| blob | 141de75a6c7c48b72e5b51e8d6991956fa4ac64a |
1 /*
2 * raid1.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5 *
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7 *
8 * RAID-1 management functions.
9 *
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11 *
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
27 * any later version.
28 *
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32 */
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/seq_file.h>
38 #include <linux/ratelimit.h>
43 #define DEBUG 0
44 #define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
46 /*
47 * Number of guaranteed r1bios in case of extreme VM load:
48 */
49 #define NR_RAID1_BIOS 256
56 {
60 /* allocate a r1bio with room for raid_disks entries in the bios array */
62 }
65 {
67 }
69 #define RESYNC_BLOCK_SIZE (64*1024)
70 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
71 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
72 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
73 #define RESYNC_WINDOW (2048*1024)
76 {
87 /*
88 * Allocate bios : 1 for reading, n-1 for writing
89 */
95 }
96 /*
97 * Allocate RESYNC_PAGES data pages and attach them to
98 * the first bio.
99 * If this is a user-requested check/repair, allocate
100 * RESYNC_PAGES for each bio.
101 */
104 else
115 }
116 }
117 /* If not user-requests, copy the page pointers to all bios */
123 }
129 out_free_pages:
134 out_free_bio:
139 }
142 {
153 }
158 }
161 {
169 }
170 }
173 {
178 }
181 {
189 }
194 }
197 {
209 }
211 /*
212 * raid_end_bio_io() is called when we have finished servicing a mirrored
213 * operation and are ready to return a success/failure code to the buffer
214 * cache layer.
215 */
217 {
235 /*
236 * Wake up any possible resync thread that waits for the device
237 * to go idle.
238 */
240 }
241 }
244 {
247 /* if nobody has done the final endio yet, do it now */
256 }
258 }
260 /*
261 * Update disk head position estimator based on IRQ completion info.
262 */
264 {
269 }
272 {
279 /*
280 * this branch is our 'one mirror IO has finished' event handler:
281 */
287 /* If all other devices have failed, we want to return
288 * the error upwards rather than fail the last device.
289 * Here we redefine "uptodate" to mean "Don't want to retry"
290 */
298 }
303 /*
304 * oops, read error:
305 */
308 KERN_ERR "md/raid1:%s: %s: "
312 b),
316 }
319 }
322 {
324 {
325 /* it really is the end of this request */
327 /* free extra copy of the data pages */
333 }
334 /* clear the bitmap if all writes complete successfully */
342 else
344 }
345 }
348 {
360 /*
361 * 'one mirror IO has finished' event handler:
362 */
367 /* an I/O failed, we can't clear the bitmap */
370 /*
371 * Set R1BIO_Uptodate in our master bio, so that we
372 * will return a good error code for to the higher
373 * levels even if IO on some other mirrored buffer
374 * fails.
375 *
376 * The 'master' represents the composite IO operation
377 * to user-side. So if something waits for IO, then it
378 * will wait for the 'master' bio.
379 */
380 sector_t first_bad;
385 /* Maybe we can clear some bad blocks. */
391 }
392 }
400 /*
401 * In behind mode, we ACK the master bio once the I/O
402 * has safely reached all non-writemostly
403 * disks. Setting the Returned bit ensures that this
404 * gets done only once -- we don't ever want to return
405 * -EIO here, instead we'll wait
406 */
409 /* Maybe we can return now */
417 }
418 }
419 }
424 /*
425 * Let's see if all mirrored write operations have finished
426 * already.
427 */
432 }
435 /*
436 * This routine returns the disk from which the requested read should
437 * be done. There is a per-array 'next expected sequential IO' sector
438 * number - if this matches on the next IO then we use the last disk.
439 * There is also a per-disk 'last know head position' sector that is
440 * maintained from IRQ contexts, both the normal and the resync IO
441 * completion handlers update this position correctly. If there is no
442 * perfect sequential match then we pick the disk whose head is closest.
443 *
444 * If there are 2 mirrors in the same 2 devices, performance degrades
445 * because position is mirror, not device based.
446 *
447 * The rdev for the device selected will have nr_pending incremented.
448 */
450 {
457 sector_t best_dist;
462 /*
463 * Check if we can balance. We can balance on the whole
464 * device if no resync is going on, or below the resync window.
465 * We take the first readable disk when above the resync window.
466 */
467 retry:
480 }
483 sector_t dist;
484 sector_t first_bad;
500 /* Don't balance among write-mostly, just
501 * use the first as a last resort */
505 }
506 /* This is a reasonable device to use. It might
507 * even be best.
508 */
512 /* already have a better device */
515 /* cannot read here. If this is the 'primary'
516 * device, then we must not read beyond
517 * bad_sectors from another device..
518 */
530 }
533 }
540 /* Don't change to another disk for sequential reads */
543 /* If device is idle, use it */
547 }
551 }
552 }
560 /* cannot risk returning a device that failed
561 * before we inc'ed nr_pending
562 */
565 }
569 }
574 }
577 {
589 /* Note the '|| 1' - when read_balance prefers
590 * non-congested targets, it can be removed
591 */
594 else
596 }
597 }
600 }
604 {
609 }
612 {
613 /* Any writes that have been queued but are awaiting
614 * bitmap updates get flushed here.
615 */
622 /* flush any pending bitmap writes to
623 * disk before proceeding w/ I/O */
631 }
634 }
636 /* Barriers....
637 * Sometimes we need to suspend IO while we do something else,
638 * either some resync/recovery, or reconfigure the array.
639 * To do this we raise a 'barrier'.
640 * The 'barrier' is a counter that can be raised multiple times
641 * to count how many activities are happening which preclude
642 * normal IO.
643 * We can only raise the barrier if there is no pending IO.
644 * i.e. if nr_pending == 0.
645 * We choose only to raise the barrier if no-one is waiting for the
646 * barrier to go down. This means that as soon as an IO request
647 * is ready, no other operations which require a barrier will start
648 * until the IO request has had a chance.
649 *
650 * So: regular IO calls 'wait_barrier'. When that returns there
651 * is no backgroup IO happening, It must arrange to call
652 * allow_barrier when it has finished its IO.
653 * backgroup IO calls must call raise_barrier. Once that returns
654 * there is no normal IO happeing. It must arrange to call
655 * lower_barrier when the particular background IO completes.
656 */
657 #define RESYNC_DEPTH 32
660 {
663 /* Wait until no block IO is waiting */
667 /* block any new IO from starting */
670 /* Now wait for all pending IO to complete */
676 }
679 {
686 }
689 {
695 );
697 }
700 }
703 {
709 }
712 {
713 /* stop syncio and normal IO and wait for everything to
714 * go quite.
715 * We increment barrier and nr_waiting, and then
716 * wait until nr_pending match nr_queued+1
717 * This is called in the context of one normal IO request
718 * that has failed. Thus any sync request that might be pending
719 * will be blocked by nr_pending, and we need to wait for
720 * pending IO requests to complete or be queued for re-try.
721 * Thus the number queued (nr_queued) plus this request (1)
722 * must match the number of pending IOs (nr_pending) before
723 * we continue.
724 */
733 }
735 {
736 /* reverse the effect of the freeze */
742 }
745 /* duplicate the data pages for behind I/O
746 */
748 {
752 GFP_NOIO);
764 }
770 do_sync_io:
776 }
779 {
796 /*
797 * Register the new request and wait if the reconstruction
798 * thread has put up a bar for new requests.
799 * Continue immediately if no resync is active currently.
800 */
807 /* As the suspend_* range is controlled by
808 * userspace, we want an interruptible
809 * wait.
810 */
820 }
822 }
828 /*
829 * make_request() can abort the operation when READA is being
830 * used and no empty request is available.
831 *
832 */
841 /* We might need to issue multiple reads to different
842 * devices if there are bad blocks around, so we keep
843 * track of the number of reads in bio->bi_phys_segments.
844 * If this is 0, there is only one r1_bio and no locking
845 * will be needed when requests complete. If it is
846 * non-zero, then it is the number of not-completed requests.
847 */
852 /*
853 * read balancing logic:
854 */
857 read_again:
861 /* couldn't find anywhere to read from */
864 }
868 bitmap) {
869 /* Reading from a write-mostly device must
870 * take care not to over-take any writes
871 * that are 'behind'
872 */
875 }
880 max_sectors);
891 /* could not read all from this device, so we will
892 * need another r1_bio.
893 */
901 else
904 /* Cannot call generic_make_request directly
905 * as that will be queued in __make_request
906 * and subsequent mempool_alloc might block waiting
907 * for it. So hand bio over to raid1d.
908 */
922 }
924 /*
925 * WRITE:
926 */
927 /* first select target devices under rcu_lock and
928 * inc refcount on their rdev. Record them by setting
929 * bios[x] to bio
930 * If there are known/acknowledged bad blocks on any device on
931 * which we have seen a write error, we want to avoid writing those
932 * blocks.
933 * This potentially requires several writes to write around
934 * the bad blocks. Each set of writes gets it's own r1bio
935 * with a set of bios attached.
936 */
940 retry_write:
950 }
955 }
959 sector_t first_bad;
964 max_sectors,
967 /* mustn't write here until the bad block is
968 * acknowledged*/
972 }
974 /* Cannot write here at all */
977 /* mustn't write more than bad_sectors
978 * to other devices yet
979 */
982 /* We don't set R1BIO_Degraded as that
983 * only applies if the disk is
984 * missing, so it might be re-added,
985 * and we want to know to recover this
986 * chunk.
987 * In this case the device is here,
988 * and the fact that this chunk is not
989 * in-sync is recorded in the bad
990 * block log
991 */
993 }
998 }
999 }
1001 }
1005 /* Wait for this device to become unblocked */
1016 }
1019 /* We are splitting this write into multiple parts, so
1020 * we need to prepare for allocating another r1_bio.
1021 */
1026 else
1029 }
1045 /* do behind I/O ?
1046 * Not if there are too many, or cannot
1047 * allocate memory, or a reader on WriteMostly
1048 * is waiting for behind writes to flush */
1060 }
1065 /* Yes, I really want the '__' version so that
1066 * we clear any unused pointer in the io_vec, rather
1067 * than leave them unchanged. This is important
1068 * because when we come to free the pages, we won't
1069 * know the original bi_idx, so we just free
1070 * them all
1071 */
1076 }
1091 }
1094 /* In case raid1d snuck in to freeze_array */
1098 /* We need another r1_bio. It has already been counted
1099 * in bio->bi_phys_segments
1100 */
1108 }
1114 }
1117 {
1128 }
1131 }
1135 {
1139 /*
1140 * If it is not operational, then we have already marked it as dead
1141 * else if it is the last working disks, ignore the error, let the
1142 * next level up know.
1143 * else mark the drive as failed
1144 */
1147 /*
1148 * Don't fail the drive, act as though we were just a
1149 * normal single drive.
1150 * However don't try a recovery from this drive as
1151 * it is very likely to fail.
1152 */
1155 }
1163 /*
1164 * if recovery is running, make sure it aborts.
1165 */
1175 }
1178 {
1185 }
1198 }
1200 }
1203 {
1209 }
1212 {
1218 /*
1219 * Find all failed disks within the RAID1 configuration
1220 * and mark them readable.
1221 * Called under mddev lock, so rcu protection not needed.
1222 */
1228 count++;
1230 }
1231 }
1238 }
1242 {
1261 /* as we don't honour merge_bvec_fn, we must
1262 * never risk violating it, so limit
1263 * ->max_segments to one lying with a single
1264 * page, as a one page request is never in
1265 * violation.
1266 */
1271 }
1276 /* As all devices are equivalent, we don't need a full recovery
1277 * if this was recently any drive of the array
1278 */
1283 }
1287 }
1290 {
1303 }
1304 /* Only remove non-faulty devices if recovery
1305 * is not possible.
1306 */
1312 }
1316 /* lost the race, try later */
1320 }
1322 }
1323 abort:
1327 }
1331 {
1340 /*
1341 * we have read a block, now it needs to be re-written,
1342 * or re-read if the read failed.
1343 * We don't do much here, just schedule handling by raid1d
1344 */
1350 }
1353 {
1360 sector_t first_bad;
1367 }
1372 /* make sure these bits doesn't get cleared. */
1395 }
1396 }
1397 }
1400 {
1401 /* Try some synchronous reads of other devices to get
1402 * good data, much like with normal read errors. Only
1403 * read into the pages we already have so we don't
1404 * need to re-issue the read request.
1405 * We don't need to freeze the array, because being in an
1406 * active sync request, there is no normal IO, and
1407 * no overlapping syncs.
1408 * We don't need to check is_badblock() again as we
1409 * made sure that anything with a bad block in range
1410 * will have bi_end_io clear.
1411 */
1430 /* No rcu protection needed here devices
1431 * can only be removed when no resync is
1432 * active, and resync is currently active
1433 */
1440 }
1441 }
1442 d++;
1449 /* Cannot read from anywhere, array is toast */
1459 }
1462 /* write it back and re-read */
1466 d--;
1476 }
1477 }
1482 d--;
1490 else
1492 }
1495 idx ++;
1496 }
1500 }
1503 {
1504 /* We have read all readable devices. If we haven't
1505 * got the block, then there is no hope left.
1506 * If we have, then we want to do a comparison
1507 * and skip the write if everything is the same.
1508 * If any blocks failed to read, then we need to
1509 * attempt an over-write
1510 */
1522 }
1541 PAGE_SIZE))
1543 }
1550 /* No need to write to this device. */
1554 }
1555 /* fixup the bio for reuse */
1573 else
1578 PAGE_SIZE);
1579 }
1580 }
1582 }
1585 {
1594 /* ouch - failed to read all of that. */
1601 /*
1602 * schedule writes
1603 */
1619 }
1622 /* if we're here, all write(s) have completed, so clean up */
1625 }
1626 }
1628 /*
1629 * This is a kernel thread which:
1630 *
1631 * 1. Retries failed read operations on working mirrors.
1632 * 2. Updates the raid superblock when problems encounter.
1633 * 3. Performs writes following reads for array synchronising.
1634 */
1638 {
1651 /* Note: no rcu protection needed here
1652 * as this is synchronous in the raid1d thread
1653 * which is the thread that might remove
1654 * a device. If raid1d ever becomes multi-threaded....
1655 */
1656 sector_t first_bad;
1668 d++;
1671 }
1675 /* Cannot read from anywhere -- bye bye array */
1678 }
1679 /* write it back and re-read */
1684 d--;
1691 /* Well, this device is dead */
1693 }
1694 }
1700 d--;
1707 /* Well, this device is dead */
1712 "md/raid1:%s: read error corrected "
1718 }
1719 }
1720 }
1723 }
1724 }
1727 {
1749 }
1768 rdev,
1771 }
1772 }
1783 rdev,
1787 }
1794 /* we got a read error. Maybe the drive is bad. Maybe just
1795 * the block and we can fix it.
1796 * We freeze all other IO, and try reading the block from
1797 * other devices. When we find one, we re-write
1798 * and check it that fixes the read error.
1799 * This is all done synchronously while the array is
1800 * frozen
1801 */
1814 read_more:
1828 }
1834 max_sectors);
1838 KERN_ERR
1839 "md/raid1:%s: redirecting sector %llu"
1850 /* Drat - have to split this up more */
1859 else
1866 GFP_NOIO);
1881 }
1883 /* just a partial read to be scheduled from separate
1884 * context
1885 */
1887 }
1891 }
1893 }
1897 {
1908 }
1910 /*
1911 * perform a "sync" on one "block"
1912 *
1913 * We need to make sure that no normal I/O request - particularly write
1914 * requests - conflict with active sync requests.
1915 *
1916 * This is achieved by tracking pending requests and a 'barrier' concept
1917 * that can be installed to exclude normal IO requests.
1918 */
1921 {
1930 sector_t sync_blocks;
1941 /* If we aborted, we need to abort the
1942 * sync on the 'current' bitmap chunk (there will
1943 * only be one in raid1 resync.
1944 * We can find the current addess in mddev->curr_resync
1945 */
1955 }
1963 }
1964 /* before building a request, check if we can skip these blocks..
1965 * This call the bitmap_start_sync doesn't actually record anything
1966 */
1969 /* We can skip this block, and probably several more */
1972 }
1973 /*
1974 * If there is non-resync activity waiting for a turn,
1975 * and resync is going fast enough,
1976 * then let it though before starting on this new sync request.
1977 */
1988 /*
1989 * If we get a correctably read error during resync or recovery,
1990 * we might want to read from a different device. So we
1991 * flag all drives that could conceivably be read from for READ,
1992 * and any others (which will be non-In_sync devices) for WRITE.
1993 * If a read fails, we try reading from something else for which READ
1994 * is OK.
1995 */
2006 /* take from bio_init */
2026 write_targets ++;
2028 /* may need to read from here */
2041 }
2042 }
2050 }
2053 read_targets++;
2054 }
2055 }
2061 }
2062 }
2069 /* These sectors are bad on all InSync devices, so we
2070 * need to mark them bad on all write targets
2071 */
2080 }
2086 /* Cannot record the badblocks, so need to
2087 * abort the resync.
2088 * If there are multiple read targets, could just
2089 * fail the really bad ones ???
2090 */
2097 }
2099 /* only resync enough to reach the next bad->good
2100 * transition */
2102 }
2105 /* extra read targets are also write targets */
2109 /* There is nowhere to write, so all non-sync
2110 * drives must be failed - so we are finished
2111 */
2116 }
2140 }
2147 /* stop here */
2150 i--;
2154 /* remove last page from this bio */
2158 }
2160 }
2161 }
2162 }
2167 bio_full:
2170 /* For a user-requested sync, we read all readable devices and do a
2171 * compare
2172 */
2180 }
2181 }
2188 }
2190 }
2193 {
2198 }
2201 {
2213 GFP_KERNEL);
2226 r1bio_pool_free,
2244 }
2265 /*
2266 * The first working device is used as a
2267 * starting point to read balancing.
2268 */
2270 }
2277 }
2285 }
2289 abort:
2297 }
2299 }
2302 {
2311 }
2316 }
2317 /*
2318 * copy the already verified devices into our private RAID1
2319 * bookkeeping area. [whatever we allocate in run(),
2320 * should be freed in stop()]
2321 */
2324 else
2335 /* as we don't honour merge_bvec_fn, we must never risk
2336 * violating it, so limit ->max_segments to 1 lying within
2337 * a single page, as a one page request is never in violation.
2338 */
2343 }
2344 }
2365 /*
2366 * Ok, everything is just fine now
2367 */
2377 }
2379 }
2382 {
2386 /* wait for behind writes to complete */
2390 /* need to kick something here to make sure I/O goes? */
2393 }
2407 }
2410 {
2411 /* no resync is happening, and there is enough space
2412 * on all devices, so we can resize.
2413 * We need to make sure resync covers any new space.
2414 * If the array is shrinking we should possibly wait until
2415 * any io in the removed space completes, but it hardly seems
2416 * worth it.
2417 */
2427 }
2431 }
2434 {
2435 /* We need to:
2436 * 1/ resize the r1bio_pool
2437 * 2/ resize conf->mirrors
2438 *
2439 * We allocate a new r1bio_pool if we can.
2440 * Then raise a device barrier and wait until all IO stops.
2441 * Then resize conf->mirrors and swap in the new r1bio pool.
2442 *
2443 * At the same time, we "pack" the devices so that all the missing
2444 * devices have the higher raid_disk numbers.
2445 */
2454 /* Cannot change chunk_size, layout, or level */
2462 }
2474 cnt++;
2477 }
2490 }
2496 }
2500 /* ok, everything is stopped */
2514 }
2517 }
2537 }
2540 {
2553 }
2554 }
2557 {
2558 /* raid1 can take over:
2559 * raid5 with 2 devices, any layout or chunk size
2560 */
2570 }
2572 }
2575 {
2593 };
2596 {
2598 }
2601 {
2603 }