| blob | e89a8d78a9ed537f417c414b2081ef5f9a97f291 |
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 */
983 }
986 }
989 {
993 }
996 {
997 /* stop syncio and normal IO and wait for everything to
998 * go quiet.
999 * We increment barrier and nr_waiting, and then
1000 * wait until nr_pending match nr_queued+extra
1001 * This is called in the context of one normal IO request
1002 * that has failed. Thus any sync request that might be pending
1003 * will be blocked by nr_pending, and we need to wait for
1004 * pending IO requests to complete or be queued for re-try.
1005 * Thus the number queued (nr_queued) plus this request (extra)
1006 * must match the number of pending IOs (nr_pending) before
1007 * we continue.
1008 */
1020 }
1023 {
1024 /* reverse the effect of the freeze */
1030 }
1034 {
1038 else
1040 }
1046 };
1049 {
1051 cb);
1065 }
1067 /* we aren't scheduling, so we can do the write-out directly. */
1082 /* Just ignore it */
1084 else
1087 }
1089 }
1093 {
1100 sector_t sectors;
1104 /*
1105 * Register the new request and wait if the reconstruction
1106 * thread has put up a bar for new requests.
1107 * Continue immediately if no resync is active currently.
1108 */
1115 /*
1116 * IO spans the reshape position. Need to wait for reshape to
1117 * pass
1118 */
1124 sectors);
1126 }
1128 read_again:
1133 }
1138 max_sectors);
1158 /*
1159 * Could not read all from this device, so we will need another
1160 * r10_bio.
1161 */
1168 else
1171 /*
1172 * Cannot call generic_make_request directly as that will be
1173 * queued in __generic_make_request and subsequent
1174 * mempool_alloc might block waiting for it. so hand bio over
1175 * to raid10d.
1176 */
1190 }
1194 {
1204 sector_t sectors;
1210 /*
1211 * Register the new request and wait if the reconstruction
1212 * thread has put up a bar for new requests.
1213 * Continue immediately if no resync is active currently.
1214 */
1221 /*
1222 * IO spans the reshape position. Need to wait for reshape to
1223 * pass
1224 */
1230 sectors);
1232 }
1240 /* Need to update reshape_position in metadata */
1250 }
1257 }
1258 /* first select target devices under rcu_lock and
1259 * inc refcount on their rdev. Record them by setting
1260 * bios[x] to bio
1261 * If there are known/acknowledged bad blocks on any device
1262 * on which we have seen a write error, we want to avoid
1263 * writing to those blocks. This potentially requires several
1264 * writes to write around the bad blocks. Each set of writes
1265 * gets its own r10_bio with a set of bios attached. The number
1266 * of r10_bios is recored in bio->bi_phys_segments just as with
1267 * the read case.
1268 */
1272 retry_write:
1288 }
1293 }
1305 }
1307 sector_t first_bad;
1315 /* Mustn't write here until the bad block
1316 * is acknowledged
1317 */
1322 }
1324 /* Cannot write here at all */
1327 /* Mustn't write more than bad_sectors
1328 * to other devices yet
1329 */
1331 /* We don't set R10BIO_Degraded as that
1332 * only applies if the disk is missing,
1333 * so it might be re-added, and we want to
1334 * know to recover this chunk.
1335 * In this case the device is here, and the
1336 * fact that this chunk is not in-sync is
1337 * recorded in the bad block log.
1338 */
1340 }
1345 }
1346 }
1350 }
1354 }
1355 }
1359 /* Have to wait for this device to get unblocked, then retry */
1367 }
1373 /* Race with remove_disk */
1376 }
1378 }
1379 }
1385 }
1388 /* We are splitting this into multiple parts, so
1389 * we need to prepare for allocating another r10_bio.
1390 */
1395 else
1398 }
1412 max_sectors);
1429 /* flush_pending_writes() needs access to the rdev so...*/
1438 cb);
1439 else
1448 }
1452 }
1457 /* Replacement just got moved to main 'rdev' */
1460 }
1463 max_sectors);
1477 /* flush_pending_writes() needs access to the rdev so...*/
1486 cb);
1487 else
1496 }
1500 }
1501 }
1503 /* Don't remove the bias on 'remaining' (one_write_done) until
1504 * after checking if we need to go around again.
1505 */
1509 /* We need another r10_bio. It has already been counted
1510 * in bio->bi_phys_segments.
1511 */
1521 }
1523 }
1526 {
1539 /*
1540 * We might need to issue multiple reads to different devices if there
1541 * are bad blocks around, so we keep track of the number of reads in
1542 * bio->bi_phys_segments. If this is 0, there is only one r10_bio and
1543 * no locking will be needed when the request completes. If it is
1544 * non-zero, then it is the number of not-completed requests.
1545 */
1551 else
1553 }
1556 {
1566 }
1570 /*
1571 * If this request crosses a chunk boundary, we need to split
1572 * it.
1573 */
1586 }
1588 /*
1589 * If a bio is splitted, the first part of bio will pass
1590 * barrier but the bio is queued in current->bio_list (see
1591 * generic_make_request). If there is a raise_barrier() called
1592 * here, the second part of bio can't pass barrier. But since
1593 * the first part bio isn't dispatched to underlaying disks
1594 * yet, the barrier is never released, hence raise_barrier will
1595 * alays wait. We have a deadlock.
1596 * Note, this only happens in read path. For write path, the
1597 * first part of bio is dispatched in a schedule() call
1598 * (because of blk plug) or offloaded to raid10d.
1599 * Quitting from the function immediately can change the bio
1600 * order queued in bio_list and avoid the deadlock.
1601 */
1606 }
1609 /* In case raid10d snuck in to freeze_array */
1611 }
1614 {
1625 else
1629 }
1636 }
1639 }
1641 /* check if there are enough drives for
1642 * every block to appear on atleast one.
1643 * Don't consider the device numbered 'ignore'
1644 * as we might be about to remove it.
1645 */
1647 {
1657 }
1669 cnt++;
1671 }
1677 out:
1680 }
1683 {
1684 /* when calling 'enough', both 'prev' and 'geo' must
1685 * be stable.
1686 * This is ensured if ->reconfig_mutex or ->device_lock
1687 * is held.
1688 */
1691 }
1694 {
1699 /*
1700 * If it is not operational, then we have already marked it as dead
1701 * else if it is the last working disks, ignore the error, let the
1702 * next level up know.
1703 * else mark the drive as failed
1704 */
1708 /*
1709 * Don't fail the drive, just return an IO error.
1710 */
1713 }
1716 /*
1717 * If recovery is running, make sure it aborts.
1718 */
1729 }
1732 {
1740 }
1744 /* This is only called with ->reconfix_mutex held, so
1745 * rcu protection of rdev is not needed */
1754 }
1755 }
1758 {
1764 }
1767 {
1774 /*
1775 * Find all non-in_sync disks within the RAID10 configuration
1776 * and mark them in_sync
1777 */
1784 /* Replacement has just become active */
1787 count++;
1789 /* Replaced device not technically faulty,
1790 * but we need to be sure it gets removed
1791 * and never re-added.
1792 */
1796 }
1802 count++;
1804 }
1805 }
1812 }
1815 {
1823 /* only hot-add to in-sync arrays, as recovery is
1824 * very different from resync
1825 */
1839 else
1859 }
1873 }
1879 }
1882 {
1894 else
1901 }
1902 /* Only remove non-faulty devices if recovery
1903 * is not possible.
1904 */
1912 }
1917 /* lost the race, try later */
1921 }
1922 }
1924 /* We must have just cleared 'rdev' */
1928 * but will never see neither -- if they are careful.
1929 */
1933 /* We might have just remove the Replacement as faulty
1934 * Clear the flag just in case
1935 */
1940 abort:
1944 }
1947 {
1953 /* this is a reshape read */
1960 else
1961 /* The write handler will notice the lack of
1962 * R10BIO_Uptodate and record any errors etc
1963 */
1967 /* for reconstruct, we always reschedule after a read.
1968 * for resync, only after all reads
1969 */
1973 /* we have read all the blocks,
1974 * do the comparison in process context in raid10d
1975 */
1977 }
1978 }
1981 {
1986 /* the primary of several recovery bios */
1991 else
2000 else
2003 }
2004 }
2005 }
2008 {
2013 sector_t first_bad;
2022 else
2034 }
2044 }
2046 /*
2047 * Note: sync and recover and handled very differently for raid10
2048 * This code is for resync.
2049 * For resync, we read through virtual addresses and read all blocks.
2050 * If there is any error, we schedule a write. The lowest numbered
2051 * drive is authoritative.
2052 * However requests come for physical address, so we need to map.
2053 * For every physical address there are raid_disks/copies virtual addresses,
2054 * which is always are least one, but is not necessarly an integer.
2055 * This means that a physical address can span multiple chunks, so we may
2056 * have to submit multiple io requests for a single sync request.
2057 */
2058 /*
2059 * We check if all blocks are in-sync and only write to blocks that
2060 * aren't in sync
2061 */
2063 {
2071 /* find the first device with a block */
2085 /* now find blocks with errors */
2099 /* We know that the bi_io_vec layout is the same for
2100 * both 'first' and 'i', so we just compare them.
2101 * All vec entries are PAGE_SIZE;
2102 */
2110 len))
2113 }
2118 /* Don't fix anything. */
2121 /* Just give up on this device */
2124 }
2125 /* Ok, we need to write this bio, either to correct an
2126 * inconsistency or to correct an unreadable block.
2127 * First we need to fixup bv_offset, bv_len and
2128 * bi_vecs, as the read request might have corrupted these
2129 */
2150 }
2152 /* Now write out to any replacement devices
2153 * that are active
2154 */
2169 }
2171 done:
2175 }
2176 }
2178 /*
2179 * Now for the recovery code.
2180 * Recovery happens across physical sectors.
2181 * We recover all non-is_sync drives by finding the virtual address of
2182 * each, and then choose a working drive that also has that virt address.
2183 * There is a separate r10_bio for each non-in_sync drive.
2184 * Only the first two slots are in use. The first for reading,
2185 * The second for writing.
2186 *
2187 */
2189 {
2190 /* We got a read error during recovery.
2191 * We repeat the read in smaller page-sized sections.
2192 * If a read succeeds, write it to the new device or record
2193 * a bad block if we cannot.
2194 * If a read fails, record a bad block on both old and
2195 * new devices.
2196 */
2209 sector_t addr;
2218 addr,
2226 addr,
2236 }
2237 }
2239 /* We don't worry if we cannot set a bad block -
2240 * it really is bad so there is no loss in not
2241 * recording it yet
2242 */
2246 /* need bad block on destination too */
2251 /* just abort the recovery */
2260 }
2261 }
2262 }
2266 idx++;
2267 }
2268 }
2271 {
2280 }
2282 /*
2283 * share the pages with the first bio
2284 * and submit the write request
2285 */
2289 /* Need to test wbio2->bi_end_io before we call
2290 * generic_make_request as if the former is NULL,
2291 * the latter is free to free wbio2.
2292 */
2299 }
2305 }
2306 }
2308 /*
2309 * Used by fix_read_error() to decay the per rdev read_errors.
2310 * We halve the read error count for every hour that has elapsed
2311 * since the last recorded read error.
2312 *
2313 */
2315 {
2323 /* first time we've seen a read error */
2326 }
2333 /*
2334 * if hours_since_last is > the number of bits in read_errors
2335 * just set read errors to 0. We do this to avoid
2336 * overflowing the shift of read_errors by hours_since_last.
2337 */
2340 else
2342 }
2346 {
2347 sector_t first_bad;
2354 /* success */
2361 }
2362 /* need to record an error - either for the block or the device */
2366 }
2368 /*
2369 * This is a kernel thread which:
2370 *
2371 * 1. Retries failed read operations on working mirrors.
2372 * 2. Updates the raid superblock when problems encounter.
2373 * 3. Performs writes following reads for array synchronising.
2374 */
2377 {
2384 /* still own a reference to this rdev, so it cannot
2385 * have been cleared recently.
2386 */
2390 /* drive has already been failed, just ignore any
2391 more fix_read_error() attempts */
2408 }
2421 sector_t first_bad;
2435 sect,
2443 }
2444 sl++;
2451 /* Cannot read from anywhere, just mark the block
2452 * as bad on the first device to discourage future
2453 * reads.
2454 */
2459 rdev,
2466 }
2468 }
2471 /* write it back and re-read */
2478 sl--;
2490 sect,
2493 /* Well, this device is dead */
2497 sect +
2499 rdev)),
2504 }
2507 }
2514 sl--;
2526 sect,
2528 READ)) {
2530 /* Well, this device is dead */
2534 sect +
2545 sect +
2549 }
2553 }
2558 }
2559 }
2562 {
2567 /* bio has the data to be written to slot 'i' where
2568 * we just recently had a write error.
2569 * We repeatedly clone the bio and trim down to one block,
2570 * then try the write. Where the write fails we record
2571 * a bad block.
2572 * It is conceivable that the bio doesn't exactly align with
2573 * blocks. We must handle this.
2574 *
2575 * We currently own a reference to the rdev.
2576 */
2579 sector_t sector;
2596 sector_t wsector;
2599 /* Write at 'sector' for 'sectors' */
2609 /* Failure! */
2618 }
2620 }
2623 {
2631 dev_t bio_dev;
2632 sector_t bio_last_sector;
2634 /* we got a read error. Maybe the drive is bad. Maybe just
2635 * the block and we can fix it.
2636 * We freeze all other IO, and try reading the block from
2637 * other devices. When we find one, we re-write
2638 * and check it that fixes the read error.
2639 * This is all done synchronously while the array is
2640 * frozen.
2641 */
2660 read_more:
2668 }
2694 /* Drat - have to split this up more */
2703 else
2709 GFP_NOIO);
2722 }
2725 {
2726 /* Some sort of write request has finished and it
2727 * succeeded in writing where we thought there was a
2728 * bad block. So forget the bad block.
2729 * Or possibly if failed and we need to record
2730 * a bad block.
2731 */
2744 rdev,
2749 rdev,
2753 }
2760 rdev,
2765 rdev,
2769 }
2770 }
2780 rdev,
2790 }
2792 }
2797 rdev,
2801 }
2802 }
2814 }
2815 }
2816 }
2819 {
2837 }
2838 }
2842 retry_list);
2851 }
2852 }
2863 }
2883 /* just a partial read to be scheduled from a
2884 * separate context
2885 */
2888 }
2893 }
2895 }
2898 {
2913 }
2915 /*
2916 * perform a "sync" on one "block"
2917 *
2918 * We need to make sure that no normal I/O request - particularly write
2919 * requests - conflict with active sync requests.
2920 *
2921 * This is achieved by tracking pending requests and a 'barrier' concept
2922 * that can be installed to exclude normal IO requests.
2923 *
2924 * Resync and recovery are handled very differently.
2925 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2926 *
2927 * For resync, we iterate over virtual addresses, read all copies,
2928 * and update if there are differences. If only one copy is live,
2929 * skip it.
2930 * For recovery, we iterate over physical addresses, read a good
2931 * value for each non-in_sync drive, and over-write.
2932 *
2933 * So, for recovery we may have several outstanding complex requests for a
2934 * given address, one for each out-of-sync device. We model this by allocating
2935 * a number of r10_bio structures, one for each out-of-sync device.
2936 * As we setup these structures, we collect all bio's together into a list
2937 * which we then process collectively to add pages, and then process again
2938 * to pass to generic_make_request.
2939 *
2940 * The r10_bio structures are linked using a borrowed master_bio pointer.
2941 * This link is counted in ->remaining. When the r10_bio that points to NULL
2942 * has its remaining count decremented to 0, the whole complex operation
2943 * is complete.
2944 *
2945 */
2949 {
2956 sector_t sync_blocks;
2965 /*
2966 * Allow skipping a full rebuild for incremental assembly
2967 * of a clean array, like RAID1 does.
2968 */
2978 }
2980 skipped:
2986 /* If we aborted, we need to abort the
2987 * sync on the 'current' bitmap chucks (there can
2988 * be several when recovering multiple devices).
2989 * as we may have started syncing it but not finished.
2990 * We can find the current address in
2991 * mddev->curr_resync, but for recovery,
2992 * we need to convert that to several
2993 * virtual addresses.
2994 */
2999 }
3006 sector_t sect =
3010 }
3012 /* completed sync */
3016 /* Completed a full sync so the replacements
3017 * are now fully recovered.
3018 */
3025 }
3027 }
3029 }
3034 }
3040 /* if there has been nothing to do on any drive,
3041 * then there is nothing to do at all..
3042 */
3045 }
3050 /* make sure whole request will fit in a chunk - if chunks
3051 * are meaningful
3052 */
3057 /*
3058 * If there is non-resync activity waiting for a turn, then let it
3059 * though before starting on this new sync request.
3060 */
3064 /* Again, very different code for resync and recovery.
3065 * Both must result in an r10bio with a list of bios that
3066 * have bi_end_io, bi_sector, bi_bdev set,
3067 * and bi_private set to the r10bio.
3068 * For recovery, we may actually create several r10bios
3069 * with 2 bios in each, that correspond to the bios in the main one.
3070 * In this case, the subordinate r10bios link back through a
3071 * borrowed master_bio pointer, and the counter in the master
3072 * includes a ref from each subordinate.
3073 */
3074 /* First, we decide what to do and set ->bi_end_io
3075 * To end_sync_read if we want to read, and
3076 * end_sync_write if we will want to write.
3077 */
3081 /* recovery... the complicated one */
3088 sector_t sect;
3105 }
3108 /* want to reconstruct this device */
3112 /* last stripe is not complete - don't
3113 * try to recover this sector.
3114 */
3117 }
3120 /* Unless we are doing a full sync, or a replacement
3121 * we only need to recover the block if it is set in
3122 * the bitmap
3123 */
3131 /* yep, skip the sync_blocks here, but don't assume
3132 * that there will never be anything to do here
3133 */
3137 }
3157 /* Need to check if the array will still be
3158 * degraded
3159 */
3167 }
3168 }
3185 /* This is where we read from */
3194 bad_sectors -= (sector
3199 }
3200 }
3215 /* and we write to 'i' (if not in_sync) */
3242 /* and maybe write to replacement */
3246 /* Note: if mreplace != NULL, then bio
3247 * cannot be NULL as r10buf_pool_alloc will
3248 * have allocated it.
3249 * So the second test here is pointless.
3250 * But it keeps semantic-checkers happy, and
3251 * this comment keeps human reviewers
3252 * happy.
3253 */
3268 }
3271 /* Cannot recover, so abort the recovery or
3272 * record a bad block */
3274 /* problem is that there are bad blocks
3275 * on other device(s)
3276 */
3284 mrdev,
3290 mreplace,
3294 }
3300 mirror->recovery_disabled
3302 }
3311 }
3316 /* Only want this if there is elsewhere to
3317 * read from. 'j' is currently the first
3318 * readable copy.
3319 */
3326 targets++;
3327 }
3331 }
3332 }
3339 }
3341 }
3343 /* resync. Schedule a read for every block at this virt offset */
3352 /* We can skip this block */
3355 }
3389 }
3401 }
3402 }
3414 count++;
3420 }
3424 /* Need to set up for writing to the replacement */
3439 count++;
3440 }
3447 mddev);
3452 mddev);
3453 }
3457 }
3458 }
3476 /* stop here */
3481 /* remove last page from this bio */
3485 }
3487 }
3491 bio_full:
3506 }
3507 }
3510 /* pretend they weren't skipped, it makes
3511 * no important difference in this case
3512 */
3516 giveup:
3517 /* There is nowhere to write, so all non-sync
3518 * drives must be failed or in resync, all drives
3519 * have a bad block, so try the next chunk...
3520 */
3525 chunks_skipped ++;
3528 }
3530 static sector_t
3532 {
3533 sector_t size;
3548 }
3551 {
3552 /* Calculate the number of sectors-per-device that will
3553 * actually be used, and set conf->dev_sectors and
3554 * conf->stride
3555 */
3561 /* 'size' is now the number of chunks in the array */
3562 /* calculate "used chunks per device" */
3565 /* We need to round up when dividing by raid_disks to
3566 * get the stride size.
3567 */
3577 }
3578 }
3582 {
3598 * updated yet. */
3603 }
3621 * actually using this, but leave code here just in case.*/
3631 }
3635 }
3638 {
3650 }
3656 }
3663 /* FIXME calc properly */
3666 GFP_KERNEL);
3689 }
3693 else
3694 /* far_copies must be 1 */
3696 }
3713 out:
3719 }
3721 }
3724 {
3729 sector_t size;
3739 }
3755 else
3758 }
3780 }
3798 }
3804 else
3807 }
3808 /* need to check that every block has at least one working mirror */
3813 }
3816 /* must ensure that shape change is supported */
3823 }
3829 i++) {
3834 /* The replacement is all we have - use it */
3838 }
3847 }
3849 }
3857 /*
3858 * Ok, everything is just fine now
3859 */
3870 /* Calculate max read-ahead size.
3871 * We need to readahead at least twice a whole stripe....
3872 * maybe...
3873 */
3877 }
3891 /* This cannot work */
3894 }
3903 }
3907 out_free_conf:
3914 out:
3916 }
3919 {
3928 }
3931 {
3941 }
3942 }
3945 {
3946 /* Resize of 'far' arrays is not supported.
3947 * For 'near' and 'offset' arrays we can set the
3948 * number of sectors used to be an appropriate multiple
3949 * of the chunk size.
3950 * For 'offset', this is far_copies*chunksize.
3951 * For 'near' the multiplier is the LCM of
3952 * near_copies and raid_disks.
3953 * So if far_copies > 1 && !far_offset, fail.
3954 * Else find LCM(raid_disks, near_copy)*far_copies and
3955 * multiply by chunk_size. Then round to this number.
3956 * This is mostly done by raid10_size()
3957 */
3976 }
3982 }
3987 }
3990 {
3998 }
4001 /* Set new parameters */
4003 /* new layout: far_copies = 1, near_copies = 2 */
4008 /* make sure it will be not marked as dirty */
4018 }
4020 }
4023 }
4026 {
4029 /* raid10 can take over:
4030 * raid0 - providing it has only two drives
4031 */
4033 /* for raid0 takeover only one zone is supported */
4039 }
4043 }
4045 }
4048 {
4049 /* Called when there is a request to change
4050 * - layout (to ->new_layout)
4051 * - chunk size (to ->new_chunk_sectors)
4052 * - raid_disks (by delta_disks)
4053 * or when trying to restart a reshape that was ongoing.
4054 *
4055 * We need to validate the request and possibly allocate
4056 * space if that might be an issue later.
4057 *
4058 * Currently we reject any reshape of a 'far' mode array,
4059 * allow chunk size to change if new is generally acceptable,
4060 * allow raid_disks to increase, and allow
4061 * a switch between 'near' mode and 'offset' mode.
4062 */
4070 /* mustn't change number of copies */
4073 /* Cannot switch to 'far' mode */
4077 /* not factor of array size */
4086 /* allocate new 'mirrors' list */
4091 GFP_KERNEL);
4094 }
4096 }
4098 /*
4099 * Need to check if array has failed when deciding whether to:
4100 * - start an array
4101 * - remove non-faulty devices
4102 * - add a spare
4103 * - allow a reshape
4104 * This determination is simple when no reshape is happening.
4105 * However if there is a reshape, we need to carefully check
4106 * both the before and after sections.
4107 * This is because some failed devices may only affect one
4108 * of the two sections, and some non-in_sync devices may
4109 * be insync in the section most affected by failed devices.
4110 */
4112 {
4118 /* 'prev' section first */
4122 degraded++;
4124 /* When we can reduce the number of devices in
4125 * an array, this might not contribute to
4126 * 'degraded'. It does now.
4127 */
4128 degraded++;
4129 }
4138 degraded2++;
4140 /* If reshape is increasing the number of devices,
4141 * this section has already been recovered, so
4142 * it doesn't contribute to degraded.
4143 * else it does.
4144 */
4146 degraded2++;
4147 }
4148 }
4153 }
4156 {
4157 /* A 'reshape' has been requested. This commits
4158 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4159 * This also checks if there are enough spares and adds them
4160 * to the array.
4161 * We currently require enough spares to make the final
4162 * array non-degraded. We also require that the difference
4163 * between old and new data_offset - on each device - is
4164 * enough that we never risk over-writing.
4165 */
4190 spares++;
4200 }
4201 }
4219 }
4229 }
4244 }
4253 else
4258 }
4261 /* This is a spare that was manually added */
4263 }
4264 }
4265 /* When a reshape changes the number of devices,
4266 * ->degraded is measured against the larger of the
4267 * pre and post numbers.
4268 */
4287 }
4293 abort:
4306 }
4308 /* Calculate the last device-address that could contain
4309 * any block from the chunk that includes the array-address 's'
4310 * and report the next address.
4311 * i.e. the address returned will be chunk-aligned and after
4312 * any data that is in the chunk containing 's'.
4313 */
4315 {
4323 }
4325 /* Calculate the first device-address that could contain
4326 * any block from the chunk that includes the array-address 's'.
4327 * This too will be the start of a chunk
4328 */
4330 {
4337 }
4341 {
4342 /* We simply copy at most one chunk (smallest of old and new)
4343 * at a time, possibly less if that exceeds RESYNC_PAGES,
4344 * or we hit a bad block or something.
4345 * This might mean we pause for normal IO in the middle of
4346 * a chunk, but that is not a problem as mddev->reshape_position
4347 * can record any location.
4348 *
4349 * If we will want to write to a location that isn't
4350 * yet recorded as 'safe' (i.e. in metadata on disk) then
4351 * we need to flush all reshape requests and update the metadata.
4352 *
4353 * When reshaping forwards (e.g. to more devices), we interpret
4354 * 'safe' as the earliest block which might not have been copied
4355 * down yet. We divide this by previous stripe size and multiply
4356 * by previous stripe length to get lowest device offset that we
4357 * cannot write to yet.
4358 * We interpret 'sector_nr' as an address that we want to write to.
4359 * From this we use last_device_address() to find where we might
4360 * write to, and first_device_address on the 'safe' position.
4361 * If this 'next' write position is after the 'safe' position,
4362 * we must update the metadata to increase the 'safe' position.
4363 *
4364 * When reshaping backwards, we round in the opposite direction
4365 * and perform the reverse test: next write position must not be
4366 * less than current safe position.
4367 *
4368 * In all this the minimum difference in data offsets
4369 * (conf->offset_diff - always positive) allows a bit of slack,
4370 * so next can be after 'safe', but not by more than offset_diff
4371 *
4372 * We need to prepare all the bios here before we start any IO
4373 * to ensure the size we choose is acceptable to all devices.
4374 * The means one for each copy for write-out and an extra one for
4375 * read-in.
4376 * We store the read-in bio in ->master_bio and the others in
4377 * ->devs[x].bio and ->devs[x].repl_bio.
4378 */
4392 /* If restarting in the middle, skip the initial sectors */
4405 }
4406 }
4408 /* We don't use sector_nr to track where we are up to
4409 * as that doesn't work well for ->reshape_backwards.
4410 * So just use ->reshape_progress.
4411 */
4413 /* 'next' is the earliest device address that we might
4414 * write to for this chunk in the new layout
4415 */
4419 /* 'safe' is the last device address that we might read from
4420 * in the old layout after a restart
4421 */
4434 /* 'next' is after the last device address that we
4435 * might write to for this chunk in the new layout
4436 */
4439 /* 'safe' is the earliest device address that we might
4440 * read from in the old layout after a restart
4441 */
4444 /* Need to update metadata if 'next' might be beyond 'safe'
4445 * as that would possibly corrupt data
4446 */
4456 }
4460 /* Need to update reshape_position in metadata */
4466 else
4476 }
4479 }
4481 read_more:
4482 /* Now schedule reads for blocks from sector_nr to last */
4495 /* Cannot read from here, so need to record bad blocks
4496 * on all the target devices.
4497 */
4498 // FIXME
4502 }
4519 /* Now find the locations in the new layout */
4536 }
4549 }
4551 /* Now add as many pages as possible to all of these bios. */
4564 /* Didn't fit, must stop */
4568 /* Remove last page from this bio */
4572 }
4574 }
4577 }
4578 bio_full:
4582 /* Now submit the read */
4592 /* Now that we have done the whole section we can
4593 * update reshape_progress
4594 */
4597 else
4601 }
4607 {
4608 /* Reshape read completed. Hopefully we have a block
4609 * to write out.
4610 * If we got a read error then we do sync 1-page reads from
4611 * elsewhere until we find the data - or give up.
4612 */
4618 /* Reshape has been aborted */
4621 }
4623 /* We definitely have the data in the pages, schedule the
4624 * writes.
4625 */
4638 }
4642 }
4649 }
4651 }
4654 {
4666 /* read-ahead size must cover two whole stripes, which is
4667 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4668 */
4675 }
4677 }
4681 {
4682 /* Use sync reads to get the blocks from somewhere else */
4709 sector_t addr;
4719 addr,
4727 failed:
4728 slot++;
4733 }
4736 /* couldn't read this block, must give up */
4740 }
4742 idx++;
4743 }
4745 }
4748 {
4763 }
4766 /* FIXME should record badblock */
4768 }
4772 }
4775 {
4781 }
4784 {
4796 }
4801 }
4807 d++) {
4814 }
4816 }
4822 }
4825 {
4846 };
4849 {
4851 }
4854 {
4856 }