| blob | be1a9fca3b2d2ade369359d109d1a53ddf30d077 |
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 * use_far_sets_bugfixed (stored in bit 18 of layout )
43 *
44 * The data to be stored is divided into chunks using chunksize. Each device
45 * is divided into far_copies sections. In each section, chunks are laid out
46 * in a style similar to raid0, but near_copies copies of each chunk is stored
47 * (each on a different drive). The starting device for each section is offset
48 * near_copies from the starting device of the previous section. Thus there
49 * are (near_copies * far_copies) of each chunk, and each is on a different
50 * drive. near_copies and far_copies must be at least one, and their product
51 * is at most raid_disks.
52 *
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of being very far
55 * apart on disk, there are adjacent stripes.
56 *
57 * The far and offset algorithms are handled slightly differently if
58 * 'use_far_sets' is true. In this case, the array's devices are grouped into
59 * sets that are (near_copies * far_copies) in size. The far copied stripes
60 * are still shifted by 'near_copies' devices, but this shifting stays confined
61 * to the set rather than the entire array. This is done to improve the number
62 * of device combinations that can fail without causing the array to fail.
63 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
64 * on a device):
65 * A B C D A B C D E
66 * ... ...
67 * D A B C E A B C D
68 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69 * [A B] [C D] [A B] [C D E]
70 * |...| |...| |...| | ... |
71 * [B A] [D C] [B A] [E C D]
72 */
74 /*
75 * Number of guaranteed r10bios in case of extreme VM load:
76 */
77 #define NR_RAID10_BIOS 256
79 /* when we get a read error on a read-only array, we redirect to another
80 * device without failing the first device, or trying to over-write to
81 * correct the read error. To keep track of bad blocks on a per-bio
82 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 */
84 #define IO_BLOCKED ((struct bio *)1)
85 /* When we successfully write to a known bad-block, we need to remove the
86 * bad-block marking which must be done from process context. So we record
87 * the success by setting devs[n].bio to IO_MADE_GOOD
88 */
89 #define IO_MADE_GOOD ((struct bio *)2)
91 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
93 /* When there are this many requests queued to be written by
94 * the raid10 thread, we become 'congested' to provide back-pressure
95 * for writeback.
96 */
109 {
113 /* allocate a r10bio with room for raid_disks entries in the
114 * bios array */
116 }
119 {
121 }
123 /* Maximum size of each resync request */
124 #define RESYNC_BLOCK_SIZE (64*1024)
125 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
126 /* amount of memory to reserve for resync requests */
127 #define RESYNC_WINDOW (1024*1024)
128 /* maximum number of concurrent requests, memory permitting */
129 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
131 /*
132 * When performing a resync, we need to read and compare, so
133 * we need as many pages are there are copies.
134 * When performing a recovery, we need 2 bios, one for read,
135 * one for write (we recover only one drive per r10buf)
136 *
137 */
139 {
154 else
157 /*
158 * Allocate bios.
159 */
171 }
172 /*
173 * Allocate RESYNC_PAGES data pages and attach them
174 * where needed.
175 */
182 /* we can share bv_page's during recovery
183 * and reshape */
195 }
196 }
200 out_free_pages:
207 out_free_bio:
213 }
216 }
219 {
231 }
233 }
237 }
239 }
242 {
254 }
255 }
258 {
263 }
266 {
272 }
275 {
285 /* wake up frozen array... */
289 }
291 /*
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
294 * cache layer.
295 */
297 {
314 /*
315 * Wake up any possible resync thread that waits for the device
316 * to go idle.
317 */
319 }
321 }
323 /*
324 * Update disk head position estimator based on IRQ completion info.
325 */
327 {
332 }
334 /*
335 * Find the disk number which triggered given bio
336 */
339 {
349 }
350 }
360 }
363 {
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 */
398 }
403 /*
404 * oops, read error - keep the refcount on the rdev
405 */
414 }
415 }
418 {
419 /* clear the bitmap if all writes complete successfully */
425 }
428 {
436 else
438 }
439 }
440 }
443 {
459 }
460 /*
461 * this branch is our 'one mirror IO has finished' event handler:
462 */
465 /* Never record new bad blocks to replacement,
466 * just fail it.
467 */
476 }
478 /*
479 * Set R10BIO_Uptodate in our master bio, so that
480 * we will return a good error code for to the higher
481 * levels even if IO on some other mirrored buffer fails.
482 *
483 * The 'master' represents the composite IO operation to
484 * user-side. So if something waits for IO, then it will
485 * wait for the 'master' bio.
486 */
487 sector_t first_bad;
490 /*
491 * Do not set R10BIO_Uptodate if the current device is
492 * rebuilding or Faulty. This is because we cannot use
493 * such device for properly reading the data back (we could
494 * potentially use it, if the current write would have felt
495 * before rdev->recovery_offset, but for simplicity we don't
496 * check this here.
497 */
502 /* Maybe we can clear some bad blocks. */
510 else
514 }
515 }
517 /*
518 *
519 * Let's see if all mirrored write operations have finished
520 * already.
521 */
525 }
527 /*
528 * RAID10 layout manager
529 * As well as the chunksize and raid_disks count, there are two
530 * parameters: near_copies and far_copies.
531 * near_copies * far_copies must be <= raid_disks.
532 * Normally one of these will be 1.
533 * If both are 1, we get raid0.
534 * If near_copies == raid_disks, we get raid1.
535 *
536 * Chunks are laid out in raid0 style with near_copies copies of the
537 * first chunk, followed by near_copies copies of the next chunk and
538 * so on.
539 * If far_copies > 1, then after 1/far_copies of the array has been assigned
540 * as described above, we start again with a device offset of near_copies.
541 * So we effectively have another copy of the whole array further down all
542 * the drives, but with blocks on different drives.
543 * With this layout, and block is never stored twice on the one device.
544 *
545 * raid10_find_phys finds the sector offset of a given virtual sector
546 * on each device that it is on.
547 *
548 * raid10_find_virt does the reverse mapping, from a device and a
549 * sector offset to a virtual address
550 */
553 {
555 sector_t sector;
556 sector_t chunk;
557 sector_t stripe;
568 /* now calculate first sector/dev */
580 /* and calculate all the others */
587 slot++;
601 }
605 slot++;
606 }
607 dev++;
611 }
612 }
613 }
616 {
628 }
631 {
633 /* Never use conf->prev as this is only called during resync
634 * or recovery, so reshape isn't happening
635 */
649 }
650 }
665 else
667 }
669 }
673 }
675 /*
676 * This routine returns the disk from which the requested read should
677 * be done. There is a per-array 'next expected sequential IO' sector
678 * number - if this matches on the next IO then we use the last disk.
679 * There is also a per-disk 'last know head position' sector that is
680 * maintained from IRQ contexts, both the normal and the resync IO
681 * completion handlers update this position correctly. If there is no
682 * perfect sequential match then we pick the disk whose head is closest.
683 *
684 * If there are 2 mirrors in the same 2 devices, performance degrades
685 * because position is mirror, not device based.
686 *
687 * The rdev for the device selected will have nr_pending incremented.
688 */
690 /*
691 * FIXME: possibly should rethink readbalancing and do it differently
692 * depending on near_copies / far_copies geometry.
693 */
697 {
716 /*
717 * Check if we can balance. We can balance on the whole
718 * device if no resync is going on (recovery is ok), or below
719 * the resync window. We take the first readable disk when
720 * above the resync window.
721 */
727 sector_t first_bad;
729 sector_t dev_sector;
749 /* Already have a better slot */
752 /* Cannot read here. If this is the
753 * 'primary' device, then we must not read
754 * beyond 'bad_sectors' from another device.
755 */
762 sector_t good_sectors =
768 }
770 /* Must read from here */
772 }
780 /* This optimisation is debatable, and completely destroys
781 * sequential read speed for 'far copies' arrays. So only
782 * keep it for 'near' arrays, and review those later.
783 */
787 /* for far > 1 always use the lowest address */
790 else
797 }
798 }
802 }
813 }
816 {
828 i++) {
834 }
835 }
838 }
841 {
842 /* Any writes that have been queued but are awaiting
843 * bitmap updates get flushed here.
844 */
852 /* flush any pending bitmap writes to disk
853 * before proceeding w/ I/O */
862 /* Just ignore it */
864 else
867 }
870 }
872 /* Barriers....
873 * Sometimes we need to suspend IO while we do something else,
874 * either some resync/recovery, or reconfigure the array.
875 * To do this we raise a 'barrier'.
876 * The 'barrier' is a counter that can be raised multiple times
877 * to count how many activities are happening which preclude
878 * normal IO.
879 * We can only raise the barrier if there is no pending IO.
880 * i.e. if nr_pending == 0.
881 * We choose only to raise the barrier if no-one is waiting for the
882 * barrier to go down. This means that as soon as an IO request
883 * is ready, no other operations which require a barrier will start
884 * until the IO request has had a chance.
885 *
886 * So: regular IO calls 'wait_barrier'. When that returns there
887 * is no backgroup IO happening, It must arrange to call
888 * allow_barrier when it has finished its IO.
889 * backgroup IO calls must call raise_barrier. Once that returns
890 * there is no normal IO happeing. It must arrange to call
891 * lower_barrier when the particular background IO completes.
892 */
895 {
899 /* Wait until no block IO is waiting (unless 'force') */
903 /* block any new IO from starting */
906 /* Now wait for all pending IO to complete */
912 }
915 {
921 }
924 {
928 /* Wait for the barrier to drop.
929 * However if there are already pending
930 * requests (preventing the barrier from
931 * rising completely), and the
932 * pre-process bio queue isn't empty,
933 * then don't wait, as we need to empty
934 * that queue to get the nr_pending
935 * count down.
936 */
946 }
949 }
952 {
956 }
959 {
960 /* stop syncio and normal IO and wait for everything to
961 * go quiet.
962 * We increment barrier and nr_waiting, and then
963 * wait until nr_pending match nr_queued+extra
964 * This is called in the context of one normal IO request
965 * that has failed. Thus any sync request that might be pending
966 * will be blocked by nr_pending, and we need to wait for
967 * pending IO requests to complete or be queued for re-try.
968 * Thus the number queued (nr_queued) plus this request (extra)
969 * must match the number of pending IOs (nr_pending) before
970 * we continue.
971 */
983 }
986 {
987 /* reverse the effect of the freeze */
993 }
997 {
1001 else
1003 }
1009 };
1012 {
1014 cb);
1028 }
1030 /* we aren't scheduling, so we can do the write-out directly. */
1040 /* Just ignore it */
1042 else
1045 }
1047 }
1050 {
1069 /*
1070 * Register the new request and wait if the reconstruction
1071 * thread has put up a bar for new requests.
1072 * Continue immediately if no resync is active currently.
1073 */
1080 /* IO spans the reshape position. Need to wait for
1081 * reshape to pass
1082 */
1087 sectors);
1089 }
1097 /* Need to update reshape_position in metadata */
1106 }
1117 /* We might need to issue multiple reads to different
1118 * devices if there are bad blocks around, so we keep
1119 * track of the number of reads in bio->bi_phys_segments.
1120 * If this is 0, there is only one r10_bio and no locking
1121 * will be needed when the request completes. If it is
1122 * non-zero, then it is the number of not-completed requests.
1123 */
1128 /*
1129 * read balancing logic:
1130 */
1134 read_again:
1139 }
1144 max_sectors);
1157 /* Could not read all from this device, so we will
1158 * need another r10_bio.
1159 */
1166 else
1169 /* Cannot call generic_make_request directly
1170 * as that will be queued in __generic_make_request
1171 * and subsequent mempool_alloc might block
1172 * waiting for it. so hand bio over to raid10d.
1173 */
1183 sectors_handled;
1188 }
1190 /*
1191 * WRITE:
1192 */
1197 }
1198 /* first select target devices under rcu_lock and
1199 * inc refcount on their rdev. Record them by setting
1200 * bios[x] to bio
1201 * If there are known/acknowledged bad blocks on any device
1202 * on which we have seen a write error, we want to avoid
1203 * writing to those blocks. This potentially requires several
1204 * writes to write around the bad blocks. Each set of writes
1205 * gets its own r10_bio with a set of bios attached. The number
1206 * of r10_bios is recored in bio->bi_phys_segments just as with
1207 * the read case.
1208 */
1212 retry_write:
1228 }
1233 }
1245 }
1247 sector_t first_bad;
1253 max_sectors,
1256 /* Mustn't write here until the bad block
1257 * is acknowledged
1258 */
1263 }
1265 /* Cannot write here at all */
1268 /* Mustn't write more than bad_sectors
1269 * to other devices yet
1270 */
1272 /* We don't set R10BIO_Degraded as that
1273 * only applies if the disk is missing,
1274 * so it might be re-added, and we want to
1275 * know to recover this chunk.
1276 * In this case the device is here, and the
1277 * fact that this chunk is not in-sync is
1278 * recorded in the bad block log.
1279 */
1281 }
1286 }
1287 }
1291 }
1295 }
1296 }
1300 /* Have to wait for this device to get unblocked, then retry */
1308 }
1314 /* Race with remove_disk */
1317 }
1319 }
1320 }
1325 }
1328 /* We are splitting this into multiple parts, so
1329 * we need to prepare for allocating another r10_bio.
1330 */
1335 else
1338 }
1352 max_sectors);
1357 rdev));
1369 cb);
1370 else
1379 }
1383 }
1388 /* Replacement just got moved to main 'rdev' */
1391 }
1394 max_sectors);
1412 }
1413 }
1415 /* Don't remove the bias on 'remaining' (one_write_done) until
1416 * after checking if we need to go around again.
1417 */
1421 /* We need another r10_bio. It has already been counted
1422 * in bio->bi_phys_segments.
1423 */
1433 }
1435 }
1438 {
1448 }
1452 /*
1453 * If this request crosses a chunk boundary, we need to split
1454 * it.
1455 */
1468 }
1473 /* In case raid10d snuck in to freeze_array */
1475 }
1478 {
1489 else
1493 }
1500 }
1503 }
1505 /* check if there are enough drives for
1506 * every block to appear on atleast one.
1507 * Don't consider the device numbered 'ignore'
1508 * as we might be about to remove it.
1509 */
1511 {
1521 }
1533 cnt++;
1535 }
1541 out:
1544 }
1547 {
1548 /* when calling 'enough', both 'prev' and 'geo' must
1549 * be stable.
1550 * This is ensured if ->reconfig_mutex or ->device_lock
1551 * is held.
1552 */
1555 }
1558 {
1563 /*
1564 * If it is not operational, then we have already marked it as dead
1565 * else if it is the last working disks, ignore the error, let the
1566 * next level up know.
1567 * else mark the drive as failed
1568 */
1572 /*
1573 * Don't fail the drive, just return an IO error.
1574 */
1577 }
1580 /*
1581 * If recovery is running, make sure it aborts.
1582 */
1594 }
1597 {
1605 }
1609 /* This is only called with ->reconfix_mutex held, so
1610 * rcu protection of rdev is not needed */
1619 }
1620 }
1623 {
1629 }
1632 {
1639 /*
1640 * Find all non-in_sync disks within the RAID10 configuration
1641 * and mark them in_sync
1642 */
1649 /* Replacement has just become active */
1652 count++;
1654 /* Replaced device not technically faulty,
1655 * but we need to be sure it gets removed
1656 * and never re-added.
1657 */
1661 }
1667 count++;
1669 }
1670 }
1677 }
1680 {
1688 /* only hot-add to in-sync arrays, as recovery is
1689 * very different from resync
1690 */
1704 else
1724 }
1738 }
1744 }
1747 {
1759 else
1766 }
1767 /* Only remove non-faulty devices if recovery
1768 * is not possible.
1769 */
1777 }
1782 /* lost the race, try later */
1786 }
1787 }
1789 /* We must have just cleared 'rdev' */
1793 * but will never see neither -- if they are careful.
1794 */
1798 /* We might have just remove the Replacement as faulty
1799 * Clear the flag just in case
1800 */
1805 abort:
1809 }
1812 {
1818 /* this is a reshape read */
1825 else
1826 /* The write handler will notice the lack of
1827 * R10BIO_Uptodate and record any errors etc
1828 */
1832 /* for reconstruct, we always reschedule after a read.
1833 * for resync, only after all reads
1834 */
1838 /* we have read all the blocks,
1839 * do the comparison in process context in raid10d
1840 */
1842 }
1843 }
1846 {
1851 /* the primary of several recovery bios */
1856 else
1865 else
1868 }
1869 }
1870 }
1873 {
1878 sector_t first_bad;
1887 else
1899 }
1909 }
1911 /*
1912 * Note: sync and recover and handled very differently for raid10
1913 * This code is for resync.
1914 * For resync, we read through virtual addresses and read all blocks.
1915 * If there is any error, we schedule a write. The lowest numbered
1916 * drive is authoritative.
1917 * However requests come for physical address, so we need to map.
1918 * For every physical address there are raid_disks/copies virtual addresses,
1919 * which is always are least one, but is not necessarly an integer.
1920 * This means that a physical address can span multiple chunks, so we may
1921 * have to submit multiple io requests for a single sync request.
1922 */
1923 /*
1924 * We check if all blocks are in-sync and only write to blocks that
1925 * aren't in sync
1926 */
1928 {
1936 /* find the first device with a block */
1950 /* now find blocks with errors */
1961 /* We know that the bi_io_vec layout is the same for
1962 * both 'first' and 'i', so we just compare them.
1963 * All vec entries are PAGE_SIZE;
1964 */
1972 len))
1975 }
1980 /* Don't fix anything. */
1982 }
1983 /* Ok, we need to write this bio, either to correct an
1984 * inconsistency or to correct an unreadable block.
1985 * First we need to fixup bv_offset, bv_len and
1986 * bi_vecs, as the read request might have corrupted these
1987 */
2007 }
2009 /* Now write out to any replacement devices
2010 * that are active
2011 */
2026 }
2028 done:
2032 }
2033 }
2035 /*
2036 * Now for the recovery code.
2037 * Recovery happens across physical sectors.
2038 * We recover all non-is_sync drives by finding the virtual address of
2039 * each, and then choose a working drive that also has that virt address.
2040 * There is a separate r10_bio for each non-in_sync drive.
2041 * Only the first two slots are in use. The first for reading,
2042 * The second for writing.
2043 *
2044 */
2046 {
2047 /* We got a read error during recovery.
2048 * We repeat the read in smaller page-sized sections.
2049 * If a read succeeds, write it to the new device or record
2050 * a bad block if we cannot.
2051 * If a read fails, record a bad block on both old and
2052 * new devices.
2053 */
2066 sector_t addr;
2075 addr,
2083 addr,
2093 }
2094 }
2096 /* We don't worry if we cannot set a bad block -
2097 * it really is bad so there is no loss in not
2098 * recording it yet
2099 */
2103 /* need bad block on destination too */
2108 /* just abort the recovery */
2110 "md/raid10:%s: recovery aborted"
2119 }
2120 }
2121 }
2125 idx++;
2126 }
2127 }
2130 {
2139 }
2141 /*
2142 * share the pages with the first bio
2143 * and submit the write request
2144 */
2148 /* Need to test wbio2->bi_end_io before we call
2149 * generic_make_request as if the former is NULL,
2150 * the latter is free to free wbio2.
2151 */
2158 }
2164 }
2165 }
2167 /*
2168 * Used by fix_read_error() to decay the per rdev read_errors.
2169 * We halve the read error count for every hour that has elapsed
2170 * since the last recorded read error.
2171 *
2172 */
2174 {
2182 /* first time we've seen a read error */
2185 }
2192 /*
2193 * if hours_since_last is > the number of bits in read_errors
2194 * just set read errors to 0. We do this to avoid
2195 * overflowing the shift of read_errors by hours_since_last.
2196 */
2199 else
2201 }
2205 {
2206 sector_t first_bad;
2213 /* success */
2220 }
2221 /* need to record an error - either for the block or the device */
2225 }
2227 /*
2228 * This is a kernel thread which:
2229 *
2230 * 1. Retries failed read operations on working mirrors.
2231 * 2. Updates the raid superblock when problems encounter.
2232 * 3. Performs writes following reads for array synchronising.
2233 */
2236 {
2243 /* still own a reference to this rdev, so it cannot
2244 * have been cleared recently.
2245 */
2249 /* drive has already been failed, just ignore any
2250 more fix_read_error() attempts */
2260 "md/raid10:%s: %s: Raid device exceeded "
2270 }
2283 sector_t first_bad;
2297 sect,
2305 }
2306 sl++;
2313 /* Cannot read from anywhere, just mark the block
2314 * as bad on the first device to discourage future
2315 * reads.
2316 */
2321 rdev,
2328 }
2330 }
2333 /* write it back and re-read */
2340 sl--;
2352 sect,
2355 /* Well, this device is dead */
2357 "md/raid10:%s: read correction "
2358 "write failed"
2362 sect +
2364 rdev)),
2370 }
2373 }
2380 sl--;
2392 sect,
2394 READ)) {
2396 /* Well, this device is dead */
2398 "md/raid10:%s: unable to read back "
2399 "corrected sectors"
2403 sect +
2413 "md/raid10:%s: read error corrected"
2417 sect +
2421 }
2425 }
2430 }
2431 }
2434 {
2439 /* bio has the data to be written to slot 'i' where
2440 * we just recently had a write error.
2441 * We repeatedly clone the bio and trim down to one block,
2442 * then try the write. Where the write fails we record
2443 * a bad block.
2444 * It is conceivable that the bio doesn't exactly align with
2445 * blocks. We must handle this.
2446 *
2447 * We currently own a reference to the rdev.
2448 */
2451 sector_t sector;
2468 sector_t wsector;
2471 /* Write at 'sector' for 'sectors' */
2481 /* Failure! */
2490 }
2492 }
2495 {
2504 /* we got a read error. Maybe the drive is bad. Maybe just
2505 * the block and we can fix it.
2506 * We freeze all other IO, and try reading the block from
2507 * other devices. When we find one, we re-write
2508 * and check it that fixes the read error.
2509 * This is all done synchronously while the array is
2510 * frozen.
2511 */
2526 read_more:
2535 }
2540 KERN_ERR
2541 "md/raid10:%s: %s: redirecting "
2558 /* Drat - have to split this up more */
2567 else
2573 GFP_NOIO);
2586 }
2589 {
2590 /* Some sort of write request has finished and it
2591 * succeeded in writing where we thought there was a
2592 * bad block. So forget the bad block.
2593 * Or possibly if failed and we need to record
2594 * a bad block.
2595 */
2608 rdev,
2613 rdev,
2617 }
2624 rdev,
2629 rdev,
2633 }
2634 }
2644 rdev,
2654 }
2656 }
2661 rdev,
2665 }
2666 }
2678 }
2679 }
2680 }
2683 {
2701 }
2702 }
2706 retry_list);
2715 }
2716 }
2727 }
2747 /* just a partial read to be scheduled from a
2748 * separate context
2749 */
2752 }
2757 }
2759 }
2762 {
2777 }
2779 /*
2780 * perform a "sync" on one "block"
2781 *
2782 * We need to make sure that no normal I/O request - particularly write
2783 * requests - conflict with active sync requests.
2784 *
2785 * This is achieved by tracking pending requests and a 'barrier' concept
2786 * that can be installed to exclude normal IO requests.
2787 *
2788 * Resync and recovery are handled very differently.
2789 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2790 *
2791 * For resync, we iterate over virtual addresses, read all copies,
2792 * and update if there are differences. If only one copy is live,
2793 * skip it.
2794 * For recovery, we iterate over physical addresses, read a good
2795 * value for each non-in_sync drive, and over-write.
2796 *
2797 * So, for recovery we may have several outstanding complex requests for a
2798 * given address, one for each out-of-sync device. We model this by allocating
2799 * a number of r10_bio structures, one for each out-of-sync device.
2800 * As we setup these structures, we collect all bio's together into a list
2801 * which we then process collectively to add pages, and then process again
2802 * to pass to generic_make_request.
2803 *
2804 * The r10_bio structures are linked using a borrowed master_bio pointer.
2805 * This link is counted in ->remaining. When the r10_bio that points to NULL
2806 * has its remaining count decremented to 0, the whole complex operation
2807 * is complete.
2808 *
2809 */
2813 {
2820 sector_t sync_blocks;
2829 /*
2830 * Allow skipping a full rebuild for incremental assembly
2831 * of a clean array, like RAID1 does.
2832 */
2842 }
2844 skipped:
2850 /* If we aborted, we need to abort the
2851 * sync on the 'current' bitmap chucks (there can
2852 * be several when recovering multiple devices).
2853 * as we may have started syncing it but not finished.
2854 * We can find the current address in
2855 * mddev->curr_resync, but for recovery,
2856 * we need to convert that to several
2857 * virtual addresses.
2858 */
2863 }
2870 sector_t sect =
2874 }
2876 /* completed sync */
2880 /* Completed a full sync so the replacements
2881 * are now fully recovered.
2882 */
2889 }
2891 }
2893 }
2898 }
2904 /* if there has been nothing to do on any drive,
2905 * then there is nothing to do at all..
2906 */
2909 }
2914 /* make sure whole request will fit in a chunk - if chunks
2915 * are meaningful
2916 */
2921 /*
2922 * If there is non-resync activity waiting for a turn, then let it
2923 * though before starting on this new sync request.
2924 */
2928 /* Again, very different code for resync and recovery.
2929 * Both must result in an r10bio with a list of bios that
2930 * have bi_end_io, bi_sector, bi_bdev set,
2931 * and bi_private set to the r10bio.
2932 * For recovery, we may actually create several r10bios
2933 * with 2 bios in each, that correspond to the bios in the main one.
2934 * In this case, the subordinate r10bios link back through a
2935 * borrowed master_bio pointer, and the counter in the master
2936 * includes a ref from each subordinate.
2937 */
2938 /* First, we decide what to do and set ->bi_end_io
2939 * To end_sync_read if we want to read, and
2940 * end_sync_write if we will want to write.
2941 */
2945 /* recovery... the complicated one */
2952 sector_t sect;
2969 }
2972 /* want to reconstruct this device */
2976 /* last stripe is not complete - don't
2977 * try to recover this sector.
2978 */
2981 }
2984 /* Unless we are doing a full sync, or a replacement
2985 * we only need to recover the block if it is set in
2986 * the bitmap
2987 */
2995 /* yep, skip the sync_blocks here, but don't assume
2996 * that there will never be anything to do here
2997 */
3001 }
3021 /* Need to check if the array will still be
3022 * degraded
3023 */
3031 }
3032 }
3049 /* This is where we read from */
3058 bad_sectors -= (sector
3063 }
3064 }
3077 /* and we write to 'i' (if not in_sync) */
3104 /* and maybe write to replacement */
3108 /* Note: if mreplace != NULL, then bio
3109 * cannot be NULL as r10buf_pool_alloc will
3110 * have allocated it.
3111 * So the second test here is pointless.
3112 * But it keeps semantic-checkers happy, and
3113 * this comment keeps human reviewers
3114 * happy.
3115 */
3130 }
3133 /* Cannot recover, so abort the recovery or
3134 * record a bad block */
3136 /* problem is that there are bad blocks
3137 * on other device(s)
3138 */
3146 mrdev,
3152 mreplace,
3156 }
3163 mirror->recovery_disabled
3165 }
3174 }
3178 }
3185 }
3187 }
3189 /* resync. Schedule a read for every block at this virt offset */
3198 /* We can skip this block */
3201 }
3235 }
3247 }
3248 }
3258 count++;
3264 }
3268 /* Need to set up for writing to the replacement */
3281 count++;
3282 }
3289 mddev);
3294 mddev);
3295 }
3299 }
3300 }
3318 /* stop here */
3323 /* remove last page from this bio */
3327 }
3329 }
3333 bio_full:
3348 }
3349 }
3352 /* pretend they weren't skipped, it makes
3353 * no important difference in this case
3354 */
3358 giveup:
3359 /* There is nowhere to write, so all non-sync
3360 * drives must be failed or in resync, all drives
3361 * have a bad block, so try the next chunk...
3362 */
3367 chunks_skipped ++;
3370 }
3372 static sector_t
3374 {
3375 sector_t size;
3390 }
3393 {
3394 /* Calculate the number of sectors-per-device that will
3395 * actually be used, and set conf->dev_sectors and
3396 * conf->stride
3397 */
3403 /* 'size' is now the number of chunks in the array */
3404 /* calculate "used chunks per device" */
3407 /* We need to round up when dividing by raid_disks to
3408 * get the stride size.
3409 */
3419 }
3420 }
3424 {
3440 * updated yet. */
3445 }
3463 * actually using this, but leave code here just in case.*/
3473 }
3477 }
3480 {
3493 }
3499 }
3506 /* FIXME calc properly */
3509 GFP_KERNEL);
3532 }
3536 else
3537 /* far_copies must be 1 */
3539 }
3556 out:
3565 }
3567 }
3570 {
3575 sector_t size;
3585 }
3601 else
3604 }
3626 }
3644 }
3650 else
3653 }
3654 /* need to check that every block has at least one working mirror */
3659 }
3662 /* must ensure that shape change is supported */
3669 }
3675 i++) {
3680 /* The replacement is all we have - use it */
3684 }
3693 }
3695 }
3705 /*
3706 * Ok, everything is just fine now
3707 */
3717 /* Calculate max read-ahead size.
3718 * We need to readahead at least twice a whole stripe....
3719 * maybe...
3720 */
3724 }
3738 /* This cannot work */
3741 }
3750 }
3754 out_free_conf:
3761 out:
3763 }
3766 {
3775 }
3778 {
3788 }
3789 }
3792 {
3793 /* Resize of 'far' arrays is not supported.
3794 * For 'near' and 'offset' arrays we can set the
3795 * number of sectors used to be an appropriate multiple
3796 * of the chunk size.
3797 * For 'offset', this is far_copies*chunksize.
3798 * For 'near' the multiplier is the LCM of
3799 * near_copies and raid_disks.
3800 * So if far_copies > 1 && !far_offset, fail.
3801 * Else find LCM(raid_disks, near_copy)*far_copies and
3802 * multiply by chunk_size. Then round to this number.
3803 * This is mostly done by raid10_size()
3804 */
3823 }
3828 }
3833 }
3838 }
3841 {
3849 }
3852 /* Set new parameters */
3854 /* new layout: far_copies = 1, near_copies = 2 */
3859 /* make sure it will be not marked as dirty */
3869 }
3871 }
3874 }
3877 {
3880 /* raid10 can take over:
3881 * raid0 - providing it has only two drives
3882 */
3884 /* for raid0 takeover only one zone is supported */
3891 }
3895 }
3897 }
3900 {
3901 /* Called when there is a request to change
3902 * - layout (to ->new_layout)
3903 * - chunk size (to ->new_chunk_sectors)
3904 * - raid_disks (by delta_disks)
3905 * or when trying to restart a reshape that was ongoing.
3906 *
3907 * We need to validate the request and possibly allocate
3908 * space if that might be an issue later.
3909 *
3910 * Currently we reject any reshape of a 'far' mode array,
3911 * allow chunk size to change if new is generally acceptable,
3912 * allow raid_disks to increase, and allow
3913 * a switch between 'near' mode and 'offset' mode.
3914 */
3922 /* mustn't change number of copies */
3925 /* Cannot switch to 'far' mode */
3929 /* not factor of array size */
3938 /* allocate new 'mirrors' list */
3943 GFP_KERNEL);
3946 }
3948 }
3950 /*
3951 * Need to check if array has failed when deciding whether to:
3952 * - start an array
3953 * - remove non-faulty devices
3954 * - add a spare
3955 * - allow a reshape
3956 * This determination is simple when no reshape is happening.
3957 * However if there is a reshape, we need to carefully check
3958 * both the before and after sections.
3959 * This is because some failed devices may only affect one
3960 * of the two sections, and some non-in_sync devices may
3961 * be insync in the section most affected by failed devices.
3962 */
3964 {
3970 /* 'prev' section first */
3974 degraded++;
3976 /* When we can reduce the number of devices in
3977 * an array, this might not contribute to
3978 * 'degraded'. It does now.
3979 */
3980 degraded++;
3981 }
3990 degraded2++;
3992 /* If reshape is increasing the number of devices,
3993 * this section has already been recovered, so
3994 * it doesn't contribute to degraded.
3995 * else it does.
3996 */
3998 degraded2++;
3999 }
4000 }
4005 }
4008 {
4009 /* A 'reshape' has been requested. This commits
4010 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4011 * This also checks if there are enough spares and adds them
4012 * to the array.
4013 * We currently require enough spares to make the final
4014 * array non-degraded. We also require that the difference
4015 * between old and new data_offset - on each device - is
4016 * enough that we never risk over-writing.
4017 */
4042 spares++;
4052 }
4053 }
4071 }
4081 }
4096 }
4105 else
4110 }
4113 /* This is a spare that was manually added */
4115 }
4116 }
4117 /* When a reshape changes the number of devices,
4118 * ->degraded is measured against the larger of the
4119 * pre and post numbers.
4120 */
4139 }
4145 abort:
4158 }
4160 /* Calculate the last device-address that could contain
4161 * any block from the chunk that includes the array-address 's'
4162 * and report the next address.
4163 * i.e. the address returned will be chunk-aligned and after
4164 * any data that is in the chunk containing 's'.
4165 */
4167 {
4175 }
4177 /* Calculate the first device-address that could contain
4178 * any block from the chunk that includes the array-address 's'.
4179 * This too will be the start of a chunk
4180 */
4182 {
4189 }
4193 {
4194 /* We simply copy at most one chunk (smallest of old and new)
4195 * at a time, possibly less if that exceeds RESYNC_PAGES,
4196 * or we hit a bad block or something.
4197 * This might mean we pause for normal IO in the middle of
4198 * a chunk, but that is not a problem as mddev->reshape_position
4199 * can record any location.
4200 *
4201 * If we will want to write to a location that isn't
4202 * yet recorded as 'safe' (i.e. in metadata on disk) then
4203 * we need to flush all reshape requests and update the metadata.
4204 *
4205 * When reshaping forwards (e.g. to more devices), we interpret
4206 * 'safe' as the earliest block which might not have been copied
4207 * down yet. We divide this by previous stripe size and multiply
4208 * by previous stripe length to get lowest device offset that we
4209 * cannot write to yet.
4210 * We interpret 'sector_nr' as an address that we want to write to.
4211 * From this we use last_device_address() to find where we might
4212 * write to, and first_device_address on the 'safe' position.
4213 * If this 'next' write position is after the 'safe' position,
4214 * we must update the metadata to increase the 'safe' position.
4215 *
4216 * When reshaping backwards, we round in the opposite direction
4217 * and perform the reverse test: next write position must not be
4218 * less than current safe position.
4219 *
4220 * In all this the minimum difference in data offsets
4221 * (conf->offset_diff - always positive) allows a bit of slack,
4222 * so next can be after 'safe', but not by more than offset_diff
4223 *
4224 * We need to prepare all the bios here before we start any IO
4225 * to ensure the size we choose is acceptable to all devices.
4226 * The means one for each copy for write-out and an extra one for
4227 * read-in.
4228 * We store the read-in bio in ->master_bio and the others in
4229 * ->devs[x].bio and ->devs[x].repl_bio.
4230 */
4244 /* If restarting in the middle, skip the initial sectors */
4257 }
4258 }
4260 /* We don't use sector_nr to track where we are up to
4261 * as that doesn't work well for ->reshape_backwards.
4262 * So just use ->reshape_progress.
4263 */
4265 /* 'next' is the earliest device address that we might
4266 * write to for this chunk in the new layout
4267 */
4271 /* 'safe' is the last device address that we might read from
4272 * in the old layout after a restart
4273 */
4286 /* 'next' is after the last device address that we
4287 * might write to for this chunk in the new layout
4288 */
4291 /* 'safe' is the earliest device address that we might
4292 * read from in the old layout after a restart
4293 */
4296 /* Need to update metadata if 'next' might be beyond 'safe'
4297 * as that would possibly corrupt data
4298 */
4308 }
4312 /* Need to update reshape_position in metadata */
4318 else
4328 }
4331 }
4333 read_more:
4334 /* Now schedule reads for blocks from sector_nr to last */
4347 /* Cannot read from here, so need to record bad blocks
4348 * on all the target devices.
4349 */
4350 // FIXME
4354 }
4371 /* Now find the locations in the new layout */
4388 }
4401 }
4403 /* Now add as many pages as possible to all of these bios. */
4416 /* Didn't fit, must stop */
4420 /* Remove last page from this bio */
4424 }
4426 }
4429 }
4430 bio_full:
4434 /* Now submit the read */
4444 /* Now that we have done the whole section we can
4445 * update reshape_progress
4446 */
4449 else
4453 }
4459 {
4460 /* Reshape read completed. Hopefully we have a block
4461 * to write out.
4462 * If we got a read error then we do sync 1-page reads from
4463 * elsewhere until we find the data - or give up.
4464 */
4470 /* Reshape has been aborted */
4473 }
4475 /* We definitely have the data in the pages, schedule the
4476 * writes.
4477 */
4490 }
4494 }
4501 }
4503 }
4506 {
4518 /* read-ahead size must cover two whole stripes, which is
4519 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4520 */
4527 }
4529 }
4533 {
4534 /* Use sync reads to get the blocks from somewhere else */
4561 sector_t addr;
4571 addr,
4579 failed:
4580 slot++;
4585 }
4588 /* couldn't read this block, must give up */
4592 }
4594 idx++;
4595 }
4597 }
4600 {
4615 }
4618 /* FIXME should record badblock */
4620 }
4624 }
4627 {
4633 }
4636 {
4648 }
4653 }
4659 d++) {
4666 }
4668 }
4674 }
4677 {
4698 };
4701 {
4703 }
4706 {
4708 }