| blob | 4602fc57c961fd16edc5d558a050493a878fd50e |
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 /*
44 * Number of guaranteed r1bios in case of extreme VM load:
45 */
46 #define NR_RAID1_BIOS 256
48 /* When there are this many requests queue to be written by
49 * the raid1 thread, we become 'congested' to provide back-pressure
50 * for writeback.
51 */
58 {
62 /* allocate a r1bio with room for raid_disks entries in the bios array */
64 }
67 {
69 }
71 #define RESYNC_BLOCK_SIZE (64*1024)
72 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
73 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75 #define RESYNC_WINDOW (2048*1024)
78 {
89 /*
90 * Allocate bios : 1 for reading, n-1 for writing
91 */
97 }
98 /*
99 * Allocate RESYNC_PAGES data pages and attach them to
100 * the first bio.
101 * If this is a user-requested check/repair, allocate
102 * RESYNC_PAGES for each bio.
103 */
106 else
117 }
118 }
119 /* If not user-requests, copy the page pointers to all bios */
125 }
131 out_free_pages:
136 out_free_bio:
141 }
144 {
155 }
160 }
163 {
171 }
172 }
175 {
180 }
183 {
191 }
196 }
199 {
211 }
213 /*
214 * raid_end_bio_io() is called when we have finished servicing a mirrored
215 * operation and are ready to return a success/failure code to the buffer
216 * cache layer.
217 */
219 {
237 /*
238 * Wake up any possible resync thread that waits for the device
239 * to go idle.
240 */
242 }
243 }
246 {
249 /* if nobody has done the final endio yet, do it now */
258 }
260 }
262 /*
263 * Update disk head position estimator based on IRQ completion info.
264 */
266 {
271 }
273 /*
274 * Find the disk number which triggered given bio
275 */
277 {
289 }
292 {
299 /*
300 * this branch is our 'one mirror IO has finished' event handler:
301 */
307 /* If all other devices have failed, we want to return
308 * the error upwards rather than fail the last device.
309 * Here we redefine "uptodate" to mean "Don't want to retry"
310 */
318 }
323 /*
324 * oops, read error:
325 */
328 KERN_ERR "md/raid1:%s: %s: "
332 b),
336 }
339 }
342 {
343 /* it really is the end of this request */
345 /* free extra copy of the data pages */
351 }
352 /* clear the bitmap if all writes complete successfully */
358 }
361 {
371 else
373 }
374 }
377 {
386 /*
387 * 'one mirror IO has finished' event handler:
388 */
394 /*
395 * Set R1BIO_Uptodate in our master bio, so that we
396 * will return a good error code for to the higher
397 * levels even if IO on some other mirrored buffer
398 * fails.
399 *
400 * The 'master' represents the composite IO operation
401 * to user-side. So if something waits for IO, then it
402 * will wait for the 'master' bio.
403 */
404 sector_t first_bad;
411 /* Maybe we can clear some bad blocks. */
417 }
418 }
424 /*
425 * In behind mode, we ACK the master bio once the I/O
426 * has safely reached all non-writemostly
427 * disks. Setting the Returned bit ensures that this
428 * gets done only once -- we don't ever want to return
429 * -EIO here, instead we'll wait
430 */
433 /* Maybe we can return now */
442 }
443 }
444 }
449 /*
450 * Let's see if all mirrored write operations have finished
451 * already.
452 */
457 }
460 /*
461 * This routine returns the disk from which the requested read should
462 * be done. There is a per-array 'next expected sequential IO' sector
463 * number - if this matches on the next IO then we use the last disk.
464 * There is also a per-disk 'last know head position' sector that is
465 * maintained from IRQ contexts, both the normal and the resync IO
466 * completion handlers update this position correctly. If there is no
467 * perfect sequential match then we pick the disk whose head is closest.
468 *
469 * If there are 2 mirrors in the same 2 devices, performance degrades
470 * because position is mirror, not device based.
471 *
472 * The rdev for the device selected will have nr_pending incremented.
473 */
475 {
482 sector_t best_dist;
487 /*
488 * Check if we can balance. We can balance on the whole
489 * device if no resync is going on, or below the resync window.
490 * We take the first readable disk when above the resync window.
491 */
492 retry:
505 }
508 sector_t dist;
509 sector_t first_bad;
525 /* Don't balance among write-mostly, just
526 * use the first as a last resort */
530 }
531 /* This is a reasonable device to use. It might
532 * even be best.
533 */
537 /* already have a better device */
540 /* cannot read here. If this is the 'primary'
541 * device, then we must not read beyond
542 * bad_sectors from another device..
543 */
555 }
558 }
565 /* Don't change to another disk for sequential reads */
568 /* If device is idle, use it */
572 }
576 }
577 }
585 /* cannot risk returning a device that failed
586 * before we inc'ed nr_pending
587 */
590 }
594 }
599 }
602 {
618 /* Note the '|| 1' - when read_balance prefers
619 * non-congested targets, it can be removed
620 */
623 else
625 }
626 }
629 }
633 {
638 }
641 {
642 /* Any writes that have been queued but are awaiting
643 * bitmap updates get flushed here.
644 */
652 /* flush any pending bitmap writes to
653 * disk before proceeding w/ I/O */
662 }
665 }
667 /* Barriers....
668 * Sometimes we need to suspend IO while we do something else,
669 * either some resync/recovery, or reconfigure the array.
670 * To do this we raise a 'barrier'.
671 * The 'barrier' is a counter that can be raised multiple times
672 * to count how many activities are happening which preclude
673 * normal IO.
674 * We can only raise the barrier if there is no pending IO.
675 * i.e. if nr_pending == 0.
676 * We choose only to raise the barrier if no-one is waiting for the
677 * barrier to go down. This means that as soon as an IO request
678 * is ready, no other operations which require a barrier will start
679 * until the IO request has had a chance.
680 *
681 * So: regular IO calls 'wait_barrier'. When that returns there
682 * is no backgroup IO happening, It must arrange to call
683 * allow_barrier when it has finished its IO.
684 * backgroup IO calls must call raise_barrier. Once that returns
685 * there is no normal IO happeing. It must arrange to call
686 * lower_barrier when the particular background IO completes.
687 */
688 #define RESYNC_DEPTH 32
691 {
694 /* Wait until no block IO is waiting */
698 /* block any new IO from starting */
701 /* Now wait for all pending IO to complete */
707 }
710 {
717 }
720 {
726 );
728 }
731 }
734 {
740 }
743 {
744 /* stop syncio and normal IO and wait for everything to
745 * go quite.
746 * We increment barrier and nr_waiting, and then
747 * wait until nr_pending match nr_queued+1
748 * This is called in the context of one normal IO request
749 * that has failed. Thus any sync request that might be pending
750 * will be blocked by nr_pending, and we need to wait for
751 * pending IO requests to complete or be queued for re-try.
752 * Thus the number queued (nr_queued) plus this request (1)
753 * must match the number of pending IOs (nr_pending) before
754 * we continue.
755 */
764 }
766 {
767 /* reverse the effect of the freeze */
773 }
776 /* duplicate the data pages for behind I/O
777 */
779 {
783 GFP_NOIO);
796 }
802 do_sync_io:
808 }
811 {
828 /*
829 * Register the new request and wait if the reconstruction
830 * thread has put up a bar for new requests.
831 * Continue immediately if no resync is active currently.
832 */
839 /* As the suspend_* range is controlled by
840 * userspace, we want an interruptible
841 * wait.
842 */
852 }
854 }
860 /*
861 * make_request() can abort the operation when READA is being
862 * used and no empty request is available.
863 *
864 */
873 /* We might need to issue multiple reads to different
874 * devices if there are bad blocks around, so we keep
875 * track of the number of reads in bio->bi_phys_segments.
876 * If this is 0, there is only one r1_bio and no locking
877 * will be needed when requests complete. If it is
878 * non-zero, then it is the number of not-completed requests.
879 */
884 /*
885 * read balancing logic:
886 */
889 read_again:
893 /* couldn't find anywhere to read from */
896 }
900 bitmap) {
901 /* Reading from a write-mostly device must
902 * take care not to over-take any writes
903 * that are 'behind'
904 */
907 }
912 max_sectors);
923 /* could not read all from this device, so we will
924 * need another r1_bio.
925 */
933 else
936 /* Cannot call generic_make_request directly
937 * as that will be queued in __make_request
938 * and subsequent mempool_alloc might block waiting
939 * for it. So hand bio over to raid1d.
940 */
954 }
956 /*
957 * WRITE:
958 */
963 }
964 /* first select target devices under rcu_lock and
965 * inc refcount on their rdev. Record them by setting
966 * bios[x] to bio
967 * If there are known/acknowledged bad blocks on any device on
968 * which we have seen a write error, we want to avoid writing those
969 * blocks.
970 * This potentially requires several writes to write around
971 * the bad blocks. Each set of writes gets it's own r1bio
972 * with a set of bios attached.
973 */
977 retry_write:
987 }
992 }
996 sector_t first_bad;
1001 max_sectors,
1004 /* mustn't write here until the bad block is
1005 * acknowledged*/
1009 }
1011 /* Cannot write here at all */
1014 /* mustn't write more than bad_sectors
1015 * to other devices yet
1016 */
1019 /* We don't set R1BIO_Degraded as that
1020 * only applies if the disk is
1021 * missing, so it might be re-added,
1022 * and we want to know to recover this
1023 * chunk.
1024 * In this case the device is here,
1025 * and the fact that this chunk is not
1026 * in-sync is recorded in the bad
1027 * block log
1028 */
1030 }
1035 }
1036 }
1038 }
1042 /* Wait for this device to become unblocked */
1053 }
1056 /* We are splitting this write into multiple parts, so
1057 * we need to prepare for allocating another r1_bio.
1058 */
1063 else
1066 }
1082 /* do behind I/O ?
1083 * Not if there are too many, or cannot
1084 * allocate memory, or a reader on WriteMostly
1085 * is waiting for behind writes to flush */
1097 }
1102 /* Yes, I really want the '__' version so that
1103 * we clear any unused pointer in the io_vec, rather
1104 * than leave them unchanged. This is important
1105 * because when we come to free the pages, we won't
1106 * know the original bi_idx, so we just free
1107 * them all
1108 */
1113 }
1129 }
1130 /* Mustn't call r1_bio_write_done before this next test,
1131 * as it could result in the bio being freed.
1132 */
1135 /* We need another r1_bio. It has already been counted
1136 * in bio->bi_phys_segments
1137 */
1145 }
1149 /* In case raid1d snuck in to freeze_array */
1156 }
1159 {
1170 }
1173 }
1177 {
1181 /*
1182 * If it is not operational, then we have already marked it as dead
1183 * else if it is the last working disks, ignore the error, let the
1184 * next level up know.
1185 * else mark the drive as failed
1186 */
1189 /*
1190 * Don't fail the drive, act as though we were just a
1191 * normal single drive.
1192 * However don't try a recovery from this drive as
1193 * it is very likely to fail.
1194 */
1197 }
1205 /*
1206 * if recovery is running, make sure it aborts.
1207 */
1217 }
1220 {
1227 }
1240 }
1242 }
1245 {
1251 }
1254 {
1260 /*
1261 * Find all failed disks within the RAID1 configuration
1262 * and mark them readable.
1263 * Called under mddev lock, so rcu protection not needed.
1264 */
1270 count++;
1272 }
1273 }
1280 }
1284 {
1303 /* as we don't honour merge_bvec_fn, we must
1304 * never risk violating it, so limit
1305 * ->max_segments to one lying with a single
1306 * page, as a one page request is never in
1307 * violation.
1308 */
1313 }
1318 /* As all devices are equivalent, we don't need a full recovery
1319 * if this was recently any drive of the array
1320 */
1325 }
1329 }
1332 {
1345 }
1346 /* Only remove non-faulty devices if recovery
1347 * is not possible.
1348 */
1354 }
1358 /* lost the race, try later */
1362 }
1364 }
1365 abort:
1369 }
1373 {
1378 /*
1379 * we have read a block, now it needs to be re-written,
1380 * or re-read if the read failed.
1381 * We don't do much here, just schedule handling by raid1d
1382 */
1388 }
1391 {
1397 sector_t first_bad;
1406 /* make sure these bits doesn't get cleared. */
1424 )
1435 }
1436 }
1437 }
1441 {
1443 /* success */
1447 /* need to record an error - either for the block or the device */
1451 }
1454 {
1455 /* Try some synchronous reads of other devices to get
1456 * good data, much like with normal read errors. Only
1457 * read into the pages we already have so we don't
1458 * need to re-issue the read request.
1459 * We don't need to freeze the array, because being in an
1460 * active sync request, there is no normal IO, and
1461 * no overlapping syncs.
1462 * We don't need to check is_badblock() again as we
1463 * made sure that anything with a bad block in range
1464 * will have bi_end_io clear.
1465 */
1484 /* No rcu protection needed here devices
1485 * can only be removed when no resync is
1486 * active, and resync is currently active
1487 */
1494 }
1495 }
1496 d++;
1504 /* Cannot read from anywhere, this block is lost.
1505 * Record a bad block on each device. If that doesn't
1506 * work just disable and interrupt the recovery.
1507 * Don't fail devices as that won't really help.
1508 */
1520 }
1528 }
1529 /* Try next page */
1532 idx++;
1534 }
1537 /* write it back and re-read */
1541 d--;
1550 }
1551 }
1556 d--;
1564 }
1567 idx ++;
1568 }
1572 }
1575 {
1576 /* We have read all readable devices. If we haven't
1577 * got the block, then there is no hope left.
1578 * If we have, then we want to do a comparison
1579 * and skip the write if everything is the same.
1580 * If any blocks failed to read, then we need to
1581 * attempt an over-write
1582 */
1594 }
1613 PAGE_SIZE))
1615 }
1622 /* No need to write to this device. */
1626 }
1627 /* fixup the bio for reuse */
1645 else
1650 PAGE_SIZE);
1651 }
1652 }
1654 }
1657 {
1666 /* ouch - failed to read all of that. */
1673 /*
1674 * schedule writes
1675 */
1691 }
1694 /* if we're here, all write(s) have completed, so clean up */
1697 }
1698 }
1700 /*
1701 * This is a kernel thread which:
1702 *
1703 * 1. Retries failed read operations on working mirrors.
1704 * 2. Updates the raid superblock when problems encounter.
1705 * 3. Performs writes following reads for array synchronising.
1706 */
1710 {
1723 /* Note: no rcu protection needed here
1724 * as this is synchronous in the raid1d thread
1725 * which is the thread that might remove
1726 * a device. If raid1d ever becomes multi-threaded....
1727 */
1728 sector_t first_bad;
1740 d++;
1743 }
1747 /* Cannot read from anywhere - mark it bad */
1752 }
1753 /* write it back and re-read */
1758 d--;
1764 }
1770 d--;
1778 "md/raid1:%s: read error corrected "
1784 }
1785 }
1786 }
1789 }
1790 }
1793 {
1795 }
1798 {
1809 }
1812 {
1819 /* bio has the data to be written to device 'i' where
1820 * we just recently had a write error.
1821 * We repeatedly clone the bio and trim down to one block,
1822 * then try the write. Where the write fails we record
1823 * a bad block.
1824 * It is conceivable that the bio doesn't exactly align with
1825 * blocks. We must handle this somehow.
1826 *
1827 * We currently own a reference on the rdev.
1828 */
1831 sector_t sector;
1850 idx++;
1855 }
1860 /* Write at 'sector' for 'sectors'*/
1874 /* failure! */
1883 }
1885 }
1888 {
1899 }
1904 }
1905 }
1908 }
1911 {
1921 /* This drive got a write error. We need to
1922 * narrow down and record precise write
1923 * errors.
1924 */
1928 /* an I/O failed, we can't clear the bitmap */
1930 }
1933 }
1937 }
1940 {
1949 /* we got a read error. Maybe the drive is bad. Maybe just
1950 * the block and we can fix it.
1951 * We freeze all other IO, and try reading the block from
1952 * other devices. When we find one, we re-write
1953 * and check it that fixes the read error.
1954 * This is all done synchronously while the array is
1955 * frozen
1956 */
1967 read_more:
1975 const unsigned long do_sync
1981 }
1988 "md/raid1:%s: redirecting sector %llu"
1999 /* Drat - have to split this up more */
2007 else
2026 }
2027 }
2030 {
2049 }
2061 else
2068 else
2069 /* just a partial read to be scheduled from separate
2070 * context
2071 */
2077 }
2079 }
2083 {
2094 }
2096 /*
2097 * perform a "sync" on one "block"
2098 *
2099 * We need to make sure that no normal I/O request - particularly write
2100 * requests - conflict with active sync requests.
2101 *
2102 * This is achieved by tracking pending requests and a 'barrier' concept
2103 * that can be installed to exclude normal IO requests.
2104 */
2106 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2107 {
2116 sector_t sync_blocks;
2127 /* If we aborted, we need to abort the
2128 * sync on the 'current' bitmap chunk (there will
2129 * only be one in raid1 resync.
2130 * We can find the current addess in mddev->curr_resync
2131 */
2141 }
2149 }
2150 /* before building a request, check if we can skip these blocks..
2151 * This call the bitmap_start_sync doesn't actually record anything
2152 */
2155 /* We can skip this block, and probably several more */
2158 }
2159 /*
2160 * If there is non-resync activity waiting for a turn,
2161 * and resync is going fast enough,
2162 * then let it though before starting on this new sync request.
2163 */
2174 /*
2175 * If we get a correctably read error during resync or recovery,
2176 * we might want to read from a different device. So we
2177 * flag all drives that could conceivably be read from for READ,
2178 * and any others (which will be non-In_sync devices) for WRITE.
2179 * If a read fails, we try reading from something else for which READ
2180 * is OK.
2181 */
2192 /* take from bio_init */
2212 write_targets ++;
2214 /* may need to read from here */
2227 }
2228 }
2236 }
2239 read_targets++;
2240 }
2241 }
2247 }
2248 }
2255 /* These sectors are bad on all InSync devices, so we
2256 * need to mark them bad on all write targets
2257 */
2266 }
2272 /* Cannot record the badblocks, so need to
2273 * abort the resync.
2274 * If there are multiple read targets, could just
2275 * fail the really bad ones ???
2276 */
2283 }
2285 /* only resync enough to reach the next bad->good
2286 * transition */
2288 }
2291 /* extra read targets are also write targets */
2295 /* There is nowhere to write, so all non-sync
2296 * drives must be failed - so we are finished
2297 */
2302 }
2326 }
2333 /* stop here */
2336 i--;
2340 /* remove last page from this bio */
2344 }
2346 }
2347 }
2348 }
2353 bio_full:
2356 /* For a user-requested sync, we read all readable devices and do a
2357 * compare
2358 */
2366 }
2367 }
2374 }
2376 }
2379 {
2384 }
2387 {
2399 GFP_KERNEL);
2412 r1bio_pool_free,
2430 }
2453 /*
2454 * The first working device is used as a
2455 * starting point to read balancing.
2456 */
2458 }
2465 }
2473 }
2477 abort:
2485 }
2487 }
2490 {
2499 }
2504 }
2505 /*
2506 * copy the already verified devices into our private RAID1
2507 * bookkeeping area. [whatever we allocate in run(),
2508 * should be freed in stop()]
2509 */
2512 else
2523 /* as we don't honour merge_bvec_fn, we must never risk
2524 * violating it, so limit ->max_segments to 1 lying within
2525 * a single page, as a one page request is never in violation.
2526 */
2531 }
2532 }
2553 /*
2554 * Ok, everything is just fine now
2555 */
2565 }
2567 }
2570 {
2574 /* wait for behind writes to complete */
2578 /* need to kick something here to make sure I/O goes? */
2581 }
2594 }
2597 {
2598 /* no resync is happening, and there is enough space
2599 * on all devices, so we can resize.
2600 * We need to make sure resync covers any new space.
2601 * If the array is shrinking we should possibly wait until
2602 * any io in the removed space completes, but it hardly seems
2603 * worth it.
2604 */
2614 }
2618 }
2621 {
2622 /* We need to:
2623 * 1/ resize the r1bio_pool
2624 * 2/ resize conf->mirrors
2625 *
2626 * We allocate a new r1bio_pool if we can.
2627 * Then raise a device barrier and wait until all IO stops.
2628 * Then resize conf->mirrors and swap in the new r1bio pool.
2629 *
2630 * At the same time, we "pack" the devices so that all the missing
2631 * devices have the higher raid_disk numbers.
2632 */
2641 /* Cannot change chunk_size, layout, or level */
2649 }
2661 cnt++;
2664 }
2677 }
2683 }
2687 /* ok, everything is stopped */
2701 }
2704 }
2724 }
2727 {
2740 }
2741 }
2744 {
2745 /* raid1 can take over:
2746 * raid5 with 2 devices, any layout or chunk size
2747 */
2757 }
2759 }
2762 {
2780 };
2783 {
2785 }
2788 {
2790 }