| blob | 59d4daa5f4c7a32c245ef954f24650fe75084117 |
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>
33 /*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
42 *
43 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most raid_disks.
51 *
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of being very far
54 * apart on disk, there are adjacent stripes.
55 *
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
63 * on a device):
64 * A B C D A B C D E
65 * ... ...
66 * D A B C E A B C D
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
71 */
73 /*
74 * Number of guaranteed r10bios in case of extreme VM load:
75 */
76 #define NR_RAID10_BIOS 256
78 /* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
82 */
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
87 */
88 #define IO_MADE_GOOD ((struct bio *)2)
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92 /* When there are this many requests queued to be written by
93 * the raid10 thread, we become 'congested' to provide back-pressure
94 * for writeback.
95 */
108 {
112 /* allocate a r10bio with room for raid_disks entries in the
113 * bios array */
115 }
118 {
120 }
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
130 /*
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
135 *
136 */
138 {
153 else
156 /*
157 * Allocate bios.
158 */
170 }
171 /*
172 * Allocate RESYNC_PAGES data pages and attach them
173 * where needed.
174 */
181 /* we can share bv_page's during recovery
182 * and reshape */
194 }
195 }
199 out_free_pages:
206 out_free_bio:
212 }
215 }
218 {
230 }
232 }
236 }
238 }
241 {
253 }
254 }
257 {
262 }
265 {
271 }
274 {
284 /* wake up frozen array... */
288 }
290 /*
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
293 * cache layer.
294 */
296 {
313 /*
314 * Wake up any possible resync thread that waits for the device
315 * to go idle.
316 */
318 }
320 }
322 /*
323 * Update disk head position estimator based on IRQ completion info.
324 */
326 {
331 }
333 /*
334 * Find the disk number which triggered given bio
335 */
338 {
348 }
349 }
359 }
362 {
373 /*
374 * this branch is our 'one mirror IO has finished' event handler:
375 */
379 /*
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
383 *
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
387 */
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
394 */
400 }
405 /*
406 * oops, read error - keep the refcount on the rdev
407 */
416 }
417 }
420 {
421 /* clear the bitmap if all writes complete successfully */
427 }
430 {
438 else
440 }
441 }
442 }
445 {
462 }
463 /*
464 * this branch is our 'one mirror IO has finished' event handler:
465 */
468 /* Never record new bad blocks to replacement,
469 * just fail it.
470 */
479 }
481 /*
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
485 *
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
489 */
490 sector_t first_bad;
495 /* Maybe we can clear some bad blocks. */
503 else
507 }
508 }
510 /*
511 *
512 * Let's see if all mirrored write operations have finished
513 * already.
514 */
518 }
520 /*
521 * RAID10 layout manager
522 * As well as the chunksize and raid_disks count, there are two
523 * parameters: near_copies and far_copies.
524 * near_copies * far_copies must be <= raid_disks.
525 * Normally one of these will be 1.
526 * If both are 1, we get raid0.
527 * If near_copies == raid_disks, we get raid1.
528 *
529 * Chunks are laid out in raid0 style with near_copies copies of the
530 * first chunk, followed by near_copies copies of the next chunk and
531 * so on.
532 * If far_copies > 1, then after 1/far_copies of the array has been assigned
533 * as described above, we start again with a device offset of near_copies.
534 * So we effectively have another copy of the whole array further down all
535 * the drives, but with blocks on different drives.
536 * With this layout, and block is never stored twice on the one device.
537 *
538 * raid10_find_phys finds the sector offset of a given virtual sector
539 * on each device that it is on.
540 *
541 * raid10_find_virt does the reverse mapping, from a device and a
542 * sector offset to a virtual address
543 */
546 {
548 sector_t sector;
549 sector_t chunk;
550 sector_t stripe;
561 /* now calculate first sector/dev */
573 /* and calculate all the others */
580 slot++;
594 }
598 slot++;
599 }
600 dev++;
604 }
605 }
606 }
609 {
621 }
624 {
626 /* Never use conf->prev as this is only called during resync
627 * or recovery, so reshape isn't happening
628 */
642 }
643 }
658 else
660 }
662 }
666 }
668 /**
669 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
670 * @q: request queue
671 * @bvm: properties of new bio
672 * @biovec: the request that could be merged to it.
673 *
674 * Return amount of bytes we can accept at this offset
675 * This requires checking for end-of-chunk if near_copies != raid_disks,
676 * and for subordinate merge_bvec_fns if merge_check_needed.
677 */
681 {
700 /* bio_add cannot handle a negative return */
715 /* Cannot give any guidance during reshape */
719 }
736 }
737 }
748 }
749 }
750 }
752 }
754 }
756 /*
757 * This routine returns the disk from which the requested read should
758 * be done. There is a per-array 'next expected sequential IO' sector
759 * number - if this matches on the next IO then we use the last disk.
760 * There is also a per-disk 'last know head position' sector that is
761 * maintained from IRQ contexts, both the normal and the resync IO
762 * completion handlers update this position correctly. If there is no
763 * perfect sequential match then we pick the disk whose head is closest.
764 *
765 * If there are 2 mirrors in the same 2 devices, performance degrades
766 * because position is mirror, not device based.
767 *
768 * The rdev for the device selected will have nr_pending incremented.
769 */
771 /*
772 * FIXME: possibly should rethink readbalancing and do it differently
773 * depending on near_copies / far_copies geometry.
774 */
778 {
791 retry:
798 /*
799 * Check if we can balance. We can balance on the whole
800 * device if no resync is going on (recovery is ok), or below
801 * the resync window. We take the first readable disk when
802 * above the resync window.
803 */
809 sector_t first_bad;
811 sector_t dev_sector;
833 /* Already have a better slot */
836 /* Cannot read here. If this is the
837 * 'primary' device, then we must not read
838 * beyond 'bad_sectors' from another device.
839 */
846 sector_t good_sectors =
852 }
854 /* Must read from here */
856 }
864 /* This optimisation is debatable, and completely destroys
865 * sequential read speed for 'far copies' arrays. So only
866 * keep it for 'near' arrays, and review those later.
867 */
871 /* for far > 1 always use the lowest address */
874 else
881 }
882 }
886 }
891 /* Cannot risk returning a device that failed
892 * before we inc'ed nr_pending
893 */
896 }
904 }
907 {
919 i++) {
925 }
926 }
929 }
933 {
938 }
941 {
942 /* Any writes that have been queued but are awaiting
943 * bitmap updates get flushed here.
944 */
952 /* flush any pending bitmap writes to disk
953 * before proceeding w/ I/O */
962 /* Just ignore it */
964 else
967 }
970 }
972 /* Barriers....
973 * Sometimes we need to suspend IO while we do something else,
974 * either some resync/recovery, or reconfigure the array.
975 * To do this we raise a 'barrier'.
976 * The 'barrier' is a counter that can be raised multiple times
977 * to count how many activities are happening which preclude
978 * normal IO.
979 * We can only raise the barrier if there is no pending IO.
980 * i.e. if nr_pending == 0.
981 * We choose only to raise the barrier if no-one is waiting for the
982 * barrier to go down. This means that as soon as an IO request
983 * is ready, no other operations which require a barrier will start
984 * until the IO request has had a chance.
985 *
986 * So: regular IO calls 'wait_barrier'. When that returns there
987 * is no backgroup IO happening, It must arrange to call
988 * allow_barrier when it has finished its IO.
989 * backgroup IO calls must call raise_barrier. Once that returns
990 * there is no normal IO happeing. It must arrange to call
991 * lower_barrier when the particular background IO completes.
992 */
995 {
999 /* Wait until no block IO is waiting (unless 'force') */
1003 /* block any new IO from starting */
1006 /* Now wait for all pending IO to complete */
1012 }
1015 {
1021 }
1024 {
1028 /* Wait for the barrier to drop.
1029 * However if there are already pending
1030 * requests (preventing the barrier from
1031 * rising completely), and the
1032 * pre-process bio queue isn't empty,
1033 * then don't wait, as we need to empty
1034 * that queue to get the nr_pending
1035 * count down.
1036 */
1044 }
1047 }
1050 {
1056 }
1059 {
1060 /* stop syncio and normal IO and wait for everything to
1061 * go quiet.
1062 * We increment barrier and nr_waiting, and then
1063 * wait until nr_pending match nr_queued+1
1064 * This is called in the context of one normal IO request
1065 * that has failed. Thus any sync request that might be pending
1066 * will be blocked by nr_pending, and we need to wait for
1067 * pending IO requests to complete or be queued for re-try.
1068 * Thus the number queued (nr_queued) plus this request (1)
1069 * must match the number of pending IOs (nr_pending) before
1070 * we continue.
1071 */
1081 }
1084 {
1085 /* reverse the effect of the freeze */
1091 }
1095 {
1099 else
1101 }
1107 };
1110 {
1112 cb);
1126 }
1128 /* we aren't scheduling, so we can do the write-out directly. */
1138 /* Just ignore it */
1140 else
1143 }
1145 }
1148 {
1172 }
1174 /* If this request crosses a chunk boundary, we need to
1175 * split it. This will only happen for 1 PAGE (or less) requests.
1176 */
1178 > chunk_sects
1182 /* Sanity check -- queue functions should prevent this happening */
1185 /* This is a one page bio that upper layers
1186 * refuse to split for us, so we need to split it.
1187 */
1191 /* Each of these 'make_request' calls will call 'wait_barrier'.
1192 * If the first succeeds but the second blocks due to the resync
1193 * thread raising the barrier, we will deadlock because the
1194 * IO to the underlying device will be queued in generic_make_request
1195 * and will never complete, so will never reduce nr_pending.
1196 * So increment nr_waiting here so no new raise_barriers will
1197 * succeed, and so the second wait_barrier cannot block.
1198 */
1213 bad_map:
1220 }
1224 /*
1225 * Register the new request and wait if the reconstruction
1226 * thread has put up a bar for new requests.
1227 * Continue immediately if no resync is active currently.
1228 */
1235 /* IO spans the reshape position. Need to wait for
1236 * reshape to pass
1237 */
1243 }
1251 /* Need to update reshape_position in metadata */
1260 }
1271 /* We might need to issue multiple reads to different
1272 * devices if there are bad blocks around, so we keep
1273 * track of the number of reads in bio->bi_phys_segments.
1274 * If this is 0, there is only one r10_bio and no locking
1275 * will be needed when the request completes. If it is
1276 * non-zero, then it is the number of not-completed requests.
1277 */
1282 /*
1283 * read balancing logic:
1284 */
1288 read_again:
1293 }
1298 max_sectors);
1311 /* Could not read all from this device, so we will
1312 * need another r10_bio.
1313 */
1320 else
1323 /* Cannot call generic_make_request directly
1324 * as that will be queued in __generic_make_request
1325 * and subsequent mempool_alloc might block
1326 * waiting for it. so hand bio over to raid10d.
1327 */
1341 }
1343 /*
1344 * WRITE:
1345 */
1350 }
1351 /* first select target devices under rcu_lock and
1352 * inc refcount on their rdev. Record them by setting
1353 * bios[x] to bio
1354 * If there are known/acknowledged bad blocks on any device
1355 * on which we have seen a write error, we want to avoid
1356 * writing to those blocks. This potentially requires several
1357 * writes to write around the bad blocks. Each set of writes
1358 * gets its own r10_bio with a set of bios attached. The number
1359 * of r10_bios is recored in bio->bi_phys_segments just as with
1360 * the read case.
1361 */
1365 retry_write:
1381 }
1386 }
1400 }
1402 sector_t first_bad;
1408 max_sectors,
1411 /* Mustn't write here until the bad block
1412 * is acknowledged
1413 */
1418 }
1420 /* Cannot write here at all */
1423 /* Mustn't write more than bad_sectors
1424 * to other devices yet
1425 */
1427 /* We don't set R10BIO_Degraded as that
1428 * only applies if the disk is missing,
1429 * so it might be re-added, and we want to
1430 * know to recover this chunk.
1431 * In this case the device is here, and the
1432 * fact that this chunk is not in-sync is
1433 * recorded in the bad block log.
1434 */
1436 }
1441 }
1442 }
1446 }
1450 }
1451 }
1455 /* Have to wait for this device to get unblocked, then retry */
1463 }
1469 /* Race with remove_disk */
1472 }
1474 }
1475 }
1480 }
1483 /* We are splitting this into multiple parts, so
1484 * we need to prepare for allocating another r10_bio.
1485 */
1490 else
1493 }
1506 max_sectors);
1511 rdev));
1524 cb);
1525 else
1534 }
1538 }
1543 /* Replacement just got moved to main 'rdev' */
1546 }
1549 max_sectors);
1568 }
1569 }
1571 /* Don't remove the bias on 'remaining' (one_write_done) until
1572 * after checking if we need to go around again.
1573 */
1577 /* We need another r10_bio. It has already been counted
1578 * in bio->bi_phys_segments.
1579 */
1589 }
1592 /* In case raid10d snuck in to freeze_array */
1594 }
1597 {
1608 else
1610 }
1618 }
1620 /* check if there are enough drives for
1621 * every block to appear on atleast one.
1622 * Don't consider the device numbered 'ignore'
1623 * as we might be about to remove it.
1624 */
1626 {
1636 cnt++;
1638 }
1644 }
1647 {
1650 }
1653 {
1657 /*
1658 * If it is not operational, then we have already marked it as dead
1659 * else if it is the last working disks, ignore the error, let the
1660 * next level up know.
1661 * else mark the drive as failed
1662 */
1665 /*
1666 * Don't fail the drive, just return an IO error.
1667 */
1674 /*
1675 * if recovery is running, make sure it aborts.
1676 */
1678 }
1687 }
1690 {
1698 }
1710 }
1711 }
1714 {
1720 }
1723 {
1730 /*
1731 * Find all non-in_sync disks within the RAID10 configuration
1732 * and mark them in_sync
1733 */
1740 /* Replacement has just become active */
1743 count++;
1745 /* Replaced device not technically faulty,
1746 * but we need to be sure it gets removed
1747 * and never re-added.
1748 */
1752 }
1757 count++;
1759 }
1760 }
1767 }
1771 {
1780 /* only hot-add to in-sync arrays, as recovery is
1781 * very different from resync
1782 */
1793 }
1798 else
1817 }
1830 }
1832 /* Some requests might not have seen this new
1833 * merge_bvec_fn. We must wait for them to complete
1834 * before merging the device fully.
1835 * First we make sure any code which has tested
1836 * our function has submitted the request, then
1837 * we wait for all outstanding requests to complete.
1838 */
1843 }
1850 }
1853 {
1865 else
1872 }
1873 /* Only remove faulty devices if recovery
1874 * is not possible.
1875 */
1883 }
1887 /* lost the race, try later */
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 }
1916 {
1922 /* this is a reshape read */
1929 else
1930 /* The write handler will notice the lack of
1931 * R10BIO_Uptodate and record any errors etc
1932 */
1936 /* for reconstruct, we always reschedule after a read.
1937 * for resync, only after all reads
1938 */
1942 /* we have read all the blocks,
1943 * do the comparison in process context in raid10d
1944 */
1946 }
1947 }
1950 {
1955 /* the primary of several recovery bios */
1960 else
1969 else
1972 }
1973 }
1974 }
1977 {
1983 sector_t first_bad;
1992 else
2004 }
2014 }
2016 /*
2017 * Note: sync and recover and handled very differently for raid10
2018 * This code is for resync.
2019 * For resync, we read through virtual addresses and read all blocks.
2020 * If there is any error, we schedule a write. The lowest numbered
2021 * drive is authoritative.
2022 * However requests come for physical address, so we need to map.
2023 * For every physical address there are raid_disks/copies virtual addresses,
2024 * which is always are least one, but is not necessarly an integer.
2025 * This means that a physical address can span multiple chunks, so we may
2026 * have to submit multiple io requests for a single sync request.
2027 */
2028 /*
2029 * We check if all blocks are in-sync and only write to blocks that
2030 * aren't in sync
2031 */
2033 {
2041 /* find the first device with a block */
2053 /* now find blocks with errors */
2064 /* We know that the bi_io_vec layout is the same for
2065 * both 'first' and 'i', so we just compare them.
2066 * All vec entries are PAGE_SIZE;
2067 */
2077 /* Don't fix anything. */
2079 }
2080 /* Ok, we need to write this bio, either to correct an
2081 * inconsistency or to correct an unreadable block.
2082 * First we need to fixup bv_offset, bv_len and
2083 * bi_vecs, as the read request might have corrupted these
2084 */
2099 PAGE_SIZE);
2100 }
2111 }
2113 /* Now write out to any replacement devices
2114 * that are active
2115 */
2127 PAGE_SIZE);
2133 }
2135 done:
2139 }
2140 }
2142 /*
2143 * Now for the recovery code.
2144 * Recovery happens across physical sectors.
2145 * We recover all non-is_sync drives by finding the virtual address of
2146 * each, and then choose a working drive that also has that virt address.
2147 * There is a separate r10_bio for each non-in_sync drive.
2148 * Only the first two slots are in use. The first for reading,
2149 * The second for writing.
2150 *
2151 */
2153 {
2154 /* We got a read error during recovery.
2155 * We repeat the read in smaller page-sized sections.
2156 * If a read succeeds, write it to the new device or record
2157 * a bad block if we cannot.
2158 * If a read fails, record a bad block on both old and
2159 * new devices.
2160 */
2173 sector_t addr;
2182 addr,
2190 addr,
2200 }
2201 }
2203 /* We don't worry if we cannot set a bad block -
2204 * it really is bad so there is no loss in not
2205 * recording it yet
2206 */
2210 /* need bad block on destination too */
2215 /* just abort the recovery */
2217 "md/raid10:%s: recovery aborted"
2226 }
2227 }
2228 }
2232 idx++;
2233 }
2234 }
2237 {
2246 }
2248 /*
2249 * share the pages with the first bio
2250 * and submit the write request
2251 */
2259 }
2265 }
2266 }
2269 /*
2270 * Used by fix_read_error() to decay the per rdev read_errors.
2271 * We halve the read error count for every hour that has elapsed
2272 * since the last recorded read error.
2273 *
2274 */
2276 {
2285 /* first time we've seen a read error */
2288 }
2295 /*
2296 * if hours_since_last is > the number of bits in read_errors
2297 * just set read errors to 0. We do this to avoid
2298 * overflowing the shift of read_errors by hours_since_last.
2299 */
2302 else
2304 }
2308 {
2309 sector_t first_bad;
2316 /* success */
2323 }
2324 /* need to record an error - either for the block or the device */
2328 }
2330 /*
2331 * This is a kernel thread which:
2332 *
2333 * 1. Retries failed read operations on working mirrors.
2334 * 2. Updates the raid superblock when problems encounter.
2335 * 3. Performs writes following reads for array synchronising.
2336 */
2339 {
2346 /* still own a reference to this rdev, so it cannot
2347 * have been cleared recently.
2348 */
2352 /* drive has already been failed, just ignore any
2353 more fix_read_error() attempts */
2363 "md/raid10:%s: %s: Raid device exceeded "
2373 }
2386 sector_t first_bad;
2400 sect,
2407 }
2408 sl++;
2415 /* Cannot read from anywhere, just mark the block
2416 * as bad on the first device to discourage future
2417 * reads.
2418 */
2423 rdev,
2430 }
2432 }
2435 /* write it back and re-read */
2442 sl--;
2454 sect,
2457 /* Well, this device is dead */
2459 "md/raid10:%s: read correction "
2460 "write failed"
2464 sect +
2466 rdev)),
2472 }
2475 }
2482 sl--;
2493 sect,
2495 READ)) {
2497 /* Well, this device is dead */
2499 "md/raid10:%s: unable to read back "
2500 "corrected sectors"
2504 sect +
2514 "md/raid10:%s: read error corrected"
2518 sect +
2522 }
2526 }
2531 }
2532 }
2535 {
2540 /* bio has the data to be written to slot 'i' where
2541 * we just recently had a write error.
2542 * We repeatedly clone the bio and trim down to one block,
2543 * then try the write. Where the write fails we record
2544 * a bad block.
2545 * It is conceivable that the bio doesn't exactly align with
2546 * blocks. We must handle this.
2547 *
2548 * We currently own a reference to the rdev.
2549 */
2552 sector_t sector;
2570 /* Write at 'sector' for 'sectors' */
2578 /* Failure! */
2587 }
2589 }
2592 {
2601 /* we got a read error. Maybe the drive is bad. Maybe just
2602 * the block and we can fix it.
2603 * We freeze all other IO, and try reading the block from
2604 * other devices. When we find one, we re-write
2605 * and check it that fixes the read error.
2606 * This is all done synchronously while the array is
2607 * frozen.
2608 */
2623 read_more:
2632 }
2637 KERN_ERR
2638 "md/raid10:%s: %s: redirecting "
2647 max_sectors);
2657 /* Drat - have to split this up more */
2666 else
2672 GFP_NOIO);
2685 }
2688 {
2689 /* Some sort of write request has finished and it
2690 * succeeded in writing where we thought there was a
2691 * bad block. So forget the bad block.
2692 * Or possibly if failed and we need to record
2693 * a bad block.
2694 */
2708 rdev,
2713 rdev,
2717 }
2724 rdev,
2729 rdev,
2733 }
2734 }
2743 rdev,
2753 }
2755 }
2760 rdev,
2764 }
2765 }
2770 }
2771 }
2774 {
2793 }
2813 /* just a partial read to be scheduled from a
2814 * separate context
2815 */
2818 }
2823 }
2825 }
2829 {
2844 }
2846 /*
2847 * perform a "sync" on one "block"
2848 *
2849 * We need to make sure that no normal I/O request - particularly write
2850 * requests - conflict with active sync requests.
2851 *
2852 * This is achieved by tracking pending requests and a 'barrier' concept
2853 * that can be installed to exclude normal IO requests.
2854 *
2855 * Resync and recovery are handled very differently.
2856 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2857 *
2858 * For resync, we iterate over virtual addresses, read all copies,
2859 * and update if there are differences. If only one copy is live,
2860 * skip it.
2861 * For recovery, we iterate over physical addresses, read a good
2862 * value for each non-in_sync drive, and over-write.
2863 *
2864 * So, for recovery we may have several outstanding complex requests for a
2865 * given address, one for each out-of-sync device. We model this by allocating
2866 * a number of r10_bio structures, one for each out-of-sync device.
2867 * As we setup these structures, we collect all bio's together into a list
2868 * which we then process collectively to add pages, and then process again
2869 * to pass to generic_make_request.
2870 *
2871 * The r10_bio structures are linked using a borrowed master_bio pointer.
2872 * This link is counted in ->remaining. When the r10_bio that points to NULL
2873 * has its remaining count decremented to 0, the whole complex operation
2874 * is complete.
2875 *
2876 */
2880 {
2887 sector_t sync_blocks;
2896 /*
2897 * Allow skipping a full rebuild for incremental assembly
2898 * of a clean array, like RAID1 does.
2899 */
2910 }
2912 skipped:
2918 /* If we aborted, we need to abort the
2919 * sync on the 'current' bitmap chucks (there can
2920 * be several when recovering multiple devices).
2921 * as we may have started syncing it but not finished.
2922 * We can find the current address in
2923 * mddev->curr_resync, but for recovery,
2924 * we need to convert that to several
2925 * virtual addresses.
2926 */
2930 }
2937 sector_t sect =
2941 }
2943 /* completed sync */
2947 /* Completed a full sync so the replacements
2948 * are now fully recovered.
2949 */
2953 ->recovery_offset
2955 }
2957 }
2962 }
2968 /* if there has been nothing to do on any drive,
2969 * then there is nothing to do at all..
2970 */
2973 }
2978 /* make sure whole request will fit in a chunk - if chunks
2979 * are meaningful
2980 */
2984 /*
2985 * If there is non-resync activity waiting for us then
2986 * put in a delay to throttle resync.
2987 */
2991 /* Again, very different code for resync and recovery.
2992 * Both must result in an r10bio with a list of bios that
2993 * have bi_end_io, bi_sector, bi_bdev set,
2994 * and bi_private set to the r10bio.
2995 * For recovery, we may actually create several r10bios
2996 * with 2 bios in each, that correspond to the bios in the main one.
2997 * In this case, the subordinate r10bios link back through a
2998 * borrowed master_bio pointer, and the counter in the master
2999 * includes a ref from each subordinate.
3000 */
3001 /* First, we decide what to do and set ->bi_end_io
3002 * To end_sync_read if we want to read, and
3003 * end_sync_write if we will want to write.
3004 */
3008 /* recovery... the complicated one */
3015 sector_t sect;
3022 &&
3029 /* want to reconstruct this device */
3033 /* last stripe is not complete - don't
3034 * try to recover this sector.
3035 */
3037 }
3038 /* Unless we are doing a full sync, or a replacement
3039 * we only need to recover the block if it is set in
3040 * the bitmap
3041 */
3049 /* yep, skip the sync_blocks here, but don't assume
3050 * that there will never be anything to do here
3051 */
3054 }
3069 /* Need to check if the array will still be
3070 * degraded
3071 */
3077 }
3093 /* This is where we read from */
3103 bad_sectors -= (sector
3108 }
3109 }
3121 /* and we write to 'i' (if not in_sync) */
3149 /* and maybe write to replacement */
3154 /* Note: if rdev != NULL, then bio
3155 * cannot be NULL as r10buf_pool_alloc will
3156 * have allocated it.
3157 * So the second test here is pointless.
3158 * But it keeps semantic-checkers happy, and
3159 * this comment keeps human reviewers
3160 * happy.
3161 */
3175 }
3177 /* Cannot recover, so abort the recovery or
3178 * record a bad block */
3184 /* problem is that there are bad blocks
3185 * on other device(s)
3186 */
3204 }
3211 mirror->recovery_disabled
3213 }
3215 }
3216 }
3223 }
3225 }
3227 /* resync. Schedule a read for every block at this virt offset */
3236 /* We can skip this block */
3239 }
3280 }
3281 }
3292 count++;
3299 /* Need to set up for writing to the replacement */
3314 count++;
3315 }
3322 mddev);
3327 mddev);
3328 }
3332 }
3333 }
3351 /* stop here */
3356 /* remove last page from this bio */
3360 }
3362 }
3366 bio_full:
3380 }
3381 }
3384 /* pretend they weren't skipped, it makes
3385 * no important difference in this case
3386 */
3390 giveup:
3391 /* There is nowhere to write, so all non-sync
3392 * drives must be failed or in resync, all drives
3393 * have a bad block, so try the next chunk...
3394 */
3399 chunks_skipped ++;
3402 }
3404 static sector_t
3406 {
3407 sector_t size;
3422 }
3425 {
3426 /* Calculate the number of sectors-per-device that will
3427 * actually be used, and set conf->dev_sectors and
3428 * conf->stride
3429 */
3435 /* 'size' is now the number of chunks in the array */
3436 /* calculate "used chunks per device" */
3439 /* We need to round up when dividing by raid_disks to
3440 * get the stride size.
3441 */
3451 }
3452 }
3456 {
3472 * updated yet. */
3477 }
3494 }
3497 {
3510 }
3516 }
3523 /* FIXME calc properly */
3526 GFP_KERNEL);
3549 }
3553 else
3554 /* far_copies must be 1 */
3556 }
3570 out:
3580 }
3582 }
3585 {
3590 sector_t size;
3600 }
3617 else
3620 }
3642 }
3662 }
3668 else
3671 }
3672 /* need to check that every block has at least one working mirror */
3677 }
3680 /* must ensure that shape change is supported */
3687 }
3693 i++) {
3698 /* The replacement is all we have - use it */
3702 }
3710 }
3712 }
3722 /*
3723 * Ok, everything is just fine now
3724 */
3736 /* Calculate max read-ahead size.
3737 * We need to readahead at least twice a whole stripe....
3738 * maybe...
3739 */
3744 }
3759 /* This cannot work */
3762 }
3772 }
3776 out_free_conf:
3784 out:
3786 }
3789 {
3797 /* the unplug fn references 'conf'*/
3807 }
3810 {
3820 }
3821 }
3824 {
3825 /* Resize of 'far' arrays is not supported.
3826 * For 'near' and 'offset' arrays we can set the
3827 * number of sectors used to be an appropriate multiple
3828 * of the chunk size.
3829 * For 'offset', this is far_copies*chunksize.
3830 * For 'near' the multiplier is the LCM of
3831 * near_copies and raid_disks.
3832 * So if far_copies > 1 && !far_offset, fail.
3833 * Else find LCM(raid_disks, near_copy)*far_copies and
3834 * multiply by chunk_size. Then round to this number.
3835 * This is mostly done by raid10_size()
3836 */
3855 }
3863 }
3868 }
3871 {
3879 }
3881 /* Set new parameters */
3883 /* new layout: far_copies = 1, near_copies = 2 */
3888 /* make sure it will be not marked as dirty */
3897 }
3900 }
3903 {
3906 /* raid10 can take over:
3907 * raid0 - providing it has only two drives
3908 */
3910 /* for raid0 takeover only one zone is supported */
3917 }
3919 }
3921 }
3924 {
3925 /* Called when there is a request to change
3926 * - layout (to ->new_layout)
3927 * - chunk size (to ->new_chunk_sectors)
3928 * - raid_disks (by delta_disks)
3929 * or when trying to restart a reshape that was ongoing.
3930 *
3931 * We need to validate the request and possibly allocate
3932 * space if that might be an issue later.
3933 *
3934 * Currently we reject any reshape of a 'far' mode array,
3935 * allow chunk size to change if new is generally acceptable,
3936 * allow raid_disks to increase, and allow
3937 * a switch between 'near' mode and 'offset' mode.
3938 */
3946 /* mustn't change number of copies */
3949 /* Cannot switch to 'far' mode */
3953 /* not factor of array size */
3962 /* allocate new 'mirrors' list */
3967 GFP_KERNEL);
3970 }
3972 }
3974 /*
3975 * Need to check if array has failed when deciding whether to:
3976 * - start an array
3977 * - remove non-faulty devices
3978 * - add a spare
3979 * - allow a reshape
3980 * This determination is simple when no reshape is happening.
3981 * However if there is a reshape, we need to carefully check
3982 * both the before and after sections.
3983 * This is because some failed devices may only affect one
3984 * of the two sections, and some non-in_sync devices may
3985 * be insync in the section most affected by failed devices.
3986 */
3988 {
3994 /* 'prev' section first */
3998 degraded++;
4000 /* When we can reduce the number of devices in
4001 * an array, this might not contribute to
4002 * 'degraded'. It does now.
4003 */
4004 degraded++;
4005 }
4014 degraded2++;
4016 /* If reshape is increasing the number of devices,
4017 * this section has already been recovered, so
4018 * it doesn't contribute to degraded.
4019 * else it does.
4020 */
4022 degraded2++;
4023 }
4024 }
4029 }
4032 {
4033 /* A 'reshape' has been requested. This commits
4034 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4035 * This also checks if there are enough spares and adds them
4036 * to the array.
4037 * We currently require enough spares to make the final
4038 * array non-degraded. We also require that the difference
4039 * between old and new data_offset - on each device - is
4040 * enough that we never risk over-writing.
4041 */
4066 spares++;
4076 }
4077 }
4095 }
4105 }
4119 }
4128 else
4133 }
4136 /* This is a spare that was manually added */
4138 }
4139 }
4140 /* When a reshape changes the number of devices,
4141 * ->degraded is measured against the larger of the
4142 * pre and post numbers.
4143 */
4161 }
4167 abort:
4179 }
4181 /* Calculate the last device-address that could contain
4182 * any block from the chunk that includes the array-address 's'
4183 * and report the next address.
4184 * i.e. the address returned will be chunk-aligned and after
4185 * any data that is in the chunk containing 's'.
4186 */
4188 {
4196 }
4198 /* Calculate the first device-address that could contain
4199 * any block from the chunk that includes the array-address 's'.
4200 * This too will be the start of a chunk
4201 */
4203 {
4210 }
4214 {
4215 /* We simply copy at most one chunk (smallest of old and new)
4216 * at a time, possibly less if that exceeds RESYNC_PAGES,
4217 * or we hit a bad block or something.
4218 * This might mean we pause for normal IO in the middle of
4219 * a chunk, but that is not a problem was mddev->reshape_position
4220 * can record any location.
4221 *
4222 * If we will want to write to a location that isn't
4223 * yet recorded as 'safe' (i.e. in metadata on disk) then
4224 * we need to flush all reshape requests and update the metadata.
4225 *
4226 * When reshaping forwards (e.g. to more devices), we interpret
4227 * 'safe' as the earliest block which might not have been copied
4228 * down yet. We divide this by previous stripe size and multiply
4229 * by previous stripe length to get lowest device offset that we
4230 * cannot write to yet.
4231 * We interpret 'sector_nr' as an address that we want to write to.
4232 * From this we use last_device_address() to find where we might
4233 * write to, and first_device_address on the 'safe' position.
4234 * If this 'next' write position is after the 'safe' position,
4235 * we must update the metadata to increase the 'safe' position.
4236 *
4237 * When reshaping backwards, we round in the opposite direction
4238 * and perform the reverse test: next write position must not be
4239 * less than current safe position.
4240 *
4241 * In all this the minimum difference in data offsets
4242 * (conf->offset_diff - always positive) allows a bit of slack,
4243 * so next can be after 'safe', but not by more than offset_disk
4244 *
4245 * We need to prepare all the bios here before we start any IO
4246 * to ensure the size we choose is acceptable to all devices.
4247 * The means one for each copy for write-out and an extra one for
4248 * read-in.
4249 * We store the read-in bio in ->master_bio and the others in
4250 * ->devs[x].bio and ->devs[x].repl_bio.
4251 */
4265 /* If restarting in the middle, skip the initial sectors */
4278 }
4279 }
4281 /* We don't use sector_nr to track where we are up to
4282 * as that doesn't work well for ->reshape_backwards.
4283 * So just use ->reshape_progress.
4284 */
4286 /* 'next' is the earliest device address that we might
4287 * write to for this chunk in the new layout
4288 */
4292 /* 'safe' is the last device address that we might read from
4293 * in the old layout after a restart
4294 */
4307 /* 'next' is after the last device address that we
4308 * might write to for this chunk in the new layout
4309 */
4312 /* 'safe' is the earliest device address that we might
4313 * read from in the old layout after a restart
4314 */
4317 /* Need to update metadata if 'next' might be beyond 'safe'
4318 * as that would possibly corrupt data
4319 */
4329 }
4333 /* Need to update reshape_position in metadata */
4339 else
4348 }
4350 read_more:
4351 /* Now schedule reads for blocks from sector_nr to last */
4363 /* Cannot read from here, so need to record bad blocks
4364 * on all the target devices.
4365 */
4366 // FIXME
4369 }
4386 /* Now find the locations in the new layout */
4402 }
4414 }
4416 /* Now add as many pages as possible to all of these bios. */
4429 /* Didn't fit, must stop */
4433 /* Remove last page from this bio */
4437 }
4439 }
4442 }
4443 bio_full:
4446 /* Now submit the read */
4456 /* Now that we have done the whole section we can
4457 * update reshape_progress
4458 */
4461 else
4465 }
4471 {
4472 /* Reshape read completed. Hopefully we have a block
4473 * to write out.
4474 * If we got a read error then we do sync 1-page reads from
4475 * elsewhere until we find the data - or give up.
4476 */
4482 /* Reshape has been aborted */
4485 }
4487 /* We definitely have the data in the pages, schedule the
4488 * writes.
4489 */
4501 }
4509 }
4511 }
4514 {
4525 /* read-ahead size must cover two whole stripes, which is
4526 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4527 */
4534 }
4536 }
4541 {
4542 /* Use sync reads to get the blocks from somewhere else */
4568 sector_t addr;
4576 addr,
4582 failed:
4583 slot++;
4588 }
4590 /* couldn't read this block, must give up */
4594 }
4596 idx++;
4597 }
4599 }
4602 {
4618 }
4621 /* FIXME should record badblock */
4623 }
4627 }
4630 {
4636 }
4639 {
4651 }
4659 d++) {
4666 }
4667 }
4673 }
4676 {
4696 };
4699 {
4701 }
4704 {
4706 }