| blob | f3bf1116794a4e069377fc6d5965101c98fd3ed8 |
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 {
1164 /*
1165 * This is an error retry, but we cannot
1166 * safely dereference the rdev in the r10_bio,
1167 * we must use the one in conf.
1168 * If it has already been disconnected (unlikely)
1169 * we lose the device name in error messages.
1170 */
1172 /*
1173 * As we are blocking raid10, it is a little safer to
1174 * use __GFP_HIGH.
1175 */
1184 /* This never gets dereferenced */
1186 }
1187 }
1197 }
1200 }
1216 }
1222 }
1239 }
1244 {
1260 else
1273 /* flush_pending_writes() needs access to the rdev so...*/
1283 }
1284 }
1287 {
1292 retry_wait:
1303 }
1308 }
1313 /*
1314 * Discard request doesn't care the write result
1315 * so it doesn't need to wait blocked disk here.
1316 */
1322 /*
1323 * Mustn't write here until the bad block
1324 * is acknowledged
1325 */
1330 }
1331 }
1332 }
1335 /* Have to wait for this device to get unblocked, then retry */
1343 }
1344 }
1348 {
1351 sector_t sectors;
1359 /* Bail out if REQ_NOWAIT is set for the bio */
1363 }
1371 }
1373 }
1384 /* Need to update reshape_position in metadata */
1393 }
1400 }
1402 /* first select target devices under rcu_lock and
1403 * inc refcount on their rdev. Record them by setting
1404 * bios[x] to bio
1405 * If there are known/acknowledged bad blocks on any device
1406 * on which we have seen a write error, we want to avoid
1407 * writing to those blocks. This potentially requires several
1408 * writes to write around the bad blocks. Each set of writes
1409 * gets its own r10_bio with a set of bios attached.
1410 */
1436 }
1438 sector_t first_bad;
1446 /* Cannot write here at all */
1449 /* Mustn't write more than bad_sectors
1450 * to other devices yet
1451 */
1453 /* We don't set R10BIO_Degraded as that
1454 * only applies if the disk is missing,
1455 * so it might be re-added, and we want to
1456 * know to recover this chunk.
1457 * In this case the device is here, and the
1458 * fact that this chunk is not in-sync is
1459 * recorded in the bad block log.
1460 */
1462 }
1467 }
1468 }
1472 }
1476 }
1477 }
1491 }
1504 }
1506 }
1509 {
1527 else
1529 }
1532 {
1549 }
1550 }
1551 }
1554 {
1561 /*
1562 * We don't care the return value of discard bio
1563 */
1573 }
1575 /*
1576 * There are some limitations to handle discard bio
1577 * 1st, the discard size is bigger than stripe_size*2.
1578 * 2st, if the discard bio spans reshape progress, we use the old way to
1579 * handle discard bio
1580 */
1582 {
1590 sector_t chunk;
1593 sector_t split_size;
1596 sector_t start_disk_offset;
1598 sector_t end_disk_offset;
1608 }
1611 /*
1612 * Check reshape again to avoid reshape happens after checking
1613 * MD_RECOVERY_RESHAPE and before wait_barrier
1614 */
1621 else
1629 /*
1630 * Maybe one discard bio is smaller than strip size or across one
1631 * stripe and discard region is larger than one stripe size. For far
1632 * offset layout, if the discard region is not aligned with stripe
1633 * size, there is hole when we submit discard bio to member disk.
1634 * For simplicity, we only handle discard bio which discard region
1635 * is bigger than stripe_size * 2
1636 */
1640 /*
1641 * Keep bio aligned with strip size.
1642 */
1649 /* Resend the fist split part */
1652 }
1659 /* Resend the second split part */
1663 }
1668 /*
1669 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1670 * One stripe contains the chunks from all member disk (one chunk from
1671 * one disk at the same HBA address). For layout detail, see 'man md 4'
1672 */
1691 retry_discard:
1699 /*
1700 * For far layout it needs more than one r10bio to cover all regions.
1701 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1702 * to record the discard bio. Other r10bio->master_bio record the first
1703 * r10bio. The first r10bio only release after all other r10bios finish.
1704 * The discard bio returns only first r10bio finishes
1705 */
1714 /*
1715 * first select target devices under rcu_lock and
1716 * inc refcount on their rdev. Record them by setting
1717 * bios[x] to bio
1718 */
1737 }
1741 }
1742 }
1749 /*
1750 * Now start to calculate the start and end address for each disk.
1751 * The space between dev_start and dev_end is the discard region.
1752 *
1753 * For dev_start, it needs to consider three conditions:
1754 * 1st, the disk is before start_disk, you can imagine the disk in
1755 * the next stripe. So the dev_start is the start address of next
1756 * stripe.
1757 * 2st, the disk is after start_disk, it means the disk is at the
1758 * same stripe of first disk
1759 * 3st, the first disk itself, we can use start_disk_offset directly
1760 */
1765 else
1772 else
1775 /*
1776 * It only handles discard bio which size is >= stripe size, so
1777 * dev_end > dev_start all the time.
1778 * It doesn't need to use rcu lock to get rdev here. We already
1779 * add rdev->nr_pending in the first loop.
1780 */
1794 }
1808 }
1809 }
1820 }
1825 out:
1828 }
1831 {
1847 /*
1848 * If this request crosses a chunk boundary, we need to split
1849 * it.
1850 */
1852 sectors > chunk_sects
1861 /* In case raid10d snuck in to freeze_array */
1864 }
1867 {
1880 else
1884 }
1891 }
1893 }
1895 /* check if there are enough drives for
1896 * every block to appear on atleast one.
1897 * Don't consider the device numbered 'ignore'
1898 * as we might be about to remove it.
1899 */
1901 {
1911 }
1922 cnt++;
1924 }
1930 out:
1932 }
1935 {
1936 /* when calling 'enough', both 'prev' and 'geo' must
1937 * be stable.
1938 * This is ensured if ->reconfig_mutex or ->device_lock
1939 * is held.
1940 */
1943 }
1945 /**
1946 * raid10_error() - RAID10 error handler.
1947 * @mddev: affected md device.
1948 * @rdev: member device to fail.
1949 *
1950 * The routine acknowledges &rdev failure and determines new @mddev state.
1951 * If it failed, then:
1952 * - &MD_BROKEN flag is set in &mddev->flags.
1953 * Otherwise, it must be degraded:
1954 * - recovery is interrupted.
1955 * - &mddev->degraded is bumped.
1956 *
1957 * @rdev is marked as &Faulty excluding case when array is failed and
1958 * &mddev->fail_last_dev is off.
1959 */
1961 {
1973 }
1974 }
1988 }
1991 {
1999 }
2011 }
2012 }
2015 {
2020 }
2023 {
2030 /*
2031 * Find all non-in_sync disks within the RAID10 configuration
2032 * and mark them in_sync
2033 */
2040 /* Replacement has just become active */
2043 count++;
2045 /* Replaced device not technically faulty,
2046 * but we need to be sure it gets removed
2047 * and never re-added.
2048 */
2052 }
2058 count++;
2060 }
2061 }
2068 }
2071 {
2080 /* only hot-add to in-sync arrays, as recovery is
2081 * very different from resync
2082 */
2094 else
2105 }
2118 }
2130 }
2134 }
2137 {
2152 else
2159 }
2160 /* Only remove non-faulty devices if recovery
2161 * is not possible.
2162 */
2170 }
2173 /* We must have just cleared 'rdev' */
2177 }
2182 abort:
2186 }
2189 {
2194 else
2195 /* The write handler will notice the lack of
2196 * R10BIO_Uptodate and record any errors etc
2197 */
2201 /* for reconstruct, we always reschedule after a read.
2202 * for resync, only after all reads
2203 */
2207 /* we have read all the blocks,
2208 * do the comparison in process context in raid10d
2209 */
2211 }
2212 }
2215 {
2221 }
2224 {
2225 /* reshape read bio isn't allocated from r10buf_pool */
2229 }
2232 {
2237 /* the primary of several recovery bios */
2242 else
2251 else
2254 }
2255 }
2256 }
2259 {
2271 else
2283 }
2287 }
2292 }
2294 /*
2295 * Note: sync and recover and handled very differently for raid10
2296 * This code is for resync.
2297 * For resync, we read through virtual addresses and read all blocks.
2298 * If there is any error, we schedule a write. The lowest numbered
2299 * drive is authoritative.
2300 * However requests come for physical address, so we need to map.
2301 * For every physical address there are raid_disks/copies virtual addresses,
2302 * which is always are least one, but is not necessarly an integer.
2303 * This means that a physical address can span multiple chunks, so we may
2304 * have to submit multiple io requests for a single sync request.
2305 */
2306 /*
2307 * We check if all blocks are in-sync and only write to blocks that
2308 * aren't in sync
2309 */
2311 {
2320 /* find the first device with a block */
2335 /* now find blocks with errors */
2352 /* We know that the bi_io_vec layout is the same for
2353 * both 'first' and 'i', so we just compare them.
2354 * All vec entries are PAGE_SIZE;
2355 */
2363 len))
2366 }
2371 /* Don't fix anything. */
2374 /* Just give up on this device */
2377 }
2378 /* Ok, we need to write this bio, either to correct an
2379 * inconsistency or to correct an unreadable block.
2380 * First we need to fixup bv_offset, bv_len and
2381 * bi_vecs, as the read request might have corrupted these
2382 */
2403 }
2405 /* Now write out to any replacement devices
2406 * that are active
2407 */
2422 }
2424 done:
2428 }
2429 }
2431 /*
2432 * Now for the recovery code.
2433 * Recovery happens across physical sectors.
2434 * We recover all non-is_sync drives by finding the virtual address of
2435 * each, and then choose a working drive that also has that virt address.
2436 * There is a separate r10_bio for each non-in_sync drive.
2437 * Only the first two slots are in use. The first for reading,
2438 * The second for writing.
2439 *
2440 */
2442 {
2443 /* We got a read error during recovery.
2444 * We repeat the read in smaller page-sized sections.
2445 * If a read succeeds, write it to the new device or record
2446 * a bad block if we cannot.
2447 * If a read fails, record a bad block on both old and
2448 * new devices.
2449 */
2463 sector_t addr;
2472 addr,
2480 addr,
2490 }
2491 }
2493 /* We don't worry if we cannot set a bad block -
2494 * it really is bad so there is no loss in not
2495 * recording it yet
2496 */
2500 /* need bad block on destination too */
2505 /* just abort the recovery */
2514 }
2515 }
2516 }
2520 idx++;
2521 }
2522 }
2525 {
2531 /* Need to test wbio2->bi_end_io before we call
2532 * submit_bio_noacct as if the former is NULL,
2533 * the latter is free to free wbio2.
2534 */
2545 }
2547 /*
2548 * share the pages with the first bio
2549 * and submit the write request
2550 */
2556 }
2562 }
2563 }
2567 {
2572 /* success */
2579 }
2580 /* need to record an error - either for the block or the device */
2584 }
2586 /*
2587 * This is a kernel thread which:
2588 *
2589 * 1. Retries failed read operations on working mirrors.
2590 * 2. Updates the raid superblock when problems encounter.
2591 * 3. Performs writes following reads for array synchronising.
2592 */
2595 {
2601 /* still own a reference to this rdev, so it cannot
2602 * have been cleared recently.
2603 */
2607 /* drive has already been failed, just ignore any
2608 more fix_read_error() attempts */
2614 }
2637 sect,
2644 }
2645 sl++;
2651 /* Cannot read from anywhere, just mark the block
2652 * as bad on the first device to discourage future
2653 * reads.
2654 */
2659 rdev,
2666 }
2668 }
2671 /* write it back and re-read */
2675 sl--;
2686 sect,
2689 /* Well, this device is dead */
2693 sect +
2695 rdev)),
2700 }
2702 }
2707 sl--;
2718 sect,
2721 /* Well, this device is dead */
2725 sect +
2736 sect +
2740 }
2743 }
2747 }
2748 }
2751 {
2756 /* bio has the data to be written to slot 'i' where
2757 * we just recently had a write error.
2758 * We repeatedly clone the bio and trim down to one block,
2759 * then try the write. Where the write fails we record
2760 * a bad block.
2761 * It is conceivable that the bio doesn't exactly align with
2762 * blocks. We must handle this.
2763 *
2764 * We currently own a reference to the rdev.
2765 */
2768 sector_t sector;
2785 sector_t wsector;
2788 /* Write at 'sector' for 'sectors' */
2798 /* Failure! */
2807 }
2809 }
2812 {
2818 /* we got a read error. Maybe the drive is bad. Maybe just
2819 * the block and we can fix it.
2820 * We freeze all other IO, and try reading the block from
2821 * other devices. When we find one, we re-write
2822 * and check it that fixes the read error.
2823 * This is all done synchronously while the array is
2824 * frozen.
2825 */
2842 /*
2843 * allow_barrier after re-submit to ensure no sync io
2844 * can be issued while regular io pending.
2845 */
2847 }
2850 {
2851 /* Some sort of write request has finished and it
2852 * succeeded in writing where we thought there was a
2853 * bad block. So forget the bad block.
2854 * Or possibly if failed and we need to record
2855 * a bad block.
2856 */
2870 rdev,
2875 rdev,
2879 }
2887 rdev,
2892 rdev,
2896 }
2897 }
2907 rdev,
2917 }
2919 }
2924 rdev,
2928 }
2929 }
2935 /*
2936 * In case freeze_array() is waiting for condition
2937 * nr_pending == nr_queued + extra to be true.
2938 */
2946 }
2947 }
2948 }
2951 {
2969 }
2970 }
2974 retry_list);
2983 }
2984 }
2995 }
3014 else
3020 }
3022 }
3025 {
3040 }
3043 {
3053 else
3066 }
3067 }
3069 }
3071 /*
3072 * Set cluster_sync_high since we need other nodes to add the
3073 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3074 */
3076 {
3077 sector_t window_size;
3080 /*
3081 * First, here we define "stripe" as a unit which across
3082 * all member devices one time, so we get chunks by use
3083 * raid_disks / near_copies. Otherwise, if near_copies is
3084 * close to raid_disks, then resync window could increases
3085 * linearly with the increase of raid_disks, which means
3086 * we will suspend a really large IO window while it is not
3087 * necessary. If raid_disks is not divisible by near_copies,
3088 * an extra chunk is needed to ensure the whole "stripe" is
3089 * covered.
3090 */
3095 else
3099 /*
3100 * At least use a 32M window to align with raid1's resync window
3101 */
3106 }
3108 /*
3109 * perform a "sync" on one "block"
3110 *
3111 * We need to make sure that no normal I/O request - particularly write
3112 * requests - conflict with active sync requests.
3113 *
3114 * This is achieved by tracking pending requests and a 'barrier' concept
3115 * that can be installed to exclude normal IO requests.
3116 *
3117 * Resync and recovery are handled very differently.
3118 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3119 *
3120 * For resync, we iterate over virtual addresses, read all copies,
3121 * and update if there are differences. If only one copy is live,
3122 * skip it.
3123 * For recovery, we iterate over physical addresses, read a good
3124 * value for each non-in_sync drive, and over-write.
3125 *
3126 * So, for recovery we may have several outstanding complex requests for a
3127 * given address, one for each out-of-sync device. We model this by allocating
3128 * a number of r10_bio structures, one for each out-of-sync device.
3129 * As we setup these structures, we collect all bio's together into a list
3130 * which we then process collectively to add pages, and then process again
3131 * to pass to submit_bio_noacct.
3132 *
3133 * The r10_bio structures are linked using a borrowed master_bio pointer.
3134 * This link is counted in ->remaining. When the r10_bio that points to NULL
3135 * has its remaining count decremented to 0, the whole complex operation
3136 * is complete.
3137 *
3138 */
3142 {
3146 sector_t nr_sectors;
3149 sector_t sync_blocks;
3156 /*
3157 * Allow skipping a full rebuild for incremental assembly
3158 * of a clean array, like RAID1 does.
3159 */
3169 }
3175 skipped:
3180 /* If we aborted, we need to abort the
3181 * sync on the 'current' bitmap chucks (there can
3182 * be several when recovering multiple devices).
3183 * as we may have started syncing it but not finished.
3184 * We can find the current address in
3185 * mddev->curr_resync, but for recovery,
3186 * we need to convert that to several
3187 * virtual addresses.
3188 */
3193 }
3201 sector_t sect =
3206 }
3208 /* completed sync */
3212 /* Completed a full sync so the replacements
3213 * are now fully recovered.
3214 */
3221 }
3222 }
3224 }
3229 }
3239 /*
3240 * recovery fails, set mirrors.recovery_disabled,
3241 * device shouldn't be added to there.
3242 */
3246 }
3247 /*
3248 * if there has been nothing to do on any drive,
3249 * then there is nothing to do at all.
3250 */
3253 }
3258 /* make sure whole request will fit in a chunk - if chunks
3259 * are meaningful
3260 */
3265 /*
3266 * If there is non-resync activity waiting for a turn, then let it
3267 * though before starting on this new sync request.
3268 */
3272 /* Again, very different code for resync and recovery.
3273 * Both must result in an r10bio with a list of bios that
3274 * have bi_end_io, bi_sector, bi_bdev set,
3275 * and bi_private set to the r10bio.
3276 * For recovery, we may actually create several r10bios
3277 * with 2 bios in each, that correspond to the bios in the main one.
3278 * In this case, the subordinate r10bios link back through a
3279 * borrowed master_bio pointer, and the counter in the master
3280 * includes a ref from each subordinate.
3281 */
3282 /* First, we decide what to do and set ->bi_end_io
3283 * To end_sync_read if we want to read, and
3284 * end_sync_write if we will want to write.
3285 */
3289 /* recovery... the complicated one */
3296 sector_t sect;
3315 /* want to reconstruct this device */
3319 /* last stripe is not complete - don't
3320 * try to recover this sector.
3321 */
3323 /* Unless we are doing a full sync, or a replacement
3324 * we only need to recover the block if it is set in
3325 * the bitmap
3326 */
3335 /* yep, skip the sync_blocks here, but don't assume
3336 * that there will never be anything to do here
3337 */
3340 }
3360 /* Need to check if the array will still be
3361 * degraded
3362 */
3369 }
3370 }
3386 /* This is where we read from */
3395 bad_sectors -= (sector
3400 }
3401 }
3414 /* and we write to 'i' (if not in_sync) */
3439 /* and maybe write to replacement */
3443 /* Note: if replace is not NULL, then bio
3444 * cannot be NULL as r10buf_pool_alloc will
3445 * have allocated it.
3446 */
3458 }
3460 /* Cannot recover, so abort the recovery or
3461 * record a bad block */
3463 /* problem is that there are bad blocks
3464 * on other device(s)
3465 */
3473 mrdev,
3479 mreplace,
3483 }
3489 mirror->recovery_disabled
3493 }
3503 }
3509 /* Only want this if there is elsewhere to
3510 * read from. 'j' is currently the first
3511 * readable copy.
3512 */
3519 targets++;
3520 }
3524 }
3525 }
3532 }
3534 }
3536 /* resync. Schedule a read for every block at this virt offset */
3539 /*
3540 * Since curr_resync_completed could probably not update in
3541 * time, and we will set cluster_sync_low based on it.
3542 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3543 * safety reason, which ensures curr_resync_completed is
3544 * updated in bitmap_cond_end_sync.
3545 */
3555 /* We can skip this block */
3558 }
3600 }
3601 }
3612 count++;
3620 /* Need to set up for writing to the replacement */
3633 count++;
3634 }
3641 mddev);
3646 mddev);
3647 }
3651 }
3652 }
3671 }
3672 }
3680 /* It is resync not recovery */
3684 /* Send resync message */
3688 }
3690 /* This is recovery not resync */
3695 /*
3696 * sector_nr is a device address for recovery, so we
3697 * need translate it to array address before compare
3698 * with cluster_sync_high.
3699 */
3704 /*
3705 * curr_resync_completed is similar as
3706 * sector_nr, so make the translation too.
3707 */
3714 }
3715 }
3721 }
3722 }
3736 }
3737 }
3740 /* pretend they weren't skipped, it makes
3741 * no important difference in this case
3742 */
3746 giveup:
3747 /* There is nowhere to write, so all non-sync
3748 * drives must be failed or in resync, all drives
3749 * have a bad block, so try the next chunk...
3750 */
3755 chunks_skipped ++;
3758 }
3760 static sector_t
3762 {
3763 sector_t size;
3778 }
3781 {
3782 /* Calculate the number of sectors-per-device that will
3783 * actually be used, and set conf->dev_sectors and
3784 * conf->stride
3785 */
3791 /* 'size' is now the number of chunks in the array */
3792 /* calculate "used chunks per device" */
3795 /* We need to round up when dividing by raid_disks to
3796 * get the stride size.
3797 */
3807 }
3808 }
3812 {
3828 * updated yet. */
3833 }
3851 * actually using this, but leave code here just in case.*/
3861 }
3865 }
3868 {
3879 }
3882 {
3894 }
3900 }
3907 /* FIXME calc properly */
3910 GFP_KERNEL);
3937 }
3941 else
3942 /* far_copies must be 1 */
3944 }
3963 out:
3966 }
3969 {
3975 }
3978 {
3991 }
3993 }
3996 {
4001 sector_t size;
4011 }
4028 }
4029 }
4050 }
4061 }
4067 }
4069 /* need to check that every block has at least one working mirror */
4074 }
4077 /* must ensure that shape change is supported */
4084 }
4090 i++) {
4095 /* The replacement is all we have - use it */
4099 }
4108 }
4114 }
4117 }
4125 /*
4126 * Ok, everything is just fine now
4127 */
4146 /* This cannot work */
4149 }
4156 }
4160 out_free_conf:
4164 out:
4166 }
4169 {
4171 }
4174 {
4179 else
4181 }
4184 {
4185 /* Resize of 'far' arrays is not supported.
4186 * For 'near' and 'offset' arrays we can set the
4187 * number of sectors used to be an appropriate multiple
4188 * of the chunk size.
4189 * For 'offset', this is far_copies*chunksize.
4190 * For 'near' the multiplier is the LCM of
4191 * near_copies and raid_disks.
4192 * So if far_copies > 1 && !far_offset, fail.
4193 * Else find LCM(raid_disks, near_copy)*far_copies and
4194 * multiply by chunk_size. Then round to this number.
4195 * This is mostly done by raid10_size()
4196 */
4222 }
4227 }
4230 {
4238 }
4241 /* Set new parameters */
4243 /* new layout: far_copies = 1, near_copies = 2 */
4248 /* make sure it will be not marked as dirty */
4258 }
4259 }
4262 }
4265 {
4268 /* raid10 can take over:
4269 * raid0 - providing it has only two drives
4270 */
4272 /* for raid0 takeover only one zone is supported */
4278 }
4282 }
4284 }
4287 {
4288 /* Called when there is a request to change
4289 * - layout (to ->new_layout)
4290 * - chunk size (to ->new_chunk_sectors)
4291 * - raid_disks (by delta_disks)
4292 * or when trying to restart a reshape that was ongoing.
4293 *
4294 * We need to validate the request and possibly allocate
4295 * space if that might be an issue later.
4296 *
4297 * Currently we reject any reshape of a 'far' mode array,
4298 * allow chunk size to change if new is generally acceptable,
4299 * allow raid_disks to increase, and allow
4300 * a switch between 'near' mode and 'offset' mode.
4301 */
4309 /* mustn't change number of copies */
4312 /* Cannot switch to 'far' mode */
4316 /* not factor of array size */
4325 /* allocate new 'mirrors' list */
4329 GFP_KERNEL);
4332 }
4334 }
4336 /*
4337 * Need to check if array has failed when deciding whether to:
4338 * - start an array
4339 * - remove non-faulty devices
4340 * - add a spare
4341 * - allow a reshape
4342 * This determination is simple when no reshape is happening.
4343 * However if there is a reshape, we need to carefully check
4344 * both the before and after sections.
4345 * This is because some failed devices may only affect one
4346 * of the two sections, and some non-in_sync devices may
4347 * be insync in the section most affected by failed devices.
4348 */
4350 {
4355 /* 'prev' section first */
4360 degraded++;
4362 /* When we can reduce the number of devices in
4363 * an array, this might not contribute to
4364 * 'degraded'. It does now.
4365 */
4366 degraded++;
4367 }
4375 degraded2++;
4377 /* If reshape is increasing the number of devices,
4378 * this section has already been recovered, so
4379 * it doesn't contribute to degraded.
4380 * else it does.
4381 */
4383 degraded2++;
4384 }
4385 }
4389 }
4392 {
4393 /* A 'reshape' has been requested. This commits
4394 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4395 * This also checks if there are enough spares and adds them
4396 * to the array.
4397 * We currently require enough spares to make the final
4398 * array non-degraded. We also require that the difference
4399 * between old and new data_offset - on each device - is
4400 * enough that we never risk over-writing.
4401 */
4426 spares++;
4437 }
4438 }
4456 }
4466 }
4485 else
4487 }
4493 }
4495 /*
4496 * some node is already performing reshape, and no need to
4497 * call bitmap_ops->resize again since it should be called when
4498 * receiving BITMAP_RESIZE msg
4499 */
4512 }
4513 }
4514 out:
4523 else
4526 /* Failure here is OK */
4528 }
4531 /* This is a spare that was manually added */
4533 }
4534 }
4535 /* When a reshape changes the number of devices,
4536 * ->degraded is measured against the larger of the
4537 * pre and post numbers.
4538 */
4555 abort:
4568 }
4570 /* Calculate the last device-address that could contain
4571 * any block from the chunk that includes the array-address 's'
4572 * and report the next address.
4573 * i.e. the address returned will be chunk-aligned and after
4574 * any data that is in the chunk containing 's'.
4575 */
4577 {
4585 }
4587 /* Calculate the first device-address that could contain
4588 * any block from the chunk that includes the array-address 's'.
4589 * This too will be the start of a chunk
4590 */
4592 {
4599 }
4603 {
4604 /* We simply copy at most one chunk (smallest of old and new)
4605 * at a time, possibly less if that exceeds RESYNC_PAGES,
4606 * or we hit a bad block or something.
4607 * This might mean we pause for normal IO in the middle of
4608 * a chunk, but that is not a problem as mddev->reshape_position
4609 * can record any location.
4610 *
4611 * If we will want to write to a location that isn't
4612 * yet recorded as 'safe' (i.e. in metadata on disk) then
4613 * we need to flush all reshape requests and update the metadata.
4614 *
4615 * When reshaping forwards (e.g. to more devices), we interpret
4616 * 'safe' as the earliest block which might not have been copied
4617 * down yet. We divide this by previous stripe size and multiply
4618 * by previous stripe length to get lowest device offset that we
4619 * cannot write to yet.
4620 * We interpret 'sector_nr' as an address that we want to write to.
4621 * From this we use last_device_address() to find where we might
4622 * write to, and first_device_address on the 'safe' position.
4623 * If this 'next' write position is after the 'safe' position,
4624 * we must update the metadata to increase the 'safe' position.
4625 *
4626 * When reshaping backwards, we round in the opposite direction
4627 * and perform the reverse test: next write position must not be
4628 * less than current safe position.
4629 *
4630 * In all this the minimum difference in data offsets
4631 * (conf->offset_diff - always positive) allows a bit of slack,
4632 * so next can be after 'safe', but not by more than offset_diff
4633 *
4634 * We need to prepare all the bios here before we start any IO
4635 * to ensure the size we choose is acceptable to all devices.
4636 * The means one for each copy for write-out and an extra one for
4637 * read-in.
4638 * We store the read-in bio in ->master_bio and the others in
4639 * ->devs[x].bio and ->devs[x].repl_bio.
4640 */
4655 /* If restarting in the middle, skip the initial sectors */
4668 }
4669 }
4671 /* We don't use sector_nr to track where we are up to
4672 * as that doesn't work well for ->reshape_backwards.
4673 * So just use ->reshape_progress.
4674 */
4676 /* 'next' is the earliest device address that we might
4677 * write to for this chunk in the new layout
4678 */
4682 /* 'safe' is the last device address that we might read from
4683 * in the old layout after a restart
4684 */
4697 /* 'next' is after the last device address that we
4698 * might write to for this chunk in the new layout
4699 */
4702 /* 'safe' is the earliest device address that we might
4703 * read from in the old layout after a restart
4704 */
4707 /* Need to update metadata if 'next' might be beyond 'safe'
4708 * as that would possibly corrupt data
4709 */
4719 }
4723 /* Need to update reshape_position in metadata */
4729 else
4739 }
4742 }
4745 read_more:
4746 /* Now schedule reads for blocks from sector_nr to last */
4759 /* Cannot read from here, so need to record bad blocks
4760 * on all the target devices.
4761 */
4762 // FIXME
4766 }
4777 /*
4778 * Broadcast RESYNC message to other nodes, so all nodes would not
4779 * write to the region to avoid conflict.
4780 */
4790 /*
4791 * Set cluster_sync_low again if next address for array
4792 * reshape is less than cluster_sync_low. Since we can't
4793 * update cluster_sync_low until it has finished reshape.
4794 */
4797 }
4801 }
4803 /* Now find the locations in the new layout */
4819 }
4830 }
4832 /* Now add as many pages as possible to all of these bios. */
4846 }
4847 }
4850 }
4853 /* Now submit the read */
4864 /* Now that we have done the whole section we can
4865 * update reshape_progress
4866 */
4869 else
4873 }
4879 {
4880 /* Reshape read completed. Hopefully we have a block
4881 * to write out.
4882 * If we got a read error then we do sync 1-page reads from
4883 * elsewhere until we find the data - or give up.
4884 */
4890 /* Reshape has been aborted */
4893 }
4895 /* We definitely have the data in the pages, schedule the
4896 * writes.
4897 */
4909 }
4918 }
4920 }
4923 {
4937 }
4940 {
4948 else
4950 }
4954 {
4955 /* Use sync reads to get the blocks from somewhere else */
4967 }
4969 /* reshape IOs share pages from .devs[0].bio */
4986 sector_t addr;
4995 addr,
5002 failed:
5003 slot++;
5008 }
5010 /* couldn't read this block, must give up */
5015 }
5017 idx++;
5018 }
5021 }
5024 {
5038 /* FIXME should record badblock */
5040 }
5044 }
5047 {
5053 }
5056 {
5066 }
5072 d++) {
5079 }
5080 }
5086 }
5089 {
5110 };
5113 {
5115 }
5118 {
5120 }