| blob | eb711060a6f2b771d4b4e68613ea4890273633dd |
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>
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/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>
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 }
142 {
149 }
154 }
158 }
161 {
167 repeat:
175 }
180 }
182 }
185 {
194 }
195 }
198 {
209 }
210 }
211 }
214 b_blocknr_t block,
216 {
222 }
225 }
229 {
238 }
239 }
240 }
242 /*
243 ** only call this on FS unmount.
244 */
247 {
256 }
258 }
261 {
273 }
276 }
278 /*
279 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
280 ** jb_array is the array to be filled in.
281 */
285 {
300 }
301 }
305 }
307 }
309 /*
310 ** find an available list bitmap. If you can't find one, flush a commit list
311 ** and try again
312 */
314 struct reiserfs_journal_list
316 {
331 }
334 }
335 }
338 }
341 }
343 /*
344 ** allocates a new chunk of X nodes, and links them all together as a list.
345 ** Uses the cnode->next and cnode->prev pointers
346 ** returns NULL on failure
347 */
349 {
354 }
358 }
364 }
367 }
369 /*
370 ** pulls a cnode off the free list, or returns NULL on failure
371 */
373 {
381 }
387 }
390 }
394 }
396 /*
397 ** returns a cnode to the free list
398 */
401 {
408 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
412 }
413 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
415 }
418 {
421 }
423 /* return a cnode with same dev, block number and size in table, or null if not found */
425 super_block
427 struct
428 reiserfs_journal_cnode
431 {
438 }
440 }
442 /*
443 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
444 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
445 ** being overwritten by a replay after crashing.
446 **
447 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
448 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
449 ** sure you never write the block without logging it.
450 **
451 ** next_zero_bit is a suggestion about the next block to try for find_forward.
452 ** when bl is rejected because it is set in a journal list bitmap, we search
453 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
454 ** through next_zero_bit for find_forward to try.
455 **
456 ** Just because we return something in next_zero_bit does not mean we won't
457 ** reject it on the next call to reiserfs_in_journal
458 **
459 */
463 {
473 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
474 ** if we crash before the transaction that freed it commits, this transaction won't
475 ** have committed either, and the block will never be written
476 */
484 data)) {
488 data),
492 }
493 }
494 }
497 /* is it in any old transactions? */
502 }
504 /* is it in the current transaction. This should never happen */
508 }
511 /* safe for reuse */
513 }
515 /* insert cn into table
516 */
519 {
527 }
529 }
531 /* lock the current transaction */
533 {
537 }
539 /* unlock the current transaction */
541 {
543 }
546 {
548 }
552 {
556 }
559 }
561 /*
562 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
563 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
564 ** transaction.
565 */
568 {
573 }
576 }
580 {
589 }
590 }
592 }
594 /*
595 * If page->mapping was null, we failed to truncate this page for
596 * some reason. Most likely because it was truncated after being
597 * logged via data=journal.
598 *
599 * This does a check to see if the buffer belongs to one of these
600 * lost pages before doing the final put_bh. If page->mapping was
601 * null, it tries to free buffers on the page, which should make the
602 * final page_cache_release drop the page from the lru.
603 */
605 {
616 }
617 }
620 {
627 }
630 else
635 }
638 {
641 else
645 }
648 {
658 }
661 {
668 }
670 #define CHUNK_SIZE 32
674 };
677 {
681 }
683 }
686 {
690 }
692 }
696 {
707 }
709 }
713 {
720 }
722 }
723 }
725 /*
726 * we want to free the jh when the buffer has been written
727 * and waited on
728 */
730 {
743 }
744 }
748 {
756 }
760 no_jh:
764 /* buffer must be locked for __add_jh, should be able to have
765 * two adds at the same time
766 */
770 }
776 }
779 }
782 {
784 }
786 {
788 }
790 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
795 {
813 }
821 }
822 /* in theory, dirty non-uptodate buffers should never get here,
823 * but the upper layer io error paths still have a few quirks.
824 * Handle them here as gracefully as we can
825 */
829 }
836 }
837 loop_next:
840 }
845 }
856 }
859 }
860 /* ugly interaction with invalidatepage here.
861 * reiserfs_invalidate_page will pin any buffer that has a valid
862 * journal head from an older transaction. If someone else sets
863 * our buffer dirty after we write it in the first loop, and
864 * then someone truncates the page away, nobody will ever write
865 * the buffer. We're safe if we write the page one last time
866 * after freeing the journal header.
867 */
872 }
875 }
878 }
882 {
891 find_first:
892 /*
893 * first we walk backwards to find the oldest uncommitted transation
894 */
905 }
907 /* if we didn't find any older uncommitted transactions, return now */
910 }
923 /* list we were called with is gone, return */
927 /* the one we just flushed is gone, this means all
928 * older lists are also gone, so first_jl is no longer
929 * valid either. Go back to the beginning.
930 */
934 }
935 }
941 }
942 }
944 }
947 {
954 }
957 }
959 /*
960 ** if this journal list still has commit blocks unflushed, send them to disk.
961 **
962 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
963 ** Before the commit block can by written, every other log block must be safely on disk
964 **
965 */
968 {
970 b_blocknr_t bn;
981 }
983 /* before we can put our commit blocks on disk, we have to make sure everyone older than
984 ** us is on disk too
985 */
992 /* list disappeared during flush_older_commits. return */
994 }
995 }
997 /* make sure nobody is trying to flush this one at the same time */
1003 }
1006 /* this commit is done, exit */
1010 }
1013 }
1018 /*
1019 * We might sleep in numerous places inside
1020 * write_ordered_buffers. Relax the write lock.
1021 */
1028 }
1030 /*
1031 * for the description block and all the log blocks, submit any buffers
1032 * that haven't already reached the disk. Try to write at least 256
1033 * log blocks. later on, we will only wait on blocks that correspond
1034 * to this transaction, but while we're unplugging we might as well
1035 * get a chunk of data on there.
1036 */
1050 }
1052 }
1053 }
1064 // since we're using ll_rw_blk above, it might have skipped over
1065 // a locked buffer. Double check here
1066 //
1067 /* redundant, sync_dirty_buffer() checks */
1072 }
1074 #ifdef CONFIG_REISERFS_CHECK
1077 #endif
1079 }
1083 }
1087 /* If there was a write error in the journal - we can't commit
1088 * this transaction - it will be invalid and, if successful,
1089 * will just end up propagating the write error out to
1090 * the file system. */
1098 else
1101 }
1103 /* If there was a write error in the journal - we can't commit this
1104 * transaction - it will be invalid and, if successful, will just end
1105 * up propagating the write error out to the filesystem. */
1107 #ifdef CONFIG_REISERFS_CHECK
1109 #endif
1111 }
1117 }
1120 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1125 /* mark the metadata dirty */
1132 }
1134 put_jl:
1139 __func__);
1141 }
1143 /*
1144 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1145 ** returns NULL if it can't find anything
1146 */
1148 reiserfs_journal_cnode
1150 {
1158 }
1160 }
1162 }
1165 {
1175 }
1177 }
1184 /*
1185 ** once all the real blocks have been flushed, it is safe to remove them from the
1186 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1187 ** block to be reallocated for data blocks if it had been deleted.
1188 */
1192 {
1197 /* which is better, to lock once around the whole loop, or
1198 ** to lock for each call to remove_journal_hash?
1199 */
1207 }
1211 }
1215 }
1217 }
1219 /*
1220 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1221 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1222 ** releasing blocks in this transaction for reuse as data blocks.
1223 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1224 **
1225 */
1229 {
1242 #ifdef CONFIG_REISERFS_CHECK
1245 #endif
1247 }
1248 }
1252 b_data);
1262 else
1270 }
1271 }
1273 }
1278 {
1280 }
1282 /*
1283 ** flush any and all journal lists older than you are
1284 ** can only be called from flush_journal_list
1285 */
1288 {
1294 /* we know we are the only ones flushing things, no extra race
1295 * protection is required.
1296 */
1297 restart:
1299 /* Did we wrap? */
1305 /* do not flush all */
1308 /* other_jl is now deleted from the list */
1310 }
1312 }
1316 {
1321 }
1322 }
1324 /* flush a journal list, both commit and real blocks
1325 **
1326 ** always set flushall to 1, unless you are calling from inside
1327 ** flush_journal_list
1328 **
1329 ** IMPORTANT. This can only be called while there are no journal writers,
1330 ** and the journal is locked. That means it can only be called from
1331 ** do_journal_end, or by journal_release
1332 */
1335 {
1351 }
1354 /* if flushall == 0, the lock is already held */
1359 }
1366 }
1368 /* if all the work is already done, get out of here */
1372 }
1374 /* start by putting the commit list on disk. This will also flush
1375 ** the commit lists of any olders transactions
1376 */
1383 /* are we done now? */
1387 }
1389 /* loop through each cnode, see if we need to write it,
1390 ** or wait on a more recent transaction, or just ignore it
1391 */
1395 }
1401 /* blocknr of 0 is no longer in the hash, ignore it */
1404 }
1406 /* This transaction failed commit. Don't write out to the disk */
1411 /* the order is important here. We check pjl to make sure we
1412 ** don't clear BH_JDirty_wait if we aren't the one writing this
1413 ** block to disk
1414 */
1418 /* we do this to make sure nobody releases the buffer while
1419 ** we are working with it
1420 */
1428 /* everything with !pjl && jwait should be writable */
1430 }
1431 }
1433 /* if someone has this block in a newer transaction, just make
1434 ** sure they are committed, and don't try writing it to disk
1435 */
1440 }
1442 /* bh == NULL when the block got to disk on its own, OR,
1443 ** the block got freed in a future transaction
1444 */
1447 }
1449 /* this should never happen. kupdate_one_transaction has this list
1450 ** locked while it works, so we should never see a buffer here that
1451 ** is not marked JDirty_wait
1452 */
1455 "BAD! buffer %llu %cdirty %cjwait, "
1460 }
1462 /* we inc again because saved_bh gets decremented at free_cnode */
1469 else
1471 count++;
1477 }
1478 free_cnode:
1482 /* we incremented this to keep others from taking the buffer head away */
1487 }
1488 }
1489 }
1497 }
1506 }
1508 #ifdef CONFIG_REISERFS_CHECK
1511 #endif
1513 }
1514 /* note, we must clear the JDirty_wait bit after the up to date
1515 ** check, otherwise we race against our flushpage routine
1516 */
1520 /* drop one ref for us */
1522 /* drop one ref for journal_mark_dirty */
1524 }
1526 }
1527 }
1532 __func__);
1533 flush_older_and_return:
1535 /* before we can update the journal header block, we _must_ flush all
1536 ** real blocks from all older transactions to disk. This is because
1537 ** once the header block is updated, this transaction will not be
1538 ** replayed after a crash
1539 */
1542 }
1545 /* before we can remove everything from the hash tables for this
1546 ** transaction, we must make sure it can never be replayed
1547 **
1548 ** since we are only called from do_journal_end, we know for sure there
1549 ** are no allocations going on while we are flushing journal lists. So,
1550 ** we only need to update the journal header block for the last list
1551 ** being flushed
1552 */
1554 err =
1562 __func__);
1563 }
1573 }
1576 /* not strictly required since we are freeing the list, but it should
1577 * help find code using dead lists later on
1578 */
1590 }
1594 {
1602 /* if the blocknr == 0, this has been cleared from the hash,
1603 ** skip it
1604 */
1607 }
1610 next:
1613 }
1615 }
1620 {
1628 }
1632 /* if the blocknr == 0, this has been cleared from the hash,
1633 ** skip it
1634 */
1637 }
1640 /* we can race against journal_mark_freed when we try
1641 * to lock_buffer(cn->bh), so we have to inc the buffer
1642 * count, and recheck things after locking
1643 */
1652 ret++;
1654 /* note, cn->bh might be null now */
1656 }
1658 }
1659 next:
1662 }
1664 }
1666 /* used by flush_commit_list */
1669 {
1677 /* look for a more recent transaction that logged this
1678 ** buffer. Only the most recent transaction with a buffer in
1679 ** it is allowed to send that buffer to disk
1680 */
1685 /* if the buffer is prepared, it will either be logged
1686 * or restored. If restored, we need to make sure
1687 * it actually gets marked dirty
1688 */
1695 }
1696 }
1698 }
1700 }
1707 {
1720 }
1722 /* we've got j_flush_mutex held, nobody is going to delete any
1723 * of these lists out from underneath us
1724 */
1733 }
1738 transactions_flushed++;
1742 /* did we wrap? */
1745 }
1748 /* don't bother with older transactions */
1751 }
1754 }
1756 done:
1759 }
1761 /* for o_sync and fsync heavy applications, they tend to use
1762 ** all the journa list slots with tiny transactions. These
1763 ** trigger lots and lots of calls to update the header block, which
1764 ** adds seeks and slows things down.
1765 **
1766 ** This function tries to clear out a large chunk of the journal lists
1767 ** at once, which makes everything faster since only the newest journal
1768 ** list updates the header block
1769 */
1772 {
1785 /* in data logging mode, try harder to flush a lot of blocks */
1788 /* flush for 256 transactions or limit blocks, whichever comes first */
1793 }
1797 }
1803 }
1804 /* try to find a group of blocks we can flush across all the
1805 ** transactions, but only bother if we've actually spanned
1806 ** across multiple lists
1807 */
1810 }
1813 }
1815 /*
1816 ** removes any nodes in table with name block and dev as bh.
1817 ** only touchs the hnext and hprev pointers.
1818 */
1823 {
1830 }
1838 }
1843 }
1851 }
1853 }
1854 }
1857 {
1867 }
1868 /* j_header_bh is on the journal dev, make sure not to release the journal
1869 * dev until we brelse j_header_bh
1870 */
1873 }
1875 /*
1876 ** call on unmount. Only set error to 1 if you haven't made your way out
1877 ** of read_super() yet. Any other caller must keep error at 0.
1878 */
1881 {
1886 /* we only want to flush out transactions if we were called with error == 0
1887 */
1889 /* end the current trans */
1893 /* make sure something gets logged to force our way into the flush code */
1902 }
1903 }
1905 /* this also catches errors during the do_journal_end above */
1915 }
1916 }
1918 reiserfs_mounted_fs_count--;
1919 /* wait for all commits to finish */
1922 /*
1923 * We must release the write lock here because
1924 * the workqueue job (flush_async_commit) needs this lock
1925 */
1932 }
1939 }
1941 /*
1942 ** call on unmount. flush all journal trans, release all alloc'd ram
1943 */
1946 {
1948 }
1950 /*
1951 ** only call from an error condition inside reiserfs_read_super!
1952 */
1955 {
1957 }
1959 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1963 {
1969 }
1971 }
1973 /* returns 0 if it did not find a description block
1974 ** returns -1 if it found a corrupt commit block
1975 ** returns 1 if both desc and commit were valid
1976 */
1981 {
1996 "journal-986: transaction "
1997 "is valid returning because trans_id %d is greater than "
2002 }
2006 "journal-1087: transaction "
2007 "is valid returning because mount_id %d is less than "
2012 }
2015 "Bad transaction length %d "
2019 }
2022 /* ok, we have a journal description block, lets see if the transaction was valid */
2023 c_bh =
2033 "journal_transaction_is_valid, commit offset %ld had bad "
2044 "journal-1004: "
2045 "transaction_is_valid setting oldest invalid trans_id "
2048 }
2050 }
2053 "journal-1006: found valid "
2062 }
2063 }
2066 {
2070 }
2071 }
2073 /*
2074 ** given the start, and values for the oldest acceptable transactions,
2075 ** this either reads in a replays a transaction, or returns because the transaction
2076 ** is invalid, or too old.
2077 */
2083 {
2108 cur_dblock -
2112 }
2115 "journal_read_trans skipping because %d is != "
2117 newest_mount_id);
2120 }
2127 }
2131 "journal_read_transaction, "
2140 }
2148 }
2151 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2164 }
2165 /* get all the buffer heads */
2182 }
2185 "REPLAY FAILURE fsck required! "
2186 "Block to replay is outside of "
2189 }
2190 /* make sure we don't try to replay onto log or reserved area */
2194 "REPLAY FAILURE fsck required! "
2196 abort_replay:
2204 }
2205 }
2206 /* read in the log blocks, memcpy to the corresponding real block */
2216 "REPLAY FAILURE fsck required! "
2226 }
2231 }
2232 /* flush out the real blocks */
2236 }
2241 "REPLAY FAILURE, fsck required! "
2250 }
2252 }
2253 cur_dblock =
2261 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2265 /* check for trans_id overflow */
2273 }
2275 /* This function reads blocks starting from block and to max_block of bufsize
2276 size (but no more than BUFNR blocks at a time). This proved to improve
2277 mounting speed on self-rebuilding raid5 arrays at least.
2278 Right now it is only used from journal code. But later we might use it
2279 from other places.
2280 Note: Do not use journal_getblk/sb_getblk functions here! */
2283 b_blocknr_t max_block)
2284 {
2296 }
2306 }
2316 }
2318 /*
2319 ** read and replay the log
2320 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2321 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2322 **
2323 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2324 **
2325 ** On exit, it sets things up so the first transaction will work correctly.
2326 */
2328 {
2350 /* step 1, read in the journal header block. Check the transaction it says
2351 ** is the first unflushed, and if that transaction is not valid,
2352 ** replay is done
2353 */
2359 }
2364 oldest_start =
2370 "journal-1153: found in "
2371 "header: first_unflushed_offset %d, last_flushed_trans_id "
2376 /* now, we try to read the first unflushed offset. If it is not valid,
2377 ** there is nothing more we can do, and it makes no sense to read
2378 ** through the whole log.
2379 */
2380 d_bh =
2387 }
2390 }
2392 /* ok, there are transactions that need to be replayed. start with the first log block, find
2393 ** all the valid transactions, and pick out the oldest.
2394 */
2399 /* Note that it is required for blocksize of primary fs device and journal
2400 device to be the same */
2401 d_bh =
2406 ret =
2417 "journal-1179: Setting "
2419 oldest_start -
2420 SB_ONDISK_JOURNAL_1st_BLOCK
2423 /* one we just read was older */
2427 "journal-1180: Resetting "
2429 oldest_start -
2430 SB_ONDISK_JOURNAL_1st_BLOCK
2432 }
2436 "journal-1299: Setting "
2439 }
2442 cur_dblock++;
2443 }
2445 }
2447 start_log_replay:
2451 "journal-1206: Starting replay "
2454 oldest_trans_id);
2456 }
2459 ret =
2466 }
2467 cur_dblock =
2469 replay_count++;
2472 }
2477 }
2478 /* j_start does not get set correctly if we don't replay any transactions.
2479 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2480 ** copy the trans_id from the header
2481 */
2486 /* check for trans_id overflow */
2494 }
2502 }
2506 /* replay failed, caller must call free_journal_ram and abort
2507 ** the mount
2508 */
2510 }
2512 }
2515 {
2527 }
2530 {
2532 }
2536 {
2544 }
2549 }
2551 }
2556 {
2558 dev_t jdev;
2571 /* there is no "jdev" option and journal is on separate device */
2576 journal);
2589 }
2600 }
2607 }
2609 /**
2610 * When creating/tuning a file system user can assign some
2611 * journal params within boundaries which depend on the ratio
2612 * blocksize/standard_blocksize.
2613 *
2614 * For blocks >= standard_blocksize transaction size should
2615 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2616 * then JOURNAL_TRANS_MAX_DEFAULT.
2617 *
2618 * For blocks < standard_blocksize these boundaries should be
2619 * decreased proportionally.
2620 */
2621 #define REISERFS_STANDARD_BLKSIZE (4096)
2625 {
2627 /* Non-default journal params.
2628 Do sanity check for them. */
2636 JOURNAL_MIN_RATIO) {
2638 "bad transaction max size (%u). "
2641 }
2645 "bad transaction max batch (%u). "
2648 }
2650 /* Default journal params.
2651 The file system was created by old version
2652 of mkreiserfs, so some fields contain zeros,
2653 and we need to advise proper values for them */
2658 }
2662 }
2664 }
2666 /*
2667 ** must be called once on fs mount. calls journal_read for you
2668 */
2671 {
2681 /*
2682 * Unlock here to avoid various RECLAIM-FS-ON <-> IN-RECLAIM-FS
2683 * dependency inversion warnings.
2684 */
2692 }
2706 /* reserved for journal area support */
2708 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2712 REISERFS_DISK_OFFSET_IN_BYTES /
2715 /* Sanity check to see is the standard journal fitting within first bitmap
2716 (actual for small blocksizes) */
2721 "journal does not fit for area addressed "
2722 "by first of bitmap blocks. It starts at "
2728 }
2730 /*
2731 * We need to unlock here to avoid creating the following
2732 * dependency:
2733 * reiserfs_lock -> sysfs_mutex
2734 * Because the reiserfs mmap path creates the following dependency:
2735 * mm->mmap -> reiserfs_lock, hence we have
2736 * mm->mmap -> reiserfs_lock ->sysfs_mutex
2737 * This would ends up in a circular dependency with sysfs readdir path
2738 * which does sysfs_mutex -> mm->mmap_sem
2739 * This is fine because the reiserfs lock is useless in mount path,
2740 * at least until we call journal_begin. We keep it for paranoid
2741 * reasons.
2742 */
2749 }
2754 /* read journal header */
2762 }
2765 /* make sure that journal matches to the super block */
2770 "journal header magic %x (device %s) does "
2777 }
2792 }
2795 "journal first block %u, max trans len %u, max batch %u, "
2842 "allocation failed (%ld bytes). Journal is "
2843 "too large for available memory. Usually "
2847 }
2856 }
2861 }
2863 reiserfs_mounted_fs_count++;
2868 }
2873 free_and_return:
2876 }
2878 /*
2879 ** test for a polite end of the current transaction. Used by file_write, and should
2880 ** be used by delete to make sure they don't write more than can fit inside a single
2881 ** transaction
2882 */
2885 {
2888 /* cannot restart while nested */
2898 }
2899 /* protected by the BKL here */
2903 }
2905 /* this must be called inside a transaction, and requires the
2906 ** kernel_lock to be held
2907 */
2909 {
2915 }
2917 /* this must be called without a transaction started, and does not
2918 ** require BKL
2919 */
2921 {
2925 }
2927 /* this must be called without a transaction started, and does not
2928 ** require BKL
2929 */
2931 {
2935 }
2938 {
2939 wait_queue_t wait;
2943 /*
2944 * we don't want to use wait_event here because
2945 * we only want to wait once.
2946 */
2954 }
2957 }
2960 {
2964 }
2967 {
2979 }
2985 }
2986 }
2988 /* join == true if you must join an existing transaction.
2989 ** join == false if you can deal with waiting for others to finish
2990 **
2991 ** this will block until the transaction is joinable. send the number of blocks you
2992 ** expect to use in nblocks.
2993 */
2997 {
3009 /* set here for journal_join */
3013 relock:
3019 }
3029 }
3032 /* if there is no room in the journal OR
3033 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3034 ** we don't sleep if there aren't other writers
3035 */
3038 (!join
3055 sched_count++;
3058 }
3059 }
3060 /* don't mess with joining the transaction if all we have to do is
3061 * wait for someone else to do a commit
3062 */
3067 }
3069 }
3074 /* someone might have ended the transaction while we joined */
3079 }
3086 }
3087 /* we are the first writer, set trans_id */
3090 }
3100 out_fail:
3102 /* Re-set th->t_super, so we can properly keep track of how many
3103 * persistent transactions there are. We need to do this so if this
3104 * call is part of a failed restart_transaction, we can free it later */
3107 }
3110 super_block
3113 {
3117 /* if we're nesting into an existing transaction. It will be
3118 ** persistent on its own
3119 */
3126 }
3134 }
3138 }
3141 {
3146 else
3151 }
3153 }
3157 {
3160 /* this keeps do_journal_end from NULLing out the current->journal_info
3161 ** pointer
3162 */
3166 }
3170 {
3173 /* this keeps do_journal_end from NULLing out the current->journal_info
3174 ** pointer
3175 */
3179 }
3183 {
3189 /* we are nesting into the current transaction */
3196 "BAD: refcount <= 1, but "
3200 /* we've ended up with a handle from a different filesystem.
3201 ** save it and restore on journal_end. This should never
3202 ** really happen...
3203 */
3208 }
3211 }
3215 /* I guess this boils down to being the reciprocal of clm-2100 above.
3216 * If do_journal_begin_r fails, we need to put it back, since journal_end
3217 * won't be called to do it. */
3220 else
3224 }
3226 /*
3227 ** puts bh into the current transaction. If it was already there, reorders removes the
3228 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3229 **
3230 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3231 ** transaction is committed.
3232 **
3233 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3234 */
3237 {
3249 }
3255 /* already in this transaction, we are done */
3259 }
3261 /* this must be turned into a panic instead of a warning. We can't allow
3262 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3263 ** could get to disk too early. NOT GOOD.
3264 */
3267 "buffer %llu bad state "
3274 }
3281 }
3282 /* this error means I've screwed up, and we've overflowed the transaction.
3283 ** Nothing can be done here, except make the FS readonly or panic.
3284 */
3289 }
3295 }
3299 }
3303 /* now put this guy on the end */
3308 }
3313 }
3324 }
3325 }
3335 }
3337 }
3341 {
3349 }
3356 /* we aren't allowed to close a nested transaction on a different
3357 ** filesystem from the one in the task struct
3358 */
3364 }
3368 }
3369 }
3371 /* removes from the current transaction, relsing and descrementing any counters.
3372 ** also files the removed buffer directly onto the clean list
3373 **
3374 ** called by journal_mark_freed when a block has been deleted
3375 **
3376 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3377 */
3380 {
3389 }
3393 }
3396 }
3399 }
3402 }
3416 }
3418 }
3423 }
3425 /*
3426 ** for any cnode in a journal list, it can only be dirtied of all the
3427 ** transactions that include it are committed to disk.
3428 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3429 ** and 0 if you aren't
3430 **
3431 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3432 ** blocks for a given transaction on disk
3433 **
3434 */
3436 {
3442 /* first test hprev. These are all newer than cn, so any node here
3443 ** with the same block number and dev means this node can't be sent
3444 ** to disk right now.
3445 */
3450 }
3452 }
3453 /* then test hnext. These are all older than cn. As long as they
3454 ** are committed to the log, it is safe to write cn to disk
3455 */
3462 }
3464 }
3466 }
3468 /* syncs the commit blocks, but does not force the real buffers to disk
3469 ** will wait until the current transaction is done/committed before returning
3470 */
3473 {
3477 /* you can sync while nested, very, very bad */
3483 }
3485 }
3487 /*
3488 ** writeback the pending async commits to disk
3489 */
3491 {
3500 /* last entry is the youngest, commit it and you get everything */
3504 }
3506 }
3508 /*
3509 ** flushes any old transactions to disk
3510 ** ends the current transaction if it is too old
3511 */
3513 {
3519 /* safety check so we don't flush while we are replaying the log during
3520 * mount
3521 */
3524 }
3526 /* check the current transaction. If there are no writers, and it is
3527 * too old, finish it, and force the commit blocks to disk
3528 */
3540 /* we're only being called from kreiserfsd, it makes no sense to do
3541 ** an async commit so that kreiserfsd can do it later
3542 */
3544 }
3545 }
3547 }
3549 /*
3550 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3551 **
3552 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3553 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3554 ** flushes the commit list and returns 0.
3555 **
3556 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3557 **
3558 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3559 */
3563 {
3578 }
3581 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3583 }
3585 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3586 ** will be dealt with by next transaction that actually writes something, but should be taken
3587 ** care of in this trans
3588 */
3591 /* if wcount > 0, and we are called to with flush or commit_now,
3592 ** we wait on j_join_wait. We will wake up when the last writer has
3593 ** finished the transaction, and started it on its way to the disk.
3594 ** Then, we flush the commit or journal list, and just return 0
3595 ** because the rest of journal end was already done for this transaction.
3596 */
3608 }
3611 /* sleep while the current transaction is still j_jlocked */
3620 }
3622 }
3623 }
3630 }
3632 }
3635 }
3637 /* deal with old transactions where we are the last writers */
3642 }
3643 /* don't batch when someone is waiting on j_join_wait */
3644 /* don't batch when syncing the commit or flushing the whole trans */
3652 }
3658 }
3660 }
3662 /*
3663 ** Does all the work that makes deleting blocks safe.
3664 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3665 **
3666 ** otherwise:
3667 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3668 ** before this transaction has finished.
3669 **
3670 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3671 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3672 ** the block can't be reallocated yet.
3673 **
3674 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3675 */
3678 {
3690 }
3691 /* if it is journal new, we just remove it from this transaction */
3698 /* set the bit for this block in the journal bitmap for this transaction */
3703 }
3706 /* Note, the entire while loop is not allowed to schedule. */
3711 }
3714 /* find all older transactions with this block, make sure they don't try to write it out */
3716 blocknr);
3722 /* remove_from_transaction will brelse the buffer if it was
3723 ** in the current trans
3724 */
3726 bh);
3729 bh);
3737 }
3738 }
3742 j_nonzerolen));
3743 }
3745 }
3746 }
3748 }
3749 }
3754 }
3757 {
3761 }
3763 /*
3764 * returns -1 on error, 0 if no commits/barriers were done and 1
3765 * if a transaction was actually committed and the barrier was done
3766 */
3769 {
3775 /* is it from the current transaction, or from an unknown transaction? */
3778 /* try to let other writers come in and grow this transaction */
3782 }
3788 /* someone might have ended this transaction while we joined */
3795 }
3802 /* this gets tricky, we have to make sure the journal list in
3803 * the inode still exists. We know the list is still around
3804 * if we've got a larger transaction id than the oldest list
3805 */
3806 flush_commit_only:
3808 /*
3809 * we only set ret to 1 when we know for sure
3810 * the barrier hasn't been started yet on the commit
3811 * block.
3812 */
3818 }
3819 }
3820 /* otherwise the list is gone, and long since committed */
3822 }
3825 {
3829 /* for the whole inode, assume unset id means it was
3830 * changed in the current transaction. More conservative
3831 */
3835 /* jl will be updated in __commit_trans_jl */
3836 }
3839 }
3843 {
3848 }
3858 }
3859 }
3861 }
3864 /*
3865 ** before we can change a metadata block, we have to make sure it won't
3866 ** be written to disk while we are altering it. So, we must:
3867 ** clean it
3868 ** wait on it.
3869 **
3870 */
3873 {
3880 }
3885 }
3888 }
3891 {
3900 /* this check should always be run, to send old lists to disk */
3907 }
3908 }
3909 }
3911 /*
3912 ** long and ugly. If flush, will not return until all commit
3913 ** blocks and all real buffers in the trans are on disk.
3914 ** If no_async, won't return until all commit blocks are on disk.
3915 **
3916 ** keep reading, there are comments as you go along
3917 **
3918 ** If the journal is aborted, we just clean up. Things like flushing
3919 ** journal lists, etc just won't happen.
3920 */
3924 {
3946 /* protect flush_older_commits from doing mistakes if the
3947 transaction ID counter gets overflowed. */
3959 }
3965 }
3969 }
3971 /* check_journal_end locks the journal, and unlocks if it does not return 1
3972 ** it tells us if we should continue with the journal_end, or just return
3973 */
3979 }
3981 /* check_journal_end might set these, check again */
3984 }
3986 /*
3987 ** j must wait means we have to flush the log blocks, and the real blocks for
3988 ** this transaction
3989 */
3992 }
3993 #ifdef REISERFS_PREALLOCATE
3994 /* quota ops might need to nest, setup the journal_info pointer for them
3995 * and raise the refcount so that it is > 0. */
3999 * the transaction */
4002 #endif
4004 /* setup description block */
4005 d_bh =
4015 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4024 /* init this journal list */
4027 /* we lock the commit before doing anything because
4028 * we want to make sure nobody tries to run flush_commit_list until
4029 * the new transaction is fully setup, and we've already flushed the
4030 * ordered bh list
4031 */
4034 /* save the transaction id in case we need to commit it later */
4047 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4048 ** for each real block, add it to the journal list hash,
4049 ** copy into real block index array in the commit or desc block
4050 */
4058 }
4061 }
4066 }
4068 /* make sure the block we are trying to log is not a block
4069 of journal or reserved area */
4074 "Trying to log block %lu, "
4077 }
4090 }
4092 i--;
4093 }
4094 }
4100 /* special check in case all buffers in the journal were marked for not logging */
4103 /* we're about to dirty all the log blocks, mark the description block
4104 * dirty now too. Don't mark the commit block dirty until all the
4105 * others are on disk
4106 */
4109 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4115 /* copy all the real blocks into log area. dirty log blocks */
4120 tmp_bh =
4124 jindex) %
4134 jindex++;
4138 /* JDirty cleared sometime during transaction. don't log this one */
4140 "BAD, buffer in journal hash, "
4143 }
4150 }
4152 /* we are done with both the c_bh and d_bh, but
4153 ** c_bh must be written after all other commit blocks,
4154 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4155 */
4159 /* now it is safe to insert this transaction on the main list */
4164 /* reset journal values for the next transaction */
4175 /* check for trans_id overflow */
4185 // make sure reiserfs_add_jh sees the new current_jl before we
4186 // write out the tails
4189 /* tail conversion targets have to hit the disk before we end the
4190 * transaction. Otherwise a later transaction might repack the tail
4191 * before this transaction commits, leaving the data block unflushed and
4192 * clean, if we crash before the later transaction commits, the data block
4193 * is lost.
4194 */
4200 }
4204 /* honor the flush wishes from the caller, simple commits can
4205 ** be done outside the journal lock, they are done below
4206 **
4207 ** if we don't flush the commit list right now, we put it into
4208 ** the work queue so the people waiting on the async progress work
4209 ** queue don't wait for this proc to flush journal lists and such.
4210 */
4217 /* if the next transaction has any chance of wrapping, flush
4218 ** transactions that might get overwritten. If any journal lists are very
4219 ** old flush them as well.
4220 */
4221 first_jl:
4232 /* if we don't cross into the next transaction and we don't
4233 * wrap, there is no way we can overlap any later transactions
4234 * break now
4235 */
4237 }
4247 /* we don't overlap anything from out start to the end of the
4248 * log, and our wrapped portion doesn't overlap anything at
4249 * the start of the log. We can break
4250 */
4252 }
4253 }
4254 }
4263 }
4267 /* wake up any body waiting to join. */
4274 }
4275 out:
4279 /* Re-set th->t_super, so we can properly keep track of how many
4280 * persistent transactions there are. We need to do this so if this
4281 * call is part of a failed restart_transaction, we can free it later */
4285 }
4287 /* Send the file system read only and refuse new transactions */
4289 {
4300 #ifdef CONFIG_REISERFS_CHECK
4302 #endif
4303 }