| blob | fd777032c2ba7551dd10fbf9572289b36d09ad27 |
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 /* the number of mounted filesystems. This is used to decide when to
62 ** start and kill the commit workqueue
63 */
69 structs at 4k */
72 /* cnode stat bits. Move these into reiserfs_fs.h */
75 #define BLOCK_FREED_HOLDER 3 /* this block was freed during this transaction, and can't be written */
78 #define BLOCK_DIRTIED 5
80 /* journal list state bits */
81 #define LIST_TOUCHED 1
82 #define LIST_DIRTY 2
85 /* flags for do_journal_end */
107 /* values for join in do_journal_begin_r */
112 };
119 {
123 }
125 /*
126 ** clears BH_Dirty and sticks the buffer on the clean list. Called because I can't allow refile_buffer to
127 ** make schedule happen after I've freed a block. Look at remove_from_transaction and journal_mark_freed for
128 ** more details.
129 */
131 {
135 }
137 }
141 {
148 }
153 }
157 }
160 {
166 repeat:
174 }
179 }
181 }
184 {
193 }
194 }
197 {
208 }
209 }
210 }
213 b_blocknr_t block,
215 {
221 }
224 }
228 {
237 }
238 }
239 }
241 /*
242 ** only call this on FS unmount.
243 */
246 {
255 }
257 }
260 {
272 }
275 }
277 /*
278 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
279 ** jb_array is the array to be filled in.
280 */
284 {
299 }
300 }
304 }
306 }
308 /*
309 ** find an available list bitmap. If you can't find one, flush a commit list
310 ** and try again
311 */
313 struct reiserfs_journal_list
315 {
330 }
333 }
334 }
337 }
340 }
342 /*
343 ** allocates a new chunk of X nodes, and links them all together as a list.
344 ** Uses the cnode->next and cnode->prev pointers
345 ** returns NULL on failure
346 */
348 {
353 }
357 }
363 }
366 }
368 /*
369 ** pulls a cnode off the free list, or returns NULL on failure
370 */
372 {
380 }
386 }
389 }
393 }
395 /*
396 ** returns a cnode to the free list
397 */
400 {
407 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
411 }
412 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
414 }
417 {
420 }
422 /* return a cnode with same dev, block number and size in table, or null if not found */
424 super_block
426 struct
427 reiserfs_journal_cnode
430 {
437 }
439 }
441 /*
442 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
443 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
444 ** 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 in the current transaction. Since deleting
447 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
448 ** sure you never write the block without logging it.
449 **
450 ** next_zero_bit is a suggestion about the next block to try for find_forward.
451 ** when bl is rejected because it is set in a journal list bitmap, we search
452 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
453 ** through next_zero_bit for find_forward to try.
454 **
455 ** Just because we return something in next_zero_bit does not mean we won't
456 ** reject it on the next call to reiserfs_in_journal
457 **
458 */
462 {
472 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
473 ** if we crash before the transaction that freed it commits, this transaction won't
474 ** have committed either, and the block will never be written
475 */
483 data)) {
487 data),
491 }
492 }
493 }
496 /* is it in any old transactions? */
501 }
503 /* is it in the current transaction. This should never happen */
507 }
510 /* safe for reuse */
512 }
514 /* insert cn into table
515 */
518 {
526 }
528 }
530 /* lock the current transaction */
532 {
536 }
538 /* unlock the current transaction */
540 {
542 }
545 {
547 }
551 {
555 }
558 }
560 /*
561 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
562 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
563 ** transaction.
564 */
567 {
572 }
575 }
579 {
588 }
589 }
591 }
593 /*
594 * If page->mapping was null, we failed to truncate this page for
595 * some reason. Most likely because it was truncated after being
596 * logged via data=journal.
597 *
598 * This does a check to see if the buffer belongs to one of these
599 * lost pages before doing the final put_bh. If page->mapping was
600 * null, it tries to free buffers on the page, which should make the
601 * final page_cache_release drop the page from the lru.
602 */
604 {
615 }
616 }
619 {
626 }
629 else
634 }
637 {
640 else
644 }
647 {
657 }
660 {
667 }
669 #define CHUNK_SIZE 32
673 };
676 {
680 }
682 }
685 {
689 }
691 }
695 {
706 }
708 }
712 {
719 }
721 }
722 }
724 /*
725 * we want to free the jh when the buffer has been written
726 * and waited on
727 */
729 {
742 }
743 }
747 {
755 }
759 no_jh:
763 /* buffer must be locked for __add_jh, should be able to have
764 * two adds at the same time
765 */
769 }
775 }
778 }
781 {
783 }
785 {
787 }
789 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
794 {
812 }
820 }
821 /* in theory, dirty non-uptodate buffers should never get here,
822 * but the upper layer io error paths still have a few quirks.
823 * Handle them here as gracefully as we can
824 */
828 }
835 }
836 loop_next:
839 }
844 }
855 }
858 }
859 /* ugly interaction with invalidatepage here.
860 * reiserfs_invalidate_page will pin any buffer that has a valid
861 * journal head from an older transaction. If someone else sets
862 * our buffer dirty after we write it in the first loop, and
863 * then someone truncates the page away, nobody will ever write
864 * the buffer. We're safe if we write the page one last time
865 * after freeing the journal header.
866 */
871 }
874 }
877 }
881 {
890 find_first:
891 /*
892 * first we walk backwards to find the oldest uncommitted transation
893 */
904 }
906 /* if we didn't find any older uncommitted transactions, return now */
909 }
922 /* list we were called with is gone, return */
926 /* the one we just flushed is gone, this means all
927 * older lists are also gone, so first_jl is no longer
928 * valid either. Go back to the beginning.
929 */
933 }
934 }
940 }
941 }
943 }
946 {
955 }
958 }
960 /*
961 ** if this journal list still has commit blocks unflushed, send them to disk.
962 **
963 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
964 ** Before the commit block can by written, every other log block must be safely on disk
965 **
966 */
969 {
971 b_blocknr_t bn;
983 }
985 /* before we can put our commit blocks on disk, we have to make sure everyone older than
986 ** us is on disk too
987 */
994 /* list disappeared during flush_older_commits. return */
996 }
997 }
999 /* make sure nobody is trying to flush this one at the same time */
1005 }
1008 /* this commit is done, exit */
1012 }
1015 }
1020 /*
1021 * We might sleep in numerous places inside
1022 * write_ordered_buffers. Relax the write lock.
1023 */
1030 }
1032 /*
1033 * for the description block and all the log blocks, submit any buffers
1034 * that haven't already reached the disk. Try to write at least 256
1035 * log blocks. later on, we will only wait on blocks that correspond
1036 * to this transaction, but while we're unplugging we might as well
1037 * get a chunk of data on there.
1038 */
1052 }
1054 }
1055 }
1066 // since we're using ll_rw_blk above, it might have skipped over
1067 // a locked buffer. Double check here
1068 //
1069 /* redundant, sync_dirty_buffer() checks */
1074 }
1076 #ifdef CONFIG_REISERFS_CHECK
1079 #endif
1081 }
1085 }
1089 /* If there was a write error in the journal - we can't commit
1090 * this transaction - it will be invalid and, if successful,
1091 * will just end up propagating the write error out to
1092 * the file system. */
1100 else
1103 }
1105 /* If there was a write error in the journal - we can't commit this
1106 * transaction - it will be invalid and, if successful, will just end
1107 * up propagating the write error out to the filesystem. */
1109 #ifdef CONFIG_REISERFS_CHECK
1111 #endif
1113 }
1119 }
1122 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1127 /* mark the metadata dirty */
1134 }
1136 put_jl:
1141 __func__);
1143 }
1145 /*
1146 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1147 ** returns NULL if it can't find anything
1148 */
1150 reiserfs_journal_cnode
1152 {
1160 }
1162 }
1164 }
1171 /*
1172 ** once all the real blocks have been flushed, it is safe to remove them from the
1173 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1174 ** block to be reallocated for data blocks if it had been deleted.
1175 */
1179 {
1184 /* which is better, to lock once around the whole loop, or
1185 ** to lock for each call to remove_journal_hash?
1186 */
1194 }
1198 }
1202 }
1204 }
1206 /*
1207 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1208 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1209 ** releasing blocks in this transaction for reuse as data blocks.
1210 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1211 **
1212 */
1216 {
1230 #ifdef CONFIG_REISERFS_CHECK
1233 #endif
1235 }
1236 }
1240 b_data);
1250 else
1258 }
1259 }
1261 }
1266 {
1268 }
1270 /*
1271 ** flush any and all journal lists older than you are
1272 ** can only be called from flush_journal_list
1273 */
1276 {
1282 /* we know we are the only ones flushing things, no extra race
1283 * protection is required.
1284 */
1285 restart:
1287 /* Did we wrap? */
1293 /* do not flush all */
1296 /* other_jl is now deleted from the list */
1298 }
1300 }
1304 {
1309 }
1310 }
1312 /* flush a journal list, both commit and real blocks
1313 **
1314 ** always set flushall to 1, unless you are calling from inside
1315 ** flush_journal_list
1316 **
1317 ** IMPORTANT. This can only be called while there are no journal writers,
1318 ** and the journal is locked. That means it can only be called from
1319 ** do_journal_end, or by journal_release
1320 */
1323 {
1340 }
1342 /* if flushall == 0, the lock is already held */
1347 }
1354 }
1356 /* if all the work is already done, get out of here */
1360 }
1362 /* start by putting the commit list on disk. This will also flush
1363 ** the commit lists of any olders transactions
1364 */
1371 /* are we done now? */
1375 }
1377 /* loop through each cnode, see if we need to write it,
1378 ** or wait on a more recent transaction, or just ignore it
1379 */
1383 }
1389 /* blocknr of 0 is no longer in the hash, ignore it */
1392 }
1394 /* This transaction failed commit. Don't write out to the disk */
1399 /* the order is important here. We check pjl to make sure we
1400 ** don't clear BH_JDirty_wait if we aren't the one writing this
1401 ** block to disk
1402 */
1406 /* we do this to make sure nobody releases the buffer while
1407 ** we are working with it
1408 */
1416 /* everything with !pjl && jwait should be writable */
1418 }
1419 }
1421 /* if someone has this block in a newer transaction, just make
1422 ** sure they are committed, and don't try writing it to disk
1423 */
1428 }
1430 /* bh == NULL when the block got to disk on its own, OR,
1431 ** the block got freed in a future transaction
1432 */
1435 }
1437 /* this should never happen. kupdate_one_transaction has this list
1438 ** locked while it works, so we should never see a buffer here that
1439 ** is not marked JDirty_wait
1440 */
1443 "BAD! buffer %llu %cdirty %cjwait, "
1448 }
1450 /* we inc again because saved_bh gets decremented at free_cnode */
1457 else
1459 count++;
1465 }
1466 free_cnode:
1470 /* we incremented this to keep others from taking the buffer head away */
1475 }
1476 }
1477 }
1485 }
1494 }
1496 #ifdef CONFIG_REISERFS_CHECK
1499 #endif
1501 }
1502 /* note, we must clear the JDirty_wait bit after the up to date
1503 ** check, otherwise we race against our flushpage routine
1504 */
1508 /* drop one ref for us */
1510 /* drop one ref for journal_mark_dirty */
1512 }
1514 }
1515 }
1520 __func__);
1521 flush_older_and_return:
1523 /* before we can update the journal header block, we _must_ flush all
1524 ** real blocks from all older transactions to disk. This is because
1525 ** once the header block is updated, this transaction will not be
1526 ** replayed after a crash
1527 */
1530 }
1533 /* before we can remove everything from the hash tables for this
1534 ** transaction, we must make sure it can never be replayed
1535 **
1536 ** since we are only called from do_journal_end, we know for sure there
1537 ** are no allocations going on while we are flushing journal lists. So,
1538 ** we only need to update the journal header block for the last list
1539 ** being flushed
1540 */
1542 err =
1550 __func__);
1551 }
1561 }
1564 /* not strictly required since we are freeing the list, but it should
1565 * help find code using dead lists later on
1566 */
1578 }
1583 {
1591 }
1595 /* if the blocknr == 0, this has been cleared from the hash,
1596 ** skip it
1597 */
1600 }
1603 /* we can race against journal_mark_freed when we try
1604 * to lock_buffer(cn->bh), so we have to inc the buffer
1605 * count, and recheck things after locking
1606 */
1615 ret++;
1617 /* note, cn->bh might be null now */
1619 }
1621 }
1622 next:
1625 }
1627 }
1629 /* used by flush_commit_list */
1632 {
1640 /* look for a more recent transaction that logged this
1641 ** buffer. Only the most recent transaction with a buffer in
1642 ** it is allowed to send that buffer to disk
1643 */
1648 /* if the buffer is prepared, it will either be logged
1649 * or restored. If restored, we need to make sure
1650 * it actually gets marked dirty
1651 */
1658 }
1659 }
1661 }
1663 }
1670 {
1683 }
1685 /* we've got j_flush_mutex held, nobody is going to delete any
1686 * of these lists out from underneath us
1687 */
1696 }
1701 transactions_flushed++;
1705 /* did we wrap? */
1708 }
1711 /* don't bother with older transactions */
1714 }
1717 }
1719 done:
1722 }
1724 /* for o_sync and fsync heavy applications, they tend to use
1725 ** all the journa list slots with tiny transactions. These
1726 ** trigger lots and lots of calls to update the header block, which
1727 ** adds seeks and slows things down.
1728 **
1729 ** This function tries to clear out a large chunk of the journal lists
1730 ** at once, which makes everything faster since only the newest journal
1731 ** list updates the header block
1732 */
1735 {
1748 /* in data logging mode, try harder to flush a lot of blocks */
1751 /* flush for 256 transactions or limit blocks, whichever comes first */
1756 }
1760 }
1766 }
1769 /* try to find a group of blocks we can flush across all the
1770 ** transactions, but only bother if we've actually spanned
1771 ** across multiple lists
1772 */
1775 }
1780 }
1782 /*
1783 ** removes any nodes in table with name block and dev as bh.
1784 ** only touchs the hnext and hprev pointers.
1785 */
1790 {
1797 }
1805 }
1810 }
1818 }
1820 }
1821 }
1824 {
1834 }
1835 /* j_header_bh is on the journal dev, make sure not to release the journal
1836 * dev until we brelse j_header_bh
1837 */
1840 }
1842 /*
1843 ** call on unmount. Only set error to 1 if you haven't made your way out
1844 ** of read_super() yet. Any other caller must keep error at 0.
1845 */
1848 {
1853 /* we only want to flush out transactions if we were called with error == 0
1854 */
1856 /* end the current trans */
1860 /* make sure something gets logged to force our way into the flush code */
1869 }
1870 }
1872 /* this also catches errors during the do_journal_end above */
1882 }
1883 }
1885 reiserfs_mounted_fs_count--;
1886 /* wait for all commits to finish */
1889 /*
1890 * We must release the write lock here because
1891 * the workqueue job (flush_async_commit) needs this lock
1892 */
1901 }
1908 }
1910 /*
1911 ** call on unmount. flush all journal trans, release all alloc'd ram
1912 */
1915 {
1917 }
1919 /*
1920 ** only call from an error condition inside reiserfs_read_super!
1921 */
1924 {
1926 }
1928 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1932 {
1938 }
1940 }
1942 /* returns 0 if it did not find a description block
1943 ** returns -1 if it found a corrupt commit block
1944 ** returns 1 if both desc and commit were valid
1945 ** NOTE: only called during fs mount
1946 */
1951 {
1966 "journal-986: transaction "
1967 "is valid returning because trans_id %d is greater than "
1972 }
1976 "journal-1087: transaction "
1977 "is valid returning because mount_id %d is less than "
1982 }
1985 "Bad transaction length %d "
1989 }
1992 /* ok, we have a journal description block, lets see if the transaction was valid */
1993 c_bh =
2003 "journal_transaction_is_valid, commit offset %ld had bad "
2014 "journal-1004: "
2015 "transaction_is_valid setting oldest invalid trans_id "
2018 }
2020 }
2023 "journal-1006: found valid "
2032 }
2033 }
2036 {
2040 }
2041 }
2043 /*
2044 ** given the start, and values for the oldest acceptable transactions,
2045 ** this either reads in a replays a transaction, or returns because the
2046 ** transaction is invalid, or too old.
2047 ** NOTE: only called during fs mount
2048 */
2054 {
2079 cur_dblock -
2083 }
2086 "journal_read_trans skipping because %d is != "
2088 newest_mount_id);
2091 }
2098 }
2102 "journal_read_transaction, "
2111 }
2119 }
2122 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2135 }
2136 /* get all the buffer heads */
2153 }
2156 "REPLAY FAILURE fsck required! "
2157 "Block to replay is outside of "
2160 }
2161 /* make sure we don't try to replay onto log or reserved area */
2165 "REPLAY FAILURE fsck required! "
2167 abort_replay:
2175 }
2176 }
2177 /* read in the log blocks, memcpy to the corresponding real block */
2184 "REPLAY FAILURE fsck required! "
2194 }
2199 }
2200 /* flush out the real blocks */
2204 }
2209 "REPLAY FAILURE, fsck required! "
2218 }
2220 }
2221 cur_dblock =
2229 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2233 /* check for trans_id overflow */
2241 }
2243 /* This function reads blocks starting from block and to max_block of bufsize
2244 size (but no more than BUFNR blocks at a time). This proved to improve
2245 mounting speed on self-rebuilding raid5 arrays at least.
2246 Right now it is only used from journal code. But later we might use it
2247 from other places.
2248 Note: Do not use journal_getblk/sb_getblk functions here! */
2251 b_blocknr_t max_block)
2252 {
2264 }
2274 }
2284 }
2286 /*
2287 ** read and replay the log
2288 ** on a clean unmount, the journal header's next unflushed pointer will
2289 ** be to an invalid transaction. This tests that before finding all the
2290 ** transactions in the log, which makes normal mount times fast.
2291 ** After a crash, this starts with the next unflushed transaction, and
2292 ** replays until it finds one too old, or invalid.
2293 ** On exit, it sets things up so the first transaction will work correctly.
2294 ** NOTE: only called during fs mount
2295 */
2297 {
2319 /* step 1, read in the journal header block. Check the transaction it says
2320 ** is the first unflushed, and if that transaction is not valid,
2321 ** replay is done
2322 */
2328 }
2333 oldest_start =
2339 "journal-1153: found in "
2340 "header: first_unflushed_offset %d, last_flushed_trans_id "
2345 /* now, we try to read the first unflushed offset. If it is not valid,
2346 ** there is nothing more we can do, and it makes no sense to read
2347 ** through the whole log.
2348 */
2349 d_bh =
2356 }
2359 }
2361 /* ok, there are transactions that need to be replayed. start with the first log block, find
2362 ** all the valid transactions, and pick out the oldest.
2363 */
2368 /* Note that it is required for blocksize of primary fs device and journal
2369 device to be the same */
2370 d_bh =
2375 ret =
2386 "journal-1179: Setting "
2388 oldest_start -
2389 SB_ONDISK_JOURNAL_1st_BLOCK
2392 /* one we just read was older */
2396 "journal-1180: Resetting "
2398 oldest_start -
2399 SB_ONDISK_JOURNAL_1st_BLOCK
2401 }
2405 "journal-1299: Setting "
2408 }
2411 cur_dblock++;
2412 }
2414 }
2416 start_log_replay:
2420 "journal-1206: Starting replay "
2423 oldest_trans_id);
2425 }
2428 ret =
2435 }
2436 cur_dblock =
2438 replay_count++;
2441 }
2446 }
2447 /* j_start does not get set correctly if we don't replay any transactions.
2448 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2449 ** copy the trans_id from the header
2450 */
2455 /* check for trans_id overflow */
2463 }
2471 }
2472 /* needed to satisfy the locking in _update_journal_header_block */
2478 /* replay failed, caller must call free_journal_ram and abort
2479 ** the mount
2480 */
2482 }
2485 }
2488 {
2500 }
2503 {
2505 }
2509 {
2513 }
2514 }
2519 {
2521 dev_t jdev;
2534 /* there is no "jdev" option and journal is on separate device */
2539 journal);
2552 }
2563 }
2570 }
2572 /**
2573 * When creating/tuning a file system user can assign some
2574 * journal params within boundaries which depend on the ratio
2575 * blocksize/standard_blocksize.
2576 *
2577 * For blocks >= standard_blocksize transaction size should
2578 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2579 * then JOURNAL_TRANS_MAX_DEFAULT.
2580 *
2581 * For blocks < standard_blocksize these boundaries should be
2582 * decreased proportionally.
2583 */
2584 #define REISERFS_STANDARD_BLKSIZE (4096)
2588 {
2590 /* Non-default journal params.
2591 Do sanity check for them. */
2599 JOURNAL_MIN_RATIO) {
2601 "bad transaction max size (%u). "
2604 }
2608 "bad transaction max batch (%u). "
2611 }
2613 /* Default journal params.
2614 The file system was created by old version
2615 of mkreiserfs, so some fields contain zeros,
2616 and we need to advise proper values for them */
2621 }
2625 }
2627 }
2629 /*
2630 ** must be called once on fs mount. calls journal_read for you
2631 */
2634 {
2649 }
2661 /* reserved for journal area support */
2663 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2667 REISERFS_DISK_OFFSET_IN_BYTES /
2670 /* Sanity check to see is the standard journal fitting within first bitmap
2671 (actual for small blocksizes) */
2676 "journal does not fit for area addressed "
2677 "by first of bitmap blocks. It starts at "
2683 }
2689 }
2693 /* read journal header */
2701 }
2704 /* make sure that journal matches to the super block */
2709 "journal header magic %x (device %s) does "
2716 }
2731 }
2734 "journal first block %u, max trans len %u, max batch %u, "
2779 "allocation failed (%ld bytes). Journal is "
2780 "too large for available memory. Usually "
2784 }
2789 /*
2790 * get_list_bitmap() may call flush_commit_list() which
2791 * requires the lock. Calling flush_commit_list() shouldn't happen
2792 * this early but I like to be paranoid.
2793 */
2801 }
2808 }
2810 reiserfs_mounted_fs_count++;
2817 free_and_return:
2820 }
2822 /*
2823 ** test for a polite end of the current transaction. Used by file_write, and should
2824 ** be used by delete to make sure they don't write more than can fit inside a single
2825 ** transaction
2826 */
2829 {
2832 /* cannot restart while nested */
2842 }
2847 }
2849 /* this must be called inside a transaction
2850 */
2852 {
2858 }
2860 /* this must be called without a transaction started
2861 */
2863 {
2867 }
2869 /* this must be called without a transaction started
2870 */
2872 {
2876 }
2879 {
2880 wait_queue_t wait;
2884 /*
2885 * we don't want to use wait_event here because
2886 * we only want to wait once.
2887 */
2895 }
2898 }
2901 {
2905 }
2908 {
2923 }
2929 }
2930 }
2932 /* join == true if you must join an existing transaction.
2933 ** join == false if you can deal with waiting for others to finish
2934 **
2935 ** this will block until the transaction is joinable. send the number of blocks you
2936 ** expect to use in nblocks.
2937 */
2941 {
2954 /* set here for journal_join */
2958 relock:
2964 }
2974 }
2977 /* if there is no room in the journal OR
2978 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
2979 ** we don't sleep if there aren't other writers
2980 */
2983 (!join
3000 sched_count++;
3003 }
3004 }
3005 /* don't mess with joining the transaction if all we have to do is
3006 * wait for someone else to do a commit
3007 */
3012 }
3014 }
3019 /* someone might have ended the transaction while we joined */
3024 }
3031 }
3032 /* we are the first writer, set trans_id */
3035 }
3045 out_fail:
3047 /* Re-set th->t_super, so we can properly keep track of how many
3048 * persistent transactions there are. We need to do this so if this
3049 * call is part of a failed restart_transaction, we can free it later */
3052 }
3055 super_block
3058 {
3062 /* if we're nesting into an existing transaction. It will be
3063 ** persistent on its own
3064 */
3071 }
3079 }
3083 }
3086 {
3091 else
3096 }
3098 }
3102 {
3105 /* this keeps do_journal_end from NULLing out the current->journal_info
3106 ** pointer
3107 */
3111 }
3115 {
3118 /* this keeps do_journal_end from NULLing out the current->journal_info
3119 ** pointer
3120 */
3124 }
3128 {
3134 /* we are nesting into the current transaction */
3141 "BAD: refcount <= 1, but "
3145 /* we've ended up with a handle from a different filesystem.
3146 ** save it and restore on journal_end. This should never
3147 ** really happen...
3148 */
3153 }
3156 }
3160 /* I guess this boils down to being the reciprocal of clm-2100 above.
3161 * If do_journal_begin_r fails, we need to put it back, since journal_end
3162 * won't be called to do it. */
3165 else
3169 }
3171 /*
3172 ** puts bh into the current transaction. If it was already there, reorders removes the
3173 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3174 **
3175 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3176 ** transaction is committed.
3177 **
3178 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3179 */
3182 {
3194 }
3198 /* already in this transaction, we are done */
3202 }
3204 /* this must be turned into a panic instead of a warning. We can't allow
3205 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3206 ** could get to disk too early. NOT GOOD.
3207 */
3210 "buffer %llu bad state "
3217 }
3224 }
3225 /* this error means I've screwed up, and we've overflowed the transaction.
3226 ** Nothing can be done here, except make the FS readonly or panic.
3227 */
3232 }
3238 }
3242 }
3246 /* now put this guy on the end */
3251 }
3256 }
3267 }
3268 }
3278 }
3281 }
3285 {
3293 }
3300 /* we aren't allowed to close a nested transaction on a different
3301 ** filesystem from the one in the task struct
3302 */
3308 }
3312 }
3313 }
3315 /* removes from the current transaction, relsing and descrementing any counters.
3316 ** also files the removed buffer directly onto the clean list
3317 **
3318 ** called by journal_mark_freed when a block has been deleted
3319 **
3320 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3321 */
3324 {
3333 }
3337 }
3340 }
3343 }
3346 }
3360 }
3362 }
3367 }
3369 /*
3370 ** for any cnode in a journal list, it can only be dirtied of all the
3371 ** transactions that include it are committed to disk.
3372 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3373 ** and 0 if you aren't
3374 **
3375 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3376 ** blocks for a given transaction on disk
3377 **
3378 */
3380 {
3386 /* first test hprev. These are all newer than cn, so any node here
3387 ** with the same block number and dev means this node can't be sent
3388 ** to disk right now.
3389 */
3394 }
3396 }
3397 /* then test hnext. These are all older than cn. As long as they
3398 ** are committed to the log, it is safe to write cn to disk
3399 */
3406 }
3408 }
3410 }
3412 /* syncs the commit blocks, but does not force the real buffers to disk
3413 ** will wait until the current transaction is done/committed before returning
3414 */
3417 {
3421 /* you can sync while nested, very, very bad */
3427 }
3429 }
3431 /*
3432 ** writeback the pending async commits to disk
3433 */
3435 {
3444 /* last entry is the youngest, commit it and you get everything */
3448 }
3450 }
3452 /*
3453 ** flushes any old transactions to disk
3454 ** ends the current transaction if it is too old
3455 */
3457 {
3463 /* safety check so we don't flush while we are replaying the log during
3464 * mount
3465 */
3469 /* check the current transaction. If there are no writers, and it is
3470 * too old, finish it, and force the commit blocks to disk
3471 */
3483 /* we're only being called from kreiserfsd, it makes no sense to do
3484 ** an async commit so that kreiserfsd can do it later
3485 */
3487 }
3488 }
3489 }
3491 /*
3492 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3493 **
3494 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3495 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3496 ** flushes the commit list and returns 0.
3497 **
3498 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3499 **
3500 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3501 */
3505 {
3520 }
3523 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3525 }
3527 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3528 ** will be dealt with by next transaction that actually writes something, but should be taken
3529 ** care of in this trans
3530 */
3533 /* if wcount > 0, and we are called to with flush or commit_now,
3534 ** we wait on j_join_wait. We will wake up when the last writer has
3535 ** finished the transaction, and started it on its way to the disk.
3536 ** Then, we flush the commit or journal list, and just return 0
3537 ** because the rest of journal end was already done for this transaction.
3538 */
3550 }
3553 /* sleep while the current transaction is still j_jlocked */
3562 }
3564 }
3565 }
3572 }
3574 }
3577 }
3579 /* deal with old transactions where we are the last writers */
3584 }
3585 /* don't batch when someone is waiting on j_join_wait */
3586 /* don't batch when syncing the commit or flushing the whole trans */
3594 }
3600 }
3602 }
3604 /*
3605 ** Does all the work that makes deleting blocks safe.
3606 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3607 **
3608 ** otherwise:
3609 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3610 ** before this transaction has finished.
3611 **
3612 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3613 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3614 ** the block can't be reallocated yet.
3615 **
3616 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3617 */
3620 {
3632 }
3633 /* if it is journal new, we just remove it from this transaction */
3640 /* set the bit for this block in the journal bitmap for this transaction */
3645 }
3648 /* Note, the entire while loop is not allowed to schedule. */
3653 }
3656 /* find all older transactions with this block, make sure they don't try to write it out */
3658 blocknr);
3664 /* remove_from_transaction will brelse the buffer if it was
3665 ** in the current trans
3666 */
3668 bh);
3671 bh);
3679 }
3680 }
3684 j_nonzerolen));
3685 }
3687 }
3688 }
3690 }
3691 }
3696 }
3699 {
3703 }
3705 /*
3706 * returns -1 on error, 0 if no commits/barriers were done and 1
3707 * if a transaction was actually committed and the barrier was done
3708 */
3711 {
3717 /* is it from the current transaction, or from an unknown transaction? */
3720 /* try to let other writers come in and grow this transaction */
3724 }
3730 /* someone might have ended this transaction while we joined */
3737 }
3744 /* this gets tricky, we have to make sure the journal list in
3745 * the inode still exists. We know the list is still around
3746 * if we've got a larger transaction id than the oldest list
3747 */
3748 flush_commit_only:
3750 /*
3751 * we only set ret to 1 when we know for sure
3752 * the barrier hasn't been started yet on the commit
3753 * block.
3754 */
3760 }
3761 }
3762 /* otherwise the list is gone, and long since committed */
3764 }
3767 {
3771 /* for the whole inode, assume unset id means it was
3772 * changed in the current transaction. More conservative
3773 */
3777 /* jl will be updated in __commit_trans_jl */
3778 }
3781 }
3785 {
3790 }
3801 }
3803 }
3805 }
3808 /*
3809 ** before we can change a metadata block, we have to make sure it won't
3810 ** be written to disk while we are altering it. So, we must:
3811 ** clean it
3812 ** wait on it.
3813 **
3814 */
3817 {
3824 }
3829 }
3832 }
3834 /*
3835 ** long and ugly. If flush, will not return until all commit
3836 ** blocks and all real buffers in the trans are on disk.
3837 ** If no_async, won't return until all commit blocks are on disk.
3838 **
3839 ** keep reading, there are comments as you go along
3840 **
3841 ** If the journal is aborted, we just clean up. Things like flushing
3842 ** journal lists, etc just won't happen.
3843 */
3847 {
3870 /* protect flush_older_commits from doing mistakes if the
3871 transaction ID counter gets overflowed. */
3883 }
3889 }
3893 }
3895 /* check_journal_end locks the journal, and unlocks if it does not return 1
3896 ** it tells us if we should continue with the journal_end, or just return
3897 */
3903 }
3905 /* check_journal_end might set these, check again */
3908 }
3910 /*
3911 ** j must wait means we have to flush the log blocks, and the real blocks for
3912 ** this transaction
3913 */
3916 }
3917 #ifdef REISERFS_PREALLOCATE
3918 /* quota ops might need to nest, setup the journal_info pointer for them
3919 * and raise the refcount so that it is > 0. */
3923 * the transaction */
3926 #endif
3928 /* setup description block */
3929 d_bh =
3939 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
3948 /* init this journal list */
3951 /* we lock the commit before doing anything because
3952 * we want to make sure nobody tries to run flush_commit_list until
3953 * the new transaction is fully setup, and we've already flushed the
3954 * ordered bh list
3955 */
3958 /* save the transaction id in case we need to commit it later */
3971 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
3972 ** for each real block, add it to the journal list hash,
3973 ** copy into real block index array in the commit or desc block
3974 */
3982 }
3985 }
3990 }
3992 /* make sure the block we are trying to log is not a block
3993 of journal or reserved area */
3998 "Trying to log block %lu, "
4001 }
4014 }
4016 i--;
4017 }
4018 }
4024 /* special check in case all buffers in the journal were marked for not logging */
4027 /* we're about to dirty all the log blocks, mark the description block
4028 * dirty now too. Don't mark the commit block dirty until all the
4029 * others are on disk
4030 */
4033 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4039 /* copy all the real blocks into log area. dirty log blocks */
4044 tmp_bh =
4048 jindex) %
4058 jindex++;
4062 /* JDirty cleared sometime during transaction. don't log this one */
4064 "BAD, buffer in journal hash, "
4067 }
4072 }
4074 /* we are done with both the c_bh and d_bh, but
4075 ** c_bh must be written after all other commit blocks,
4076 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4077 */
4081 /* now it is safe to insert this transaction on the main list */
4086 /* reset journal values for the next transaction */
4097 /* check for trans_id overflow */
4107 // make sure reiserfs_add_jh sees the new current_jl before we
4108 // write out the tails
4111 /* tail conversion targets have to hit the disk before we end the
4112 * transaction. Otherwise a later transaction might repack the tail
4113 * before this transaction commits, leaving the data block unflushed and
4114 * clean, if we crash before the later transaction commits, the data block
4115 * is lost.
4116 */
4122 }
4126 /* honor the flush wishes from the caller, simple commits can
4127 ** be done outside the journal lock, they are done below
4128 **
4129 ** if we don't flush the commit list right now, we put it into
4130 ** the work queue so the people waiting on the async progress work
4131 ** queue don't wait for this proc to flush journal lists and such.
4132 */
4139 /* if the next transaction has any chance of wrapping, flush
4140 ** transactions that might get overwritten. If any journal lists are very
4141 ** old flush them as well.
4142 */
4143 first_jl:
4154 /* if we don't cross into the next transaction and we don't
4155 * wrap, there is no way we can overlap any later transactions
4156 * break now
4157 */
4159 }
4169 /* we don't overlap anything from out start to the end of the
4170 * log, and our wrapped portion doesn't overlap anything at
4171 * the start of the log. We can break
4172 */
4174 }
4175 }
4176 }
4184 }
4188 /* wake up any body waiting to join. */
4195 }
4196 out:
4200 /* Re-set th->t_super, so we can properly keep track of how many
4201 * persistent transactions there are. We need to do this so if this
4202 * call is part of a failed restart_transaction, we can free it later */
4206 }
4208 /* Send the file system read only and refuse new transactions */
4210 {
4221 #ifdef CONFIG_REISERFS_CHECK
4223 #endif
4224 }