| blob | 99c9207899a777c615880f286a6f617ffc4a70bf |
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 (16 * 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 */
2662 rdev,
2667 rdev,
2671 }
2678 rdev,
2683 rdev,
2687 }
2688 }
2698 rdev,
2708 }
2710 }
2715 rdev,
2719 }
2720 }
2726 /*
2727 * In case freeze_array() is waiting for condition
2728 * nr_pending == nr_queued + extra to be true.
2729 */
2737 }
2738 }
2739 }
2742 {
2760 }
2761 }
2765 retry_list);
2774 }
2775 }
2786 }
2805 else
2811 }
2813 }
2816 {
2831 }
2834 {
2844 else
2857 }
2858 }
2860 }
2862 /*
2863 * Set cluster_sync_high since we need other nodes to add the
2864 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2865 */
2867 {
2868 sector_t window_size;
2871 /*
2872 * First, here we define "stripe" as a unit which across
2873 * all member devices one time, so we get chunks by use
2874 * raid_disks / near_copies. Otherwise, if near_copies is
2875 * close to raid_disks, then resync window could increases
2876 * linearly with the increase of raid_disks, which means
2877 * we will suspend a really large IO window while it is not
2878 * necessary. If raid_disks is not divisible by near_copies,
2879 * an extra chunk is needed to ensure the whole "stripe" is
2880 * covered.
2881 */
2886 else
2890 /*
2891 * At least use a 32M window to align with raid1's resync window
2892 */
2897 }
2899 /*
2900 * perform a "sync" on one "block"
2901 *
2902 * We need to make sure that no normal I/O request - particularly write
2903 * requests - conflict with active sync requests.
2904 *
2905 * This is achieved by tracking pending requests and a 'barrier' concept
2906 * that can be installed to exclude normal IO requests.
2907 *
2908 * Resync and recovery are handled very differently.
2909 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2910 *
2911 * For resync, we iterate over virtual addresses, read all copies,
2912 * and update if there are differences. If only one copy is live,
2913 * skip it.
2914 * For recovery, we iterate over physical addresses, read a good
2915 * value for each non-in_sync drive, and over-write.
2916 *
2917 * So, for recovery we may have several outstanding complex requests for a
2918 * given address, one for each out-of-sync device. We model this by allocating
2919 * a number of r10_bio structures, one for each out-of-sync device.
2920 * As we setup these structures, we collect all bio's together into a list
2921 * which we then process collectively to add pages, and then process again
2922 * to pass to generic_make_request.
2923 *
2924 * The r10_bio structures are linked using a borrowed master_bio pointer.
2925 * This link is counted in ->remaining. When the r10_bio that points to NULL
2926 * has its remaining count decremented to 0, the whole complex operation
2927 * is complete.
2928 *
2929 */
2933 {
2940 sector_t sync_blocks;
2950 /*
2951 * Allow skipping a full rebuild for incremental assembly
2952 * of a clean array, like RAID1 does.
2953 */
2963 }
2965 skipped:
2974 /* If we aborted, we need to abort the
2975 * sync on the 'current' bitmap chucks (there can
2976 * be several when recovering multiple devices).
2977 * as we may have started syncing it but not finished.
2978 * We can find the current address in
2979 * mddev->curr_resync, but for recovery,
2980 * we need to convert that to several
2981 * virtual addresses.
2982 */
2987 }
2994 sector_t sect =
2998 }
3000 /* completed sync */
3004 /* Completed a full sync so the replacements
3005 * are now fully recovered.
3006 */
3013 }
3015 }
3017 }
3022 }
3028 /* if there has been nothing to do on any drive,
3029 * then there is nothing to do at all..
3030 */
3033 }
3038 /* make sure whole request will fit in a chunk - if chunks
3039 * are meaningful
3040 */
3045 /*
3046 * If there is non-resync activity waiting for a turn, then let it
3047 * though before starting on this new sync request.
3048 */
3052 /* Again, very different code for resync and recovery.
3053 * Both must result in an r10bio with a list of bios that
3054 * have bi_end_io, bi_sector, bi_disk set,
3055 * and bi_private set to the r10bio.
3056 * For recovery, we may actually create several r10bios
3057 * with 2 bios in each, that correspond to the bios in the main one.
3058 * In this case, the subordinate r10bios link back through a
3059 * borrowed master_bio pointer, and the counter in the master
3060 * includes a ref from each subordinate.
3061 */
3062 /* First, we decide what to do and set ->bi_end_io
3063 * To end_sync_read if we want to read, and
3064 * end_sync_write if we will want to write.
3065 */
3069 /* recovery... the complicated one */
3076 sector_t sect;
3093 }
3096 /* want to reconstruct this device */
3100 /* last stripe is not complete - don't
3101 * try to recover this sector.
3102 */
3105 }
3108 /* Unless we are doing a full sync, or a replacement
3109 * we only need to recover the block if it is set in
3110 * the bitmap
3111 */
3119 /* yep, skip the sync_blocks here, but don't assume
3120 * that there will never be anything to do here
3121 */
3125 }
3145 /* Need to check if the array will still be
3146 * degraded
3147 */
3155 }
3156 }
3173 /* This is where we read from */
3182 bad_sectors -= (sector
3187 }
3188 }
3201 /* and we write to 'i' (if not in_sync) */
3226 /* and maybe write to replacement */
3230 /* Note: if mreplace != NULL, then bio
3231 * cannot be NULL as r10buf_pool_alloc will
3232 * have allocated it.
3233 * So the second test here is pointless.
3234 * But it keeps semantic-checkers happy, and
3235 * this comment keeps human reviewers
3236 * happy.
3237 */
3250 }
3253 /* Cannot recover, so abort the recovery or
3254 * record a bad block */
3256 /* problem is that there are bad blocks
3257 * on other device(s)
3258 */
3266 mrdev,
3272 mreplace,
3276 }
3282 mirror->recovery_disabled
3284 }
3293 }
3298 /* Only want this if there is elsewhere to
3299 * read from. 'j' is currently the first
3300 * readable copy.
3301 */
3308 targets++;
3309 }
3313 }
3314 }
3321 }
3323 }
3325 /* resync. Schedule a read for every block at this virt offset */
3328 /*
3329 * Since curr_resync_completed could probably not update in
3330 * time, and we will set cluster_sync_low based on it.
3331 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3332 * safety reason, which ensures curr_resync_completed is
3333 * updated in bitmap_cond_end_sync.
3334 */
3344 /* We can skip this block */
3347 }
3380 }
3392 }
3393 }
3404 count++;
3410 }
3413 /* Need to set up for writing to the replacement */
3426 count++;
3428 }
3435 mddev);
3440 mddev);
3441 }
3445 }
3446 }
3461 /*
3462 * won't fail because the vec table is big enough
3463 * to hold all these pages
3464 */
3466 }
3474 /* It is resync not recovery */
3478 /* Send resync message */
3482 }
3484 /* This is recovery not resync */
3489 /*
3490 * sector_nr is a device address for recovery, so we
3491 * need translate it to array address before compare
3492 * with cluster_sync_high.
3493 */
3498 /*
3499 * curr_resync_completed is similar as
3500 * sector_nr, so make the translation too.
3501 */
3508 }
3509 }
3515 }
3516 }
3530 }
3531 }
3534 /* pretend they weren't skipped, it makes
3535 * no important difference in this case
3536 */
3540 giveup:
3541 /* There is nowhere to write, so all non-sync
3542 * drives must be failed or in resync, all drives
3543 * have a bad block, so try the next chunk...
3544 */
3549 chunks_skipped ++;
3552 }
3554 static sector_t
3556 {
3557 sector_t size;
3572 }
3575 {
3576 /* Calculate the number of sectors-per-device that will
3577 * actually be used, and set conf->dev_sectors and
3578 * conf->stride
3579 */
3585 /* 'size' is now the number of chunks in the array */
3586 /* calculate "used chunks per device" */
3589 /* We need to round up when dividing by raid_disks to
3590 * get the stride size.
3591 */
3601 }
3602 }
3606 {
3622 * updated yet. */
3627 }
3645 * actually using this, but leave code here just in case.*/
3655 }
3659 }
3662 {
3674 }
3680 }
3687 /* FIXME calc properly */
3690 GFP_KERNEL);
3717 }
3721 else
3722 /* far_copies must be 1 */
3724 }
3741 out:
3749 }
3751 }
3754 {
3759 sector_t size;
3772 }
3786 }
3787 }
3801 else
3804 }
3825 }
3843 }
3849 else
3852 }
3853 /* need to check that every block has at least one working mirror */
3858 }
3861 /* must ensure that shape change is supported */
3868 }
3874 i++) {
3879 /* The replacement is all we have - use it */
3883 }
3892 }
3894 }
3902 /*
3903 * Ok, everything is just fine now
3904 */
3915 /* Calculate max read-ahead size.
3916 * We need to readahead at least twice a whole stripe....
3917 * maybe...
3918 */
3922 }
3936 /* This cannot work */
3939 }
3948 }
3952 out_free_conf:
3959 out:
3961 }
3964 {
3975 }
3978 {
3983 else
3985 }
3988 {
3989 /* Resize of 'far' arrays is not supported.
3990 * For 'near' and 'offset' arrays we can set the
3991 * number of sectors used to be an appropriate multiple
3992 * of the chunk size.
3993 * For 'offset', this is far_copies*chunksize.
3994 * For 'near' the multiplier is the LCM of
3995 * near_copies and raid_disks.
3996 * So if far_copies > 1 && !far_offset, fail.
3997 * Else find LCM(raid_disks, near_copy)*far_copies and
3998 * multiply by chunk_size. Then round to this number.
3999 * This is mostly done by raid10_size()
4000 */
4019 }
4025 }
4030 }
4033 {
4041 }
4044 /* Set new parameters */
4046 /* new layout: far_copies = 1, near_copies = 2 */
4051 /* make sure it will be not marked as dirty */
4061 }
4063 }
4066 }
4069 {
4072 /* raid10 can take over:
4073 * raid0 - providing it has only two drives
4074 */
4076 /* for raid0 takeover only one zone is supported */
4082 }
4086 }
4088 }
4091 {
4092 /* Called when there is a request to change
4093 * - layout (to ->new_layout)
4094 * - chunk size (to ->new_chunk_sectors)
4095 * - raid_disks (by delta_disks)
4096 * or when trying to restart a reshape that was ongoing.
4097 *
4098 * We need to validate the request and possibly allocate
4099 * space if that might be an issue later.
4100 *
4101 * Currently we reject any reshape of a 'far' mode array,
4102 * allow chunk size to change if new is generally acceptable,
4103 * allow raid_disks to increase, and allow
4104 * a switch between 'near' mode and 'offset' mode.
4105 */
4113 /* mustn't change number of copies */
4116 /* Cannot switch to 'far' mode */
4120 /* not factor of array size */
4129 /* allocate new 'mirrors' list */
4134 GFP_KERNEL);
4137 }
4139 }
4141 /*
4142 * Need to check if array has failed when deciding whether to:
4143 * - start an array
4144 * - remove non-faulty devices
4145 * - add a spare
4146 * - allow a reshape
4147 * This determination is simple when no reshape is happening.
4148 * However if there is a reshape, we need to carefully check
4149 * both the before and after sections.
4150 * This is because some failed devices may only affect one
4151 * of the two sections, and some non-in_sync devices may
4152 * be insync in the section most affected by failed devices.
4153 */
4155 {
4161 /* 'prev' section first */
4165 degraded++;
4167 /* When we can reduce the number of devices in
4168 * an array, this might not contribute to
4169 * 'degraded'. It does now.
4170 */
4171 degraded++;
4172 }
4181 degraded2++;
4183 /* If reshape is increasing the number of devices,
4184 * this section has already been recovered, so
4185 * it doesn't contribute to degraded.
4186 * else it does.
4187 */
4189 degraded2++;
4190 }
4191 }
4196 }
4199 {
4200 /* A 'reshape' has been requested. This commits
4201 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4202 * This also checks if there are enough spares and adds them
4203 * to the array.
4204 * We currently require enough spares to make the final
4205 * array non-degraded. We also require that the difference
4206 * between old and new data_offset - on each device - is
4207 * enough that we never risk over-writing.
4208 */
4233 spares++;
4244 }
4245 }
4263 }
4273 }
4288 }
4297 else
4302 }
4305 /* This is a spare that was manually added */
4307 }
4308 }
4309 /* When a reshape changes the number of devices,
4310 * ->degraded is measured against the larger of the
4311 * pre and post numbers.
4312 */
4331 }
4337 abort:
4350 }
4352 /* Calculate the last device-address that could contain
4353 * any block from the chunk that includes the array-address 's'
4354 * and report the next address.
4355 * i.e. the address returned will be chunk-aligned and after
4356 * any data that is in the chunk containing 's'.
4357 */
4359 {
4367 }
4369 /* Calculate the first device-address that could contain
4370 * any block from the chunk that includes the array-address 's'.
4371 * This too will be the start of a chunk
4372 */
4374 {
4381 }
4385 {
4386 /* We simply copy at most one chunk (smallest of old and new)
4387 * at a time, possibly less if that exceeds RESYNC_PAGES,
4388 * or we hit a bad block or something.
4389 * This might mean we pause for normal IO in the middle of
4390 * a chunk, but that is not a problem as mddev->reshape_position
4391 * can record any location.
4392 *
4393 * If we will want to write to a location that isn't
4394 * yet recorded as 'safe' (i.e. in metadata on disk) then
4395 * we need to flush all reshape requests and update the metadata.
4396 *
4397 * When reshaping forwards (e.g. to more devices), we interpret
4398 * 'safe' as the earliest block which might not have been copied
4399 * down yet. We divide this by previous stripe size and multiply
4400 * by previous stripe length to get lowest device offset that we
4401 * cannot write to yet.
4402 * We interpret 'sector_nr' as an address that we want to write to.
4403 * From this we use last_device_address() to find where we might
4404 * write to, and first_device_address on the 'safe' position.
4405 * If this 'next' write position is after the 'safe' position,
4406 * we must update the metadata to increase the 'safe' position.
4407 *
4408 * When reshaping backwards, we round in the opposite direction
4409 * and perform the reverse test: next write position must not be
4410 * less than current safe position.
4411 *
4412 * In all this the minimum difference in data offsets
4413 * (conf->offset_diff - always positive) allows a bit of slack,
4414 * so next can be after 'safe', but not by more than offset_diff
4415 *
4416 * We need to prepare all the bios here before we start any IO
4417 * to ensure the size we choose is acceptable to all devices.
4418 * The means one for each copy for write-out and an extra one for
4419 * read-in.
4420 * We store the read-in bio in ->master_bio and the others in
4421 * ->devs[x].bio and ->devs[x].repl_bio.
4422 */
4437 /* If restarting in the middle, skip the initial sectors */
4450 }
4451 }
4453 /* We don't use sector_nr to track where we are up to
4454 * as that doesn't work well for ->reshape_backwards.
4455 * So just use ->reshape_progress.
4456 */
4458 /* 'next' is the earliest device address that we might
4459 * write to for this chunk in the new layout
4460 */
4464 /* 'safe' is the last device address that we might read from
4465 * in the old layout after a restart
4466 */
4479 /* 'next' is after the last device address that we
4480 * might write to for this chunk in the new layout
4481 */
4484 /* 'safe' is the earliest device address that we might
4485 * read from in the old layout after a restart
4486 */
4489 /* Need to update metadata if 'next' might be beyond 'safe'
4490 * as that would possibly corrupt data
4491 */
4501 }
4505 /* Need to update reshape_position in metadata */
4511 else
4521 }
4524 }
4526 read_more:
4527 /* Now schedule reads for blocks from sector_nr to last */
4540 /* Cannot read from here, so need to record bad blocks
4541 * on all the target devices.
4542 */
4543 // FIXME
4547 }
4564 /* Now find the locations in the new layout */
4581 }
4592 }
4594 /* Now add as many pages as possible to all of these bios. */
4604 /*
4605 * won't fail because the vec table is big enough
4606 * to hold all these pages
4607 */
4609 }
4612 }
4616 /* Now submit the read */
4626 /* Now that we have done the whole section we can
4627 * update reshape_progress
4628 */
4631 else
4635 }
4641 {
4642 /* Reshape read completed. Hopefully we have a block
4643 * to write out.
4644 * If we got a read error then we do sync 1-page reads from
4645 * elsewhere until we find the data - or give up.
4646 */
4652 /* Reshape has been aborted */
4655 }
4657 /* We definitely have the data in the pages, schedule the
4658 * writes.
4659 */
4672 }
4676 }
4683 }
4685 }
4688 {
4700 /* read-ahead size must cover two whole stripes, which is
4701 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4702 */
4709 }
4711 }
4715 {
4716 /* Use sync reads to get the blocks from somewhere else */
4729 }
4731 /* reshape IOs share pages from .devs[0].bio */
4749 sector_t addr;
4759 addr,
4767 failed:
4768 slot++;
4773 }
4776 /* couldn't read this block, must give up */
4781 }
4783 idx++;
4784 }
4787 }
4790 {
4805 }
4808 /* FIXME should record badblock */
4810 }
4814 }
4817 {
4823 }
4826 {
4838 }
4843 }
4849 d++) {
4856 }
4858 }
4864 }
4867 {
4888 };
4891 {
4893 }
4896 {
4898 }