| blob | 063c43d83b72c2f0f753edb7b08f8dd608fa15ad |
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...) \
114 {
118 /* allocate a r10bio with room for raid_disks entries in the
119 * bios array */
121 }
124 {
126 }
128 /* Maximum size of each resync request */
129 #define RESYNC_BLOCK_SIZE (64*1024)
130 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
131 /* amount of memory to reserve for resync requests */
132 #define RESYNC_WINDOW (1024*1024)
133 /* maximum number of concurrent requests, memory permitting */
134 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
136 /*
137 * When performing a resync, we need to read and compare, so
138 * we need as many pages are there are copies.
139 * When performing a recovery, we need 2 bios, one for read,
140 * one for write (we recover only one drive per r10buf)
141 *
142 */
144 {
159 else
162 /*
163 * Allocate bios.
164 */
176 }
177 /*
178 * Allocate RESYNC_PAGES data pages and attach them
179 * where needed.
180 */
187 /* we can share bv_page's during recovery
188 * and reshape */
200 }
201 }
205 out_free_pages:
212 out_free_bio:
218 }
221 }
224 {
236 }
238 }
242 }
244 }
247 {
259 }
260 }
263 {
268 }
271 {
277 }
280 {
290 /* wake up frozen array... */
294 }
296 /*
297 * raid_end_bio_io() is called when we have finished servicing a mirrored
298 * operation and are ready to return a success/failure code to the buffer
299 * cache layer.
300 */
302 {
319 /*
320 * Wake up any possible resync thread that waits for the device
321 * to go idle.
322 */
324 }
326 }
328 /*
329 * Update disk head position estimator based on IRQ completion info.
330 */
332 {
337 }
339 /*
340 * Find the disk number which triggered given bio
341 */
344 {
354 }
355 }
365 }
368 {
378 /*
379 * this branch is our 'one mirror IO has finished' event handler:
380 */
384 /*
385 * Set R10BIO_Uptodate in our master bio, so that
386 * we will return a good error code to the higher
387 * levels even if IO on some other mirrored buffer fails.
388 *
389 * The 'master' represents the composite IO operation to
390 * user-side. So if something waits for IO, then it will
391 * wait for the 'master' bio.
392 */
395 /* If all other devices that store this block have
396 * failed, we want to return the error upwards rather
397 * than fail the last device. Here we redefine
398 * "uptodate" to mean "Don't want to retry"
399 */
403 }
408 /*
409 * oops, read error - keep the refcount on the rdev
410 */
418 }
419 }
422 {
423 /* clear the bitmap if all writes complete successfully */
429 }
432 {
440 else
442 }
443 }
444 }
447 {
467 }
468 /*
469 * this branch is our 'one mirror IO has finished' event handler:
470 */
473 /* Never record new bad blocks to replacement,
474 * just fail it.
475 */
488 /* This is the only remaining device,
489 * We need to retry the write without
490 * FailFast
491 */
497 }
500 }
502 /*
503 * Set R10BIO_Uptodate in our master bio, so that
504 * we will return a good error code for to the higher
505 * levels even if IO on some other mirrored buffer fails.
506 *
507 * The 'master' represents the composite IO operation to
508 * user-side. So if something waits for IO, then it will
509 * wait for the 'master' bio.
510 */
511 sector_t first_bad;
514 /*
515 * Do not set R10BIO_Uptodate if the current device is
516 * rebuilding or Faulty. This is because we cannot use
517 * such device for properly reading the data back (we could
518 * potentially use it, if the current write would have felt
519 * before rdev->recovery_offset, but for simplicity we don't
520 * check this here.
521 */
526 /* Maybe we can clear some bad blocks. */
534 else
538 }
539 }
541 /*
542 *
543 * Let's see if all mirrored write operations have finished
544 * already.
545 */
551 }
553 /*
554 * RAID10 layout manager
555 * As well as the chunksize and raid_disks count, there are two
556 * parameters: near_copies and far_copies.
557 * near_copies * far_copies must be <= raid_disks.
558 * Normally one of these will be 1.
559 * If both are 1, we get raid0.
560 * If near_copies == raid_disks, we get raid1.
561 *
562 * Chunks are laid out in raid0 style with near_copies copies of the
563 * first chunk, followed by near_copies copies of the next chunk and
564 * so on.
565 * If far_copies > 1, then after 1/far_copies of the array has been assigned
566 * as described above, we start again with a device offset of near_copies.
567 * So we effectively have another copy of the whole array further down all
568 * the drives, but with blocks on different drives.
569 * With this layout, and block is never stored twice on the one device.
570 *
571 * raid10_find_phys finds the sector offset of a given virtual sector
572 * on each device that it is on.
573 *
574 * raid10_find_virt does the reverse mapping, from a device and a
575 * sector offset to a virtual address
576 */
579 {
581 sector_t sector;
582 sector_t chunk;
583 sector_t stripe;
594 /* now calculate first sector/dev */
606 /* and calculate all the others */
613 slot++;
627 }
631 slot++;
632 }
633 dev++;
637 }
638 }
639 }
642 {
654 }
657 {
659 /* Never use conf->prev as this is only called during resync
660 * or recovery, so reshape isn't happening
661 */
675 }
676 }
691 else
693 }
695 }
699 }
701 /*
702 * This routine returns the disk from which the requested read should
703 * be done. There is a per-array 'next expected sequential IO' sector
704 * number - if this matches on the next IO then we use the last disk.
705 * There is also a per-disk 'last know head position' sector that is
706 * maintained from IRQ contexts, both the normal and the resync IO
707 * completion handlers update this position correctly. If there is no
708 * perfect sequential match then we pick the disk whose head is closest.
709 *
710 * If there are 2 mirrors in the same 2 devices, performance degrades
711 * because position is mirror, not device based.
712 *
713 * The rdev for the device selected will have nr_pending incremented.
714 */
716 /*
717 * FIXME: possibly should rethink readbalancing and do it differently
718 * depending on near_copies / far_copies geometry.
719 */
723 {
743 /*
744 * Check if we can balance. We can balance on the whole
745 * device if no resync is going on (recovery is ok), or below
746 * the resync window. We take the first readable disk when
747 * above the resync window.
748 */
754 sector_t first_bad;
756 sector_t dev_sector;
776 /* Already have a better slot */
779 /* Cannot read here. If this is the
780 * 'primary' device, then we must not read
781 * beyond 'bad_sectors' from another device.
782 */
789 sector_t good_sectors =
795 }
797 /* Must read from here */
799 }
808 /* At least 2 disks to choose from so failfast is OK */
810 /* This optimisation is debatable, and completely destroys
811 * sequential read speed for 'far copies' arrays. So only
812 * keep it for 'near' arrays, and review those later.
813 */
817 /* for far > 1 always use the lowest address */
820 else
827 }
828 }
832 }
843 }
846 {
858 i++) {
864 }
865 }
868 }
871 {
872 /* Any writes that have been queued but are awaiting
873 * bitmap updates get flushed here.
874 */
882 /* flush any pending bitmap writes to disk
883 * before proceeding w/ I/O */
897 /* Just ignore it */
899 else
902 }
905 }
907 /* Barriers....
908 * Sometimes we need to suspend IO while we do something else,
909 * either some resync/recovery, or reconfigure the array.
910 * To do this we raise a 'barrier'.
911 * The 'barrier' is a counter that can be raised multiple times
912 * to count how many activities are happening which preclude
913 * normal IO.
914 * We can only raise the barrier if there is no pending IO.
915 * i.e. if nr_pending == 0.
916 * We choose only to raise the barrier if no-one is waiting for the
917 * barrier to go down. This means that as soon as an IO request
918 * is ready, no other operations which require a barrier will start
919 * until the IO request has had a chance.
920 *
921 * So: regular IO calls 'wait_barrier'. When that returns there
922 * is no backgroup IO happening, It must arrange to call
923 * allow_barrier when it has finished its IO.
924 * backgroup IO calls must call raise_barrier. Once that returns
925 * there is no normal IO happeing. It must arrange to call
926 * lower_barrier when the particular background IO completes.
927 */
930 {
934 /* Wait until no block IO is waiting (unless 'force') */
938 /* block any new IO from starting */
941 /* Now wait for all pending IO to complete */
947 }
950 {
956 }
959 {
963 /* Wait for the barrier to drop.
964 * However if there are already pending
965 * requests (preventing the barrier from
966 * rising completely), and the
967 * pre-process bio queue isn't empty,
968 * then don't wait, as we need to empty
969 * that queue to get the nr_pending
970 * count down.
971 */
982 }
985 }
988 {
992 }
995 {
996 /* stop syncio and normal IO and wait for everything to
997 * go quiet.
998 * We increment barrier and nr_waiting, and then
999 * wait until nr_pending match nr_queued+extra
1000 * This is called in the context of one normal IO request
1001 * that has failed. Thus any sync request that might be pending
1002 * will be blocked by nr_pending, and we need to wait for
1003 * pending IO requests to complete or be queued for re-try.
1004 * Thus the number queued (nr_queued) plus this request (extra)
1005 * must match the number of pending IOs (nr_pending) before
1006 * we continue.
1007 */
1019 }
1022 {
1023 /* reverse the effect of the freeze */
1029 }
1033 {
1037 else
1039 }
1045 };
1048 {
1050 cb);
1064 }
1066 /* we aren't scheduling, so we can do the write-out directly. */
1081 /* Just ignore it */
1083 else
1086 }
1088 }
1092 {
1099 sector_t sectors;
1103 /*
1104 * Register the new request and wait if the reconstruction
1105 * thread has put up a bar for new requests.
1106 * Continue immediately if no resync is active currently.
1107 */
1114 /*
1115 * IO spans the reshape position. Need to wait for reshape to
1116 * pass
1117 */
1123 sectors);
1125 }
1127 read_again:
1132 }
1137 max_sectors);
1157 /*
1158 * Could not read all from this device, so we will need another
1159 * r10_bio.
1160 */
1167 else
1170 /*
1171 * Cannot call generic_make_request directly as that will be
1172 * queued in __generic_make_request and subsequent
1173 * mempool_alloc might block waiting for it. so hand bio over
1174 * to raid10d.
1175 */
1189 }
1193 {
1203 sector_t sectors;
1209 /*
1210 * Register the new request and wait if the reconstruction
1211 * thread has put up a bar for new requests.
1212 * Continue immediately if no resync is active currently.
1213 */
1220 /*
1221 * IO spans the reshape position. Need to wait for reshape to
1222 * pass
1223 */
1229 sectors);
1231 }
1239 /* Need to update reshape_position in metadata */
1249 }
1256 }
1257 /* first select target devices under rcu_lock and
1258 * inc refcount on their rdev. Record them by setting
1259 * bios[x] to bio
1260 * If there are known/acknowledged bad blocks on any device
1261 * on which we have seen a write error, we want to avoid
1262 * writing to those blocks. This potentially requires several
1263 * writes to write around the bad blocks. Each set of writes
1264 * gets its own r10_bio with a set of bios attached. The number
1265 * of r10_bios is recored in bio->bi_phys_segments just as with
1266 * the read case.
1267 */
1271 retry_write:
1287 }
1292 }
1304 }
1306 sector_t first_bad;
1314 /* Mustn't write here until the bad block
1315 * is acknowledged
1316 */
1321 }
1323 /* Cannot write here at all */
1326 /* Mustn't write more than bad_sectors
1327 * to other devices yet
1328 */
1330 /* We don't set R10BIO_Degraded as that
1331 * only applies if the disk is missing,
1332 * so it might be re-added, and we want to
1333 * know to recover this chunk.
1334 * In this case the device is here, and the
1335 * fact that this chunk is not in-sync is
1336 * recorded in the bad block log.
1337 */
1339 }
1344 }
1345 }
1349 }
1353 }
1354 }
1358 /* Have to wait for this device to get unblocked, then retry */
1366 }
1372 /* Race with remove_disk */
1375 }
1377 }
1378 }
1384 }
1387 /* We are splitting this into multiple parts, so
1388 * we need to prepare for allocating another r10_bio.
1389 */
1394 else
1397 }
1411 max_sectors);
1428 /* flush_pending_writes() needs access to the rdev so...*/
1437 cb);
1438 else
1447 }
1451 }
1456 /* Replacement just got moved to main 'rdev' */
1459 }
1462 max_sectors);
1476 /* flush_pending_writes() needs access to the rdev so...*/
1486 }
1487 }
1489 /* Don't remove the bias on 'remaining' (one_write_done) until
1490 * after checking if we need to go around again.
1491 */
1495 /* We need another r10_bio. It has already been counted
1496 * in bio->bi_phys_segments.
1497 */
1507 }
1509 }
1512 {
1525 /*
1526 * We might need to issue multiple reads to different devices if there
1527 * are bad blocks around, so we keep track of the number of reads in
1528 * bio->bi_phys_segments. If this is 0, there is only one r10_bio and
1529 * no locking will be needed when the request completes. If it is
1530 * non-zero, then it is the number of not-completed requests.
1531 */
1537 else
1539 }
1542 {
1552 }
1556 /*
1557 * If this request crosses a chunk boundary, we need to split
1558 * it.
1559 */
1572 }
1577 /* In case raid10d snuck in to freeze_array */
1579 }
1582 {
1593 else
1597 }
1604 }
1607 }
1609 /* check if there are enough drives for
1610 * every block to appear on atleast one.
1611 * Don't consider the device numbered 'ignore'
1612 * as we might be about to remove it.
1613 */
1615 {
1625 }
1637 cnt++;
1639 }
1645 out:
1648 }
1651 {
1652 /* when calling 'enough', both 'prev' and 'geo' must
1653 * be stable.
1654 * This is ensured if ->reconfig_mutex or ->device_lock
1655 * is held.
1656 */
1659 }
1662 {
1667 /*
1668 * If it is not operational, then we have already marked it as dead
1669 * else if it is the last working disks, ignore the error, let the
1670 * next level up know.
1671 * else mark the drive as failed
1672 */
1676 /*
1677 * Don't fail the drive, just return an IO error.
1678 */
1681 }
1684 /*
1685 * If recovery is running, make sure it aborts.
1686 */
1697 }
1700 {
1708 }
1712 /* This is only called with ->reconfix_mutex held, so
1713 * rcu protection of rdev is not needed */
1722 }
1723 }
1726 {
1732 }
1735 {
1742 /*
1743 * Find all non-in_sync disks within the RAID10 configuration
1744 * and mark them in_sync
1745 */
1752 /* Replacement has just become active */
1755 count++;
1757 /* Replaced device not technically faulty,
1758 * but we need to be sure it gets removed
1759 * and never re-added.
1760 */
1764 }
1770 count++;
1772 }
1773 }
1780 }
1783 {
1791 /* only hot-add to in-sync arrays, as recovery is
1792 * very different from resync
1793 */
1807 else
1827 }
1841 }
1847 }
1850 {
1862 else
1869 }
1870 /* Only remove non-faulty devices if recovery
1871 * is not possible.
1872 */
1880 }
1885 /* lost the race, try later */
1889 }
1890 }
1892 /* We must have just cleared 'rdev' */
1896 * but will never see neither -- if they are careful.
1897 */
1901 /* We might have just remove the Replacement as faulty
1902 * Clear the flag just in case
1903 */
1908 abort:
1912 }
1915 {
1921 /* this is a reshape read */
1928 else
1929 /* The write handler will notice the lack of
1930 * R10BIO_Uptodate and record any errors etc
1931 */
1935 /* for reconstruct, we always reschedule after a read.
1936 * for resync, only after all reads
1937 */
1941 /* we have read all the blocks,
1942 * do the comparison in process context in raid10d
1943 */
1945 }
1946 }
1949 {
1954 /* the primary of several recovery bios */
1959 else
1968 else
1971 }
1972 }
1973 }
1976 {
1981 sector_t first_bad;
1990 else
2002 }
2012 }
2014 /*
2015 * Note: sync and recover and handled very differently for raid10
2016 * This code is for resync.
2017 * For resync, we read through virtual addresses and read all blocks.
2018 * If there is any error, we schedule a write. The lowest numbered
2019 * drive is authoritative.
2020 * However requests come for physical address, so we need to map.
2021 * For every physical address there are raid_disks/copies virtual addresses,
2022 * which is always are least one, but is not necessarly an integer.
2023 * This means that a physical address can span multiple chunks, so we may
2024 * have to submit multiple io requests for a single sync request.
2025 */
2026 /*
2027 * We check if all blocks are in-sync and only write to blocks that
2028 * aren't in sync
2029 */
2031 {
2039 /* find the first device with a block */
2053 /* now find blocks with errors */
2067 /* We know that the bi_io_vec layout is the same for
2068 * both 'first' and 'i', so we just compare them.
2069 * All vec entries are PAGE_SIZE;
2070 */
2078 len))
2081 }
2086 /* Don't fix anything. */
2089 /* Just give up on this device */
2092 }
2093 /* Ok, we need to write this bio, either to correct an
2094 * inconsistency or to correct an unreadable block.
2095 * First we need to fixup bv_offset, bv_len and
2096 * bi_vecs, as the read request might have corrupted these
2097 */
2118 }
2120 /* Now write out to any replacement devices
2121 * that are active
2122 */
2137 }
2139 done:
2143 }
2144 }
2146 /*
2147 * Now for the recovery code.
2148 * Recovery happens across physical sectors.
2149 * We recover all non-is_sync drives by finding the virtual address of
2150 * each, and then choose a working drive that also has that virt address.
2151 * There is a separate r10_bio for each non-in_sync drive.
2152 * Only the first two slots are in use. The first for reading,
2153 * The second for writing.
2154 *
2155 */
2157 {
2158 /* We got a read error during recovery.
2159 * We repeat the read in smaller page-sized sections.
2160 * If a read succeeds, write it to the new device or record
2161 * a bad block if we cannot.
2162 * If a read fails, record a bad block on both old and
2163 * new devices.
2164 */
2177 sector_t addr;
2186 addr,
2194 addr,
2204 }
2205 }
2207 /* We don't worry if we cannot set a bad block -
2208 * it really is bad so there is no loss in not
2209 * recording it yet
2210 */
2214 /* need bad block on destination too */
2219 /* just abort the recovery */
2228 }
2229 }
2230 }
2234 idx++;
2235 }
2236 }
2239 {
2248 }
2250 /*
2251 * share the pages with the first bio
2252 * and submit the write request
2253 */
2257 /* Need to test wbio2->bi_end_io before we call
2258 * generic_make_request as if the former is NULL,
2259 * the latter is free to free wbio2.
2260 */
2267 }
2273 }
2274 }
2276 /*
2277 * Used by fix_read_error() to decay the per rdev read_errors.
2278 * We halve the read error count for every hour that has elapsed
2279 * since the last recorded read error.
2280 *
2281 */
2283 {
2291 /* first time we've seen a read error */
2294 }
2301 /*
2302 * if hours_since_last is > the number of bits in read_errors
2303 * just set read errors to 0. We do this to avoid
2304 * overflowing the shift of read_errors by hours_since_last.
2305 */
2308 else
2310 }
2314 {
2315 sector_t first_bad;
2322 /* success */
2329 }
2330 /* need to record an error - either for the block or the device */
2334 }
2336 /*
2337 * This is a kernel thread which:
2338 *
2339 * 1. Retries failed read operations on working mirrors.
2340 * 2. Updates the raid superblock when problems encounter.
2341 * 3. Performs writes following reads for array synchronising.
2342 */
2345 {
2352 /* still own a reference to this rdev, so it cannot
2353 * have been cleared recently.
2354 */
2358 /* drive has already been failed, just ignore any
2359 more fix_read_error() attempts */
2376 }
2389 sector_t first_bad;
2403 sect,
2411 }
2412 sl++;
2419 /* Cannot read from anywhere, just mark the block
2420 * as bad on the first device to discourage future
2421 * reads.
2422 */
2427 rdev,
2434 }
2436 }
2439 /* write it back and re-read */
2446 sl--;
2458 sect,
2461 /* Well, this device is dead */
2465 sect +
2467 rdev)),
2472 }
2475 }
2482 sl--;
2494 sect,
2496 READ)) {
2498 /* Well, this device is dead */
2502 sect +
2513 sect +
2517 }
2521 }
2526 }
2527 }
2530 {
2535 /* bio has the data to be written to slot 'i' where
2536 * we just recently had a write error.
2537 * We repeatedly clone the bio and trim down to one block,
2538 * then try the write. Where the write fails we record
2539 * a bad block.
2540 * It is conceivable that the bio doesn't exactly align with
2541 * blocks. We must handle this.
2542 *
2543 * We currently own a reference to the rdev.
2544 */
2547 sector_t sector;
2564 sector_t wsector;
2567 /* Write at 'sector' for 'sectors' */
2577 /* Failure! */
2586 }
2588 }
2591 {
2599 dev_t bio_dev;
2600 sector_t bio_last_sector;
2602 /* we got a read error. Maybe the drive is bad. Maybe just
2603 * the block and we can fix it.
2604 * We freeze all other IO, and try reading the block from
2605 * other devices. When we find one, we re-write
2606 * and check it that fixes the read error.
2607 * This is all done synchronously while the array is
2608 * frozen.
2609 */
2628 read_more:
2636 }
2662 /* Drat - have to split this up more */
2671 else
2677 GFP_NOIO);
2690 }
2693 {
2694 /* Some sort of write request has finished and it
2695 * succeeded in writing where we thought there was a
2696 * bad block. So forget the bad block.
2697 * Or possibly if failed and we need to record
2698 * a bad block.
2699 */
2712 rdev,
2717 rdev,
2721 }
2728 rdev,
2733 rdev,
2737 }
2738 }
2748 rdev,
2758 }
2760 }
2765 rdev,
2769 }
2770 }
2782 }
2783 }
2784 }
2787 {
2805 }
2806 }
2810 retry_list);
2819 }
2820 }
2831 }
2851 /* just a partial read to be scheduled from a
2852 * separate context
2853 */
2856 }
2861 }
2863 }
2866 {
2881 }
2883 /*
2884 * perform a "sync" on one "block"
2885 *
2886 * We need to make sure that no normal I/O request - particularly write
2887 * requests - conflict with active sync requests.
2888 *
2889 * This is achieved by tracking pending requests and a 'barrier' concept
2890 * that can be installed to exclude normal IO requests.
2891 *
2892 * Resync and recovery are handled very differently.
2893 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2894 *
2895 * For resync, we iterate over virtual addresses, read all copies,
2896 * and update if there are differences. If only one copy is live,
2897 * skip it.
2898 * For recovery, we iterate over physical addresses, read a good
2899 * value for each non-in_sync drive, and over-write.
2900 *
2901 * So, for recovery we may have several outstanding complex requests for a
2902 * given address, one for each out-of-sync device. We model this by allocating
2903 * a number of r10_bio structures, one for each out-of-sync device.
2904 * As we setup these structures, we collect all bio's together into a list
2905 * which we then process collectively to add pages, and then process again
2906 * to pass to generic_make_request.
2907 *
2908 * The r10_bio structures are linked using a borrowed master_bio pointer.
2909 * This link is counted in ->remaining. When the r10_bio that points to NULL
2910 * has its remaining count decremented to 0, the whole complex operation
2911 * is complete.
2912 *
2913 */
2917 {
2924 sector_t sync_blocks;
2933 /*
2934 * Allow skipping a full rebuild for incremental assembly
2935 * of a clean array, like RAID1 does.
2936 */
2946 }
2948 skipped:
2954 /* If we aborted, we need to abort the
2955 * sync on the 'current' bitmap chucks (there can
2956 * be several when recovering multiple devices).
2957 * as we may have started syncing it but not finished.
2958 * We can find the current address in
2959 * mddev->curr_resync, but for recovery,
2960 * we need to convert that to several
2961 * virtual addresses.
2962 */
2967 }
2974 sector_t sect =
2978 }
2980 /* completed sync */
2984 /* Completed a full sync so the replacements
2985 * are now fully recovered.
2986 */
2993 }
2995 }
2997 }
3002 }
3008 /* if there has been nothing to do on any drive,
3009 * then there is nothing to do at all..
3010 */
3013 }
3018 /* make sure whole request will fit in a chunk - if chunks
3019 * are meaningful
3020 */
3025 /*
3026 * If there is non-resync activity waiting for a turn, then let it
3027 * though before starting on this new sync request.
3028 */
3032 /* Again, very different code for resync and recovery.
3033 * Both must result in an r10bio with a list of bios that
3034 * have bi_end_io, bi_sector, bi_bdev set,
3035 * and bi_private set to the r10bio.
3036 * For recovery, we may actually create several r10bios
3037 * with 2 bios in each, that correspond to the bios in the main one.
3038 * In this case, the subordinate r10bios link back through a
3039 * borrowed master_bio pointer, and the counter in the master
3040 * includes a ref from each subordinate.
3041 */
3042 /* First, we decide what to do and set ->bi_end_io
3043 * To end_sync_read if we want to read, and
3044 * end_sync_write if we will want to write.
3045 */
3049 /* recovery... the complicated one */
3056 sector_t sect;
3073 }
3076 /* want to reconstruct this device */
3080 /* last stripe is not complete - don't
3081 * try to recover this sector.
3082 */
3085 }
3088 /* Unless we are doing a full sync, or a replacement
3089 * we only need to recover the block if it is set in
3090 * the bitmap
3091 */
3099 /* yep, skip the sync_blocks here, but don't assume
3100 * that there will never be anything to do here
3101 */
3105 }
3125 /* Need to check if the array will still be
3126 * degraded
3127 */
3135 }
3136 }
3153 /* This is where we read from */
3162 bad_sectors -= (sector
3167 }
3168 }
3183 /* and we write to 'i' (if not in_sync) */
3210 /* and maybe write to replacement */
3214 /* Note: if mreplace != NULL, then bio
3215 * cannot be NULL as r10buf_pool_alloc will
3216 * have allocated it.
3217 * So the second test here is pointless.
3218 * But it keeps semantic-checkers happy, and
3219 * this comment keeps human reviewers
3220 * happy.
3221 */
3236 }
3239 /* Cannot recover, so abort the recovery or
3240 * record a bad block */
3242 /* problem is that there are bad blocks
3243 * on other device(s)
3244 */
3252 mrdev,
3258 mreplace,
3262 }
3268 mirror->recovery_disabled
3270 }
3279 }
3284 /* Only want this if there is elsewhere to
3285 * read from. 'j' is currently the first
3286 * readable copy.
3287 */
3294 targets++;
3295 }
3299 }
3300 }
3307 }
3309 }
3311 /* resync. Schedule a read for every block at this virt offset */
3320 /* We can skip this block */
3323 }
3357 }
3369 }
3370 }
3382 count++;
3388 }
3392 /* Need to set up for writing to the replacement */
3407 count++;
3408 }
3415 mddev);
3420 mddev);
3421 }
3425 }
3426 }
3444 /* stop here */
3449 /* remove last page from this bio */
3453 }
3455 }
3459 bio_full:
3474 }
3475 }
3478 /* pretend they weren't skipped, it makes
3479 * no important difference in this case
3480 */
3484 giveup:
3485 /* There is nowhere to write, so all non-sync
3486 * drives must be failed or in resync, all drives
3487 * have a bad block, so try the next chunk...
3488 */
3493 chunks_skipped ++;
3496 }
3498 static sector_t
3500 {
3501 sector_t size;
3516 }
3519 {
3520 /* Calculate the number of sectors-per-device that will
3521 * actually be used, and set conf->dev_sectors and
3522 * conf->stride
3523 */
3529 /* 'size' is now the number of chunks in the array */
3530 /* calculate "used chunks per device" */
3533 /* We need to round up when dividing by raid_disks to
3534 * get the stride size.
3535 */
3545 }
3546 }
3550 {
3566 * updated yet. */
3571 }
3589 * actually using this, but leave code here just in case.*/
3599 }
3603 }
3606 {
3618 }
3624 }
3631 /* FIXME calc properly */
3634 GFP_KERNEL);
3657 }
3661 else
3662 /* far_copies must be 1 */
3664 }
3681 out:
3687 }
3689 }
3692 {
3697 sector_t size;
3707 }
3723 else
3726 }
3748 }
3766 }
3772 else
3775 }
3776 /* need to check that every block has at least one working mirror */
3781 }
3784 /* must ensure that shape change is supported */
3791 }
3797 i++) {
3802 /* The replacement is all we have - use it */
3806 }
3815 }
3817 }
3825 /*
3826 * Ok, everything is just fine now
3827 */
3838 /* Calculate max read-ahead size.
3839 * We need to readahead at least twice a whole stripe....
3840 * maybe...
3841 */
3845 }
3859 /* This cannot work */
3862 }
3871 }
3875 out_free_conf:
3882 out:
3884 }
3887 {
3896 }
3899 {
3909 }
3910 }
3913 {
3914 /* Resize of 'far' arrays is not supported.
3915 * For 'near' and 'offset' arrays we can set the
3916 * number of sectors used to be an appropriate multiple
3917 * of the chunk size.
3918 * For 'offset', this is far_copies*chunksize.
3919 * For 'near' the multiplier is the LCM of
3920 * near_copies and raid_disks.
3921 * So if far_copies > 1 && !far_offset, fail.
3922 * Else find LCM(raid_disks, near_copy)*far_copies and
3923 * multiply by chunk_size. Then round to this number.
3924 * This is mostly done by raid10_size()
3925 */
3944 }
3949 }
3954 }
3959 }
3962 {
3970 }
3973 /* Set new parameters */
3975 /* new layout: far_copies = 1, near_copies = 2 */
3980 /* make sure it will be not marked as dirty */
3990 }
3992 }
3995 }
3998 {
4001 /* raid10 can take over:
4002 * raid0 - providing it has only two drives
4003 */
4005 /* for raid0 takeover only one zone is supported */
4011 }
4015 }
4017 }
4020 {
4021 /* Called when there is a request to change
4022 * - layout (to ->new_layout)
4023 * - chunk size (to ->new_chunk_sectors)
4024 * - raid_disks (by delta_disks)
4025 * or when trying to restart a reshape that was ongoing.
4026 *
4027 * We need to validate the request and possibly allocate
4028 * space if that might be an issue later.
4029 *
4030 * Currently we reject any reshape of a 'far' mode array,
4031 * allow chunk size to change if new is generally acceptable,
4032 * allow raid_disks to increase, and allow
4033 * a switch between 'near' mode and 'offset' mode.
4034 */
4042 /* mustn't change number of copies */
4045 /* Cannot switch to 'far' mode */
4049 /* not factor of array size */
4058 /* allocate new 'mirrors' list */
4063 GFP_KERNEL);
4066 }
4068 }
4070 /*
4071 * Need to check if array has failed when deciding whether to:
4072 * - start an array
4073 * - remove non-faulty devices
4074 * - add a spare
4075 * - allow a reshape
4076 * This determination is simple when no reshape is happening.
4077 * However if there is a reshape, we need to carefully check
4078 * both the before and after sections.
4079 * This is because some failed devices may only affect one
4080 * of the two sections, and some non-in_sync devices may
4081 * be insync in the section most affected by failed devices.
4082 */
4084 {
4090 /* 'prev' section first */
4094 degraded++;
4096 /* When we can reduce the number of devices in
4097 * an array, this might not contribute to
4098 * 'degraded'. It does now.
4099 */
4100 degraded++;
4101 }
4110 degraded2++;
4112 /* If reshape is increasing the number of devices,
4113 * this section has already been recovered, so
4114 * it doesn't contribute to degraded.
4115 * else it does.
4116 */
4118 degraded2++;
4119 }
4120 }
4125 }
4128 {
4129 /* A 'reshape' has been requested. This commits
4130 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4131 * This also checks if there are enough spares and adds them
4132 * to the array.
4133 * We currently require enough spares to make the final
4134 * array non-degraded. We also require that the difference
4135 * between old and new data_offset - on each device - is
4136 * enough that we never risk over-writing.
4137 */
4162 spares++;
4172 }
4173 }
4191 }
4201 }
4216 }
4225 else
4230 }
4233 /* This is a spare that was manually added */
4235 }
4236 }
4237 /* When a reshape changes the number of devices,
4238 * ->degraded is measured against the larger of the
4239 * pre and post numbers.
4240 */
4259 }
4265 abort:
4278 }
4280 /* Calculate the last device-address that could contain
4281 * any block from the chunk that includes the array-address 's'
4282 * and report the next address.
4283 * i.e. the address returned will be chunk-aligned and after
4284 * any data that is in the chunk containing 's'.
4285 */
4287 {
4295 }
4297 /* Calculate the first device-address that could contain
4298 * any block from the chunk that includes the array-address 's'.
4299 * This too will be the start of a chunk
4300 */
4302 {
4309 }
4313 {
4314 /* We simply copy at most one chunk (smallest of old and new)
4315 * at a time, possibly less if that exceeds RESYNC_PAGES,
4316 * or we hit a bad block or something.
4317 * This might mean we pause for normal IO in the middle of
4318 * a chunk, but that is not a problem as mddev->reshape_position
4319 * can record any location.
4320 *
4321 * If we will want to write to a location that isn't
4322 * yet recorded as 'safe' (i.e. in metadata on disk) then
4323 * we need to flush all reshape requests and update the metadata.
4324 *
4325 * When reshaping forwards (e.g. to more devices), we interpret
4326 * 'safe' as the earliest block which might not have been copied
4327 * down yet. We divide this by previous stripe size and multiply
4328 * by previous stripe length to get lowest device offset that we
4329 * cannot write to yet.
4330 * We interpret 'sector_nr' as an address that we want to write to.
4331 * From this we use last_device_address() to find where we might
4332 * write to, and first_device_address on the 'safe' position.
4333 * If this 'next' write position is after the 'safe' position,
4334 * we must update the metadata to increase the 'safe' position.
4335 *
4336 * When reshaping backwards, we round in the opposite direction
4337 * and perform the reverse test: next write position must not be
4338 * less than current safe position.
4339 *
4340 * In all this the minimum difference in data offsets
4341 * (conf->offset_diff - always positive) allows a bit of slack,
4342 * so next can be after 'safe', but not by more than offset_diff
4343 *
4344 * We need to prepare all the bios here before we start any IO
4345 * to ensure the size we choose is acceptable to all devices.
4346 * The means one for each copy for write-out and an extra one for
4347 * read-in.
4348 * We store the read-in bio in ->master_bio and the others in
4349 * ->devs[x].bio and ->devs[x].repl_bio.
4350 */
4364 /* If restarting in the middle, skip the initial sectors */
4377 }
4378 }
4380 /* We don't use sector_nr to track where we are up to
4381 * as that doesn't work well for ->reshape_backwards.
4382 * So just use ->reshape_progress.
4383 */
4385 /* 'next' is the earliest device address that we might
4386 * write to for this chunk in the new layout
4387 */
4391 /* 'safe' is the last device address that we might read from
4392 * in the old layout after a restart
4393 */
4406 /* 'next' is after the last device address that we
4407 * might write to for this chunk in the new layout
4408 */
4411 /* 'safe' is the earliest device address that we might
4412 * read from in the old layout after a restart
4413 */
4416 /* Need to update metadata if 'next' might be beyond 'safe'
4417 * as that would possibly corrupt data
4418 */
4428 }
4432 /* Need to update reshape_position in metadata */
4438 else
4448 }
4451 }
4453 read_more:
4454 /* Now schedule reads for blocks from sector_nr to last */
4467 /* Cannot read from here, so need to record bad blocks
4468 * on all the target devices.
4469 */
4470 // FIXME
4474 }
4491 /* Now find the locations in the new layout */
4508 }
4521 }
4523 /* Now add as many pages as possible to all of these bios. */
4536 /* Didn't fit, must stop */
4540 /* Remove last page from this bio */
4544 }
4546 }
4549 }
4550 bio_full:
4554 /* Now submit the read */
4564 /* Now that we have done the whole section we can
4565 * update reshape_progress
4566 */
4569 else
4573 }
4579 {
4580 /* Reshape read completed. Hopefully we have a block
4581 * to write out.
4582 * If we got a read error then we do sync 1-page reads from
4583 * elsewhere until we find the data - or give up.
4584 */
4590 /* Reshape has been aborted */
4593 }
4595 /* We definitely have the data in the pages, schedule the
4596 * writes.
4597 */
4610 }
4614 }
4621 }
4623 }
4626 {
4638 /* read-ahead size must cover two whole stripes, which is
4639 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4640 */
4647 }
4649 }
4653 {
4654 /* Use sync reads to get the blocks from somewhere else */
4681 sector_t addr;
4691 addr,
4699 failed:
4700 slot++;
4705 }
4708 /* couldn't read this block, must give up */
4712 }
4714 idx++;
4715 }
4717 }
4720 {
4735 }
4738 /* FIXME should record badblock */
4740 }
4744 }
4747 {
4753 }
4756 {
4768 }
4773 }
4779 d++) {
4786 }
4788 }
4794 }
4797 {
4818 };
4821 {
4823 }
4826 {
4828 }