| blob | 19fbc810e8e74f951a33e3f6611ca7ec2ee2f6e1 |
1 /*
2 ** Write ahead logging implementation copyright Chris Mason 2000
3 **
4 ** The background commits make this code very interelated, and
5 ** overly complex. I need to rethink things a bit....The major players:
6 **
7 ** journal_begin -- call with the number of blocks you expect to log.
8 ** If the current transaction is too
9 ** old, it will block until the current transaction is
10 ** finished, and then start a new one.
11 ** Usually, your transaction will get joined in with
12 ** previous ones for speed.
13 **
14 ** journal_join -- same as journal_begin, but won't block on the current
15 ** transaction regardless of age. Don't ever call
16 ** this. Ever. There are only two places it should be
17 ** called from, and they are both inside this file.
18 **
19 ** journal_mark_dirty -- adds blocks into this transaction. clears any flags
20 ** that might make them get sent to disk
21 ** and then marks them BH_JDirty. Puts the buffer head
22 ** into the current transaction hash.
23 **
24 ** journal_end -- if the current transaction is batchable, it does nothing
25 ** otherwise, it could do an async/synchronous commit, or
26 ** a full flush of all log and real blocks in the
27 ** transaction.
28 **
29 ** flush_old_commits -- if the current transaction is too old, it is ended and
30 ** commit blocks are sent to disk. Forces commit blocks
31 ** to disk for all backgrounded commits that have been
32 ** around too long.
33 ** -- Note, if you call this as an immediate flush from
34 ** from within kupdate, it will ignore the immediate flag
35 */
37 #include <linux/time.h>
38 #include <linux/semaphore.h>
39 #include <linux/vmalloc.h>
40 #include <linux/reiserfs_fs.h>
41 #include <linux/kernel.h>
42 #include <linux/errno.h>
43 #include <linux/fcntl.h>
44 #include <linux/stat.h>
45 #include <linux/string.h>
46 #include <linux/smp_lock.h>
47 #include <linux/buffer_head.h>
48 #include <linux/workqueue.h>
49 #include <linux/writeback.h>
50 #include <linux/blkdev.h>
51 #include <linux/backing-dev.h>
52 #include <linux/uaccess.h>
53 #include <linux/slab.h>
55 #include <asm/system.h>
57 /* gets a struct reiserfs_journal_list * from a list head */
58 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
59 j_list))
60 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
61 j_working_list))
63 /* the number of mounted filesystems. This is used to decide when to
64 ** start and kill the commit workqueue
65 */
71 structs at 4k */
74 /* cnode stat bits. Move these into reiserfs_fs.h */
77 #define BLOCK_FREED_HOLDER 3 /* this block was freed during this transaction, and can't be written */
80 #define BLOCK_DIRTIED 5
82 /* journal list state bits */
83 #define LIST_TOUCHED 1
84 #define LIST_DIRTY 2
87 /* flags for do_journal_end */
109 /* values for join in do_journal_begin_r */
114 };
121 {
125 }
127 /*
128 ** clears BH_Dirty and sticks the buffer on the clean list. Called because I can't allow refile_buffer to
129 ** make schedule happen after I've freed a block. Look at remove_from_transaction and journal_mark_freed for
130 ** more details.
131 */
133 {
137 }
139 }
142 {
146 }
150 {
157 }
162 }
166 }
169 {
175 repeat:
183 }
188 }
190 }
193 {
202 }
203 }
206 {
217 }
218 }
219 }
222 b_blocknr_t block,
224 {
230 }
233 }
237 {
246 }
247 }
248 }
250 /*
251 ** only call this on FS unmount.
252 */
255 {
264 }
266 }
269 {
281 }
284 }
286 /*
287 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
288 ** jb_array is the array to be filled in.
289 */
293 {
308 }
310 }
314 }
316 }
318 /*
319 ** find an available list bitmap. If you can't find one, flush a commit list
320 ** and try again
321 */
323 struct reiserfs_journal_list
325 {
340 }
343 }
344 }
347 }
350 }
352 /*
353 ** allocates a new chunk of X nodes, and links them all together as a list.
354 ** Uses the cnode->next and cnode->prev pointers
355 ** returns NULL on failure
356 */
358 {
363 }
367 }
374 }
377 }
379 /*
380 ** pulls a cnode off the free list, or returns NULL on failure
381 */
383 {
391 }
397 }
400 }
404 }
406 /*
407 ** returns a cnode to the free list
408 */
411 {
418 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
422 }
423 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
425 }
428 {
431 }
433 /* return a cnode with same dev, block number and size in table, or null if not found */
435 super_block
437 struct
438 reiserfs_journal_cnode
441 {
448 }
450 }
452 /*
453 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
454 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
455 ** being overwritten by a replay after crashing.
456 **
457 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
458 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
459 ** sure you never write the block without logging it.
460 **
461 ** next_zero_bit is a suggestion about the next block to try for find_forward.
462 ** when bl is rejected because it is set in a journal list bitmap, we search
463 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
464 ** through next_zero_bit for find_forward to try.
465 **
466 ** Just because we return something in next_zero_bit does not mean we won't
467 ** reject it on the next call to reiserfs_in_journal
468 **
469 */
473 {
483 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
484 ** if we crash before the transaction that freed it commits, this transaction won't
485 ** have committed either, and the block will never be written
486 */
494 data)) {
498 data),
502 }
503 }
504 }
507 /* is it in any old transactions? */
512 }
514 /* is it in the current transaction. This should never happen */
518 }
521 /* safe for reuse */
523 }
525 /* insert cn into table
526 */
529 {
537 }
539 }
541 /* lock the current transaction */
543 {
547 }
549 /* unlock the current transaction */
551 {
553 }
556 {
558 }
562 {
566 }
569 }
571 /*
572 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
573 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
574 ** transaction.
575 */
578 {
583 }
586 }
590 {
599 }
600 }
602 }
604 /*
605 * If page->mapping was null, we failed to truncate this page for
606 * some reason. Most likely because it was truncated after being
607 * logged via data=journal.
608 *
609 * This does a check to see if the buffer belongs to one of these
610 * lost pages before doing the final put_bh. If page->mapping was
611 * null, it tries to free buffers on the page, which should make the
612 * final page_cache_release drop the page from the lru.
613 */
615 {
626 }
627 }
630 {
637 }
640 else
645 }
648 {
651 else
655 }
658 {
668 }
671 {
678 }
681 {
688 }
692 {
701 }
702 }
704 #define CHUNK_SIZE 32
708 };
711 {
716 }
719 }
722 {
727 }
730 }
734 {
745 }
747 }
751 {
758 }
760 }
761 }
763 /*
764 * we want to free the jh when the buffer has been written
765 * and waited on
766 */
768 {
781 }
782 }
786 {
794 }
798 no_jh:
802 /* buffer must be locked for __add_jh, should be able to have
803 * two adds at the same time
804 */
808 }
814 }
817 }
820 {
822 }
824 {
826 }
828 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
833 {
851 }
859 }
860 /* in theory, dirty non-uptodate buffers should never get here,
861 * but the upper layer io error paths still have a few quirks.
862 * Handle them here as gracefully as we can
863 */
867 }
874 }
875 loop_next:
878 }
883 }
894 }
897 }
898 /* ugly interaction with invalidatepage here.
899 * reiserfs_invalidate_page will pin any buffer that has a valid
900 * journal head from an older transaction. If someone else sets
901 * our buffer dirty after we write it in the first loop, and
902 * then someone truncates the page away, nobody will ever write
903 * the buffer. We're safe if we write the page one last time
904 * after freeing the journal header.
905 */
910 }
913 }
916 }
920 {
929 find_first:
930 /*
931 * first we walk backwards to find the oldest uncommitted transation
932 */
943 }
945 /* if we didn't find any older uncommitted transactions, return now */
948 }
961 /* list we were called with is gone, return */
965 /* the one we just flushed is gone, this means all
966 * older lists are also gone, so first_jl is no longer
967 * valid either. Go back to the beginning.
968 */
972 }
973 }
979 }
980 }
982 }
985 {
993 }
996 }
998 /*
999 ** if this journal list still has commit blocks unflushed, send them to disk.
1000 **
1001 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
1002 ** Before the commit block can by written, every other log block must be safely on disk
1003 **
1004 */
1007 {
1009 b_blocknr_t bn;
1021 }
1025 /* before we can put our commit blocks on disk, we have to make sure everyone older than
1026 ** us is on disk too
1027 */
1034 /* list disappeared during flush_older_commits. return */
1036 }
1037 }
1039 /* make sure nobody is trying to flush this one at the same time */
1045 }
1048 /* this commit is done, exit */
1052 }
1055 }
1060 /*
1061 * We might sleep in numerous places inside
1062 * write_ordered_buffers. Relax the write lock.
1063 */
1070 }
1072 /*
1073 * for the description block and all the log blocks, submit any buffers
1074 * that haven't already reached the disk. Try to write at least 256
1075 * log blocks. later on, we will only wait on blocks that correspond
1076 * to this transaction, but while we're unplugging we might as well
1077 * get a chunk of data on there.
1078 */
1092 }
1094 }
1095 }
1098 /* We're skipping the commit if there's an error */
1102 /* wait on everything written so far before writing the commit
1103 * if we are in barrier mode, send the commit down now
1104 */
1114 }
1115 }
1124 // since we're using ll_rw_blk above, it might have skipped over
1125 // a locked buffer. Double check here
1126 //
1127 /* redundant, sync_dirty_buffer() checks */
1132 }
1134 #ifdef CONFIG_REISERFS_CHECK
1137 #endif
1139 }
1143 }
1148 /* If there was a write error in the journal - we can't commit
1149 * this transaction - it will be invalid and, if successful,
1150 * will just end up propagating the write error out to
1151 * the file system. */
1159 }
1164 }
1168 /* If there was a write error in the journal - we can't commit this
1169 * transaction - it will be invalid and, if successful, will just end
1170 * up propagating the write error out to the filesystem. */
1172 #ifdef CONFIG_REISERFS_CHECK
1174 #endif
1176 }
1182 }
1185 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1190 /* mark the metadata dirty */
1197 }
1199 put_jl:
1204 __func__);
1207 }
1209 /*
1210 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1211 ** returns NULL if it can't find anything
1212 */
1214 reiserfs_journal_cnode
1216 {
1224 }
1226 }
1228 }
1231 {
1241 }
1243 }
1250 /*
1251 ** once all the real blocks have been flushed, it is safe to remove them from the
1252 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1253 ** block to be reallocated for data blocks if it had been deleted.
1254 */
1258 {
1263 /* which is better, to lock once around the whole loop, or
1264 ** to lock for each call to remove_journal_hash?
1265 */
1273 }
1277 }
1281 }
1283 }
1285 /*
1286 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1287 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1288 ** releasing blocks in this transaction for reuse as data blocks.
1289 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1290 **
1291 */
1295 {
1308 #ifdef CONFIG_REISERFS_CHECK
1311 #endif
1313 }
1314 }
1318 b_data);
1331 }
1337 sync:
1342 }
1347 }
1348 }
1350 }
1355 {
1357 }
1359 /*
1360 ** flush any and all journal lists older than you are
1361 ** can only be called from flush_journal_list
1362 */
1365 {
1371 /* we know we are the only ones flushing things, no extra race
1372 * protection is required.
1373 */
1374 restart:
1376 /* Did we wrap? */
1382 /* do not flush all */
1385 /* other_jl is now deleted from the list */
1387 }
1389 }
1393 {
1398 }
1399 }
1401 /* flush a journal list, both commit and real blocks
1402 **
1403 ** always set flushall to 1, unless you are calling from inside
1404 ** flush_journal_list
1405 **
1406 ** IMPORTANT. This can only be called while there are no journal writers,
1407 ** and the journal is locked. That means it can only be called from
1408 ** do_journal_end, or by journal_release
1409 */
1412 {
1428 }
1431 /* if flushall == 0, the lock is already held */
1436 }
1443 }
1447 /* if all the work is already done, get out of here */
1451 }
1453 /* start by putting the commit list on disk. This will also flush
1454 ** the commit lists of any olders transactions
1455 */
1462 /* are we done now? */
1466 }
1468 /* loop through each cnode, see if we need to write it,
1469 ** or wait on a more recent transaction, or just ignore it
1470 */
1474 }
1480 /* blocknr of 0 is no longer in the hash, ignore it */
1483 }
1485 /* This transaction failed commit. Don't write out to the disk */
1490 /* the order is important here. We check pjl to make sure we
1491 ** don't clear BH_JDirty_wait if we aren't the one writing this
1492 ** block to disk
1493 */
1497 /* we do this to make sure nobody releases the buffer while
1498 ** we are working with it
1499 */
1507 /* everything with !pjl && jwait should be writable */
1509 }
1510 }
1512 /* if someone has this block in a newer transaction, just make
1513 ** sure they are committed, and don't try writing it to disk
1514 */
1519 }
1521 /* bh == NULL when the block got to disk on its own, OR,
1522 ** the block got freed in a future transaction
1523 */
1526 }
1528 /* this should never happen. kupdate_one_transaction has this list
1529 ** locked while it works, so we should never see a buffer here that
1530 ** is not marked JDirty_wait
1531 */
1534 "BAD! buffer %llu %cdirty %cjwait, "
1539 }
1541 /* we inc again because saved_bh gets decremented at free_cnode */
1548 else
1550 count++;
1556 }
1557 free_cnode:
1561 /* we incremented this to keep others from taking the buffer head away */
1566 }
1567 }
1568 }
1576 }
1585 }
1587 #ifdef CONFIG_REISERFS_CHECK
1590 #endif
1592 }
1593 /* note, we must clear the JDirty_wait bit after the up to date
1594 ** check, otherwise we race against our flushpage routine
1595 */
1599 /* drop one ref for us */
1601 /* drop one ref for journal_mark_dirty */
1603 }
1605 }
1606 }
1611 __func__);
1612 flush_older_and_return:
1614 /* before we can update the journal header block, we _must_ flush all
1615 ** real blocks from all older transactions to disk. This is because
1616 ** once the header block is updated, this transaction will not be
1617 ** replayed after a crash
1618 */
1621 }
1624 /* before we can remove everything from the hash tables for this
1625 ** transaction, we must make sure it can never be replayed
1626 **
1627 ** since we are only called from do_journal_end, we know for sure there
1628 ** are no allocations going on while we are flushing journal lists. So,
1629 ** we only need to update the journal header block for the last list
1630 ** being flushed
1631 */
1633 err =
1641 __func__);
1642 }
1652 }
1655 /* not strictly required since we are freeing the list, but it should
1656 * help find code using dead lists later on
1657 */
1670 }
1674 {
1682 /* if the blocknr == 0, this has been cleared from the hash,
1683 ** skip it
1684 */
1687 }
1690 next:
1693 }
1695 }
1700 {
1708 }
1712 /* if the blocknr == 0, this has been cleared from the hash,
1713 ** skip it
1714 */
1717 }
1720 /* we can race against journal_mark_freed when we try
1721 * to lock_buffer(cn->bh), so we have to inc the buffer
1722 * count, and recheck things after locking
1723 */
1732 ret++;
1734 /* note, cn->bh might be null now */
1736 }
1738 }
1739 next:
1742 }
1744 }
1746 /* used by flush_commit_list */
1749 {
1757 /* look for a more recent transaction that logged this
1758 ** buffer. Only the most recent transaction with a buffer in
1759 ** it is allowed to send that buffer to disk
1760 */
1765 /* if the buffer is prepared, it will either be logged
1766 * or restored. If restored, we need to make sure
1767 * it actually gets marked dirty
1768 */
1775 }
1776 }
1778 }
1780 }
1787 {
1800 }
1802 /* we've got j_flush_mutex held, nobody is going to delete any
1803 * of these lists out from underneath us
1804 */
1813 }
1818 transactions_flushed++;
1822 /* did we wrap? */
1825 }
1828 /* don't bother with older transactions */
1831 }
1834 }
1836 done:
1839 }
1841 /* for o_sync and fsync heavy applications, they tend to use
1842 ** all the journa list slots with tiny transactions. These
1843 ** trigger lots and lots of calls to update the header block, which
1844 ** adds seeks and slows things down.
1845 **
1846 ** This function tries to clear out a large chunk of the journal lists
1847 ** at once, which makes everything faster since only the newest journal
1848 ** list updates the header block
1849 */
1852 {
1865 /* in data logging mode, try harder to flush a lot of blocks */
1868 /* flush for 256 transactions or limit blocks, whichever comes first */
1873 }
1877 }
1883 }
1884 /* try to find a group of blocks we can flush across all the
1885 ** transactions, but only bother if we've actually spanned
1886 ** across multiple lists
1887 */
1890 }
1893 }
1895 /*
1896 ** removes any nodes in table with name block and dev as bh.
1897 ** only touchs the hnext and hprev pointers.
1898 */
1903 {
1910 }
1918 }
1923 }
1931 }
1933 }
1934 }
1937 {
1947 }
1948 /* j_header_bh is on the journal dev, make sure not to release the journal
1949 * dev until we brelse j_header_bh
1950 */
1953 }
1955 /*
1956 ** call on unmount. Only set error to 1 if you haven't made your way out
1957 ** of read_super() yet. Any other caller must keep error at 0.
1958 */
1961 {
1966 /* we only want to flush out transactions if we were called with error == 0
1967 */
1969 /* end the current trans */
1973 /* make sure something gets logged to force our way into the flush code */
1982 }
1983 }
1985 /* this also catches errors during the do_journal_end above */
1995 }
1996 }
1998 reiserfs_mounted_fs_count--;
1999 /* wait for all commits to finish */
2002 /*
2003 * We must release the write lock here because
2004 * the workqueue job (flush_async_commit) needs this lock
2005 */
2012 }
2019 }
2021 /*
2022 ** call on unmount. flush all journal trans, release all alloc'd ram
2023 */
2026 {
2028 }
2030 /*
2031 ** only call from an error condition inside reiserfs_read_super!
2032 */
2035 {
2037 }
2039 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
2043 {
2049 }
2051 }
2053 /* returns 0 if it did not find a description block
2054 ** returns -1 if it found a corrupt commit block
2055 ** returns 1 if both desc and commit were valid
2056 */
2061 {
2076 "journal-986: transaction "
2077 "is valid returning because trans_id %d is greater than "
2082 }
2086 "journal-1087: transaction "
2087 "is valid returning because mount_id %d is less than "
2092 }
2095 "Bad transaction length %d "
2099 }
2102 /* ok, we have a journal description block, lets see if the transaction was valid */
2103 c_bh =
2113 "journal_transaction_is_valid, commit offset %ld had bad "
2124 "journal-1004: "
2125 "transaction_is_valid setting oldest invalid trans_id "
2128 }
2130 }
2133 "journal-1006: found valid "
2142 }
2143 }
2146 {
2150 }
2151 }
2153 /*
2154 ** given the start, and values for the oldest acceptable transactions,
2155 ** this either reads in a replays a transaction, or returns because the transaction
2156 ** is invalid, or too old.
2157 */
2163 {
2188 cur_dblock -
2192 }
2195 "journal_read_trans skipping because %d is != "
2197 newest_mount_id);
2200 }
2207 }
2211 "journal_read_transaction, "
2220 }
2228 }
2231 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2244 }
2245 /* get all the buffer heads */
2262 }
2265 "REPLAY FAILURE fsck required! "
2266 "Block to replay is outside of "
2269 }
2270 /* make sure we don't try to replay onto log or reserved area */
2274 "REPLAY FAILURE fsck required! "
2276 abort_replay:
2284 }
2285 }
2286 /* read in the log blocks, memcpy to the corresponding real block */
2296 "REPLAY FAILURE fsck required! "
2306 }
2311 }
2312 /* flush out the real blocks */
2316 }
2321 "REPLAY FAILURE, fsck required! "
2330 }
2332 }
2333 cur_dblock =
2341 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2345 /* check for trans_id overflow */
2353 }
2355 /* This function reads blocks starting from block and to max_block of bufsize
2356 size (but no more than BUFNR blocks at a time). This proved to improve
2357 mounting speed on self-rebuilding raid5 arrays at least.
2358 Right now it is only used from journal code. But later we might use it
2359 from other places.
2360 Note: Do not use journal_getblk/sb_getblk functions here! */
2363 b_blocknr_t max_block)
2364 {
2376 }
2386 }
2396 }
2398 /*
2399 ** read and replay the log
2400 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2401 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2402 **
2403 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2404 **
2405 ** On exit, it sets things up so the first transaction will work correctly.
2406 */
2408 {
2430 /* step 1, read in the journal header block. Check the transaction it says
2431 ** is the first unflushed, and if that transaction is not valid,
2432 ** replay is done
2433 */
2439 }
2444 oldest_start =
2450 "journal-1153: found in "
2451 "header: first_unflushed_offset %d, last_flushed_trans_id "
2456 /* now, we try to read the first unflushed offset. If it is not valid,
2457 ** there is nothing more we can do, and it makes no sense to read
2458 ** through the whole log.
2459 */
2460 d_bh =
2467 }
2470 }
2472 /* ok, there are transactions that need to be replayed. start with the first log block, find
2473 ** all the valid transactions, and pick out the oldest.
2474 */
2479 /* Note that it is required for blocksize of primary fs device and journal
2480 device to be the same */
2481 d_bh =
2486 ret =
2497 "journal-1179: Setting "
2499 oldest_start -
2500 SB_ONDISK_JOURNAL_1st_BLOCK
2503 /* one we just read was older */
2507 "journal-1180: Resetting "
2509 oldest_start -
2510 SB_ONDISK_JOURNAL_1st_BLOCK
2512 }
2516 "journal-1299: Setting "
2519 }
2522 cur_dblock++;
2523 }
2525 }
2527 start_log_replay:
2531 "journal-1206: Starting replay "
2534 oldest_trans_id);
2536 }
2539 ret =
2546 }
2547 cur_dblock =
2549 replay_count++;
2552 }
2557 }
2558 /* j_start does not get set correctly if we don't replay any transactions.
2559 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2560 ** copy the trans_id from the header
2561 */
2566 /* check for trans_id overflow */
2574 }
2582 }
2586 /* replay failed, caller must call free_journal_ram and abort
2587 ** the mount
2588 */
2590 }
2592 }
2595 {
2607 }
2610 {
2612 }
2616 {
2626 }
2631 }
2633 }
2638 {
2640 dev_t jdev;
2653 /* there is no "jdev" option and journal is on separate device */
2669 }
2672 }
2675 }
2687 }
2694 }
2696 /**
2697 * When creating/tuning a file system user can assign some
2698 * journal params within boundaries which depend on the ratio
2699 * blocksize/standard_blocksize.
2700 *
2701 * For blocks >= standard_blocksize transaction size should
2702 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2703 * then JOURNAL_TRANS_MAX_DEFAULT.
2704 *
2705 * For blocks < standard_blocksize these boundaries should be
2706 * decreased proportionally.
2707 */
2708 #define REISERFS_STANDARD_BLKSIZE (4096)
2712 {
2714 /* Non-default journal params.
2715 Do sanity check for them. */
2723 JOURNAL_MIN_RATIO) {
2725 "bad transaction max size (%u). "
2728 }
2732 "bad transaction max batch (%u). "
2735 }
2737 /* Default journal params.
2738 The file system was created by old version
2739 of mkreiserfs, so some fields contain zeros,
2740 and we need to advise proper values for them */
2745 }
2749 }
2751 }
2753 /*
2754 ** must be called once on fs mount. calls journal_read for you
2755 */
2758 {
2768 /*
2769 * Unlock here to avoid various RECLAIM-FS-ON <-> IN-RECLAIM-FS
2770 * dependency inversion warnings.
2771 */
2779 }
2794 /* reserved for journal area support */
2796 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2800 REISERFS_DISK_OFFSET_IN_BYTES /
2803 /* Sanity check to see is the standard journal fitting withing first bitmap
2804 (actual for small blocksizes) */
2809 "journal does not fit for area addressed "
2810 "by first of bitmap blocks. It starts at "
2816 }
2818 /*
2819 * We need to unlock here to avoid creating the following
2820 * dependency:
2821 * reiserfs_lock -> sysfs_mutex
2822 * Because the reiserfs mmap path creates the following dependency:
2823 * mm->mmap -> reiserfs_lock, hence we have
2824 * mm->mmap -> reiserfs_lock ->sysfs_mutex
2825 * This would ends up in a circular dependency with sysfs readdir path
2826 * which does sysfs_mutex -> mm->mmap_sem
2827 * This is fine because the reiserfs lock is useless in mount path,
2828 * at least until we call journal_begin. We keep it for paranoid
2829 * reasons.
2830 */
2837 }
2842 /* read journal header */
2850 }
2853 /* make sure that journal matches to the super block */
2858 "journal header magic %x (device %s) does "
2865 }
2880 }
2883 "journal first block %u, max trans len %u, max batch %u, "
2930 "allocation failed (%ld bytes). Journal is "
2931 "too large for available memory. Usually "
2935 }
2944 }
2949 }
2951 reiserfs_mounted_fs_count++;
2956 }
2961 free_and_return:
2964 }
2966 /*
2967 ** test for a polite end of the current transaction. Used by file_write, and should
2968 ** be used by delete to make sure they don't write more than can fit inside a single
2969 ** transaction
2970 */
2973 {
2976 /* cannot restart while nested */
2986 }
2987 /* protected by the BKL here */
2991 }
2993 /* this must be called inside a transaction, and requires the
2994 ** kernel_lock to be held
2995 */
2997 {
3003 }
3005 /* this must be called without a transaction started, and does not
3006 ** require BKL
3007 */
3009 {
3013 }
3015 /* this must be called without a transaction started, and does not
3016 ** require BKL
3017 */
3019 {
3023 }
3026 {
3027 wait_queue_t wait;
3031 /*
3032 * we don't want to use wait_event here because
3033 * we only want to wait once.
3034 */
3042 }
3045 }
3048 {
3052 }
3055 {
3067 }
3073 }
3074 }
3076 /* join == true if you must join an existing transaction.
3077 ** join == false if you can deal with waiting for others to finish
3078 **
3079 ** this will block until the transaction is joinable. send the number of blocks you
3080 ** expect to use in nblocks.
3081 */
3085 {
3097 /* set here for journal_join */
3101 relock:
3107 }
3117 }
3120 /* if there is no room in the journal OR
3121 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3122 ** we don't sleep if there aren't other writers
3123 */
3126 (!join
3143 sched_count++;
3146 }
3147 }
3148 /* don't mess with joining the transaction if all we have to do is
3149 * wait for someone else to do a commit
3150 */
3155 }
3157 }
3162 /* someone might have ended the transaction while we joined */
3167 }
3174 }
3175 /* we are the first writer, set trans_id */
3178 }
3189 out_fail:
3191 /* Re-set th->t_super, so we can properly keep track of how many
3192 * persistent transactions there are. We need to do this so if this
3193 * call is part of a failed restart_transaction, we can free it later */
3196 }
3199 super_block
3202 {
3206 /* if we're nesting into an existing transaction. It will be
3207 ** persistent on its own
3208 */
3215 }
3223 }
3227 }
3230 {
3235 else
3240 }
3242 }
3246 {
3249 /* this keeps do_journal_end from NULLing out the current->journal_info
3250 ** pointer
3251 */
3255 }
3259 {
3262 /* this keeps do_journal_end from NULLing out the current->journal_info
3263 ** pointer
3264 */
3268 }
3272 {
3278 /* we are nesting into the current transaction */
3285 "BAD: refcount <= 1, but "
3289 /* we've ended up with a handle from a different filesystem.
3290 ** save it and restore on journal_end. This should never
3291 ** really happen...
3292 */
3297 }
3300 }
3304 /* I guess this boils down to being the reciprocal of clm-2100 above.
3305 * If do_journal_begin_r fails, we need to put it back, since journal_end
3306 * won't be called to do it. */
3309 else
3313 }
3315 /*
3316 ** puts bh into the current transaction. If it was already there, reorders removes the
3317 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3318 **
3319 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3320 ** transaction is committed.
3321 **
3322 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3323 */
3326 {
3338 }
3344 /* already in this transaction, we are done */
3348 }
3350 /* this must be turned into a panic instead of a warning. We can't allow
3351 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3352 ** could get to disk too early. NOT GOOD.
3353 */
3356 "buffer %llu bad state "
3363 }
3370 }
3371 /* this error means I've screwed up, and we've overflowed the transaction.
3372 ** Nothing can be done here, except make the FS readonly or panic.
3373 */
3378 }
3384 }
3388 }
3392 /* now put this guy on the end */
3397 }
3402 }
3413 }
3414 }
3424 }
3426 }
3430 {
3438 }
3445 /* we aren't allowed to close a nested transaction on a different
3446 ** filesystem from the one in the task struct
3447 */
3453 }
3457 }
3458 }
3460 /* removes from the current transaction, relsing and descrementing any counters.
3461 ** also files the removed buffer directly onto the clean list
3462 **
3463 ** called by journal_mark_freed when a block has been deleted
3464 **
3465 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3466 */
3469 {
3478 }
3482 }
3485 }
3488 }
3491 }
3505 }
3507 }
3512 }
3514 /*
3515 ** for any cnode in a journal list, it can only be dirtied of all the
3516 ** transactions that include it are committed to disk.
3517 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3518 ** and 0 if you aren't
3519 **
3520 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3521 ** blocks for a given transaction on disk
3522 **
3523 */
3525 {
3531 /* first test hprev. These are all newer than cn, so any node here
3532 ** with the same block number and dev means this node can't be sent
3533 ** to disk right now.
3534 */
3539 }
3541 }
3542 /* then test hnext. These are all older than cn. As long as they
3543 ** are committed to the log, it is safe to write cn to disk
3544 */
3551 }
3553 }
3555 }
3557 /* syncs the commit blocks, but does not force the real buffers to disk
3558 ** will wait until the current transaction is done/committed before returning
3559 */
3562 {
3566 /* you can sync while nested, very, very bad */
3572 }
3574 }
3576 /*
3577 ** writeback the pending async commits to disk
3578 */
3580 {
3589 /* last entry is the youngest, commit it and you get everything */
3593 }
3595 }
3597 /*
3598 ** flushes any old transactions to disk
3599 ** ends the current transaction if it is too old
3600 */
3602 {
3608 /* safety check so we don't flush while we are replaying the log during
3609 * mount
3610 */
3613 }
3615 /* check the current transaction. If there are no writers, and it is
3616 * too old, finish it, and force the commit blocks to disk
3617 */
3629 /* we're only being called from kreiserfsd, it makes no sense to do
3630 ** an async commit so that kreiserfsd can do it later
3631 */
3633 }
3634 }
3636 }
3638 /*
3639 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3640 **
3641 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3642 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3643 ** flushes the commit list and returns 0.
3644 **
3645 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3646 **
3647 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3648 */
3652 {
3667 }
3670 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3672 }
3674 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3675 ** will be dealt with by next transaction that actually writes something, but should be taken
3676 ** care of in this trans
3677 */
3680 /* if wcount > 0, and we are called to with flush or commit_now,
3681 ** we wait on j_join_wait. We will wake up when the last writer has
3682 ** finished the transaction, and started it on its way to the disk.
3683 ** Then, we flush the commit or journal list, and just return 0
3684 ** because the rest of journal end was already done for this transaction.
3685 */
3697 }
3700 /* sleep while the current transaction is still j_jlocked */
3709 }
3711 }
3712 }
3719 }
3721 }
3724 }
3726 /* deal with old transactions where we are the last writers */
3731 }
3732 /* don't batch when someone is waiting on j_join_wait */
3733 /* don't batch when syncing the commit or flushing the whole trans */
3741 }
3747 }
3749 }
3751 /*
3752 ** Does all the work that makes deleting blocks safe.
3753 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3754 **
3755 ** otherwise:
3756 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3757 ** before this transaction has finished.
3758 **
3759 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3760 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3761 ** the block can't be reallocated yet.
3762 **
3763 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3764 */
3767 {
3779 }
3780 /* if it is journal new, we just remove it from this transaction */
3787 /* set the bit for this block in the journal bitmap for this transaction */
3792 }
3795 /* Note, the entire while loop is not allowed to schedule. */
3800 }
3803 /* find all older transactions with this block, make sure they don't try to write it out */
3805 blocknr);
3811 /* remove_from_transaction will brelse the buffer if it was
3812 ** in the current trans
3813 */
3815 bh);
3818 bh);
3826 }
3827 }
3831 j_nonzerolen));
3832 }
3834 }
3835 }
3837 }
3838 }
3843 }
3846 {
3850 }
3852 /*
3853 * returns -1 on error, 0 if no commits/barriers were done and 1
3854 * if a transaction was actually committed and the barrier was done
3855 */
3858 {
3864 /* is it from the current transaction, or from an unknown transaction? */
3867 /* try to let other writers come in and grow this transaction */
3871 }
3877 /* someone might have ended this transaction while we joined */
3884 }
3891 /* this gets tricky, we have to make sure the journal list in
3892 * the inode still exists. We know the list is still around
3893 * if we've got a larger transaction id than the oldest list
3894 */
3895 flush_commit_only:
3897 /*
3898 * we only set ret to 1 when we know for sure
3899 * the barrier hasn't been started yet on the commit
3900 * block.
3901 */
3907 }
3908 }
3909 /* otherwise the list is gone, and long since committed */
3911 }
3914 {
3918 /* for the whole inode, assume unset id means it was
3919 * changed in the current transaction. More conservative
3920 */
3924 /* jl will be updated in __commit_trans_jl */
3925 }
3928 }
3932 {
3937 }
3947 }
3948 }
3950 }
3953 /*
3954 ** before we can change a metadata block, we have to make sure it won't
3955 ** be written to disk while we are altering it. So, we must:
3956 ** clean it
3957 ** wait on it.
3958 **
3959 */
3962 {
3969 }
3974 }
3977 }
3980 {
3989 /* this check should always be run, to send old lists to disk */
3996 }
3997 }
3998 }
4000 /*
4001 ** long and ugly. If flush, will not return until all commit
4002 ** blocks and all real buffers in the trans are on disk.
4003 ** If no_async, won't return until all commit blocks are on disk.
4004 **
4005 ** keep reading, there are comments as you go along
4006 **
4007 ** If the journal is aborted, we just clean up. Things like flushing
4008 ** journal lists, etc just won't happen.
4009 */
4013 {
4035 /* protect flush_older_commits from doing mistakes if the
4036 transaction ID counter gets overflowed. */
4049 }
4055 }
4059 }
4061 /* check_journal_end locks the journal, and unlocks if it does not return 1
4062 ** it tells us if we should continue with the journal_end, or just return
4063 */
4069 }
4071 /* check_journal_end might set these, check again */
4074 }
4076 /*
4077 ** j must wait means we have to flush the log blocks, and the real blocks for
4078 ** this transaction
4079 */
4082 }
4083 #ifdef REISERFS_PREALLOCATE
4084 /* quota ops might need to nest, setup the journal_info pointer for them
4085 * and raise the refcount so that it is > 0. */
4089 * the transaction */
4092 #endif
4094 /* setup description block */
4095 d_bh =
4105 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4114 /* init this journal list */
4117 /* we lock the commit before doing anything because
4118 * we want to make sure nobody tries to run flush_commit_list until
4119 * the new transaction is fully setup, and we've already flushed the
4120 * ordered bh list
4121 */
4124 /* save the transaction id in case we need to commit it later */
4137 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4138 ** for each real block, add it to the journal list hash,
4139 ** copy into real block index array in the commit or desc block
4140 */
4148 }
4151 }
4156 }
4158 /* make sure the block we are trying to log is not a block
4159 of journal or reserved area */
4164 "Trying to log block %lu, "
4167 }
4180 }
4182 i--;
4183 }
4184 }
4190 /* special check in case all buffers in the journal were marked for not logging */
4193 /* we're about to dirty all the log blocks, mark the description block
4194 * dirty now too. Don't mark the commit block dirty until all the
4195 * others are on disk
4196 */
4199 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4205 /* copy all the real blocks into log area. dirty log blocks */
4210 tmp_bh =
4214 jindex) %
4224 jindex++;
4228 /* JDirty cleared sometime during transaction. don't log this one */
4230 "BAD, buffer in journal hash, "
4233 }
4240 }
4242 /* we are done with both the c_bh and d_bh, but
4243 ** c_bh must be written after all other commit blocks,
4244 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4245 */
4249 /* now it is safe to insert this transaction on the main list */
4254 /* reset journal values for the next transaction */
4265 /* check for trans_id overflow */
4275 // make sure reiserfs_add_jh sees the new current_jl before we
4276 // write out the tails
4279 /* tail conversion targets have to hit the disk before we end the
4280 * transaction. Otherwise a later transaction might repack the tail
4281 * before this transaction commits, leaving the data block unflushed and
4282 * clean, if we crash before the later transaction commits, the data block
4283 * is lost.
4284 */
4290 }
4294 /* honor the flush wishes from the caller, simple commits can
4295 ** be done outside the journal lock, they are done below
4296 **
4297 ** if we don't flush the commit list right now, we put it into
4298 ** the work queue so the people waiting on the async progress work
4299 ** queue don't wait for this proc to flush journal lists and such.
4300 */
4307 /* if the next transaction has any chance of wrapping, flush
4308 ** transactions that might get overwritten. If any journal lists are very
4309 ** old flush them as well.
4310 */
4311 first_jl:
4322 /* if we don't cross into the next transaction and we don't
4323 * wrap, there is no way we can overlap any later transactions
4324 * break now
4325 */
4327 }
4337 /* we don't overlap anything from out start to the end of the
4338 * log, and our wrapped portion doesn't overlap anything at
4339 * the start of the log. We can break
4340 */
4342 }
4343 }
4344 }
4353 }
4357 /* wake up any body waiting to join. */
4364 }
4365 out:
4369 /* Re-set th->t_super, so we can properly keep track of how many
4370 * persistent transactions there are. We need to do this so if this
4371 * call is part of a failed restart_transaction, we can free it later */
4375 }
4377 /* Send the file system read only and refuse new transactions */
4379 {
4390 #ifdef CONFIG_REISERFS_CHECK
4392 #endif
4393 }