| blob | b5fe0aa033e7148a877d1bb7bf170f068c6f6f40 |
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>
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 }
141 {
145 }
149 {
156 }
161 }
165 }
168 {
174 repeat:
182 }
187 }
189 }
192 {
201 }
202 }
205 {
216 }
217 }
218 }
221 b_blocknr_t block,
223 {
229 }
232 }
236 {
245 }
246 }
247 }
249 /*
250 ** only call this on FS unmount.
251 */
254 {
263 }
265 }
268 {
280 }
283 }
285 /*
286 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
287 ** jb_array is the array to be filled in.
288 */
292 {
307 }
309 }
313 }
315 }
317 /*
318 ** find an available list bitmap. If you can't find one, flush a commit list
319 ** and try again
320 */
322 struct reiserfs_journal_list
324 {
339 }
342 }
343 }
346 }
349 }
351 /*
352 ** allocates a new chunk of X nodes, and links them all together as a list.
353 ** Uses the cnode->next and cnode->prev pointers
354 ** returns NULL on failure
355 */
357 {
362 }
366 }
373 }
376 }
378 /*
379 ** pulls a cnode off the free list, or returns NULL on failure
380 */
382 {
390 }
396 }
399 }
403 }
405 /*
406 ** returns a cnode to the free list
407 */
410 {
417 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
421 }
422 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
424 }
427 {
430 }
432 /* utility function to force a BUG if it is called without the big
433 ** kernel lock held. caller is the string printed just before calling BUG()
434 */
436 {
437 #ifdef CONFIG_SMP
441 }
442 #else
443 ;
444 #endif
445 }
447 /* return a cnode with same dev, block number and size in table, or null if not found */
449 super_block
451 struct
452 reiserfs_journal_cnode
455 {
462 }
464 }
466 /*
467 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
468 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
469 ** being overwritten by a replay after crashing.
470 **
471 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
472 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
473 ** sure you never write the block without logging it.
474 **
475 ** next_zero_bit is a suggestion about the next block to try for find_forward.
476 ** when bl is rejected because it is set in a journal list bitmap, we search
477 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
478 ** through next_zero_bit for find_forward to try.
479 **
480 ** Just because we return something in next_zero_bit does not mean we won't
481 ** reject it on the next call to reiserfs_in_journal
482 **
483 */
487 {
497 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
498 ** if we crash before the transaction that freed it commits, this transaction won't
499 ** have committed either, and the block will never be written
500 */
508 data)) {
512 data),
516 }
517 }
518 }
521 /* is it in any old transactions? */
526 }
528 /* is it in the current transaction. This should never happen */
532 }
535 /* safe for reuse */
537 }
539 /* insert cn into table
540 */
543 {
551 }
553 }
555 /* lock the current transaction */
557 {
560 }
562 /* unlock the current transaction */
564 {
566 }
569 {
571 }
575 {
579 }
582 }
584 /*
585 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
586 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
587 ** transaction.
588 */
591 {
596 }
599 }
603 {
612 }
613 }
615 }
617 /*
618 * If page->mapping was null, we failed to truncate this page for
619 * some reason. Most likely because it was truncated after being
620 * logged via data=journal.
621 *
622 * This does a check to see if the buffer belongs to one of these
623 * lost pages before doing the final put_bh. If page->mapping was
624 * null, it tries to free buffers on the page, which should make the
625 * final page_cache_release drop the page from the lru.
626 */
628 {
639 }
640 }
643 {
650 }
653 else
658 }
661 {
664 else
668 }
671 {
681 }
684 {
691 }
694 {
701 }
705 {
712 }
713 }
715 #define CHUNK_SIZE 32
719 };
722 {
727 }
730 }
733 {
738 }
741 }
745 {
756 }
758 }
762 {
769 }
771 }
772 }
774 /*
775 * we want to free the jh when the buffer has been written
776 * and waited on
777 */
779 {
792 }
793 }
797 {
805 }
809 no_jh:
813 /* buffer must be locked for __add_jh, should be able to have
814 * two adds at the same time
815 */
819 }
825 }
828 }
831 {
833 }
835 {
837 }
839 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
844 {
862 }
870 }
871 /* in theory, dirty non-uptodate buffers should never get here,
872 * but the upper layer io error paths still have a few quirks.
873 * Handle them here as gracefully as we can
874 */
878 }
885 }
886 loop_next:
889 }
894 }
905 }
908 }
909 /* ugly interaction with invalidatepage here.
910 * reiserfs_invalidate_page will pin any buffer that has a valid
911 * journal head from an older transaction. If someone else sets
912 * our buffer dirty after we write it in the first loop, and
913 * then someone truncates the page away, nobody will ever write
914 * the buffer. We're safe if we write the page one last time
915 * after freeing the journal header.
916 */
921 }
924 }
927 }
931 {
940 find_first:
941 /*
942 * first we walk backwards to find the oldest uncommitted transation
943 */
954 }
956 /* if we didn't find any older uncommitted transactions, return now */
959 }
972 /* list we were called with is gone, return */
976 /* the one we just flushed is gone, this means all
977 * older lists are also gone, so first_jl is no longer
978 * valid either. Go back to the beginning.
979 */
983 }
984 }
990 }
991 }
993 }
996 {
1002 }
1004 /*
1005 ** if this journal list still has commit blocks unflushed, send them to disk.
1006 **
1007 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
1008 ** Before the commit block can by written, every other log block must be safely on disk
1009 **
1010 */
1013 {
1015 b_blocknr_t bn;
1027 }
1031 /* before we can put our commit blocks on disk, we have to make sure everyone older than
1032 ** us is on disk too
1033 */
1040 /* list disappeared during flush_older_commits. return */
1042 }
1043 }
1045 /* make sure nobody is trying to flush this one at the same time */
1050 }
1053 /* this commit is done, exit */
1057 }
1060 }
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 */
1091 }
1092 }
1095 /* We're skipping the commit if there's an error */
1099 /* wait on everything written so far before writing the commit
1100 * if we are in barrier mode, send the commit down now
1101 */
1111 }
1112 }
1118 // since we're using ll_rw_blk above, it might have skipped over
1119 // a locked buffer. Double check here
1120 //
1124 #ifdef CONFIG_REISERFS_CHECK
1127 #endif
1129 }
1133 }
1138 /* If there was a write error in the journal - we can't commit
1139 * this transaction - it will be invalid and, if successful,
1140 * will just end up propagating the write error out to
1141 * the file system. */
1147 }
1153 /* If there was a write error in the journal - we can't commit this
1154 * transaction - it will be invalid and, if successful, will just end
1155 * up propagating the write error out to the filesystem. */
1157 #ifdef CONFIG_REISERFS_CHECK
1159 #endif
1161 }
1167 }
1170 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1175 /* mark the metadata dirty */
1182 }
1184 put_jl:
1189 __func__);
1192 }
1194 /*
1195 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1196 ** returns NULL if it can't find anything
1197 */
1199 reiserfs_journal_cnode
1201 {
1209 }
1211 }
1213 }
1216 {
1226 }
1228 }
1235 /*
1236 ** once all the real blocks have been flushed, it is safe to remove them from the
1237 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1238 ** block to be reallocated for data blocks if it had been deleted.
1239 */
1243 {
1248 /* which is better, to lock once around the whole loop, or
1249 ** to lock for each call to remove_journal_hash?
1250 */
1258 }
1262 }
1266 }
1268 }
1270 /*
1271 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1272 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1273 ** releasing blocks in this transaction for reuse as data blocks.
1274 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1275 **
1276 */
1280 {
1291 #ifdef CONFIG_REISERFS_CHECK
1294 #endif
1296 }
1297 }
1301 b_data);
1314 }
1318 sync:
1321 }
1326 }
1327 }
1329 }
1334 {
1336 }
1338 /*
1339 ** flush any and all journal lists older than you are
1340 ** can only be called from flush_journal_list
1341 */
1344 {
1350 /* we know we are the only ones flushing things, no extra race
1351 * protection is required.
1352 */
1353 restart:
1355 /* Did we wrap? */
1361 /* do not flush all */
1364 /* other_jl is now deleted from the list */
1366 }
1368 }
1372 {
1377 }
1378 }
1380 /* flush a journal list, both commit and real blocks
1381 **
1382 ** always set flushall to 1, unless you are calling from inside
1383 ** flush_journal_list
1384 **
1385 ** IMPORTANT. This can only be called while there are no journal writers,
1386 ** and the journal is locked. That means it can only be called from
1387 ** do_journal_end, or by journal_release
1388 */
1391 {
1407 }
1410 /* if flushall == 0, the lock is already held */
1415 }
1422 }
1426 /* if all the work is already done, get out of here */
1430 }
1432 /* start by putting the commit list on disk. This will also flush
1433 ** the commit lists of any olders transactions
1434 */
1441 /* are we done now? */
1445 }
1447 /* loop through each cnode, see if we need to write it,
1448 ** or wait on a more recent transaction, or just ignore it
1449 */
1453 }
1459 /* blocknr of 0 is no longer in the hash, ignore it */
1462 }
1464 /* This transaction failed commit. Don't write out to the disk */
1469 /* the order is important here. We check pjl to make sure we
1470 ** don't clear BH_JDirty_wait if we aren't the one writing this
1471 ** block to disk
1472 */
1476 /* we do this to make sure nobody releases the buffer while
1477 ** we are working with it
1478 */
1486 /* everything with !pjl && jwait should be writable */
1488 }
1489 }
1491 /* if someone has this block in a newer transaction, just make
1492 ** sure they are committed, and don't try writing it to disk
1493 */
1498 }
1500 /* bh == NULL when the block got to disk on its own, OR,
1501 ** the block got freed in a future transaction
1502 */
1505 }
1507 /* this should never happen. kupdate_one_transaction has this list
1508 ** locked while it works, so we should never see a buffer here that
1509 ** is not marked JDirty_wait
1510 */
1513 "BAD! buffer %llu %cdirty %cjwait, "
1518 }
1520 /* we inc again because saved_bh gets decremented at free_cnode */
1527 else
1529 count++;
1535 }
1536 free_cnode:
1540 /* we incremented this to keep others from taking the buffer head away */
1545 }
1546 }
1547 }
1555 }
1560 }
1562 #ifdef CONFIG_REISERFS_CHECK
1565 #endif
1567 }
1568 /* note, we must clear the JDirty_wait bit after the up to date
1569 ** check, otherwise we race against our flushpage routine
1570 */
1574 /* drop one ref for us */
1576 /* drop one ref for journal_mark_dirty */
1578 }
1580 }
1581 }
1586 __func__);
1587 flush_older_and_return:
1589 /* before we can update the journal header block, we _must_ flush all
1590 ** real blocks from all older transactions to disk. This is because
1591 ** once the header block is updated, this transaction will not be
1592 ** replayed after a crash
1593 */
1596 }
1599 /* before we can remove everything from the hash tables for this
1600 ** transaction, we must make sure it can never be replayed
1601 **
1602 ** since we are only called from do_journal_end, we know for sure there
1603 ** are no allocations going on while we are flushing journal lists. So,
1604 ** we only need to update the journal header block for the last list
1605 ** being flushed
1606 */
1608 err =
1616 __func__);
1617 }
1627 }
1630 /* not strictly required since we are freeing the list, but it should
1631 * help find code using dead lists later on
1632 */
1645 }
1649 {
1657 /* if the blocknr == 0, this has been cleared from the hash,
1658 ** skip it
1659 */
1662 }
1665 next:
1668 }
1670 }
1675 {
1683 }
1687 /* if the blocknr == 0, this has been cleared from the hash,
1688 ** skip it
1689 */
1692 }
1695 /* we can race against journal_mark_freed when we try
1696 * to lock_buffer(cn->bh), so we have to inc the buffer
1697 * count, and recheck things after locking
1698 */
1707 ret++;
1709 /* note, cn->bh might be null now */
1711 }
1713 }
1714 next:
1717 }
1719 }
1721 /* used by flush_commit_list */
1724 {
1732 /* look for a more recent transaction that logged this
1733 ** buffer. Only the most recent transaction with a buffer in
1734 ** it is allowed to send that buffer to disk
1735 */
1740 /* if the buffer is prepared, it will either be logged
1741 * or restored. If restored, we need to make sure
1742 * it actually gets marked dirty
1743 */
1750 }
1751 }
1753 }
1755 }
1762 {
1775 }
1777 /* we've got j_flush_mutex held, nobody is going to delete any
1778 * of these lists out from underneath us
1779 */
1788 }
1793 transactions_flushed++;
1797 /* did we wrap? */
1800 }
1803 /* don't bother with older transactions */
1806 }
1809 }
1811 done:
1814 }
1816 /* for o_sync and fsync heavy applications, they tend to use
1817 ** all the journa list slots with tiny transactions. These
1818 ** trigger lots and lots of calls to update the header block, which
1819 ** adds seeks and slows things down.
1820 **
1821 ** This function tries to clear out a large chunk of the journal lists
1822 ** at once, which makes everything faster since only the newest journal
1823 ** list updates the header block
1824 */
1827 {
1840 /* in data logging mode, try harder to flush a lot of blocks */
1843 /* flush for 256 transactions or limit blocks, whichever comes first */
1848 }
1852 }
1858 }
1859 /* try to find a group of blocks we can flush across all the
1860 ** transactions, but only bother if we've actually spanned
1861 ** across multiple lists
1862 */
1865 }
1868 }
1870 /*
1871 ** removes any nodes in table with name block and dev as bh.
1872 ** only touchs the hnext and hprev pointers.
1873 */
1878 {
1885 }
1893 }
1898 }
1906 }
1908 }
1909 }
1912 {
1922 }
1923 /* j_header_bh is on the journal dev, make sure not to release the journal
1924 * dev until we brelse j_header_bh
1925 */
1928 }
1930 /*
1931 ** call on unmount. Only set error to 1 if you haven't made your way out
1932 ** of read_super() yet. Any other caller must keep error at 0.
1933 */
1936 {
1941 /* we only want to flush out transactions if we were called with error == 0
1942 */
1944 /* end the current trans */
1948 /* make sure something gets logged to force our way into the flush code */
1957 }
1958 }
1960 /* this also catches errors during the do_journal_end above */
1970 }
1971 }
1973 reiserfs_mounted_fs_count--;
1974 /* wait for all commits to finish */
1980 }
1985 }
1987 /*
1988 ** call on unmount. flush all journal trans, release all alloc'd ram
1989 */
1992 {
1994 }
1996 /*
1997 ** only call from an error condition inside reiserfs_read_super!
1998 */
2001 {
2003 }
2005 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
2009 {
2015 }
2017 }
2019 /* returns 0 if it did not find a description block
2020 ** returns -1 if it found a corrupt commit block
2021 ** returns 1 if both desc and commit were valid
2022 */
2027 {
2042 "journal-986: transaction "
2043 "is valid returning because trans_id %d is greater than "
2048 }
2052 "journal-1087: transaction "
2053 "is valid returning because mount_id %d is less than "
2058 }
2061 "Bad transaction length %d "
2065 }
2068 /* ok, we have a journal description block, lets see if the transaction was valid */
2069 c_bh =
2079 "journal_transaction_is_valid, commit offset %ld had bad "
2090 "journal-1004: "
2091 "transaction_is_valid setting oldest invalid trans_id "
2094 }
2096 }
2099 "journal-1006: found valid "
2108 }
2109 }
2112 {
2116 }
2117 }
2119 /*
2120 ** given the start, and values for the oldest acceptable transactions,
2121 ** this either reads in a replays a transaction, or returns because the transaction
2122 ** is invalid, or too old.
2123 */
2129 {
2154 cur_dblock -
2158 }
2161 "journal_read_trans skipping because %d is != "
2163 newest_mount_id);
2166 }
2173 }
2177 "journal_read_transaction, "
2186 }
2194 }
2197 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2210 }
2211 /* get all the buffer heads */
2228 }
2231 "REPLAY FAILURE fsck required! "
2232 "Block to replay is outside of "
2235 }
2236 /* make sure we don't try to replay onto log or reserved area */
2240 "REPLAY FAILURE fsck required! "
2242 abort_replay:
2250 }
2251 }
2252 /* read in the log blocks, memcpy to the corresponding real block */
2258 "REPLAY FAILURE fsck required! "
2268 }
2273 }
2274 /* flush out the real blocks */
2278 }
2283 "REPLAY FAILURE, fsck required! "
2292 }
2294 }
2295 cur_dblock =
2303 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2307 /* check for trans_id overflow */
2315 }
2317 /* This function reads blocks starting from block and to max_block of bufsize
2318 size (but no more than BUFNR blocks at a time). This proved to improve
2319 mounting speed on self-rebuilding raid5 arrays at least.
2320 Right now it is only used from journal code. But later we might use it
2321 from other places.
2322 Note: Do not use journal_getblk/sb_getblk functions here! */
2325 b_blocknr_t max_block)
2326 {
2338 }
2348 }
2358 }
2360 /*
2361 ** read and replay the log
2362 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2363 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2364 **
2365 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2366 **
2367 ** On exit, it sets things up so the first transaction will work correctly.
2368 */
2370 {
2392 /* step 1, read in the journal header block. Check the transaction it says
2393 ** is the first unflushed, and if that transaction is not valid,
2394 ** replay is done
2395 */
2401 }
2406 oldest_start =
2412 "journal-1153: found in "
2413 "header: first_unflushed_offset %d, last_flushed_trans_id "
2418 /* now, we try to read the first unflushed offset. If it is not valid,
2419 ** there is nothing more we can do, and it makes no sense to read
2420 ** through the whole log.
2421 */
2422 d_bh =
2429 }
2432 }
2434 /* ok, there are transactions that need to be replayed. start with the first log block, find
2435 ** all the valid transactions, and pick out the oldest.
2436 */
2441 /* Note that it is required for blocksize of primary fs device and journal
2442 device to be the same */
2443 d_bh =
2448 ret =
2459 "journal-1179: Setting "
2461 oldest_start -
2462 SB_ONDISK_JOURNAL_1st_BLOCK
2465 /* one we just read was older */
2469 "journal-1180: Resetting "
2471 oldest_start -
2472 SB_ONDISK_JOURNAL_1st_BLOCK
2474 }
2478 "journal-1299: Setting "
2481 }
2484 cur_dblock++;
2485 }
2487 }
2489 start_log_replay:
2493 "journal-1206: Starting replay "
2496 oldest_trans_id);
2498 }
2501 ret =
2508 }
2509 cur_dblock =
2511 replay_count++;
2514 }
2519 }
2520 /* j_start does not get set correctly if we don't replay any transactions.
2521 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2522 ** copy the trans_id from the header
2523 */
2528 /* check for trans_id overflow */
2536 }
2544 }
2548 /* replay failed, caller must call free_journal_ram and abort
2549 ** the mount
2550 */
2552 }
2554 }
2557 {
2569 }
2572 {
2574 }
2578 {
2588 }
2593 }
2595 }
2600 {
2602 dev_t jdev;
2615 /* there is no "jdev" option and journal is on separate device */
2631 }
2634 }
2637 }
2649 }
2656 }
2658 /**
2659 * When creating/tuning a file system user can assign some
2660 * journal params within boundaries which depend on the ratio
2661 * blocksize/standard_blocksize.
2662 *
2663 * For blocks >= standard_blocksize transaction size should
2664 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2665 * then JOURNAL_TRANS_MAX_DEFAULT.
2666 *
2667 * For blocks < standard_blocksize these boundaries should be
2668 * decreased proportionally.
2669 */
2670 #define REISERFS_STANDARD_BLKSIZE (4096)
2674 {
2676 /* Non-default journal params.
2677 Do sanity check for them. */
2685 JOURNAL_MIN_RATIO) {
2687 "bad transaction max size (%u). "
2690 }
2694 "bad transaction max batch (%u). "
2697 }
2699 /* Default journal params.
2700 The file system was created by old version
2701 of mkreiserfs, so some fields contain zeros,
2702 and we need to advise proper values for them */
2707 }
2711 }
2713 }
2715 /*
2716 ** must be called once on fs mount. calls journal_read for you
2717 */
2720 {
2734 }
2747 /* reserved for journal area support */
2749 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2753 REISERFS_DISK_OFFSET_IN_BYTES /
2756 /* Sanity check to see is the standard journal fitting withing first bitmap
2757 (actual for small blocksizes) */
2762 "journal does not fit for area addressed "
2763 "by first of bitmap blocks. It starts at "
2769 }
2775 }
2779 /* read journal header */
2787 }
2790 /* make sure that journal matches to the super block */
2795 "journal header magic %x (device %s) does "
2802 }
2817 }
2820 "journal first block %u, max trans len %u, max batch %u, "
2865 "allocation failed (%ld bytes). Journal is "
2866 "too large for available memory. Usually "
2870 }
2879 }
2884 }
2886 reiserfs_mounted_fs_count++;
2893 free_and_return:
2896 }
2898 /*
2899 ** test for a polite end of the current transaction. Used by file_write, and should
2900 ** be used by delete to make sure they don't write more than can fit inside a single
2901 ** transaction
2902 */
2905 {
2908 /* cannot restart while nested */
2918 }
2919 /* protected by the BKL here */
2923 }
2925 /* this must be called inside a transaction, and requires the
2926 ** kernel_lock to be held
2927 */
2929 {
2935 }
2937 /* this must be called without a transaction started, and does not
2938 ** require BKL
2939 */
2941 {
2945 }
2947 /* this must be called without a transaction started, and does not
2948 ** require BKL
2949 */
2951 {
2955 }
2958 {
2959 wait_queue_t wait;
2963 /*
2964 * we don't want to use wait_event here because
2965 * we only want to wait once.
2966 */
2974 }
2977 {
2981 }
2984 {
2994 }
3000 }
3001 }
3003 /* join == true if you must join an existing transaction.
3004 ** join == false if you can deal with waiting for others to finish
3005 **
3006 ** this will block until the transaction is joinable. send the number of blocks you
3007 ** expect to use in nblocks.
3008 */
3012 {
3024 /* set here for journal_join */
3028 relock:
3034 }
3042 }
3045 /* if there is no room in the journal OR
3046 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3047 ** we don't sleep if there aren't other writers
3048 */
3051 (!join
3068 sched_count++;
3071 }
3072 }
3073 /* don't mess with joining the transaction if all we have to do is
3074 * wait for someone else to do a commit
3075 */
3080 }
3082 }
3087 /* someone might have ended the transaction while we joined */
3092 }
3099 }
3100 /* we are the first writer, set trans_id */
3103 }
3114 out_fail:
3116 /* Re-set th->t_super, so we can properly keep track of how many
3117 * persistent transactions there are. We need to do this so if this
3118 * call is part of a failed restart_transaction, we can free it later */
3121 }
3124 super_block
3127 {
3131 /* if we're nesting into an existing transaction. It will be
3132 ** persistent on its own
3133 */
3140 }
3148 }
3152 }
3155 {
3160 else
3165 }
3167 }
3171 {
3174 /* this keeps do_journal_end from NULLing out the current->journal_info
3175 ** pointer
3176 */
3180 }
3184 {
3187 /* this keeps do_journal_end from NULLing out the current->journal_info
3188 ** pointer
3189 */
3193 }
3197 {
3203 /* we are nesting into the current transaction */
3210 "BAD: refcount <= 1, but "
3214 /* we've ended up with a handle from a different filesystem.
3215 ** save it and restore on journal_end. This should never
3216 ** really happen...
3217 */
3222 }
3225 }
3229 /* I guess this boils down to being the reciprocal of clm-2100 above.
3230 * If do_journal_begin_r fails, we need to put it back, since journal_end
3231 * won't be called to do it. */
3234 else
3238 }
3240 /*
3241 ** puts bh into the current transaction. If it was already there, reorders removes the
3242 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3243 **
3244 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3245 ** transaction is committed.
3246 **
3247 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3248 */
3251 {
3263 }
3269 /* already in this transaction, we are done */
3273 }
3275 /* this must be turned into a panic instead of a warning. We can't allow
3276 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3277 ** could get to disk too early. NOT GOOD.
3278 */
3281 "buffer %llu bad state "
3288 }
3295 }
3296 /* this error means I've screwed up, and we've overflowed the transaction.
3297 ** Nothing can be done here, except make the FS readonly or panic.
3298 */
3303 }
3309 }
3313 }
3317 /* now put this guy on the end */
3322 }
3327 }
3338 }
3339 }
3349 }
3351 }
3355 {
3363 }
3370 /* we aren't allowed to close a nested transaction on a different
3371 ** filesystem from the one in the task struct
3372 */
3378 }
3382 }
3383 }
3385 /* removes from the current transaction, relsing and descrementing any counters.
3386 ** also files the removed buffer directly onto the clean list
3387 **
3388 ** called by journal_mark_freed when a block has been deleted
3389 **
3390 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3391 */
3394 {
3403 }
3407 }
3410 }
3413 }
3416 }
3430 }
3432 }
3437 }
3439 /*
3440 ** for any cnode in a journal list, it can only be dirtied of all the
3441 ** transactions that include it are committed to disk.
3442 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3443 ** and 0 if you aren't
3444 **
3445 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3446 ** blocks for a given transaction on disk
3447 **
3448 */
3450 {
3456 /* first test hprev. These are all newer than cn, so any node here
3457 ** with the same block number and dev means this node can't be sent
3458 ** to disk right now.
3459 */
3464 }
3466 }
3467 /* then test hnext. These are all older than cn. As long as they
3468 ** are committed to the log, it is safe to write cn to disk
3469 */
3476 }
3478 }
3480 }
3482 /* syncs the commit blocks, but does not force the real buffers to disk
3483 ** will wait until the current transaction is done/committed before returning
3484 */
3487 {
3491 /* you can sync while nested, very, very bad */
3497 }
3499 }
3501 /*
3502 ** writeback the pending async commits to disk
3503 */
3505 {
3514 /* last entry is the youngest, commit it and you get everything */
3518 }
3520 }
3522 /*
3523 ** flushes any old transactions to disk
3524 ** ends the current transaction if it is too old
3525 */
3527 {
3533 /* safety check so we don't flush while we are replaying the log during
3534 * mount
3535 */
3538 }
3540 /* check the current transaction. If there are no writers, and it is
3541 * too old, finish it, and force the commit blocks to disk
3542 */
3554 /* we're only being called from kreiserfsd, it makes no sense to do
3555 ** an async commit so that kreiserfsd can do it later
3556 */
3558 }
3559 }
3561 }
3563 /*
3564 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3565 **
3566 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3567 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3568 ** flushes the commit list and returns 0.
3569 **
3570 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3571 **
3572 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3573 */
3577 {
3592 }
3595 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3597 }
3599 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3600 ** will be dealt with by next transaction that actually writes something, but should be taken
3601 ** care of in this trans
3602 */
3605 /* if wcount > 0, and we are called to with flush or commit_now,
3606 ** we wait on j_join_wait. We will wake up when the last writer has
3607 ** finished the transaction, and started it on its way to the disk.
3608 ** Then, we flush the commit or journal list, and just return 0
3609 ** because the rest of journal end was already done for this transaction.
3610 */
3622 }
3625 /* sleep while the current transaction is still j_jlocked */
3634 }
3636 }
3637 }
3644 }
3646 }
3649 }
3651 /* deal with old transactions where we are the last writers */
3656 }
3657 /* don't batch when someone is waiting on j_join_wait */
3658 /* don't batch when syncing the commit or flushing the whole trans */
3666 }
3672 }
3674 }
3676 /*
3677 ** Does all the work that makes deleting blocks safe.
3678 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3679 **
3680 ** otherwise:
3681 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3682 ** before this transaction has finished.
3683 **
3684 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3685 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3686 ** the block can't be reallocated yet.
3687 **
3688 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3689 */
3692 {
3704 }
3705 /* if it is journal new, we just remove it from this transaction */
3712 /* set the bit for this block in the journal bitmap for this transaction */
3717 }
3720 /* Note, the entire while loop is not allowed to schedule. */
3725 }
3728 /* find all older transactions with this block, make sure they don't try to write it out */
3730 blocknr);
3736 /* remove_from_transaction will brelse the buffer if it was
3737 ** in the current trans
3738 */
3740 bh);
3743 bh);
3751 }
3752 }
3756 j_nonzerolen));
3757 }
3759 }
3760 }
3762 }
3763 }
3768 }
3771 {
3775 }
3777 /*
3778 * returns -1 on error, 0 if no commits/barriers were done and 1
3779 * if a transaction was actually committed and the barrier was done
3780 */
3783 {
3789 /* is it from the current transaction, or from an unknown transaction? */
3792 /* try to let other writers come in and grow this transaction */
3796 }
3802 /* someone might have ended this transaction while we joined */
3809 }
3816 /* this gets tricky, we have to make sure the journal list in
3817 * the inode still exists. We know the list is still around
3818 * if we've got a larger transaction id than the oldest list
3819 */
3820 flush_commit_only:
3822 /*
3823 * we only set ret to 1 when we know for sure
3824 * the barrier hasn't been started yet on the commit
3825 * block.
3826 */
3832 }
3833 }
3834 /* otherwise the list is gone, and long since committed */
3836 }
3839 {
3843 /* for the whole inode, assume unset id means it was
3844 * changed in the current transaction. More conservative
3845 */
3849 /* jl will be updated in __commit_trans_jl */
3850 }
3853 }
3857 {
3862 }
3872 }
3873 }
3875 }
3878 /*
3879 ** before we can change a metadata block, we have to make sure it won't
3880 ** be written to disk while we are altering it. So, we must:
3881 ** clean it
3882 ** wait on it.
3883 **
3884 */
3887 {
3894 }
3899 }
3902 }
3905 {
3914 /* this check should always be run, to send old lists to disk */
3921 }
3922 }
3923 }
3925 /*
3926 ** long and ugly. If flush, will not return until all commit
3927 ** blocks and all real buffers in the trans are on disk.
3928 ** If no_async, won't return until all commit blocks are on disk.
3929 **
3930 ** keep reading, there are comments as you go along
3931 **
3932 ** If the journal is aborted, we just clean up. Things like flushing
3933 ** journal lists, etc just won't happen.
3934 */
3938 {
3960 /* protect flush_older_commits from doing mistakes if the
3961 transaction ID counter gets overflowed. */
3974 }
3980 }
3984 }
3986 /* check_journal_end locks the journal, and unlocks if it does not return 1
3987 ** it tells us if we should continue with the journal_end, or just return
3988 */
3994 }
3996 /* check_journal_end might set these, check again */
3999 }
4001 /*
4002 ** j must wait means we have to flush the log blocks, and the real blocks for
4003 ** this transaction
4004 */
4007 }
4008 #ifdef REISERFS_PREALLOCATE
4009 /* quota ops might need to nest, setup the journal_info pointer for them
4010 * and raise the refcount so that it is > 0. */
4014 * the transaction */
4017 #endif
4019 /* setup description block */
4020 d_bh =
4030 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4039 /* init this journal list */
4042 /* we lock the commit before doing anything because
4043 * we want to make sure nobody tries to run flush_commit_list until
4044 * the new transaction is fully setup, and we've already flushed the
4045 * ordered bh list
4046 */
4049 /* save the transaction id in case we need to commit it later */
4062 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4063 ** for each real block, add it to the journal list hash,
4064 ** copy into real block index array in the commit or desc block
4065 */
4073 }
4076 }
4081 }
4083 /* make sure the block we are trying to log is not a block
4084 of journal or reserved area */
4089 "Trying to log block %lu, "
4092 }
4105 }
4107 i--;
4108 }
4109 }
4115 /* special check in case all buffers in the journal were marked for not logging */
4118 /* we're about to dirty all the log blocks, mark the description block
4119 * dirty now too. Don't mark the commit block dirty until all the
4120 * others are on disk
4121 */
4124 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4130 /* copy all the real blocks into log area. dirty log blocks */
4135 tmp_bh =
4139 jindex) %
4149 jindex++;
4153 /* JDirty cleared sometime during transaction. don't log this one */
4155 "BAD, buffer in journal hash, "
4158 }
4163 }
4165 /* we are done with both the c_bh and d_bh, but
4166 ** c_bh must be written after all other commit blocks,
4167 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4168 */
4172 /* now it is safe to insert this transaction on the main list */
4177 /* reset journal values for the next transaction */
4188 /* check for trans_id overflow */
4198 // make sure reiserfs_add_jh sees the new current_jl before we
4199 // write out the tails
4202 /* tail conversion targets have to hit the disk before we end the
4203 * transaction. Otherwise a later transaction might repack the tail
4204 * before this transaction commits, leaving the data block unflushed and
4205 * clean, if we crash before the later transaction commits, the data block
4206 * is lost.
4207 */
4213 }
4217 /* honor the flush wishes from the caller, simple commits can
4218 ** be done outside the journal lock, they are done below
4219 **
4220 ** if we don't flush the commit list right now, we put it into
4221 ** the work queue so the people waiting on the async progress work
4222 ** queue don't wait for this proc to flush journal lists and such.
4223 */
4230 /* if the next transaction has any chance of wrapping, flush
4231 ** transactions that might get overwritten. If any journal lists are very
4232 ** old flush them as well.
4233 */
4234 first_jl:
4245 /* if we don't cross into the next transaction and we don't
4246 * wrap, there is no way we can overlap any later transactions
4247 * break now
4248 */
4250 }
4260 /* we don't overlap anything from out start to the end of the
4261 * log, and our wrapped portion doesn't overlap anything at
4262 * the start of the log. We can break
4263 */
4265 }
4266 }
4267 }
4276 }
4280 /* wake up any body waiting to join. */
4287 }
4288 out:
4292 /* Re-set th->t_super, so we can properly keep track of how many
4293 * persistent transactions there are. We need to do this so if this
4294 * call is part of a failed restart_transaction, we can free it later */
4298 }
4300 /* Send the file system read only and refuse new transactions */
4302 {
4313 #ifdef CONFIG_REISERFS_CHECK
4315 #endif
4316 }