| blob | 060eb3f598e7d58b90d9595a9e9d4dcac33c4046 |
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 <linux/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 }
223 b_blocknr_t block,
225 {
231 }
234 }
238 {
247 }
248 }
249 }
251 /*
252 ** only call this on FS unmount.
253 */
256 {
265 }
267 }
270 {
282 }
285 }
287 /*
288 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
289 ** jb_array is the array to be filled in.
290 */
294 {
309 }
311 }
315 }
317 }
319 /*
320 ** find an available list bitmap. If you can't find one, flush a commit list
321 ** and try again
322 */
324 struct reiserfs_journal_list
326 {
341 }
344 }
345 }
348 }
351 }
353 /*
354 ** allocates a new chunk of X nodes, and links them all together as a list.
355 ** Uses the cnode->next and cnode->prev pointers
356 ** returns NULL on failure
357 */
359 {
364 }
368 }
375 }
378 }
380 /*
381 ** pulls a cnode off the free list, or returns NULL on failure
382 */
384 {
392 }
398 }
401 }
405 }
407 /*
408 ** returns a cnode to the free list
409 */
412 {
419 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
423 }
424 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */
426 }
429 {
432 }
434 /* utility function to force a BUG if it is called without the big
435 ** kernel lock held. caller is the string printed just before calling BUG()
436 */
438 {
439 #ifdef CONFIG_SMP
442 caller);
443 }
444 #else
445 ;
446 #endif
447 }
449 /* return a cnode with same dev, block number and size in table, or null if not found */
451 super_block
453 struct
454 reiserfs_journal_cnode
457 {
464 }
466 }
468 /*
469 ** this actually means 'can this block be reallocated yet?'. If you set search_all, a block can only be allocated
470 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
471 ** being overwritten by a replay after crashing.
472 **
473 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction. Since deleting
474 ** a block removes it from the current transaction, this case should never happen. If you don't set search_all, make
475 ** sure you never write the block without logging it.
476 **
477 ** next_zero_bit is a suggestion about the next block to try for find_forward.
478 ** when bl is rejected because it is set in a journal list bitmap, we search
479 ** for the next zero bit in the bitmap that rejected bl. Then, we return that
480 ** through next_zero_bit for find_forward to try.
481 **
482 ** Just because we return something in next_zero_bit does not mean we won't
483 ** reject it on the next call to reiserfs_in_journal
484 **
485 */
489 {
499 /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
500 ** if we crash before the transaction that freed it commits, this transaction won't
501 ** have committed either, and the block will never be written
502 */
510 data)) {
514 data),
518 }
519 }
520 }
523 /* is it in any old transactions? */
528 }
530 /* is it in the current transaction. This should never happen */
534 }
537 /* safe for reuse */
539 }
541 /* insert cn into table
542 */
545 {
553 }
555 }
557 /* lock the current transaction */
559 {
562 }
564 /* unlock the current transaction */
566 {
568 }
571 {
573 }
577 {
581 }
584 }
586 /*
587 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
588 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
589 ** transaction.
590 */
593 {
598 }
601 }
605 {
614 }
615 }
617 }
619 /*
620 * If page->mapping was null, we failed to truncate this page for
621 * some reason. Most likely because it was truncated after being
622 * logged via data=journal.
623 *
624 * This does a check to see if the buffer belongs to one of these
625 * lost pages before doing the final put_bh. If page->mapping was
626 * null, it tries to free buffers on the page, which should make the
627 * final page_cache_release drop the page from the lru.
628 */
630 {
641 }
642 }
645 {
652 }
655 else
660 }
663 {
666 else
670 }
673 {
683 }
686 {
693 }
696 {
703 }
707 {
714 }
715 }
717 #define CHUNK_SIZE 32
721 };
724 {
729 }
732 }
735 {
740 }
743 }
747 {
758 }
760 }
764 {
771 }
773 }
774 }
776 /*
777 * we want to free the jh when the buffer has been written
778 * and waited on
779 */
781 {
794 }
795 }
799 {
807 }
811 no_jh:
815 /* buffer must be locked for __add_jh, should be able to have
816 * two adds at the same time
817 */
821 }
827 }
830 }
833 {
835 }
837 {
839 }
841 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
846 {
864 }
872 }
873 /* in theory, dirty non-uptodate buffers should never get here,
874 * but the upper layer io error paths still have a few quirks.
875 * Handle them here as gracefully as we can
876 */
880 }
887 }
888 loop_next:
891 }
896 }
907 }
910 }
911 /* ugly interaction with invalidatepage here.
912 * reiserfs_invalidate_page will pin any buffer that has a valid
913 * journal head from an older transaction. If someone else sets
914 * our buffer dirty after we write it in the first loop, and
915 * then someone truncates the page away, nobody will ever write
916 * the buffer. We're safe if we write the page one last time
917 * after freeing the journal header.
918 */
923 }
926 }
929 }
933 {
942 find_first:
943 /*
944 * first we walk backwards to find the oldest uncommitted transation
945 */
956 }
958 /* if we didn't find any older uncommitted transactions, return now */
961 }
974 /* list we were called with is gone, return */
978 /* the one we just flushed is gone, this means all
979 * older lists are also gone, so first_jl is no longer
980 * valid either. Go back to the beginning.
981 */
985 }
986 }
992 }
993 }
995 }
998 {
1004 }
1006 /*
1007 ** if this journal list still has commit blocks unflushed, send them to disk.
1008 **
1009 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
1010 ** Before the commit block can by written, every other log block must be safely on disk
1011 **
1012 */
1015 {
1017 b_blocknr_t bn;
1029 }
1033 /* before we can put our commit blocks on disk, we have to make sure everyone older than
1034 ** us is on disk too
1035 */
1042 /* list disappeared during flush_older_commits. return */
1044 }
1045 }
1047 /* make sure nobody is trying to flush this one at the same time */
1052 }
1055 /* this commit is done, exit */
1059 }
1062 }
1072 }
1074 /*
1075 * for the description block and all the log blocks, submit any buffers
1076 * that haven't already reached the disk. Try to write at least 256
1077 * log blocks. later on, we will only wait on blocks that correspond
1078 * to this transaction, but while we're unplugging we might as well
1079 * get a chunk of data on there.
1080 */
1093 }
1094 }
1097 /* We're skipping the commit if there's an error */
1101 /* wait on everything written so far before writing the commit
1102 * if we are in barrier mode, send the commit down now
1103 */
1113 }
1114 }
1120 // since we're using ll_rw_blk above, it might have skipped over
1121 // a locked buffer. Double check here
1122 //
1126 #ifdef CONFIG_REISERFS_CHECK
1128 #endif
1130 }
1134 }
1139 /* If there was a write error in the journal - we can't commit
1140 * this transaction - it will be invalid and, if successful,
1141 * will just end up propagating the write error out to
1142 * the file system. */
1148 }
1154 /* If there was a write error in the journal - we can't commit this
1155 * transaction - it will be invalid and, if successful, will just end
1156 * up propagating the write error out to the filesystem. */
1158 #ifdef CONFIG_REISERFS_CHECK
1160 #endif
1162 }
1168 }
1171 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1176 /* mark the metadata dirty */
1183 }
1185 put_jl:
1190 __FUNCTION__);
1193 }
1195 /*
1196 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1197 ** returns NULL if it can't find anything
1198 */
1200 reiserfs_journal_cnode
1202 {
1210 }
1212 }
1214 }
1217 {
1227 }
1229 }
1236 /*
1237 ** once all the real blocks have been flushed, it is safe to remove them from the
1238 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1239 ** block to be reallocated for data blocks if it had been deleted.
1240 */
1244 {
1249 /* which is better, to lock once around the whole loop, or
1250 ** to lock for each call to remove_journal_hash?
1251 */
1259 }
1263 }
1267 }
1269 }
1271 /*
1272 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1273 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1274 ** releasing blocks in this transaction for reuse as data blocks.
1275 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1276 **
1277 */
1281 {
1292 #ifdef CONFIG_REISERFS_CHECK
1295 #endif
1297 }
1298 }
1302 b_data);
1315 }
1319 sync:
1322 }
1327 }
1328 }
1330 }
1335 {
1337 }
1339 /*
1340 ** flush any and all journal lists older than you are
1341 ** can only be called from flush_journal_list
1342 */
1345 {
1351 /* we know we are the only ones flushing things, no extra race
1352 * protection is required.
1353 */
1354 restart:
1356 /* Did we wrap? */
1362 /* do not flush all */
1365 /* other_jl is now deleted from the list */
1367 }
1369 }
1373 {
1378 }
1379 }
1381 /* flush a journal list, both commit and real blocks
1382 **
1383 ** always set flushall to 1, unless you are calling from inside
1384 ** flush_journal_list
1385 **
1386 ** IMPORTANT. This can only be called while there are no journal writers,
1387 ** and the journal is locked. That means it can only be called from
1388 ** do_journal_end, or by journal_release
1389 */
1392 {
1409 }
1412 /* if flushall == 0, the lock is already held */
1417 }
1425 }
1429 /* if all the work is already done, get out of here */
1433 }
1435 /* start by putting the commit list on disk. This will also flush
1436 ** the commit lists of any olders transactions
1437 */
1444 /* are we done now? */
1448 }
1450 /* loop through each cnode, see if we need to write it,
1451 ** or wait on a more recent transaction, or just ignore it
1452 */
1456 }
1462 /* blocknr of 0 is no longer in the hash, ignore it */
1465 }
1467 /* This transaction failed commit. Don't write out to the disk */
1472 /* the order is important here. We check pjl to make sure we
1473 ** don't clear BH_JDirty_wait if we aren't the one writing this
1474 ** block to disk
1475 */
1479 /* we do this to make sure nobody releases the buffer while
1480 ** we are working with it
1481 */
1489 /* everything with !pjl && jwait should be writable */
1491 }
1492 }
1494 /* if someone has this block in a newer transaction, just make
1495 ** sure they are committed, and don't try writing it to disk
1496 */
1501 }
1503 /* bh == NULL when the block got to disk on its own, OR,
1504 ** the block got freed in a future transaction
1505 */
1508 }
1510 /* this should never happen. kupdate_one_transaction has this list
1511 ** locked while it works, so we should never see a buffer here that
1512 ** is not marked JDirty_wait
1513 */
1516 "journal-813: BAD! buffer %llu %cdirty %cjwait, "
1521 }
1523 /* we inc again because saved_bh gets decremented at free_cnode */
1530 else
1532 count++;
1538 }
1539 free_cnode:
1543 /* we incremented this to keep others from taking the buffer head away */
1548 }
1549 }
1550 }
1558 }
1563 }
1565 #ifdef CONFIG_REISERFS_CHECK
1568 #endif
1570 }
1571 /* note, we must clear the JDirty_wait bit after the up to date
1572 ** check, otherwise we race against our flushpage routine
1573 */
1577 /* drop one ref for us */
1579 /* drop one ref for journal_mark_dirty */
1581 }
1583 }
1584 }
1589 __FUNCTION__);
1590 flush_older_and_return:
1592 /* before we can update the journal header block, we _must_ flush all
1593 ** real blocks from all older transactions to disk. This is because
1594 ** once the header block is updated, this transaction will not be
1595 ** replayed after a crash
1596 */
1599 }
1602 /* before we can remove everything from the hash tables for this
1603 ** transaction, we must make sure it can never be replayed
1604 **
1605 ** since we are only called from do_journal_end, we know for sure there
1606 ** are no allocations going on while we are flushing journal lists. So,
1607 ** we only need to update the journal header block for the last list
1608 ** being flushed
1609 */
1611 err =
1619 __FUNCTION__);
1620 }
1630 }
1633 /* not strictly required since we are freeing the list, but it should
1634 * help find code using dead lists later on
1635 */
1648 }
1652 {
1660 /* if the blocknr == 0, this has been cleared from the hash,
1661 ** skip it
1662 */
1665 }
1668 next:
1671 }
1673 }
1678 {
1686 }
1690 /* if the blocknr == 0, this has been cleared from the hash,
1691 ** skip it
1692 */
1695 }
1698 /* we can race against journal_mark_freed when we try
1699 * to lock_buffer(cn->bh), so we have to inc the buffer
1700 * count, and recheck things after locking
1701 */
1710 ret++;
1712 /* note, cn->bh might be null now */
1714 }
1716 }
1717 next:
1720 }
1722 }
1724 /* used by flush_commit_list */
1727 {
1735 /* look for a more recent transaction that logged this
1736 ** buffer. Only the most recent transaction with a buffer in
1737 ** it is allowed to send that buffer to disk
1738 */
1743 /* if the buffer is prepared, it will either be logged
1744 * or restored. If restored, we need to make sure
1745 * it actually gets marked dirty
1746 */
1753 }
1754 }
1756 }
1758 }
1765 {
1778 }
1780 /* we've got j_flush_sem held, nobody is going to delete any
1781 * of these lists out from underneath us
1782 */
1791 }
1796 transactions_flushed++;
1800 /* did we wrap? */
1803 }
1806 /* don't bother with older transactions */
1809 }
1812 }
1814 done:
1817 }
1819 /* for o_sync and fsync heavy applications, they tend to use
1820 ** all the journa list slots with tiny transactions. These
1821 ** trigger lots and lots of calls to update the header block, which
1822 ** adds seeks and slows things down.
1823 **
1824 ** This function tries to clear out a large chunk of the journal lists
1825 ** at once, which makes everything faster since only the newest journal
1826 ** list updates the header block
1827 */
1830 {
1843 /* in data logging mode, try harder to flush a lot of blocks */
1846 /* flush for 256 transactions or limit blocks, whichever comes first */
1851 }
1855 }
1861 }
1862 /* try to find a group of blocks we can flush across all the
1863 ** transactions, but only bother if we've actually spanned
1864 ** across multiple lists
1865 */
1868 }
1871 }
1873 /*
1874 ** removes any nodes in table with name block and dev as bh.
1875 ** only touchs the hnext and hprev pointers.
1876 */
1881 {
1888 }
1896 }
1901 }
1909 }
1911 }
1912 }
1915 {
1925 }
1926 /* j_header_bh is on the journal dev, make sure not to release the journal
1927 * dev until we brelse j_header_bh
1928 */
1931 }
1933 /*
1934 ** call on unmount. Only set error to 1 if you haven't made your way out
1935 ** of read_super() yet. Any other caller must keep error at 0.
1936 */
1939 {
1944 /* we only want to flush out transactions if we were called with error == 0
1945 */
1947 /* end the current trans */
1951 /* make sure something gets logged to force our way into the flush code */
1960 }
1961 }
1963 /* this also catches errors during the do_journal_end above */
1973 }
1974 }
1976 reiserfs_mounted_fs_count--;
1977 /* wait for all commits to finish */
1983 }
1988 }
1990 /*
1991 ** call on unmount. flush all journal trans, release all alloc'd ram
1992 */
1995 {
1997 }
1999 /*
2000 ** only call from an error condition inside reiserfs_read_super!
2001 */
2004 {
2006 }
2008 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
2012 {
2018 }
2020 }
2022 /* returns 0 if it did not find a description block
2023 ** returns -1 if it found a corrupt commit block
2024 ** returns 1 if both desc and commit were valid
2025 */
2030 {
2045 "journal-986: transaction "
2046 "is valid returning because trans_id %d is greater than "
2051 }
2055 "journal-1087: transaction "
2056 "is valid returning because mount_id %d is less than "
2061 }
2067 }
2070 /* ok, we have a journal description block, lets see if the transaction was valid */
2071 c_bh =
2081 "journal_transaction_is_valid, commit offset %ld had bad "
2092 "journal-1004: "
2093 "transaction_is_valid setting oldest invalid trans_id "
2096 }
2098 }
2101 "journal-1006: found valid "
2110 }
2111 }
2114 {
2118 }
2119 }
2121 /*
2122 ** given the start, and values for the oldest acceptable transactions,
2123 ** this either reads in a replays a transaction, or returns because the transaction
2124 ** is invalid, or too old.
2125 */
2131 {
2156 cur_dblock -
2160 }
2163 "journal_read_trans skipping because %d is != "
2165 newest_mount_id);
2168 }
2175 }
2179 "journal_read_transaction, "
2188 }
2190 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2203 }
2204 /* get all the buffer heads */
2221 }
2226 }
2227 /* make sure we don't try to replay onto log or reserved area */
2232 abort_replay:
2240 }
2241 }
2242 /* read in the log blocks, memcpy to the corresponding real block */
2257 }
2262 }
2263 /* flush out the real blocks */
2267 }
2280 }
2282 }
2283 cur_dblock =
2291 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2295 /* check for trans_id overflow */
2303 }
2305 /* This function reads blocks starting from block and to max_block of bufsize
2306 size (but no more than BUFNR blocks at a time). This proved to improve
2307 mounting speed on self-rebuilding raid5 arrays at least.
2308 Right now it is only used from journal code. But later we might use it
2309 from other places.
2310 Note: Do not use journal_getblk/sb_getblk functions here! */
2313 b_blocknr_t max_block)
2314 {
2326 }
2336 }
2346 }
2348 /*
2349 ** read and replay the log
2350 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2351 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2352 **
2353 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2354 **
2355 ** On exit, it sets things up so the first transaction will work correctly.
2356 */
2358 {
2380 /* step 1, read in the journal header block. Check the transaction it says
2381 ** is the first unflushed, and if that transaction is not valid,
2382 ** replay is done
2383 */
2389 }
2394 oldest_start =
2400 "journal-1153: found in "
2401 "header: first_unflushed_offset %d, last_flushed_trans_id "
2406 /* now, we try to read the first unflushed offset. If it is not valid,
2407 ** there is nothing more we can do, and it makes no sense to read
2408 ** through the whole log.
2409 */
2410 d_bh =
2417 }
2420 }
2426 }
2428 /* ok, there are transactions that need to be replayed. start with the first log block, find
2429 ** all the valid transactions, and pick out the oldest.
2430 */
2435 /* Note that it is required for blocksize of primary fs device and journal
2436 device to be the same */
2437 d_bh =
2442 ret =
2453 "journal-1179: Setting "
2455 oldest_start -
2456 SB_ONDISK_JOURNAL_1st_BLOCK
2459 /* one we just read was older */
2463 "journal-1180: Resetting "
2465 oldest_start -
2466 SB_ONDISK_JOURNAL_1st_BLOCK
2468 }
2472 "journal-1299: Setting "
2475 }
2478 cur_dblock++;
2479 }
2481 }
2483 start_log_replay:
2487 "journal-1206: Starting replay "
2490 oldest_trans_id);
2492 }
2495 ret =
2502 }
2503 cur_dblock =
2505 replay_count++;
2508 }
2513 }
2514 /* j_start does not get set correctly if we don't replay any transactions.
2515 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2516 ** copy the trans_id from the header
2517 */
2522 /* check for trans_id overflow */
2530 }
2538 }
2542 /* replay failed, caller must call free_journal_ram and abort
2543 ** the mount
2544 */
2546 }
2548 }
2551 {
2563 }
2566 {
2568 }
2572 {
2584 }
2589 result);
2590 }
2592 }
2597 {
2599 dev_t jdev;
2613 /* there is no "jdev" option and journal is on separate device */
2626 }
2637 /* ok */
2643 }
2650 }
2652 }
2654 /**
2655 * When creating/tuning a file system user can assign some
2656 * journal params within boundaries which depend on the ratio
2657 * blocksize/standard_blocksize.
2658 *
2659 * For blocks >= standard_blocksize transaction size should
2660 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2661 * then JOURNAL_TRANS_MAX_DEFAULT.
2662 *
2663 * For blocks < standard_blocksize these boundaries should be
2664 * decreased proportionally.
2665 */
2666 #define REISERFS_STANDARD_BLKSIZE (4096)
2670 {
2672 /* Non-default journal params.
2673 Do sanity check for them. */
2681 JOURNAL_MIN_RATIO) {
2686 }
2693 }
2695 /* Default journal params.
2696 The file system was created by old version
2697 of mkreiserfs, so some fields contain zeros,
2698 and we need to advise proper values for them */
2705 }
2707 }
2709 /*
2710 ** must be called once on fs mount. calls journal_read for you
2711 */
2714 {
2728 }
2741 /* reserved for journal area support */
2743 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2747 REISERFS_DISK_OFFSET_IN_BYTES /
2750 /* Sanity check to see is the standard journal fitting withing first bitmap
2751 (actual for small blocksizes) */
2756 "journal-1393: journal does not fit for area "
2757 "addressed by first of bitmap blocks. It starts at "
2763 }
2769 }
2773 /* read journal header */
2781 }
2784 /* make sure that journal matches to the super block */
2789 "sh-460: journal header magic %x "
2790 "(device %s) does not match to magic found in super "
2796 }
2811 }
2814 "journal first block %u, max trans len %u, max batch %u, "
2859 "allocation failed (%ld bytes). Journal is "
2860 "too large for available memory. Usually "
2864 }
2873 }
2877 }
2879 reiserfs_mounted_fs_count++;
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 {
2974 }
2977 {
2987 }
2993 }
2994 }
2996 /* join == true if you must join an existing transaction.
2997 ** join == false if you can deal with waiting for others to finish
2998 **
2999 ** this will block until the transaction is joinable. send the number of blocks you
3000 ** expect to use in nblocks.
3001 */
3005 {
3017 /* set here for journal_join */
3021 relock:
3027 }
3035 }
3038 /* if there is no room in the journal OR
3039 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3040 ** we don't sleep if there aren't other writers
3041 */
3044 (!join
3061 sched_count++;
3064 }
3065 }
3066 /* don't mess with joining the transaction if all we have to do is
3067 * wait for someone else to do a commit
3068 */
3073 }
3075 }
3080 /* someone might have ended the transaction while we joined */
3085 }
3092 }
3093 /* we are the first writer, set trans_id */
3096 }
3107 out_fail:
3109 /* Re-set th->t_super, so we can properly keep track of how many
3110 * persistent transactions there are. We need to do this so if this
3111 * call is part of a failed restart_transaction, we can free it later */
3114 }
3117 super_block
3120 {
3124 /* if we're nesting into an existing transaction. It will be
3125 ** persistent on its own
3126 */
3133 }
3141 }
3145 }
3148 {
3153 else
3158 }
3160 }
3164 {
3167 /* this keeps do_journal_end from NULLing out the current->journal_info
3168 ** pointer
3169 */
3173 }
3177 {
3180 /* this keeps do_journal_end from NULLing out the current->journal_info
3181 ** pointer
3182 */
3186 }
3190 {
3196 /* we are nesting into the current transaction */
3206 /* we've ended up with a handle from a different filesystem.
3207 ** save it and restore on journal_end. This should never
3208 ** really happen...
3209 */
3214 }
3217 }
3221 /* I guess this boils down to being the reciprocal of clm-2100 above.
3222 * If do_journal_begin_r fails, we need to put it back, since journal_end
3223 * won't be called to do it. */
3226 else
3230 }
3232 /*
3233 ** puts bh into the current transaction. If it was already there, reorders removes the
3234 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3235 **
3236 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3237 ** transaction is committed.
3238 **
3239 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3240 */
3243 {
3255 }
3261 /* already in this transaction, we are done */
3265 }
3267 /* this must be turned into a panic instead of a warning. We can't allow
3268 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3269 ** could get to disk too early. NOT GOOD.
3270 */
3279 }
3286 }
3287 /* this error means I've screwed up, and we've overflowed the transaction.
3288 ** Nothing can be done here, except make the FS readonly or panic.
3289 */
3294 }
3300 }
3304 }
3308 /* now put this guy on the end */
3313 }
3318 }
3329 }
3330 }
3340 }
3342 }
3346 {
3354 }
3361 /* we aren't allowed to close a nested transaction on a different
3362 ** filesystem from the one in the task struct
3363 */
3369 }
3373 }
3374 }
3376 /* removes from the current transaction, relsing and descrementing any counters.
3377 ** also files the removed buffer directly onto the clean list
3378 **
3379 ** called by journal_mark_freed when a block has been deleted
3380 **
3381 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3382 */
3385 {
3394 }
3398 }
3401 }
3404 }
3407 }
3421 }
3423 }
3428 }
3430 /*
3431 ** for any cnode in a journal list, it can only be dirtied of all the
3432 ** transactions that include it are committed to disk.
3433 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3434 ** and 0 if you aren't
3435 **
3436 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3437 ** blocks for a given transaction on disk
3438 **
3439 */
3441 {
3447 /* first test hprev. These are all newer than cn, so any node here
3448 ** with the same block number and dev means this node can't be sent
3449 ** to disk right now.
3450 */
3455 }
3457 }
3458 /* then test hnext. These are all older than cn. As long as they
3459 ** are committed to the log, it is safe to write cn to disk
3460 */
3467 }
3469 }
3471 }
3473 /* syncs the commit blocks, but does not force the real buffers to disk
3474 ** will wait until the current transaction is done/committed before returning
3475 */
3478 {
3482 /* you can sync while nested, very, very bad */
3488 }
3490 }
3492 /*
3493 ** writeback the pending async commits to disk
3494 */
3496 {
3505 /* last entry is the youngest, commit it and you get everything */
3509 }
3511 }
3513 /*
3514 ** flushes any old transactions to disk
3515 ** ends the current transaction if it is too old
3516 */
3518 {
3524 /* safety check so we don't flush while we are replaying the log during
3525 * mount
3526 */
3529 }
3531 /* check the current transaction. If there are no writers, and it is
3532 * too old, finish it, and force the commit blocks to disk
3533 */
3545 /* we're only being called from kreiserfsd, it makes no sense to do
3546 ** an async commit so that kreiserfsd can do it later
3547 */
3549 }
3550 }
3552 }
3554 /*
3555 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3556 **
3557 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3558 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3559 ** flushes the commit list and returns 0.
3560 **
3561 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3562 **
3563 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3564 */
3568 {
3583 }
3586 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3588 }
3590 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3591 ** will be dealt with by next transaction that actually writes something, but should be taken
3592 ** care of in this trans
3593 */
3596 /* if wcount > 0, and we are called to with flush or commit_now,
3597 ** we wait on j_join_wait. We will wake up when the last writer has
3598 ** finished the transaction, and started it on its way to the disk.
3599 ** Then, we flush the commit or journal list, and just return 0
3600 ** because the rest of journal end was already done for this transaction.
3601 */
3613 }
3616 /* sleep while the current transaction is still j_jlocked */
3625 }
3627 }
3628 }
3635 }
3637 }
3640 }
3642 /* deal with old transactions where we are the last writers */
3647 }
3648 /* don't batch when someone is waiting on j_join_wait */
3649 /* don't batch when syncing the commit or flushing the whole trans */
3657 }
3663 }
3665 }
3667 /*
3668 ** Does all the work that makes deleting blocks safe.
3669 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3670 **
3671 ** otherwise:
3672 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3673 ** before this transaction has finished.
3674 **
3675 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3676 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3677 ** the block can't be reallocated yet.
3678 **
3679 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3680 */
3683 {
3695 }
3696 /* if it is journal new, we just remove it from this transaction */
3703 /* set the bit for this block in the journal bitmap for this transaction */
3708 }
3711 /* Note, the entire while loop is not allowed to schedule. */
3716 }
3719 /* find all older transactions with this block, make sure they don't try to write it out */
3721 blocknr);
3727 /* remove_from_transaction will brelse the buffer if it was
3728 ** in the current trans
3729 */
3731 bh);
3734 bh);
3741 }
3742 }
3746 j_nonzerolen));
3747 }
3749 }
3750 }
3752 }
3753 }
3758 }
3761 {
3765 }
3767 /*
3768 * returns -1 on error, 0 if no commits/barriers were done and 1
3769 * if a transaction was actually committed and the barrier was done
3770 */
3773 {
3779 /* is it from the current transaction, or from an unknown transaction? */
3782 /* try to let other writers come in and grow this transaction */
3786 }
3792 /* someone might have ended this transaction while we joined */
3799 }
3806 /* this gets tricky, we have to make sure the journal list in
3807 * the inode still exists. We know the list is still around
3808 * if we've got a larger transaction id than the oldest list
3809 */
3810 flush_commit_only:
3812 /*
3813 * we only set ret to 1 when we know for sure
3814 * the barrier hasn't been started yet on the commit
3815 * block.
3816 */
3822 }
3823 }
3824 /* otherwise the list is gone, and long since committed */
3826 }
3829 {
3833 /* for the whole inode, assume unset id means it was
3834 * changed in the current transaction. More conservative
3835 */
3839 /* jl will be updated in __commit_trans_jl */
3840 }
3843 }
3847 {
3852 }
3862 }
3863 }
3865 }
3868 /*
3869 ** before we can change a metadata block, we have to make sure it won't
3870 ** be written to disk while we are altering it. So, we must:
3871 ** clean it
3872 ** wait on it.
3873 **
3874 */
3877 {
3884 }
3889 }
3892 }
3895 {
3904 /* this check should always be run, to send old lists to disk */
3911 }
3912 }
3913 }
3915 /*
3916 ** long and ugly. If flush, will not return until all commit
3917 ** blocks and all real buffers in the trans are on disk.
3918 ** If no_async, won't return until all commit blocks are on disk.
3919 **
3920 ** keep reading, there are comments as you go along
3921 **
3922 ** If the journal is aborted, we just clean up. Things like flushing
3923 ** journal lists, etc just won't happen.
3924 */
3928 {
3950 /* protect flush_older_commits from doing mistakes if the
3951 transaction ID counter gets overflowed. */
3964 }
3970 }
3974 }
3976 /* check_journal_end locks the journal, and unlocks if it does not return 1
3977 ** it tells us if we should continue with the journal_end, or just return
3978 */
3984 }
3986 /* check_journal_end might set these, check again */
3989 }
3991 /*
3992 ** j must wait means we have to flush the log blocks, and the real blocks for
3993 ** this transaction
3994 */
3997 }
3998 #ifdef REISERFS_PREALLOCATE
3999 /* quota ops might need to nest, setup the journal_info pointer for them
4000 * and raise the refcount so that it is > 0. */
4004 * the transaction */
4007 #endif
4009 /* setup description block */
4010 d_bh =
4020 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
4029 /* init this journal list */
4032 /* we lock the commit before doing anything because
4033 * we want to make sure nobody tries to run flush_commit_list until
4034 * the new transaction is fully setup, and we've already flushed the
4035 * ordered bh list
4036 */
4039 /* save the transaction id in case we need to commit it later */
4052 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4053 ** for each real block, add it to the journal list hash,
4054 ** copy into real block index array in the commit or desc block
4055 */
4063 }
4066 }
4071 }
4073 /* make sure the block we are trying to log is not a block
4074 of journal or reserved area */
4081 }
4094 }
4096 i--;
4097 }
4098 }
4104 /* special check in case all buffers in the journal were marked for not logging */
4107 /* we're about to dirty all the log blocks, mark the description block
4108 * dirty now too. Don't mark the commit block dirty until all the
4109 * others are on disk
4110 */
4113 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4119 /* copy all the real blocks into log area. dirty log blocks */
4124 tmp_bh =
4128 jindex) %
4138 jindex++;
4142 /* JDirty cleared sometime during transaction. don't log this one */
4146 }
4151 }
4153 /* we are done with both the c_bh and d_bh, but
4154 ** c_bh must be written after all other commit blocks,
4155 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4156 */
4160 /* now it is safe to insert this transaction on the main list */
4165 /* reset journal values for the next transaction */
4176 /* check for trans_id overflow */
4186 // make sure reiserfs_add_jh sees the new current_jl before we
4187 // write out the tails
4190 /* tail conversion targets have to hit the disk before we end the
4191 * transaction. Otherwise a later transaction might repack the tail
4192 * before this transaction commits, leaving the data block unflushed and
4193 * clean, if we crash before the later transaction commits, the data block
4194 * is lost.
4195 */
4201 }
4205 /* honor the flush wishes from the caller, simple commits can
4206 ** be done outside the journal lock, they are done below
4207 **
4208 ** if we don't flush the commit list right now, we put it into
4209 ** the work queue so the people waiting on the async progress work
4210 ** queue don't wait for this proc to flush journal lists and such.
4211 */
4218 /* if the next transaction has any chance of wrapping, flush
4219 ** transactions that might get overwritten. If any journal lists are very
4220 ** old flush them as well.
4221 */
4222 first_jl:
4233 /* if we don't cross into the next transaction and we don't
4234 * wrap, there is no way we can overlap any later transactions
4235 * break now
4236 */
4238 }
4248 /* we don't overlap anything from out start to the end of the
4249 * log, and our wrapped portion doesn't overlap anything at
4250 * the start of the log. We can break
4251 */
4253 }
4254 }
4255 }
4264 }
4268 /* wake up any body waiting to join. */
4275 }
4276 out:
4280 /* Re-set th->t_super, so we can properly keep track of how many
4281 * persistent transactions there are. We need to do this so if this
4282 * call is part of a failed restart_transaction, we can free it later */
4286 }
4289 {
4300 #ifdef CONFIG_REISERFS_CHECK
4302 #endif
4303 }
4306 {
4315 }
4318 {
4320 }