| blob | c5e6c60fc0d41b53a578874087ae87259e3ebcce |
1 /*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include <trace/events/block.h>
34 /*
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
37 * chunk_size
38 * raid_disks
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
44 *
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
53 *
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
57 *
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
65 * on a device):
66 * A B C D A B C D E
67 * ... ...
68 * D A B C E A B C D
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
73 */
75 /*
76 * Number of guaranteed r10bios in case of extreme VM load:
77 */
78 #define NR_RAID10_BIOS 256
80 /* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84 */
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
89 */
90 #define IO_MADE_GOOD ((struct bio *)2)
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
94 /* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
96 * for writeback.
97 */
110 #define raid10_log(md, fmt, args...) \
115 /*
116 * for resync bio, r10bio pointer can be retrieved from the per-bio
117 * 'struct resync_pages'.
118 */
120 {
122 }
125 {
129 /* allocate a r10bio with room for raid_disks entries in the
130 * bios array */
132 }
135 {
137 }
139 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
140 /* amount of memory to reserve for resync requests */
141 #define RESYNC_WINDOW (1024*1024)
142 /* maximum number of concurrent requests, memory permitting */
143 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
145 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
147 /*
148 * When performing a resync, we need to read and compare, so
149 * we need as many pages are there are copies.
150 * When performing a recovery, we need 2 bios, one for read,
151 * one for write (we recover only one drive per r10buf)
152 *
153 */
155 {
170 else
173 /* allocate once for all bios */
176 else
182 /*
183 * Allocate bios.
184 */
196 }
197 /*
198 * Allocate RESYNC_PAGES data pages and attach them
199 * where needed.
200 */
218 }
225 }
226 }
230 out_free_pages:
235 out_free_bio:
241 }
243 out_free_r10bio:
246 }
249 {
265 }
267 /* resync pages array stored in the 1st bio's .bi_private */
271 }
274 {
286 }
287 }
290 {
295 }
298 {
304 }
307 {
317 /* wake up frozen array... */
321 }
323 /*
324 * raid_end_bio_io() is called when we have finished servicing a mirrored
325 * operation and are ready to return a success/failure code to the buffer
326 * cache layer.
327 */
329 {
337 /*
338 * Wake up any possible resync thread that waits for the device
339 * to go idle.
340 */
344 }
346 /*
347 * Update disk head position estimator based on IRQ completion info.
348 */
350 {
355 }
357 /*
358 * Find the disk number which triggered given bio
359 */
362 {
372 }
373 }
383 }
386 {
395 /*
396 * this branch is our 'one mirror IO has finished' event handler:
397 */
401 /*
402 * Set R10BIO_Uptodate in our master bio, so that
403 * we will return a good error code to the higher
404 * levels even if IO on some other mirrored buffer fails.
405 *
406 * The 'master' represents the composite IO operation to
407 * user-side. So if something waits for IO, then it will
408 * wait for the 'master' bio.
409 */
412 /* If all other devices that store this block have
413 * failed, we want to return the error upwards rather
414 * than fail the last device. Here we redefine
415 * "uptodate" to mean "Don't want to retry"
416 */
420 }
425 /*
426 * oops, read error - keep the refcount on the rdev
427 */
435 }
436 }
439 {
440 /* clear the bitmap if all writes complete successfully */
446 }
449 {
457 else
459 }
460 }
461 }
464 {
484 }
485 /*
486 * this branch is our 'one mirror IO has finished' event handler:
487 */
490 /* Never record new bad blocks to replacement,
491 * just fail it.
492 */
505 /* This is the only remaining device,
506 * We need to retry the write without
507 * FailFast
508 */
514 }
517 }
519 /*
520 * Set R10BIO_Uptodate in our master bio, so that
521 * we will return a good error code for to the higher
522 * levels even if IO on some other mirrored buffer fails.
523 *
524 * The 'master' represents the composite IO operation to
525 * user-side. So if something waits for IO, then it will
526 * wait for the 'master' bio.
527 */
528 sector_t first_bad;
531 /*
532 * Do not set R10BIO_Uptodate if the current device is
533 * rebuilding or Faulty. This is because we cannot use
534 * such device for properly reading the data back (we could
535 * potentially use it, if the current write would have felt
536 * before rdev->recovery_offset, but for simplicity we don't
537 * check this here.
538 */
543 /* Maybe we can clear some bad blocks. */
551 else
555 }
556 }
558 /*
559 *
560 * Let's see if all mirrored write operations have finished
561 * already.
562 */
568 }
570 /*
571 * RAID10 layout manager
572 * As well as the chunksize and raid_disks count, there are two
573 * parameters: near_copies and far_copies.
574 * near_copies * far_copies must be <= raid_disks.
575 * Normally one of these will be 1.
576 * If both are 1, we get raid0.
577 * If near_copies == raid_disks, we get raid1.
578 *
579 * Chunks are laid out in raid0 style with near_copies copies of the
580 * first chunk, followed by near_copies copies of the next chunk and
581 * so on.
582 * If far_copies > 1, then after 1/far_copies of the array has been assigned
583 * as described above, we start again with a device offset of near_copies.
584 * So we effectively have another copy of the whole array further down all
585 * the drives, but with blocks on different drives.
586 * With this layout, and block is never stored twice on the one device.
587 *
588 * raid10_find_phys finds the sector offset of a given virtual sector
589 * on each device that it is on.
590 *
591 * raid10_find_virt does the reverse mapping, from a device and a
592 * sector offset to a virtual address
593 */
596 {
598 sector_t sector;
599 sector_t chunk;
600 sector_t stripe;
611 /* now calculate first sector/dev */
623 /* and calculate all the others */
630 slot++;
644 }
648 slot++;
649 }
650 dev++;
654 }
655 }
656 }
659 {
671 }
674 {
676 /* Never use conf->prev as this is only called during resync
677 * or recovery, so reshape isn't happening
678 */
692 }
693 }
708 else
710 }
712 }
716 }
718 /*
719 * This routine returns the disk from which the requested read should
720 * be done. There is a per-array 'next expected sequential IO' sector
721 * number - if this matches on the next IO then we use the last disk.
722 * There is also a per-disk 'last know head position' sector that is
723 * maintained from IRQ contexts, both the normal and the resync IO
724 * completion handlers update this position correctly. If there is no
725 * perfect sequential match then we pick the disk whose head is closest.
726 *
727 * If there are 2 mirrors in the same 2 devices, performance degrades
728 * because position is mirror, not device based.
729 *
730 * The rdev for the device selected will have nr_pending incremented.
731 */
733 /*
734 * FIXME: possibly should rethink readbalancing and do it differently
735 * depending on near_copies / far_copies geometry.
736 */
740 {
759 /*
760 * Check if we can balance. We can balance on the whole
761 * device if no resync is going on (recovery is ok), or below
762 * the resync window. We take the first readable disk when
763 * above the resync window.
764 */
773 sector_t first_bad;
775 sector_t dev_sector;
795 /* Already have a better slot */
798 /* Cannot read here. If this is the
799 * 'primary' device, then we must not read
800 * beyond 'bad_sectors' from another device.
801 */
808 sector_t good_sectors =
814 }
816 /* Must read from here */
818 }
827 /* At least 2 disks to choose from so failfast is OK */
829 /* This optimisation is debatable, and completely destroys
830 * sequential read speed for 'far copies' arrays. So only
831 * keep it for 'near' arrays, and review those later.
832 */
836 /* for far > 1 always use the lowest address */
839 else
846 }
847 }
851 }
862 }
865 {
877 i++) {
883 }
884 }
887 }
890 {
891 /* Any writes that have been queued but are awaiting
892 * bitmap updates get flushed here.
893 */
904 /*
905 * As this is called in a wait_event() loop (see freeze_array),
906 * current->state might be TASK_UNINTERRUPTIBLE which will
907 * cause a warning when we prepare to wait again. As it is
908 * rare that this path is taken, it is perfectly safe to force
909 * us to go around the wait_event() loop again, so the warning
910 * is a false-positive. Silence the warning by resetting
911 * thread state
912 */
916 /* flush any pending bitmap writes to disk
917 * before proceeding w/ I/O */
930 /* Just ignore it */
932 else
935 }
939 }
941 /* Barriers....
942 * Sometimes we need to suspend IO while we do something else,
943 * either some resync/recovery, or reconfigure the array.
944 * To do this we raise a 'barrier'.
945 * The 'barrier' is a counter that can be raised multiple times
946 * to count how many activities are happening which preclude
947 * normal IO.
948 * We can only raise the barrier if there is no pending IO.
949 * i.e. if nr_pending == 0.
950 * We choose only to raise the barrier if no-one is waiting for the
951 * barrier to go down. This means that as soon as an IO request
952 * is ready, no other operations which require a barrier will start
953 * until the IO request has had a chance.
954 *
955 * So: regular IO calls 'wait_barrier'. When that returns there
956 * is no backgroup IO happening, It must arrange to call
957 * allow_barrier when it has finished its IO.
958 * backgroup IO calls must call raise_barrier. Once that returns
959 * there is no normal IO happeing. It must arrange to call
960 * lower_barrier when the particular background IO completes.
961 */
964 {
968 /* Wait until no block IO is waiting (unless 'force') */
972 /* block any new IO from starting */
975 /* Now wait for all pending IO to complete */
981 }
984 {
990 }
993 {
997 /* Wait for the barrier to drop.
998 * However if there are already pending
999 * requests (preventing the barrier from
1000 * rising completely), and the
1001 * pre-process bio queue isn't empty,
1002 * then don't wait, as we need to empty
1003 * that queue to get the nr_pending
1004 * count down.
1005 */
1017 }
1020 }
1023 {
1027 }
1030 {
1031 /* stop syncio and normal IO and wait for everything to
1032 * go quiet.
1033 * We increment barrier and nr_waiting, and then
1034 * wait until nr_pending match nr_queued+extra
1035 * This is called in the context of one normal IO request
1036 * that has failed. Thus any sync request that might be pending
1037 * will be blocked by nr_pending, and we need to wait for
1038 * pending IO requests to complete or be queued for re-try.
1039 * Thus the number queued (nr_queued) plus this request (extra)
1040 * must match the number of pending IOs (nr_pending) before
1041 * we continue.
1042 */
1054 }
1057 {
1058 /* reverse the effect of the freeze */
1064 }
1068 {
1072 else
1074 }
1080 };
1083 {
1085 cb);
1099 }
1101 /* we aren't scheduling, so we can do the write-out directly. */
1115 /* Just ignore it */
1117 else
1120 }
1122 }
1126 {
1132 sector_t sectors;
1140 /*
1141 * This is an error retry, but we cannot
1142 * safely dereference the rdev in the r10_bio,
1143 * we must use the one in conf.
1144 * If it has already been disconnected (unlikely)
1145 * we lose the device name in error messages.
1146 */
1148 /*
1149 * As we are blocking raid10, it is a little safer to
1150 * use __GFP_HIGH.
1151 */
1161 /* This never gets dereferenced */
1163 }
1165 }
1166 /*
1167 * Register the new request and wait if the reconstruction
1168 * thread has put up a bar for new requests.
1169 * Continue immediately if no resync is active currently.
1170 */
1177 /*
1178 * IO spans the reshape position. Need to wait for reshape to
1179 * pass
1180 */
1186 sectors);
1188 }
1196 }
1199 }
1213 }
1237 }
1242 {
1257 /* Replacement just got moved to main 'rdev' */
1260 }
1267 else
1285 /* flush_pending_writes() needs access to the rdev so...*/
1293 else
1304 }
1305 }
1309 {
1313 sector_t sectors;
1328 }
1330 }
1332 /*
1333 * Register the new request and wait if the reconstruction
1334 * thread has put up a bar for new requests.
1335 * Continue immediately if no resync is active currently.
1336 */
1343 /*
1344 * IO spans the reshape position. Need to wait for reshape to
1345 * pass
1346 */
1352 sectors);
1354 }
1362 /* Need to update reshape_position in metadata */
1372 }
1379 }
1380 /* first select target devices under rcu_lock and
1381 * inc refcount on their rdev. Record them by setting
1382 * bios[x] to bio
1383 * If there are known/acknowledged bad blocks on any device
1384 * on which we have seen a write error, we want to avoid
1385 * writing to those blocks. This potentially requires several
1386 * writes to write around the bad blocks. Each set of writes
1387 * gets its own r10_bio with a set of bios attached.
1388 */
1392 retry_write:
1408 }
1413 }
1425 }
1427 sector_t first_bad;
1435 /* Mustn't write here until the bad block
1436 * is acknowledged
1437 */
1442 }
1444 /* Cannot write here at all */
1447 /* Mustn't write more than bad_sectors
1448 * to other devices yet
1449 */
1451 /* We don't set R10BIO_Degraded as that
1452 * only applies if the disk is missing,
1453 * so it might be re-added, and we want to
1454 * know to recover this chunk.
1455 * In this case the device is here, and the
1456 * fact that this chunk is not in-sync is
1457 * recorded in the bad block log.
1458 */
1460 }
1465 }
1466 }
1470 }
1474 }
1475 }
1479 /* Have to wait for this device to get unblocked, then retry */
1487 }
1493 /* Race with remove_disk */
1496 }
1498 }
1499 }
1505 }
1517 }
1527 }
1529 }
1532 {
1548 else
1550 }
1553 {
1562 }
1567 /*
1568 * If this request crosses a chunk boundary, we need to split
1569 * it.
1570 */
1572 sectors > chunk_sects
1581 /* In case raid10d snuck in to freeze_array */
1584 }
1587 {
1598 else
1602 }
1609 }
1612 }
1614 /* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1618 */
1620 {
1630 }
1642 cnt++;
1644 }
1650 out:
1653 }
1656 {
1657 /* when calling 'enough', both 'prev' and 'geo' must
1658 * be stable.
1659 * This is ensured if ->reconfig_mutex or ->device_lock
1660 * is held.
1661 */
1664 }
1667 {
1672 /*
1673 * If it is not operational, then we have already marked it as dead
1674 * else if it is the last working disks, ignore the error, let the
1675 * next level up know.
1676 * else mark the drive as failed
1677 */
1681 /*
1682 * Don't fail the drive, just return an IO error.
1683 */
1686 }
1689 /*
1690 * If recovery is running, make sure it aborts.
1691 */
1702 }
1705 {
1713 }
1717 /* This is only called with ->reconfix_mutex held, so
1718 * rcu protection of rdev is not needed */
1727 }
1728 }
1731 {
1737 }
1740 {
1747 /*
1748 * Find all non-in_sync disks within the RAID10 configuration
1749 * and mark them in_sync
1750 */
1757 /* Replacement has just become active */
1760 count++;
1762 /* Replaced device not technically faulty,
1763 * but we need to be sure it gets removed
1764 * and never re-added.
1765 */
1769 }
1775 count++;
1777 }
1778 }
1785 }
1788 {
1796 /* only hot-add to in-sync arrays, as recovery is
1797 * very different from resync
1798 */
1812 else
1832 }
1846 }
1852 }
1855 {
1867 else
1874 }
1875 /* Only remove non-faulty devices if recovery
1876 * is not possible.
1877 */
1885 }
1890 /* lost the race, try later */
1894 }
1895 }
1897 /* We must have just cleared 'rdev' */
1901 * but will never see neither -- if they are careful.
1902 */
1904 }
1909 abort:
1913 }
1916 {
1921 else
1922 /* The write handler will notice the lack of
1923 * R10BIO_Uptodate and record any errors etc
1924 */
1928 /* for reconstruct, we always reschedule after a read.
1929 * for resync, only after all reads
1930 */
1934 /* we have read all the blocks,
1935 * do the comparison in process context in raid10d
1936 */
1938 }
1939 }
1942 {
1948 }
1951 {
1952 /* reshape read bio isn't allocated from r10buf_pool */
1956 }
1959 {
1964 /* the primary of several recovery bios */
1969 else
1978 else
1981 }
1982 }
1983 }
1986 {
1991 sector_t first_bad;
2000 else
2012 }
2022 }
2024 /*
2025 * Note: sync and recover and handled very differently for raid10
2026 * This code is for resync.
2027 * For resync, we read through virtual addresses and read all blocks.
2028 * If there is any error, we schedule a write. The lowest numbered
2029 * drive is authoritative.
2030 * However requests come for physical address, so we need to map.
2031 * For every physical address there are raid_disks/copies virtual addresses,
2032 * which is always are least one, but is not necessarly an integer.
2033 * This means that a physical address can span multiple chunks, so we may
2034 * have to submit multiple io requests for a single sync request.
2035 */
2036 /*
2037 * We check if all blocks are in-sync and only write to blocks that
2038 * aren't in sync
2039 */
2041 {
2050 /* find the first device with a block */
2065 /* now find blocks with errors */
2082 /* We know that the bi_io_vec layout is the same for
2083 * both 'first' and 'i', so we just compare them.
2084 * All vec entries are PAGE_SIZE;
2085 */
2093 len))
2096 }
2101 /* Don't fix anything. */
2104 /* Just give up on this device */
2107 }
2108 /* Ok, we need to write this bio, either to correct an
2109 * inconsistency or to correct an unreadable block.
2110 * First we need to fixup bv_offset, bv_len and
2111 * bi_vecs, as the read request might have corrupted these
2112 */
2135 }
2137 /* Now write out to any replacement devices
2138 * that are active
2139 */
2154 }
2156 done:
2160 }
2161 }
2163 /*
2164 * Now for the recovery code.
2165 * Recovery happens across physical sectors.
2166 * We recover all non-is_sync drives by finding the virtual address of
2167 * each, and then choose a working drive that also has that virt address.
2168 * There is a separate r10_bio for each non-in_sync drive.
2169 * Only the first two slots are in use. The first for reading,
2170 * The second for writing.
2171 *
2172 */
2174 {
2175 /* We got a read error during recovery.
2176 * We repeat the read in smaller page-sized sections.
2177 * If a read succeeds, write it to the new device or record
2178 * a bad block if we cannot.
2179 * If a read fails, record a bad block on both old and
2180 * new devices.
2181 */
2195 sector_t addr;
2204 addr,
2212 addr,
2222 }
2223 }
2225 /* We don't worry if we cannot set a bad block -
2226 * it really is bad so there is no loss in not
2227 * recording it yet
2228 */
2232 /* need bad block on destination too */
2237 /* just abort the recovery */
2246 }
2247 }
2248 }
2252 idx++;
2253 }
2254 }
2257 {
2266 }
2268 /*
2269 * share the pages with the first bio
2270 * and submit the write request
2271 */
2275 /* Need to test wbio2->bi_end_io before we call
2276 * generic_make_request as if the former is NULL,
2277 * the latter is free to free wbio2.
2278 */
2285 }
2291 }
2292 }
2294 /*
2295 * Used by fix_read_error() to decay the per rdev read_errors.
2296 * We halve the read error count for every hour that has elapsed
2297 * since the last recorded read error.
2298 *
2299 */
2301 {
2309 /* first time we've seen a read error */
2312 }
2319 /*
2320 * if hours_since_last is > the number of bits in read_errors
2321 * just set read errors to 0. We do this to avoid
2322 * overflowing the shift of read_errors by hours_since_last.
2323 */
2326 else
2328 }
2332 {
2333 sector_t first_bad;
2340 /* success */
2347 }
2348 /* need to record an error - either for the block or the device */
2352 }
2354 /*
2355 * This is a kernel thread which:
2356 *
2357 * 1. Retries failed read operations on working mirrors.
2358 * 2. Updates the raid superblock when problems encounter.
2359 * 3. Performs writes following reads for array synchronising.
2360 */
2363 {
2370 /* still own a reference to this rdev, so it cannot
2371 * have been cleared recently.
2372 */
2376 /* drive has already been failed, just ignore any
2377 more fix_read_error() attempts */
2394 }
2407 sector_t first_bad;
2421 sect,
2429 }
2430 sl++;
2437 /* Cannot read from anywhere, just mark the block
2438 * as bad on the first device to discourage future
2439 * reads.
2440 */
2445 rdev,
2452 }
2454 }
2457 /* write it back and re-read */
2464 sl--;
2476 sect,
2479 /* Well, this device is dead */
2483 sect +
2485 rdev)),
2490 }
2493 }
2500 sl--;
2512 sect,
2514 READ)) {
2516 /* Well, this device is dead */
2520 sect +
2531 sect +
2535 }
2539 }
2544 }
2545 }
2548 {
2553 /* bio has the data to be written to slot 'i' where
2554 * we just recently had a write error.
2555 * We repeatedly clone the bio and trim down to one block,
2556 * then try the write. Where the write fails we record
2557 * a bad block.
2558 * It is conceivable that the bio doesn't exactly align with
2559 * blocks. We must handle this.
2560 *
2561 * We currently own a reference to the rdev.
2562 */
2565 sector_t sector;
2582 sector_t wsector;
2585 /* Write at 'sector' for 'sectors' */
2595 /* Failure! */
2604 }
2606 }
2609 {
2615 /* we got a read error. Maybe the drive is bad. Maybe just
2616 * the block and we can fix it.
2617 * We freeze all other IO, and try reading the block from
2618 * other devices. When we find one, we re-write
2619 * and check it that fixes the read error.
2620 * This is all done synchronously while the array is
2621 * frozen.
2622 */
2640 }
2643 {
2644 /* Some sort of write request has finished and it
2645 * succeeded in writing where we thought there was a
2646 * bad block. So forget the bad block.
2647 * Or possibly if failed and we need to record
2648 * a bad block.
2649 */
2663 rdev,
2668 rdev,
2672 }
2680 rdev,
2685 rdev,
2689 }
2690 }
2700 rdev,
2710 }
2712 }
2717 rdev,
2721 }
2722 }
2728 /*
2729 * In case freeze_array() is waiting for condition
2730 * nr_pending == nr_queued + extra to be true.
2731 */
2739 }
2740 }
2741 }
2744 {
2762 }
2763 }
2767 retry_list);
2776 }
2777 }
2788 }
2807 else
2813 }
2815 }
2818 {
2833 }
2836 {
2846 else
2859 }
2860 }
2862 }
2864 /*
2865 * Set cluster_sync_high since we need other nodes to add the
2866 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2867 */
2869 {
2870 sector_t window_size;
2873 /*
2874 * First, here we define "stripe" as a unit which across
2875 * all member devices one time, so we get chunks by use
2876 * raid_disks / near_copies. Otherwise, if near_copies is
2877 * close to raid_disks, then resync window could increases
2878 * linearly with the increase of raid_disks, which means
2879 * we will suspend a really large IO window while it is not
2880 * necessary. If raid_disks is not divisible by near_copies,
2881 * an extra chunk is needed to ensure the whole "stripe" is
2882 * covered.
2883 */
2888 else
2892 /*
2893 * At least use a 32M window to align with raid1's resync window
2894 */
2899 }
2901 /*
2902 * perform a "sync" on one "block"
2903 *
2904 * We need to make sure that no normal I/O request - particularly write
2905 * requests - conflict with active sync requests.
2906 *
2907 * This is achieved by tracking pending requests and a 'barrier' concept
2908 * that can be installed to exclude normal IO requests.
2909 *
2910 * Resync and recovery are handled very differently.
2911 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2912 *
2913 * For resync, we iterate over virtual addresses, read all copies,
2914 * and update if there are differences. If only one copy is live,
2915 * skip it.
2916 * For recovery, we iterate over physical addresses, read a good
2917 * value for each non-in_sync drive, and over-write.
2918 *
2919 * So, for recovery we may have several outstanding complex requests for a
2920 * given address, one for each out-of-sync device. We model this by allocating
2921 * a number of r10_bio structures, one for each out-of-sync device.
2922 * As we setup these structures, we collect all bio's together into a list
2923 * which we then process collectively to add pages, and then process again
2924 * to pass to generic_make_request.
2925 *
2926 * The r10_bio structures are linked using a borrowed master_bio pointer.
2927 * This link is counted in ->remaining. When the r10_bio that points to NULL
2928 * has its remaining count decremented to 0, the whole complex operation
2929 * is complete.
2930 *
2931 */
2935 {
2942 sector_t sync_blocks;
2952 /*
2953 * Allow skipping a full rebuild for incremental assembly
2954 * of a clean array, like RAID1 does.
2955 */
2965 }
2967 skipped:
2976 /* If we aborted, we need to abort the
2977 * sync on the 'current' bitmap chucks (there can
2978 * be several when recovering multiple devices).
2979 * as we may have started syncing it but not finished.
2980 * We can find the current address in
2981 * mddev->curr_resync, but for recovery,
2982 * we need to convert that to several
2983 * virtual addresses.
2984 */
2989 }
2996 sector_t sect =
3000 }
3002 /* completed sync */
3006 /* Completed a full sync so the replacements
3007 * are now fully recovered.
3008 */
3015 }
3017 }
3019 }
3024 }
3030 /* if there has been nothing to do on any drive,
3031 * then there is nothing to do at all..
3032 */
3035 }
3040 /* make sure whole request will fit in a chunk - if chunks
3041 * are meaningful
3042 */
3047 /*
3048 * If there is non-resync activity waiting for a turn, then let it
3049 * though before starting on this new sync request.
3050 */
3054 /* Again, very different code for resync and recovery.
3055 * Both must result in an r10bio with a list of bios that
3056 * have bi_end_io, bi_sector, bi_disk set,
3057 * and bi_private set to the r10bio.
3058 * For recovery, we may actually create several r10bios
3059 * with 2 bios in each, that correspond to the bios in the main one.
3060 * In this case, the subordinate r10bios link back through a
3061 * borrowed master_bio pointer, and the counter in the master
3062 * includes a ref from each subordinate.
3063 */
3064 /* First, we decide what to do and set ->bi_end_io
3065 * To end_sync_read if we want to read, and
3066 * end_sync_write if we will want to write.
3067 */
3071 /* recovery... the complicated one */
3078 sector_t sect;
3095 }
3098 /* want to reconstruct this device */
3102 /* last stripe is not complete - don't
3103 * try to recover this sector.
3104 */
3107 }
3110 /* Unless we are doing a full sync, or a replacement
3111 * we only need to recover the block if it is set in
3112 * the bitmap
3113 */
3121 /* yep, skip the sync_blocks here, but don't assume
3122 * that there will never be anything to do here
3123 */
3127 }
3147 /* Need to check if the array will still be
3148 * degraded
3149 */
3157 }
3158 }
3175 /* This is where we read from */
3184 bad_sectors -= (sector
3189 }
3190 }
3203 /* and we write to 'i' (if not in_sync) */
3228 /* and maybe write to replacement */
3232 /* Note: if mreplace != NULL, then bio
3233 * cannot be NULL as r10buf_pool_alloc will
3234 * have allocated it.
3235 * So the second test here is pointless.
3236 * But it keeps semantic-checkers happy, and
3237 * this comment keeps human reviewers
3238 * happy.
3239 */
3252 }
3255 /* Cannot recover, so abort the recovery or
3256 * record a bad block */
3258 /* problem is that there are bad blocks
3259 * on other device(s)
3260 */
3268 mrdev,
3274 mreplace,
3278 }
3284 mirror->recovery_disabled
3286 }
3295 }
3300 /* Only want this if there is elsewhere to
3301 * read from. 'j' is currently the first
3302 * readable copy.
3303 */
3310 targets++;
3311 }
3315 }
3316 }
3323 }
3325 }
3327 /* resync. Schedule a read for every block at this virt offset */
3330 /*
3331 * Since curr_resync_completed could probably not update in
3332 * time, and we will set cluster_sync_low based on it.
3333 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3334 * safety reason, which ensures curr_resync_completed is
3335 * updated in bitmap_cond_end_sync.
3336 */
3346 /* We can skip this block */
3349 }
3382 }
3394 }
3395 }
3406 count++;
3412 }
3415 /* Need to set up for writing to the replacement */
3428 count++;
3430 }
3437 mddev);
3442 mddev);
3443 }
3447 }
3448 }
3463 /*
3464 * won't fail because the vec table is big enough
3465 * to hold all these pages
3466 */
3468 }
3476 /* It is resync not recovery */
3480 /* Send resync message */
3484 }
3486 /* This is recovery not resync */
3491 /*
3492 * sector_nr is a device address for recovery, so we
3493 * need translate it to array address before compare
3494 * with cluster_sync_high.
3495 */
3500 /*
3501 * curr_resync_completed is similar as
3502 * sector_nr, so make the translation too.
3503 */
3510 }
3511 }
3517 }
3518 }
3532 }
3533 }
3536 /* pretend they weren't skipped, it makes
3537 * no important difference in this case
3538 */
3542 giveup:
3543 /* There is nowhere to write, so all non-sync
3544 * drives must be failed or in resync, all drives
3545 * have a bad block, so try the next chunk...
3546 */
3551 chunks_skipped ++;
3554 }
3556 static sector_t
3558 {
3559 sector_t size;
3574 }
3577 {
3578 /* Calculate the number of sectors-per-device that will
3579 * actually be used, and set conf->dev_sectors and
3580 * conf->stride
3581 */
3587 /* 'size' is now the number of chunks in the array */
3588 /* calculate "used chunks per device" */
3591 /* We need to round up when dividing by raid_disks to
3592 * get the stride size.
3593 */
3603 }
3604 }
3608 {
3624 * updated yet. */
3629 }
3647 * actually using this, but leave code here just in case.*/
3657 }
3661 }
3664 {
3676 }
3682 }
3689 /* FIXME calc properly */
3692 GFP_KERNEL);
3719 }
3723 else
3724 /* far_copies must be 1 */
3726 }
3743 out:
3751 }
3753 }
3756 {
3761 sector_t size;
3774 }
3788 }
3789 }
3803 else
3806 }
3827 }
3845 }
3851 else
3854 }
3855 /* need to check that every block has at least one working mirror */
3860 }
3863 /* must ensure that shape change is supported */
3870 }
3876 i++) {
3881 /* The replacement is all we have - use it */
3885 }
3894 }
3896 }
3904 /*
3905 * Ok, everything is just fine now
3906 */
3917 /* Calculate max read-ahead size.
3918 * We need to readahead at least twice a whole stripe....
3919 * maybe...
3920 */
3924 }
3938 /* This cannot work */
3941 }
3950 }
3954 out_free_conf:
3961 out:
3963 }
3966 {
3977 }
3980 {
3985 else
3987 }
3990 {
3991 /* Resize of 'far' arrays is not supported.
3992 * For 'near' and 'offset' arrays we can set the
3993 * number of sectors used to be an appropriate multiple
3994 * of the chunk size.
3995 * For 'offset', this is far_copies*chunksize.
3996 * For 'near' the multiplier is the LCM of
3997 * near_copies and raid_disks.
3998 * So if far_copies > 1 && !far_offset, fail.
3999 * Else find LCM(raid_disks, near_copy)*far_copies and
4000 * multiply by chunk_size. Then round to this number.
4001 * This is mostly done by raid10_size()
4002 */
4021 }
4027 }
4032 }
4035 {
4043 }
4046 /* Set new parameters */
4048 /* new layout: far_copies = 1, near_copies = 2 */
4053 /* make sure it will be not marked as dirty */
4063 }
4065 }
4068 }
4071 {
4074 /* raid10 can take over:
4075 * raid0 - providing it has only two drives
4076 */
4078 /* for raid0 takeover only one zone is supported */
4084 }
4088 }
4090 }
4093 {
4094 /* Called when there is a request to change
4095 * - layout (to ->new_layout)
4096 * - chunk size (to ->new_chunk_sectors)
4097 * - raid_disks (by delta_disks)
4098 * or when trying to restart a reshape that was ongoing.
4099 *
4100 * We need to validate the request and possibly allocate
4101 * space if that might be an issue later.
4102 *
4103 * Currently we reject any reshape of a 'far' mode array,
4104 * allow chunk size to change if new is generally acceptable,
4105 * allow raid_disks to increase, and allow
4106 * a switch between 'near' mode and 'offset' mode.
4107 */
4115 /* mustn't change number of copies */
4118 /* Cannot switch to 'far' mode */
4122 /* not factor of array size */
4131 /* allocate new 'mirrors' list */
4136 GFP_KERNEL);
4139 }
4141 }
4143 /*
4144 * Need to check if array has failed when deciding whether to:
4145 * - start an array
4146 * - remove non-faulty devices
4147 * - add a spare
4148 * - allow a reshape
4149 * This determination is simple when no reshape is happening.
4150 * However if there is a reshape, we need to carefully check
4151 * both the before and after sections.
4152 * This is because some failed devices may only affect one
4153 * of the two sections, and some non-in_sync devices may
4154 * be insync in the section most affected by failed devices.
4155 */
4157 {
4163 /* 'prev' section first */
4167 degraded++;
4169 /* When we can reduce the number of devices in
4170 * an array, this might not contribute to
4171 * 'degraded'. It does now.
4172 */
4173 degraded++;
4174 }
4183 degraded2++;
4185 /* If reshape is increasing the number of devices,
4186 * this section has already been recovered, so
4187 * it doesn't contribute to degraded.
4188 * else it does.
4189 */
4191 degraded2++;
4192 }
4193 }
4198 }
4201 {
4202 /* A 'reshape' has been requested. This commits
4203 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4204 * This also checks if there are enough spares and adds them
4205 * to the array.
4206 * We currently require enough spares to make the final
4207 * array non-degraded. We also require that the difference
4208 * between old and new data_offset - on each device - is
4209 * enough that we never risk over-writing.
4210 */
4235 spares++;
4246 }
4247 }
4265 }
4275 }
4290 }
4299 else
4304 }
4307 /* This is a spare that was manually added */
4309 }
4310 }
4311 /* When a reshape changes the number of devices,
4312 * ->degraded is measured against the larger of the
4313 * pre and post numbers.
4314 */
4333 }
4339 abort:
4352 }
4354 /* Calculate the last device-address that could contain
4355 * any block from the chunk that includes the array-address 's'
4356 * and report the next address.
4357 * i.e. the address returned will be chunk-aligned and after
4358 * any data that is in the chunk containing 's'.
4359 */
4361 {
4369 }
4371 /* Calculate the first device-address that could contain
4372 * any block from the chunk that includes the array-address 's'.
4373 * This too will be the start of a chunk
4374 */
4376 {
4383 }
4387 {
4388 /* We simply copy at most one chunk (smallest of old and new)
4389 * at a time, possibly less if that exceeds RESYNC_PAGES,
4390 * or we hit a bad block or something.
4391 * This might mean we pause for normal IO in the middle of
4392 * a chunk, but that is not a problem as mddev->reshape_position
4393 * can record any location.
4394 *
4395 * If we will want to write to a location that isn't
4396 * yet recorded as 'safe' (i.e. in metadata on disk) then
4397 * we need to flush all reshape requests and update the metadata.
4398 *
4399 * When reshaping forwards (e.g. to more devices), we interpret
4400 * 'safe' as the earliest block which might not have been copied
4401 * down yet. We divide this by previous stripe size and multiply
4402 * by previous stripe length to get lowest device offset that we
4403 * cannot write to yet.
4404 * We interpret 'sector_nr' as an address that we want to write to.
4405 * From this we use last_device_address() to find where we might
4406 * write to, and first_device_address on the 'safe' position.
4407 * If this 'next' write position is after the 'safe' position,
4408 * we must update the metadata to increase the 'safe' position.
4409 *
4410 * When reshaping backwards, we round in the opposite direction
4411 * and perform the reverse test: next write position must not be
4412 * less than current safe position.
4413 *
4414 * In all this the minimum difference in data offsets
4415 * (conf->offset_diff - always positive) allows a bit of slack,
4416 * so next can be after 'safe', but not by more than offset_diff
4417 *
4418 * We need to prepare all the bios here before we start any IO
4419 * to ensure the size we choose is acceptable to all devices.
4420 * The means one for each copy for write-out and an extra one for
4421 * read-in.
4422 * We store the read-in bio in ->master_bio and the others in
4423 * ->devs[x].bio and ->devs[x].repl_bio.
4424 */
4439 /* If restarting in the middle, skip the initial sectors */
4452 }
4453 }
4455 /* We don't use sector_nr to track where we are up to
4456 * as that doesn't work well for ->reshape_backwards.
4457 * So just use ->reshape_progress.
4458 */
4460 /* 'next' is the earliest device address that we might
4461 * write to for this chunk in the new layout
4462 */
4466 /* 'safe' is the last device address that we might read from
4467 * in the old layout after a restart
4468 */
4481 /* 'next' is after the last device address that we
4482 * might write to for this chunk in the new layout
4483 */
4486 /* 'safe' is the earliest device address that we might
4487 * read from in the old layout after a restart
4488 */
4491 /* Need to update metadata if 'next' might be beyond 'safe'
4492 * as that would possibly corrupt data
4493 */
4503 }
4507 /* Need to update reshape_position in metadata */
4513 else
4523 }
4526 }
4528 read_more:
4529 /* Now schedule reads for blocks from sector_nr to last */
4542 /* Cannot read from here, so need to record bad blocks
4543 * on all the target devices.
4544 */
4545 // FIXME
4549 }
4566 /* Now find the locations in the new layout */
4583 }
4594 }
4596 /* Now add as many pages as possible to all of these bios. */
4606 /*
4607 * won't fail because the vec table is big enough
4608 * to hold all these pages
4609 */
4611 }
4614 }
4618 /* Now submit the read */
4628 /* Now that we have done the whole section we can
4629 * update reshape_progress
4630 */
4633 else
4637 }
4643 {
4644 /* Reshape read completed. Hopefully we have a block
4645 * to write out.
4646 * If we got a read error then we do sync 1-page reads from
4647 * elsewhere until we find the data - or give up.
4648 */
4654 /* Reshape has been aborted */
4657 }
4659 /* We definitely have the data in the pages, schedule the
4660 * writes.
4661 */
4674 }
4678 }
4685 }
4687 }
4690 {
4702 /* read-ahead size must cover two whole stripes, which is
4703 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4704 */
4711 }
4713 }
4717 {
4718 /* Use sync reads to get the blocks from somewhere else */
4731 }
4733 /* reshape IOs share pages from .devs[0].bio */
4751 sector_t addr;
4761 addr,
4769 failed:
4770 slot++;
4775 }
4778 /* couldn't read this block, must give up */
4783 }
4785 idx++;
4786 }
4789 }
4792 {
4807 }
4810 /* FIXME should record badblock */
4812 }
4816 }
4819 {
4825 }
4828 {
4838 }
4845 d++) {
4852 }
4854 }
4860 }
4863 {
4884 };
4887 {
4889 }
4892 {
4894 }