| blob | 9643c3bbeb3b611d3817bc17698d76c5d54e89b5 |
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
440 caller);
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
1126 #endif
1128 }
1132 }
1137 /* If there was a write error in the journal - we can't commit
1138 * this transaction - it will be invalid and, if successful,
1139 * will just end up propagating the write error out to
1140 * the file system. */
1146 }
1152 /* If there was a write error in the journal - we can't commit this
1153 * transaction - it will be invalid and, if successful, will just end
1154 * up propagating the write error out to the filesystem. */
1156 #ifdef CONFIG_REISERFS_CHECK
1158 #endif
1160 }
1166 }
1169 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1174 /* mark the metadata dirty */
1181 }
1183 put_jl:
1188 __func__);
1191 }
1193 /*
1194 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1195 ** returns NULL if it can't find anything
1196 */
1198 reiserfs_journal_cnode
1200 {
1208 }
1210 }
1212 }
1215 {
1225 }
1227 }
1234 /*
1235 ** once all the real blocks have been flushed, it is safe to remove them from the
1236 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1237 ** block to be reallocated for data blocks if it had been deleted.
1238 */
1242 {
1247 /* which is better, to lock once around the whole loop, or
1248 ** to lock for each call to remove_journal_hash?
1249 */
1257 }
1261 }
1265 }
1267 }
1269 /*
1270 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1271 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1272 ** releasing blocks in this transaction for reuse as data blocks.
1273 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1274 **
1275 */
1279 {
1290 #ifdef CONFIG_REISERFS_CHECK
1293 #endif
1295 }
1296 }
1300 b_data);
1313 }
1317 sync:
1320 }
1325 }
1326 }
1328 }
1333 {
1335 }
1337 /*
1338 ** flush any and all journal lists older than you are
1339 ** can only be called from flush_journal_list
1340 */
1343 {
1349 /* we know we are the only ones flushing things, no extra race
1350 * protection is required.
1351 */
1352 restart:
1354 /* Did we wrap? */
1360 /* do not flush all */
1363 /* other_jl is now deleted from the list */
1365 }
1367 }
1371 {
1376 }
1377 }
1379 /* flush a journal list, both commit and real blocks
1380 **
1381 ** always set flushall to 1, unless you are calling from inside
1382 ** flush_journal_list
1383 **
1384 ** IMPORTANT. This can only be called while there are no journal writers,
1385 ** and the journal is locked. That means it can only be called from
1386 ** do_journal_end, or by journal_release
1387 */
1390 {
1407 }
1410 /* if flushall == 0, the lock is already held */
1415 }
1423 }
1427 /* if all the work is already done, get out of here */
1431 }
1433 /* start by putting the commit list on disk. This will also flush
1434 ** the commit lists of any olders transactions
1435 */
1442 /* are we done now? */
1446 }
1448 /* loop through each cnode, see if we need to write it,
1449 ** or wait on a more recent transaction, or just ignore it
1450 */
1454 }
1460 /* blocknr of 0 is no longer in the hash, ignore it */
1463 }
1465 /* This transaction failed commit. Don't write out to the disk */
1470 /* the order is important here. We check pjl to make sure we
1471 ** don't clear BH_JDirty_wait if we aren't the one writing this
1472 ** block to disk
1473 */
1477 /* we do this to make sure nobody releases the buffer while
1478 ** we are working with it
1479 */
1487 /* everything with !pjl && jwait should be writable */
1489 }
1490 }
1492 /* if someone has this block in a newer transaction, just make
1493 ** sure they are committed, and don't try writing it to disk
1494 */
1499 }
1501 /* bh == NULL when the block got to disk on its own, OR,
1502 ** the block got freed in a future transaction
1503 */
1506 }
1508 /* this should never happen. kupdate_one_transaction has this list
1509 ** locked while it works, so we should never see a buffer here that
1510 ** is not marked JDirty_wait
1511 */
1514 "journal-813: BAD! buffer %llu %cdirty %cjwait, "
1519 }
1521 /* we inc again because saved_bh gets decremented at free_cnode */
1528 else
1530 count++;
1536 }
1537 free_cnode:
1541 /* we incremented this to keep others from taking the buffer head away */
1546 }
1547 }
1548 }
1556 }
1561 }
1563 #ifdef CONFIG_REISERFS_CHECK
1566 #endif
1568 }
1569 /* note, we must clear the JDirty_wait bit after the up to date
1570 ** check, otherwise we race against our flushpage routine
1571 */
1575 /* drop one ref for us */
1577 /* drop one ref for journal_mark_dirty */
1579 }
1581 }
1582 }
1587 __func__);
1588 flush_older_and_return:
1590 /* before we can update the journal header block, we _must_ flush all
1591 ** real blocks from all older transactions to disk. This is because
1592 ** once the header block is updated, this transaction will not be
1593 ** replayed after a crash
1594 */
1597 }
1600 /* before we can remove everything from the hash tables for this
1601 ** transaction, we must make sure it can never be replayed
1602 **
1603 ** since we are only called from do_journal_end, we know for sure there
1604 ** are no allocations going on while we are flushing journal lists. So,
1605 ** we only need to update the journal header block for the last list
1606 ** being flushed
1607 */
1609 err =
1617 __func__);
1618 }
1628 }
1631 /* not strictly required since we are freeing the list, but it should
1632 * help find code using dead lists later on
1633 */
1646 }
1650 {
1658 /* if the blocknr == 0, this has been cleared from the hash,
1659 ** skip it
1660 */
1663 }
1666 next:
1669 }
1671 }
1676 {
1684 }
1688 /* if the blocknr == 0, this has been cleared from the hash,
1689 ** skip it
1690 */
1693 }
1696 /* we can race against journal_mark_freed when we try
1697 * to lock_buffer(cn->bh), so we have to inc the buffer
1698 * count, and recheck things after locking
1699 */
1708 ret++;
1710 /* note, cn->bh might be null now */
1712 }
1714 }
1715 next:
1718 }
1720 }
1722 /* used by flush_commit_list */
1725 {
1733 /* look for a more recent transaction that logged this
1734 ** buffer. Only the most recent transaction with a buffer in
1735 ** it is allowed to send that buffer to disk
1736 */
1741 /* if the buffer is prepared, it will either be logged
1742 * or restored. If restored, we need to make sure
1743 * it actually gets marked dirty
1744 */
1751 }
1752 }
1754 }
1756 }
1763 {
1776 }
1778 /* we've got j_flush_mutex held, nobody is going to delete any
1779 * of these lists out from underneath us
1780 */
1789 }
1794 transactions_flushed++;
1798 /* did we wrap? */
1801 }
1804 /* don't bother with older transactions */
1807 }
1810 }
1812 done:
1815 }
1817 /* for o_sync and fsync heavy applications, they tend to use
1818 ** all the journa list slots with tiny transactions. These
1819 ** trigger lots and lots of calls to update the header block, which
1820 ** adds seeks and slows things down.
1821 **
1822 ** This function tries to clear out a large chunk of the journal lists
1823 ** at once, which makes everything faster since only the newest journal
1824 ** list updates the header block
1825 */
1828 {
1841 /* in data logging mode, try harder to flush a lot of blocks */
1844 /* flush for 256 transactions or limit blocks, whichever comes first */
1849 }
1853 }
1859 }
1860 /* try to find a group of blocks we can flush across all the
1861 ** transactions, but only bother if we've actually spanned
1862 ** across multiple lists
1863 */
1866 }
1869 }
1871 /*
1872 ** removes any nodes in table with name block and dev as bh.
1873 ** only touchs the hnext and hprev pointers.
1874 */
1879 {
1886 }
1894 }
1899 }
1907 }
1909 }
1910 }
1913 {
1923 }
1924 /* j_header_bh is on the journal dev, make sure not to release the journal
1925 * dev until we brelse j_header_bh
1926 */
1929 }
1931 /*
1932 ** call on unmount. Only set error to 1 if you haven't made your way out
1933 ** of read_super() yet. Any other caller must keep error at 0.
1934 */
1937 {
1942 /* we only want to flush out transactions if we were called with error == 0
1943 */
1945 /* end the current trans */
1949 /* make sure something gets logged to force our way into the flush code */
1958 }
1959 }
1961 /* this also catches errors during the do_journal_end above */
1971 }
1972 }
1974 reiserfs_mounted_fs_count--;
1975 /* wait for all commits to finish */
1981 }
1986 }
1988 /*
1989 ** call on unmount. flush all journal trans, release all alloc'd ram
1990 */
1993 {
1995 }
1997 /*
1998 ** only call from an error condition inside reiserfs_read_super!
1999 */
2002 {
2004 }
2006 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
2010 {
2016 }
2018 }
2020 /* returns 0 if it did not find a description block
2021 ** returns -1 if it found a corrupt commit block
2022 ** returns 1 if both desc and commit were valid
2023 */
2028 {
2043 "journal-986: transaction "
2044 "is valid returning because trans_id %d is greater than "
2049 }
2053 "journal-1087: transaction "
2054 "is valid returning because mount_id %d is less than "
2059 }
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 }
2188 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2201 }
2202 /* get all the buffer heads */
2219 }
2224 }
2225 /* make sure we don't try to replay onto log or reserved area */
2230 abort_replay:
2238 }
2239 }
2240 /* read in the log blocks, memcpy to the corresponding real block */
2255 }
2260 }
2261 /* flush out the real blocks */
2265 }
2278 }
2280 }
2281 cur_dblock =
2289 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2293 /* check for trans_id overflow */
2301 }
2303 /* This function reads blocks starting from block and to max_block of bufsize
2304 size (but no more than BUFNR blocks at a time). This proved to improve
2305 mounting speed on self-rebuilding raid5 arrays at least.
2306 Right now it is only used from journal code. But later we might use it
2307 from other places.
2308 Note: Do not use journal_getblk/sb_getblk functions here! */
2311 b_blocknr_t max_block)
2312 {
2324 }
2334 }
2344 }
2346 /*
2347 ** read and replay the log
2348 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2349 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2350 **
2351 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2352 **
2353 ** On exit, it sets things up so the first transaction will work correctly.
2354 */
2356 {
2378 /* step 1, read in the journal header block. Check the transaction it says
2379 ** is the first unflushed, and if that transaction is not valid,
2380 ** replay is done
2381 */
2387 }
2392 oldest_start =
2398 "journal-1153: found in "
2399 "header: first_unflushed_offset %d, last_flushed_trans_id "
2404 /* now, we try to read the first unflushed offset. If it is not valid,
2405 ** there is nothing more we can do, and it makes no sense to read
2406 ** through the whole log.
2407 */
2408 d_bh =
2415 }
2418 }
2424 }
2426 /* ok, there are transactions that need to be replayed. start with the first log block, find
2427 ** all the valid transactions, and pick out the oldest.
2428 */
2433 /* Note that it is required for blocksize of primary fs device and journal
2434 device to be the same */
2435 d_bh =
2440 ret =
2451 "journal-1179: Setting "
2453 oldest_start -
2454 SB_ONDISK_JOURNAL_1st_BLOCK
2457 /* one we just read was older */
2461 "journal-1180: Resetting "
2463 oldest_start -
2464 SB_ONDISK_JOURNAL_1st_BLOCK
2466 }
2470 "journal-1299: Setting "
2473 }
2476 cur_dblock++;
2477 }
2479 }
2481 start_log_replay:
2485 "journal-1206: Starting replay "
2488 oldest_trans_id);
2490 }
2493 ret =
2500 }
2501 cur_dblock =
2503 replay_count++;
2506 }
2511 }
2512 /* j_start does not get set correctly if we don't replay any transactions.
2513 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2514 ** copy the trans_id from the header
2515 */
2520 /* check for trans_id overflow */
2528 }
2536 }
2540 /* replay failed, caller must call free_journal_ram and abort
2541 ** the mount
2542 */
2544 }
2546 }
2549 {
2561 }
2564 {
2566 }
2570 {
2580 }
2585 result);
2586 }
2588 }
2593 {
2595 dev_t jdev;
2608 /* there is no "jdev" option and journal is on separate device */
2624 }
2627 }
2630 }
2642 }
2649 }
2651 /**
2652 * When creating/tuning a file system user can assign some
2653 * journal params within boundaries which depend on the ratio
2654 * blocksize/standard_blocksize.
2655 *
2656 * For blocks >= standard_blocksize transaction size should
2657 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2658 * then JOURNAL_TRANS_MAX_DEFAULT.
2659 *
2660 * For blocks < standard_blocksize these boundaries should be
2661 * decreased proportionally.
2662 */
2663 #define REISERFS_STANDARD_BLKSIZE (4096)
2667 {
2669 /* Non-default journal params.
2670 Do sanity check for them. */
2678 JOURNAL_MIN_RATIO) {
2683 }
2690 }
2692 /* Default journal params.
2693 The file system was created by old version
2694 of mkreiserfs, so some fields contain zeros,
2695 and we need to advise proper values for them */
2702 }
2704 }
2706 /*
2707 ** must be called once on fs mount. calls journal_read for you
2708 */
2711 {
2725 }
2738 /* reserved for journal area support */
2740 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2744 REISERFS_DISK_OFFSET_IN_BYTES /
2747 /* Sanity check to see is the standard journal fitting withing first bitmap
2748 (actual for small blocksizes) */
2753 "journal-1393: journal does not fit for area "
2754 "addressed by first of bitmap blocks. It starts at "
2760 }
2766 }
2770 /* read journal header */
2778 }
2781 /* make sure that journal matches to the super block */
2786 "sh-460: journal header magic %x "
2787 "(device %s) does not match to magic found in super "
2793 }
2808 }
2811 "journal first block %u, max trans len %u, max batch %u, "
2856 "allocation failed (%ld bytes). Journal is "
2857 "too large for available memory. Usually "
2861 }
2870 }
2874 }
2876 reiserfs_mounted_fs_count++;
2883 free_and_return:
2886 }
2888 /*
2889 ** test for a polite end of the current transaction. Used by file_write, and should
2890 ** be used by delete to make sure they don't write more than can fit inside a single
2891 ** transaction
2892 */
2895 {
2898 /* cannot restart while nested */
2908 }
2909 /* protected by the BKL here */
2913 }
2915 /* this must be called inside a transaction, and requires the
2916 ** kernel_lock to be held
2917 */
2919 {
2925 }
2927 /* this must be called without a transaction started, and does not
2928 ** require BKL
2929 */
2931 {
2935 }
2937 /* this must be called without a transaction started, and does not
2938 ** require BKL
2939 */
2941 {
2945 }
2948 {
2949 wait_queue_t wait;
2953 /*
2954 * we don't want to use wait_event here because
2955 * we only want to wait once.
2956 */
2964 }
2967 {
2971 }
2974 {
2984 }
2990 }
2991 }
2993 /* join == true if you must join an existing transaction.
2994 ** join == false if you can deal with waiting for others to finish
2995 **
2996 ** this will block until the transaction is joinable. send the number of blocks you
2997 ** expect to use in nblocks.
2998 */
3002 {
3014 /* set here for journal_join */
3018 relock:
3024 }
3032 }
3035 /* if there is no room in the journal OR
3036 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3037 ** we don't sleep if there aren't other writers
3038 */
3041 (!join
3058 sched_count++;
3061 }
3062 }
3063 /* don't mess with joining the transaction if all we have to do is
3064 * wait for someone else to do a commit
3065 */
3070 }
3072 }
3077 /* someone might have ended the transaction while we joined */
3082 }
3089 }
3090 /* we are the first writer, set trans_id */
3093 }
3104 out_fail:
3106 /* Re-set th->t_super, so we can properly keep track of how many
3107 * persistent transactions there are. We need to do this so if this
3108 * call is part of a failed restart_transaction, we can free it later */
3111 }
3114 super_block
3117 {
3121 /* if we're nesting into an existing transaction. It will be
3122 ** persistent on its own
3123 */
3130 }
3138 }
3142 }
3145 {
3150 else
3155 }
3157 }
3161 {
3164 /* this keeps do_journal_end from NULLing out the current->journal_info
3165 ** pointer
3166 */
3170 }
3174 {
3177 /* this keeps do_journal_end from NULLing out the current->journal_info
3178 ** pointer
3179 */
3183 }
3187 {
3193 /* we are nesting into the current transaction */
3203 /* we've ended up with a handle from a different filesystem.
3204 ** save it and restore on journal_end. This should never
3205 ** really happen...
3206 */
3211 }
3214 }
3218 /* I guess this boils down to being the reciprocal of clm-2100 above.
3219 * If do_journal_begin_r fails, we need to put it back, since journal_end
3220 * won't be called to do it. */
3223 else
3227 }
3229 /*
3230 ** puts bh into the current transaction. If it was already there, reorders removes the
3231 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3232 **
3233 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3234 ** transaction is committed.
3235 **
3236 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3237 */
3240 {
3252 }
3258 /* already in this transaction, we are done */
3262 }
3264 /* this must be turned into a panic instead of a warning. We can't allow
3265 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3266 ** could get to disk too early. NOT GOOD.
3267 */
3276 }
3283 }
3284 /* this error means I've screwed up, and we've overflowed the transaction.
3285 ** Nothing can be done here, except make the FS readonly or panic.
3286 */
3291 }
3297 }
3301 }
3305 /* now put this guy on the end */
3310 }
3315 }
3326 }
3327 }
3337 }
3339 }
3343 {
3351 }
3358 /* we aren't allowed to close a nested transaction on a different
3359 ** filesystem from the one in the task struct
3360 */
3366 }
3370 }
3371 }
3373 /* removes from the current transaction, relsing and descrementing any counters.
3374 ** also files the removed buffer directly onto the clean list
3375 **
3376 ** called by journal_mark_freed when a block has been deleted
3377 **
3378 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3379 */
3382 {
3391 }
3395 }
3398 }
3401 }
3404 }
3418 }
3420 }
3425 }
3427 /*
3428 ** for any cnode in a journal list, it can only be dirtied of all the
3429 ** transactions that include it are committed to disk.
3430 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3431 ** and 0 if you aren't
3432 **
3433 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3434 ** blocks for a given transaction on disk
3435 **
3436 */
3438 {
3444 /* first test hprev. These are all newer than cn, so any node here
3445 ** with the same block number and dev means this node can't be sent
3446 ** to disk right now.
3447 */
3452 }
3454 }
3455 /* then test hnext. These are all older than cn. As long as they
3456 ** are committed to the log, it is safe to write cn to disk
3457 */
3464 }
3466 }
3468 }
3470 /* syncs the commit blocks, but does not force the real buffers to disk
3471 ** will wait until the current transaction is done/committed before returning
3472 */
3475 {
3479 /* you can sync while nested, very, very bad */
3485 }
3487 }
3489 /*
3490 ** writeback the pending async commits to disk
3491 */
3493 {
3502 /* last entry is the youngest, commit it and you get everything */
3506 }
3508 }
3510 /*
3511 ** flushes any old transactions to disk
3512 ** ends the current transaction if it is too old
3513 */
3515 {
3521 /* safety check so we don't flush while we are replaying the log during
3522 * mount
3523 */
3526 }
3528 /* check the current transaction. If there are no writers, and it is
3529 * too old, finish it, and force the commit blocks to disk
3530 */
3542 /* we're only being called from kreiserfsd, it makes no sense to do
3543 ** an async commit so that kreiserfsd can do it later
3544 */
3546 }
3547 }
3549 }
3551 /*
3552 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3553 **
3554 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3555 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3556 ** flushes the commit list and returns 0.
3557 **
3558 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3559 **
3560 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3561 */
3565 {
3580 }
3583 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3585 }
3587 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3588 ** will be dealt with by next transaction that actually writes something, but should be taken
3589 ** care of in this trans
3590 */
3593 /* if wcount > 0, and we are called to with flush or commit_now,
3594 ** we wait on j_join_wait. We will wake up when the last writer has
3595 ** finished the transaction, and started it on its way to the disk.
3596 ** Then, we flush the commit or journal list, and just return 0
3597 ** because the rest of journal end was already done for this transaction.
3598 */
3610 }
3613 /* sleep while the current transaction is still j_jlocked */
3622 }
3624 }
3625 }
3632 }
3634 }
3637 }
3639 /* deal with old transactions where we are the last writers */
3644 }
3645 /* don't batch when someone is waiting on j_join_wait */
3646 /* don't batch when syncing the commit or flushing the whole trans */
3654 }
3660 }
3662 }
3664 /*
3665 ** Does all the work that makes deleting blocks safe.
3666 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3667 **
3668 ** otherwise:
3669 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3670 ** before this transaction has finished.
3671 **
3672 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3673 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3674 ** the block can't be reallocated yet.
3675 **
3676 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3677 */
3680 {
3692 }
3693 /* if it is journal new, we just remove it from this transaction */
3700 /* set the bit for this block in the journal bitmap for this transaction */
3705 }
3708 /* Note, the entire while loop is not allowed to schedule. */
3713 }
3716 /* find all older transactions with this block, make sure they don't try to write it out */
3718 blocknr);
3724 /* remove_from_transaction will brelse the buffer if it was
3725 ** in the current trans
3726 */
3728 bh);
3731 bh);
3738 }
3739 }
3743 j_nonzerolen));
3744 }
3746 }
3747 }
3749 }
3750 }
3755 }
3758 {
3762 }
3764 /*
3765 * returns -1 on error, 0 if no commits/barriers were done and 1
3766 * if a transaction was actually committed and the barrier was done
3767 */
3770 {
3776 /* is it from the current transaction, or from an unknown transaction? */
3779 /* try to let other writers come in and grow this transaction */
3783 }
3789 /* someone might have ended this transaction while we joined */
3796 }
3803 /* this gets tricky, we have to make sure the journal list in
3804 * the inode still exists. We know the list is still around
3805 * if we've got a larger transaction id than the oldest list
3806 */
3807 flush_commit_only:
3809 /*
3810 * we only set ret to 1 when we know for sure
3811 * the barrier hasn't been started yet on the commit
3812 * block.
3813 */
3819 }
3820 }
3821 /* otherwise the list is gone, and long since committed */
3823 }
3826 {
3830 /* for the whole inode, assume unset id means it was
3831 * changed in the current transaction. More conservative
3832 */
3836 /* jl will be updated in __commit_trans_jl */
3837 }
3840 }
3844 {
3849 }
3859 }
3860 }
3862 }
3865 /*
3866 ** before we can change a metadata block, we have to make sure it won't
3867 ** be written to disk while we are altering it. So, we must:
3868 ** clean it
3869 ** wait on it.
3870 **
3871 */
3874 {
3881 }
3886 }
3889 }
3892 {
3901 /* this check should always be run, to send old lists to disk */
3908 }
3909 }
3910 }
3912 /*
3913 ** long and ugly. If flush, will not return until all commit
3914 ** blocks and all real buffers in the trans are on disk.
3915 ** If no_async, won't return until all commit blocks are on disk.
3916 **
3917 ** keep reading, there are comments as you go along
3918 **
3919 ** If the journal is aborted, we just clean up. Things like flushing
3920 ** journal lists, etc just won't happen.
3921 */
3925 {
3947 /* protect flush_older_commits from doing mistakes if the
3948 transaction ID counter gets overflowed. */
3961 }
3967 }
3971 }
3973 /* check_journal_end locks the journal, and unlocks if it does not return 1
3974 ** it tells us if we should continue with the journal_end, or just return
3975 */
3981 }
3983 /* check_journal_end might set these, check again */
3986 }
3988 /*
3989 ** j must wait means we have to flush the log blocks, and the real blocks for
3990 ** this transaction
3991 */
3994 }
3995 #ifdef REISERFS_PREALLOCATE
3996 /* quota ops might need to nest, setup the journal_info pointer for them
3997 * and raise the refcount so that it is > 0. */
4001 * the transaction */
4004 #endif
4006 /* setup description block */
4007 d_bh =
4017 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4026 /* init this journal list */
4029 /* we lock the commit before doing anything because
4030 * we want to make sure nobody tries to run flush_commit_list until
4031 * the new transaction is fully setup, and we've already flushed the
4032 * ordered bh list
4033 */
4036 /* save the transaction id in case we need to commit it later */
4049 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4050 ** for each real block, add it to the journal list hash,
4051 ** copy into real block index array in the commit or desc block
4052 */
4060 }
4063 }
4068 }
4070 /* make sure the block we are trying to log is not a block
4071 of journal or reserved area */
4078 }
4091 }
4093 i--;
4094 }
4095 }
4101 /* special check in case all buffers in the journal were marked for not logging */
4104 /* we're about to dirty all the log blocks, mark the description block
4105 * dirty now too. Don't mark the commit block dirty until all the
4106 * others are on disk
4107 */
4110 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4116 /* copy all the real blocks into log area. dirty log blocks */
4121 tmp_bh =
4125 jindex) %
4135 jindex++;
4139 /* JDirty cleared sometime during transaction. don't log this one */
4143 }
4148 }
4150 /* we are done with both the c_bh and d_bh, but
4151 ** c_bh must be written after all other commit blocks,
4152 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4153 */
4157 /* now it is safe to insert this transaction on the main list */
4162 /* reset journal values for the next transaction */
4173 /* check for trans_id overflow */
4183 // make sure reiserfs_add_jh sees the new current_jl before we
4184 // write out the tails
4187 /* tail conversion targets have to hit the disk before we end the
4188 * transaction. Otherwise a later transaction might repack the tail
4189 * before this transaction commits, leaving the data block unflushed and
4190 * clean, if we crash before the later transaction commits, the data block
4191 * is lost.
4192 */
4198 }
4202 /* honor the flush wishes from the caller, simple commits can
4203 ** be done outside the journal lock, they are done below
4204 **
4205 ** if we don't flush the commit list right now, we put it into
4206 ** the work queue so the people waiting on the async progress work
4207 ** queue don't wait for this proc to flush journal lists and such.
4208 */
4215 /* if the next transaction has any chance of wrapping, flush
4216 ** transactions that might get overwritten. If any journal lists are very
4217 ** old flush them as well.
4218 */
4219 first_jl:
4230 /* if we don't cross into the next transaction and we don't
4231 * wrap, there is no way we can overlap any later transactions
4232 * break now
4233 */
4235 }
4245 /* we don't overlap anything from out start to the end of the
4246 * log, and our wrapped portion doesn't overlap anything at
4247 * the start of the log. We can break
4248 */
4250 }
4251 }
4252 }
4261 }
4265 /* wake up any body waiting to join. */
4272 }
4273 out:
4277 /* Re-set th->t_super, so we can properly keep track of how many
4278 * persistent transactions there are. We need to do this so if this
4279 * call is part of a failed restart_transaction, we can free it later */
4283 }
4286 {
4297 #ifdef CONFIG_REISERFS_CHECK
4299 #endif
4300 }
4303 {
4312 }
4315 {
4317 }