| blob | 4cad9e75ef560f6238087bf2cc49ccd5cc571412 |
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 <asm/uaccess.h>
38 #include <asm/system.h>
40 #include <linux/time.h>
41 #include <asm/semaphore.h>
43 #include <linux/vmalloc.h>
44 #include <linux/reiserfs_fs.h>
46 #include <linux/kernel.h>
47 #include <linux/errno.h>
48 #include <linux/fcntl.h>
49 #include <linux/stat.h>
50 #include <linux/string.h>
51 #include <linux/smp_lock.h>
52 #include <linux/buffer_head.h>
53 #include <linux/workqueue.h>
54 #include <linux/writeback.h>
55 #include <linux/blkdev.h>
56 #include <linux/backing-dev.h>
58 /* gets a struct reiserfs_journal_list * from a list head */
59 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
60 j_list))
61 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
62 j_working_list))
64 /* the number of mounted filesystems. This is used to decide when to
65 ** start and kill the commit workqueue
66 */
72 structs at 4k */
75 /* cnode stat bits. Move these into reiserfs_fs.h */
78 #define BLOCK_FREED_HOLDER 3 /* this block was freed during this transaction, and can't be written */
81 #define BLOCK_DIRTIED 5
83 /* journal list state bits */
84 #define LIST_TOUCHED 1
85 #define LIST_DIRTY 2
88 /* flags for do_journal_end */
110 /* values for join in do_journal_begin_r */
115 };
122 {
126 }
128 /*
129 ** clears BH_Dirty and sticks the buffer on the clean list. Called because I can't allow refile_buffer to
130 ** make schedule happen after I've freed a block. Look at remove_from_transaction and journal_mark_freed for
131 ** more details.
132 */
134 {
138 }
140 }
143 {
147 }
151 {
158 }
163 }
167 }
170 {
176 repeat:
184 }
189 }
191 }
194 {
203 }
204 }
207 {
218 }
219 }
220 }
224 {
230 }
233 }
237 {
246 }
247 }
248 }
250 /*
251 ** only call this on FS unmount.
252 */
255 {
264 }
266 }
269 {
281 }
284 }
286 /*
287 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
288 ** jb_array is the array to be filled in.
289 */
293 {
308 }
310 }
314 }
316 }
318 /*
319 ** find an available list bitmap. If you can't find one, flush a commit list
320 ** and try again
321 */
323 struct reiserfs_journal_list
325 {
340 }
343 }
344 }
347 }
350 }
352 /*
353 ** allocates a new chunk of X nodes, and links them all together as a list.
354 ** Uses the cnode->next and cnode->prev pointers
355 ** returns NULL on failure
356 */
358 {
363 }
367 }
374 }
377 }
379 /*
380 ** pulls a cnode off the free list, or returns NULL on failure
381 */
383 {
391 }
397 }
400 }
404 }
406 /*
407 ** returns a cnode to the free list
408 */
411 {
418 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
422 }
423 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
425 }
428 {
431 }
433 /* utility function to force a BUG if it is called without the big
434 ** kernel lock held. caller is the string printed just before calling BUG()
435 */
437 {
438 #ifdef CONFIG_SMP
441 caller);
442 }
443 #else
444 ;
445 #endif
446 }
448 /* return a cnode with same dev, block number and size in table, or null if not found */
450 super_block
452 struct
453 reiserfs_journal_cnode
456 {
463 }
465 }
467 /*
468 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
469 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
470 ** being overwritten by a replay after crashing.
471 **
472 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
473 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
474 ** sure you never write the block without logging it.
475 **
476 ** next_zero_bit is a suggestion about the next block to try for find_forward.
477 ** when bl is rejected because it is set in a journal list bitmap, we search
478 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
479 ** through next_zero_bit for find_forward to try.
480 **
481 ** Just because we return something in next_zero_bit does not mean we won't
482 ** reject it on the next call to reiserfs_in_journal
483 **
484 */
488 {
498 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
499 ** if we crash before the transaction that freed it commits, this transaction won't
500 ** have committed either, and the block will never be written
501 */
509 data)) {
513 data),
517 }
518 }
519 }
522 /* is it in any old transactions? */
527 }
529 /* is it in the current transaction. This should never happen */
533 }
536 /* safe for reuse */
538 }
540 /* insert cn into table
541 */
544 {
552 }
554 }
556 /* lock the current transaction */
558 {
561 }
563 /* unlock the current transaction */
565 {
567 }
570 {
572 }
576 {
580 }
583 }
585 /*
586 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
587 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
588 ** transaction.
589 */
592 {
597 }
600 }
604 {
613 }
614 }
616 }
618 /*
619 * If page->mapping was null, we failed to truncate this page for
620 * some reason. Most likely because it was truncated after being
621 * logged via data=journal.
622 *
623 * This does a check to see if the buffer belongs to one of these
624 * lost pages before doing the final put_bh. If page->mapping was
625 * null, it tries to free buffers on the page, which should make the
626 * final page_cache_release drop the page from the lru.
627 */
629 {
640 }
641 }
644 {
651 }
654 else
659 }
662 {
665 else
669 }
672 {
682 }
685 {
692 }
695 {
702 }
706 {
713 }
714 }
716 #define CHUNK_SIZE 32
720 };
723 {
728 }
731 }
734 {
739 }
742 }
746 {
757 }
759 }
763 {
770 }
772 }
773 }
775 /*
776 * we want to free the jh when the buffer has been written
777 * and waited on
778 */
780 {
793 }
794 }
798 {
806 }
810 no_jh:
814 /* buffer must be locked for __add_jh, should be able to have
815 * two adds at the same time
816 */
820 }
826 }
829 }
832 {
834 }
836 {
838 }
840 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
845 {
863 }
871 }
872 /* in theory, dirty non-uptodate buffers should never get here,
873 * but the upper layer io error paths still have a few quirks.
874 * Handle them here as gracefully as we can
875 */
879 }
886 }
887 loop_next:
890 }
895 }
906 }
909 }
910 /* ugly interaction with invalidatepage here.
911 * reiserfs_invalidate_page will pin any buffer that has a valid
912 * journal head from an older transaction. If someone else sets
913 * our buffer dirty after we write it in the first loop, and
914 * then someone truncates the page away, nobody will ever write
915 * the buffer. We're safe if we write the page one last time
916 * after freeing the journal header.
917 */
922 }
925 }
928 }
932 {
941 find_first:
942 /*
943 * first we walk backwards to find the oldest uncommitted transation
944 */
955 }
957 /* if we didn't find any older uncommitted transactions, return now */
960 }
973 /* list we were called with is gone, return */
977 /* the one we just flushed is gone, this means all
978 * older lists are also gone, so first_jl is no longer
979 * valid either. Go back to the beginning.
980 */
984 }
985 }
991 }
992 }
994 }
997 {
1003 }
1005 /*
1006 ** if this journal list still has commit blocks unflushed, send them to disk.
1007 **
1008 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
1009 ** Before the commit block can by written, every other log block must be safely on disk
1010 **
1011 */
1014 {
1028 }
1032 /* before we can put our commit blocks on disk, we have to make sure everyone older than
1033 ** us is on disk too
1034 */
1041 /* list disappeared during flush_older_commits. return */
1043 }
1044 }
1046 /* make sure nobody is trying to flush this one at the same time */
1051 }
1054 /* this commit is done, exit */
1058 }
1061 }
1071 }
1073 /*
1074 * for the description block and all the log blocks, submit any buffers
1075 * that haven't already reached the disk. Try to write at least 256
1076 * log blocks. later on, we will only wait on blocks that correspond
1077 * to this transaction, but while we're unplugging we might as well
1078 * get a chunk of data on there.
1079 */
1092 }
1093 }
1096 /* We're skipping the commit if there's an error */
1100 /* wait on everything written so far before writing the commit
1101 * if we are in barrier mode, send the commit down now
1102 */
1112 }
1113 }
1119 // since we're using ll_rw_blk above, it might have skipped over
1120 // a locked buffer. Double check here
1121 //
1125 #ifdef CONFIG_REISERFS_CHECK
1127 #endif
1129 }
1133 }
1138 /* If there was a write error in the journal - we can't commit
1139 * this transaction - it will be invalid and, if successful,
1140 * will just end up propagating the write error out to
1141 * the file system. */
1147 }
1153 /* If there was a write error in the journal - we can't commit this
1154 * transaction - it will be invalid and, if successful, will just end
1155 * up propagating the write error out to the filesystem. */
1157 #ifdef CONFIG_REISERFS_CHECK
1159 #endif
1161 }
1167 }
1170 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1175 /* mark the metadata dirty */
1182 }
1184 put_jl:
1189 __FUNCTION__);
1192 }
1194 /*
1195 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1196 ** returns NULL if it can't find anything
1197 */
1199 reiserfs_journal_cnode
1201 {
1209 }
1211 }
1213 }
1216 {
1226 }
1228 }
1235 /*
1236 ** once all the real blocks have been flushed, it is safe to remove them from the
1237 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1238 ** block to be reallocated for data blocks if it had been deleted.
1239 */
1243 {
1248 /* which is better, to lock once around the whole loop, or
1249 ** to lock for each call to remove_journal_hash?
1250 */
1258 }
1262 }
1266 }
1268 }
1270 /*
1271 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1272 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1273 ** releasing blocks in this transaction for reuse as data blocks.
1274 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1275 **
1276 */
1280 {
1291 #ifdef CONFIG_REISERFS_CHECK
1294 #endif
1296 }
1297 }
1301 b_data);
1314 }
1318 sync:
1321 }
1326 }
1327 }
1329 }
1334 {
1336 }
1338 /*
1339 ** flush any and all journal lists older than you are
1340 ** can only be called from flush_journal_list
1341 */
1344 {
1350 /* we know we are the only ones flushing things, no extra race
1351 * protection is required.
1352 */
1353 restart:
1355 /* Did we wrap? */
1361 /* do not flush all */
1364 /* other_jl is now deleted from the list */
1366 }
1368 }
1372 {
1377 }
1378 }
1380 /* flush a journal list, both commit and real blocks
1381 **
1382 ** always set flushall to 1, unless you are calling from inside
1383 ** flush_journal_list
1384 **
1385 ** IMPORTANT. This can only be called while there are no journal writers,
1386 ** and the journal is locked. That means it can only be called from
1387 ** do_journal_end, or by journal_release
1388 */
1391 {
1408 }
1411 /* if flushall == 0, the lock is already held */
1416 }
1424 }
1428 /* if all the work is already done, get out of here */
1432 }
1434 /* start by putting the commit list on disk. This will also flush
1435 ** the commit lists of any olders transactions
1436 */
1443 /* are we done now? */
1447 }
1449 /* loop through each cnode, see if we need to write it,
1450 ** or wait on a more recent transaction, or just ignore it
1451 */
1455 }
1461 /* blocknr of 0 is no longer in the hash, ignore it */
1464 }
1466 /* This transaction failed commit. Don't write out to the disk */
1471 /* the order is important here. We check pjl to make sure we
1472 ** don't clear BH_JDirty_wait if we aren't the one writing this
1473 ** block to disk
1474 */
1478 /* we do this to make sure nobody releases the buffer while
1479 ** we are working with it
1480 */
1488 /* everything with !pjl && jwait should be writable */
1490 }
1491 }
1493 /* if someone has this block in a newer transaction, just make
1494 ** sure they are committed, and don't try writing it to disk
1495 */
1500 }
1502 /* bh == NULL when the block got to disk on its own, OR,
1503 ** the block got freed in a future transaction
1504 */
1507 }
1509 /* this should never happen. kupdate_one_transaction has this list
1510 ** locked while it works, so we should never see a buffer here that
1511 ** is not marked JDirty_wait
1512 */
1515 "journal-813: BAD! buffer %llu %cdirty %cjwait, "
1520 }
1522 /* we inc again because saved_bh gets decremented at free_cnode */
1529 else
1531 count++;
1537 }
1538 free_cnode:
1542 /* we incremented this to keep others from taking the buffer head away */
1547 }
1548 }
1549 }
1557 }
1562 }
1564 #ifdef CONFIG_REISERFS_CHECK
1567 #endif
1569 }
1570 /* note, we must clear the JDirty_wait bit after the up to date
1571 ** check, otherwise we race against our flushpage routine
1572 */
1576 /* drop one ref for us */
1578 /* drop one ref for journal_mark_dirty */
1580 }
1582 }
1583 }
1588 __FUNCTION__);
1589 flush_older_and_return:
1591 /* before we can update the journal header block, we _must_ flush all
1592 ** real blocks from all older transactions to disk. This is because
1593 ** once the header block is updated, this transaction will not be
1594 ** replayed after a crash
1595 */
1598 }
1601 /* before we can remove everything from the hash tables for this
1602 ** transaction, we must make sure it can never be replayed
1603 **
1604 ** since we are only called from do_journal_end, we know for sure there
1605 ** are no allocations going on while we are flushing journal lists. So,
1606 ** we only need to update the journal header block for the last list
1607 ** being flushed
1608 */
1610 err =
1618 __FUNCTION__);
1619 }
1629 }
1632 /* not strictly required since we are freeing the list, but it should
1633 * help find code using dead lists later on
1634 */
1647 }
1651 {
1659 /* if the blocknr == 0, this has been cleared from the hash,
1660 ** skip it
1661 */
1664 }
1667 next:
1670 }
1672 }
1677 {
1685 }
1689 /* if the blocknr == 0, this has been cleared from the hash,
1690 ** skip it
1691 */
1694 }
1697 /* we can race against journal_mark_freed when we try
1698 * to lock_buffer(cn->bh), so we have to inc the buffer
1699 * count, and recheck things after locking
1700 */
1709 ret++;
1711 /* note, cn->bh might be null now */
1713 }
1715 }
1716 next:
1719 }
1721 }
1723 /* used by flush_commit_list */
1726 {
1734 /* look for a more recent transaction that logged this
1735 ** buffer. Only the most recent transaction with a buffer in
1736 ** it is allowed to send that buffer to disk
1737 */
1742 /* if the buffer is prepared, it will either be logged
1743 * or restored. If restored, we need to make sure
1744 * it actually gets marked dirty
1745 */
1752 }
1753 }
1755 }
1757 }
1764 {
1777 }
1779 /* we've got j_flush_sem held, nobody is going to delete any
1780 * of these lists out from underneath us
1781 */
1790 }
1795 transactions_flushed++;
1799 /* did we wrap? */
1802 }
1805 /* don't bother with older transactions */
1808 }
1811 }
1813 done:
1816 }
1818 /* for o_sync and fsync heavy applications, they tend to use
1819 ** all the journa list slots with tiny transactions. These
1820 ** trigger lots and lots of calls to update the header block, which
1821 ** adds seeks and slows things down.
1822 **
1823 ** This function tries to clear out a large chunk of the journal lists
1824 ** at once, which makes everything faster since only the newest journal
1825 ** list updates the header block
1826 */
1829 {
1842 /* in data logging mode, try harder to flush a lot of blocks */
1845 /* flush for 256 transactions or limit blocks, whichever comes first */
1850 }
1854 }
1860 }
1861 /* try to find a group of blocks we can flush across all the
1862 ** transactions, but only bother if we've actually spanned
1863 ** across multiple lists
1864 */
1867 }
1870 }
1872 /*
1873 ** removes any nodes in table with name block and dev as bh.
1874 ** only touchs the hnext and hprev pointers.
1875 */
1880 {
1887 }
1895 }
1900 }
1908 }
1910 }
1911 }
1914 {
1924 }
1925 /* j_header_bh is on the journal dev, make sure not to release the journal
1926 * dev until we brelse j_header_bh
1927 */
1930 }
1932 /*
1933 ** call on unmount. Only set error to 1 if you haven't made your way out
1934 ** of read_super() yet. Any other caller must keep error at 0.
1935 */
1938 {
1943 /* we only want to flush out transactions if we were called with error == 0
1944 */
1946 /* end the current trans */
1950 /* make sure something gets logged to force our way into the flush code */
1959 }
1960 }
1962 /* this also catches errors during the do_journal_end above */
1972 }
1973 }
1975 reiserfs_mounted_fs_count--;
1976 /* wait for all commits to finish */
1982 }
1987 }
1989 /*
1990 ** call on unmount. flush all journal trans, release all alloc'd ram
1991 */
1994 {
1996 }
1998 /*
1999 ** only call from an error condition inside reiserfs_read_super!
2000 */
2003 {
2005 }
2007 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
2011 {
2017 }
2019 }
2021 /* returns 0 if it did not find a description block
2022 ** returns -1 if it found a corrupt commit block
2023 ** returns 1 if both desc and commit were valid
2024 */
2029 {
2044 "journal-986: transaction "
2045 "is valid returning because trans_id %d is greater than "
2050 }
2054 "journal-1087: transaction "
2055 "is valid returning because mount_id %d is less than "
2060 }
2066 }
2069 /* ok, we have a journal description block, lets see if the transaction was valid */
2070 c_bh =
2080 "journal_transaction_is_valid, commit offset %ld had bad "
2091 "journal-1004: "
2092 "transaction_is_valid setting oldest invalid trans_id "
2095 }
2097 }
2100 "journal-1006: found valid "
2109 }
2110 }
2113 {
2117 }
2118 }
2120 /*
2121 ** given the start, and values for the oldest acceptable transactions,
2122 ** this either reads in a replays a transaction, or returns because the transaction
2123 ** is invalid, or too old.
2124 */
2130 {
2155 cur_dblock -
2159 }
2162 "journal_read_trans skipping because %d is != "
2164 newest_mount_id);
2167 }
2174 }
2178 "journal_read_transaction, "
2187 }
2189 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2202 }
2203 /* get all the buffer heads */
2220 }
2225 }
2226 /* make sure we don't try to replay onto log or reserved area */
2231 abort_replay:
2239 }
2240 }
2241 /* read in the log blocks, memcpy to the corresponding real block */
2256 }
2261 }
2262 /* flush out the real blocks */
2266 }
2279 }
2281 }
2282 cur_dblock =
2290 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2294 /* check for trans_id overflow */
2302 }
2304 /* This function reads blocks starting from block and to max_block of bufsize
2305 size (but no more than BUFNR blocks at a time). This proved to improve
2306 mounting speed on self-rebuilding raid5 arrays at least.
2307 Right now it is only used from journal code. But later we might use it
2308 from other places.
2309 Note: Do not use journal_getblk/sb_getblk functions here! */
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 {
2582 }
2587 result);
2588 }
2590 }
2595 {
2597 dev_t jdev;
2611 /* there is no "jdev" option and journal is on separate device */
2624 }
2635 /* ok */
2641 }
2648 }
2650 }
2652 /**
2653 * When creating/tuning a file system user can assign some
2654 * journal params within boundaries which depend on the ratio
2655 * blocksize/standard_blocksize.
2656 *
2657 * For blocks >= standard_blocksize transaction size should
2658 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2659 * then JOURNAL_TRANS_MAX_DEFAULT.
2660 *
2661 * For blocks < standard_blocksize these boundaries should be
2662 * decreased proportionally.
2663 */
2664 #define REISERFS_STANDARD_BLKSIZE (4096)
2668 {
2670 /* Non-default journal params.
2671 Do sanity check for them. */
2679 JOURNAL_MIN_RATIO) {
2684 }
2691 }
2693 /* Default journal params.
2694 The file system was created by old version
2695 of mkreiserfs, so some fields contain zeros,
2696 and we need to advise proper values for them */
2703 }
2705 }
2707 /*
2708 ** must be called once on fs mount. calls journal_read for you
2709 */
2712 {
2726 }
2739 /* reserved for journal area support */
2741 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2745 REISERFS_DISK_OFFSET_IN_BYTES /
2748 /* Sanity check to see is the standard journal fitting withing first bitmap
2749 (actual for small blocksizes) */
2754 "journal-1393: journal does not fit for area "
2755 "addressed by first of bitmap blocks. It starts at "
2761 }
2767 }
2771 /* read journal header */
2779 }
2782 /* make sure that journal matches to the super block */
2787 "sh-460: journal header magic %x "
2788 "(device %s) does not match to magic found in super "
2794 }
2809 }
2812 "journal first block %u, max trans len %u, max batch %u, "
2857 "allocation failed (%ld bytes). Journal is "
2858 "too large for available memory. Usually "
2862 }
2871 }
2875 }
2877 reiserfs_mounted_fs_count++;
2884 free_and_return:
2887 }
2889 /*
2890 ** test for a polite end of the current transaction. Used by file_write, and should
2891 ** be used by delete to make sure they don't write more than can fit inside a single
2892 ** transaction
2893 */
2896 {
2899 /* cannot restart while nested */
2909 }
2910 /* protected by the BKL here */
2914 }
2916 /* this must be called inside a transaction, and requires the
2917 ** kernel_lock to be held
2918 */
2920 {
2926 }
2928 /* this must be called without a transaction started, and does not
2929 ** require BKL
2930 */
2932 {
2936 }
2938 /* this must be called without a transaction started, and does not
2939 ** require BKL
2940 */
2942 {
2946 }
2949 {
2950 wait_queue_t wait;
2954 /*
2955 * we don't want to use wait_event here because
2956 * we only want to wait once.
2957 */
2965 }
2968 {
2972 }
2975 {
2985 }
2991 }
2992 }
2994 /* join == true if you must join an existing transaction.
2995 ** join == false if you can deal with waiting for others to finish
2996 **
2997 ** this will block until the transaction is joinable. send the number of blocks you
2998 ** expect to use in nblocks.
2999 */
3003 {
3015 /* set here for journal_join */
3019 relock:
3025 }
3033 }
3036 /* if there is no room in the journal OR
3037 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3038 ** we don't sleep if there aren't other writers
3039 */
3042 (!join
3059 sched_count++;
3062 }
3063 }
3064 /* don't mess with joining the transaction if all we have to do is
3065 * wait for someone else to do a commit
3066 */
3071 }
3073 }
3078 /* someone might have ended the transaction while we joined */
3083 }
3090 }
3091 /* we are the first writer, set trans_id */
3094 }
3105 out_fail:
3107 /* Re-set th->t_super, so we can properly keep track of how many
3108 * persistent transactions there are. We need to do this so if this
3109 * call is part of a failed restart_transaction, we can free it later */
3112 }
3115 super_block
3118 {
3122 /* if we're nesting into an existing transaction. It will be
3123 ** persistent on its own
3124 */
3131 }
3139 }
3143 }
3146 {
3151 else
3156 }
3158 }
3162 {
3165 /* this keeps do_journal_end from NULLing out the current->journal_info
3166 ** pointer
3167 */
3171 }
3175 {
3178 /* this keeps do_journal_end from NULLing out the current->journal_info
3179 ** pointer
3180 */
3184 }
3188 {
3194 /* we are nesting into the current transaction */
3204 /* we've ended up with a handle from a different filesystem.
3205 ** save it and restore on journal_end. This should never
3206 ** really happen...
3207 */
3212 }
3215 }
3219 /* I guess this boils down to being the reciprocal of clm-2100 above.
3220 * If do_journal_begin_r fails, we need to put it back, since journal_end
3221 * won't be called to do it. */
3224 else
3228 }
3230 /*
3231 ** puts bh into the current transaction. If it was already there, reorders removes the
3232 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3233 **
3234 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3235 ** transaction is committed.
3236 **
3237 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3238 */
3241 {
3253 }
3259 /* already in this transaction, we are done */
3263 }
3265 /* this must be turned into a panic instead of a warning. We can't allow
3266 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3267 ** could get to disk too early. NOT GOOD.
3268 */
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);
3739 }
3740 }
3744 j_nonzerolen));
3745 }
3747 }
3748 }
3750 }
3751 }
3756 }
3759 {
3763 }
3765 /*
3766 * returns -1 on error, 0 if no commits/barriers were done and 1
3767 * if a transaction was actually committed and the barrier was done
3768 */
3771 {
3777 /* is it from the current transaction, or from an unknown transaction? */
3780 /* try to let other writers come in and grow this transaction */
3784 }
3790 /* someone might have ended this transaction while we joined */
3797 }
3804 /* this gets tricky, we have to make sure the journal list in
3805 * the inode still exists. We know the list is still around
3806 * if we've got a larger transaction id than the oldest list
3807 */
3808 flush_commit_only:
3810 /*
3811 * we only set ret to 1 when we know for sure
3812 * the barrier hasn't been started yet on the commit
3813 * block.
3814 */
3820 }
3821 }
3822 /* otherwise the list is gone, and long since committed */
3824 }
3827 {
3831 /* for the whole inode, assume unset id means it was
3832 * changed in the current transaction. More conservative
3833 */
3837 /* jl will be updated in __commit_trans_jl */
3838 }
3841 }
3845 {
3850 }
3860 }
3861 }
3863 }
3866 /*
3867 ** before we can change a metadata block, we have to make sure it won't
3868 ** be written to disk while we are altering it. So, we must:
3869 ** clean it
3870 ** wait on it.
3871 **
3872 */
3875 {
3882 }
3887 }
3890 }
3893 {
3902 /* this check should always be run, to send old lists to disk */
3909 }
3910 }
3911 }
3913 /*
3914 ** long and ugly. If flush, will not return until all commit
3915 ** blocks and all real buffers in the trans are on disk.
3916 ** If no_async, won't return until all commit blocks are on disk.
3917 **
3918 ** keep reading, there are comments as you go along
3919 **
3920 ** If the journal is aborted, we just clean up. Things like flushing
3921 ** journal lists, etc just won't happen.
3922 */
3926 {
3948 /* protect flush_older_commits from doing mistakes if the
3949 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 */
4079 }
4092 }
4094 i--;
4095 }
4096 }
4102 /* special check in case all buffers in the journal were marked for not logging */
4105 /* we're about to dirty all the log blocks, mark the description block
4106 * dirty now too. Don't mark the commit block dirty until all the
4107 * others are on disk
4108 */
4111 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4117 /* copy all the real blocks into log area. dirty log blocks */
4122 tmp_bh =
4126 jindex) %
4136 jindex++;
4140 /* JDirty cleared sometime during transaction. don't log this one */
4144 }
4149 }
4151 /* we are done with both the c_bh and d_bh, but
4152 ** c_bh must be written after all other commit blocks,
4153 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4154 */
4158 /* now it is safe to insert this transaction on the main list */
4163 /* reset journal values for the next transaction */
4174 /* check for trans_id overflow */
4184 // make sure reiserfs_add_jh sees the new current_jl before we
4185 // write out the tails
4188 /* tail conversion targets have to hit the disk before we end the
4189 * transaction. Otherwise a later transaction might repack the tail
4190 * before this transaction commits, leaving the data block unflushed and
4191 * clean, if we crash before the later transaction commits, the data block
4192 * is lost.
4193 */
4199 }
4203 /* honor the flush wishes from the caller, simple commits can
4204 ** be done outside the journal lock, they are done below
4205 **
4206 ** if we don't flush the commit list right now, we put it into
4207 ** the work queue so the people waiting on the async progress work
4208 ** queue don't wait for this proc to flush journal lists and such.
4209 */
4216 /* if the next transaction has any chance of wrapping, flush
4217 ** transactions that might get overwritten. If any journal lists are very
4218 ** old flush them as well.
4219 */
4220 first_jl:
4231 /* if we don't cross into the next transaction and we don't
4232 * wrap, there is no way we can overlap any later transactions
4233 * break now
4234 */
4236 }
4246 /* we don't overlap anything from out start to the end of the
4247 * log, and our wrapped portion doesn't overlap anything at
4248 * the start of the log. We can break
4249 */
4251 }
4252 }
4253 }
4262 }
4266 /* wake up any body waiting to join. */
4273 }
4274 out:
4278 /* Re-set th->t_super, so we can properly keep track of how many
4279 * persistent transactions there are. We need to do this so if this
4280 * call is part of a failed restart_transaction, we can free it later */
4284 }
4287 {
4298 #ifdef CONFIG_REISERFS_CHECK
4300 #endif
4301 }
4304 {
4313 }
4316 {
4318 }