| blob | 7d7a8a2524dcab91c7d84bb9af49d827fd370122 |
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * raid10.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 2000-2004 Neil Brown
6 *
7 * RAID-10 support for md.
8 *
9 * Base on code in raid1.c. See raid1.c for further copyright information.
10 */
12 #include <linux/slab.h>
13 #include <linux/delay.h>
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/ratelimit.h>
18 #include <linux/kthread.h>
19 #include <linux/raid/md_p.h>
20 #include <trace/events/block.h>
28 /*
29 * RAID10 provides a combination of RAID0 and RAID1 functionality.
30 * The layout of data is defined by
31 * chunk_size
32 * raid_disks
33 * near_copies (stored in low byte of layout)
34 * far_copies (stored in second byte of layout)
35 * far_offset (stored in bit 16 of layout )
36 * use_far_sets (stored in bit 17 of layout )
37 * use_far_sets_bugfixed (stored in bit 18 of layout )
38 *
39 * The data to be stored is divided into chunks using chunksize. Each device
40 * is divided into far_copies sections. In each section, chunks are laid out
41 * in a style similar to raid0, but near_copies copies of each chunk is stored
42 * (each on a different drive). The starting device for each section is offset
43 * near_copies from the starting device of the previous section. Thus there
44 * are (near_copies * far_copies) of each chunk, and each is on a different
45 * drive. near_copies and far_copies must be at least one, and their product
46 * is at most raid_disks.
47 *
48 * If far_offset is true, then the far_copies are handled a bit differently.
49 * The copies are still in different stripes, but instead of being very far
50 * apart on disk, there are adjacent stripes.
51 *
52 * The far and offset algorithms are handled slightly differently if
53 * 'use_far_sets' is true. In this case, the array's devices are grouped into
54 * sets that are (near_copies * far_copies) in size. The far copied stripes
55 * are still shifted by 'near_copies' devices, but this shifting stays confined
56 * to the set rather than the entire array. This is done to improve the number
57 * of device combinations that can fail without causing the array to fail.
58 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
59 * on a device):
60 * A B C D A B C D E
61 * ... ...
62 * D A B C E A B C D
63 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
64 * [A B] [C D] [A B] [C D E]
65 * |...| |...| |...| | ... |
66 * [B A] [D C] [B A] [E C D]
67 */
81 #define NULL_CMD
82 #define cmd_before(conf, cmd) \
83 do { \
84 write_sequnlock_irq(&(conf)->resync_lock); \
85 cmd; \
86 } while (0)
87 #define cmd_after(conf) write_seqlock_irq(&(conf)->resync_lock)
89 #define wait_event_barrier_cmd(conf, cond, cmd) \
90 wait_event_cmd((conf)->wait_barrier, cond, cmd_before(conf, cmd), \
91 cmd_after(conf))
93 #define wait_event_barrier(conf, cond) \
94 wait_event_barrier_cmd(conf, cond, NULL_CMD)
96 /*
97 * for resync bio, r10bio pointer can be retrieved from the per-bio
98 * 'struct resync_pages'.
99 */
101 {
103 }
106 {
110 /* allocate a r10bio with room for raid_disks entries in the
111 * bios array */
113 }
115 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
116 /* amount of memory to reserve for resync requests */
117 #define RESYNC_WINDOW (1024*1024)
118 /* maximum number of concurrent requests, memory permitting */
119 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
120 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
121 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
123 /*
124 * When performing a resync, we need to read and compare, so
125 * we need as many pages are there are copies.
126 * When performing a recovery, we need 2 bios, one for read,
127 * one for write (we recover only one drive per r10buf)
128 *
129 */
131 {
146 else
149 /* allocate once for all bios */
152 else
158 /*
159 * Allocate bios.
160 */
174 }
175 /*
176 * Allocate RESYNC_PAGES data pages and attach them
177 * where needed.
178 */
196 }
203 }
204 }
208 out_free_pages:
213 out_free_bio:
221 }
223 out_free_r10bio:
226 }
229 {
243 }
249 }
250 }
252 /* resync pages array stored in the 1st bio's .bi_private */
256 }
259 {
271 }
272 }
275 {
280 }
283 {
289 }
292 {
295 }
298 {
308 /* wake up frozen array... */
312 }
314 /*
315 * raid_end_bio_io() is called when we have finished servicing a mirrored
316 * operation and are ready to return a success/failure code to the buffer
317 * cache layer.
318 */
320 {
328 /*
329 * Wake up any possible resync thread that waits for the device
330 * to go idle.
331 */
335 }
337 /*
338 * Update disk head position estimator based on IRQ completion info.
339 */
341 {
346 }
348 /*
349 * Find the disk number which triggered given bio
350 */
353 {
363 }
364 }
373 }
376 {
385 /*
386 * this branch is our 'one mirror IO has finished' event handler:
387 */
391 /*
392 * Set R10BIO_Uptodate in our master bio, so that
393 * we will return a good error code to the higher
394 * levels even if IO on some other mirrored buffer fails.
395 *
396 * The 'master' represents the composite IO operation to
397 * user-side. So if something waits for IO, then it will
398 * wait for the 'master' bio.
399 */
402 /* If all other devices that store this block have
403 * failed, we want to return the error upwards rather
404 * than fail the last device. Here we redefine
405 * "uptodate" to mean "Don't want to retry"
406 */
410 }
415 /*
416 * oops, read error - keep the refcount on the rdev
417 */
424 }
425 }
428 {
431 /* clear the bitmap if all writes complete successfully */
436 }
439 {
447 else
449 }
450 }
451 }
454 {
474 }
475 /*
476 * this branch is our 'one mirror IO has finished' event handler:
477 */
480 /* Never record new bad blocks to replacement,
481 * just fail it.
482 */
494 }
496 /*
497 * When the device is faulty, it is not necessary to
498 * handle write error.
499 */
503 /* Fail the request */
508 }
509 }
511 /*
512 * Set R10BIO_Uptodate in our master bio, so that
513 * we will return a good error code for to the higher
514 * levels even if IO on some other mirrored buffer fails.
515 *
516 * The 'master' represents the composite IO operation to
517 * user-side. So if something waits for IO, then it will
518 * wait for the 'master' bio.
519 *
520 * Do not set R10BIO_Uptodate if the current device is
521 * rebuilding or Faulty. This is because we cannot use
522 * such device for properly reading the data back (we could
523 * potentially use it, if the current write would have felt
524 * before rdev->recovery_offset, but for simplicity we don't
525 * check this here.
526 */
531 /* Maybe we can clear some bad blocks. */
538 else
542 }
543 }
545 /*
546 *
547 * Let's see if all mirrored write operations have finished
548 * already.
549 */
555 }
557 /*
558 * RAID10 layout manager
559 * As well as the chunksize and raid_disks count, there are two
560 * parameters: near_copies and far_copies.
561 * near_copies * far_copies must be <= raid_disks.
562 * Normally one of these will be 1.
563 * If both are 1, we get raid0.
564 * If near_copies == raid_disks, we get raid1.
565 *
566 * Chunks are laid out in raid0 style with near_copies copies of the
567 * first chunk, followed by near_copies copies of the next chunk and
568 * so on.
569 * If far_copies > 1, then after 1/far_copies of the array has been assigned
570 * as described above, we start again with a device offset of near_copies.
571 * So we effectively have another copy of the whole array further down all
572 * the drives, but with blocks on different drives.
573 * With this layout, and block is never stored twice on the one device.
574 *
575 * raid10_find_phys finds the sector offset of a given virtual sector
576 * on each device that it is on.
577 *
578 * raid10_find_virt does the reverse mapping, from a device and a
579 * sector offset to a virtual address
580 */
583 {
585 sector_t sector;
586 sector_t chunk;
587 sector_t stripe;
598 /* now calculate first sector/dev */
610 /* and calculate all the others */
617 slot++;
631 }
635 slot++;
636 }
637 dev++;
641 }
642 }
643 }
646 {
658 }
661 {
663 /* Never use conf->prev as this is only called during resync
664 * or recovery, so reshape isn't happening
665 */
679 }
680 }
695 else
697 }
699 }
703 }
705 /*
706 * This routine returns the disk from which the requested read should
707 * be done. There is a per-array 'next expected sequential IO' sector
708 * number - if this matches on the next IO then we use the last disk.
709 * There is also a per-disk 'last know head position' sector that is
710 * maintained from IRQ contexts, both the normal and the resync IO
711 * completion handlers update this position correctly. If there is no
712 * perfect sequential match then we pick the disk whose head is closest.
713 *
714 * If there are 2 mirrors in the same 2 devices, performance degrades
715 * because position is mirror, not device based.
716 *
717 * The rdev for the device selected will have nr_pending incremented.
718 */
720 /*
721 * FIXME: possibly should rethink readbalancing and do it differently
722 * depending on near_copies / far_copies geometry.
723 */
727 {
754 sector_t first_bad;
756 sector_t dev_sector;
779 /* Already have a better slot */
782 /* Cannot read here. If this is the
783 * 'primary' device, then we must not read
784 * beyond 'bad_sectors' from another device.
785 */
792 sector_t good_sectors =
798 }
800 /* Must read from here */
802 }
817 }
820 /* At least 2 disks to choose from so failfast is OK */
822 /* This optimisation is debatable, and completely destroys
823 * sequential read speed for 'far copies' arrays. So only
824 * keep it for 'near' arrays, and review those later.
825 */
829 /* for far > 1 always use the lowest address */
832 else
840 }
841 }
849 }
850 }
860 }
863 {
864 /* Any writes that have been queued but are awaiting
865 * bitmap updates get flushed here.
866 */
876 /*
877 * As this is called in a wait_event() loop (see freeze_array),
878 * current->state might be TASK_UNINTERRUPTIBLE which will
879 * cause a warning when we prepare to wait again. As it is
880 * rare that this path is taken, it is perfectly safe to force
881 * us to go around the wait_event() loop again, so the warning
882 * is a false-positive. Silence the warning by resetting
883 * thread state
884 */
897 }
901 }
903 /* Barriers....
904 * Sometimes we need to suspend IO while we do something else,
905 * either some resync/recovery, or reconfigure the array.
906 * To do this we raise a 'barrier'.
907 * The 'barrier' is a counter that can be raised multiple times
908 * to count how many activities are happening which preclude
909 * normal IO.
910 * We can only raise the barrier if there is no pending IO.
911 * i.e. if nr_pending == 0.
912 * We choose only to raise the barrier if no-one is waiting for the
913 * barrier to go down. This means that as soon as an IO request
914 * is ready, no other operations which require a barrier will start
915 * until the IO request has had a chance.
916 *
917 * So: regular IO calls 'wait_barrier'. When that returns there
918 * is no backgroup IO happening, It must arrange to call
919 * allow_barrier when it has finished its IO.
920 * backgroup IO calls must call raise_barrier. Once that returns
921 * there is no normal IO happeing. It must arrange to call
922 * lower_barrier when the particular background IO completes.
923 */
926 {
932 /* Wait until no block IO is waiting (unless 'force') */
935 /* block any new IO from starting */
938 /* Now wait for all pending IO to complete */
943 }
946 {
953 }
956 {
960 /* barrier is dropped */
964 /*
965 * If there are already pending requests (preventing the barrier from
966 * rising completely), and the pre-process bio queue isn't empty, then
967 * don't wait, as we need to empty that queue to get the nr_pending
968 * count down.
969 */
974 /* daemon thread must exist while handling io */
976 /*
977 * move on if io is issued from raid10d(), nr_pending is not released
978 * from original io(see handle_read_error()). All raise barrier is
979 * blocked until this io is done.
980 */
984 }
987 }
990 {
1004 }
1007 {
1015 /* Return false when nowait flag is set */
1023 }
1026 }
1027 /* Only increment nr_pending when we wait */
1032 }
1035 {
1039 }
1042 {
1043 /* stop syncio and normal IO and wait for everything to
1044 * go quiet.
1045 * We increment barrier and nr_waiting, and then
1046 * wait until nr_pending match nr_queued+extra
1047 * This is called in the context of one normal IO request
1048 * that has failed. Thus any sync request that might be pending
1049 * will be blocked by nr_pending, and we need to wait for
1050 * pending IO requests to complete or be queued for re-try.
1051 * Thus the number queued (nr_queued) plus this request (extra)
1052 * must match the number of pending IOs (nr_pending) before
1053 * we continue.
1054 */
1063 }
1066 {
1067 /* reverse the effect of the freeze */
1073 }
1077 {
1081 else
1083 }
1086 {
1100 }
1102 /* we aren't scheduling, so we can do the write-out directly. */
1113 }
1115 }
1117 /*
1118 * 1. Register the new request and wait if the reconstruction thread has put
1119 * up a bar for new requests. Continue immediately if no resync is active
1120 * currently.
1121 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1122 */
1125 {
1126 /* Bail out if REQ_NOWAIT is set for the bio */
1130 }
1138 }
1143 sectors);
1145 }
1147 }
1151 {
1165 /*
1166 * This is an error retry, but we cannot
1167 * safely dereference the rdev in the r10_bio,
1168 * we must use the one in conf.
1169 * If it has already been disconnected (unlikely)
1170 * we lose the device name in error messages.
1171 */
1173 /*
1174 * As we are blocking raid10, it is a little safer to
1175 * use __GFP_HIGH.
1176 */
1185 /* This never gets dereferenced */
1187 }
1188 }
1198 }
1201 }
1213 }
1221 }
1227 }
1244 err_handle:
1249 }
1254 {
1271 else
1284 /* flush_pending_writes() needs access to the rdev so...*/
1294 }
1295 }
1298 {
1303 retry_wait:
1312 /*
1313 * Discard request doesn't care the write result
1314 * so it doesn't need to wait blocked disk here.
1315 */
1320 /*
1321 * Mustn't write here until the bad
1322 * block is acknowledged
1323 */
1330 }
1331 }
1338 }
1339 }
1342 /* Have to wait for this device to get unblocked, then retry */
1350 }
1351 }
1355 {
1358 sector_t sectors;
1367 /* Bail out if REQ_NOWAIT is set for the bio */
1371 }
1379 }
1381 }
1392 /* Need to update reshape_position in metadata */
1401 }
1408 }
1410 /* first select target devices under rcu_lock and
1411 * inc refcount on their rdev. Record them by setting
1412 * bios[x] to bio
1413 * If there are known/acknowledged bad blocks on any device
1414 * on which we have seen a write error, we want to avoid
1415 * writing to those blocks. This potentially requires several
1416 * writes to write around the bad blocks. Each set of writes
1417 * gets its own r10_bio with a set of bios attached.
1418 */
1444 }
1446 sector_t first_bad;
1454 /* Cannot write here at all */
1457 /* Mustn't write more than bad_sectors
1458 * to other devices yet
1459 */
1461 /* We don't set R10BIO_Degraded as that
1462 * only applies if the disk is missing,
1463 * so it might be re-added, and we want to
1464 * know to recover this chunk.
1465 * In this case the device is here, and the
1466 * fact that this chunk is not in-sync is
1467 * recorded in the bad block log.
1468 */
1470 }
1474 /*
1475 * We cannot atomically write this, so just
1476 * error in that case. It could be possible to
1477 * atomically write other mirrors, but the
1478 * complexity of supporting that is not worth
1479 * the benefit.
1480 */
1484 }
1489 }
1490 }
1494 }
1498 }
1499 }
1510 }
1517 }
1530 }
1533 err_handle:
1542 }
1546 }
1547 }
1552 }
1555 {
1573 else
1575 }
1578 {
1595 }
1596 }
1597 }
1600 {
1607 /*
1608 * We don't care the return value of discard bio
1609 */
1619 }
1621 /*
1622 * There are some limitations to handle discard bio
1623 * 1st, the discard size is bigger than stripe_size*2.
1624 * 2st, if the discard bio spans reshape progress, we use the old way to
1625 * handle discard bio
1626 */
1628 {
1636 sector_t chunk;
1639 sector_t split_size;
1642 sector_t start_disk_offset;
1644 sector_t end_disk_offset;
1654 }
1657 /*
1658 * Check reshape again to avoid reshape happens after checking
1659 * MD_RECOVERY_RESHAPE and before wait_barrier
1660 */
1667 else
1675 /*
1676 * Maybe one discard bio is smaller than strip size or across one
1677 * stripe and discard region is larger than one stripe size. For far
1678 * offset layout, if the discard region is not aligned with stripe
1679 * size, there is hole when we submit discard bio to member disk.
1680 * For simplicity, we only handle discard bio which discard region
1681 * is bigger than stripe_size * 2
1682 */
1686 /*
1687 * Keep bio aligned with strip size.
1688 */
1697 }
1700 /* Resend the fist split part */
1703 }
1712 }
1715 /* Resend the second split part */
1719 }
1724 /*
1725 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1726 * One stripe contains the chunks from all member disk (one chunk from
1727 * one disk at the same HBA address). For layout detail, see 'man md 4'
1728 */
1747 retry_discard:
1755 /*
1756 * For far layout it needs more than one r10bio to cover all regions.
1757 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1758 * to record the discard bio. Other r10bio->master_bio record the first
1759 * r10bio. The first r10bio only release after all other r10bios finish.
1760 * The discard bio returns only first r10bio finishes
1761 */
1770 /*
1771 * first select target devices under rcu_lock and
1772 * inc refcount on their rdev. Record them by setting
1773 * bios[x] to bio
1774 */
1793 }
1797 }
1798 }
1805 /*
1806 * Now start to calculate the start and end address for each disk.
1807 * The space between dev_start and dev_end is the discard region.
1808 *
1809 * For dev_start, it needs to consider three conditions:
1810 * 1st, the disk is before start_disk, you can imagine the disk in
1811 * the next stripe. So the dev_start is the start address of next
1812 * stripe.
1813 * 2st, the disk is after start_disk, it means the disk is at the
1814 * same stripe of first disk
1815 * 3st, the first disk itself, we can use start_disk_offset directly
1816 */
1821 else
1828 else
1831 /*
1832 * It only handles discard bio which size is >= stripe size, so
1833 * dev_end > dev_start all the time.
1834 * It doesn't need to use rcu lock to get rdev here. We already
1835 * add rdev->nr_pending in the first loop.
1836 */
1850 }
1864 }
1865 }
1876 }
1881 out:
1884 }
1887 {
1903 /*
1904 * If this request crosses a chunk boundary, we need to split
1905 * it.
1906 */
1908 sectors > chunk_sects
1917 /* In case raid10d snuck in to freeze_array */
1920 }
1923 {
1936 else
1940 }
1947 }
1949 }
1951 /* check if there are enough drives for
1952 * every block to appear on atleast one.
1953 * Don't consider the device numbered 'ignore'
1954 * as we might be about to remove it.
1955 */
1957 {
1967 }
1978 cnt++;
1980 }
1986 out:
1988 }
1991 {
1992 /* when calling 'enough', both 'prev' and 'geo' must
1993 * be stable.
1994 * This is ensured if ->reconfig_mutex or ->device_lock
1995 * is held.
1996 */
1999 }
2001 /**
2002 * raid10_error() - RAID10 error handler.
2003 * @mddev: affected md device.
2004 * @rdev: member device to fail.
2005 *
2006 * The routine acknowledges &rdev failure and determines new @mddev state.
2007 * If it failed, then:
2008 * - &MD_BROKEN flag is set in &mddev->flags.
2009 * Otherwise, it must be degraded:
2010 * - recovery is interrupted.
2011 * - &mddev->degraded is bumped.
2012 *
2013 * @rdev is marked as &Faulty excluding case when array is failed and
2014 * &mddev->fail_last_dev is off.
2015 */
2017 {
2029 }
2030 }
2044 }
2047 {
2055 }
2067 }
2068 }
2071 {
2076 }
2079 {
2086 /*
2087 * Find all non-in_sync disks within the RAID10 configuration
2088 * and mark them in_sync
2089 */
2096 /* Replacement has just become active */
2099 count++;
2101 /* Replaced device not technically faulty,
2102 * but we need to be sure it gets removed
2103 * and never re-added.
2104 */
2108 }
2114 count++;
2116 }
2117 }
2124 }
2127 {
2136 /* only hot-add to in-sync arrays, as recovery is
2137 * very different from resync
2138 */
2150 else
2161 }
2174 }
2186 }
2190 }
2193 {
2208 else
2215 }
2216 /* Only remove non-faulty devices if recovery
2217 * is not possible.
2218 */
2226 }
2229 /* We must have just cleared 'rdev' */
2233 }
2238 abort:
2242 }
2245 {
2250 else
2251 /* The write handler will notice the lack of
2252 * R10BIO_Uptodate and record any errors etc
2253 */
2257 /* for reconstruct, we always reschedule after a read.
2258 * for resync, only after all reads
2259 */
2263 /* we have read all the blocks,
2264 * do the comparison in process context in raid10d
2265 */
2267 }
2268 }
2271 {
2277 }
2280 {
2281 /* reshape read bio isn't allocated from r10buf_pool */
2285 }
2288 {
2293 /* the primary of several recovery bios */
2298 else
2307 else
2310 }
2311 }
2312 }
2315 {
2327 else
2339 }
2343 }
2348 }
2350 /*
2351 * Note: sync and recover and handled very differently for raid10
2352 * This code is for resync.
2353 * For resync, we read through virtual addresses and read all blocks.
2354 * If there is any error, we schedule a write. The lowest numbered
2355 * drive is authoritative.
2356 * However requests come for physical address, so we need to map.
2357 * For every physical address there are raid_disks/copies virtual addresses,
2358 * which is always are least one, but is not necessarly an integer.
2359 * This means that a physical address can span multiple chunks, so we may
2360 * have to submit multiple io requests for a single sync request.
2361 */
2362 /*
2363 * We check if all blocks are in-sync and only write to blocks that
2364 * aren't in sync
2365 */
2367 {
2376 /* find the first device with a block */
2391 /* now find blocks with errors */
2408 /* We know that the bi_io_vec layout is the same for
2409 * both 'first' and 'i', so we just compare them.
2410 * All vec entries are PAGE_SIZE;
2411 */
2419 len))
2422 }
2427 /* Don't fix anything. */
2430 /* Just give up on this device */
2433 }
2434 /* Ok, we need to write this bio, either to correct an
2435 * inconsistency or to correct an unreadable block.
2436 * First we need to fixup bv_offset, bv_len and
2437 * bi_vecs, as the read request might have corrupted these
2438 */
2459 }
2461 /* Now write out to any replacement devices
2462 * that are active
2463 */
2478 }
2480 done:
2484 }
2485 }
2487 /*
2488 * Now for the recovery code.
2489 * Recovery happens across physical sectors.
2490 * We recover all non-is_sync drives by finding the virtual address of
2491 * each, and then choose a working drive that also has that virt address.
2492 * There is a separate r10_bio for each non-in_sync drive.
2493 * Only the first two slots are in use. The first for reading,
2494 * The second for writing.
2495 *
2496 */
2498 {
2499 /* We got a read error during recovery.
2500 * We repeat the read in smaller page-sized sections.
2501 * If a read succeeds, write it to the new device or record
2502 * a bad block if we cannot.
2503 * If a read fails, record a bad block on both old and
2504 * new devices.
2505 */
2519 sector_t addr;
2528 addr,
2536 addr,
2546 }
2547 }
2549 /* We don't worry if we cannot set a bad block -
2550 * it really is bad so there is no loss in not
2551 * recording it yet
2552 */
2556 /* need bad block on destination too */
2561 /* just abort the recovery */
2570 }
2571 }
2572 }
2576 idx++;
2577 }
2578 }
2581 {
2587 /* Need to test wbio2->bi_end_io before we call
2588 * submit_bio_noacct as if the former is NULL,
2589 * the latter is free to free wbio2.
2590 */
2601 }
2603 /*
2604 * share the pages with the first bio
2605 * and submit the write request
2606 */
2612 }
2618 }
2619 }
2623 {
2628 /* success */
2635 }
2636 /* need to record an error - either for the block or the device */
2640 }
2642 /*
2643 * This is a kernel thread which:
2644 *
2645 * 1. Retries failed read operations on working mirrors.
2646 * 2. Updates the raid superblock when problems encounter.
2647 * 3. Performs writes following reads for array synchronising.
2648 */
2651 {
2657 /* still own a reference to this rdev, so it cannot
2658 * have been cleared recently.
2659 */
2663 /* drive has already been failed, just ignore any
2664 more fix_read_error() attempts */
2670 }
2693 sect,
2700 }
2701 sl++;
2707 /* Cannot read from anywhere, just mark the block
2708 * as bad on the first device to discourage future
2709 * reads.
2710 */
2715 rdev,
2722 }
2724 }
2727 /* write it back and re-read */
2731 sl--;
2742 sect,
2745 /* Well, this device is dead */
2749 sect +
2751 rdev)),
2756 }
2758 }
2763 sl--;
2774 sect,
2777 /* Well, this device is dead */
2781 sect +
2792 sect +
2796 }
2799 }
2803 }
2804 }
2807 {
2812 /* bio has the data to be written to slot 'i' where
2813 * we just recently had a write error.
2814 * We repeatedly clone the bio and trim down to one block,
2815 * then try the write. Where the write fails we record
2816 * a bad block.
2817 * It is conceivable that the bio doesn't exactly align with
2818 * blocks. We must handle this.
2819 *
2820 * We currently own a reference to the rdev.
2821 */
2824 sector_t sector;
2841 sector_t wsector;
2844 /* Write at 'sector' for 'sectors' */
2854 /* Failure! */
2863 }
2865 }
2868 {
2874 /* we got a read error. Maybe the drive is bad. Maybe just
2875 * the block and we can fix it.
2876 * We freeze all other IO, and try reading the block from
2877 * other devices. When we find one, we re-write
2878 * and check it that fixes the read error.
2879 * This is all done synchronously while the array is
2880 * frozen.
2881 */
2898 /*
2899 * allow_barrier after re-submit to ensure no sync io
2900 * can be issued while regular io pending.
2901 */
2903 }
2906 {
2907 /* Some sort of write request has finished and it
2908 * succeeded in writing where we thought there was a
2909 * bad block. So forget the bad block.
2910 * Or possibly if failed and we need to record
2911 * a bad block.
2912 */
2926 rdev,
2931 rdev,
2935 }
2943 rdev,
2948 rdev,
2952 }
2953 }
2963 rdev,
2973 }
2975 }
2980 rdev,
2984 }
2985 }
2991 /*
2992 * In case freeze_array() is waiting for condition
2993 * nr_pending == nr_queued + extra to be true.
2994 */
3002 }
3003 }
3004 }
3007 {
3025 }
3026 }
3030 retry_list);
3039 }
3040 }
3051 }
3070 else
3076 }
3078 }
3081 {
3096 }
3099 {
3109 else
3122 }
3123 }
3125 }
3127 /*
3128 * Set cluster_sync_high since we need other nodes to add the
3129 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3130 */
3132 {
3133 sector_t window_size;
3136 /*
3137 * First, here we define "stripe" as a unit which across
3138 * all member devices one time, so we get chunks by use
3139 * raid_disks / near_copies. Otherwise, if near_copies is
3140 * close to raid_disks, then resync window could increases
3141 * linearly with the increase of raid_disks, which means
3142 * we will suspend a really large IO window while it is not
3143 * necessary. If raid_disks is not divisible by near_copies,
3144 * an extra chunk is needed to ensure the whole "stripe" is
3145 * covered.
3146 */
3151 else
3155 /*
3156 * At least use a 32M window to align with raid1's resync window
3157 */
3162 }
3164 /*
3165 * perform a "sync" on one "block"
3166 *
3167 * We need to make sure that no normal I/O request - particularly write
3168 * requests - conflict with active sync requests.
3169 *
3170 * This is achieved by tracking pending requests and a 'barrier' concept
3171 * that can be installed to exclude normal IO requests.
3172 *
3173 * Resync and recovery are handled very differently.
3174 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3175 *
3176 * For resync, we iterate over virtual addresses, read all copies,
3177 * and update if there are differences. If only one copy is live,
3178 * skip it.
3179 * For recovery, we iterate over physical addresses, read a good
3180 * value for each non-in_sync drive, and over-write.
3181 *
3182 * So, for recovery we may have several outstanding complex requests for a
3183 * given address, one for each out-of-sync device. We model this by allocating
3184 * a number of r10_bio structures, one for each out-of-sync device.
3185 * As we setup these structures, we collect all bio's together into a list
3186 * which we then process collectively to add pages, and then process again
3187 * to pass to submit_bio_noacct.
3188 *
3189 * The r10_bio structures are linked using a borrowed master_bio pointer.
3190 * This link is counted in ->remaining. When the r10_bio that points to NULL
3191 * has its remaining count decremented to 0, the whole complex operation
3192 * is complete.
3193 *
3194 */
3198 {
3202 sector_t nr_sectors;
3205 sector_t sync_blocks;
3212 /*
3213 * Allow skipping a full rebuild for incremental assembly
3214 * of a clean array, like RAID1 does.
3215 */
3225 }
3231 skipped:
3236 /* If we aborted, we need to abort the
3237 * sync on the 'current' bitmap chucks (there can
3238 * be several when recovering multiple devices).
3239 * as we may have started syncing it but not finished.
3240 * We can find the current address in
3241 * mddev->curr_resync, but for recovery,
3242 * we need to convert that to several
3243 * virtual addresses.
3244 */
3249 }
3257 sector_t sect =
3262 }
3264 /* completed sync */
3268 /* Completed a full sync so the replacements
3269 * are now fully recovered.
3270 */
3277 }
3278 }
3280 }
3285 }
3295 /*
3296 * recovery fails, set mirrors.recovery_disabled,
3297 * device shouldn't be added to there.
3298 */
3302 }
3303 /*
3304 * if there has been nothing to do on any drive,
3305 * then there is nothing to do at all.
3306 */
3309 }
3314 /* make sure whole request will fit in a chunk - if chunks
3315 * are meaningful
3316 */
3321 /*
3322 * If there is non-resync activity waiting for a turn, then let it
3323 * though before starting on this new sync request.
3324 */
3328 /* Again, very different code for resync and recovery.
3329 * Both must result in an r10bio with a list of bios that
3330 * have bi_end_io, bi_sector, bi_bdev set,
3331 * and bi_private set to the r10bio.
3332 * For recovery, we may actually create several r10bios
3333 * with 2 bios in each, that correspond to the bios in the main one.
3334 * In this case, the subordinate r10bios link back through a
3335 * borrowed master_bio pointer, and the counter in the master
3336 * includes a ref from each subordinate.
3337 */
3338 /* First, we decide what to do and set ->bi_end_io
3339 * To end_sync_read if we want to read, and
3340 * end_sync_write if we will want to write.
3341 */
3345 /* recovery... the complicated one */
3352 sector_t sect;
3371 /* want to reconstruct this device */
3375 /* last stripe is not complete - don't
3376 * try to recover this sector.
3377 */
3379 /* Unless we are doing a full sync, or a replacement
3380 * we only need to recover the block if it is set in
3381 * the bitmap
3382 */
3391 /* yep, skip the sync_blocks here, but don't assume
3392 * that there will never be anything to do here
3393 */
3396 }
3416 /* Need to check if the array will still be
3417 * degraded
3418 */
3425 }
3426 }
3442 /* This is where we read from */
3451 bad_sectors -= (sector
3456 }
3457 }
3470 /* and we write to 'i' (if not in_sync) */
3495 /* and maybe write to replacement */
3499 /* Note: if replace is not NULL, then bio
3500 * cannot be NULL as r10buf_pool_alloc will
3501 * have allocated it.
3502 */
3514 }
3516 /* Cannot recover, so abort the recovery or
3517 * record a bad block */
3519 /* problem is that there are bad blocks
3520 * on other device(s)
3521 */
3529 mrdev,
3535 mreplace,
3539 }
3545 mirror->recovery_disabled
3549 }
3559 }
3565 /* Only want this if there is elsewhere to
3566 * read from. 'j' is currently the first
3567 * readable copy.
3568 */
3575 targets++;
3576 }
3580 }
3581 }
3588 }
3590 }
3592 /* resync. Schedule a read for every block at this virt offset */
3595 /*
3596 * Since curr_resync_completed could probably not update in
3597 * time, and we will set cluster_sync_low based on it.
3598 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3599 * safety reason, which ensures curr_resync_completed is
3600 * updated in bitmap_cond_end_sync.
3601 */
3611 /* We can skip this block */
3614 }
3656 }
3657 }
3668 count++;
3676 /* Need to set up for writing to the replacement */
3689 count++;
3690 }
3697 mddev);
3702 mddev);
3703 }
3707 }
3708 }
3727 }
3728 }
3736 /* It is resync not recovery */
3740 /* Send resync message */
3744 }
3746 /* This is recovery not resync */
3751 /*
3752 * sector_nr is a device address for recovery, so we
3753 * need translate it to array address before compare
3754 * with cluster_sync_high.
3755 */
3760 /*
3761 * curr_resync_completed is similar as
3762 * sector_nr, so make the translation too.
3763 */
3770 }
3771 }
3777 }
3778 }
3792 }
3793 }
3796 /* pretend they weren't skipped, it makes
3797 * no important difference in this case
3798 */
3802 giveup:
3803 /* There is nowhere to write, so all non-sync
3804 * drives must be failed or in resync, all drives
3805 * have a bad block, so try the next chunk...
3806 */
3811 chunks_skipped ++;
3814 }
3816 static sector_t
3818 {
3819 sector_t size;
3834 }
3837 {
3838 /* Calculate the number of sectors-per-device that will
3839 * actually be used, and set conf->dev_sectors and
3840 * conf->stride
3841 */
3847 /* 'size' is now the number of chunks in the array */
3848 /* calculate "used chunks per device" */
3851 /* We need to round up when dividing by raid_disks to
3852 * get the stride size.
3853 */
3863 }
3864 }
3868 {
3884 * updated yet. */
3889 }
3907 * actually using this, but leave code here just in case.*/
3917 }
3921 }
3924 {
3935 }
3938 {
3950 }
3956 }
3963 /* FIXME calc properly */
3966 GFP_KERNEL);
3993 }
3997 else
3998 /* far_copies must be 1 */
4000 }
4019 out:
4022 }
4025 {
4031 }
4034 {
4048 }
4050 }
4053 {
4058 sector_t size;
4068 }
4085 }
4086 }
4107 }
4118 }
4126 }
4127 }
4129 /* need to check that every block has at least one working mirror */
4134 }
4137 /* must ensure that shape change is supported */
4144 }
4150 i++) {
4155 /* The replacement is all we have - use it */
4159 }
4168 }
4174 }
4177 }
4185 /*
4186 * Ok, everything is just fine now
4187 */
4206 /* This cannot work */
4209 }
4216 }
4220 out_free_conf:
4224 out:
4226 }
4229 {
4231 }
4234 {
4239 else
4241 }
4244 {
4245 /* Resize of 'far' arrays is not supported.
4246 * For 'near' and 'offset' arrays we can set the
4247 * number of sectors used to be an appropriate multiple
4248 * of the chunk size.
4249 * For 'offset', this is far_copies*chunksize.
4250 * For 'near' the multiplier is the LCM of
4251 * near_copies and raid_disks.
4252 * So if far_copies > 1 && !far_offset, fail.
4253 * Else find LCM(raid_disks, near_copy)*far_copies and
4254 * multiply by chunk_size. Then round to this number.
4255 * This is mostly done by raid10_size()
4256 */
4282 }
4287 }
4290 {
4298 }
4301 /* Set new parameters */
4303 /* new layout: far_copies = 1, near_copies = 2 */
4308 /* make sure it will be not marked as dirty */
4318 }
4319 }
4322 }
4325 {
4328 /* raid10 can take over:
4329 * raid0 - providing it has only two drives
4330 */
4332 /* for raid0 takeover only one zone is supported */
4338 }
4342 }
4344 }
4347 {
4348 /* Called when there is a request to change
4349 * - layout (to ->new_layout)
4350 * - chunk size (to ->new_chunk_sectors)
4351 * - raid_disks (by delta_disks)
4352 * or when trying to restart a reshape that was ongoing.
4353 *
4354 * We need to validate the request and possibly allocate
4355 * space if that might be an issue later.
4356 *
4357 * Currently we reject any reshape of a 'far' mode array,
4358 * allow chunk size to change if new is generally acceptable,
4359 * allow raid_disks to increase, and allow
4360 * a switch between 'near' mode and 'offset' mode.
4361 */
4369 /* mustn't change number of copies */
4372 /* Cannot switch to 'far' mode */
4376 /* not factor of array size */
4385 /* allocate new 'mirrors' list */
4389 GFP_KERNEL);
4392 }
4394 }
4396 /*
4397 * Need to check if array has failed when deciding whether to:
4398 * - start an array
4399 * - remove non-faulty devices
4400 * - add a spare
4401 * - allow a reshape
4402 * This determination is simple when no reshape is happening.
4403 * However if there is a reshape, we need to carefully check
4404 * both the before and after sections.
4405 * This is because some failed devices may only affect one
4406 * of the two sections, and some non-in_sync devices may
4407 * be insync in the section most affected by failed devices.
4408 */
4410 {
4415 /* 'prev' section first */
4420 degraded++;
4422 /* When we can reduce the number of devices in
4423 * an array, this might not contribute to
4424 * 'degraded'. It does now.
4425 */
4426 degraded++;
4427 }
4435 degraded2++;
4437 /* If reshape is increasing the number of devices,
4438 * this section has already been recovered, so
4439 * it doesn't contribute to degraded.
4440 * else it does.
4441 */
4443 degraded2++;
4444 }
4445 }
4449 }
4452 {
4453 /* A 'reshape' has been requested. This commits
4454 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4455 * This also checks if there are enough spares and adds them
4456 * to the array.
4457 * We currently require enough spares to make the final
4458 * array non-degraded. We also require that the difference
4459 * between old and new data_offset - on each device - is
4460 * enough that we never risk over-writing.
4461 */
4486 spares++;
4497 }
4498 }
4516 }
4526 }
4545 else
4547 }
4553 }
4555 /*
4556 * some node is already performing reshape, and no need to
4557 * call bitmap_ops->resize again since it should be called when
4558 * receiving BITMAP_RESIZE msg
4559 */
4572 }
4573 }
4574 out:
4583 else
4586 /* Failure here is OK */
4588 }
4591 /* This is a spare that was manually added */
4593 }
4594 }
4595 /* When a reshape changes the number of devices,
4596 * ->degraded is measured against the larger of the
4597 * pre and post numbers.
4598 */
4615 abort:
4628 }
4630 /* Calculate the last device-address that could contain
4631 * any block from the chunk that includes the array-address 's'
4632 * and report the next address.
4633 * i.e. the address returned will be chunk-aligned and after
4634 * any data that is in the chunk containing 's'.
4635 */
4637 {
4645 }
4647 /* Calculate the first device-address that could contain
4648 * any block from the chunk that includes the array-address 's'.
4649 * This too will be the start of a chunk
4650 */
4652 {
4659 }
4663 {
4664 /* We simply copy at most one chunk (smallest of old and new)
4665 * at a time, possibly less if that exceeds RESYNC_PAGES,
4666 * or we hit a bad block or something.
4667 * This might mean we pause for normal IO in the middle of
4668 * a chunk, but that is not a problem as mddev->reshape_position
4669 * can record any location.
4670 *
4671 * If we will want to write to a location that isn't
4672 * yet recorded as 'safe' (i.e. in metadata on disk) then
4673 * we need to flush all reshape requests and update the metadata.
4674 *
4675 * When reshaping forwards (e.g. to more devices), we interpret
4676 * 'safe' as the earliest block which might not have been copied
4677 * down yet. We divide this by previous stripe size and multiply
4678 * by previous stripe length to get lowest device offset that we
4679 * cannot write to yet.
4680 * We interpret 'sector_nr' as an address that we want to write to.
4681 * From this we use last_device_address() to find where we might
4682 * write to, and first_device_address on the 'safe' position.
4683 * If this 'next' write position is after the 'safe' position,
4684 * we must update the metadata to increase the 'safe' position.
4685 *
4686 * When reshaping backwards, we round in the opposite direction
4687 * and perform the reverse test: next write position must not be
4688 * less than current safe position.
4689 *
4690 * In all this the minimum difference in data offsets
4691 * (conf->offset_diff - always positive) allows a bit of slack,
4692 * so next can be after 'safe', but not by more than offset_diff
4693 *
4694 * We need to prepare all the bios here before we start any IO
4695 * to ensure the size we choose is acceptable to all devices.
4696 * The means one for each copy for write-out and an extra one for
4697 * read-in.
4698 * We store the read-in bio in ->master_bio and the others in
4699 * ->devs[x].bio and ->devs[x].repl_bio.
4700 */
4715 /* If restarting in the middle, skip the initial sectors */
4728 }
4729 }
4731 /* We don't use sector_nr to track where we are up to
4732 * as that doesn't work well for ->reshape_backwards.
4733 * So just use ->reshape_progress.
4734 */
4736 /* 'next' is the earliest device address that we might
4737 * write to for this chunk in the new layout
4738 */
4742 /* 'safe' is the last device address that we might read from
4743 * in the old layout after a restart
4744 */
4757 /* 'next' is after the last device address that we
4758 * might write to for this chunk in the new layout
4759 */
4762 /* 'safe' is the earliest device address that we might
4763 * read from in the old layout after a restart
4764 */
4767 /* Need to update metadata if 'next' might be beyond 'safe'
4768 * as that would possibly corrupt data
4769 */
4779 }
4783 /* Need to update reshape_position in metadata */
4789 else
4799 }
4802 }
4805 read_more:
4806 /* Now schedule reads for blocks from sector_nr to last */
4819 /* Cannot read from here, so need to record bad blocks
4820 * on all the target devices.
4821 */
4822 // FIXME
4826 }
4837 /*
4838 * Broadcast RESYNC message to other nodes, so all nodes would not
4839 * write to the region to avoid conflict.
4840 */
4850 /*
4851 * Set cluster_sync_low again if next address for array
4852 * reshape is less than cluster_sync_low. Since we can't
4853 * update cluster_sync_low until it has finished reshape.
4854 */
4857 }
4861 }
4863 /* Now find the locations in the new layout */
4879 }
4890 }
4892 /* Now add as many pages as possible to all of these bios. */
4906 }
4907 }
4910 }
4913 /* Now submit the read */
4924 /* Now that we have done the whole section we can
4925 * update reshape_progress
4926 */
4929 else
4933 }
4939 {
4940 /* Reshape read completed. Hopefully we have a block
4941 * to write out.
4942 * If we got a read error then we do sync 1-page reads from
4943 * elsewhere until we find the data - or give up.
4944 */
4950 /* Reshape has been aborted */
4953 }
4955 /* We definitely have the data in the pages, schedule the
4956 * writes.
4957 */
4969 }
4978 }
4980 }
4983 {
4997 }
5000 {
5008 else
5010 }
5014 {
5015 /* Use sync reads to get the blocks from somewhere else */
5027 }
5029 /* reshape IOs share pages from .devs[0].bio */
5046 sector_t addr;
5055 addr,
5062 failed:
5063 slot++;
5068 }
5070 /* couldn't read this block, must give up */
5075 }
5077 idx++;
5078 }
5081 }
5084 {
5098 /* FIXME should record badblock */
5100 }
5104 }
5107 {
5113 }
5116 {
5126 }
5132 d++) {
5139 }
5140 }
5146 }
5149 {
5170 };
5173 {
5175 }
5178 {
5180 }