| blob | 076c8b19468297dc71add1987b249ab054d6e20d |
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>
53 #include <linux/slab.h>
55 #include <asm/system.h>
57 /* gets a struct reiserfs_journal_list * from a list head */
58 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
59 j_list))
60 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
61 j_working_list))
63 /* the number of mounted filesystems. This is used to decide when to
64 ** start and kill the commit workqueue
65 */
71 structs at 4k */
74 /* cnode stat bits. Move these into reiserfs_fs.h */
77 #define BLOCK_FREED_HOLDER 3 /* this block was freed during this transaction, and can't be written */
80 #define BLOCK_DIRTIED 5
82 /* journal list state bits */
83 #define LIST_TOUCHED 1
84 #define LIST_DIRTY 2
87 /* flags for do_journal_end */
109 /* values for join in do_journal_begin_r */
114 };
121 {
125 }
127 /*
128 ** clears BH_Dirty and sticks the buffer on the clean list. Called because I can't allow refile_buffer to
129 ** make schedule happen after I've freed a block. Look at remove_from_transaction and journal_mark_freed for
130 ** more details.
131 */
133 {
137 }
139 }
143 {
150 }
155 }
159 }
162 {
168 repeat:
176 }
181 }
183 }
186 {
195 }
196 }
199 {
210 }
211 }
212 }
215 b_blocknr_t block,
217 {
223 }
226 }
230 {
239 }
240 }
241 }
243 /*
244 ** only call this on FS unmount.
245 */
248 {
257 }
259 }
262 {
274 }
277 }
279 /*
280 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
281 ** jb_array is the array to be filled in.
282 */
286 {
301 }
303 }
307 }
309 }
311 /*
312 ** find an available list bitmap. If you can't find one, flush a commit list
313 ** and try again
314 */
316 struct reiserfs_journal_list
318 {
333 }
336 }
337 }
340 }
343 }
345 /*
346 ** allocates a new chunk of X nodes, and links them all together as a list.
347 ** Uses the cnode->next and cnode->prev pointers
348 ** returns NULL on failure
349 */
351 {
356 }
360 }
367 }
370 }
372 /*
373 ** pulls a cnode off the free list, or returns NULL on failure
374 */
376 {
384 }
390 }
393 }
397 }
399 /*
400 ** returns a cnode to the free list
401 */
404 {
411 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
415 }
416 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
418 }
421 {
424 }
426 /* return a cnode with same dev, block number and size in table, or null if not found */
428 super_block
430 struct
431 reiserfs_journal_cnode
434 {
441 }
443 }
445 /*
446 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
447 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
448 ** being overwritten by a replay after crashing.
449 **
450 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
451 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
452 ** sure you never write the block without logging it.
453 **
454 ** next_zero_bit is a suggestion about the next block to try for find_forward.
455 ** when bl is rejected because it is set in a journal list bitmap, we search
456 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
457 ** through next_zero_bit for find_forward to try.
458 **
459 ** Just because we return something in next_zero_bit does not mean we won't
460 ** reject it on the next call to reiserfs_in_journal
461 **
462 */
466 {
476 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
477 ** if we crash before the transaction that freed it commits, this transaction won't
478 ** have committed either, and the block will never be written
479 */
487 data)) {
491 data),
495 }
496 }
497 }
500 /* is it in any old transactions? */
505 }
507 /* is it in the current transaction. This should never happen */
511 }
514 /* safe for reuse */
516 }
518 /* insert cn into table
519 */
522 {
530 }
532 }
534 /* lock the current transaction */
536 {
540 }
542 /* unlock the current transaction */
544 {
546 }
549 {
551 }
555 {
559 }
562 }
564 /*
565 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
566 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
567 ** transaction.
568 */
571 {
576 }
579 }
583 {
592 }
593 }
595 }
597 /*
598 * If page->mapping was null, we failed to truncate this page for
599 * some reason. Most likely because it was truncated after being
600 * logged via data=journal.
601 *
602 * This does a check to see if the buffer belongs to one of these
603 * lost pages before doing the final put_bh. If page->mapping was
604 * null, it tries to free buffers on the page, which should make the
605 * final page_cache_release drop the page from the lru.
606 */
608 {
619 }
620 }
623 {
630 }
633 else
638 }
641 {
644 else
648 }
651 {
661 }
664 {
671 }
673 #define CHUNK_SIZE 32
677 };
680 {
685 }
688 }
691 {
696 }
699 }
703 {
714 }
716 }
720 {
727 }
729 }
730 }
732 /*
733 * we want to free the jh when the buffer has been written
734 * and waited on
735 */
737 {
750 }
751 }
755 {
763 }
767 no_jh:
771 /* buffer must be locked for __add_jh, should be able to have
772 * two adds at the same time
773 */
777 }
783 }
786 }
789 {
791 }
793 {
795 }
797 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
802 {
820 }
828 }
829 /* in theory, dirty non-uptodate buffers should never get here,
830 * but the upper layer io error paths still have a few quirks.
831 * Handle them here as gracefully as we can
832 */
836 }
843 }
844 loop_next:
847 }
852 }
863 }
866 }
867 /* ugly interaction with invalidatepage here.
868 * reiserfs_invalidate_page will pin any buffer that has a valid
869 * journal head from an older transaction. If someone else sets
870 * our buffer dirty after we write it in the first loop, and
871 * then someone truncates the page away, nobody will ever write
872 * the buffer. We're safe if we write the page one last time
873 * after freeing the journal header.
874 */
879 }
882 }
885 }
889 {
898 find_first:
899 /*
900 * first we walk backwards to find the oldest uncommitted transation
901 */
912 }
914 /* if we didn't find any older uncommitted transactions, return now */
917 }
930 /* list we were called with is gone, return */
934 /* the one we just flushed is gone, this means all
935 * older lists are also gone, so first_jl is no longer
936 * valid either. Go back to the beginning.
937 */
941 }
942 }
948 }
949 }
951 }
954 {
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 transaction 1)
970 ** Before the commit block can by written, every other log block must be safely on disk
971 **
972 */
975 {
977 b_blocknr_t bn;
988 }
992 /* before we can put our commit blocks on disk, we have to make sure everyone older than
993 ** us is on disk too
994 */
1001 /* list disappeared during flush_older_commits. return */
1003 }
1004 }
1006 /* make sure nobody is trying to flush this one at the same time */
1012 }
1015 /* this commit is done, exit */
1019 }
1022 }
1027 /*
1028 * We might sleep in numerous places inside
1029 * write_ordered_buffers. Relax the write lock.
1030 */
1037 }
1039 /*
1040 * for the description block and all the log blocks, submit any buffers
1041 * that haven't already reached the disk. Try to write at least 256
1042 * log blocks. later on, we will only wait on blocks that correspond
1043 * to this transaction, but while we're unplugging we might as well
1044 * get a chunk of data on there.
1045 */
1059 }
1061 }
1062 }
1073 // since we're using ll_rw_blk above, it might have skipped over
1074 // a locked buffer. Double check here
1075 //
1076 /* redundant, sync_dirty_buffer() checks */
1081 }
1083 #ifdef CONFIG_REISERFS_CHECK
1086 #endif
1088 }
1092 }
1096 /* If there was a write error in the journal - we can't commit
1097 * this transaction - it will be invalid and, if successful,
1098 * will just end up propagating the write error out to
1099 * the file system. */
1107 else
1110 }
1112 /* If there was a write error in the journal - we can't commit this
1113 * transaction - it will be invalid and, if successful, will just end
1114 * up propagating the write error out to the filesystem. */
1116 #ifdef CONFIG_REISERFS_CHECK
1118 #endif
1120 }
1126 }
1129 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1134 /* mark the metadata dirty */
1141 }
1143 put_jl:
1148 __func__);
1151 }
1153 /*
1154 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1155 ** returns NULL if it can't find anything
1156 */
1158 reiserfs_journal_cnode
1160 {
1168 }
1170 }
1172 }
1175 {
1185 }
1187 }
1194 /*
1195 ** once all the real blocks have been flushed, it is safe to remove them from the
1196 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1197 ** block to be reallocated for data blocks if it had been deleted.
1198 */
1202 {
1207 /* which is better, to lock once around the whole loop, or
1208 ** to lock for each call to remove_journal_hash?
1209 */
1217 }
1221 }
1225 }
1227 }
1229 /*
1230 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1231 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1232 ** releasing blocks in this transaction for reuse as data blocks.
1233 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1234 **
1235 */
1239 {
1252 #ifdef CONFIG_REISERFS_CHECK
1255 #endif
1257 }
1258 }
1262 b_data);
1272 else
1280 }
1281 }
1283 }
1288 {
1290 }
1292 /*
1293 ** flush any and all journal lists older than you are
1294 ** can only be called from flush_journal_list
1295 */
1298 {
1304 /* we know we are the only ones flushing things, no extra race
1305 * protection is required.
1306 */
1307 restart:
1309 /* Did we wrap? */
1315 /* do not flush all */
1318 /* other_jl is now deleted from the list */
1320 }
1322 }
1326 {
1331 }
1332 }
1334 /* flush a journal list, both commit and real blocks
1335 **
1336 ** always set flushall to 1, unless you are calling from inside
1337 ** flush_journal_list
1338 **
1339 ** IMPORTANT. This can only be called while there are no journal writers,
1340 ** and the journal is locked. That means it can only be called from
1341 ** do_journal_end, or by journal_release
1342 */
1345 {
1361 }
1364 /* if flushall == 0, the lock is already held */
1369 }
1376 }
1380 /* if all the work is already done, get out of here */
1384 }
1386 /* start by putting the commit list on disk. This will also flush
1387 ** the commit lists of any olders transactions
1388 */
1395 /* are we done now? */
1399 }
1401 /* loop through each cnode, see if we need to write it,
1402 ** or wait on a more recent transaction, or just ignore it
1403 */
1407 }
1413 /* blocknr of 0 is no longer in the hash, ignore it */
1416 }
1418 /* This transaction failed commit. Don't write out to the disk */
1423 /* the order is important here. We check pjl to make sure we
1424 ** don't clear BH_JDirty_wait if we aren't the one writing this
1425 ** block to disk
1426 */
1430 /* we do this to make sure nobody releases the buffer while
1431 ** we are working with it
1432 */
1440 /* everything with !pjl && jwait should be writable */
1442 }
1443 }
1445 /* if someone has this block in a newer transaction, just make
1446 ** sure they are committed, and don't try writing it to disk
1447 */
1452 }
1454 /* bh == NULL when the block got to disk on its own, OR,
1455 ** the block got freed in a future transaction
1456 */
1459 }
1461 /* this should never happen. kupdate_one_transaction has this list
1462 ** locked while it works, so we should never see a buffer here that
1463 ** is not marked JDirty_wait
1464 */
1467 "BAD! buffer %llu %cdirty %cjwait, "
1472 }
1474 /* we inc again because saved_bh gets decremented at free_cnode */
1481 else
1483 count++;
1489 }
1490 free_cnode:
1494 /* we incremented this to keep others from taking the buffer head away */
1499 }
1500 }
1501 }
1509 }
1518 }
1520 #ifdef CONFIG_REISERFS_CHECK
1523 #endif
1525 }
1526 /* note, we must clear the JDirty_wait bit after the up to date
1527 ** check, otherwise we race against our flushpage routine
1528 */
1532 /* drop one ref for us */
1534 /* drop one ref for journal_mark_dirty */
1536 }
1538 }
1539 }
1544 __func__);
1545 flush_older_and_return:
1547 /* before we can update the journal header block, we _must_ flush all
1548 ** real blocks from all older transactions to disk. This is because
1549 ** once the header block is updated, this transaction will not be
1550 ** replayed after a crash
1551 */
1554 }
1557 /* before we can remove everything from the hash tables for this
1558 ** transaction, we must make sure it can never be replayed
1559 **
1560 ** since we are only called from do_journal_end, we know for sure there
1561 ** are no allocations going on while we are flushing journal lists. So,
1562 ** we only need to update the journal header block for the last list
1563 ** being flushed
1564 */
1566 err =
1574 __func__);
1575 }
1585 }
1588 /* not strictly required since we are freeing the list, but it should
1589 * help find code using dead lists later on
1590 */
1603 }
1607 {
1615 /* if the blocknr == 0, this has been cleared from the hash,
1616 ** skip it
1617 */
1620 }
1623 next:
1626 }
1628 }
1633 {
1641 }
1645 /* if the blocknr == 0, this has been cleared from the hash,
1646 ** skip it
1647 */
1650 }
1653 /* we can race against journal_mark_freed when we try
1654 * to lock_buffer(cn->bh), so we have to inc the buffer
1655 * count, and recheck things after locking
1656 */
1665 ret++;
1667 /* note, cn->bh might be null now */
1669 }
1671 }
1672 next:
1675 }
1677 }
1679 /* used by flush_commit_list */
1682 {
1690 /* look for a more recent transaction that logged this
1691 ** buffer. Only the most recent transaction with a buffer in
1692 ** it is allowed to send that buffer to disk
1693 */
1698 /* if the buffer is prepared, it will either be logged
1699 * or restored. If restored, we need to make sure
1700 * it actually gets marked dirty
1701 */
1708 }
1709 }
1711 }
1713 }
1720 {
1733 }
1735 /* we've got j_flush_mutex held, nobody is going to delete any
1736 * of these lists out from underneath us
1737 */
1746 }
1751 transactions_flushed++;
1755 /* did we wrap? */
1758 }
1761 /* don't bother with older transactions */
1764 }
1767 }
1769 done:
1772 }
1774 /* for o_sync and fsync heavy applications, they tend to use
1775 ** all the journa list slots with tiny transactions. These
1776 ** trigger lots and lots of calls to update the header block, which
1777 ** adds seeks and slows things down.
1778 **
1779 ** This function tries to clear out a large chunk of the journal lists
1780 ** at once, which makes everything faster since only the newest journal
1781 ** list updates the header block
1782 */
1785 {
1798 /* in data logging mode, try harder to flush a lot of blocks */
1801 /* flush for 256 transactions or limit blocks, whichever comes first */
1806 }
1810 }
1816 }
1817 /* try to find a group of blocks we can flush across all the
1818 ** transactions, but only bother if we've actually spanned
1819 ** across multiple lists
1820 */
1823 }
1826 }
1828 /*
1829 ** removes any nodes in table with name block and dev as bh.
1830 ** only touchs the hnext and hprev pointers.
1831 */
1836 {
1843 }
1851 }
1856 }
1864 }
1866 }
1867 }
1870 {
1880 }
1881 /* j_header_bh is on the journal dev, make sure not to release the journal
1882 * dev until we brelse j_header_bh
1883 */
1886 }
1888 /*
1889 ** call on unmount. Only set error to 1 if you haven't made your way out
1890 ** of read_super() yet. Any other caller must keep error at 0.
1891 */
1894 {
1899 /* we only want to flush out transactions if we were called with error == 0
1900 */
1902 /* end the current trans */
1906 /* make sure something gets logged to force our way into the flush code */
1915 }
1916 }
1918 /* this also catches errors during the do_journal_end above */
1928 }
1929 }
1931 reiserfs_mounted_fs_count--;
1932 /* wait for all commits to finish */
1935 /*
1936 * We must release the write lock here because
1937 * the workqueue job (flush_async_commit) needs this lock
1938 */
1945 }
1952 }
1954 /*
1955 ** call on unmount. flush all journal trans, release all alloc'd ram
1956 */
1959 {
1961 }
1963 /*
1964 ** only call from an error condition inside reiserfs_read_super!
1965 */
1968 {
1970 }
1972 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1976 {
1982 }
1984 }
1986 /* returns 0 if it did not find a description block
1987 ** returns -1 if it found a corrupt commit block
1988 ** returns 1 if both desc and commit were valid
1989 */
1994 {
2009 "journal-986: transaction "
2010 "is valid returning because trans_id %d is greater than "
2015 }
2019 "journal-1087: transaction "
2020 "is valid returning because mount_id %d is less than "
2025 }
2028 "Bad transaction length %d "
2032 }
2035 /* ok, we have a journal description block, lets see if the transaction was valid */
2036 c_bh =
2046 "journal_transaction_is_valid, commit offset %ld had bad "
2057 "journal-1004: "
2058 "transaction_is_valid setting oldest invalid trans_id "
2061 }
2063 }
2066 "journal-1006: found valid "
2075 }
2076 }
2079 {
2083 }
2084 }
2086 /*
2087 ** given the start, and values for the oldest acceptable transactions,
2088 ** this either reads in a replays a transaction, or returns because the transaction
2089 ** is invalid, or too old.
2090 */
2096 {
2121 cur_dblock -
2125 }
2128 "journal_read_trans skipping because %d is != "
2130 newest_mount_id);
2133 }
2140 }
2144 "journal_read_transaction, "
2153 }
2161 }
2164 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2177 }
2178 /* get all the buffer heads */
2195 }
2198 "REPLAY FAILURE fsck required! "
2199 "Block to replay is outside of "
2202 }
2203 /* make sure we don't try to replay onto log or reserved area */
2207 "REPLAY FAILURE fsck required! "
2209 abort_replay:
2217 }
2218 }
2219 /* read in the log blocks, memcpy to the corresponding real block */
2229 "REPLAY FAILURE fsck required! "
2239 }
2244 }
2245 /* flush out the real blocks */
2249 }
2254 "REPLAY FAILURE, fsck required! "
2263 }
2265 }
2266 cur_dblock =
2274 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2278 /* check for trans_id overflow */
2286 }
2288 /* This function reads blocks starting from block and to max_block of bufsize
2289 size (but no more than BUFNR blocks at a time). This proved to improve
2290 mounting speed on self-rebuilding raid5 arrays at least.
2291 Right now it is only used from journal code. But later we might use it
2292 from other places.
2293 Note: Do not use journal_getblk/sb_getblk functions here! */
2296 b_blocknr_t max_block)
2297 {
2309 }
2319 }
2329 }
2331 /*
2332 ** read and replay the log
2333 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2334 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2335 **
2336 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2337 **
2338 ** On exit, it sets things up so the first transaction will work correctly.
2339 */
2341 {
2363 /* step 1, read in the journal header block. Check the transaction it says
2364 ** is the first unflushed, and if that transaction is not valid,
2365 ** replay is done
2366 */
2372 }
2377 oldest_start =
2383 "journal-1153: found in "
2384 "header: first_unflushed_offset %d, last_flushed_trans_id "
2389 /* now, we try to read the first unflushed offset. If it is not valid,
2390 ** there is nothing more we can do, and it makes no sense to read
2391 ** through the whole log.
2392 */
2393 d_bh =
2400 }
2403 }
2405 /* ok, there are transactions that need to be replayed. start with the first log block, find
2406 ** all the valid transactions, and pick out the oldest.
2407 */
2412 /* Note that it is required for blocksize of primary fs device and journal
2413 device to be the same */
2414 d_bh =
2419 ret =
2430 "journal-1179: Setting "
2432 oldest_start -
2433 SB_ONDISK_JOURNAL_1st_BLOCK
2436 /* one we just read was older */
2440 "journal-1180: Resetting "
2442 oldest_start -
2443 SB_ONDISK_JOURNAL_1st_BLOCK
2445 }
2449 "journal-1299: Setting "
2452 }
2455 cur_dblock++;
2456 }
2458 }
2460 start_log_replay:
2464 "journal-1206: Starting replay "
2467 oldest_trans_id);
2469 }
2472 ret =
2479 }
2480 cur_dblock =
2482 replay_count++;
2485 }
2490 }
2491 /* j_start does not get set correctly if we don't replay any transactions.
2492 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2493 ** copy the trans_id from the header
2494 */
2499 /* check for trans_id overflow */
2507 }
2515 }
2519 /* replay failed, caller must call free_journal_ram and abort
2520 ** the mount
2521 */
2523 }
2525 }
2528 {
2540 }
2543 {
2545 }
2549 {
2559 }
2564 }
2566 }
2571 {
2573 dev_t jdev;
2586 /* there is no "jdev" option and journal is on separate device */
2602 }
2605 }
2608 }
2620 }
2627 }
2629 /**
2630 * When creating/tuning a file system user can assign some
2631 * journal params within boundaries which depend on the ratio
2632 * blocksize/standard_blocksize.
2633 *
2634 * For blocks >= standard_blocksize transaction size should
2635 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2636 * then JOURNAL_TRANS_MAX_DEFAULT.
2637 *
2638 * For blocks < standard_blocksize these boundaries should be
2639 * decreased proportionally.
2640 */
2641 #define REISERFS_STANDARD_BLKSIZE (4096)
2645 {
2647 /* Non-default journal params.
2648 Do sanity check for them. */
2656 JOURNAL_MIN_RATIO) {
2658 "bad transaction max size (%u). "
2661 }
2665 "bad transaction max batch (%u). "
2668 }
2670 /* Default journal params.
2671 The file system was created by old version
2672 of mkreiserfs, so some fields contain zeros,
2673 and we need to advise proper values for them */
2678 }
2682 }
2684 }
2686 /*
2687 ** must be called once on fs mount. calls journal_read for you
2688 */
2691 {
2701 /*
2702 * Unlock here to avoid various RECLAIM-FS-ON <-> IN-RECLAIM-FS
2703 * dependency inversion warnings.
2704 */
2712 }
2727 /* reserved for journal area support */
2729 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2733 REISERFS_DISK_OFFSET_IN_BYTES /
2736 /* Sanity check to see is the standard journal fitting withing first bitmap
2737 (actual for small blocksizes) */
2742 "journal does not fit for area addressed "
2743 "by first of bitmap blocks. It starts at "
2749 }
2751 /*
2752 * We need to unlock here to avoid creating the following
2753 * dependency:
2754 * reiserfs_lock -> sysfs_mutex
2755 * Because the reiserfs mmap path creates the following dependency:
2756 * mm->mmap -> reiserfs_lock, hence we have
2757 * mm->mmap -> reiserfs_lock ->sysfs_mutex
2758 * This would ends up in a circular dependency with sysfs readdir path
2759 * which does sysfs_mutex -> mm->mmap_sem
2760 * This is fine because the reiserfs lock is useless in mount path,
2761 * at least until we call journal_begin. We keep it for paranoid
2762 * reasons.
2763 */
2770 }
2775 /* read journal header */
2783 }
2786 /* make sure that journal matches to the super block */
2791 "journal header magic %x (device %s) does "
2798 }
2813 }
2816 "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 }
2877 }
2882 }
2884 reiserfs_mounted_fs_count++;
2889 }
2894 free_and_return:
2897 }
2899 /*
2900 ** test for a polite end of the current transaction. Used by file_write, and should
2901 ** be used by delete to make sure they don't write more than can fit inside a single
2902 ** transaction
2903 */
2906 {
2909 /* cannot restart while nested */
2919 }
2920 /* protected by the BKL here */
2924 }
2926 /* this must be called inside a transaction, and requires the
2927 ** kernel_lock to be held
2928 */
2930 {
2936 }
2938 /* this must be called without a transaction started, and does not
2939 ** require BKL
2940 */
2942 {
2946 }
2948 /* this must be called without a transaction started, and does not
2949 ** require BKL
2950 */
2952 {
2956 }
2959 {
2960 wait_queue_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 {
3000 }
3006 }
3007 }
3009 /* join == true if you must join an existing transaction.
3010 ** join == false if you can deal with waiting for others to finish
3011 **
3012 ** this will block until the transaction is joinable. send the number of blocks you
3013 ** expect to use in nblocks.
3014 */
3018 {
3030 /* set here for journal_join */
3034 relock:
3040 }
3050 }
3053 /* if there is no room in the journal OR
3054 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3055 ** we don't sleep if there aren't other writers
3056 */
3059 (!join
3076 sched_count++;
3079 }
3080 }
3081 /* don't mess with joining the transaction if all we have to do is
3082 * wait for someone else to do a commit
3083 */
3088 }
3090 }
3095 /* someone might have ended the transaction while we joined */
3100 }
3107 }
3108 /* we are the first writer, set trans_id */
3111 }
3122 out_fail:
3124 /* Re-set th->t_super, so we can properly keep track of how many
3125 * persistent transactions there are. We need to do this so if this
3126 * call is part of a failed restart_transaction, we can free it later */
3129 }
3132 super_block
3135 {
3139 /* if we're nesting into an existing transaction. It will be
3140 ** persistent on its own
3141 */
3148 }
3156 }
3160 }
3163 {
3168 else
3173 }
3175 }
3179 {
3182 /* this keeps do_journal_end from NULLing out the current->journal_info
3183 ** pointer
3184 */
3188 }
3192 {
3195 /* this keeps do_journal_end from NULLing out the current->journal_info
3196 ** pointer
3197 */
3201 }
3205 {
3211 /* we are nesting into the current transaction */
3218 "BAD: refcount <= 1, but "
3222 /* we've ended up with a handle from a different filesystem.
3223 ** save it and restore on journal_end. This should never
3224 ** really happen...
3225 */
3230 }
3233 }
3237 /* I guess this boils down to being the reciprocal of clm-2100 above.
3238 * If do_journal_begin_r fails, we need to put it back, since journal_end
3239 * won't be called to do it. */
3242 else
3246 }
3248 /*
3249 ** puts bh into the current transaction. If it was already there, reorders removes the
3250 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3251 **
3252 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3253 ** transaction is committed.
3254 **
3255 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3256 */
3259 {
3271 }
3277 /* already in this transaction, we are done */
3281 }
3283 /* this must be turned into a panic instead of a warning. We can't allow
3284 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3285 ** could get to disk too early. NOT GOOD.
3286 */
3289 "buffer %llu bad state "
3296 }
3303 }
3304 /* this error means I've screwed up, and we've overflowed the transaction.
3305 ** Nothing can be done here, except make the FS readonly or panic.
3306 */
3311 }
3317 }
3321 }
3325 /* now put this guy on the end */
3330 }
3335 }
3346 }
3347 }
3357 }
3359 }
3363 {
3371 }
3378 /* we aren't allowed to close a nested transaction on a different
3379 ** filesystem from the one in the task struct
3380 */
3386 }
3390 }
3391 }
3393 /* removes from the current transaction, relsing and descrementing any counters.
3394 ** also files the removed buffer directly onto the clean list
3395 **
3396 ** called by journal_mark_freed when a block has been deleted
3397 **
3398 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3399 */
3402 {
3411 }
3415 }
3418 }
3421 }
3424 }
3438 }
3440 }
3445 }
3447 /*
3448 ** for any cnode in a journal list, it can only be dirtied of all the
3449 ** transactions that include it are committed to disk.
3450 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3451 ** and 0 if you aren't
3452 **
3453 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3454 ** blocks for a given transaction on disk
3455 **
3456 */
3458 {
3464 /* first test hprev. These are all newer than cn, so any node here
3465 ** with the same block number and dev means this node can't be sent
3466 ** to disk right now.
3467 */
3472 }
3474 }
3475 /* then test hnext. These are all older than cn. As long as they
3476 ** are committed to the log, it is safe to write cn to disk
3477 */
3484 }
3486 }
3488 }
3490 /* syncs the commit blocks, but does not force the real buffers to disk
3491 ** will wait until the current transaction is done/committed before returning
3492 */
3495 {
3499 /* you can sync while nested, very, very bad */
3505 }
3507 }
3509 /*
3510 ** writeback the pending async commits to disk
3511 */
3513 {
3522 /* last entry is the youngest, commit it and you get everything */
3526 }
3528 }
3530 /*
3531 ** flushes any old transactions to disk
3532 ** ends the current transaction if it is too old
3533 */
3535 {
3541 /* safety check so we don't flush while we are replaying the log during
3542 * mount
3543 */
3546 }
3548 /* check the current transaction. If there are no writers, and it is
3549 * too old, finish it, and force the commit blocks to disk
3550 */
3562 /* we're only being called from kreiserfsd, it makes no sense to do
3563 ** an async commit so that kreiserfsd can do it later
3564 */
3566 }
3567 }
3569 }
3571 /*
3572 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3573 **
3574 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3575 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3576 ** flushes the commit list and returns 0.
3577 **
3578 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3579 **
3580 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3581 */
3585 {
3600 }
3603 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3605 }
3607 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3608 ** will be dealt with by next transaction that actually writes something, but should be taken
3609 ** care of in this trans
3610 */
3613 /* if wcount > 0, and we are called to with flush or commit_now,
3614 ** we wait on j_join_wait. We will wake up when the last writer has
3615 ** finished the transaction, and started it on its way to the disk.
3616 ** Then, we flush the commit or journal list, and just return 0
3617 ** because the rest of journal end was already done for this transaction.
3618 */
3630 }
3633 /* sleep while the current transaction is still j_jlocked */
3642 }
3644 }
3645 }
3652 }
3654 }
3657 }
3659 /* deal with old transactions where we are the last writers */
3664 }
3665 /* don't batch when someone is waiting on j_join_wait */
3666 /* don't batch when syncing the commit or flushing the whole trans */
3674 }
3680 }
3682 }
3684 /*
3685 ** Does all the work that makes deleting blocks safe.
3686 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3687 **
3688 ** otherwise:
3689 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3690 ** before this transaction has finished.
3691 **
3692 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3693 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3694 ** the block can't be reallocated yet.
3695 **
3696 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3697 */
3700 {
3712 }
3713 /* if it is journal new, we just remove it from this transaction */
3720 /* set the bit for this block in the journal bitmap for this transaction */
3725 }
3728 /* Note, the entire while loop is not allowed to schedule. */
3733 }
3736 /* find all older transactions with this block, make sure they don't try to write it out */
3738 blocknr);
3744 /* remove_from_transaction will brelse the buffer if it was
3745 ** in the current trans
3746 */
3748 bh);
3751 bh);
3759 }
3760 }
3764 j_nonzerolen));
3765 }
3767 }
3768 }
3770 }
3771 }
3776 }
3779 {
3783 }
3785 /*
3786 * returns -1 on error, 0 if no commits/barriers were done and 1
3787 * if a transaction was actually committed and the barrier was done
3788 */
3791 {
3797 /* is it from the current transaction, or from an unknown transaction? */
3800 /* try to let other writers come in and grow this transaction */
3804 }
3810 /* someone might have ended this transaction while we joined */
3817 }
3824 /* this gets tricky, we have to make sure the journal list in
3825 * the inode still exists. We know the list is still around
3826 * if we've got a larger transaction id than the oldest list
3827 */
3828 flush_commit_only:
3830 /*
3831 * we only set ret to 1 when we know for sure
3832 * the barrier hasn't been started yet on the commit
3833 * block.
3834 */
3840 }
3841 }
3842 /* otherwise the list is gone, and long since committed */
3844 }
3847 {
3851 /* for the whole inode, assume unset id means it was
3852 * changed in the current transaction. More conservative
3853 */
3857 /* jl will be updated in __commit_trans_jl */
3858 }
3861 }
3865 {
3870 }
3880 }
3881 }
3883 }
3886 /*
3887 ** before we can change a metadata block, we have to make sure it won't
3888 ** be written to disk while we are altering it. So, we must:
3889 ** clean it
3890 ** wait on it.
3891 **
3892 */
3895 {
3902 }
3907 }
3910 }
3913 {
3922 /* this check should always be run, to send old lists to disk */
3929 }
3930 }
3931 }
3933 /*
3934 ** long and ugly. If flush, will not return until all commit
3935 ** blocks and all real buffers in the trans are on disk.
3936 ** If no_async, won't return until all commit blocks are on disk.
3937 **
3938 ** keep reading, there are comments as you go along
3939 **
3940 ** If the journal is aborted, we just clean up. Things like flushing
3941 ** journal lists, etc just won't happen.
3942 */
3946 {
3968 /* protect flush_older_commits from doing mistakes if the
3969 transaction ID counter gets overflowed. */
3982 }
3988 }
3992 }
3994 /* check_journal_end locks the journal, and unlocks if it does not return 1
3995 ** it tells us if we should continue with the journal_end, or just return
3996 */
4002 }
4004 /* check_journal_end might set these, check again */
4007 }
4009 /*
4010 ** j must wait means we have to flush the log blocks, and the real blocks for
4011 ** this transaction
4012 */
4015 }
4016 #ifdef REISERFS_PREALLOCATE
4017 /* quota ops might need to nest, setup the journal_info pointer for them
4018 * and raise the refcount so that it is > 0. */
4022 * the transaction */
4025 #endif
4027 /* setup description block */
4028 d_bh =
4038 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4047 /* init this journal list */
4050 /* we lock the commit before doing anything because
4051 * we want to make sure nobody tries to run flush_commit_list until
4052 * the new transaction is fully setup, and we've already flushed the
4053 * ordered bh list
4054 */
4057 /* save the transaction id in case we need to commit it later */
4070 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4071 ** for each real block, add it to the journal list hash,
4072 ** copy into real block index array in the commit or desc block
4073 */
4081 }
4084 }
4089 }
4091 /* make sure the block we are trying to log is not a block
4092 of journal or reserved area */
4097 "Trying to log block %lu, "
4100 }
4113 }
4115 i--;
4116 }
4117 }
4123 /* special check in case all buffers in the journal were marked for not logging */
4126 /* we're about to dirty all the log blocks, mark the description block
4127 * dirty now too. Don't mark the commit block dirty until all the
4128 * others are on disk
4129 */
4132 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4138 /* copy all the real blocks into log area. dirty log blocks */
4143 tmp_bh =
4147 jindex) %
4157 jindex++;
4161 /* JDirty cleared sometime during transaction. don't log this one */
4163 "BAD, buffer in journal hash, "
4166 }
4173 }
4175 /* we are done with both the c_bh and d_bh, but
4176 ** c_bh must be written after all other commit blocks,
4177 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4178 */
4182 /* now it is safe to insert this transaction on the main list */
4187 /* reset journal values for the next transaction */
4198 /* check for trans_id overflow */
4208 // make sure reiserfs_add_jh sees the new current_jl before we
4209 // write out the tails
4212 /* tail conversion targets have to hit the disk before we end the
4213 * transaction. Otherwise a later transaction might repack the tail
4214 * before this transaction commits, leaving the data block unflushed and
4215 * clean, if we crash before the later transaction commits, the data block
4216 * is lost.
4217 */
4223 }
4227 /* honor the flush wishes from the caller, simple commits can
4228 ** be done outside the journal lock, they are done below
4229 **
4230 ** if we don't flush the commit list right now, we put it into
4231 ** the work queue so the people waiting on the async progress work
4232 ** queue don't wait for this proc to flush journal lists and such.
4233 */
4240 /* if the next transaction has any chance of wrapping, flush
4241 ** transactions that might get overwritten. If any journal lists are very
4242 ** old flush them as well.
4243 */
4244 first_jl:
4255 /* if we don't cross into the next transaction and we don't
4256 * wrap, there is no way we can overlap any later transactions
4257 * break now
4258 */
4260 }
4270 /* we don't overlap anything from out start to the end of the
4271 * log, and our wrapped portion doesn't overlap anything at
4272 * the start of the log. We can break
4273 */
4275 }
4276 }
4277 }
4286 }
4290 /* wake up any body waiting to join. */
4297 }
4298 out:
4302 /* Re-set th->t_super, so we can properly keep track of how many
4303 * persistent transactions there are. We need to do this so if this
4304 * call is part of a failed restart_transaction, we can free it later */
4308 }
4310 /* Send the file system read only and refuse new transactions */
4312 {
4323 #ifdef CONFIG_REISERFS_CHECK
4325 #endif
4326 }