| blob | a159ba5a35e722457f83977bf4a82c1344528fc8 |
1 /*
2 ** Write ahead logging implementation copyright Chris Mason 2000
3 **
4 ** The background commits make this code very interrelated, and
5 ** overly complex. I need to rethink things a bit....The major players:
6 **
7 ** journal_begin -- call with the number of blocks you expect to log.
8 ** If the current transaction is too
9 ** old, it will block until the current transaction is
10 ** finished, and then start a new one.
11 ** Usually, your transaction will get joined in with
12 ** previous ones for speed.
13 **
14 ** journal_join -- same as journal_begin, but won't block on the current
15 ** transaction regardless of age. Don't ever call
16 ** this. Ever. There are only two places it should be
17 ** called from, and they are both inside this file.
18 **
19 ** journal_mark_dirty -- adds blocks into this transaction. clears any flags
20 ** that might make them get sent to disk
21 ** and then marks them BH_JDirty. Puts the buffer head
22 ** into the current transaction hash.
23 **
24 ** journal_end -- if the current transaction is batchable, it does nothing
25 ** otherwise, it could do an async/synchronous commit, or
26 ** a full flush of all log and real blocks in the
27 ** transaction.
28 **
29 ** flush_old_commits -- if the current transaction is too old, it is ended and
30 ** commit blocks are sent to disk. Forces commit blocks
31 ** to disk for all backgrounded commits that have been
32 ** around too long.
33 ** -- Note, if you call this as an immediate flush from
34 ** from within kupdate, it will ignore the immediate flag
35 */
37 #include <linux/time.h>
38 #include <linux/semaphore.h>
39 #include <linux/vmalloc.h>
40 #include <linux/reiserfs_fs.h>
41 #include <linux/kernel.h>
42 #include <linux/errno.h>
43 #include <linux/fcntl.h>
44 #include <linux/stat.h>
45 #include <linux/string.h>
46 #include <linux/buffer_head.h>
47 #include <linux/workqueue.h>
48 #include <linux/writeback.h>
49 #include <linux/blkdev.h>
50 #include <linux/backing-dev.h>
51 #include <linux/uaccess.h>
52 #include <linux/slab.h>
54 #include <asm/system.h>
56 /* gets a struct reiserfs_journal_list * from a list head */
57 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
58 j_list))
59 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
60 j_working_list))
62 /* the number of mounted filesystems. This is used to decide when to
63 ** start and kill the commit workqueue
64 */
70 structs at 4k */
73 /* cnode stat bits. Move these into reiserfs_fs.h */
76 #define BLOCK_FREED_HOLDER 3 /* this block was freed during this transaction, and can't be written */
79 #define BLOCK_DIRTIED 5
81 /* journal list state bits */
82 #define LIST_TOUCHED 1
83 #define LIST_DIRTY 2
86 /* flags for do_journal_end */
108 /* values for join in do_journal_begin_r */
113 };
120 {
124 }
126 /*
127 ** clears BH_Dirty and sticks the buffer on the clean list. Called because I can't allow refile_buffer to
128 ** make schedule happen after I've freed a block. Look at remove_from_transaction and journal_mark_freed for
129 ** more details.
130 */
132 {
136 }
138 }
142 {
149 }
154 }
158 }
161 {
167 repeat:
175 }
180 }
182 }
185 {
194 }
195 }
198 {
209 }
210 }
211 }
214 b_blocknr_t block,
216 {
222 }
225 }
229 {
238 }
239 }
240 }
242 /*
243 ** only call this on FS unmount.
244 */
247 {
256 }
258 }
261 {
273 }
276 }
278 /*
279 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
280 ** jb_array is the array to be filled in.
281 */
285 {
300 }
302 }
306 }
308 }
310 /*
311 ** find an available list bitmap. If you can't find one, flush a commit list
312 ** and try again
313 */
315 struct reiserfs_journal_list
317 {
332 }
335 }
336 }
339 }
342 }
344 /*
345 ** allocates a new chunk of X nodes, and links them all together as a list.
346 ** Uses the cnode->next and cnode->prev pointers
347 ** returns NULL on failure
348 */
350 {
355 }
359 }
366 }
369 }
371 /*
372 ** pulls a cnode off the free list, or returns NULL on failure
373 */
375 {
383 }
389 }
392 }
396 }
398 /*
399 ** returns a cnode to the free list
400 */
403 {
410 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
414 }
415 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
417 }
420 {
423 }
425 /* return a cnode with same dev, block number and size in table, or null if not found */
427 super_block
429 struct
430 reiserfs_journal_cnode
433 {
440 }
442 }
444 /*
445 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
446 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
447 ** being overwritten by a replay after crashing.
448 **
449 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
450 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
451 ** sure you never write the block without logging it.
452 **
453 ** next_zero_bit is a suggestion about the next block to try for find_forward.
454 ** when bl is rejected because it is set in a journal list bitmap, we search
455 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
456 ** through next_zero_bit for find_forward to try.
457 **
458 ** Just because we return something in next_zero_bit does not mean we won't
459 ** reject it on the next call to reiserfs_in_journal
460 **
461 */
465 {
475 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
476 ** if we crash before the transaction that freed it commits, this transaction won't
477 ** have committed either, and the block will never be written
478 */
486 data)) {
490 data),
494 }
495 }
496 }
499 /* is it in any old transactions? */
504 }
506 /* is it in the current transaction. This should never happen */
510 }
513 /* safe for reuse */
515 }
517 /* insert cn into table
518 */
521 {
529 }
531 }
533 /* lock the current transaction */
535 {
539 }
541 /* unlock the current transaction */
543 {
545 }
548 {
550 }
554 {
558 }
561 }
563 /*
564 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
565 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
566 ** transaction.
567 */
570 {
575 }
578 }
582 {
591 }
592 }
594 }
596 /*
597 * If page->mapping was null, we failed to truncate this page for
598 * some reason. Most likely because it was truncated after being
599 * logged via data=journal.
600 *
601 * This does a check to see if the buffer belongs to one of these
602 * lost pages before doing the final put_bh. If page->mapping was
603 * null, it tries to free buffers on the page, which should make the
604 * final page_cache_release drop the page from the lru.
605 */
607 {
618 }
619 }
622 {
629 }
632 else
637 }
640 {
643 else
647 }
650 {
660 }
663 {
670 }
672 #define CHUNK_SIZE 32
676 };
679 {
683 }
685 }
688 {
692 }
694 }
698 {
709 }
711 }
715 {
722 }
724 }
725 }
727 /*
728 * we want to free the jh when the buffer has been written
729 * and waited on
730 */
732 {
745 }
746 }
750 {
758 }
762 no_jh:
766 /* buffer must be locked for __add_jh, should be able to have
767 * two adds at the same time
768 */
772 }
778 }
781 }
784 {
786 }
788 {
790 }
792 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
797 {
815 }
823 }
824 /* in theory, dirty non-uptodate buffers should never get here,
825 * but the upper layer io error paths still have a few quirks.
826 * Handle them here as gracefully as we can
827 */
831 }
838 }
839 loop_next:
842 }
847 }
858 }
861 }
862 /* ugly interaction with invalidatepage here.
863 * reiserfs_invalidate_page will pin any buffer that has a valid
864 * journal head from an older transaction. If someone else sets
865 * our buffer dirty after we write it in the first loop, and
866 * then someone truncates the page away, nobody will ever write
867 * the buffer. We're safe if we write the page one last time
868 * after freeing the journal header.
869 */
874 }
877 }
880 }
884 {
893 find_first:
894 /*
895 * first we walk backwards to find the oldest uncommitted transation
896 */
907 }
909 /* if we didn't find any older uncommitted transactions, return now */
912 }
925 /* list we were called with is gone, return */
929 /* the one we just flushed is gone, this means all
930 * older lists are also gone, so first_jl is no longer
931 * valid either. Go back to the beginning.
932 */
936 }
937 }
943 }
944 }
946 }
949 {
956 }
959 }
961 /*
962 ** if this journal list still has commit blocks unflushed, send them to disk.
963 **
964 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
965 ** Before the commit block can by written, every other log block must be safely on disk
966 **
967 */
970 {
972 b_blocknr_t bn;
983 }
985 /* before we can put our commit blocks on disk, we have to make sure everyone older than
986 ** us is on disk too
987 */
994 /* list disappeared during flush_older_commits. return */
996 }
997 }
999 /* make sure nobody is trying to flush this one at the same time */
1005 }
1008 /* this commit is done, exit */
1012 }
1015 }
1020 /*
1021 * We might sleep in numerous places inside
1022 * write_ordered_buffers. Relax the write lock.
1023 */
1030 }
1032 /*
1033 * for the description block and all the log blocks, submit any buffers
1034 * that haven't already reached the disk. Try to write at least 256
1035 * log blocks. later on, we will only wait on blocks that correspond
1036 * to this transaction, but while we're unplugging we might as well
1037 * get a chunk of data on there.
1038 */
1052 }
1054 }
1055 }
1066 // since we're using ll_rw_blk above, it might have skipped over
1067 // a locked buffer. Double check here
1068 //
1069 /* redundant, sync_dirty_buffer() checks */
1074 }
1076 #ifdef CONFIG_REISERFS_CHECK
1079 #endif
1081 }
1085 }
1089 /* If there was a write error in the journal - we can't commit
1090 * this transaction - it will be invalid and, if successful,
1091 * will just end up propagating the write error out to
1092 * the file system. */
1100 else
1103 }
1105 /* If there was a write error in the journal - we can't commit this
1106 * transaction - it will be invalid and, if successful, will just end
1107 * up propagating the write error out to the filesystem. */
1109 #ifdef CONFIG_REISERFS_CHECK
1111 #endif
1113 }
1119 }
1122 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1127 /* mark the metadata dirty */
1134 }
1136 put_jl:
1141 __func__);
1143 }
1145 /*
1146 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1147 ** returns NULL if it can't find anything
1148 */
1150 reiserfs_journal_cnode
1152 {
1160 }
1162 }
1164 }
1167 {
1177 }
1179 }
1186 /*
1187 ** once all the real blocks have been flushed, it is safe to remove them from the
1188 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1189 ** block to be reallocated for data blocks if it had been deleted.
1190 */
1194 {
1199 /* which is better, to lock once around the whole loop, or
1200 ** to lock for each call to remove_journal_hash?
1201 */
1209 }
1213 }
1217 }
1219 }
1221 /*
1222 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1223 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1224 ** releasing blocks in this transaction for reuse as data blocks.
1225 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1226 **
1227 */
1231 {
1244 #ifdef CONFIG_REISERFS_CHECK
1247 #endif
1249 }
1250 }
1254 b_data);
1264 else
1272 }
1273 }
1275 }
1280 {
1282 }
1284 /*
1285 ** flush any and all journal lists older than you are
1286 ** can only be called from flush_journal_list
1287 */
1290 {
1296 /* we know we are the only ones flushing things, no extra race
1297 * protection is required.
1298 */
1299 restart:
1301 /* Did we wrap? */
1307 /* do not flush all */
1310 /* other_jl is now deleted from the list */
1312 }
1314 }
1318 {
1323 }
1324 }
1326 /* flush a journal list, both commit and real blocks
1327 **
1328 ** always set flushall to 1, unless you are calling from inside
1329 ** flush_journal_list
1330 **
1331 ** IMPORTANT. This can only be called while there are no journal writers,
1332 ** and the journal is locked. That means it can only be called from
1333 ** do_journal_end, or by journal_release
1334 */
1337 {
1353 }
1356 /* if flushall == 0, the lock is already held */
1361 }
1368 }
1370 /* if all the work is already done, get out of here */
1374 }
1376 /* start by putting the commit list on disk. This will also flush
1377 ** the commit lists of any olders transactions
1378 */
1385 /* are we done now? */
1389 }
1391 /* loop through each cnode, see if we need to write it,
1392 ** or wait on a more recent transaction, or just ignore it
1393 */
1397 }
1403 /* blocknr of 0 is no longer in the hash, ignore it */
1406 }
1408 /* This transaction failed commit. Don't write out to the disk */
1413 /* the order is important here. We check pjl to make sure we
1414 ** don't clear BH_JDirty_wait if we aren't the one writing this
1415 ** block to disk
1416 */
1420 /* we do this to make sure nobody releases the buffer while
1421 ** we are working with it
1422 */
1430 /* everything with !pjl && jwait should be writable */
1432 }
1433 }
1435 /* if someone has this block in a newer transaction, just make
1436 ** sure they are committed, and don't try writing it to disk
1437 */
1442 }
1444 /* bh == NULL when the block got to disk on its own, OR,
1445 ** the block got freed in a future transaction
1446 */
1449 }
1451 /* this should never happen. kupdate_one_transaction has this list
1452 ** locked while it works, so we should never see a buffer here that
1453 ** is not marked JDirty_wait
1454 */
1457 "BAD! buffer %llu %cdirty %cjwait, "
1462 }
1464 /* we inc again because saved_bh gets decremented at free_cnode */
1471 else
1473 count++;
1479 }
1480 free_cnode:
1484 /* we incremented this to keep others from taking the buffer head away */
1489 }
1490 }
1491 }
1499 }
1508 }
1510 #ifdef CONFIG_REISERFS_CHECK
1513 #endif
1515 }
1516 /* note, we must clear the JDirty_wait bit after the up to date
1517 ** check, otherwise we race against our flushpage routine
1518 */
1522 /* drop one ref for us */
1524 /* drop one ref for journal_mark_dirty */
1526 }
1528 }
1529 }
1534 __func__);
1535 flush_older_and_return:
1537 /* before we can update the journal header block, we _must_ flush all
1538 ** real blocks from all older transactions to disk. This is because
1539 ** once the header block is updated, this transaction will not be
1540 ** replayed after a crash
1541 */
1544 }
1547 /* before we can remove everything from the hash tables for this
1548 ** transaction, we must make sure it can never be replayed
1549 **
1550 ** since we are only called from do_journal_end, we know for sure there
1551 ** are no allocations going on while we are flushing journal lists. So,
1552 ** we only need to update the journal header block for the last list
1553 ** being flushed
1554 */
1556 err =
1564 __func__);
1565 }
1575 }
1578 /* not strictly required since we are freeing the list, but it should
1579 * help find code using dead lists later on
1580 */
1592 }
1596 {
1604 /* if the blocknr == 0, this has been cleared from the hash,
1605 ** skip it
1606 */
1609 }
1612 next:
1615 }
1617 }
1622 {
1630 }
1634 /* if the blocknr == 0, this has been cleared from the hash,
1635 ** skip it
1636 */
1639 }
1642 /* we can race against journal_mark_freed when we try
1643 * to lock_buffer(cn->bh), so we have to inc the buffer
1644 * count, and recheck things after locking
1645 */
1654 ret++;
1656 /* note, cn->bh might be null now */
1658 }
1660 }
1661 next:
1664 }
1666 }
1668 /* used by flush_commit_list */
1671 {
1679 /* look for a more recent transaction that logged this
1680 ** buffer. Only the most recent transaction with a buffer in
1681 ** it is allowed to send that buffer to disk
1682 */
1687 /* if the buffer is prepared, it will either be logged
1688 * or restored. If restored, we need to make sure
1689 * it actually gets marked dirty
1690 */
1697 }
1698 }
1700 }
1702 }
1709 {
1722 }
1724 /* we've got j_flush_mutex held, nobody is going to delete any
1725 * of these lists out from underneath us
1726 */
1735 }
1740 transactions_flushed++;
1744 /* did we wrap? */
1747 }
1750 /* don't bother with older transactions */
1753 }
1756 }
1758 done:
1761 }
1763 /* for o_sync and fsync heavy applications, they tend to use
1764 ** all the journa list slots with tiny transactions. These
1765 ** trigger lots and lots of calls to update the header block, which
1766 ** adds seeks and slows things down.
1767 **
1768 ** This function tries to clear out a large chunk of the journal lists
1769 ** at once, which makes everything faster since only the newest journal
1770 ** list updates the header block
1771 */
1774 {
1787 /* in data logging mode, try harder to flush a lot of blocks */
1790 /* flush for 256 transactions or limit blocks, whichever comes first */
1795 }
1799 }
1805 }
1806 /* try to find a group of blocks we can flush across all the
1807 ** transactions, but only bother if we've actually spanned
1808 ** across multiple lists
1809 */
1812 }
1815 }
1817 /*
1818 ** removes any nodes in table with name block and dev as bh.
1819 ** only touchs the hnext and hprev pointers.
1820 */
1825 {
1832 }
1840 }
1845 }
1853 }
1855 }
1856 }
1859 {
1869 }
1870 /* j_header_bh is on the journal dev, make sure not to release the journal
1871 * dev until we brelse j_header_bh
1872 */
1875 }
1877 /*
1878 ** call on unmount. Only set error to 1 if you haven't made your way out
1879 ** of read_super() yet. Any other caller must keep error at 0.
1880 */
1883 {
1888 /* we only want to flush out transactions if we were called with error == 0
1889 */
1891 /* end the current trans */
1895 /* make sure something gets logged to force our way into the flush code */
1904 }
1905 }
1907 /* this also catches errors during the do_journal_end above */
1917 }
1918 }
1920 reiserfs_mounted_fs_count--;
1921 /* wait for all commits to finish */
1924 /*
1925 * We must release the write lock here because
1926 * the workqueue job (flush_async_commit) needs this lock
1927 */
1934 }
1941 }
1943 /*
1944 ** call on unmount. flush all journal trans, release all alloc'd ram
1945 */
1948 {
1950 }
1952 /*
1953 ** only call from an error condition inside reiserfs_read_super!
1954 */
1957 {
1959 }
1961 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1965 {
1971 }
1973 }
1975 /* returns 0 if it did not find a description block
1976 ** returns -1 if it found a corrupt commit block
1977 ** returns 1 if both desc and commit were valid
1978 */
1983 {
1998 "journal-986: transaction "
1999 "is valid returning because trans_id %d is greater than "
2004 }
2008 "journal-1087: transaction "
2009 "is valid returning because mount_id %d is less than "
2014 }
2017 "Bad transaction length %d "
2021 }
2024 /* ok, we have a journal description block, lets see if the transaction was valid */
2025 c_bh =
2035 "journal_transaction_is_valid, commit offset %ld had bad "
2046 "journal-1004: "
2047 "transaction_is_valid setting oldest invalid trans_id "
2050 }
2052 }
2055 "journal-1006: found valid "
2064 }
2065 }
2068 {
2072 }
2073 }
2075 /*
2076 ** given the start, and values for the oldest acceptable transactions,
2077 ** this either reads in a replays a transaction, or returns because the transaction
2078 ** is invalid, or too old.
2079 */
2085 {
2110 cur_dblock -
2114 }
2117 "journal_read_trans skipping because %d is != "
2119 newest_mount_id);
2122 }
2129 }
2133 "journal_read_transaction, "
2142 }
2150 }
2153 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2166 }
2167 /* get all the buffer heads */
2184 }
2187 "REPLAY FAILURE fsck required! "
2188 "Block to replay is outside of "
2191 }
2192 /* make sure we don't try to replay onto log or reserved area */
2196 "REPLAY FAILURE fsck required! "
2198 abort_replay:
2206 }
2207 }
2208 /* read in the log blocks, memcpy to the corresponding real block */
2218 "REPLAY FAILURE fsck required! "
2228 }
2233 }
2234 /* flush out the real blocks */
2238 }
2243 "REPLAY FAILURE, fsck required! "
2252 }
2254 }
2255 cur_dblock =
2263 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2267 /* check for trans_id overflow */
2275 }
2277 /* This function reads blocks starting from block and to max_block of bufsize
2278 size (but no more than BUFNR blocks at a time). This proved to improve
2279 mounting speed on self-rebuilding raid5 arrays at least.
2280 Right now it is only used from journal code. But later we might use it
2281 from other places.
2282 Note: Do not use journal_getblk/sb_getblk functions here! */
2285 b_blocknr_t max_block)
2286 {
2298 }
2308 }
2318 }
2320 /*
2321 ** read and replay the log
2322 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2323 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2324 **
2325 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2326 **
2327 ** On exit, it sets things up so the first transaction will work correctly.
2328 */
2330 {
2352 /* step 1, read in the journal header block. Check the transaction it says
2353 ** is the first unflushed, and if that transaction is not valid,
2354 ** replay is done
2355 */
2361 }
2366 oldest_start =
2372 "journal-1153: found in "
2373 "header: first_unflushed_offset %d, last_flushed_trans_id "
2378 /* now, we try to read the first unflushed offset. If it is not valid,
2379 ** there is nothing more we can do, and it makes no sense to read
2380 ** through the whole log.
2381 */
2382 d_bh =
2389 }
2392 }
2394 /* ok, there are transactions that need to be replayed. start with the first log block, find
2395 ** all the valid transactions, and pick out the oldest.
2396 */
2401 /* Note that it is required for blocksize of primary fs device and journal
2402 device to be the same */
2403 d_bh =
2408 ret =
2419 "journal-1179: Setting "
2421 oldest_start -
2422 SB_ONDISK_JOURNAL_1st_BLOCK
2425 /* one we just read was older */
2429 "journal-1180: Resetting "
2431 oldest_start -
2432 SB_ONDISK_JOURNAL_1st_BLOCK
2434 }
2438 "journal-1299: Setting "
2441 }
2444 cur_dblock++;
2445 }
2447 }
2449 start_log_replay:
2453 "journal-1206: Starting replay "
2456 oldest_trans_id);
2458 }
2461 ret =
2468 }
2469 cur_dblock =
2471 replay_count++;
2474 }
2479 }
2480 /* j_start does not get set correctly if we don't replay any transactions.
2481 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2482 ** copy the trans_id from the header
2483 */
2488 /* check for trans_id overflow */
2496 }
2504 }
2508 /* replay failed, caller must call free_journal_ram and abort
2509 ** the mount
2510 */
2512 }
2514 }
2517 {
2529 }
2532 {
2534 }
2538 {
2546 }
2551 }
2553 }
2558 {
2560 dev_t jdev;
2573 /* there is no "jdev" option and journal is on separate device */
2578 journal);
2591 }
2602 }
2609 }
2611 /**
2612 * When creating/tuning a file system user can assign some
2613 * journal params within boundaries which depend on the ratio
2614 * blocksize/standard_blocksize.
2615 *
2616 * For blocks >= standard_blocksize transaction size should
2617 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2618 * then JOURNAL_TRANS_MAX_DEFAULT.
2619 *
2620 * For blocks < standard_blocksize these boundaries should be
2621 * decreased proportionally.
2622 */
2623 #define REISERFS_STANDARD_BLKSIZE (4096)
2627 {
2629 /* Non-default journal params.
2630 Do sanity check for them. */
2638 JOURNAL_MIN_RATIO) {
2640 "bad transaction max size (%u). "
2643 }
2647 "bad transaction max batch (%u). "
2650 }
2652 /* Default journal params.
2653 The file system was created by old version
2654 of mkreiserfs, so some fields contain zeros,
2655 and we need to advise proper values for them */
2660 }
2664 }
2666 }
2668 /*
2669 ** must be called once on fs mount. calls journal_read for you
2670 */
2673 {
2683 /*
2684 * Unlock here to avoid various RECLAIM-FS-ON <-> IN-RECLAIM-FS
2685 * dependency inversion warnings.
2686 */
2694 }
2709 /* reserved for journal area support */
2711 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2715 REISERFS_DISK_OFFSET_IN_BYTES /
2718 /* Sanity check to see is the standard journal fitting within first bitmap
2719 (actual for small blocksizes) */
2724 "journal does not fit for area addressed "
2725 "by first of bitmap blocks. It starts at "
2731 }
2733 /*
2734 * We need to unlock here to avoid creating the following
2735 * dependency:
2736 * reiserfs_lock -> sysfs_mutex
2737 * Because the reiserfs mmap path creates the following dependency:
2738 * mm->mmap -> reiserfs_lock, hence we have
2739 * mm->mmap -> reiserfs_lock ->sysfs_mutex
2740 * This would ends up in a circular dependency with sysfs readdir path
2741 * which does sysfs_mutex -> mm->mmap_sem
2742 * This is fine because the reiserfs lock is useless in mount path,
2743 * at least until we call journal_begin. We keep it for paranoid
2744 * reasons.
2745 */
2752 }
2757 /* read journal header */
2765 }
2768 /* make sure that journal matches to the super block */
2773 "journal header magic %x (device %s) does "
2780 }
2795 }
2798 "journal first block %u, max trans len %u, max batch %u, "
2845 "allocation failed (%ld bytes). Journal is "
2846 "too large for available memory. Usually "
2850 }
2859 }
2864 }
2866 reiserfs_mounted_fs_count++;
2871 }
2876 free_and_return:
2879 }
2881 /*
2882 ** test for a polite end of the current transaction. Used by file_write, and should
2883 ** be used by delete to make sure they don't write more than can fit inside a single
2884 ** transaction
2885 */
2888 {
2891 /* cannot restart while nested */
2901 }
2902 /* protected by the BKL here */
2906 }
2908 /* this must be called inside a transaction, and requires the
2909 ** kernel_lock to be held
2910 */
2912 {
2918 }
2920 /* this must be called without a transaction started, and does not
2921 ** require BKL
2922 */
2924 {
2928 }
2930 /* this must be called without a transaction started, and does not
2931 ** require BKL
2932 */
2934 {
2938 }
2941 {
2942 wait_queue_t wait;
2946 /*
2947 * we don't want to use wait_event here because
2948 * we only want to wait once.
2949 */
2957 }
2960 }
2963 {
2967 }
2970 {
2982 }
2988 }
2989 }
2991 /* join == true if you must join an existing transaction.
2992 ** join == false if you can deal with waiting for others to finish
2993 **
2994 ** this will block until the transaction is joinable. send the number of blocks you
2995 ** expect to use in nblocks.
2996 */
3000 {
3012 /* set here for journal_join */
3016 relock:
3022 }
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 }
3103 out_fail:
3105 /* Re-set th->t_super, so we can properly keep track of how many
3106 * persistent transactions there are. We need to do this so if this
3107 * call is part of a failed restart_transaction, we can free it later */
3110 }
3113 super_block
3116 {
3120 /* if we're nesting into an existing transaction. It will be
3121 ** persistent on its own
3122 */
3129 }
3137 }
3141 }
3144 {
3149 else
3154 }
3156 }
3160 {
3163 /* this keeps do_journal_end from NULLing out the current->journal_info
3164 ** pointer
3165 */
3169 }
3173 {
3176 /* this keeps do_journal_end from NULLing out the current->journal_info
3177 ** pointer
3178 */
3182 }
3186 {
3192 /* we are nesting into the current transaction */
3199 "BAD: refcount <= 1, but "
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 */
3270 "buffer %llu bad state "
3277 }
3284 }
3285 /* this error means I've screwed up, and we've overflowed the transaction.
3286 ** Nothing can be done here, except make the FS readonly or panic.
3287 */
3292 }
3298 }
3302 }
3306 /* now put this guy on the end */
3311 }
3316 }
3327 }
3328 }
3338 }
3340 }
3344 {
3352 }
3359 /* we aren't allowed to close a nested transaction on a different
3360 ** filesystem from the one in the task struct
3361 */
3367 }
3371 }
3372 }
3374 /* removes from the current transaction, relsing and descrementing any counters.
3375 ** also files the removed buffer directly onto the clean list
3376 **
3377 ** called by journal_mark_freed when a block has been deleted
3378 **
3379 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3380 */
3383 {
3392 }
3396 }
3399 }
3402 }
3405 }
3419 }
3421 }
3426 }
3428 /*
3429 ** for any cnode in a journal list, it can only be dirtied of all the
3430 ** transactions that include it are committed to disk.
3431 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3432 ** and 0 if you aren't
3433 **
3434 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3435 ** blocks for a given transaction on disk
3436 **
3437 */
3439 {
3445 /* first test hprev. These are all newer than cn, so any node here
3446 ** with the same block number and dev means this node can't be sent
3447 ** to disk right now.
3448 */
3453 }
3455 }
3456 /* then test hnext. These are all older than cn. As long as they
3457 ** are committed to the log, it is safe to write cn to disk
3458 */
3465 }
3467 }
3469 }
3471 /* syncs the commit blocks, but does not force the real buffers to disk
3472 ** will wait until the current transaction is done/committed before returning
3473 */
3476 {
3480 /* you can sync while nested, very, very bad */
3486 }
3488 }
3490 /*
3491 ** writeback the pending async commits to disk
3492 */
3494 {
3503 /* last entry is the youngest, commit it and you get everything */
3507 }
3509 }
3511 /*
3512 ** flushes any old transactions to disk
3513 ** ends the current transaction if it is too old
3514 */
3516 {
3522 /* safety check so we don't flush while we are replaying the log during
3523 * mount
3524 */
3527 }
3529 /* check the current transaction. If there are no writers, and it is
3530 * too old, finish it, and force the commit blocks to disk
3531 */
3543 /* we're only being called from kreiserfsd, it makes no sense to do
3544 ** an async commit so that kreiserfsd can do it later
3545 */
3547 }
3548 }
3550 }
3552 /*
3553 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3554 **
3555 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3556 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3557 ** flushes the commit list and returns 0.
3558 **
3559 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3560 **
3561 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3562 */
3566 {
3581 }
3584 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3586 }
3588 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3589 ** will be dealt with by next transaction that actually writes something, but should be taken
3590 ** care of in this trans
3591 */
3594 /* if wcount > 0, and we are called to with flush or commit_now,
3595 ** we wait on j_join_wait. We will wake up when the last writer has
3596 ** finished the transaction, and started it on its way to the disk.
3597 ** Then, we flush the commit or journal list, and just return 0
3598 ** because the rest of journal end was already done for this transaction.
3599 */
3611 }
3614 /* sleep while the current transaction is still j_jlocked */
3623 }
3625 }
3626 }
3633 }
3635 }
3638 }
3640 /* deal with old transactions where we are the last writers */
3645 }
3646 /* don't batch when someone is waiting on j_join_wait */
3647 /* don't batch when syncing the commit or flushing the whole trans */
3655 }
3661 }
3663 }
3665 /*
3666 ** Does all the work that makes deleting blocks safe.
3667 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3668 **
3669 ** otherwise:
3670 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3671 ** before this transaction has finished.
3672 **
3673 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3674 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3675 ** the block can't be reallocated yet.
3676 **
3677 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3678 */
3681 {
3693 }
3694 /* if it is journal new, we just remove it from this transaction */
3701 /* set the bit for this block in the journal bitmap for this transaction */
3706 }
3709 /* Note, the entire while loop is not allowed to schedule. */
3714 }
3717 /* find all older transactions with this block, make sure they don't try to write it out */
3719 blocknr);
3725 /* remove_from_transaction will brelse the buffer if it was
3726 ** in the current trans
3727 */
3729 bh);
3732 bh);
3740 }
3741 }
3745 j_nonzerolen));
3746 }
3748 }
3749 }
3751 }
3752 }
3757 }
3760 {
3764 }
3766 /*
3767 * returns -1 on error, 0 if no commits/barriers were done and 1
3768 * if a transaction was actually committed and the barrier was done
3769 */
3772 {
3778 /* is it from the current transaction, or from an unknown transaction? */
3781 /* try to let other writers come in and grow this transaction */
3785 }
3791 /* someone might have ended this transaction while we joined */
3798 }
3805 /* this gets tricky, we have to make sure the journal list in
3806 * the inode still exists. We know the list is still around
3807 * if we've got a larger transaction id than the oldest list
3808 */
3809 flush_commit_only:
3811 /*
3812 * we only set ret to 1 when we know for sure
3813 * the barrier hasn't been started yet on the commit
3814 * block.
3815 */
3821 }
3822 }
3823 /* otherwise the list is gone, and long since committed */
3825 }
3828 {
3832 /* for the whole inode, assume unset id means it was
3833 * changed in the current transaction. More conservative
3834 */
3838 /* jl will be updated in __commit_trans_jl */
3839 }
3842 }
3846 {
3851 }
3861 }
3862 }
3864 }
3867 /*
3868 ** before we can change a metadata block, we have to make sure it won't
3869 ** be written to disk while we are altering it. So, we must:
3870 ** clean it
3871 ** wait on it.
3872 **
3873 */
3876 {
3883 }
3888 }
3891 }
3894 {
3903 /* this check should always be run, to send old lists to disk */
3910 }
3911 }
3912 }
3914 /*
3915 ** long and ugly. If flush, will not return until all commit
3916 ** blocks and all real buffers in the trans are on disk.
3917 ** If no_async, won't return until all commit blocks are on disk.
3918 **
3919 ** keep reading, there are comments as you go along
3920 **
3921 ** If the journal is aborted, we just clean up. Things like flushing
3922 ** journal lists, etc just won't happen.
3923 */
3927 {
3949 /* protect flush_older_commits from doing mistakes if the
3950 transaction ID counter gets overflowed. */
3962 }
3968 }
3972 }
3974 /* check_journal_end locks the journal, and unlocks if it does not return 1
3975 ** it tells us if we should continue with the journal_end, or just return
3976 */
3982 }
3984 /* check_journal_end might set these, check again */
3987 }
3989 /*
3990 ** j must wait means we have to flush the log blocks, and the real blocks for
3991 ** this transaction
3992 */
3995 }
3996 #ifdef REISERFS_PREALLOCATE
3997 /* quota ops might need to nest, setup the journal_info pointer for them
3998 * and raise the refcount so that it is > 0. */
4002 * the transaction */
4005 #endif
4007 /* setup description block */
4008 d_bh =
4018 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4027 /* init this journal list */
4030 /* we lock the commit before doing anything because
4031 * we want to make sure nobody tries to run flush_commit_list until
4032 * the new transaction is fully setup, and we've already flushed the
4033 * ordered bh list
4034 */
4037 /* save the transaction id in case we need to commit it later */
4050 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4051 ** for each real block, add it to the journal list hash,
4052 ** copy into real block index array in the commit or desc block
4053 */
4061 }
4064 }
4069 }
4071 /* make sure the block we are trying to log is not a block
4072 of journal or reserved area */
4077 "Trying to log block %lu, "
4080 }
4093 }
4095 i--;
4096 }
4097 }
4103 /* special check in case all buffers in the journal were marked for not logging */
4106 /* we're about to dirty all the log blocks, mark the description block
4107 * dirty now too. Don't mark the commit block dirty until all the
4108 * others are on disk
4109 */
4112 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4118 /* copy all the real blocks into log area. dirty log blocks */
4123 tmp_bh =
4127 jindex) %
4137 jindex++;
4141 /* JDirty cleared sometime during transaction. don't log this one */
4143 "BAD, buffer in journal hash, "
4146 }
4153 }
4155 /* we are done with both the c_bh and d_bh, but
4156 ** c_bh must be written after all other commit blocks,
4157 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4158 */
4162 /* now it is safe to insert this transaction on the main list */
4167 /* reset journal values for the next transaction */
4178 /* check for trans_id overflow */
4188 // make sure reiserfs_add_jh sees the new current_jl before we
4189 // write out the tails
4192 /* tail conversion targets have to hit the disk before we end the
4193 * transaction. Otherwise a later transaction might repack the tail
4194 * before this transaction commits, leaving the data block unflushed and
4195 * clean, if we crash before the later transaction commits, the data block
4196 * is lost.
4197 */
4203 }
4207 /* honor the flush wishes from the caller, simple commits can
4208 ** be done outside the journal lock, they are done below
4209 **
4210 ** if we don't flush the commit list right now, we put it into
4211 ** the work queue so the people waiting on the async progress work
4212 ** queue don't wait for this proc to flush journal lists and such.
4213 */
4220 /* if the next transaction has any chance of wrapping, flush
4221 ** transactions that might get overwritten. If any journal lists are very
4222 ** old flush them as well.
4223 */
4224 first_jl:
4235 /* if we don't cross into the next transaction and we don't
4236 * wrap, there is no way we can overlap any later transactions
4237 * break now
4238 */
4240 }
4250 /* we don't overlap anything from out start to the end of the
4251 * log, and our wrapped portion doesn't overlap anything at
4252 * the start of the log. We can break
4253 */
4255 }
4256 }
4257 }
4266 }
4270 /* wake up any body waiting to join. */
4277 }
4278 out:
4282 /* Re-set th->t_super, so we can properly keep track of how many
4283 * persistent transactions there are. We need to do this so if this
4284 * call is part of a failed restart_transaction, we can free it later */
4288 }
4290 /* Send the file system read only and refuse new transactions */
4292 {
4303 #ifdef CONFIG_REISERFS_CHECK
4305 #endif
4306 }