| blob | 1b73529b8099bbb7bc758869c09e8ff27d802ce0 |
1 /*
2 ** Write ahead logging implementation copyright Chris Mason 2000
3 **
4 ** The background commits make this code very interelated, and
5 ** overly complex. I need to rethink things a bit....The major players:
6 **
7 ** journal_begin -- call with the number of blocks you expect to log.
8 ** If the current transaction is too
9 ** old, it will block until the current transaction is
10 ** finished, and then start a new one.
11 ** Usually, your transaction will get joined in with
12 ** previous ones for speed.
13 **
14 ** journal_join -- same as journal_begin, but won't block on the current
15 ** transaction regardless of age. Don't ever call
16 ** this. Ever. There are only two places it should be
17 ** called from, and they are both inside this file.
18 **
19 ** journal_mark_dirty -- adds blocks into this transaction. clears any flags
20 ** that might make them get sent to disk
21 ** and then marks them BH_JDirty. Puts the buffer head
22 ** into the current transaction hash.
23 **
24 ** journal_end -- if the current transaction is batchable, it does nothing
25 ** otherwise, it could do an async/synchronous commit, or
26 ** a full flush of all log and real blocks in the
27 ** transaction.
28 **
29 ** flush_old_commits -- if the current transaction is too old, it is ended and
30 ** commit blocks are sent to disk. Forces commit blocks
31 ** to disk for all backgrounded commits that have been
32 ** around too long.
33 ** -- Note, if you call this as an immediate flush from
34 ** from within kupdate, it will ignore the immediate flag
35 */
37 #include <linux/config.h>
38 #include <asm/uaccess.h>
39 #include <asm/system.h>
41 #include <linux/time.h>
42 #include <asm/semaphore.h>
44 #include <linux/vmalloc.h>
45 #include <linux/reiserfs_fs.h>
47 #include <linux/kernel.h>
48 #include <linux/errno.h>
49 #include <linux/fcntl.h>
50 #include <linux/stat.h>
51 #include <linux/string.h>
52 #include <linux/smp_lock.h>
53 #include <linux/buffer_head.h>
54 #include <linux/workqueue.h>
55 #include <linux/writeback.h>
56 #include <linux/blkdev.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 }
619 {
626 }
629 else
633 }
636 {
639 else
643 }
646 {
656 }
659 {
666 }
669 {
676 }
680 {
687 }
688 }
690 #define CHUNK_SIZE 32
694 };
697 {
702 }
705 }
708 {
713 }
716 }
720 {
732 }
734 }
738 {
745 }
747 }
748 }
750 /*
751 * we want to free the jh when the buffer has been written
752 * and waited on
753 */
755 {
768 }
769 }
773 {
781 }
785 no_jh:
789 /* buffer must be locked for __add_jh, should be able to have
790 * two adds at the same time
791 */
796 }
802 }
805 }
808 {
810 }
812 {
814 }
816 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
821 {
840 }
848 }
849 /* in theory, dirty non-uptodate buffers should never get here,
850 * but the upper layer io error paths still have a few quirks.
851 * Handle them here as gracefully as we can
852 */
856 }
864 }
865 loop_next:
868 }
873 }
884 }
887 }
888 /* ugly interaction with invalidatepage here.
889 * reiserfs_invalidate_page will pin any buffer that has a valid
890 * journal head from an older transaction. If someone else sets
891 * our buffer dirty after we write it in the first loop, and
892 * then someone truncates the page away, nobody will ever write
893 * the buffer. We're safe if we write the page one last time
894 * after freeing the journal header.
895 */
900 }
903 }
906 }
910 {
919 find_first:
920 /*
921 * first we walk backwards to find the oldest uncommitted transation
922 */
933 }
935 /* if we didn't find any older uncommitted transactions, return now */
938 }
951 /* list we were called with is gone, return */
955 /* the one we just flushed is gone, this means all
956 * older lists are also gone, so first_jl is no longer
957 * valid either. Go back to the beginning.
958 */
962 }
963 }
969 }
970 }
972 }
974 {
980 }
982 /*
983 ** if this journal list still has commit blocks unflushed, send them to disk.
984 **
985 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
986 ** Before the commit block can by written, every other log block must be safely on disk
987 **
988 */
991 {
1005 }
1009 /* before we can put our commit blocks on disk, we have to make sure everyone older than
1010 ** us is on disk too
1011 */
1018 /* list disappeared during flush_older_commits. return */
1020 }
1021 }
1023 /* make sure nobody is trying to flush this one at the same time */
1028 }
1031 /* this commit is done, exit */
1035 }
1038 }
1048 }
1050 /*
1051 * for the description block and all the log blocks, submit any buffers
1052 * that haven't already reached the disk. Try to write at least 256
1053 * log blocks. later on, we will only wait on blocks that correspond
1054 * to this transaction, but while we're unplugging we might as well
1055 * get a chunk of data on there.
1056 */
1069 }
1070 }
1073 /* We're skipping the commit if there's an error */
1077 /* wait on everything written so far before writing the commit
1078 * if we are in barrier mode, send the commit down now
1079 */
1089 }
1090 }
1096 // since we're using ll_rw_blk above, it might have skipped over
1097 // a locked buffer. Double check here
1098 //
1102 #ifdef CONFIG_REISERFS_CHECK
1104 #endif
1106 }
1110 }
1115 /* If there was a write error in the journal - we can't commit
1116 * this transaction - it will be invalid and, if successful,
1117 * will just end up propogating the write error out to
1118 * the file system. */
1124 }
1130 /* If there was a write error in the journal - we can't commit this
1131 * transaction - it will be invalid and, if successful, will just end
1132 * up propogating the write error out to the filesystem. */
1134 #ifdef CONFIG_REISERFS_CHECK
1136 #endif
1138 }
1144 }
1147 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1152 /* mark the metadata dirty */
1159 }
1161 put_jl:
1166 __FUNCTION__);
1169 }
1171 /*
1172 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1173 ** returns NULL if it can't find anything
1174 */
1176 reiserfs_journal_cnode
1178 {
1186 }
1188 }
1190 }
1197 /*
1198 ** once all the real blocks have been flushed, it is safe to remove them from the
1199 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1200 ** block to be reallocated for data blocks if it had been deleted.
1201 */
1205 {
1210 /* which is better, to lock once around the whole loop, or
1211 ** to lock for each call to remove_journal_hash?
1212 */
1220 }
1224 }
1228 }
1230 }
1232 /*
1233 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1234 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1235 ** releasing blocks in this transaction for reuse as data blocks.
1236 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1237 **
1238 */
1242 {
1253 #ifdef CONFIG_REISERFS_CHECK
1256 #endif
1258 }
1259 }
1263 b_data);
1276 }
1280 sync:
1283 }
1288 }
1289 }
1291 }
1296 {
1298 }
1300 /*
1301 ** flush any and all journal lists older than you are
1302 ** can only be called from flush_journal_list
1303 */
1306 {
1312 /* we know we are the only ones flushing things, no extra race
1313 * protection is required.
1314 */
1315 restart:
1317 /* Did we wrap? */
1323 /* do not flush all */
1326 /* other_jl is now deleted from the list */
1328 }
1330 }
1334 {
1339 }
1340 }
1342 /* flush a journal list, both commit and real blocks
1343 **
1344 ** always set flushall to 1, unless you are calling from inside
1345 ** flush_journal_list
1346 **
1347 ** IMPORTANT. This can only be called while there are no journal writers,
1348 ** and the journal is locked. That means it can only be called from
1349 ** do_journal_end, or by journal_release
1350 */
1353 {
1370 }
1373 /* if flushall == 0, the lock is already held */
1378 }
1386 }
1390 /* if all the work is already done, get out of here */
1394 }
1396 /* start by putting the commit list on disk. This will also flush
1397 ** the commit lists of any olders transactions
1398 */
1405 /* are we done now? */
1409 }
1411 /* loop through each cnode, see if we need to write it,
1412 ** or wait on a more recent transaction, or just ignore it
1413 */
1417 }
1423 /* blocknr of 0 is no longer in the hash, ignore it */
1426 }
1428 /* This transaction failed commit. Don't write out to the disk */
1433 /* the order is important here. We check pjl to make sure we
1434 ** don't clear BH_JDirty_wait if we aren't the one writing this
1435 ** block to disk
1436 */
1440 /* we do this to make sure nobody releases the buffer while
1441 ** we are working with it
1442 */
1450 /* everything with !pjl && jwait should be writable */
1452 }
1453 }
1455 /* if someone has this block in a newer transaction, just make
1456 ** sure they are commited, and don't try writing it to disk
1457 */
1462 }
1464 /* bh == NULL when the block got to disk on its own, OR,
1465 ** the block got freed in a future transaction
1466 */
1469 }
1471 /* this should never happen. kupdate_one_transaction has this list
1472 ** locked while it works, so we should never see a buffer here that
1473 ** is not marked JDirty_wait
1474 */
1477 "journal-813: BAD! buffer %llu %cdirty %cjwait, "
1482 }
1484 /* we inc again because saved_bh gets decremented at free_cnode */
1491 else
1493 count++;
1499 }
1500 free_cnode:
1504 /* we incremented this to keep others from taking the buffer head away */
1509 }
1510 }
1511 }
1519 }
1524 }
1526 #ifdef CONFIG_REISERFS_CHECK
1529 #endif
1531 }
1532 /* note, we must clear the JDirty_wait bit after the up to date
1533 ** check, otherwise we race against our flushpage routine
1534 */
1538 /* undo the inc from journal_mark_dirty */
1541 }
1543 }
1544 }
1549 __FUNCTION__);
1550 flush_older_and_return:
1552 /* before we can update the journal header block, we _must_ flush all
1553 ** real blocks from all older transactions to disk. This is because
1554 ** once the header block is updated, this transaction will not be
1555 ** replayed after a crash
1556 */
1559 }
1562 /* before we can remove everything from the hash tables for this
1563 ** transaction, we must make sure it can never be replayed
1564 **
1565 ** since we are only called from do_journal_end, we know for sure there
1566 ** are no allocations going on while we are flushing journal lists. So,
1567 ** we only need to update the journal header block for the last list
1568 ** being flushed
1569 */
1571 err =
1579 __FUNCTION__);
1580 }
1590 }
1593 /* not strictly required since we are freeing the list, but it should
1594 * help find code using dead lists later on
1595 */
1608 }
1613 {
1621 }
1625 /* if the blocknr == 0, this has been cleared from the hash,
1626 ** skip it
1627 */
1630 }
1633 /* we can race against journal_mark_freed when we try
1634 * to lock_buffer(cn->bh), so we have to inc the buffer
1635 * count, and recheck things after locking
1636 */
1645 ret++;
1647 /* note, cn->bh might be null now */
1649 }
1651 }
1652 next:
1655 }
1657 }
1659 /* used by flush_commit_list */
1662 {
1670 /* look for a more recent transaction that logged this
1671 ** buffer. Only the most recent transaction with a buffer in
1672 ** it is allowed to send that buffer to disk
1673 */
1678 /* if the buffer is prepared, it will either be logged
1679 * or restored. If restored, we need to make sure
1680 * it actually gets marked dirty
1681 */
1688 }
1689 }
1691 }
1693 }
1700 {
1713 }
1715 /* we've got j_flush_sem held, nobody is going to delete any
1716 * of these lists out from underneath us
1717 */
1726 }
1731 transactions_flushed++;
1735 /* did we wrap? */
1738 }
1741 /* don't bother with older transactions */
1744 }
1747 }
1749 done:
1752 }
1754 /* for o_sync and fsync heavy applications, they tend to use
1755 ** all the journa list slots with tiny transactions. These
1756 ** trigger lots and lots of calls to update the header block, which
1757 ** adds seeks and slows things down.
1758 **
1759 ** This function tries to clear out a large chunk of the journal lists
1760 ** at once, which makes everything faster since only the newest journal
1761 ** list updates the header block
1762 */
1765 {
1778 /* in data logging mode, try harder to flush a lot of blocks */
1781 /* flush for 256 transactions or limit blocks, whichever comes first */
1786 }
1790 }
1796 }
1797 /* try to find a group of blocks we can flush across all the
1798 ** transactions, but only bother if we've actually spanned
1799 ** across multiple lists
1800 */
1803 }
1806 }
1808 /*
1809 ** removes any nodes in table with name block and dev as bh.
1810 ** only touchs the hnext and hprev pointers.
1811 */
1816 {
1823 }
1831 }
1836 }
1844 }
1846 }
1847 }
1850 {
1860 }
1861 /* j_header_bh is on the journal dev, make sure not to release the journal
1862 * dev until we brelse j_header_bh
1863 */
1866 }
1868 /*
1869 ** call on unmount. Only set error to 1 if you haven't made your way out
1870 ** of read_super() yet. Any other caller must keep error at 0.
1871 */
1874 {
1879 /* we only want to flush out transactions if we were called with error == 0
1880 */
1882 /* end the current trans */
1886 /* make sure something gets logged to force our way into the flush code */
1895 }
1896 }
1898 /* this also catches errors during the do_journal_end above */
1908 }
1909 }
1911 reiserfs_mounted_fs_count--;
1912 /* wait for all commits to finish */
1918 }
1923 }
1925 /*
1926 ** call on unmount. flush all journal trans, release all alloc'd ram
1927 */
1930 {
1932 }
1934 /*
1935 ** only call from an error condition inside reiserfs_read_super!
1936 */
1939 {
1941 }
1943 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1947 {
1953 }
1955 }
1957 /* returns 0 if it did not find a description block
1958 ** returns -1 if it found a corrupt commit block
1959 ** returns 1 if both desc and commit were valid
1960 */
1965 {
1980 "journal-986: transaction "
1981 "is valid returning because trans_id %d is greater than "
1986 }
1990 "journal-1087: transaction "
1991 "is valid returning because mount_id %d is less than "
1996 }
2002 }
2005 /* ok, we have a journal description block, lets see if the transaction was valid */
2006 c_bh =
2016 "journal_transaction_is_valid, commit offset %ld had bad "
2027 "journal-1004: "
2028 "transaction_is_valid setting oldest invalid trans_id "
2031 }
2033 }
2036 "journal-1006: found valid "
2045 }
2046 }
2049 {
2053 }
2054 }
2056 /*
2057 ** given the start, and values for the oldest acceptable transactions,
2058 ** this either reads in a replays a transaction, or returns because the transaction
2059 ** is invalid, or too old.
2060 */
2066 {
2091 cur_dblock -
2095 }
2098 "journal_read_trans skipping because %d is != "
2100 newest_mount_id);
2103 }
2110 }
2114 "journal_read_transaction, "
2123 }
2125 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2138 }
2139 /* get all the buffer heads */
2156 }
2161 }
2162 /* make sure we don't try to replay onto log or reserved area */
2167 abort_replay:
2175 }
2176 }
2177 /* read in the log blocks, memcpy to the corresponding real block */
2192 }
2197 }
2198 /* flush out the real blocks */
2202 }
2215 }
2217 }
2218 cur_dblock =
2226 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2230 /* check for trans_id overflow */
2238 }
2240 /* This function reads blocks starting from block and to max_block of bufsize
2241 size (but no more than BUFNR blocks at a time). This proved to improve
2242 mounting speed on self-rebuilding raid5 arrays at least.
2243 Right now it is only used from journal code. But later we might use it
2244 from other places.
2245 Note: Do not use journal_getblk/sb_getblk functions here! */
2248 {
2260 }
2270 }
2280 }
2282 /*
2283 ** read and replay the log
2284 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2285 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2286 **
2287 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2288 **
2289 ** On exit, it sets things up so the first transaction will work correctly.
2290 */
2292 {
2314 /* step 1, read in the journal header block. Check the transaction it says
2315 ** is the first unflushed, and if that transaction is not valid,
2316 ** replay is done
2317 */
2323 }
2328 oldest_start =
2334 "journal-1153: found in "
2335 "header: first_unflushed_offset %d, last_flushed_trans_id "
2340 /* now, we try to read the first unflushed offset. If it is not valid,
2341 ** there is nothing more we can do, and it makes no sense to read
2342 ** through the whole log.
2343 */
2344 d_bh =
2351 }
2354 }
2360 }
2362 /* ok, there are transactions that need to be replayed. start with the first log block, find
2363 ** all the valid transactions, and pick out the oldest.
2364 */
2369 /* Note that it is required for blocksize of primary fs device and journal
2370 device to be the same */
2371 d_bh =
2376 ret =
2387 "journal-1179: Setting "
2389 oldest_start -
2390 SB_ONDISK_JOURNAL_1st_BLOCK
2393 /* one we just read was older */
2397 "journal-1180: Resetting "
2399 oldest_start -
2400 SB_ONDISK_JOURNAL_1st_BLOCK
2402 }
2406 "journal-1299: Setting "
2409 }
2412 cur_dblock++;
2413 }
2415 }
2417 start_log_replay:
2421 "journal-1206: Starting replay "
2424 oldest_trans_id);
2426 }
2429 ret =
2436 }
2437 cur_dblock =
2439 replay_count++;
2442 }
2447 }
2448 /* j_start does not get set correctly if we don't replay any transactions.
2449 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2450 ** copy the trans_id from the header
2451 */
2456 /* check for trans_id overflow */
2464 }
2472 }
2476 /* replay failed, caller must call free_journal_ram and abort
2477 ** the mount
2478 */
2480 }
2482 }
2485 {
2497 }
2500 {
2502 }
2506 {
2518 }
2523 result);
2524 }
2526 }
2531 {
2533 dev_t jdev;
2547 /* there is no "jdev" option and journal is on separate device */
2560 }
2571 /* ok */
2577 }
2584 }
2586 }
2588 /*
2589 ** must be called once on fs mount. calls journal_read for you
2590 */
2593 {
2607 }
2620 /* reserved for journal area support */
2622 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2626 REISERFS_DISK_OFFSET_IN_BYTES /
2629 /* Sanity check to see is the standard journal fitting withing first bitmap
2630 (actual for small blocksizes) */
2635 "journal-1393: journal does not fit for area "
2636 "addressed by first of bitmap blocks. It starts at "
2642 }
2648 }
2652 /* read journal header */
2660 }
2663 /* make sure that journal matches to the super block */
2668 "sh-460: journal header magic %x "
2669 "(device %s) does not match to magic found in super "
2675 }
2684 /* make sure these parameters are available, assign it if they are not */
2692 JOURNAL_MIN_RATIO)
2709 }
2712 /*we have the file system was created by old version of mkreiserfs
2713 so this field contains zero value */
2718 /* for blocksize >= 4096 - max transaction size is 1024. For block size < 4096
2719 trans max size is decreased proportionally */
2723 }
2724 }
2731 }
2734 "journal first block %u, max trans len %u, max batch %u, "
2779 "allocation failed (%ld bytes). Journal is "
2780 "too large for available memory. Usually "
2784 }
2793 }
2797 }
2799 reiserfs_mounted_fs_count++;
2805 free_and_return:
2808 }
2810 /*
2811 ** test for a polite end of the current transaction. Used by file_write, and should
2812 ** be used by delete to make sure they don't write more than can fit inside a single
2813 ** transaction
2814 */
2817 {
2820 /* cannot restart while nested */
2830 }
2831 /* protected by the BKL here */
2835 }
2837 /* this must be called inside a transaction, and requires the
2838 ** kernel_lock to be held
2839 */
2841 {
2847 }
2849 /* this must be called without a transaction started, and does not
2850 ** require BKL
2851 */
2853 {
2857 }
2859 /* this must be called without a transaction started, and does not
2860 ** require BKL
2861 */
2863 {
2867 }
2870 {
2871 wait_queue_t wait;
2875 /*
2876 * we don't want to use wait_event here because
2877 * we only want to wait once.
2878 */
2886 }
2889 {
2893 }
2896 {
2906 }
2912 }
2913 }
2915 /* join == true if you must join an existing transaction.
2916 ** join == false if you can deal with waiting for others to finish
2917 **
2918 ** this will block until the transaction is joinable. send the number of blocks you
2919 ** expect to use in nblocks.
2920 */
2924 {
2937 /* set here for journal_join */
2941 relock:
2947 }
2955 }
2958 /* if there is no room in the journal OR
2959 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
2960 ** we don't sleep if there aren't other writers
2961 */
2964 (!join
2981 sched_count++;
2984 }
2985 }
2986 /* don't mess with joining the transaction if all we have to do is
2987 * wait for someone else to do a commit
2988 */
2993 }
2995 }
3000 /* someone might have ended the transaction while we joined */
3005 }
3012 }
3013 /* we are the first writer, set trans_id */
3016 }
3027 out_fail:
3029 /* Re-set th->t_super, so we can properly keep track of how many
3030 * persistent transactions there are. We need to do this so if this
3031 * call is part of a failed restart_transaction, we can free it later */
3034 }
3037 super_block
3040 {
3044 /* if we're nesting into an existing transaction. It will be
3045 ** persistent on its own
3046 */
3052 }
3054 }
3062 }
3066 }
3069 {
3074 else
3079 }
3081 }
3085 {
3088 /* this keeps do_journal_end from NULLing out the current->journal_info
3089 ** pointer
3090 */
3094 }
3096 }
3100 {
3103 /* this keeps do_journal_end from NULLing out the current->journal_info
3104 ** pointer
3105 */
3109 }
3111 }
3115 {
3121 /* we are nesting into the current transaction */
3131 /* we've ended up with a handle from a different filesystem.
3132 ** save it and restore on journal_end. This should never
3133 ** really happen...
3134 */
3139 }
3142 }
3147 /* I guess this boils down to being the reciprocal of clm-2100 above.
3148 * If do_journal_begin_r fails, we need to put it back, since journal_end
3149 * won't be called to do it. */
3152 else
3156 }
3158 /*
3159 ** puts bh into the current transaction. If it was already there, reorders removes the
3160 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3161 **
3162 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3163 ** transaction is committed.
3164 **
3165 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3166 */
3169 {
3181 }
3187 /* already in this transaction, we are done */
3191 }
3193 /* this must be turned into a panic instead of a warning. We can't allow
3194 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3195 ** could get to disk too early. NOT GOOD.
3196 */
3205 }
3212 }
3213 /* this error means I've screwed up, and we've overflowed the transaction.
3214 ** Nothing can be done here, except make the FS readonly or panic.
3215 */
3220 }
3226 }
3230 }
3234 /* now put this guy on the end */
3239 }
3244 }
3255 }
3256 }
3266 }
3268 }
3272 {
3280 }
3287 /* we aren't allowed to close a nested transaction on a different
3288 ** filesystem from the one in the task struct
3289 */
3296 }
3300 }
3301 }
3303 /* removes from the current transaction, relsing and descrementing any counters.
3304 ** also files the removed buffer directly onto the clean list
3305 **
3306 ** called by journal_mark_freed when a block has been deleted
3307 **
3308 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3309 */
3312 {
3321 }
3325 }
3328 }
3331 }
3334 }
3348 }
3350 }
3355 }
3357 /*
3358 ** for any cnode in a journal list, it can only be dirtied of all the
3359 ** transactions that include it are commited to disk.
3360 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3361 ** and 0 if you aren't
3362 **
3363 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3364 ** blocks for a given transaction on disk
3365 **
3366 */
3368 {
3374 /* first test hprev. These are all newer than cn, so any node here
3375 ** with the same block number and dev means this node can't be sent
3376 ** to disk right now.
3377 */
3382 }
3384 }
3385 /* then test hnext. These are all older than cn. As long as they
3386 ** are committed to the log, it is safe to write cn to disk
3387 */
3394 }
3396 }
3398 }
3400 /* syncs the commit blocks, but does not force the real buffers to disk
3401 ** will wait until the current transaction is done/commited before returning
3402 */
3405 {
3409 /* you can sync while nested, very, very bad */
3412 }
3417 }
3419 }
3421 /*
3422 ** writeback the pending async commits to disk
3423 */
3425 {
3433 /* last entry is the youngest, commit it and you get everything */
3437 }
3439 /*
3440 * this is a little racey, but there's no harm in missing
3441 * the filemap_fdata_write
3442 */
3448 }
3449 }
3451 /*
3452 ** flushes any old transactions to disk
3453 ** ends the current transaction if it is too old
3454 */
3456 {
3462 /* safety check so we don't flush while we are replaying the log during
3463 * mount
3464 */
3467 }
3469 /* check the current transaction. If there are no writers, and it is
3470 * too old, finish it, and force the commit blocks to disk
3471 */
3483 /* we're only being called from kreiserfsd, it makes no sense to do
3484 ** an async commit so that kreiserfsd can do it later
3485 */
3487 }
3488 }
3490 }
3492 /*
3493 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3494 **
3495 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3496 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3497 ** flushes the commit list and returns 0.
3498 **
3499 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3500 **
3501 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3502 */
3506 {
3521 }
3524 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3526 }
3528 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3529 ** will be dealt with by next transaction that actually writes something, but should be taken
3530 ** care of in this trans
3531 */
3534 }
3535 /* if wcount > 0, and we are called to with flush or commit_now,
3536 ** we wait on j_join_wait. We will wake up when the last writer has
3537 ** finished the transaction, and started it on its way to the disk.
3538 ** Then, we flush the commit or journal list, and just return 0
3539 ** because the rest of journal end was already done for this transaction.
3540 */
3552 }
3555 /* sleep while the current transaction is still j_jlocked */
3564 }
3566 }
3567 }
3570 }
3575 }
3577 }
3580 }
3582 /* deal with old transactions where we are the last writers */
3587 }
3588 /* don't batch when someone is waiting on j_join_wait */
3589 /* don't batch when syncing the commit or flushing the whole trans */
3597 }
3603 }
3605 }
3607 /*
3608 ** Does all the work that makes deleting blocks safe.
3609 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3610 **
3611 ** otherwise:
3612 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3613 ** before this transaction has finished.
3614 **
3615 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3616 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3617 ** the block can't be reallocated yet.
3618 **
3619 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3620 */
3623 {
3635 }
3636 /* if it is journal new, we just remove it from this transaction */
3643 /* set the bit for this block in the journal bitmap for this transaction */
3648 }
3651 /* Note, the entire while loop is not allowed to schedule. */
3656 }
3659 /* find all older transactions with this block, make sure they don't try to write it out */
3661 blocknr);
3667 /* remove_from_transaction will brelse the buffer if it was
3668 ** in the current trans
3669 */
3671 bh);
3674 bh);
3681 }
3682 }
3686 j_nonzerolen));
3687 }
3689 }
3690 }
3692 }
3693 }
3700 }
3701 }
3703 }
3706 {
3710 }
3712 /*
3713 * returns -1 on error, 0 if no commits/barriers were done and 1
3714 * if a transaction was actually committed and the barrier was done
3715 */
3718 {
3724 /* is it from the current transaction, or from an unknown transaction? */
3727 /* try to let other writers come in and grow this transaction */
3731 }
3737 /* someone might have ended this transaction while we joined */
3744 }
3751 /* this gets tricky, we have to make sure the journal list in
3752 * the inode still exists. We know the list is still around
3753 * if we've got a larger transaction id than the oldest list
3754 */
3755 flush_commit_only:
3757 /*
3758 * we only set ret to 1 when we know for sure
3759 * the barrier hasn't been started yet on the commit
3760 * block.
3761 */
3767 }
3768 }
3769 /* otherwise the list is gone, and long since committed */
3771 }
3774 {
3778 /* for the whole inode, assume unset id means it was
3779 * changed in the current transaction. More conservative
3780 */
3784 /* jl will be updated in __commit_trans_jl */
3785 }
3788 }
3792 {
3797 }
3807 }
3808 }
3810 }
3813 /*
3814 ** before we can change a metadata block, we have to make sure it won't
3815 ** be written to disk while we are altering it. So, we must:
3816 ** clean it
3817 ** wait on it.
3818 **
3819 */
3822 {
3829 }
3834 }
3837 }
3840 {
3849 /* this check should always be run, to send old lists to disk */
3854 }
3855 }
3856 }
3858 /*
3859 ** long and ugly. If flush, will not return until all commit
3860 ** blocks and all real buffers in the trans are on disk.
3861 ** If no_async, won't return until all commit blocks are on disk.
3862 **
3863 ** keep reading, there are comments as you go along
3864 **
3865 ** If the journal is aborted, we just clean up. Things like flushing
3866 ** journal lists, etc just won't happen.
3867 */
3871 {
3893 /* protect flush_older_commits from doing mistakes if the
3894 transaction ID counter gets overflowed. */
3907 }
3913 }
3917 }
3919 /* check_journal_end locks the journal, and unlocks if it does not return 1
3920 ** it tells us if we should continue with the journal_end, or just return
3921 */
3927 }
3929 /* check_journal_end might set these, check again */
3932 }
3934 /*
3935 ** j must wait means we have to flush the log blocks, and the real blocks for
3936 ** this transaction
3937 */
3940 }
3941 #ifdef REISERFS_PREALLOCATE
3942 /* quota ops might need to nest, setup the journal_info pointer for them
3943 * and raise the refcount so that it is > 0. */
3947 * the transaction */
3950 #endif
3952 /* setup description block */
3953 d_bh =
3963 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
3972 /* init this journal list */
3975 /* we lock the commit before doing anything because
3976 * we want to make sure nobody tries to run flush_commit_list until
3977 * the new transaction is fully setup, and we've already flushed the
3978 * ordered bh list
3979 */
3982 /* save the transaction id in case we need to commit it later */
3995 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
3996 ** for each real block, add it to the journal list hash,
3997 ** copy into real block index array in the commit or desc block
3998 */
4006 }
4009 }
4014 }
4016 /* make sure the block we are trying to log is not a block
4017 of journal or reserved area */
4024 }
4037 }
4039 i--;
4040 }
4041 }
4047 /* special check in case all buffers in the journal were marked for not logging */
4050 }
4052 /* we're about to dirty all the log blocks, mark the description block
4053 * dirty now too. Don't mark the commit block dirty until all the
4054 * others are on disk
4055 */
4058 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4064 /* copy all the real blocks into log area. dirty log blocks */
4069 tmp_bh =
4073 jindex) %
4083 jindex++;
4087 /* JDirty cleared sometime during transaction. don't log this one */
4091 }
4096 }
4098 /* we are done with both the c_bh and d_bh, but
4099 ** c_bh must be written after all other commit blocks,
4100 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4101 */
4105 /* now it is safe to insert this transaction on the main list */
4110 /* reset journal values for the next transaction */
4121 /* check for trans_id overflow */
4131 // make sure reiserfs_add_jh sees the new current_jl before we
4132 // write out the tails
4135 /* tail conversion targets have to hit the disk before we end the
4136 * transaction. Otherwise a later transaction might repack the tail
4137 * before this transaction commits, leaving the data block unflushed and
4138 * clean, if we crash before the later transaction commits, the data block
4139 * is lost.
4140 */
4146 }
4151 /* honor the flush wishes from the caller, simple commits can
4152 ** be done outside the journal lock, they are done below
4153 **
4154 ** if we don't flush the commit list right now, we put it into
4155 ** the work queue so the people waiting on the async progress work
4156 ** queue don't wait for this proc to flush journal lists and such.
4157 */
4164 /* if the next transaction has any chance of wrapping, flush
4165 ** transactions that might get overwritten. If any journal lists are very
4166 ** old flush them as well.
4167 */
4168 first_jl:
4179 /* if we don't cross into the next transaction and we don't
4180 * wrap, there is no way we can overlap any later transactions
4181 * break now
4182 */
4184 }
4194 /* we don't overlap anything from out start to the end of the
4195 * log, and our wrapped portion doesn't overlap anything at
4196 * the start of the log. We can break
4197 */
4199 }
4200 }
4201 }
4210 }
4214 /* wake up any body waiting to join. */
4221 }
4222 out:
4226 /* Re-set th->t_super, so we can properly keep track of how many
4227 * persistent transactions there are. We need to do this so if this
4228 * call is part of a failed restart_transaction, we can free it later */
4232 }
4235 {
4246 #ifdef CONFIG_REISERFS_CHECK
4248 #endif
4249 }
4252 {
4261 }
4264 {
4266 }