| blob | c5e82ece7c6c7d7afd2002d15c4293a99662bb4c |
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 {
684 }
687 }
690 {
695 }
698 }
702 {
713 }
715 }
719 {
726 }
728 }
729 }
731 /*
732 * we want to free the jh when the buffer has been written
733 * and waited on
734 */
736 {
749 }
750 }
754 {
762 }
766 no_jh:
770 /* buffer must be locked for __add_jh, should be able to have
771 * two adds at the same time
772 */
776 }
782 }
785 }
788 {
790 }
792 {
794 }
796 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
801 {
819 }
827 }
828 /* in theory, dirty non-uptodate buffers should never get here,
829 * but the upper layer io error paths still have a few quirks.
830 * Handle them here as gracefully as we can
831 */
835 }
842 }
843 loop_next:
846 }
851 }
862 }
865 }
866 /* ugly interaction with invalidatepage here.
867 * reiserfs_invalidate_page will pin any buffer that has a valid
868 * journal head from an older transaction. If someone else sets
869 * our buffer dirty after we write it in the first loop, and
870 * then someone truncates the page away, nobody will ever write
871 * the buffer. We're safe if we write the page one last time
872 * after freeing the journal header.
873 */
878 }
881 }
884 }
888 {
897 find_first:
898 /*
899 * first we walk backwards to find the oldest uncommitted transation
900 */
911 }
913 /* if we didn't find any older uncommitted transactions, return now */
916 }
929 /* list we were called with is gone, return */
933 /* the one we just flushed is gone, this means all
934 * older lists are also gone, so first_jl is no longer
935 * valid either. Go back to the beginning.
936 */
940 }
941 }
947 }
948 }
950 }
953 {
960 }
963 }
965 /*
966 ** if this journal list still has commit blocks unflushed, send them to disk.
967 **
968 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
969 ** Before the commit block can by written, every other log block must be safely on disk
970 **
971 */
974 {
976 b_blocknr_t bn;
987 }
991 /* before we can put our commit blocks on disk, we have to make sure everyone older than
992 ** us is on disk too
993 */
1000 /* list disappeared during flush_older_commits. return */
1002 }
1003 }
1005 /* make sure nobody is trying to flush this one at the same time */
1011 }
1014 /* this commit is done, exit */
1018 }
1021 }
1026 /*
1027 * We might sleep in numerous places inside
1028 * write_ordered_buffers. Relax the write lock.
1029 */
1036 }
1038 /*
1039 * for the description block and all the log blocks, submit any buffers
1040 * that haven't already reached the disk. Try to write at least 256
1041 * log blocks. later on, we will only wait on blocks that correspond
1042 * to this transaction, but while we're unplugging we might as well
1043 * get a chunk of data on there.
1044 */
1058 }
1060 }
1061 }
1072 // since we're using ll_rw_blk above, it might have skipped over
1073 // a locked buffer. Double check here
1074 //
1075 /* redundant, sync_dirty_buffer() checks */
1080 }
1082 #ifdef CONFIG_REISERFS_CHECK
1085 #endif
1087 }
1091 }
1095 /* If there was a write error in the journal - we can't commit
1096 * this transaction - it will be invalid and, if successful,
1097 * will just end up propagating the write error out to
1098 * the file system. */
1106 else
1109 }
1111 /* If there was a write error in the journal - we can't commit this
1112 * transaction - it will be invalid and, if successful, will just end
1113 * up propagating the write error out to the filesystem. */
1115 #ifdef CONFIG_REISERFS_CHECK
1117 #endif
1119 }
1125 }
1128 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1133 /* mark the metadata dirty */
1140 }
1142 put_jl:
1147 __func__);
1150 }
1152 /*
1153 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1154 ** returns NULL if it can't find anything
1155 */
1157 reiserfs_journal_cnode
1159 {
1167 }
1169 }
1171 }
1174 {
1184 }
1186 }
1193 /*
1194 ** once all the real blocks have been flushed, it is safe to remove them from the
1195 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1196 ** block to be reallocated for data blocks if it had been deleted.
1197 */
1201 {
1206 /* which is better, to lock once around the whole loop, or
1207 ** to lock for each call to remove_journal_hash?
1208 */
1216 }
1220 }
1224 }
1226 }
1228 /*
1229 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1230 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1231 ** releasing blocks in this transaction for reuse as data blocks.
1232 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1233 **
1234 */
1238 {
1251 #ifdef CONFIG_REISERFS_CHECK
1254 #endif
1256 }
1257 }
1261 b_data);
1271 else
1279 }
1280 }
1282 }
1287 {
1289 }
1291 /*
1292 ** flush any and all journal lists older than you are
1293 ** can only be called from flush_journal_list
1294 */
1297 {
1303 /* we know we are the only ones flushing things, no extra race
1304 * protection is required.
1305 */
1306 restart:
1308 /* Did we wrap? */
1314 /* do not flush all */
1317 /* other_jl is now deleted from the list */
1319 }
1321 }
1325 {
1330 }
1331 }
1333 /* flush a journal list, both commit and real blocks
1334 **
1335 ** always set flushall to 1, unless you are calling from inside
1336 ** flush_journal_list
1337 **
1338 ** IMPORTANT. This can only be called while there are no journal writers,
1339 ** and the journal is locked. That means it can only be called from
1340 ** do_journal_end, or by journal_release
1341 */
1344 {
1360 }
1363 /* if flushall == 0, the lock is already held */
1368 }
1375 }
1379 /* if all the work is already done, get out of here */
1383 }
1385 /* start by putting the commit list on disk. This will also flush
1386 ** the commit lists of any olders transactions
1387 */
1394 /* are we done now? */
1398 }
1400 /* loop through each cnode, see if we need to write it,
1401 ** or wait on a more recent transaction, or just ignore it
1402 */
1406 }
1412 /* blocknr of 0 is no longer in the hash, ignore it */
1415 }
1417 /* This transaction failed commit. Don't write out to the disk */
1422 /* the order is important here. We check pjl to make sure we
1423 ** don't clear BH_JDirty_wait if we aren't the one writing this
1424 ** block to disk
1425 */
1429 /* we do this to make sure nobody releases the buffer while
1430 ** we are working with it
1431 */
1439 /* everything with !pjl && jwait should be writable */
1441 }
1442 }
1444 /* if someone has this block in a newer transaction, just make
1445 ** sure they are committed, and don't try writing it to disk
1446 */
1451 }
1453 /* bh == NULL when the block got to disk on its own, OR,
1454 ** the block got freed in a future transaction
1455 */
1458 }
1460 /* this should never happen. kupdate_one_transaction has this list
1461 ** locked while it works, so we should never see a buffer here that
1462 ** is not marked JDirty_wait
1463 */
1466 "BAD! buffer %llu %cdirty %cjwait, "
1471 }
1473 /* we inc again because saved_bh gets decremented at free_cnode */
1480 else
1482 count++;
1488 }
1489 free_cnode:
1493 /* we incremented this to keep others from taking the buffer head away */
1498 }
1499 }
1500 }
1508 }
1517 }
1519 #ifdef CONFIG_REISERFS_CHECK
1522 #endif
1524 }
1525 /* note, we must clear the JDirty_wait bit after the up to date
1526 ** check, otherwise we race against our flushpage routine
1527 */
1531 /* drop one ref for us */
1533 /* drop one ref for journal_mark_dirty */
1535 }
1537 }
1538 }
1543 __func__);
1544 flush_older_and_return:
1546 /* before we can update the journal header block, we _must_ flush all
1547 ** real blocks from all older transactions to disk. This is because
1548 ** once the header block is updated, this transaction will not be
1549 ** replayed after a crash
1550 */
1553 }
1556 /* before we can remove everything from the hash tables for this
1557 ** transaction, we must make sure it can never be replayed
1558 **
1559 ** since we are only called from do_journal_end, we know for sure there
1560 ** are no allocations going on while we are flushing journal lists. So,
1561 ** we only need to update the journal header block for the last list
1562 ** being flushed
1563 */
1565 err =
1573 __func__);
1574 }
1584 }
1587 /* not strictly required since we are freeing the list, but it should
1588 * help find code using dead lists later on
1589 */
1602 }
1606 {
1614 /* if the blocknr == 0, this has been cleared from the hash,
1615 ** skip it
1616 */
1619 }
1622 next:
1625 }
1627 }
1632 {
1640 }
1644 /* if the blocknr == 0, this has been cleared from the hash,
1645 ** skip it
1646 */
1649 }
1652 /* we can race against journal_mark_freed when we try
1653 * to lock_buffer(cn->bh), so we have to inc the buffer
1654 * count, and recheck things after locking
1655 */
1664 ret++;
1666 /* note, cn->bh might be null now */
1668 }
1670 }
1671 next:
1674 }
1676 }
1678 /* used by flush_commit_list */
1681 {
1689 /* look for a more recent transaction that logged this
1690 ** buffer. Only the most recent transaction with a buffer in
1691 ** it is allowed to send that buffer to disk
1692 */
1697 /* if the buffer is prepared, it will either be logged
1698 * or restored. If restored, we need to make sure
1699 * it actually gets marked dirty
1700 */
1707 }
1708 }
1710 }
1712 }
1719 {
1732 }
1734 /* we've got j_flush_mutex held, nobody is going to delete any
1735 * of these lists out from underneath us
1736 */
1745 }
1750 transactions_flushed++;
1754 /* did we wrap? */
1757 }
1760 /* don't bother with older transactions */
1763 }
1766 }
1768 done:
1771 }
1773 /* for o_sync and fsync heavy applications, they tend to use
1774 ** all the journa list slots with tiny transactions. These
1775 ** trigger lots and lots of calls to update the header block, which
1776 ** adds seeks and slows things down.
1777 **
1778 ** This function tries to clear out a large chunk of the journal lists
1779 ** at once, which makes everything faster since only the newest journal
1780 ** list updates the header block
1781 */
1784 {
1797 /* in data logging mode, try harder to flush a lot of blocks */
1800 /* flush for 256 transactions or limit blocks, whichever comes first */
1805 }
1809 }
1815 }
1816 /* try to find a group of blocks we can flush across all the
1817 ** transactions, but only bother if we've actually spanned
1818 ** across multiple lists
1819 */
1822 }
1825 }
1827 /*
1828 ** removes any nodes in table with name block and dev as bh.
1829 ** only touchs the hnext and hprev pointers.
1830 */
1835 {
1842 }
1850 }
1855 }
1863 }
1865 }
1866 }
1869 {
1879 }
1880 /* j_header_bh is on the journal dev, make sure not to release the journal
1881 * dev until we brelse j_header_bh
1882 */
1885 }
1887 /*
1888 ** call on unmount. Only set error to 1 if you haven't made your way out
1889 ** of read_super() yet. Any other caller must keep error at 0.
1890 */
1893 {
1898 /* we only want to flush out transactions if we were called with error == 0
1899 */
1901 /* end the current trans */
1905 /* make sure something gets logged to force our way into the flush code */
1914 }
1915 }
1917 /* this also catches errors during the do_journal_end above */
1927 }
1928 }
1930 reiserfs_mounted_fs_count--;
1931 /* wait for all commits to finish */
1934 /*
1935 * We must release the write lock here because
1936 * the workqueue job (flush_async_commit) needs this lock
1937 */
1944 }
1951 }
1953 /*
1954 ** call on unmount. flush all journal trans, release all alloc'd ram
1955 */
1958 {
1960 }
1962 /*
1963 ** only call from an error condition inside reiserfs_read_super!
1964 */
1967 {
1969 }
1971 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1975 {
1981 }
1983 }
1985 /* returns 0 if it did not find a description block
1986 ** returns -1 if it found a corrupt commit block
1987 ** returns 1 if both desc and commit were valid
1988 */
1993 {
2008 "journal-986: transaction "
2009 "is valid returning because trans_id %d is greater than "
2014 }
2018 "journal-1087: transaction "
2019 "is valid returning because mount_id %d is less than "
2024 }
2027 "Bad transaction length %d "
2031 }
2034 /* ok, we have a journal description block, lets see if the transaction was valid */
2035 c_bh =
2045 "journal_transaction_is_valid, commit offset %ld had bad "
2056 "journal-1004: "
2057 "transaction_is_valid setting oldest invalid trans_id "
2060 }
2062 }
2065 "journal-1006: found valid "
2074 }
2075 }
2078 {
2082 }
2083 }
2085 /*
2086 ** given the start, and values for the oldest acceptable transactions,
2087 ** this either reads in a replays a transaction, or returns because the transaction
2088 ** is invalid, or too old.
2089 */
2095 {
2120 cur_dblock -
2124 }
2127 "journal_read_trans skipping because %d is != "
2129 newest_mount_id);
2132 }
2139 }
2143 "journal_read_transaction, "
2152 }
2160 }
2163 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2176 }
2177 /* get all the buffer heads */
2194 }
2197 "REPLAY FAILURE fsck required! "
2198 "Block to replay is outside of "
2201 }
2202 /* make sure we don't try to replay onto log or reserved area */
2206 "REPLAY FAILURE fsck required! "
2208 abort_replay:
2216 }
2217 }
2218 /* read in the log blocks, memcpy to the corresponding real block */
2228 "REPLAY FAILURE fsck required! "
2238 }
2243 }
2244 /* flush out the real blocks */
2248 }
2253 "REPLAY FAILURE, fsck required! "
2262 }
2264 }
2265 cur_dblock =
2273 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2277 /* check for trans_id overflow */
2285 }
2287 /* This function reads blocks starting from block and to max_block of bufsize
2288 size (but no more than BUFNR blocks at a time). This proved to improve
2289 mounting speed on self-rebuilding raid5 arrays at least.
2290 Right now it is only used from journal code. But later we might use it
2291 from other places.
2292 Note: Do not use journal_getblk/sb_getblk functions here! */
2295 b_blocknr_t max_block)
2296 {
2308 }
2318 }
2328 }
2330 /*
2331 ** read and replay the log
2332 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2333 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2334 **
2335 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2336 **
2337 ** On exit, it sets things up so the first transaction will work correctly.
2338 */
2340 {
2362 /* step 1, read in the journal header block. Check the transaction it says
2363 ** is the first unflushed, and if that transaction is not valid,
2364 ** replay is done
2365 */
2371 }
2376 oldest_start =
2382 "journal-1153: found in "
2383 "header: first_unflushed_offset %d, last_flushed_trans_id "
2388 /* now, we try to read the first unflushed offset. If it is not valid,
2389 ** there is nothing more we can do, and it makes no sense to read
2390 ** through the whole log.
2391 */
2392 d_bh =
2399 }
2402 }
2404 /* ok, there are transactions that need to be replayed. start with the first log block, find
2405 ** all the valid transactions, and pick out the oldest.
2406 */
2411 /* Note that it is required for blocksize of primary fs device and journal
2412 device to be the same */
2413 d_bh =
2418 ret =
2429 "journal-1179: Setting "
2431 oldest_start -
2432 SB_ONDISK_JOURNAL_1st_BLOCK
2435 /* one we just read was older */
2439 "journal-1180: Resetting "
2441 oldest_start -
2442 SB_ONDISK_JOURNAL_1st_BLOCK
2444 }
2448 "journal-1299: Setting "
2451 }
2454 cur_dblock++;
2455 }
2457 }
2459 start_log_replay:
2463 "journal-1206: Starting replay "
2466 oldest_trans_id);
2468 }
2471 ret =
2478 }
2479 cur_dblock =
2481 replay_count++;
2484 }
2489 }
2490 /* j_start does not get set correctly if we don't replay any transactions.
2491 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2492 ** copy the trans_id from the header
2493 */
2498 /* check for trans_id overflow */
2506 }
2514 }
2518 /* replay failed, caller must call free_journal_ram and abort
2519 ** the mount
2520 */
2522 }
2524 }
2527 {
2539 }
2542 {
2544 }
2548 {
2556 }
2561 }
2563 }
2568 {
2570 dev_t jdev;
2583 /* there is no "jdev" option and journal is on separate device */
2588 journal);
2601 }
2612 }
2619 }
2621 /**
2622 * When creating/tuning a file system user can assign some
2623 * journal params within boundaries which depend on the ratio
2624 * blocksize/standard_blocksize.
2625 *
2626 * For blocks >= standard_blocksize transaction size should
2627 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2628 * then JOURNAL_TRANS_MAX_DEFAULT.
2629 *
2630 * For blocks < standard_blocksize these boundaries should be
2631 * decreased proportionally.
2632 */
2633 #define REISERFS_STANDARD_BLKSIZE (4096)
2637 {
2639 /* Non-default journal params.
2640 Do sanity check for them. */
2648 JOURNAL_MIN_RATIO) {
2650 "bad transaction max size (%u). "
2653 }
2657 "bad transaction max batch (%u). "
2660 }
2662 /* Default journal params.
2663 The file system was created by old version
2664 of mkreiserfs, so some fields contain zeros,
2665 and we need to advise proper values for them */
2670 }
2674 }
2676 }
2678 /*
2679 ** must be called once on fs mount. calls journal_read for you
2680 */
2683 {
2693 /*
2694 * Unlock here to avoid various RECLAIM-FS-ON <-> IN-RECLAIM-FS
2695 * dependency inversion warnings.
2696 */
2704 }
2719 /* reserved for journal area support */
2721 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2725 REISERFS_DISK_OFFSET_IN_BYTES /
2728 /* Sanity check to see is the standard journal fitting within first bitmap
2729 (actual for small blocksizes) */
2734 "journal does not fit for area addressed "
2735 "by first of bitmap blocks. It starts at "
2741 }
2743 /*
2744 * We need to unlock here to avoid creating the following
2745 * dependency:
2746 * reiserfs_lock -> sysfs_mutex
2747 * Because the reiserfs mmap path creates the following dependency:
2748 * mm->mmap -> reiserfs_lock, hence we have
2749 * mm->mmap -> reiserfs_lock ->sysfs_mutex
2750 * This would ends up in a circular dependency with sysfs readdir path
2751 * which does sysfs_mutex -> mm->mmap_sem
2752 * This is fine because the reiserfs lock is useless in mount path,
2753 * at least until we call journal_begin. We keep it for paranoid
2754 * reasons.
2755 */
2762 }
2767 /* read journal header */
2775 }
2778 /* make sure that journal matches to the super block */
2783 "journal header magic %x (device %s) does "
2790 }
2805 }
2808 "journal first block %u, max trans len %u, max batch %u, "
2855 "allocation failed (%ld bytes). Journal is "
2856 "too large for available memory. Usually "
2860 }
2869 }
2874 }
2876 reiserfs_mounted_fs_count++;
2881 }
2886 free_and_return:
2889 }
2891 /*
2892 ** test for a polite end of the current transaction. Used by file_write, and should
2893 ** be used by delete to make sure they don't write more than can fit inside a single
2894 ** transaction
2895 */
2898 {
2901 /* cannot restart while nested */
2911 }
2912 /* protected by the BKL here */
2916 }
2918 /* this must be called inside a transaction, and requires the
2919 ** kernel_lock to be held
2920 */
2922 {
2928 }
2930 /* this must be called without a transaction started, and does not
2931 ** require BKL
2932 */
2934 {
2938 }
2940 /* this must be called without a transaction started, and does not
2941 ** require BKL
2942 */
2944 {
2948 }
2951 {
2952 wait_queue_t wait;
2956 /*
2957 * we don't want to use wait_event here because
2958 * we only want to wait once.
2959 */
2967 }
2970 }
2973 {
2977 }
2980 {
2992 }
2998 }
2999 }
3001 /* join == true if you must join an existing transaction.
3002 ** join == false if you can deal with waiting for others to finish
3003 **
3004 ** this will block until the transaction is joinable. send the number of blocks you
3005 ** expect to use in nblocks.
3006 */
3010 {
3022 /* set here for journal_join */
3026 relock:
3032 }
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 }
4165 }
4167 /* we are done with both the c_bh and d_bh, but
4168 ** c_bh must be written after all other commit blocks,
4169 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4170 */
4174 /* now it is safe to insert this transaction on the main list */
4179 /* reset journal values for the next transaction */
4190 /* check for trans_id overflow */
4200 // make sure reiserfs_add_jh sees the new current_jl before we
4201 // write out the tails
4204 /* tail conversion targets have to hit the disk before we end the
4205 * transaction. Otherwise a later transaction might repack the tail
4206 * before this transaction commits, leaving the data block unflushed and
4207 * clean, if we crash before the later transaction commits, the data block
4208 * is lost.
4209 */
4215 }
4219 /* honor the flush wishes from the caller, simple commits can
4220 ** be done outside the journal lock, they are done below
4221 **
4222 ** if we don't flush the commit list right now, we put it into
4223 ** the work queue so the people waiting on the async progress work
4224 ** queue don't wait for this proc to flush journal lists and such.
4225 */
4232 /* if the next transaction has any chance of wrapping, flush
4233 ** transactions that might get overwritten. If any journal lists are very
4234 ** old flush them as well.
4235 */
4236 first_jl:
4247 /* if we don't cross into the next transaction and we don't
4248 * wrap, there is no way we can overlap any later transactions
4249 * break now
4250 */
4252 }
4262 /* we don't overlap anything from out start to the end of the
4263 * log, and our wrapped portion doesn't overlap anything at
4264 * the start of the log. We can break
4265 */
4267 }
4268 }
4269 }
4278 }
4282 /* wake up any body waiting to join. */
4289 }
4290 out:
4294 /* Re-set th->t_super, so we can properly keep track of how many
4295 * persistent transactions there are. We need to do this so if this
4296 * call is part of a failed restart_transaction, we can free it later */
4300 }
4302 /* Send the file system read only and refuse new transactions */
4304 {
4315 #ifdef CONFIG_REISERFS_CHECK
4317 #endif
4318 }