOLDER 3 /* this block was freed during this transaction, and can't be written */
79 #define BLOCK_DIRTIED 5
81 /* journal list state bits */
82 #define LIST_TOUCHED 1
83 #define LIST_DIRTY 2
86 /* flags for do_journal_end */
108 /* values for join in do_journal_begin_r */
113 };
120 {
124 }
126 /*
127 ** clears BH_Dirty and sticks the buffer on the clean list. Called because I can't allow refile_buffer to
128 ** make schedule happen after I've freed a block. Look at remove_from_transaction and journal_mark_freed for
129 ** more details.
130 */
132 {
136 }
138 }
141 {
145 }
149 {
156 }
161 }
165 }
168 {
174 repeat:
182 }
187 }
189 }
192 {
201 }
202 }
205 {
216 }
217 }
218 }
221 b_blocknr_t block,
223 {
229 }
232 }
236 {
245 }
246 }
247 }
249 /*
250 ** only call this on FS unmount.
251 */
254 {
263 }
265 }
268 {
280 }
283 }
285 /*
286 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
287 ** jb_array is the array to be filled in.
288 */
292 {
307 }
309 }
313 }
315 }
317 /*
318 ** find an available list bitmap. If you can't find one, flush a commit list
319 ** and try again
320 */
322 struct reiserfs_journal_list
324 {
339 }
342 }
343 }
346 }
349 }
351 /*
352 ** allocates a new chunk of X nodes, and links them all together as a list.
353 ** Uses the cnode->next and cnode->prev pointers
354 ** returns NULL on failure
355 */
357 {
362 }
366 }
373 }
376 }
378 /*
379 ** pulls a cnode off the free list, or returns NULL on failure
380 */
382 {
390 }
396 }
399 }
403 }
405 /*
406 ** returns a cnode to the free list
407 */
410 {
417 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
421 }
422 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
424 }
427 {
430 }
432 /* return a cnode with same dev, block number and size in table, or null if not found */
434 super_block
436 struct
437 reiserfs_journal_cnode
440 {
447 }
449 }
451 /*
452 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
453 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
454 ** being overwritten by a replay after crashing.
455 **
456 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
457 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
458 ** sure you never write the block without logging it.
459 **
460 ** next_zero_bit is a suggestion about the next block to try for find_forward.
461 ** when bl is rejected because it is set in a journal list bitmap, we search
462 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
463 ** through next_zero_bit for find_forward to try.
464 **
465 ** Just because we return something in next_zero_bit does not mean we won't
466 ** reject it on the next call to reiserfs_in_journal
467 **
468 */
472 {
482 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
483 ** if we crash before the transaction that freed it commits, this transaction won't
484 ** have committed either, and the block will never be written
485 */
493 data)) {
497 data),
501 }
502 }
503 }
506 /* is it in any old transactions? */
511 }
513 /* is it in the current transaction. This should never happen */
517 }
520 /* safe for reuse */
522 }
524 /* insert cn into table
525 */
528 {
536 }
538 }
540 /* lock the current transaction */
542 {
546 }
548 /* unlock the current transaction */
550 {
552 }
555 {
557 }
561 {
565 }
568 }
570 /*
571 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
572 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
573 ** transaction.
574 */
577 {
582 }
585 }
589 {
598 }
599 }
601 }
603 /*
604 * If page->mapping was null, we failed to truncate this page for
605 * some reason. Most likely because it was truncated after being
606 * logged via data=journal.
607 *
608 * This does a check to see if the buffer belongs to one of these
609 * lost pages before doing the final put_bh. If page->mapping was
610 * null, it tries to free buffers on the page, which should make the
611 * final page_cache_release drop the page from the lru.
612 */
614 {
625 }
626 }
629 {
636 }
639 else
644 }
647 {
650 else
654 }
657 {
667 }
670 {
677 }
680 {
687 }
691 {
700 }
701 }
703 #define CHUNK_SIZE 32
707 };
710 {
715 }
718 }
721 {
726 }
729 }
733 {
744 }
746 }
750 {
757 }
759 }
760 }
762 /*
763 * we want to free the jh when the buffer has been written
764 * and waited on
765 */
767 {
780 }
781 }
785 {
793 }
797 no_jh:
801 /* buffer must be locked for __add_jh, should be able to have
802 * two adds at the same time
803 */
807 }
813 }
816 }
819 {
821 }
823 {
825 }
827 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
832 {
850 }
858 }
859 /* in theory, dirty non-uptodate buffers should never get here,
860 * but the upper layer io error paths still have a few quirks.
861 * Handle them here as gracefully as we can
862 */
866 }
873 }
874 loop_next:
877 }
882 }
893 }
896 }
897 /* ugly interaction with invalidatepage here.
898 * reiserfs_invalidate_page will pin any buffer that has a valid
899 * journal head from an older transaction. If someone else sets
900 * our buffer dirty after we write it in the first loop, and
901 * then someone truncates the page away, nobody will ever write
902 * the buffer. We're safe if we write the page one last time
903 * after freeing the journal header.
904 */
909 }
912 }
915 }
919 {
928 find_first:
929 /*
930 * first we walk backwards to find the oldest uncommitted transation
931 */
942 }
944 /* if we didn't find any older uncommitted transactions, return now */
947 }
960 /* list we were called with is gone, return */
964 /* the one we just flushed is gone, this means all
965 * older lists are also gone, so first_jl is no longer
966 * valid either. Go back to the beginning.
967 */
971 }
972 }
978 }
979 }
981 }
984 {
992 }
995 }
997 /*
998 ** if this journal list still has commit blocks unflushed, send them to disk.
999 **
1000 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
1001 ** Before the commit block can by written, every other log block must be safely on disk
1002 **
1003 */
1006 {
1008 b_blocknr_t bn;
1020 }
1024 /* before we can put our commit blocks on disk, we have to make sure everyone older than
1025 ** us is on disk too
1026 */
1033 /* list disappeared during flush_older_commits. return */
1035 }
1036 }
1038 /* make sure nobody is trying to flush this one at the same time */
1044 }
1047 /* this commit is done, exit */
1051 }
1054 }
1059 /*
1060 * We might sleep in numerous places inside
1061 * write_ordered_buffers. Relax the write lock.
1062 */
1069 }
1071 /*
1072 * for the description block and all the log blocks, submit any buffers
1073 * that haven't already reached the disk. Try to write at least 256
1074 * log blocks. later on, we will only wait on blocks that correspond
1075 * to this transaction, but while we're unplugging we might as well
1076 * get a chunk of data on there.
1077 */
1091 }
1093 }
1094 }
1097 /* We're skipping the commit if there's an error */
1101 /* wait on everything written so far before writing the commit
1102 * if we are in barrier mode, send the commit down now
1103 */
1113 }
1114 }
1123 // since we're using ll_rw_blk above, it might have skipped over
1124 // a locked buffer. Double check here
1125 //
1126 /* redundant, sync_dirty_buffer() checks */
1131 }
1133 #ifdef CONFIG_REISERFS_CHECK
1136 #endif
1138 }
1142 }
1147 /* If there was a write error in the journal - we can't commit
1148 * this transaction - it will be invalid and, if successful,
1149 * will just end up propagating the write error out to
1150 * the file system. */
1158 }
1163 }
1167 /* If there was a write error in the journal - we can't commit this
1168 * transaction - it will be invalid and, if successful, will just end
1169 * up propagating the write error out to the filesystem. */
1171 #ifdef CONFIG_REISERFS_CHECK
1173 #endif
1175 }
1181 }
1184 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1189 /* mark the metadata dirty */
1196 }
1198 put_jl:
1203 __func__);
1206 }
1208 /*
1209 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1210 ** returns NULL if it can't find anything
1211 */
1213 reiserfs_journal_cnode
1215 {
1223 }
1225 }
1227 }
1230 {
1240 }
1242 }
1249 /*
1250 ** once all the real blocks have been flushed, it is safe to remove them from the
1251 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1252 ** block to be reallocated for data blocks if it had been deleted.
1253 */
1257 {
1262 /* which is better, to lock once around the whole loop, or
1263 ** to lock for each call to remove_journal_hash?
1264 */
1272 }
1276 }
1280 }
1282 }
1284 /*
1285 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1286 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1287 ** releasing blocks in this transaction for reuse as data blocks.
1288 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1289 **
1290 */
1294 {
1307 #ifdef CONFIG_REISERFS_CHECK
1310 #endif
1312 }
1313 }
1317 b_data);
1330 }
1336 sync:
1341 }
1346 }
1347 }
1349 }
1354 {
1356 }
1358 /*
1359 ** flush any and all journal lists older than you are
1360 ** can only be called from flush_journal_list
1361 */
1364 {
1370 /* we know we are the only ones flushing things, no extra race
1371 * protection is required.
1372 */
1373 restart:
1375 /* Did we wrap? */
1381 /* do not flush all */
1384 /* other_jl is now deleted from the list */
1386 }
1388 }
1392 {
1397 }
1398 }
1400 /* flush a journal list, both commit and real blocks
1401 **
1402 ** always set flushall to 1, unless you are calling from inside
1403 ** flush_journal_list
1404 **
1405 ** IMPORTANT. This can only be called while there are no journal writers,
1406 ** and the journal is locked. That means it can only be called from
1407 ** do_journal_end, or by journal_release
1408 */
1411 {
1427 }
1430 /* if flushall == 0, the lock is already held */
1435 }
1442 }
1446 /* if all the work is already done, get out of here */
1450 }
1452 /* start by putting the commit list on disk. This will also flush
1453 ** the commit lists of any olders transactions
1454 */
1461 /* are we done now? */
1465 }
1467 /* loop through each cnode, see if we need to write it,
1468 ** or wait on a more recent transaction, or just ignore it
1469 */
1473 }
1479 /* blocknr of 0 is no longer in the hash, ignore it */
1482 }
1484 /* This transaction failed commit. Don't write out to the disk */
1489 /* the order is important here. We check pjl to make sure we
1490 ** don't clear BH_JDirty_wait if we aren't the one writing this
1491 ** block to disk
1492 */
1496 /* we do this to make sure nobody releases the buffer while
1497 ** we are working with it
1498 */
1506 /* everything with !pjl && jwait should be writable */
1508 }
1509 }
1511 /* if someone has this block in a newer transaction, just make
1512 ** sure they are committed, and don't try writing it to disk
1513 */
1518 }
1520 /* bh == NULL when the block got to disk on its own, OR,
1521 ** the block got freed in a future transaction
1522 */
1525 }
1527 /* this should never happen. kupdate_one_transaction has this list
1528 ** locked while it works, so we should never see a buffer here that
1529 ** is not marked JDirty_wait
1530 */
1533 "BAD! buffer %llu %cdirty %cjwait, "
1538 }
1540 /* we inc again because saved_bh gets decremented at free_cnode */
1547 else
1549 count++;
1555 }
1556 free_cnode:
1560 /* we incremented this to keep others from taking the buffer head away */
1565 }
1566 }
1567 }
1575 }
1584 }
1586 #ifdef CONFIG_REISERFS_CHECK
1589 #endif
1591 }
1592 /* note, we must clear the JDirty_wait bit after the up to date
1593 ** check, otherwise we race against our flushpage routine
1594 */
1598 /* drop one ref for us */
1600 /* drop one ref for journal_mark_dirty */
1602 }
1604 }
1605 }
1610 __func__);
1611 flush_older_and_return:
1613 /* before we can update the journal header block, we _must_ flush all
1614 ** real blocks from all older transactions to disk. This is because
1615 ** once the header block is updated, this transaction will not be
1616 ** replayed after a crash
1617 */
1620 }
1623 /* before we can remove everything from the hash tables for this
1624 ** transaction, we must make sure it can never be replayed
1625 **
1626 ** since we are only called from do_journal_end, we know for sure there
1627 ** are no allocations going on while we are flushing journal lists. So,
1628 ** we only need to update the journal header block for the last list
1629 ** being flushed
1630 */
1632 err =
1640 __func__);
1641 }
1651 }
1654 /* not strictly required since we are freeing the list, but it should
1655 * help find code using dead lists later on
1656 */
1669 }
1673 {
1681 /* if the blocknr == 0, this has been cleared from the hash,
1682 ** skip it
1683 */
1686 }
1689 next:
1692 }
1694 }
1699 {
1707 }
1711 /* if the blocknr == 0, this has been cleared from the hash,
1712 ** skip it
1713 */
1716 }
1719 /* we can race against journal_mark_freed when we try
1720 * to lock_buffer(cn->bh), so we have to inc the buffer
1721 * count, and recheck things after locking
1722 */
1731 ret++;
1733 /* note, cn->bh might be null now */
1735 }
1737 }
1738 next:
1741 }
1743 }
1745 /* used by flush_commit_list */
1748 {
1756 /* look for a more recent transaction that logged this
1757 ** buffer. Only the most recent transaction with a buffer in
1758 ** it is allowed to send that buffer to disk
1759 */
1764 /* if the buffer is prepared, it will either be logged
1765 * or restored. If restored, we need to make sure
1766 * it actually gets marked dirty
1767 */
1774 }
1775 }
1777 }
1779 }
1786 {
1799 }
1801 /* we've got j_flush_mutex held, nobody is going to delete any
1802 * of these lists out from underneath us
1803 */
1812 }
1817 transactions_flushed++;
1821 /* did we wrap? */
1824 }
1827 /* don't bother with older transactions */
1830 }
1833 }
1835 done:
1838 }
1840 /* for o_sync and fsync heavy applications, they tend to use
1841 ** all the journa list slots with tiny transactions. These
1842 ** trigger lots and lots of calls to update the header block, which
1843 ** adds seeks and slows things down.
1844 **
1845 ** This function tries to clear out a large chunk of the journal lists
1846 ** at once, which makes everything faster since only the newest journal
1847 ** list updates the header block
1848 */
1851 {
1864 /* in data logging mode, try harder to flush a lot of blocks */
1867 /* flush for 256 transactions or limit blocks, whichever comes first */
1872 }
1876 }
1882 }
1883 /* try to find a group of blocks we can flush across all the
1884 ** transactions, but only bother if we've actually spanned
1885 ** across multiple lists
1886 */
1889 }
1892 }
1894 /*
1895 ** removes any nodes in table with name block and dev as bh.
1896 ** only touchs the hnext and hprev pointers.
1897 */
1902 {
1909 }
1917 }
1922 }
1930 }
1932 }
1933 }
1936 {
1946 }
1947 /* j_header_bh is on the journal dev, make sure not to release the journal
1948 * dev until we brelse j_header_bh
1949 */
1952 }
1954 /*
1955 ** call on unmount. Only set error to 1 if you haven't made your way out
1956 ** of read_super() yet. Any other caller must keep error at 0.
1957 */
1960 {
1965 /* we only want to flush out transactions if we were called with error == 0
1966 */
1968 /* end the current trans */
1972 /* make sure something gets logged to force our way into the flush code */
1981 }
1982 }
1984 /* this also catches errors during the do_journal_end above */
1994 }
1995 }
1997 reiserfs_mounted_fs_count--;
1998 /* wait for all commits to finish */
2001 /*
2002 * We must release the write lock here because
2003 * the workqueue job (flush_async_commit) needs this lock
2004 */
2011 }
2018 }
2020 /*
2021 ** call on unmount. flush all journal trans, release all alloc'd ram
2022 */
2025 {
2027 }
2029 /*
2030 ** only call from an error condition inside reiserfs_read_super!
2031 */
2034 {
2036 }
2038 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
2042 {
2048 }
2050 }
2052 /* returns 0 if it did not find a description block
2053 ** returns -1 if it found a corrupt commit block
2054 ** returns 1 if both desc and commit were valid
2055 */
2060 {
2075 "journal-986: transaction "
2076 "is valid returning because trans_id %d is greater than "
2081 }
2085 "journal-1087: transaction "
2086 "is valid returning because mount_id %d is less than "
2091 }
2094 "Bad transaction length %d "
2098 }
2101 /* ok, we have a journal description block, lets see if the transaction was valid */
2102 c_bh =
2112 "journal_transaction_is_valid, commit offset %ld had bad "
2123 "journal-1004: "
2124 "transaction_is_valid setting oldest invalid trans_id "
2127 }
2129 }
2132 "journal-1006: found valid "
2141 }
2142 }
2145 {
2149 }
2150 }
2152 /*
2153 ** given the start, and values for the oldest acceptable transactions,
2154 ** this either reads in a replays a transaction, or returns because the transaction
2155 ** is invalid, or too old.
2156 */
2162 {
2187 cur_dblock -
2191 }
2194 "journal_read_trans skipping because %d is != "
2196 newest_mount_id);
2199 }
2206 }
2210 "journal_read_transaction, "
2219 }
2221 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2234 }
2235 /* get all the buffer heads */
2252 }
2255 "REPLAY FAILURE fsck required! "
2256 "Block to replay is outside of "
2259 }
2260 /* make sure we don't try to replay onto log or reserved area */
2264 "REPLAY FAILURE fsck required! "
2266 abort_replay:
2274 }
2275 }
2276 /* read in the log blocks, memcpy to the corresponding real block */
2286 "REPLAY FAILURE fsck required! "
2296 }
2301 }
2302 /* flush out the real blocks */
2306 }
2311 "REPLAY FAILURE, fsck required! "
2320 }
2322 }
2323 cur_dblock =
2331 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2335 /* check for trans_id overflow */
2343 }
2345 /* This function reads blocks starting from block and to max_block of bufsize
2346 size (but no more than BUFNR blocks at a time). This proved to improve
2347 mounting speed on self-rebuilding raid5 arrays at least.
2348 Right now it is only used from journal code. But later we might use it
2349 from other places.
2350 Note: Do not use journal_getblk/sb_getblk functions here! */
2353 b_blocknr_t max_block)
2354 {
2366 }
2376 }
2386 }
2388 /*
2389 ** read and replay the log
2390 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2391 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2392 **
2393 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2394 **
2395 ** On exit, it sets things up so the first transaction will work correctly.
2396 */
2398 {
2420 /* step 1, read in the journal header block. Check the transaction it says
2421 ** is the first unflushed, and if that transaction is not valid,
2422 ** replay is done
2423 */
2429 }
2434 oldest_start =
2440 "journal-1153: found in "
2441 "header: first_unflushed_offset %d, last_flushed_trans_id "
2446 /* now, we try to read the first unflushed offset. If it is not valid,
2447 ** there is nothing more we can do, and it makes no sense to read
2448 ** through the whole log.
2449 */
2450 d_bh =
2457 }
2460 }
2466 }
2468 /* ok, there are transactions that need to be replayed. start with the first log block, find
2469 ** all the valid transactions, and pick out the oldest.
2470 */
2475 /* Note that it is required for blocksize of primary fs device and journal
2476 device to be the same */
2477 d_bh =
2482 ret =
2493 "journal-1179: Setting "
2495 oldest_start -
2496 SB_ONDISK_JOURNAL_1st_BLOCK
2499 /* one we just read was older */
2503 "journal-1180: Resetting "
2505 oldest_start -
2506 SB_ONDISK_JOURNAL_1st_BLOCK
2508 }
2512 "journal-1299: Setting "
2515 }
2518 cur_dblock++;
2519 }
2521 }
2523 start_log_replay:
2527 "journal-1206: Starting replay "
2530 oldest_trans_id);
2532 }
2535 ret =
2542 }
2543 cur_dblock =
2545 replay_count++;
2548 }
2553 }
2554 /* j_start does not get set correctly if we don't replay any transactions.
2555 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2556 ** copy the trans_id from the header
2557 */
2562 /* check for trans_id overflow */
2570 }
2578 }
2582 /* replay failed, caller must call free_journal_ram and abort
2583 ** the mount
2584 */
2586 }
2588 }
2591 {
2603 }
2606 {
2608 }
2612 {
2622 }
2627 }
2629 }
2634 {
2636 dev_t jdev;
2649 /* there is no "jdev" option and journal is on separate device */
2665 }
2668 }
2671 }
2683 }
2690 }
2692 /**
2693 * When creating/tuning a file system user can assign some
2694 * journal params within boundaries which depend on the ratio
2695 * blocksize/standard_blocksize.
2696 *
2697 * For blocks >= standard_blocksize transaction size should
2698 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2699 * then JOURNAL_TRANS_MAX_DEFAULT.
2700 *
2701 * For blocks < standard_blocksize these boundaries should be
2702 * decreased proportionally.
2703 */
2704 #define REISERFS_STANDARD_BLKSIZE (4096)
2708 {
2710 /* Non-default journal params.
2711 Do sanity check for them. */
2719 JOURNAL_MIN_RATIO) {
2721 "bad transaction max size (%u). "
2724 }
2728 "bad transaction max batch (%u). "
2731 }
2733 /* Default journal params.
2734 The file system was created by old version
2735 of mkreiserfs, so some fields contain zeros,
2736 and we need to advise proper values for them */
2741 }
2745 }
2747 }
2749 /*
2750 ** must be called once on fs mount. calls journal_read for you
2751 */
2754 {
2764 /*
2765 * Unlock here to avoid various RECLAIM-FS-ON <-> IN-RECLAIM-FS
2766 * dependency inversion warnings.
2767 */
2775 }
2790 /* reserved for journal area support */
2792 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2796 REISERFS_DISK_OFFSET_IN_BYTES /
2799 /* Sanity check to see is the standard journal fitting withing first bitmap
2800 (actual for small blocksizes) */
2805 "journal does not fit for area addressed "
2806 "by first of bitmap blocks. It starts at "
2812 }
2814 /*
2815 * We need to unlock here to avoid creating the following
2816 * dependency:
2817 * reiserfs_lock -> sysfs_mutex
2818 * Because the reiserfs mmap path creates the following dependency:
2819 * mm->mmap -> reiserfs_lock, hence we have
2820 * mm->mmap -> reiserfs_lock ->sysfs_mutex
2821 * This would ends up in a circular dependency with sysfs readdir path
2822 * which does sysfs_mutex -> mm->mmap_sem
2823 * This is fine because the reiserfs lock is useless in mount path,
2824 * at least until we call journal_begin. We keep it for paranoid
2825 * reasons.
2826 */
2833 }
2838 /* read journal header */
2846 }
2849 /* make sure that journal matches to the super block */
2854 "journal header magic %x (device %s) does "
2861 }
2876 }
2879 "journal first block %u, max trans len %u, max batch %u, "
2926 "allocation failed (%ld bytes). Journal is "
2927 "too large for available memory. Usually "
2931 }
2940 }
2945 }
2947 reiserfs_mounted_fs_count++;
2952 }
2957 free_and_return:
2960 }
2962 /*
2963 ** test for a polite end of the current transaction. Used by file_write, and should
2964 ** be used by delete to make sure they don't write more than can fit inside a single
2965 ** transaction
2966 */
2969 {
2972 /* cannot restart while nested */
2982 }
2983 /* protected by the BKL here */
2987 }
2989 /* this must be called inside a transaction, and requires the
2990 ** kernel_lock to be held
2991 */
2993 {
2999 }
3001 /* this must be called without a transaction started, and does not
3002 ** require BKL
3003 */
3005 {
3009 }
3011 /* this must be called without a transaction started, and does not
3012 ** require BKL
3013 */
3015 {
3019 }
3022 {
3023 wait_queue_t wait;
3027 /*
3028 * we don't want to use wait_event here because
3029 * we only want to wait once.
3030 */
3038 }
3041 }
3044 {
3048 }
3051 {
3063 }
3069 }
3070 }
3072 /* join == true if you must join an existing transaction.
3073 ** join == false if you can deal with waiting for others to finish
3074 **
3075 ** this will block until the transaction is joinable. send the number of blocks you
3076 ** expect to use in nblocks.
3077 */
3081 {
3093 /* set here for journal_join */
3097 relock:
3103 }
3113 }
3116 /* if there is no room in the journal OR
3117 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3118 ** we don't sleep if there aren't other writers
3119 */
3122 (!join
3139 sched_count++;
3142 }
3143 }
3144 /* don't mess with joining the transaction if all we have to do is
3145 * wait for someone else to do a commit
3146 */
3151 }
3153 }
3158 /* someone might have ended the transaction while we joined */
3163 }
3170 }
3171 /* we are the first writer, set trans_id */
3174 }
3185 out_fail:
3187 /* Re-set th->t_super, so we can properly keep track of how many
3188 * persistent transactions there are. We need to do this so if this
3189 * call is part of a failed restart_transaction, we can free it later */
3192 }
3195 super_block
3198 {
3202 /* if we're nesting into an existing transaction. It will be
3203 ** persistent on its own
3204 */
3211 }
3219 }
3223 }
3226 {
3231 else
3236 }
3238 }
3242 {
3245 /* this keeps do_journal_end from NULLing out the current->journal_info
3246 ** pointer
3247 */
3251 }
3255 {
3258 /* this keeps do_journal_end from NULLing out the current->journal_info
3259 ** pointer
3260 */
3264 }
3268 {
3274 /* we are nesting into the current transaction */
3281 "BAD: refcount <= 1, but "
3285 /* we've ended up with a handle from a different filesystem.
3286 ** save it and restore on journal_end. This should never
3287 ** really happen...
3288 */
3293 }
3296 }
3300 /* I guess this boils down to being the reciprocal of clm-2100 above.
3301 * If do_journal_begin_r fails, we need to put it back, since journal_end
3302 * won't be called to do it. */
3305 else
3309 }
3311 /*
3312 ** puts bh into the current transaction. If it was already there, reorders removes the
3313 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3314 **
3315 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3316 ** transaction is committed.
3317 **
3318 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3319 */
3322 {
3334 }
3340 /* already in this transaction, we are done */
3344 }
3346 /* this must be turned into a panic instead of a warning. We can't allow
3347 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3348 ** could get to disk too early. NOT GOOD.
3349 */
3352 "buffer %llu bad state "
3359 }
3366 }
3367 /* this error means I've screwed up, and we've overflowed the transaction.
3368 ** Nothing can be done here, except make the FS readonly or panic.
3369 */
3374 }
3380 }
3384 }
3388 /* now put this guy on the end */
3393 }
3398 }
3409 }
3410 }
3420 }
3422 }
3426 {
3434 }
3441 /* we aren't allowed to close a nested transaction on a different
3442 ** filesystem from the one in the task struct
3443 */
3449 }
3453 }
3454 }
3456 /* removes from the current transaction, relsing and descrementing any counters.
3457 ** also files the removed buffer directly onto the clean list
3458 **
3459 ** called by journal_mark_freed when a block has been deleted
3460 **
3461 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3462 */
3465 {
3474 }
3478 }
3481 }
3484 }
3487 }
3501 }
3503 }
3508 }
3510 /*
3511 ** for any cnode in a journal list, it can only be dirtied of all the
3512 ** transactions that include it are committed to disk.
3513 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3514 ** and 0 if you aren't
3515 **
3516 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3517 ** blocks for a given transaction on disk
3518 **
3519 */
3521 {
3527 /* first test hprev. These are all newer than cn, so any node here
3528 ** with the same block number and dev means this node can't be sent
3529 ** to disk right now.
3530 */
3535 }
3537 }
3538 /* then test hnext. These are all older than cn. As long as they
3539 ** are committed to the log, it is safe to write cn to disk
3540 */
3547 }
3549 }
3551 }
3553 /* syncs the commit blocks, but does not force the real buffers to disk
3554 ** will wait until the current transaction is done/committed before returning
3555 */
3558 {
3562 /* you can sync while nested, very, very bad */
3568 }
3570 }
3572 /*
3573 ** writeback the pending async commits to disk
3574 */
3576 {
3585 /* last entry is the youngest, commit it and you get everything */
3589 }
3591 }
3593 /*
3594 ** flushes any old transactions to disk
3595 ** ends the current transaction if it is too old
3596 */
3598 {
3604 /* safety check so we don't flush while we are replaying the log during
3605 * mount
3606 */
3609 }
3611 /* check the current transaction. If there are no writers, and it is
3612 * too old, finish it, and force the commit blocks to disk
3613 */
3625 /* we're only being called from kreiserfsd, it makes no sense to do
3626 ** an async commit so that kreiserfsd can do it later
3627 */
3629 }
3630 }
3632 }
3634 /*
3635 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3636 **
3637 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3638 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3639 ** flushes the commit list and returns 0.
3640 **
3641 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3642 **
3643 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3644 */
3648 {
3663 }
3666 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3668 }
3670 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3671 ** will be dealt with by next transaction that actually writes something, but should be taken
3672 ** care of in this trans
3673 */
3676 /* if wcount > 0, and we are called to with flush or commit_now,
3677 ** we wait on j_join_wait. We will wake up when the last writer has
3678 ** finished the transaction, and started it on its way to the disk.
3679 ** Then, we flush the commit or journal list, and just return 0
3680 ** because the rest of journal end was already done for this transaction.
3681 */
3693 }
3696 /* sleep while the current transaction is still j_jlocked */
3705 }
3707 }
3708 }
3715 }
3717 }
3720 }
3722 /* deal with old transactions where we are the last writers */
3727 }
3728 /* don't batch when someone is waiting on j_join_wait */
3729 /* don't batch when syncing the commit or flushing the whole trans */
3737 }
3743 }
3745 }
3747 /*
3748 ** Does all the work that makes deleting blocks safe.
3749 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3750 **
3751 ** otherwise:
3752 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3753 ** before this transaction has finished.
3754 **
3755 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3756 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3757 ** the block can't be reallocated yet.
3758 **
3759 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3760 */
3763 {
3775 }
3776 /* if it is journal new, we just remove it from this transaction */
3783 /* set the bit for this block in the journal bitmap for this transaction */
3788 }
3791 /* Note, the entire while loop is not allowed to schedule. */
3796 }
3799 /* find all older transactions with this block, make sure they don't try to write it out */
3801 blocknr);
3807 /* remove_from_transaction will brelse the buffer if it was
3808 ** in the current trans
3809 */
3811 bh);
3814 bh);
3822 }
3823 }
3827 j_nonzerolen));
3828 }
3830 }
3831 }
3833 }
3834 }
3839 }
3842 {
3846 }
3848 /*
3849 * returns -1 on error, 0 if no commits/barriers were done and 1
3850 * if a transaction was actually committed and the barrier was done
3851 */
3854 {
3860 /* is it from the current transaction, or from an unknown transaction? */
3863 /* try to let other writers come in and grow this transaction */
3867 }
3873 /* someone might have ended this transaction while we joined */
3880 }
3887 /* this gets tricky, we have to make sure the journal list in
3888 * the inode still exists. We know the list is still around
3889 * if we've got a larger transaction id than the oldest list
3890 */
3891 flush_commit_only:
3893 /*
3894 * we only set ret to 1 when we know for sure
3895 * the barrier hasn't been started yet on the commit
3896 * block.
3897 */
3903 }
3904 }
3905 /* otherwise the list is gone, and long since committed */
3907 }
3910 {
3914 /* for the whole inode, assume unset id means it was
3915 * changed in the current transaction. More conservative
3916 */
3920 /* jl will be updated in __commit_trans_jl */
3921 }
3924 }
3928 {
3933 }
3943 }
3944 }
3946 }
3949 /*
3950 ** before we can change a metadata block, we have to make sure it won't
3951 ** be written to disk while we are altering it. So, we must:
3952 ** clean it
3953 ** wait on it.
3954 **
3955 */
3958 {
3965 }
3970 }
3973 }
3976 {
3985 /* this check should always be run, to send old lists to disk */
3992 }
3993 }
3994 }
3996 /*
3997 ** long and ugly. If flush, will not return until all commit
3998 ** blocks and all real buffers in the trans are on disk.
3999 ** If no_async, won't return until all commit blocks are on disk.
4000 **
4001 ** keep reading, there are comments as you go along
4002 **
4003 ** If the journal is aborted, we just clean up. Things like flushing
4004 ** journal lists, etc just won't happen.
4005 */
4009 {
4031 /* protect flush_older_commits from doing mistakes if the
4032 transaction ID counter gets overflowed. */
4045 }
4051 }
4055 }
4057 /* check_journal_end locks the journal, and unlocks if it does not return 1
4058 ** it tells us if we should continue with the journal_end, or just return
4059 */
4065 }
4067 /* check_journal_end might set these, check again */
4070 }
4072 /*
4073 ** j must wait means we have to flush the log blocks, and the real blocks for
4074 ** this transaction
4075 */
4078 }
4079 #ifdef REISERFS_PREALLOCATE
4080 /* quota ops might need to nest, setup the journal_info pointer for them
4081 * and raise the refcount so that it is > 0. */
4085 * the transaction */
4088 #endif
4090 /* setup description block */
4091 d_bh =
4101 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4110 /* init this journal list */
4113 /* we lock the commit before doing anything because
4114 * we want to make sure nobody tries to run flush_commit_list until
4115 * the new transaction is fully setup, and we've already flushed the
4116 * ordered bh list
4117 */
4120 /* save the transaction id in case we need to commit it later */
4133 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4134 ** for each real block, add it to the journal list hash,
4135 ** copy into real block index array in the commit or desc block
4136 */
4144 }
4147 }
4152 }
4154 /* make sure the block we are trying to log is not a block
4155 of journal or reserved area */
4160 "Trying to log block %lu, "
4163 }
4176 }
4178 i--;
4179 }
4180 }
4186 /* special check in case all buffers in the journal were marked for not logging */
4189 /* we're about to dirty all the log blocks, mark the description block
4190 * dirty now too. Don't mark the commit block dirty until all the
4191 * others are on disk
4192 */
4195 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4201 /* copy all the real blocks into log area. dirty log blocks */
4206 tmp_bh =
4210 jindex) %
4220 jindex++;
4224 /* JDirty cleared sometime during transaction. don't log this one */
4226 "BAD, buffer in journal hash, "
4229 }
4236 }
4238 /* we are done with both the c_bh and d_bh, but
4239 ** c_bh must be written after all other commit blocks,
4240 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4241 */
4245 /* now it is safe to insert this transaction on the main list */
4250 /* reset journal values for the next transaction */
4261 /* check for trans_id overflow */
4271 // make sure reiserfs_add_jh sees the new current_jl before we
4272 // write out the tails
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 and
4278 * clean, if we crash before the later transaction commits, the data block
4279 * is lost.
4280 */
4286 }
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 */
4303 /* if the next transaction has any chance of wrapping, flush
4304 ** transactions that might get overwritten. If any journal lists are very
4305 ** old flush them as well.
4306 */
4307 first_jl:
4318 /* if we don't cross into the next transaction and we don't
4319 * wrap, there is no way we can overlap any later transactions
4320 * break now
4321 */
4323 }
4333 /* we don't overlap anything from out start to the end of the
4334 * log, and our wrapped portion doesn't overlap anything at
4335 * the start of the log. We can break
4336 */
4338 }
4339 }
4340 }
4349 }
4353 /* wake up any body waiting to join. */
4360 }
4361 out:
4365 /* Re-set th->t_super, so we can properly keep track of how many
4366 * persistent transactions there are. We need to do this so if this
4367 * call is part of a failed restart_transaction, we can free it later */
4371 }
4373 /* Send the file system read only and refuse new transactions */
4375 {
4386 #ifdef CONFIG_REISERFS_CHECK
4388 #endif
4389 }