| blob | afcadcc03e8ac87c7f25f3e2393b3c108daaf91d |
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 {
953 }
956 }
958 /*
959 ** if this journal list still has commit blocks unflushed, send them to disk.
960 **
961 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
962 ** Before the commit block can by written, every other log block must be safely on disk
963 **
964 */
967 {
969 b_blocknr_t bn;
980 }
982 /* before we can put our commit blocks on disk, we have to make sure everyone older than
983 ** us is on disk too
984 */
991 /* list disappeared during flush_older_commits. return */
993 }
994 }
996 /* make sure nobody is trying to flush this one at the same time */
1002 }
1005 /* this commit is done, exit */
1009 }
1012 }
1017 /*
1018 * We might sleep in numerous places inside
1019 * write_ordered_buffers. Relax the write lock.
1020 */
1027 }
1029 /*
1030 * for the description block and all the log blocks, submit any buffers
1031 * that haven't already reached the disk. Try to write at least 256
1032 * log blocks. later on, we will only wait on blocks that correspond
1033 * to this transaction, but while we're unplugging we might as well
1034 * get a chunk of data on there.
1035 */
1049 }
1051 }
1052 }
1063 // since we're using ll_rw_blk above, it might have skipped over
1064 // a locked buffer. Double check here
1065 //
1066 /* redundant, sync_dirty_buffer() checks */
1071 }
1073 #ifdef CONFIG_REISERFS_CHECK
1076 #endif
1078 }
1082 }
1086 /* If there was a write error in the journal - we can't commit
1087 * this transaction - it will be invalid and, if successful,
1088 * will just end up propagating the write error out to
1089 * the file system. */
1097 else
1100 }
1102 /* If there was a write error in the journal - we can't commit this
1103 * transaction - it will be invalid and, if successful, will just end
1104 * up propagating the write error out to the filesystem. */
1106 #ifdef CONFIG_REISERFS_CHECK
1108 #endif
1110 }
1116 }
1119 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1124 /* mark the metadata dirty */
1131 }
1133 put_jl:
1138 __func__);
1140 }
1142 /*
1143 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1144 ** returns NULL if it can't find anything
1145 */
1147 reiserfs_journal_cnode
1149 {
1157 }
1159 }
1161 }
1164 {
1174 }
1176 }
1183 /*
1184 ** once all the real blocks have been flushed, it is safe to remove them from the
1185 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1186 ** block to be reallocated for data blocks if it had been deleted.
1187 */
1191 {
1196 /* which is better, to lock once around the whole loop, or
1197 ** to lock for each call to remove_journal_hash?
1198 */
1206 }
1210 }
1214 }
1216 }
1218 /*
1219 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1220 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1221 ** releasing blocks in this transaction for reuse as data blocks.
1222 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1223 **
1224 */
1228 {
1241 #ifdef CONFIG_REISERFS_CHECK
1244 #endif
1246 }
1247 }
1251 b_data);
1261 else
1269 }
1270 }
1272 }
1277 {
1279 }
1281 /*
1282 ** flush any and all journal lists older than you are
1283 ** can only be called from flush_journal_list
1284 */
1287 {
1293 /* we know we are the only ones flushing things, no extra race
1294 * protection is required.
1295 */
1296 restart:
1298 /* Did we wrap? */
1304 /* do not flush all */
1307 /* other_jl is now deleted from the list */
1309 }
1311 }
1315 {
1320 }
1321 }
1323 /* flush a journal list, both commit and real blocks
1324 **
1325 ** always set flushall to 1, unless you are calling from inside
1326 ** flush_journal_list
1327 **
1328 ** IMPORTANT. This can only be called while there are no journal writers,
1329 ** and the journal is locked. That means it can only be called from
1330 ** do_journal_end, or by journal_release
1331 */
1334 {
1350 }
1353 /* if flushall == 0, the lock is already held */
1358 }
1365 }
1367 /* if all the work is already done, get out of here */
1371 }
1373 /* start by putting the commit list on disk. This will also flush
1374 ** the commit lists of any olders transactions
1375 */
1382 /* are we done now? */
1386 }
1388 /* loop through each cnode, see if we need to write it,
1389 ** or wait on a more recent transaction, or just ignore it
1390 */
1394 }
1400 /* blocknr of 0 is no longer in the hash, ignore it */
1403 }
1405 /* This transaction failed commit. Don't write out to the disk */
1410 /* the order is important here. We check pjl to make sure we
1411 ** don't clear BH_JDirty_wait if we aren't the one writing this
1412 ** block to disk
1413 */
1417 /* we do this to make sure nobody releases the buffer while
1418 ** we are working with it
1419 */
1427 /* everything with !pjl && jwait should be writable */
1429 }
1430 }
1432 /* if someone has this block in a newer transaction, just make
1433 ** sure they are committed, and don't try writing it to disk
1434 */
1439 }
1441 /* bh == NULL when the block got to disk on its own, OR,
1442 ** the block got freed in a future transaction
1443 */
1446 }
1448 /* this should never happen. kupdate_one_transaction has this list
1449 ** locked while it works, so we should never see a buffer here that
1450 ** is not marked JDirty_wait
1451 */
1454 "BAD! buffer %llu %cdirty %cjwait, "
1459 }
1461 /* we inc again because saved_bh gets decremented at free_cnode */
1468 else
1470 count++;
1476 }
1477 free_cnode:
1481 /* we incremented this to keep others from taking the buffer head away */
1486 }
1487 }
1488 }
1496 }
1505 }
1507 #ifdef CONFIG_REISERFS_CHECK
1510 #endif
1512 }
1513 /* note, we must clear the JDirty_wait bit after the up to date
1514 ** check, otherwise we race against our flushpage routine
1515 */
1519 /* drop one ref for us */
1521 /* drop one ref for journal_mark_dirty */
1523 }
1525 }
1526 }
1531 __func__);
1532 flush_older_and_return:
1534 /* before we can update the journal header block, we _must_ flush all
1535 ** real blocks from all older transactions to disk. This is because
1536 ** once the header block is updated, this transaction will not be
1537 ** replayed after a crash
1538 */
1541 }
1544 /* before we can remove everything from the hash tables for this
1545 ** transaction, we must make sure it can never be replayed
1546 **
1547 ** since we are only called from do_journal_end, we know for sure there
1548 ** are no allocations going on while we are flushing journal lists. So,
1549 ** we only need to update the journal header block for the last list
1550 ** being flushed
1551 */
1553 err =
1561 __func__);
1562 }
1572 }
1575 /* not strictly required since we are freeing the list, but it should
1576 * help find code using dead lists later on
1577 */
1589 }
1593 {
1601 /* if the blocknr == 0, this has been cleared from the hash,
1602 ** skip it
1603 */
1606 }
1609 next:
1612 }
1614 }
1619 {
1627 }
1631 /* if the blocknr == 0, this has been cleared from the hash,
1632 ** skip it
1633 */
1636 }
1639 /* we can race against journal_mark_freed when we try
1640 * to lock_buffer(cn->bh), so we have to inc the buffer
1641 * count, and recheck things after locking
1642 */
1651 ret++;
1653 /* note, cn->bh might be null now */
1655 }
1657 }
1658 next:
1661 }
1663 }
1665 /* used by flush_commit_list */
1668 {
1676 /* look for a more recent transaction that logged this
1677 ** buffer. Only the most recent transaction with a buffer in
1678 ** it is allowed to send that buffer to disk
1679 */
1684 /* if the buffer is prepared, it will either be logged
1685 * or restored. If restored, we need to make sure
1686 * it actually gets marked dirty
1687 */
1694 }
1695 }
1697 }
1699 }
1706 {
1719 }
1721 /* we've got j_flush_mutex held, nobody is going to delete any
1722 * of these lists out from underneath us
1723 */
1732 }
1737 transactions_flushed++;
1741 /* did we wrap? */
1744 }
1747 /* don't bother with older transactions */
1750 }
1753 }
1755 done:
1758 }
1760 /* for o_sync and fsync heavy applications, they tend to use
1761 ** all the journa list slots with tiny transactions. These
1762 ** trigger lots and lots of calls to update the header block, which
1763 ** adds seeks and slows things down.
1764 **
1765 ** This function tries to clear out a large chunk of the journal lists
1766 ** at once, which makes everything faster since only the newest journal
1767 ** list updates the header block
1768 */
1771 {
1784 /* in data logging mode, try harder to flush a lot of blocks */
1787 /* flush for 256 transactions or limit blocks, whichever comes first */
1792 }
1796 }
1802 }
1803 /* try to find a group of blocks we can flush across all the
1804 ** transactions, but only bother if we've actually spanned
1805 ** across multiple lists
1806 */
1809 }
1812 }
1814 /*
1815 ** removes any nodes in table with name block and dev as bh.
1816 ** only touchs the hnext and hprev pointers.
1817 */
1822 {
1829 }
1837 }
1842 }
1850 }
1852 }
1853 }
1856 {
1866 }
1867 /* j_header_bh is on the journal dev, make sure not to release the journal
1868 * dev until we brelse j_header_bh
1869 */
1872 }
1874 /*
1875 ** call on unmount. Only set error to 1 if you haven't made your way out
1876 ** of read_super() yet. Any other caller must keep error at 0.
1877 */
1880 {
1885 /* we only want to flush out transactions if we were called with error == 0
1886 */
1888 /* end the current trans */
1892 /* make sure something gets logged to force our way into the flush code */
1901 }
1902 }
1904 /* this also catches errors during the do_journal_end above */
1914 }
1915 }
1917 reiserfs_mounted_fs_count--;
1918 /* wait for all commits to finish */
1921 /*
1922 * We must release the write lock here because
1923 * the workqueue job (flush_async_commit) needs this lock
1924 */
1933 }
1940 }
1942 /*
1943 ** call on unmount. flush all journal trans, release all alloc'd ram
1944 */
1947 {
1949 }
1951 /*
1952 ** only call from an error condition inside reiserfs_read_super!
1953 */
1956 {
1958 }
1960 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1964 {
1970 }
1972 }
1974 /* returns 0 if it did not find a description block
1975 ** returns -1 if it found a corrupt commit block
1976 ** returns 1 if both desc and commit were valid
1977 */
1982 {
1997 "journal-986: transaction "
1998 "is valid returning because trans_id %d is greater than "
2003 }
2007 "journal-1087: transaction "
2008 "is valid returning because mount_id %d is less than "
2013 }
2016 "Bad transaction length %d "
2020 }
2023 /* ok, we have a journal description block, lets see if the transaction was valid */
2024 c_bh =
2034 "journal_transaction_is_valid, commit offset %ld had bad "
2045 "journal-1004: "
2046 "transaction_is_valid setting oldest invalid trans_id "
2049 }
2051 }
2054 "journal-1006: found valid "
2063 }
2064 }
2067 {
2071 }
2072 }
2074 /*
2075 ** given the start, and values for the oldest acceptable transactions,
2076 ** this either reads in a replays a transaction, or returns because the transaction
2077 ** is invalid, or too old.
2078 */
2084 {
2109 cur_dblock -
2113 }
2116 "journal_read_trans skipping because %d is != "
2118 newest_mount_id);
2121 }
2128 }
2132 "journal_read_transaction, "
2141 }
2149 }
2152 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2165 }
2166 /* get all the buffer heads */
2183 }
2186 "REPLAY FAILURE fsck required! "
2187 "Block to replay is outside of "
2190 }
2191 /* make sure we don't try to replay onto log or reserved area */
2195 "REPLAY FAILURE fsck required! "
2197 abort_replay:
2205 }
2206 }
2207 /* read in the log blocks, memcpy to the corresponding real block */
2217 "REPLAY FAILURE fsck required! "
2227 }
2232 }
2233 /* flush out the real blocks */
2237 }
2242 "REPLAY FAILURE, fsck required! "
2251 }
2253 }
2254 cur_dblock =
2262 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2266 /* check for trans_id overflow */
2274 }
2276 /* This function reads blocks starting from block and to max_block of bufsize
2277 size (but no more than BUFNR blocks at a time). This proved to improve
2278 mounting speed on self-rebuilding raid5 arrays at least.
2279 Right now it is only used from journal code. But later we might use it
2280 from other places.
2281 Note: Do not use journal_getblk/sb_getblk functions here! */
2284 b_blocknr_t max_block)
2285 {
2297 }
2307 }
2317 }
2319 /*
2320 ** read and replay the log
2321 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2322 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2323 **
2324 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2325 **
2326 ** On exit, it sets things up so the first transaction will work correctly.
2327 */
2329 {
2351 /* step 1, read in the journal header block. Check the transaction it says
2352 ** is the first unflushed, and if that transaction is not valid,
2353 ** replay is done
2354 */
2360 }
2365 oldest_start =
2371 "journal-1153: found in "
2372 "header: first_unflushed_offset %d, last_flushed_trans_id "
2377 /* now, we try to read the first unflushed offset. If it is not valid,
2378 ** there is nothing more we can do, and it makes no sense to read
2379 ** through the whole log.
2380 */
2381 d_bh =
2388 }
2391 }
2393 /* ok, there are transactions that need to be replayed. start with the first log block, find
2394 ** all the valid transactions, and pick out the oldest.
2395 */
2400 /* Note that it is required for blocksize of primary fs device and journal
2401 device to be the same */
2402 d_bh =
2407 ret =
2418 "journal-1179: Setting "
2420 oldest_start -
2421 SB_ONDISK_JOURNAL_1st_BLOCK
2424 /* one we just read was older */
2428 "journal-1180: Resetting "
2430 oldest_start -
2431 SB_ONDISK_JOURNAL_1st_BLOCK
2433 }
2437 "journal-1299: Setting "
2440 }
2443 cur_dblock++;
2444 }
2446 }
2448 start_log_replay:
2452 "journal-1206: Starting replay "
2455 oldest_trans_id);
2457 }
2460 ret =
2467 }
2468 cur_dblock =
2470 replay_count++;
2473 }
2478 }
2479 /* j_start does not get set correctly if we don't replay any transactions.
2480 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2481 ** copy the trans_id from the header
2482 */
2487 /* check for trans_id overflow */
2495 }
2503 }
2507 /* replay failed, caller must call free_journal_ram and abort
2508 ** the mount
2509 */
2511 }
2513 }
2516 {
2528 }
2531 {
2533 }
2537 {
2545 }
2550 }
2552 }
2557 {
2559 dev_t jdev;
2572 /* there is no "jdev" option and journal is on separate device */
2577 journal);
2590 }
2601 }
2608 }
2610 /**
2611 * When creating/tuning a file system user can assign some
2612 * journal params within boundaries which depend on the ratio
2613 * blocksize/standard_blocksize.
2614 *
2615 * For blocks >= standard_blocksize transaction size should
2616 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2617 * then JOURNAL_TRANS_MAX_DEFAULT.
2618 *
2619 * For blocks < standard_blocksize these boundaries should be
2620 * decreased proportionally.
2621 */
2622 #define REISERFS_STANDARD_BLKSIZE (4096)
2626 {
2628 /* Non-default journal params.
2629 Do sanity check for them. */
2637 JOURNAL_MIN_RATIO) {
2639 "bad transaction max size (%u). "
2642 }
2646 "bad transaction max batch (%u). "
2649 }
2651 /* Default journal params.
2652 The file system was created by old version
2653 of mkreiserfs, so some fields contain zeros,
2654 and we need to advise proper values for them */
2659 }
2663 }
2665 }
2667 /*
2668 ** must be called once on fs mount. calls journal_read for you
2669 */
2672 {
2687 }
2699 /* reserved for journal area support */
2701 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2705 REISERFS_DISK_OFFSET_IN_BYTES /
2708 /* Sanity check to see is the standard journal fitting within first bitmap
2709 (actual for small blocksizes) */
2714 "journal does not fit for area addressed "
2715 "by first of bitmap blocks. It starts at "
2721 }
2727 }
2731 /* read journal header */
2739 }
2742 /* make sure that journal matches to the super block */
2747 "journal header magic %x (device %s) does "
2754 }
2769 }
2772 "journal first block %u, max trans len %u, max batch %u, "
2817 "allocation failed (%ld bytes). Journal is "
2818 "too large for available memory. Usually "
2822 }
2827 /*
2828 * get_list_bitmap() may call flush_commit_list() which
2829 * requires the lock. Calling flush_commit_list() shouldn't happen
2830 * this early but I like to be paranoid.
2831 */
2839 }
2841 /*
2842 * Journal_read needs to be inspected in order to push down
2843 * the lock further inside (or even remove it).
2844 */
2852 }
2854 reiserfs_mounted_fs_count++;
2861 free_and_return:
2864 }
2866 /*
2867 ** test for a polite end of the current transaction. Used by file_write, and should
2868 ** be used by delete to make sure they don't write more than can fit inside a single
2869 ** transaction
2870 */
2873 {
2876 /* cannot restart while nested */
2886 }
2891 }
2893 /* this must be called inside a transaction
2894 */
2896 {
2902 }
2904 /* this must be called without a transaction started
2905 */
2907 {
2911 }
2913 /* this must be called without a transaction started
2914 */
2916 {
2920 }
2923 {
2924 wait_queue_t wait;
2928 /*
2929 * we don't want to use wait_event here because
2930 * we only want to wait once.
2931 */
2939 }
2942 }
2945 {
2949 }
2952 {
2964 }
2970 }
2971 }
2973 /* join == true if you must join an existing transaction.
2974 ** join == false if you can deal with waiting for others to finish
2975 **
2976 ** this will block until the transaction is joinable. send the number of blocks you
2977 ** expect to use in nblocks.
2978 */
2982 {
2994 /* set here for journal_join */
2998 relock:
3004 }
3014 }
3017 /* if there is no room in the journal OR
3018 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3019 ** we don't sleep if there aren't other writers
3020 */
3023 (!join
3040 sched_count++;
3043 }
3044 }
3045 /* don't mess with joining the transaction if all we have to do is
3046 * wait for someone else to do a commit
3047 */
3052 }
3054 }
3059 /* someone might have ended the transaction while we joined */
3064 }
3071 }
3072 /* we are the first writer, set trans_id */
3075 }
3085 out_fail:
3087 /* Re-set th->t_super, so we can properly keep track of how many
3088 * persistent transactions there are. We need to do this so if this
3089 * call is part of a failed restart_transaction, we can free it later */
3092 }
3095 super_block
3098 {
3102 /* if we're nesting into an existing transaction. It will be
3103 ** persistent on its own
3104 */
3111 }
3119 }
3123 }
3126 {
3131 else
3136 }
3138 }
3142 {
3145 /* this keeps do_journal_end from NULLing out the current->journal_info
3146 ** pointer
3147 */
3151 }
3155 {
3158 /* this keeps do_journal_end from NULLing out the current->journal_info
3159 ** pointer
3160 */
3164 }
3168 {
3174 /* we are nesting into the current transaction */
3181 "BAD: refcount <= 1, but "
3185 /* we've ended up with a handle from a different filesystem.
3186 ** save it and restore on journal_end. This should never
3187 ** really happen...
3188 */
3193 }
3196 }
3200 /* I guess this boils down to being the reciprocal of clm-2100 above.
3201 * If do_journal_begin_r fails, we need to put it back, since journal_end
3202 * won't be called to do it. */
3205 else
3209 }
3211 /*
3212 ** puts bh into the current transaction. If it was already there, reorders removes the
3213 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3214 **
3215 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3216 ** transaction is committed.
3217 **
3218 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3219 */
3222 {
3234 }
3238 /* already in this transaction, we are done */
3242 }
3244 /* this must be turned into a panic instead of a warning. We can't allow
3245 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3246 ** could get to disk too early. NOT GOOD.
3247 */
3250 "buffer %llu bad state "
3257 }
3264 }
3265 /* this error means I've screwed up, and we've overflowed the transaction.
3266 ** Nothing can be done here, except make the FS readonly or panic.
3267 */
3272 }
3278 }
3282 }
3286 /* now put this guy on the end */
3291 }
3296 }
3307 }
3308 }
3318 }
3321 }
3325 {
3333 }
3340 /* we aren't allowed to close a nested transaction on a different
3341 ** filesystem from the one in the task struct
3342 */
3348 }
3352 }
3353 }
3355 /* removes from the current transaction, relsing and descrementing any counters.
3356 ** also files the removed buffer directly onto the clean list
3357 **
3358 ** called by journal_mark_freed when a block has been deleted
3359 **
3360 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3361 */
3364 {
3373 }
3377 }
3380 }
3383 }
3386 }
3400 }
3402 }
3407 }
3409 /*
3410 ** for any cnode in a journal list, it can only be dirtied of all the
3411 ** transactions that include it are committed to disk.
3412 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3413 ** and 0 if you aren't
3414 **
3415 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3416 ** blocks for a given transaction on disk
3417 **
3418 */
3420 {
3426 /* first test hprev. These are all newer than cn, so any node here
3427 ** with the same block number and dev means this node can't be sent
3428 ** to disk right now.
3429 */
3434 }
3436 }
3437 /* then test hnext. These are all older than cn. As long as they
3438 ** are committed to the log, it is safe to write cn to disk
3439 */
3446 }
3448 }
3450 }
3452 /* syncs the commit blocks, but does not force the real buffers to disk
3453 ** will wait until the current transaction is done/committed before returning
3454 */
3457 {
3461 /* you can sync while nested, very, very bad */
3467 }
3469 }
3471 /*
3472 ** writeback the pending async commits to disk
3473 */
3475 {
3484 /* last entry is the youngest, commit it and you get everything */
3488 }
3490 }
3492 /*
3493 ** flushes any old transactions to disk
3494 ** ends the current transaction if it is too old
3495 */
3497 {
3503 /* safety check so we don't flush while we are replaying the log during
3504 * mount
3505 */
3509 /* check the current transaction. If there are no writers, and it is
3510 * too old, finish it, and force the commit blocks to disk
3511 */
3523 /* we're only being called from kreiserfsd, it makes no sense to do
3524 ** an async commit so that kreiserfsd can do it later
3525 */
3527 }
3528 }
3529 }
3531 /*
3532 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3533 **
3534 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3535 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3536 ** flushes the commit list and returns 0.
3537 **
3538 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3539 **
3540 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3541 */
3545 {
3560 }
3563 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3565 }
3567 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3568 ** will be dealt with by next transaction that actually writes something, but should be taken
3569 ** care of in this trans
3570 */
3573 /* if wcount > 0, and we are called to with flush or commit_now,
3574 ** we wait on j_join_wait. We will wake up when the last writer has
3575 ** finished the transaction, and started it on its way to the disk.
3576 ** Then, we flush the commit or journal list, and just return 0
3577 ** because the rest of journal end was already done for this transaction.
3578 */
3590 }
3593 /* sleep while the current transaction is still j_jlocked */
3602 }
3604 }
3605 }
3612 }
3614 }
3617 }
3619 /* deal with old transactions where we are the last writers */
3624 }
3625 /* don't batch when someone is waiting on j_join_wait */
3626 /* don't batch when syncing the commit or flushing the whole trans */
3634 }
3640 }
3642 }
3644 /*
3645 ** Does all the work that makes deleting blocks safe.
3646 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3647 **
3648 ** otherwise:
3649 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3650 ** before this transaction has finished.
3651 **
3652 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3653 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3654 ** the block can't be reallocated yet.
3655 **
3656 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3657 */
3660 {
3672 }
3673 /* if it is journal new, we just remove it from this transaction */
3680 /* set the bit for this block in the journal bitmap for this transaction */
3685 }
3688 /* Note, the entire while loop is not allowed to schedule. */
3693 }
3696 /* find all older transactions with this block, make sure they don't try to write it out */
3698 blocknr);
3704 /* remove_from_transaction will brelse the buffer if it was
3705 ** in the current trans
3706 */
3708 bh);
3711 bh);
3719 }
3720 }
3724 j_nonzerolen));
3725 }
3727 }
3728 }
3730 }
3731 }
3736 }
3739 {
3743 }
3745 /*
3746 * returns -1 on error, 0 if no commits/barriers were done and 1
3747 * if a transaction was actually committed and the barrier was done
3748 */
3751 {
3757 /* is it from the current transaction, or from an unknown transaction? */
3760 /* try to let other writers come in and grow this transaction */
3764 }
3770 /* someone might have ended this transaction while we joined */
3777 }
3784 /* this gets tricky, we have to make sure the journal list in
3785 * the inode still exists. We know the list is still around
3786 * if we've got a larger transaction id than the oldest list
3787 */
3788 flush_commit_only:
3790 /*
3791 * we only set ret to 1 when we know for sure
3792 * the barrier hasn't been started yet on the commit
3793 * block.
3794 */
3800 }
3801 }
3802 /* otherwise the list is gone, and long since committed */
3804 }
3807 {
3811 /* for the whole inode, assume unset id means it was
3812 * changed in the current transaction. More conservative
3813 */
3817 /* jl will be updated in __commit_trans_jl */
3818 }
3821 }
3825 {
3830 }
3840 }
3841 }
3843 }
3846 /*
3847 ** before we can change a metadata block, we have to make sure it won't
3848 ** be written to disk while we are altering it. So, we must:
3849 ** clean it
3850 ** wait on it.
3851 **
3852 */
3855 {
3862 }
3867 }
3870 }
3873 {
3882 /* this check should always be run, to send old lists to disk */
3889 }
3890 }
3891 }
3893 /*
3894 ** long and ugly. If flush, will not return until all commit
3895 ** blocks and all real buffers in the trans are on disk.
3896 ** If no_async, won't return until all commit blocks are on disk.
3897 **
3898 ** keep reading, there are comments as you go along
3899 **
3900 ** If the journal is aborted, we just clean up. Things like flushing
3901 ** journal lists, etc just won't happen.
3902 */
3906 {
3928 /* protect flush_older_commits from doing mistakes if the
3929 transaction ID counter gets overflowed. */
3941 }
3947 }
3951 }
3953 /* check_journal_end locks the journal, and unlocks if it does not return 1
3954 ** it tells us if we should continue with the journal_end, or just return
3955 */
3961 }
3963 /* check_journal_end might set these, check again */
3966 }
3968 /*
3969 ** j must wait means we have to flush the log blocks, and the real blocks for
3970 ** this transaction
3971 */
3974 }
3975 #ifdef REISERFS_PREALLOCATE
3976 /* quota ops might need to nest, setup the journal_info pointer for them
3977 * and raise the refcount so that it is > 0. */
3981 * the transaction */
3984 #endif
3986 /* setup description block */
3987 d_bh =
3997 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4006 /* init this journal list */
4009 /* we lock the commit before doing anything because
4010 * we want to make sure nobody tries to run flush_commit_list until
4011 * the new transaction is fully setup, and we've already flushed the
4012 * ordered bh list
4013 */
4016 /* save the transaction id in case we need to commit it later */
4029 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4030 ** for each real block, add it to the journal list hash,
4031 ** copy into real block index array in the commit or desc block
4032 */
4040 }
4043 }
4048 }
4050 /* make sure the block we are trying to log is not a block
4051 of journal or reserved area */
4056 "Trying to log block %lu, "
4059 }
4072 }
4074 i--;
4075 }
4076 }
4082 /* special check in case all buffers in the journal were marked for not logging */
4085 /* we're about to dirty all the log blocks, mark the description block
4086 * dirty now too. Don't mark the commit block dirty until all the
4087 * others are on disk
4088 */
4091 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4097 /* copy all the real blocks into log area. dirty log blocks */
4102 tmp_bh =
4106 jindex) %
4116 jindex++;
4120 /* JDirty cleared sometime during transaction. don't log this one */
4122 "BAD, buffer in journal hash, "
4125 }
4132 }
4134 /* we are done with both the c_bh and d_bh, but
4135 ** c_bh must be written after all other commit blocks,
4136 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4137 */
4141 /* now it is safe to insert this transaction on the main list */
4146 /* reset journal values for the next transaction */
4157 /* check for trans_id overflow */
4167 // make sure reiserfs_add_jh sees the new current_jl before we
4168 // write out the tails
4171 /* tail conversion targets have to hit the disk before we end the
4172 * transaction. Otherwise a later transaction might repack the tail
4173 * before this transaction commits, leaving the data block unflushed and
4174 * clean, if we crash before the later transaction commits, the data block
4175 * is lost.
4176 */
4182 }
4186 /* honor the flush wishes from the caller, simple commits can
4187 ** be done outside the journal lock, they are done below
4188 **
4189 ** if we don't flush the commit list right now, we put it into
4190 ** the work queue so the people waiting on the async progress work
4191 ** queue don't wait for this proc to flush journal lists and such.
4192 */
4199 /* if the next transaction has any chance of wrapping, flush
4200 ** transactions that might get overwritten. If any journal lists are very
4201 ** old flush them as well.
4202 */
4203 first_jl:
4214 /* if we don't cross into the next transaction and we don't
4215 * wrap, there is no way we can overlap any later transactions
4216 * break now
4217 */
4219 }
4229 /* we don't overlap anything from out start to the end of the
4230 * log, and our wrapped portion doesn't overlap anything at
4231 * the start of the log. We can break
4232 */
4234 }
4235 }
4236 }
4245 }
4249 /* wake up any body waiting to join. */
4256 }
4257 out:
4261 /* Re-set th->t_super, so we can properly keep track of how many
4262 * persistent transactions there are. We need to do this so if this
4263 * call is part of a failed restart_transaction, we can free it later */
4267 }
4269 /* Send the file system read only and refuse new transactions */
4271 {
4282 #ifdef CONFIG_REISERFS_CHECK
4284 #endif
4285 }