| blob | 374df5796649f49de33f38f3b845f6b6aef85e0a |
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 /* amount of memory to reserve for resync requests */
140 #define RESYNC_WINDOW (1024*1024)
141 /* maximum number of concurrent requests, memory permitting */
142 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144 /*
145 * When performing a resync, we need to read and compare, so
146 * we need as many pages are there are copies.
147 * When performing a recovery, we need 2 bios, one for read,
148 * one for write (we recover only one drive per r10buf)
149 *
150 */
152 {
167 else
170 /* allocate once for all bios */
173 else
179 /*
180 * Allocate bios.
181 */
193 }
194 /*
195 * Allocate RESYNC_PAGES data pages and attach them
196 * where needed.
197 */
215 }
222 }
223 }
227 out_free_pages:
232 out_free_bio:
238 }
240 out_free_r10bio:
243 }
246 {
262 }
264 /* resync pages array stored in the 1st bio's .bi_private */
268 }
271 {
283 }
284 }
287 {
292 }
295 {
301 }
304 {
314 /* wake up frozen array... */
318 }
320 /*
321 * raid_end_bio_io() is called when we have finished servicing a mirrored
322 * operation and are ready to return a success/failure code to the buffer
323 * cache layer.
324 */
326 {
334 /*
335 * Wake up any possible resync thread that waits for the device
336 * to go idle.
337 */
341 }
343 /*
344 * Update disk head position estimator based on IRQ completion info.
345 */
347 {
352 }
354 /*
355 * Find the disk number which triggered given bio
356 */
359 {
369 }
370 }
380 }
383 {
393 /*
394 * this branch is our 'one mirror IO has finished' event handler:
395 */
399 /*
400 * Set R10BIO_Uptodate in our master bio, so that
401 * we will return a good error code to the higher
402 * levels even if IO on some other mirrored buffer fails.
403 *
404 * The 'master' represents the composite IO operation to
405 * user-side. So if something waits for IO, then it will
406 * wait for the 'master' bio.
407 */
410 /* If all other devices that store this block have
411 * failed, we want to return the error upwards rather
412 * than fail the last device. Here we redefine
413 * "uptodate" to mean "Don't want to retry"
414 */
418 }
423 /*
424 * oops, read error - keep the refcount on the rdev
425 */
433 }
434 }
437 {
438 /* clear the bitmap if all writes complete successfully */
444 }
447 {
455 else
457 }
458 }
459 }
462 {
482 }
483 /*
484 * this branch is our 'one mirror IO has finished' event handler:
485 */
488 /* Never record new bad blocks to replacement,
489 * just fail it.
490 */
503 /* This is the only remaining device,
504 * We need to retry the write without
505 * FailFast
506 */
512 }
515 }
517 /*
518 * Set R10BIO_Uptodate in our master bio, so that
519 * we will return a good error code for to the higher
520 * levels even if IO on some other mirrored buffer fails.
521 *
522 * The 'master' represents the composite IO operation to
523 * user-side. So if something waits for IO, then it will
524 * wait for the 'master' bio.
525 */
526 sector_t first_bad;
529 /*
530 * Do not set R10BIO_Uptodate if the current device is
531 * rebuilding or Faulty. This is because we cannot use
532 * such device for properly reading the data back (we could
533 * potentially use it, if the current write would have felt
534 * before rdev->recovery_offset, but for simplicity we don't
535 * check this here.
536 */
541 /* Maybe we can clear some bad blocks. */
549 else
553 }
554 }
556 /*
557 *
558 * Let's see if all mirrored write operations have finished
559 * already.
560 */
566 }
568 /*
569 * RAID10 layout manager
570 * As well as the chunksize and raid_disks count, there are two
571 * parameters: near_copies and far_copies.
572 * near_copies * far_copies must be <= raid_disks.
573 * Normally one of these will be 1.
574 * If both are 1, we get raid0.
575 * If near_copies == raid_disks, we get raid1.
576 *
577 * Chunks are laid out in raid0 style with near_copies copies of the
578 * first chunk, followed by near_copies copies of the next chunk and
579 * so on.
580 * If far_copies > 1, then after 1/far_copies of the array has been assigned
581 * as described above, we start again with a device offset of near_copies.
582 * So we effectively have another copy of the whole array further down all
583 * the drives, but with blocks on different drives.
584 * With this layout, and block is never stored twice on the one device.
585 *
586 * raid10_find_phys finds the sector offset of a given virtual sector
587 * on each device that it is on.
588 *
589 * raid10_find_virt does the reverse mapping, from a device and a
590 * sector offset to a virtual address
591 */
594 {
596 sector_t sector;
597 sector_t chunk;
598 sector_t stripe;
609 /* now calculate first sector/dev */
621 /* and calculate all the others */
628 slot++;
642 }
646 slot++;
647 }
648 dev++;
652 }
653 }
654 }
657 {
669 }
672 {
674 /* Never use conf->prev as this is only called during resync
675 * or recovery, so reshape isn't happening
676 */
690 }
691 }
706 else
708 }
710 }
714 }
716 /*
717 * This routine returns the disk from which the requested read should
718 * be done. There is a per-array 'next expected sequential IO' sector
719 * number - if this matches on the next IO then we use the last disk.
720 * There is also a per-disk 'last know head position' sector that is
721 * maintained from IRQ contexts, both the normal and the resync IO
722 * completion handlers update this position correctly. If there is no
723 * perfect sequential match then we pick the disk whose head is closest.
724 *
725 * If there are 2 mirrors in the same 2 devices, performance degrades
726 * because position is mirror, not device based.
727 *
728 * The rdev for the device selected will have nr_pending incremented.
729 */
731 /*
732 * FIXME: possibly should rethink readbalancing and do it differently
733 * depending on near_copies / far_copies geometry.
734 */
738 {
758 /*
759 * Check if we can balance. We can balance on the whole
760 * device if no resync is going on (recovery is ok), or below
761 * the resync window. We take the first readable disk when
762 * above the resync window.
763 */
769 sector_t first_bad;
771 sector_t dev_sector;
791 /* Already have a better slot */
794 /* Cannot read here. If this is the
795 * 'primary' device, then we must not read
796 * beyond 'bad_sectors' from another device.
797 */
804 sector_t good_sectors =
810 }
812 /* Must read from here */
814 }
823 /* At least 2 disks to choose from so failfast is OK */
825 /* This optimisation is debatable, and completely destroys
826 * sequential read speed for 'far copies' arrays. So only
827 * keep it for 'near' arrays, and review those later.
828 */
832 /* for far > 1 always use the lowest address */
835 else
842 }
843 }
847 }
858 }
861 {
873 i++) {
879 }
880 }
883 }
886 {
887 /* Any writes that have been queued but are awaiting
888 * bitmap updates get flushed here.
889 */
897 /* flush any pending bitmap writes to disk
898 * before proceeding w/ I/O */
911 /* Just ignore it */
913 else
916 }
919 }
921 /* Barriers....
922 * Sometimes we need to suspend IO while we do something else,
923 * either some resync/recovery, or reconfigure the array.
924 * To do this we raise a 'barrier'.
925 * The 'barrier' is a counter that can be raised multiple times
926 * to count how many activities are happening which preclude
927 * normal IO.
928 * We can only raise the barrier if there is no pending IO.
929 * i.e. if nr_pending == 0.
930 * We choose only to raise the barrier if no-one is waiting for the
931 * barrier to go down. This means that as soon as an IO request
932 * is ready, no other operations which require a barrier will start
933 * until the IO request has had a chance.
934 *
935 * So: regular IO calls 'wait_barrier'. When that returns there
936 * is no backgroup IO happening, It must arrange to call
937 * allow_barrier when it has finished its IO.
938 * backgroup IO calls must call raise_barrier. Once that returns
939 * there is no normal IO happeing. It must arrange to call
940 * lower_barrier when the particular background IO completes.
941 */
944 {
948 /* Wait until no block IO is waiting (unless 'force') */
952 /* block any new IO from starting */
955 /* Now wait for all pending IO to complete */
961 }
964 {
970 }
973 {
977 /* Wait for the barrier to drop.
978 * However if there are already pending
979 * requests (preventing the barrier from
980 * rising completely), and the
981 * pre-process bio queue isn't empty,
982 * then don't wait, as we need to empty
983 * that queue to get the nr_pending
984 * count down.
985 */
997 }
1000 }
1003 {
1007 }
1010 {
1011 /* stop syncio and normal IO and wait for everything to
1012 * go quiet.
1013 * We increment barrier and nr_waiting, and then
1014 * wait until nr_pending match nr_queued+extra
1015 * This is called in the context of one normal IO request
1016 * that has failed. Thus any sync request that might be pending
1017 * will be blocked by nr_pending, and we need to wait for
1018 * pending IO requests to complete or be queued for re-try.
1019 * Thus the number queued (nr_queued) plus this request (extra)
1020 * must match the number of pending IOs (nr_pending) before
1021 * we continue.
1022 */
1034 }
1037 {
1038 /* reverse the effect of the freeze */
1044 }
1048 {
1052 else
1054 }
1060 };
1063 {
1065 cb);
1079 }
1081 /* we aren't scheduling, so we can do the write-out directly. */
1095 /* Just ignore it */
1097 else
1100 }
1102 }
1106 {
1112 sector_t sectors;
1120 /*
1121 * This is an error retry, but we cannot
1122 * safely dereference the rdev in the r10_bio,
1123 * we must use the one in conf.
1124 * If it has already been disconnected (unlikely)
1125 * we lose the device name in error messages.
1126 */
1128 /*
1129 * As we are blocking raid10, it is a little safer to
1130 * use __GFP_HIGH.
1131 */
1141 /* This never gets dereferenced */
1143 }
1145 }
1146 /*
1147 * Register the new request and wait if the reconstruction
1148 * thread has put up a bar for new requests.
1149 * Continue immediately if no resync is active currently.
1150 */
1157 /*
1158 * IO spans the reshape position. Need to wait for reshape to
1159 * pass
1160 */
1166 sectors);
1168 }
1176 }
1179 }
1193 }
1217 }
1222 {
1237 /* Replacement just got moved to main 'rdev' */
1240 }
1247 else
1265 /* flush_pending_writes() needs access to the rdev so...*/
1273 else
1284 }
1285 }
1289 {
1293 sector_t sectors;
1296 /*
1297 * Register the new request and wait if the reconstruction
1298 * thread has put up a bar for new requests.
1299 * Continue immediately if no resync is active currently.
1300 */
1307 /*
1308 * IO spans the reshape position. Need to wait for reshape to
1309 * pass
1310 */
1316 sectors);
1318 }
1326 /* Need to update reshape_position in metadata */
1336 }
1343 }
1344 /* first select target devices under rcu_lock and
1345 * inc refcount on their rdev. Record them by setting
1346 * bios[x] to bio
1347 * If there are known/acknowledged bad blocks on any device
1348 * on which we have seen a write error, we want to avoid
1349 * writing to those blocks. This potentially requires several
1350 * writes to write around the bad blocks. Each set of writes
1351 * gets its own r10_bio with a set of bios attached.
1352 */
1356 retry_write:
1372 }
1377 }
1389 }
1391 sector_t first_bad;
1399 /* Mustn't write here until the bad block
1400 * is acknowledged
1401 */
1406 }
1408 /* Cannot write here at all */
1411 /* Mustn't write more than bad_sectors
1412 * to other devices yet
1413 */
1415 /* We don't set R10BIO_Degraded as that
1416 * only applies if the disk is missing,
1417 * so it might be re-added, and we want to
1418 * know to recover this chunk.
1419 * In this case the device is here, and the
1420 * fact that this chunk is not in-sync is
1421 * recorded in the bad block log.
1422 */
1424 }
1429 }
1430 }
1434 }
1438 }
1439 }
1443 /* Have to wait for this device to get unblocked, then retry */
1451 }
1457 /* Race with remove_disk */
1460 }
1462 }
1463 }
1469 }
1481 }
1491 }
1493 }
1496 {
1512 else
1514 }
1517 {
1526 }
1531 /*
1532 * If this request crosses a chunk boundary, we need to split
1533 * it.
1534 */
1536 sectors > chunk_sects
1545 /* In case raid10d snuck in to freeze_array */
1548 }
1551 {
1562 else
1566 }
1573 }
1576 }
1578 /* check if there are enough drives for
1579 * every block to appear on atleast one.
1580 * Don't consider the device numbered 'ignore'
1581 * as we might be about to remove it.
1582 */
1584 {
1594 }
1606 cnt++;
1608 }
1614 out:
1617 }
1620 {
1621 /* when calling 'enough', both 'prev' and 'geo' must
1622 * be stable.
1623 * This is ensured if ->reconfig_mutex or ->device_lock
1624 * is held.
1625 */
1628 }
1631 {
1636 /*
1637 * If it is not operational, then we have already marked it as dead
1638 * else if it is the last working disks, ignore the error, let the
1639 * next level up know.
1640 * else mark the drive as failed
1641 */
1645 /*
1646 * Don't fail the drive, just return an IO error.
1647 */
1650 }
1653 /*
1654 * If recovery is running, make sure it aborts.
1655 */
1666 }
1669 {
1677 }
1681 /* This is only called with ->reconfix_mutex held, so
1682 * rcu protection of rdev is not needed */
1691 }
1692 }
1695 {
1701 }
1704 {
1711 /*
1712 * Find all non-in_sync disks within the RAID10 configuration
1713 * and mark them in_sync
1714 */
1721 /* Replacement has just become active */
1724 count++;
1726 /* Replaced device not technically faulty,
1727 * but we need to be sure it gets removed
1728 * and never re-added.
1729 */
1733 }
1739 count++;
1741 }
1742 }
1749 }
1752 {
1760 /* only hot-add to in-sync arrays, as recovery is
1761 * very different from resync
1762 */
1776 else
1796 }
1810 }
1816 }
1819 {
1831 else
1838 }
1839 /* Only remove non-faulty devices if recovery
1840 * is not possible.
1841 */
1849 }
1854 /* lost the race, try later */
1858 }
1859 }
1861 /* We must have just cleared 'rdev' */
1865 * but will never see neither -- if they are careful.
1866 */
1868 }
1873 abort:
1877 }
1880 {
1885 else
1886 /* The write handler will notice the lack of
1887 * R10BIO_Uptodate and record any errors etc
1888 */
1892 /* for reconstruct, we always reschedule after a read.
1893 * for resync, only after all reads
1894 */
1898 /* we have read all the blocks,
1899 * do the comparison in process context in raid10d
1900 */
1902 }
1903 }
1906 {
1912 }
1915 {
1916 /* reshape read bio isn't allocated from r10buf_pool */
1920 }
1923 {
1928 /* the primary of several recovery bios */
1933 else
1942 else
1945 }
1946 }
1947 }
1950 {
1955 sector_t first_bad;
1964 else
1976 }
1986 }
1988 /*
1989 * Note: sync and recover and handled very differently for raid10
1990 * This code is for resync.
1991 * For resync, we read through virtual addresses and read all blocks.
1992 * If there is any error, we schedule a write. The lowest numbered
1993 * drive is authoritative.
1994 * However requests come for physical address, so we need to map.
1995 * For every physical address there are raid_disks/copies virtual addresses,
1996 * which is always are least one, but is not necessarly an integer.
1997 * This means that a physical address can span multiple chunks, so we may
1998 * have to submit multiple io requests for a single sync request.
1999 */
2000 /*
2001 * We check if all blocks are in-sync and only write to blocks that
2002 * aren't in sync
2003 */
2005 {
2014 /* find the first device with a block */
2029 /* now find blocks with errors */
2046 /* We know that the bi_io_vec layout is the same for
2047 * both 'first' and 'i', so we just compare them.
2048 * All vec entries are PAGE_SIZE;
2049 */
2057 len))
2060 }
2065 /* Don't fix anything. */
2068 /* Just give up on this device */
2071 }
2072 /* Ok, we need to write this bio, either to correct an
2073 * inconsistency or to correct an unreadable block.
2074 * First we need to fixup bv_offset, bv_len and
2075 * bi_vecs, as the read request might have corrupted these
2076 */
2099 }
2101 /* Now write out to any replacement devices
2102 * that are active
2103 */
2118 }
2120 done:
2124 }
2125 }
2127 /*
2128 * Now for the recovery code.
2129 * Recovery happens across physical sectors.
2130 * We recover all non-is_sync drives by finding the virtual address of
2131 * each, and then choose a working drive that also has that virt address.
2132 * There is a separate r10_bio for each non-in_sync drive.
2133 * Only the first two slots are in use. The first for reading,
2134 * The second for writing.
2135 *
2136 */
2138 {
2139 /* We got a read error during recovery.
2140 * We repeat the read in smaller page-sized sections.
2141 * If a read succeeds, write it to the new device or record
2142 * a bad block if we cannot.
2143 * If a read fails, record a bad block on both old and
2144 * new devices.
2145 */
2159 sector_t addr;
2168 addr,
2176 addr,
2186 }
2187 }
2189 /* We don't worry if we cannot set a bad block -
2190 * it really is bad so there is no loss in not
2191 * recording it yet
2192 */
2196 /* need bad block on destination too */
2201 /* just abort the recovery */
2210 }
2211 }
2212 }
2216 idx++;
2217 }
2218 }
2221 {
2230 }
2232 /*
2233 * share the pages with the first bio
2234 * and submit the write request
2235 */
2239 /* Need to test wbio2->bi_end_io before we call
2240 * generic_make_request as if the former is NULL,
2241 * the latter is free to free wbio2.
2242 */
2249 }
2255 }
2256 }
2258 /*
2259 * Used by fix_read_error() to decay the per rdev read_errors.
2260 * We halve the read error count for every hour that has elapsed
2261 * since the last recorded read error.
2262 *
2263 */
2265 {
2273 /* first time we've seen a read error */
2276 }
2283 /*
2284 * if hours_since_last is > the number of bits in read_errors
2285 * just set read errors to 0. We do this to avoid
2286 * overflowing the shift of read_errors by hours_since_last.
2287 */
2290 else
2292 }
2296 {
2297 sector_t first_bad;
2304 /* success */
2311 }
2312 /* need to record an error - either for the block or the device */
2316 }
2318 /*
2319 * This is a kernel thread which:
2320 *
2321 * 1. Retries failed read operations on working mirrors.
2322 * 2. Updates the raid superblock when problems encounter.
2323 * 3. Performs writes following reads for array synchronising.
2324 */
2327 {
2334 /* still own a reference to this rdev, so it cannot
2335 * have been cleared recently.
2336 */
2340 /* drive has already been failed, just ignore any
2341 more fix_read_error() attempts */
2358 }
2371 sector_t first_bad;
2385 sect,
2393 }
2394 sl++;
2401 /* Cannot read from anywhere, just mark the block
2402 * as bad on the first device to discourage future
2403 * reads.
2404 */
2409 rdev,
2416 }
2418 }
2421 /* write it back and re-read */
2428 sl--;
2440 sect,
2443 /* Well, this device is dead */
2447 sect +
2449 rdev)),
2454 }
2457 }
2464 sl--;
2476 sect,
2478 READ)) {
2480 /* Well, this device is dead */
2484 sect +
2495 sect +
2499 }
2503 }
2508 }
2509 }
2512 {
2517 /* bio has the data to be written to slot 'i' where
2518 * we just recently had a write error.
2519 * We repeatedly clone the bio and trim down to one block,
2520 * then try the write. Where the write fails we record
2521 * a bad block.
2522 * It is conceivable that the bio doesn't exactly align with
2523 * blocks. We must handle this.
2524 *
2525 * We currently own a reference to the rdev.
2526 */
2529 sector_t sector;
2546 sector_t wsector;
2549 /* Write at 'sector' for 'sectors' */
2559 /* Failure! */
2568 }
2570 }
2573 {
2578 sector_t bio_last_sector;
2580 /* we got a read error. Maybe the drive is bad. Maybe just
2581 * the block and we can fix it.
2582 * We freeze all other IO, and try reading the block from
2583 * other devices. When we find one, we re-write
2584 * and check it that fixes the read error.
2585 * This is all done synchronously while the array is
2586 * frozen.
2587 */
2606 }
2609 {
2610 /* Some sort of write request has finished and it
2611 * succeeded in writing where we thought there was a
2612 * bad block. So forget the bad block.
2613 * Or possibly if failed and we need to record
2614 * a bad block.
2615 */
2628 rdev,
2633 rdev,
2637 }
2644 rdev,
2649 rdev,
2653 }
2654 }
2664 rdev,
2674 }
2676 }
2681 rdev,
2685 }
2686 }
2692 /*
2693 * In case freeze_array() is waiting for condition
2694 * nr_pending == nr_queued + extra to be true.
2695 */
2703 }
2704 }
2705 }
2708 {
2726 }
2727 }
2731 retry_list);
2740 }
2741 }
2752 }
2771 else
2777 }
2779 }
2782 {
2797 }
2800 {
2810 else
2823 }
2824 }
2826 }
2828 /*
2829 * perform a "sync" on one "block"
2830 *
2831 * We need to make sure that no normal I/O request - particularly write
2832 * requests - conflict with active sync requests.
2833 *
2834 * This is achieved by tracking pending requests and a 'barrier' concept
2835 * that can be installed to exclude normal IO requests.
2836 *
2837 * Resync and recovery are handled very differently.
2838 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2839 *
2840 * For resync, we iterate over virtual addresses, read all copies,
2841 * and update if there are differences. If only one copy is live,
2842 * skip it.
2843 * For recovery, we iterate over physical addresses, read a good
2844 * value for each non-in_sync drive, and over-write.
2845 *
2846 * So, for recovery we may have several outstanding complex requests for a
2847 * given address, one for each out-of-sync device. We model this by allocating
2848 * a number of r10_bio structures, one for each out-of-sync device.
2849 * As we setup these structures, we collect all bio's together into a list
2850 * which we then process collectively to add pages, and then process again
2851 * to pass to generic_make_request.
2852 *
2853 * The r10_bio structures are linked using a borrowed master_bio pointer.
2854 * This link is counted in ->remaining. When the r10_bio that points to NULL
2855 * has its remaining count decremented to 0, the whole complex operation
2856 * is complete.
2857 *
2858 */
2862 {
2869 sector_t sync_blocks;
2879 /*
2880 * Allow skipping a full rebuild for incremental assembly
2881 * of a clean array, like RAID1 does.
2882 */
2892 }
2894 skipped:
2900 /* If we aborted, we need to abort the
2901 * sync on the 'current' bitmap chucks (there can
2902 * be several when recovering multiple devices).
2903 * as we may have started syncing it but not finished.
2904 * We can find the current address in
2905 * mddev->curr_resync, but for recovery,
2906 * we need to convert that to several
2907 * virtual addresses.
2908 */
2913 }
2920 sector_t sect =
2924 }
2926 /* completed sync */
2930 /* Completed a full sync so the replacements
2931 * are now fully recovered.
2932 */
2939 }
2941 }
2943 }
2948 }
2954 /* if there has been nothing to do on any drive,
2955 * then there is nothing to do at all..
2956 */
2959 }
2964 /* make sure whole request will fit in a chunk - if chunks
2965 * are meaningful
2966 */
2971 /*
2972 * If there is non-resync activity waiting for a turn, then let it
2973 * though before starting on this new sync request.
2974 */
2978 /* Again, very different code for resync and recovery.
2979 * Both must result in an r10bio with a list of bios that
2980 * have bi_end_io, bi_sector, bi_disk set,
2981 * and bi_private set to the r10bio.
2982 * For recovery, we may actually create several r10bios
2983 * with 2 bios in each, that correspond to the bios in the main one.
2984 * In this case, the subordinate r10bios link back through a
2985 * borrowed master_bio pointer, and the counter in the master
2986 * includes a ref from each subordinate.
2987 */
2988 /* First, we decide what to do and set ->bi_end_io
2989 * To end_sync_read if we want to read, and
2990 * end_sync_write if we will want to write.
2991 */
2995 /* recovery... the complicated one */
3002 sector_t sect;
3019 }
3022 /* want to reconstruct this device */
3026 /* last stripe is not complete - don't
3027 * try to recover this sector.
3028 */
3031 }
3034 /* Unless we are doing a full sync, or a replacement
3035 * we only need to recover the block if it is set in
3036 * the bitmap
3037 */
3045 /* yep, skip the sync_blocks here, but don't assume
3046 * that there will never be anything to do here
3047 */
3051 }
3071 /* Need to check if the array will still be
3072 * degraded
3073 */
3081 }
3082 }
3099 /* This is where we read from */
3108 bad_sectors -= (sector
3113 }
3114 }
3127 /* and we write to 'i' (if not in_sync) */
3152 /* and maybe write to replacement */
3156 /* Note: if mreplace != NULL, then bio
3157 * cannot be NULL as r10buf_pool_alloc will
3158 * have allocated it.
3159 * So the second test here is pointless.
3160 * But it keeps semantic-checkers happy, and
3161 * this comment keeps human reviewers
3162 * happy.
3163 */
3176 }
3179 /* Cannot recover, so abort the recovery or
3180 * record a bad block */
3182 /* problem is that there are bad blocks
3183 * on other device(s)
3184 */
3192 mrdev,
3198 mreplace,
3202 }
3208 mirror->recovery_disabled
3210 }
3219 }
3224 /* Only want this if there is elsewhere to
3225 * read from. 'j' is currently the first
3226 * readable copy.
3227 */
3234 targets++;
3235 }
3239 }
3240 }
3247 }
3249 }
3251 /* resync. Schedule a read for every block at this virt offset */
3260 /* We can skip this block */
3263 }
3296 }
3308 }
3309 }
3320 count++;
3326 }
3329 /* Need to set up for writing to the replacement */
3342 count++;
3344 }
3351 mddev);
3356 mddev);
3357 }
3361 }
3362 }
3377 /*
3378 * won't fail because the vec table is big enough
3379 * to hold all these pages
3380 */
3382 }
3400 }
3401 }
3404 /* pretend they weren't skipped, it makes
3405 * no important difference in this case
3406 */
3410 giveup:
3411 /* There is nowhere to write, so all non-sync
3412 * drives must be failed or in resync, all drives
3413 * have a bad block, so try the next chunk...
3414 */
3419 chunks_skipped ++;
3422 }
3424 static sector_t
3426 {
3427 sector_t size;
3442 }
3445 {
3446 /* Calculate the number of sectors-per-device that will
3447 * actually be used, and set conf->dev_sectors and
3448 * conf->stride
3449 */
3455 /* 'size' is now the number of chunks in the array */
3456 /* calculate "used chunks per device" */
3459 /* We need to round up when dividing by raid_disks to
3460 * get the stride size.
3461 */
3471 }
3472 }
3476 {
3492 * updated yet. */
3497 }
3515 * actually using this, but leave code here just in case.*/
3525 }
3529 }
3532 {
3544 }
3550 }
3557 /* FIXME calc properly */
3560 GFP_KERNEL);
3587 }
3591 else
3592 /* far_copies must be 1 */
3594 }
3611 out:
3619 }
3621 }
3624 {
3629 sector_t size;
3642 }
3659 else
3662 }
3683 }
3701 }
3707 else
3710 }
3711 /* need to check that every block has at least one working mirror */
3716 }
3719 /* must ensure that shape change is supported */
3726 }
3732 i++) {
3737 /* The replacement is all we have - use it */
3741 }
3750 }
3752 }
3760 /*
3761 * Ok, everything is just fine now
3762 */
3773 /* Calculate max read-ahead size.
3774 * We need to readahead at least twice a whole stripe....
3775 * maybe...
3776 */
3780 }
3794 /* This cannot work */
3797 }
3806 }
3810 out_free_conf:
3817 out:
3819 }
3822 {
3833 }
3836 {
3846 }
3847 }
3850 {
3851 /* Resize of 'far' arrays is not supported.
3852 * For 'near' and 'offset' arrays we can set the
3853 * number of sectors used to be an appropriate multiple
3854 * of the chunk size.
3855 * For 'offset', this is far_copies*chunksize.
3856 * For 'near' the multiplier is the LCM of
3857 * near_copies and raid_disks.
3858 * So if far_copies > 1 && !far_offset, fail.
3859 * Else find LCM(raid_disks, near_copy)*far_copies and
3860 * multiply by chunk_size. Then round to this number.
3861 * This is mostly done by raid10_size()
3862 */
3881 }
3887 }
3892 }
3895 {
3903 }
3906 /* Set new parameters */
3908 /* new layout: far_copies = 1, near_copies = 2 */
3913 /* make sure it will be not marked as dirty */
3923 }
3925 }
3928 }
3931 {
3934 /* raid10 can take over:
3935 * raid0 - providing it has only two drives
3936 */
3938 /* for raid0 takeover only one zone is supported */
3944 }
3948 }
3950 }
3953 {
3954 /* Called when there is a request to change
3955 * - layout (to ->new_layout)
3956 * - chunk size (to ->new_chunk_sectors)
3957 * - raid_disks (by delta_disks)
3958 * or when trying to restart a reshape that was ongoing.
3959 *
3960 * We need to validate the request and possibly allocate
3961 * space if that might be an issue later.
3962 *
3963 * Currently we reject any reshape of a 'far' mode array,
3964 * allow chunk size to change if new is generally acceptable,
3965 * allow raid_disks to increase, and allow
3966 * a switch between 'near' mode and 'offset' mode.
3967 */
3975 /* mustn't change number of copies */
3978 /* Cannot switch to 'far' mode */
3982 /* not factor of array size */
3991 /* allocate new 'mirrors' list */
3996 GFP_KERNEL);
3999 }
4001 }
4003 /*
4004 * Need to check if array has failed when deciding whether to:
4005 * - start an array
4006 * - remove non-faulty devices
4007 * - add a spare
4008 * - allow a reshape
4009 * This determination is simple when no reshape is happening.
4010 * However if there is a reshape, we need to carefully check
4011 * both the before and after sections.
4012 * This is because some failed devices may only affect one
4013 * of the two sections, and some non-in_sync devices may
4014 * be insync in the section most affected by failed devices.
4015 */
4017 {
4023 /* 'prev' section first */
4027 degraded++;
4029 /* When we can reduce the number of devices in
4030 * an array, this might not contribute to
4031 * 'degraded'. It does now.
4032 */
4033 degraded++;
4034 }
4043 degraded2++;
4045 /* If reshape is increasing the number of devices,
4046 * this section has already been recovered, so
4047 * it doesn't contribute to degraded.
4048 * else it does.
4049 */
4051 degraded2++;
4052 }
4053 }
4058 }
4061 {
4062 /* A 'reshape' has been requested. This commits
4063 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4064 * This also checks if there are enough spares and adds them
4065 * to the array.
4066 * We currently require enough spares to make the final
4067 * array non-degraded. We also require that the difference
4068 * between old and new data_offset - on each device - is
4069 * enough that we never risk over-writing.
4070 */
4095 spares++;
4106 }
4107 }
4125 }
4135 }
4150 }
4159 else
4164 }
4167 /* This is a spare that was manually added */
4169 }
4170 }
4171 /* When a reshape changes the number of devices,
4172 * ->degraded is measured against the larger of the
4173 * pre and post numbers.
4174 */
4193 }
4199 abort:
4212 }
4214 /* Calculate the last device-address that could contain
4215 * any block from the chunk that includes the array-address 's'
4216 * and report the next address.
4217 * i.e. the address returned will be chunk-aligned and after
4218 * any data that is in the chunk containing 's'.
4219 */
4221 {
4229 }
4231 /* Calculate the first device-address that could contain
4232 * any block from the chunk that includes the array-address 's'.
4233 * This too will be the start of a chunk
4234 */
4236 {
4243 }
4247 {
4248 /* We simply copy at most one chunk (smallest of old and new)
4249 * at a time, possibly less if that exceeds RESYNC_PAGES,
4250 * or we hit a bad block or something.
4251 * This might mean we pause for normal IO in the middle of
4252 * a chunk, but that is not a problem as mddev->reshape_position
4253 * can record any location.
4254 *
4255 * If we will want to write to a location that isn't
4256 * yet recorded as 'safe' (i.e. in metadata on disk) then
4257 * we need to flush all reshape requests and update the metadata.
4258 *
4259 * When reshaping forwards (e.g. to more devices), we interpret
4260 * 'safe' as the earliest block which might not have been copied
4261 * down yet. We divide this by previous stripe size and multiply
4262 * by previous stripe length to get lowest device offset that we
4263 * cannot write to yet.
4264 * We interpret 'sector_nr' as an address that we want to write to.
4265 * From this we use last_device_address() to find where we might
4266 * write to, and first_device_address on the 'safe' position.
4267 * If this 'next' write position is after the 'safe' position,
4268 * we must update the metadata to increase the 'safe' position.
4269 *
4270 * When reshaping backwards, we round in the opposite direction
4271 * and perform the reverse test: next write position must not be
4272 * less than current safe position.
4273 *
4274 * In all this the minimum difference in data offsets
4275 * (conf->offset_diff - always positive) allows a bit of slack,
4276 * so next can be after 'safe', but not by more than offset_diff
4277 *
4278 * We need to prepare all the bios here before we start any IO
4279 * to ensure the size we choose is acceptable to all devices.
4280 * The means one for each copy for write-out and an extra one for
4281 * read-in.
4282 * We store the read-in bio in ->master_bio and the others in
4283 * ->devs[x].bio and ->devs[x].repl_bio.
4284 */
4299 /* If restarting in the middle, skip the initial sectors */
4312 }
4313 }
4315 /* We don't use sector_nr to track where we are up to
4316 * as that doesn't work well for ->reshape_backwards.
4317 * So just use ->reshape_progress.
4318 */
4320 /* 'next' is the earliest device address that we might
4321 * write to for this chunk in the new layout
4322 */
4326 /* 'safe' is the last device address that we might read from
4327 * in the old layout after a restart
4328 */
4341 /* 'next' is after the last device address that we
4342 * might write to for this chunk in the new layout
4343 */
4346 /* 'safe' is the earliest device address that we might
4347 * read from in the old layout after a restart
4348 */
4351 /* Need to update metadata if 'next' might be beyond 'safe'
4352 * as that would possibly corrupt data
4353 */
4363 }
4367 /* Need to update reshape_position in metadata */
4373 else
4383 }
4386 }
4388 read_more:
4389 /* Now schedule reads for blocks from sector_nr to last */
4402 /* Cannot read from here, so need to record bad blocks
4403 * on all the target devices.
4404 */
4405 // FIXME
4409 }
4426 /* Now find the locations in the new layout */
4443 }
4454 }
4456 /* Now add as many pages as possible to all of these bios. */
4466 /*
4467 * won't fail because the vec table is big enough
4468 * to hold all these pages
4469 */
4471 }
4474 }
4478 /* Now submit the read */
4488 /* Now that we have done the whole section we can
4489 * update reshape_progress
4490 */
4493 else
4497 }
4503 {
4504 /* Reshape read completed. Hopefully we have a block
4505 * to write out.
4506 * If we got a read error then we do sync 1-page reads from
4507 * elsewhere until we find the data - or give up.
4508 */
4514 /* Reshape has been aborted */
4517 }
4519 /* We definitely have the data in the pages, schedule the
4520 * writes.
4521 */
4534 }
4538 }
4545 }
4547 }
4550 {
4562 /* read-ahead size must cover two whole stripes, which is
4563 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4564 */
4571 }
4573 }
4577 {
4578 /* Use sync reads to get the blocks from somewhere else */
4590 /* reshape IOs share pages from .devs[0].bio */
4608 sector_t addr;
4618 addr,
4626 failed:
4627 slot++;
4632 }
4635 /* couldn't read this block, must give up */
4639 }
4641 idx++;
4642 }
4644 }
4647 {
4662 }
4665 /* FIXME should record badblock */
4667 }
4671 }
4674 {
4680 }
4683 {
4695 }
4700 }
4706 d++) {
4713 }
4715 }
4721 }
4724 {
4745 };
4748 {
4750 }
4753 {
4755 }