| blob | f25086aeef5f450afa6c4aa604787cd168ae299f |
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 }
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 {
731 }
733 }
737 {
744 }
746 }
747 }
749 /*
750 * we want to free the jh when the buffer has been written
751 * and waited on
752 */
754 {
767 }
768 }
772 {
780 }
784 no_jh:
788 /* buffer must be locked for __add_jh, should be able to have
789 * two adds at the same time
790 */
794 }
800 }
803 }
806 {
808 }
810 {
812 }
814 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
819 {
837 }
845 }
846 /* in theory, dirty non-uptodate buffers should never get here,
847 * but the upper layer io error paths still have a few quirks.
848 * Handle them here as gracefully as we can
849 */
853 }
860 }
861 loop_next:
864 }
869 }
880 }
883 }
884 /* ugly interaction with invalidatepage here.
885 * reiserfs_invalidate_page will pin any buffer that has a valid
886 * journal head from an older transaction. If someone else sets
887 * our buffer dirty after we write it in the first loop, and
888 * then someone truncates the page away, nobody will ever write
889 * the buffer. We're safe if we write the page one last time
890 * after freeing the journal header.
891 */
896 }
899 }
902 }
906 {
915 find_first:
916 /*
917 * first we walk backwards to find the oldest uncommitted transation
918 */
929 }
931 /* if we didn't find any older uncommitted transactions, return now */
934 }
947 /* list we were called with is gone, return */
951 /* the one we just flushed is gone, this means all
952 * older lists are also gone, so first_jl is no longer
953 * valid either. Go back to the beginning.
954 */
958 }
959 }
965 }
966 }
968 }
970 {
976 }
978 /*
979 ** if this journal list still has commit blocks unflushed, send them to disk.
980 **
981 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
982 ** Before the commit block can by written, every other log block must be safely on disk
983 **
984 */
987 {
1001 }
1005 /* before we can put our commit blocks on disk, we have to make sure everyone older than
1006 ** us is on disk too
1007 */
1014 /* list disappeared during flush_older_commits. return */
1016 }
1017 }
1019 /* make sure nobody is trying to flush this one at the same time */
1024 }
1027 /* this commit is done, exit */
1031 }
1034 }
1044 }
1046 /*
1047 * for the description block and all the log blocks, submit any buffers
1048 * that haven't already reached the disk. Try to write at least 256
1049 * log blocks. later on, we will only wait on blocks that correspond
1050 * to this transaction, but while we're unplugging we might as well
1051 * get a chunk of data on there.
1052 */
1065 }
1066 }
1069 /* We're skipping the commit if there's an error */
1073 /* wait on everything written so far before writing the commit
1074 * if we are in barrier mode, send the commit down now
1075 */
1085 }
1086 }
1092 // since we're using ll_rw_blk above, it might have skipped over
1093 // a locked buffer. Double check here
1094 //
1098 #ifdef CONFIG_REISERFS_CHECK
1100 #endif
1102 }
1106 }
1111 /* If there was a write error in the journal - we can't commit
1112 * this transaction - it will be invalid and, if successful,
1113 * will just end up propagating the write error out to
1114 * the file system. */
1120 }
1126 /* If there was a write error in the journal - we can't commit this
1127 * transaction - it will be invalid and, if successful, will just end
1128 * up propagating the write error out to the filesystem. */
1130 #ifdef CONFIG_REISERFS_CHECK
1132 #endif
1134 }
1140 }
1143 /* now, every commit block is on the disk. It is safe to allow blocks freed during this transaction to be reallocated */
1148 /* mark the metadata dirty */
1155 }
1157 put_jl:
1162 __FUNCTION__);
1165 }
1167 /*
1168 ** flush_journal_list frequently needs to find a newer transaction for a given block. This does that, or
1169 ** returns NULL if it can't find anything
1170 */
1172 reiserfs_journal_cnode
1174 {
1182 }
1184 }
1186 }
1189 {
1199 }
1201 }
1208 /*
1209 ** once all the real blocks have been flushed, it is safe to remove them from the
1210 ** journal list for this transaction. Aside from freeing the cnode, this also allows the
1211 ** block to be reallocated for data blocks if it had been deleted.
1212 */
1216 {
1221 /* which is better, to lock once around the whole loop, or
1222 ** to lock for each call to remove_journal_hash?
1223 */
1231 }
1235 }
1239 }
1241 }
1243 /*
1244 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1245 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1246 ** releasing blocks in this transaction for reuse as data blocks.
1247 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1248 **
1249 */
1253 {
1264 #ifdef CONFIG_REISERFS_CHECK
1267 #endif
1269 }
1270 }
1274 b_data);
1287 }
1291 sync:
1294 }
1299 }
1300 }
1302 }
1307 {
1309 }
1311 /*
1312 ** flush any and all journal lists older than you are
1313 ** can only be called from flush_journal_list
1314 */
1317 {
1323 /* we know we are the only ones flushing things, no extra race
1324 * protection is required.
1325 */
1326 restart:
1328 /* Did we wrap? */
1334 /* do not flush all */
1337 /* other_jl is now deleted from the list */
1339 }
1341 }
1345 {
1350 }
1351 }
1353 /* flush a journal list, both commit and real blocks
1354 **
1355 ** always set flushall to 1, unless you are calling from inside
1356 ** flush_journal_list
1357 **
1358 ** IMPORTANT. This can only be called while there are no journal writers,
1359 ** and the journal is locked. That means it can only be called from
1360 ** do_journal_end, or by journal_release
1361 */
1364 {
1381 }
1384 /* if flushall == 0, the lock is already held */
1389 }
1397 }
1401 /* if all the work is already done, get out of here */
1405 }
1407 /* start by putting the commit list on disk. This will also flush
1408 ** the commit lists of any olders transactions
1409 */
1416 /* are we done now? */
1420 }
1422 /* loop through each cnode, see if we need to write it,
1423 ** or wait on a more recent transaction, or just ignore it
1424 */
1428 }
1434 /* blocknr of 0 is no longer in the hash, ignore it */
1437 }
1439 /* This transaction failed commit. Don't write out to the disk */
1444 /* the order is important here. We check pjl to make sure we
1445 ** don't clear BH_JDirty_wait if we aren't the one writing this
1446 ** block to disk
1447 */
1451 /* we do this to make sure nobody releases the buffer while
1452 ** we are working with it
1453 */
1461 /* everything with !pjl && jwait should be writable */
1463 }
1464 }
1466 /* if someone has this block in a newer transaction, just make
1467 ** sure they are committed, and don't try writing it to disk
1468 */
1473 }
1475 /* bh == NULL when the block got to disk on its own, OR,
1476 ** the block got freed in a future transaction
1477 */
1480 }
1482 /* this should never happen. kupdate_one_transaction has this list
1483 ** locked while it works, so we should never see a buffer here that
1484 ** is not marked JDirty_wait
1485 */
1488 "journal-813: BAD! buffer %llu %cdirty %cjwait, "
1493 }
1495 /* we inc again because saved_bh gets decremented at free_cnode */
1502 else
1504 count++;
1510 }
1511 free_cnode:
1515 /* we incremented this to keep others from taking the buffer head away */
1520 }
1521 }
1522 }
1530 }
1535 }
1537 #ifdef CONFIG_REISERFS_CHECK
1540 #endif
1542 }
1543 /* note, we must clear the JDirty_wait bit after the up to date
1544 ** check, otherwise we race against our flushpage routine
1545 */
1549 /* undo the inc from journal_mark_dirty */
1552 }
1554 }
1555 }
1560 __FUNCTION__);
1561 flush_older_and_return:
1563 /* before we can update the journal header block, we _must_ flush all
1564 ** real blocks from all older transactions to disk. This is because
1565 ** once the header block is updated, this transaction will not be
1566 ** replayed after a crash
1567 */
1570 }
1573 /* before we can remove everything from the hash tables for this
1574 ** transaction, we must make sure it can never be replayed
1575 **
1576 ** since we are only called from do_journal_end, we know for sure there
1577 ** are no allocations going on while we are flushing journal lists. So,
1578 ** we only need to update the journal header block for the last list
1579 ** being flushed
1580 */
1582 err =
1590 __FUNCTION__);
1591 }
1601 }
1604 /* not strictly required since we are freeing the list, but it should
1605 * help find code using dead lists later on
1606 */
1619 }
1623 {
1631 /* if the blocknr == 0, this has been cleared from the hash,
1632 ** skip it
1633 */
1636 }
1639 next:
1642 }
1644 }
1649 {
1657 }
1661 /* if the blocknr == 0, this has been cleared from the hash,
1662 ** skip it
1663 */
1666 }
1669 /* we can race against journal_mark_freed when we try
1670 * to lock_buffer(cn->bh), so we have to inc the buffer
1671 * count, and recheck things after locking
1672 */
1681 ret++;
1683 /* note, cn->bh might be null now */
1685 }
1687 }
1688 next:
1691 }
1693 }
1695 /* used by flush_commit_list */
1698 {
1706 /* look for a more recent transaction that logged this
1707 ** buffer. Only the most recent transaction with a buffer in
1708 ** it is allowed to send that buffer to disk
1709 */
1714 /* if the buffer is prepared, it will either be logged
1715 * or restored. If restored, we need to make sure
1716 * it actually gets marked dirty
1717 */
1724 }
1725 }
1727 }
1729 }
1736 {
1749 }
1751 /* we've got j_flush_sem held, nobody is going to delete any
1752 * of these lists out from underneath us
1753 */
1762 }
1767 transactions_flushed++;
1771 /* did we wrap? */
1774 }
1777 /* don't bother with older transactions */
1780 }
1783 }
1785 done:
1788 }
1790 /* for o_sync and fsync heavy applications, they tend to use
1791 ** all the journa list slots with tiny transactions. These
1792 ** trigger lots and lots of calls to update the header block, which
1793 ** adds seeks and slows things down.
1794 **
1795 ** This function tries to clear out a large chunk of the journal lists
1796 ** at once, which makes everything faster since only the newest journal
1797 ** list updates the header block
1798 */
1801 {
1814 /* in data logging mode, try harder to flush a lot of blocks */
1817 /* flush for 256 transactions or limit blocks, whichever comes first */
1822 }
1826 }
1832 }
1833 /* try to find a group of blocks we can flush across all the
1834 ** transactions, but only bother if we've actually spanned
1835 ** across multiple lists
1836 */
1839 }
1842 }
1844 /*
1845 ** removes any nodes in table with name block and dev as bh.
1846 ** only touchs the hnext and hprev pointers.
1847 */
1852 {
1859 }
1867 }
1872 }
1880 }
1882 }
1883 }
1886 {
1896 }
1897 /* j_header_bh is on the journal dev, make sure not to release the journal
1898 * dev until we brelse j_header_bh
1899 */
1902 }
1904 /*
1905 ** call on unmount. Only set error to 1 if you haven't made your way out
1906 ** of read_super() yet. Any other caller must keep error at 0.
1907 */
1910 {
1915 /* we only want to flush out transactions if we were called with error == 0
1916 */
1918 /* end the current trans */
1922 /* make sure something gets logged to force our way into the flush code */
1931 }
1932 }
1934 /* this also catches errors during the do_journal_end above */
1944 }
1945 }
1947 reiserfs_mounted_fs_count--;
1948 /* wait for all commits to finish */
1954 }
1959 }
1961 /*
1962 ** call on unmount. flush all journal trans, release all alloc'd ram
1963 */
1966 {
1968 }
1970 /*
1971 ** only call from an error condition inside reiserfs_read_super!
1972 */
1975 {
1977 }
1979 /* compares description block with commit block. returns 1 if they differ, 0 if they are the same */
1983 {
1989 }
1991 }
1993 /* returns 0 if it did not find a description block
1994 ** returns -1 if it found a corrupt commit block
1995 ** returns 1 if both desc and commit were valid
1996 */
2001 {
2016 "journal-986: transaction "
2017 "is valid returning because trans_id %d is greater than "
2022 }
2026 "journal-1087: transaction "
2027 "is valid returning because mount_id %d is less than "
2032 }
2038 }
2041 /* ok, we have a journal description block, lets see if the transaction was valid */
2042 c_bh =
2052 "journal_transaction_is_valid, commit offset %ld had bad "
2063 "journal-1004: "
2064 "transaction_is_valid setting oldest invalid trans_id "
2067 }
2069 }
2072 "journal-1006: found valid "
2081 }
2082 }
2085 {
2089 }
2090 }
2092 /*
2093 ** given the start, and values for the oldest acceptable transactions,
2094 ** this either reads in a replays a transaction, or returns because the transaction
2095 ** is invalid, or too old.
2096 */
2102 {
2127 cur_dblock -
2131 }
2134 "journal_read_trans skipping because %d is != "
2136 newest_mount_id);
2139 }
2146 }
2150 "journal_read_transaction, "
2159 }
2161 /* now we know we've got a good transaction, and it was inside the valid time ranges */
2174 }
2175 /* get all the buffer heads */
2192 }
2197 }
2198 /* make sure we don't try to replay onto log or reserved area */
2203 abort_replay:
2211 }
2212 }
2213 /* read in the log blocks, memcpy to the corresponding real block */
2228 }
2233 }
2234 /* flush out the real blocks */
2238 }
2251 }
2253 }
2254 cur_dblock =
2262 /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2266 /* check for trans_id overflow */
2274 }
2276 /* This function reads blocks starting from block and to max_block of bufsize
2277 size (but no more than BUFNR blocks at a time). This proved to improve
2278 mounting speed on self-rebuilding raid5 arrays at least.
2279 Right now it is only used from journal code. But later we might use it
2280 from other places.
2281 Note: Do not use journal_getblk/sb_getblk functions here! */
2284 {
2296 }
2306 }
2316 }
2318 /*
2319 ** read and replay the log
2320 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2321 ** transaction. This tests that before finding all the transactions in the log, which makes normal mount times fast.
2322 **
2323 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2324 **
2325 ** On exit, it sets things up so the first transaction will work correctly.
2326 */
2328 {
2350 /* step 1, read in the journal header block. Check the transaction it says
2351 ** is the first unflushed, and if that transaction is not valid,
2352 ** replay is done
2353 */
2359 }
2364 oldest_start =
2370 "journal-1153: found in "
2371 "header: first_unflushed_offset %d, last_flushed_trans_id "
2376 /* now, we try to read the first unflushed offset. If it is not valid,
2377 ** there is nothing more we can do, and it makes no sense to read
2378 ** through the whole log.
2379 */
2380 d_bh =
2387 }
2390 }
2396 }
2398 /* ok, there are transactions that need to be replayed. start with the first log block, find
2399 ** all the valid transactions, and pick out the oldest.
2400 */
2405 /* Note that it is required for blocksize of primary fs device and journal
2406 device to be the same */
2407 d_bh =
2412 ret =
2423 "journal-1179: Setting "
2425 oldest_start -
2426 SB_ONDISK_JOURNAL_1st_BLOCK
2429 /* one we just read was older */
2433 "journal-1180: Resetting "
2435 oldest_start -
2436 SB_ONDISK_JOURNAL_1st_BLOCK
2438 }
2442 "journal-1299: Setting "
2445 }
2448 cur_dblock++;
2449 }
2451 }
2453 start_log_replay:
2457 "journal-1206: Starting replay "
2460 oldest_trans_id);
2462 }
2465 ret =
2472 }
2473 cur_dblock =
2475 replay_count++;
2478 }
2483 }
2484 /* j_start does not get set correctly if we don't replay any transactions.
2485 ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2486 ** copy the trans_id from the header
2487 */
2492 /* check for trans_id overflow */
2500 }
2508 }
2512 /* replay failed, caller must call free_journal_ram and abort
2513 ** the mount
2514 */
2516 }
2518 }
2521 {
2533 }
2536 {
2538 }
2542 {
2554 }
2559 result);
2560 }
2562 }
2567 {
2569 dev_t jdev;
2583 /* there is no "jdev" option and journal is on separate device */
2596 }
2607 /* ok */
2613 }
2620 }
2622 }
2624 /*
2625 ** must be called once on fs mount. calls journal_read for you
2626 */
2629 {
2643 }
2656 /* reserved for journal area support */
2658 REISERFS_OLD_DISK_OFFSET_IN_BYTES
2662 REISERFS_DISK_OFFSET_IN_BYTES /
2665 /* Sanity check to see is the standard journal fitting withing first bitmap
2666 (actual for small blocksizes) */
2671 "journal-1393: journal does not fit for area "
2672 "addressed by first of bitmap blocks. It starts at "
2678 }
2684 }
2688 /* read journal header */
2696 }
2699 /* make sure that journal matches to the super block */
2704 "sh-460: journal header magic %x "
2705 "(device %s) does not match to magic found in super "
2711 }
2720 /* make sure these parameters are available, assign it if they are not */
2728 JOURNAL_MIN_RATIO)
2745 }
2748 /*we have the file system was created by old version of mkreiserfs
2749 so this field contains zero value */
2754 /* for blocksize >= 4096 - max transaction size is 1024. For block size < 4096
2755 trans max size is decreased proportionally */
2759 }
2760 }
2767 }
2770 "journal first block %u, max trans len %u, max batch %u, "
2815 "allocation failed (%ld bytes). Journal is "
2816 "too large for available memory. Usually "
2820 }
2829 }
2833 }
2835 reiserfs_mounted_fs_count++;
2842 free_and_return:
2845 }
2847 /*
2848 ** test for a polite end of the current transaction. Used by file_write, and should
2849 ** be used by delete to make sure they don't write more than can fit inside a single
2850 ** transaction
2851 */
2854 {
2857 /* cannot restart while nested */
2867 }
2868 /* protected by the BKL here */
2872 }
2874 /* this must be called inside a transaction, and requires the
2875 ** kernel_lock to be held
2876 */
2878 {
2884 }
2886 /* this must be called without a transaction started, and does not
2887 ** require BKL
2888 */
2890 {
2894 }
2896 /* this must be called without a transaction started, and does not
2897 ** require BKL
2898 */
2900 {
2904 }
2907 {
2908 wait_queue_t wait;
2912 /*
2913 * we don't want to use wait_event here because
2914 * we only want to wait once.
2915 */
2923 }
2926 {
2930 }
2933 {
2943 }
2949 }
2950 }
2952 /* join == true if you must join an existing transaction.
2953 ** join == false if you can deal with waiting for others to finish
2954 **
2955 ** this will block until the transaction is joinable. send the number of blocks you
2956 ** expect to use in nblocks.
2957 */
2961 {
2973 /* set here for journal_join */
2977 relock:
2983 }
2991 }
2994 /* if there is no room in the journal OR
2995 ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
2996 ** we don't sleep if there aren't other writers
2997 */
3000 (!join
3017 sched_count++;
3020 }
3021 }
3022 /* don't mess with joining the transaction if all we have to do is
3023 * wait for someone else to do a commit
3024 */
3029 }
3031 }
3036 /* someone might have ended the transaction while we joined */
3041 }
3048 }
3049 /* we are the first writer, set trans_id */
3052 }
3063 out_fail:
3065 /* Re-set th->t_super, so we can properly keep track of how many
3066 * persistent transactions there are. We need to do this so if this
3067 * call is part of a failed restart_transaction, we can free it later */
3070 }
3073 super_block
3076 {
3080 /* if we're nesting into an existing transaction. It will be
3081 ** persistent on its own
3082 */
3089 }
3097 }
3101 }
3104 {
3109 else
3114 }
3116 }
3120 {
3123 /* this keeps do_journal_end from NULLing out the current->journal_info
3124 ** pointer
3125 */
3129 }
3133 {
3136 /* this keeps do_journal_end from NULLing out the current->journal_info
3137 ** pointer
3138 */
3142 }
3146 {
3152 /* we are nesting into the current transaction */
3162 /* we've ended up with a handle from a different filesystem.
3163 ** save it and restore on journal_end. This should never
3164 ** really happen...
3165 */
3170 }
3173 }
3177 /* I guess this boils down to being the reciprocal of clm-2100 above.
3178 * If do_journal_begin_r fails, we need to put it back, since journal_end
3179 * won't be called to do it. */
3182 else
3186 }
3188 /*
3189 ** puts bh into the current transaction. If it was already there, reorders removes the
3190 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3191 **
3192 ** if it was dirty, cleans and files onto the clean list. I can't let it be dirty again until the
3193 ** transaction is committed.
3194 **
3195 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3196 */
3199 {
3211 }
3217 /* already in this transaction, we are done */
3221 }
3223 /* this must be turned into a panic instead of a warning. We can't allow
3224 ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3225 ** could get to disk too early. NOT GOOD.
3226 */
3235 }
3242 }
3243 /* this error means I've screwed up, and we've overflowed the transaction.
3244 ** Nothing can be done here, except make the FS readonly or panic.
3245 */
3250 }
3256 }
3260 }
3264 /* now put this guy on the end */
3269 }
3274 }
3285 }
3286 }
3296 }
3298 }
3302 {
3310 }
3317 /* we aren't allowed to close a nested transaction on a different
3318 ** filesystem from the one in the task struct
3319 */
3325 }
3329 }
3330 }
3332 /* removes from the current transaction, relsing and descrementing any counters.
3333 ** also files the removed buffer directly onto the clean list
3334 **
3335 ** called by journal_mark_freed when a block has been deleted
3336 **
3337 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3338 */
3341 {
3350 }
3354 }
3357 }
3360 }
3363 }
3377 }
3379 }
3384 }
3386 /*
3387 ** for any cnode in a journal list, it can only be dirtied of all the
3388 ** transactions that include it are committed to disk.
3389 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3390 ** and 0 if you aren't
3391 **
3392 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3393 ** blocks for a given transaction on disk
3394 **
3395 */
3397 {
3403 /* first test hprev. These are all newer than cn, so any node here
3404 ** with the same block number and dev means this node can't be sent
3405 ** to disk right now.
3406 */
3411 }
3413 }
3414 /* then test hnext. These are all older than cn. As long as they
3415 ** are committed to the log, it is safe to write cn to disk
3416 */
3423 }
3425 }
3427 }
3429 /* syncs the commit blocks, but does not force the real buffers to disk
3430 ** will wait until the current transaction is done/committed before returning
3431 */
3434 {
3438 /* you can sync while nested, very, very bad */
3444 }
3446 }
3448 /*
3449 ** writeback the pending async commits to disk
3450 */
3452 {
3461 /* last entry is the youngest, commit it and you get everything */
3465 }
3467 }
3469 /*
3470 ** flushes any old transactions to disk
3471 ** ends the current transaction if it is too old
3472 */
3474 {
3480 /* safety check so we don't flush while we are replaying the log during
3481 * mount
3482 */
3485 }
3487 /* check the current transaction. If there are no writers, and it is
3488 * too old, finish it, and force the commit blocks to disk
3489 */
3501 /* we're only being called from kreiserfsd, it makes no sense to do
3502 ** an async commit so that kreiserfsd can do it later
3503 */
3505 }
3506 }
3508 }
3510 /*
3511 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3512 **
3513 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3514 ** the writers are done. By the time it wakes up, the transaction it was called has already ended, so it just
3515 ** flushes the commit list and returns 0.
3516 **
3517 ** Won't batch when flush or commit_now is set. Also won't batch when others are waiting on j_join_wait.
3518 **
3519 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3520 */
3524 {
3539 }
3542 if (atomic_read(&(journal->j_wcount)) > 0) { /* <= 0 is allowed. unmounting might not call begin */
3544 }
3546 /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3547 ** will be dealt with by next transaction that actually writes something, but should be taken
3548 ** care of in this trans
3549 */
3552 /* if wcount > 0, and we are called to with flush or commit_now,
3553 ** we wait on j_join_wait. We will wake up when the last writer has
3554 ** finished the transaction, and started it on its way to the disk.
3555 ** Then, we flush the commit or journal list, and just return 0
3556 ** because the rest of journal end was already done for this transaction.
3557 */
3569 }
3572 /* sleep while the current transaction is still j_jlocked */
3581 }
3583 }
3584 }
3591 }
3593 }
3596 }
3598 /* deal with old transactions where we are the last writers */
3603 }
3604 /* don't batch when someone is waiting on j_join_wait */
3605 /* don't batch when syncing the commit or flushing the whole trans */
3613 }
3619 }
3621 }
3623 /*
3624 ** Does all the work that makes deleting blocks safe.
3625 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3626 **
3627 ** otherwise:
3628 ** set a bit for the block in the journal bitmap. That will prevent it from being allocated for unformatted nodes
3629 ** before this transaction has finished.
3630 **
3631 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. That will prevent any old transactions with
3632 ** this block from trying to flush to the real location. Since we aren't removing the cnode from the journal_list_hash,
3633 ** the block can't be reallocated yet.
3634 **
3635 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3636 */
3639 {
3651 }
3652 /* if it is journal new, we just remove it from this transaction */
3659 /* set the bit for this block in the journal bitmap for this transaction */
3664 }
3667 /* Note, the entire while loop is not allowed to schedule. */
3672 }
3675 /* find all older transactions with this block, make sure they don't try to write it out */
3677 blocknr);
3683 /* remove_from_transaction will brelse the buffer if it was
3684 ** in the current trans
3685 */
3687 bh);
3690 bh);
3697 }
3698 }
3702 j_nonzerolen));
3703 }
3705 }
3706 }
3708 }
3709 }
3716 }
3717 }
3719 }
3722 {
3726 }
3728 /*
3729 * returns -1 on error, 0 if no commits/barriers were done and 1
3730 * if a transaction was actually committed and the barrier was done
3731 */
3734 {
3740 /* is it from the current transaction, or from an unknown transaction? */
3743 /* try to let other writers come in and grow this transaction */
3747 }
3753 /* someone might have ended this transaction while we joined */
3760 }
3767 /* this gets tricky, we have to make sure the journal list in
3768 * the inode still exists. We know the list is still around
3769 * if we've got a larger transaction id than the oldest list
3770 */
3771 flush_commit_only:
3773 /*
3774 * we only set ret to 1 when we know for sure
3775 * the barrier hasn't been started yet on the commit
3776 * block.
3777 */
3783 }
3784 }
3785 /* otherwise the list is gone, and long since committed */
3787 }
3790 {
3794 /* for the whole inode, assume unset id means it was
3795 * changed in the current transaction. More conservative
3796 */
3800 /* jl will be updated in __commit_trans_jl */
3801 }
3804 }
3808 {
3813 }
3823 }
3824 }
3826 }
3829 /*
3830 ** before we can change a metadata block, we have to make sure it won't
3831 ** be written to disk while we are altering it. So, we must:
3832 ** clean it
3833 ** wait on it.
3834 **
3835 */
3838 {
3845 }
3850 }
3853 }
3856 {
3865 /* this check should always be run, to send old lists to disk */
3872 }
3873 }
3874 }
3876 /*
3877 ** long and ugly. If flush, will not return until all commit
3878 ** blocks and all real buffers in the trans are on disk.
3879 ** If no_async, won't return until all commit blocks are on disk.
3880 **
3881 ** keep reading, there are comments as you go along
3882 **
3883 ** If the journal is aborted, we just clean up. Things like flushing
3884 ** journal lists, etc just won't happen.
3885 */
3889 {
3911 /* protect flush_older_commits from doing mistakes if the
3912 transaction ID counter gets overflowed. */
3925 }
3931 }
3935 }
3937 /* check_journal_end locks the journal, and unlocks if it does not return 1
3938 ** it tells us if we should continue with the journal_end, or just return
3939 */
3945 }
3947 /* check_journal_end might set these, check again */
3950 }
3952 /*
3953 ** j must wait means we have to flush the log blocks, and the real blocks for
3954 ** this transaction
3955 */
3958 }
3959 #ifdef REISERFS_PREALLOCATE
3960 /* quota ops might need to nest, setup the journal_info pointer for them
3961 * and raise the refcount so that it is > 0. */
3965 * the transaction */
3968 #endif
3970 /* setup description block */
3971 d_bh =
3981 /* setup commit block. Don't write (keep it clean too) this one until after everyone else is written */
3990 /* init this journal list */
3993 /* we lock the commit before doing anything because
3994 * we want to make sure nobody tries to run flush_commit_list until
3995 * the new transaction is fully setup, and we've already flushed the
3996 * ordered bh list
3997 */
4000 /* save the transaction id in case we need to commit it later */
4013 /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4014 ** for each real block, add it to the journal list hash,
4015 ** copy into real block index array in the commit or desc block
4016 */
4024 }
4027 }
4032 }
4034 /* make sure the block we are trying to log is not a block
4035 of journal or reserved area */
4042 }
4055 }
4057 i--;
4058 }
4059 }
4065 /* special check in case all buffers in the journal were marked for not logging */
4068 /* we're about to dirty all the log blocks, mark the description block
4069 * dirty now too. Don't mark the commit block dirty until all the
4070 * others are on disk
4071 */
4074 /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4080 /* copy all the real blocks into log area. dirty log blocks */
4085 tmp_bh =
4089 jindex) %
4099 jindex++;
4103 /* JDirty cleared sometime during transaction. don't log this one */
4107 }
4112 }
4114 /* we are done with both the c_bh and d_bh, but
4115 ** c_bh must be written after all other commit blocks,
4116 ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4117 */
4121 /* now it is safe to insert this transaction on the main list */
4126 /* reset journal values for the next transaction */
4137 /* check for trans_id overflow */
4147 // make sure reiserfs_add_jh sees the new current_jl before we
4148 // write out the tails
4151 /* tail conversion targets have to hit the disk before we end the
4152 * transaction. Otherwise a later transaction might repack the tail
4153 * before this transaction commits, leaving the data block unflushed and
4154 * clean, if we crash before the later transaction commits, the data block
4155 * is lost.
4156 */
4162 }
4166 /* honor the flush wishes from the caller, simple commits can
4167 ** be done outside the journal lock, they are done below
4168 **
4169 ** if we don't flush the commit list right now, we put it into
4170 ** the work queue so the people waiting on the async progress work
4171 ** queue don't wait for this proc to flush journal lists and such.
4172 */
4179 /* if the next transaction has any chance of wrapping, flush
4180 ** transactions that might get overwritten. If any journal lists are very
4181 ** old flush them as well.
4182 */
4183 first_jl:
4194 /* if we don't cross into the next transaction and we don't
4195 * wrap, there is no way we can overlap any later transactions
4196 * break now
4197 */
4199 }
4209 /* we don't overlap anything from out start to the end of the
4210 * log, and our wrapped portion doesn't overlap anything at
4211 * the start of the log. We can break
4212 */
4214 }
4215 }
4216 }
4225 }
4229 /* wake up any body waiting to join. */
4236 }
4237 out:
4241 /* Re-set th->t_super, so we can properly keep track of how many
4242 * persistent transactions there are. We need to do this so if this
4243 * call is part of a failed restart_transaction, we can free it later */
4247 }
4250 {
4261 #ifdef CONFIG_REISERFS_CHECK
4263 #endif
4264 }
4267 {
4276 }
4279 {
4281 }