| blob | 39fddda2fef2d918863699e3f283f441f6fb52c1 |
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 {
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;
493 /*
494 * Do not set R10BIO_Uptodate if the current device is
495 * rebuilding or Faulty. This is because we cannot use
496 * such device for properly reading the data back (we could
497 * potentially use it, if the current write would have felt
498 * before rdev->recovery_offset, but for simplicity we don't
499 * check this here.
500 */
505 /* Maybe we can clear some bad blocks. */
513 else
517 }
518 }
520 /*
521 *
522 * Let's see if all mirrored write operations have finished
523 * already.
524 */
528 }
530 /*
531 * RAID10 layout manager
532 * As well as the chunksize and raid_disks count, there are two
533 * parameters: near_copies and far_copies.
534 * near_copies * far_copies must be <= raid_disks.
535 * Normally one of these will be 1.
536 * If both are 1, we get raid0.
537 * If near_copies == raid_disks, we get raid1.
538 *
539 * Chunks are laid out in raid0 style with near_copies copies of the
540 * first chunk, followed by near_copies copies of the next chunk and
541 * so on.
542 * If far_copies > 1, then after 1/far_copies of the array has been assigned
543 * as described above, we start again with a device offset of near_copies.
544 * So we effectively have another copy of the whole array further down all
545 * the drives, but with blocks on different drives.
546 * With this layout, and block is never stored twice on the one device.
547 *
548 * raid10_find_phys finds the sector offset of a given virtual sector
549 * on each device that it is on.
550 *
551 * raid10_find_virt does the reverse mapping, from a device and a
552 * sector offset to a virtual address
553 */
556 {
558 sector_t sector;
559 sector_t chunk;
560 sector_t stripe;
571 /* now calculate first sector/dev */
583 /* and calculate all the others */
590 slot++;
604 }
608 slot++;
609 }
610 dev++;
614 }
615 }
616 }
619 {
631 }
634 {
636 /* Never use conf->prev as this is only called during resync
637 * or recovery, so reshape isn't happening
638 */
652 }
653 }
668 else
670 }
672 }
676 }
678 /*
679 * This routine returns the disk from which the requested read should
680 * be done. There is a per-array 'next expected sequential IO' sector
681 * number - if this matches on the next IO then we use the last disk.
682 * There is also a per-disk 'last know head position' sector that is
683 * maintained from IRQ contexts, both the normal and the resync IO
684 * completion handlers update this position correctly. If there is no
685 * perfect sequential match then we pick the disk whose head is closest.
686 *
687 * If there are 2 mirrors in the same 2 devices, performance degrades
688 * because position is mirror, not device based.
689 *
690 * The rdev for the device selected will have nr_pending incremented.
691 */
693 /*
694 * FIXME: possibly should rethink readbalancing and do it differently
695 * depending on near_copies / far_copies geometry.
696 */
700 {
719 /*
720 * Check if we can balance. We can balance on the whole
721 * device if no resync is going on (recovery is ok), or below
722 * the resync window. We take the first readable disk when
723 * above the resync window.
724 */
730 sector_t first_bad;
732 sector_t dev_sector;
752 /* Already have a better slot */
755 /* Cannot read here. If this is the
756 * 'primary' device, then we must not read
757 * beyond 'bad_sectors' from another device.
758 */
765 sector_t good_sectors =
771 }
773 /* Must read from here */
775 }
783 /* This optimisation is debatable, and completely destroys
784 * sequential read speed for 'far copies' arrays. So only
785 * keep it for 'near' arrays, and review those later.
786 */
790 /* for far > 1 always use the lowest address */
793 else
800 }
801 }
805 }
816 }
819 {
831 i++) {
837 }
838 }
841 }
844 {
845 /* Any writes that have been queued but are awaiting
846 * bitmap updates get flushed here.
847 */
855 /* flush any pending bitmap writes to disk
856 * before proceeding w/ I/O */
865 /* Just ignore it */
867 else
870 }
873 }
875 /* Barriers....
876 * Sometimes we need to suspend IO while we do something else,
877 * either some resync/recovery, or reconfigure the array.
878 * To do this we raise a 'barrier'.
879 * The 'barrier' is a counter that can be raised multiple times
880 * to count how many activities are happening which preclude
881 * normal IO.
882 * We can only raise the barrier if there is no pending IO.
883 * i.e. if nr_pending == 0.
884 * We choose only to raise the barrier if no-one is waiting for the
885 * barrier to go down. This means that as soon as an IO request
886 * is ready, no other operations which require a barrier will start
887 * until the IO request has had a chance.
888 *
889 * So: regular IO calls 'wait_barrier'. When that returns there
890 * is no backgroup IO happening, It must arrange to call
891 * allow_barrier when it has finished its IO.
892 * backgroup IO calls must call raise_barrier. Once that returns
893 * there is no normal IO happeing. It must arrange to call
894 * lower_barrier when the particular background IO completes.
895 */
898 {
902 /* Wait until no block IO is waiting (unless 'force') */
906 /* block any new IO from starting */
909 /* Now wait for all pending IO to complete */
915 }
918 {
924 }
927 {
931 /* Wait for the barrier to drop.
932 * However if there are already pending
933 * requests (preventing the barrier from
934 * rising completely), and the
935 * pre-process bio queue isn't empty,
936 * then don't wait, as we need to empty
937 * that queue to get the nr_pending
938 * count down.
939 */
949 }
952 }
955 {
959 }
962 {
963 /* stop syncio and normal IO and wait for everything to
964 * go quiet.
965 * We increment barrier and nr_waiting, and then
966 * wait until nr_pending match nr_queued+extra
967 * This is called in the context of one normal IO request
968 * that has failed. Thus any sync request that might be pending
969 * will be blocked by nr_pending, and we need to wait for
970 * pending IO requests to complete or be queued for re-try.
971 * Thus the number queued (nr_queued) plus this request (extra)
972 * must match the number of pending IOs (nr_pending) before
973 * we continue.
974 */
986 }
989 {
990 /* reverse the effect of the freeze */
996 }
1000 {
1004 else
1006 }
1012 };
1015 {
1017 cb);
1031 }
1033 /* we aren't scheduling, so we can do the write-out directly. */
1043 /* Just ignore it */
1045 else
1048 }
1050 }
1053 {
1072 /*
1073 * Register the new request and wait if the reconstruction
1074 * thread has put up a bar for new requests.
1075 * Continue immediately if no resync is active currently.
1076 */
1083 /* IO spans the reshape position. Need to wait for
1084 * reshape to pass
1085 */
1090 sectors);
1092 }
1100 /* Need to update reshape_position in metadata */
1109 }
1120 /* We might need to issue multiple reads to different
1121 * devices if there are bad blocks around, so we keep
1122 * track of the number of reads in bio->bi_phys_segments.
1123 * If this is 0, there is only one r10_bio and no locking
1124 * will be needed when the request completes. If it is
1125 * non-zero, then it is the number of not-completed requests.
1126 */
1131 /*
1132 * read balancing logic:
1133 */
1137 read_again:
1142 }
1147 max_sectors);
1160 /* Could not read all from this device, so we will
1161 * need another r10_bio.
1162 */
1169 else
1172 /* Cannot call generic_make_request directly
1173 * as that will be queued in __generic_make_request
1174 * and subsequent mempool_alloc might block
1175 * waiting for it. so hand bio over to raid10d.
1176 */
1186 sectors_handled;
1191 }
1193 /*
1194 * WRITE:
1195 */
1200 }
1201 /* first select target devices under rcu_lock and
1202 * inc refcount on their rdev. Record them by setting
1203 * bios[x] to bio
1204 * If there are known/acknowledged bad blocks on any device
1205 * on which we have seen a write error, we want to avoid
1206 * writing to those blocks. This potentially requires several
1207 * writes to write around the bad blocks. Each set of writes
1208 * gets its own r10_bio with a set of bios attached. The number
1209 * of r10_bios is recored in bio->bi_phys_segments just as with
1210 * the read case.
1211 */
1215 retry_write:
1231 }
1236 }
1248 }
1250 sector_t first_bad;
1256 max_sectors,
1259 /* Mustn't write here until the bad block
1260 * is acknowledged
1261 */
1266 }
1268 /* Cannot write here at all */
1271 /* Mustn't write more than bad_sectors
1272 * to other devices yet
1273 */
1275 /* We don't set R10BIO_Degraded as that
1276 * only applies if the disk is missing,
1277 * so it might be re-added, and we want to
1278 * know to recover this chunk.
1279 * In this case the device is here, and the
1280 * fact that this chunk is not in-sync is
1281 * recorded in the bad block log.
1282 */
1284 }
1289 }
1290 }
1294 }
1298 }
1299 }
1303 /* Have to wait for this device to get unblocked, then retry */
1311 }
1317 /* Race with remove_disk */
1320 }
1322 }
1323 }
1328 }
1331 /* We are splitting this into multiple parts, so
1332 * we need to prepare for allocating another r10_bio.
1333 */
1338 else
1341 }
1355 max_sectors);
1360 rdev));
1372 cb);
1373 else
1382 }
1386 }
1391 /* Replacement just got moved to main 'rdev' */
1394 }
1397 max_sectors);
1415 }
1416 }
1418 /* Don't remove the bias on 'remaining' (one_write_done) until
1419 * after checking if we need to go around again.
1420 */
1424 /* We need another r10_bio. It has already been counted
1425 * in bio->bi_phys_segments.
1426 */
1436 }
1438 }
1441 {
1451 }
1455 /*
1456 * If this request crosses a chunk boundary, we need to split
1457 * it.
1458 */
1471 }
1476 /* In case raid10d snuck in to freeze_array */
1478 }
1481 {
1492 else
1496 }
1503 }
1506 }
1508 /* check if there are enough drives for
1509 * every block to appear on atleast one.
1510 * Don't consider the device numbered 'ignore'
1511 * as we might be about to remove it.
1512 */
1514 {
1524 }
1536 cnt++;
1538 }
1544 out:
1547 }
1550 {
1551 /* when calling 'enough', both 'prev' and 'geo' must
1552 * be stable.
1553 * This is ensured if ->reconfig_mutex or ->device_lock
1554 * is held.
1555 */
1558 }
1561 {
1566 /*
1567 * If it is not operational, then we have already marked it as dead
1568 * else if it is the last working disks, ignore the error, let the
1569 * next level up know.
1570 * else mark the drive as failed
1571 */
1575 /*
1576 * Don't fail the drive, just return an IO error.
1577 */
1580 }
1583 /*
1584 * If recovery is running, make sure it aborts.
1585 */
1597 }
1600 {
1608 }
1612 /* This is only called with ->reconfix_mutex held, so
1613 * rcu protection of rdev is not needed */
1622 }
1623 }
1626 {
1632 }
1635 {
1642 /*
1643 * Find all non-in_sync disks within the RAID10 configuration
1644 * and mark them in_sync
1645 */
1652 /* Replacement has just become active */
1655 count++;
1657 /* Replaced device not technically faulty,
1658 * but we need to be sure it gets removed
1659 * and never re-added.
1660 */
1664 }
1670 count++;
1672 }
1673 }
1680 }
1683 {
1691 /* only hot-add to in-sync arrays, as recovery is
1692 * very different from resync
1693 */
1707 else
1727 }
1741 }
1747 }
1750 {
1762 else
1769 }
1770 /* Only remove non-faulty devices if recovery
1771 * is not possible.
1772 */
1780 }
1785 /* lost the race, try later */
1789 }
1790 }
1792 /* We must have just cleared 'rdev' */
1796 * but will never see neither -- if they are careful.
1797 */
1801 /* We might have just remove the Replacement as faulty
1802 * Clear the flag just in case
1803 */
1808 abort:
1812 }
1815 {
1821 /* this is a reshape read */
1828 else
1829 /* The write handler will notice the lack of
1830 * R10BIO_Uptodate and record any errors etc
1831 */
1835 /* for reconstruct, we always reschedule after a read.
1836 * for resync, only after all reads
1837 */
1841 /* we have read all the blocks,
1842 * do the comparison in process context in raid10d
1843 */
1845 }
1846 }
1849 {
1854 /* the primary of several recovery bios */
1859 else
1868 else
1871 }
1872 }
1873 }
1876 {
1881 sector_t first_bad;
1890 else
1902 }
1912 }
1914 /*
1915 * Note: sync and recover and handled very differently for raid10
1916 * This code is for resync.
1917 * For resync, we read through virtual addresses and read all blocks.
1918 * If there is any error, we schedule a write. The lowest numbered
1919 * drive is authoritative.
1920 * However requests come for physical address, so we need to map.
1921 * For every physical address there are raid_disks/copies virtual addresses,
1922 * which is always are least one, but is not necessarly an integer.
1923 * This means that a physical address can span multiple chunks, so we may
1924 * have to submit multiple io requests for a single sync request.
1925 */
1926 /*
1927 * We check if all blocks are in-sync and only write to blocks that
1928 * aren't in sync
1929 */
1931 {
1939 /* find the first device with a block */
1953 /* now find blocks with errors */
1964 /* We know that the bi_io_vec layout is the same for
1965 * both 'first' and 'i', so we just compare them.
1966 * All vec entries are PAGE_SIZE;
1967 */
1975 len))
1978 }
1983 /* Don't fix anything. */
1985 }
1986 /* Ok, we need to write this bio, either to correct an
1987 * inconsistency or to correct an unreadable block.
1988 * First we need to fixup bv_offset, bv_len and
1989 * bi_vecs, as the read request might have corrupted these
1990 */
2010 }
2012 /* Now write out to any replacement devices
2013 * that are active
2014 */
2029 }
2031 done:
2035 }
2036 }
2038 /*
2039 * Now for the recovery code.
2040 * Recovery happens across physical sectors.
2041 * We recover all non-is_sync drives by finding the virtual address of
2042 * each, and then choose a working drive that also has that virt address.
2043 * There is a separate r10_bio for each non-in_sync drive.
2044 * Only the first two slots are in use. The first for reading,
2045 * The second for writing.
2046 *
2047 */
2049 {
2050 /* We got a read error during recovery.
2051 * We repeat the read in smaller page-sized sections.
2052 * If a read succeeds, write it to the new device or record
2053 * a bad block if we cannot.
2054 * If a read fails, record a bad block on both old and
2055 * new devices.
2056 */
2069 sector_t addr;
2078 addr,
2086 addr,
2096 }
2097 }
2099 /* We don't worry if we cannot set a bad block -
2100 * it really is bad so there is no loss in not
2101 * recording it yet
2102 */
2106 /* need bad block on destination too */
2111 /* just abort the recovery */
2113 "md/raid10:%s: recovery aborted"
2122 }
2123 }
2124 }
2128 idx++;
2129 }
2130 }
2133 {
2142 }
2144 /*
2145 * share the pages with the first bio
2146 * and submit the write request
2147 */
2151 /* Need to test wbio2->bi_end_io before we call
2152 * generic_make_request as if the former is NULL,
2153 * the latter is free to free wbio2.
2154 */
2161 }
2167 }
2168 }
2170 /*
2171 * Used by fix_read_error() to decay the per rdev read_errors.
2172 * We halve the read error count for every hour that has elapsed
2173 * since the last recorded read error.
2174 *
2175 */
2177 {
2185 /* first time we've seen a read error */
2188 }
2195 /*
2196 * if hours_since_last is > the number of bits in read_errors
2197 * just set read errors to 0. We do this to avoid
2198 * overflowing the shift of read_errors by hours_since_last.
2199 */
2202 else
2204 }
2208 {
2209 sector_t first_bad;
2216 /* success */
2223 }
2224 /* need to record an error - either for the block or the device */
2228 }
2230 /*
2231 * This is a kernel thread which:
2232 *
2233 * 1. Retries failed read operations on working mirrors.
2234 * 2. Updates the raid superblock when problems encounter.
2235 * 3. Performs writes following reads for array synchronising.
2236 */
2239 {
2246 /* still own a reference to this rdev, so it cannot
2247 * have been cleared recently.
2248 */
2252 /* drive has already been failed, just ignore any
2253 more fix_read_error() attempts */
2263 "md/raid10:%s: %s: Raid device exceeded "
2273 }
2286 sector_t first_bad;
2300 sect,
2308 }
2309 sl++;
2316 /* Cannot read from anywhere, just mark the block
2317 * as bad on the first device to discourage future
2318 * reads.
2319 */
2324 rdev,
2331 }
2333 }
2336 /* write it back and re-read */
2343 sl--;
2355 sect,
2358 /* Well, this device is dead */
2360 "md/raid10:%s: read correction "
2361 "write failed"
2365 sect +
2367 rdev)),
2373 }
2376 }
2383 sl--;
2395 sect,
2397 READ)) {
2399 /* Well, this device is dead */
2401 "md/raid10:%s: unable to read back "
2402 "corrected sectors"
2406 sect +
2416 "md/raid10:%s: read error corrected"
2420 sect +
2424 }
2428 }
2433 }
2434 }
2437 {
2442 /* bio has the data to be written to slot 'i' where
2443 * we just recently had a write error.
2444 * We repeatedly clone the bio and trim down to one block,
2445 * then try the write. Where the write fails we record
2446 * a bad block.
2447 * It is conceivable that the bio doesn't exactly align with
2448 * blocks. We must handle this.
2449 *
2450 * We currently own a reference to the rdev.
2451 */
2454 sector_t sector;
2471 sector_t wsector;
2474 /* Write at 'sector' for 'sectors' */
2484 /* Failure! */
2493 }
2495 }
2498 {
2507 /* we got a read error. Maybe the drive is bad. Maybe just
2508 * the block and we can fix it.
2509 * We freeze all other IO, and try reading the block from
2510 * other devices. When we find one, we re-write
2511 * and check it that fixes the read error.
2512 * This is all done synchronously while the array is
2513 * frozen.
2514 */
2529 read_more:
2538 }
2543 KERN_ERR
2544 "md/raid10:%s: %s: redirecting "
2561 /* Drat - have to split this up more */
2570 else
2576 GFP_NOIO);
2589 }
2592 {
2593 /* Some sort of write request has finished and it
2594 * succeeded in writing where we thought there was a
2595 * bad block. So forget the bad block.
2596 * Or possibly if failed and we need to record
2597 * a bad block.
2598 */
2611 rdev,
2616 rdev,
2620 }
2627 rdev,
2632 rdev,
2636 }
2637 }
2647 rdev,
2657 }
2659 }
2664 rdev,
2668 }
2669 }
2681 }
2682 }
2683 }
2686 {
2704 }
2705 }
2709 retry_list);
2718 }
2719 }
2730 }
2750 /* just a partial read to be scheduled from a
2751 * separate context
2752 */
2755 }
2760 }
2762 }
2765 {
2780 }
2782 /*
2783 * perform a "sync" on one "block"
2784 *
2785 * We need to make sure that no normal I/O request - particularly write
2786 * requests - conflict with active sync requests.
2787 *
2788 * This is achieved by tracking pending requests and a 'barrier' concept
2789 * that can be installed to exclude normal IO requests.
2790 *
2791 * Resync and recovery are handled very differently.
2792 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2793 *
2794 * For resync, we iterate over virtual addresses, read all copies,
2795 * and update if there are differences. If only one copy is live,
2796 * skip it.
2797 * For recovery, we iterate over physical addresses, read a good
2798 * value for each non-in_sync drive, and over-write.
2799 *
2800 * So, for recovery we may have several outstanding complex requests for a
2801 * given address, one for each out-of-sync device. We model this by allocating
2802 * a number of r10_bio structures, one for each out-of-sync device.
2803 * As we setup these structures, we collect all bio's together into a list
2804 * which we then process collectively to add pages, and then process again
2805 * to pass to generic_make_request.
2806 *
2807 * The r10_bio structures are linked using a borrowed master_bio pointer.
2808 * This link is counted in ->remaining. When the r10_bio that points to NULL
2809 * has its remaining count decremented to 0, the whole complex operation
2810 * is complete.
2811 *
2812 */
2816 {
2823 sector_t sync_blocks;
2832 /*
2833 * Allow skipping a full rebuild for incremental assembly
2834 * of a clean array, like RAID1 does.
2835 */
2845 }
2847 skipped:
2853 /* If we aborted, we need to abort the
2854 * sync on the 'current' bitmap chucks (there can
2855 * be several when recovering multiple devices).
2856 * as we may have started syncing it but not finished.
2857 * We can find the current address in
2858 * mddev->curr_resync, but for recovery,
2859 * we need to convert that to several
2860 * virtual addresses.
2861 */
2866 }
2873 sector_t sect =
2877 }
2879 /* completed sync */
2883 /* Completed a full sync so the replacements
2884 * are now fully recovered.
2885 */
2892 }
2894 }
2896 }
2901 }
2907 /* if there has been nothing to do on any drive,
2908 * then there is nothing to do at all..
2909 */
2912 }
2917 /* make sure whole request will fit in a chunk - if chunks
2918 * are meaningful
2919 */
2924 /*
2925 * If there is non-resync activity waiting for a turn, then let it
2926 * though before starting on this new sync request.
2927 */
2931 /* Again, very different code for resync and recovery.
2932 * Both must result in an r10bio with a list of bios that
2933 * have bi_end_io, bi_sector, bi_bdev set,
2934 * and bi_private set to the r10bio.
2935 * For recovery, we may actually create several r10bios
2936 * with 2 bios in each, that correspond to the bios in the main one.
2937 * In this case, the subordinate r10bios link back through a
2938 * borrowed master_bio pointer, and the counter in the master
2939 * includes a ref from each subordinate.
2940 */
2941 /* First, we decide what to do and set ->bi_end_io
2942 * To end_sync_read if we want to read, and
2943 * end_sync_write if we will want to write.
2944 */
2948 /* recovery... the complicated one */
2955 sector_t sect;
2972 }
2975 /* want to reconstruct this device */
2979 /* last stripe is not complete - don't
2980 * try to recover this sector.
2981 */
2984 }
2987 /* Unless we are doing a full sync, or a replacement
2988 * we only need to recover the block if it is set in
2989 * the bitmap
2990 */
2998 /* yep, skip the sync_blocks here, but don't assume
2999 * that there will never be anything to do here
3000 */
3004 }
3024 /* Need to check if the array will still be
3025 * degraded
3026 */
3034 }
3035 }
3052 /* This is where we read from */
3061 bad_sectors -= (sector
3066 }
3067 }
3080 /* and we write to 'i' (if not in_sync) */
3107 /* and maybe write to replacement */
3111 /* Note: if mreplace != NULL, then bio
3112 * cannot be NULL as r10buf_pool_alloc will
3113 * have allocated it.
3114 * So the second test here is pointless.
3115 * But it keeps semantic-checkers happy, and
3116 * this comment keeps human reviewers
3117 * happy.
3118 */
3133 }
3136 /* Cannot recover, so abort the recovery or
3137 * record a bad block */
3139 /* problem is that there are bad blocks
3140 * on other device(s)
3141 */
3149 mrdev,
3155 mreplace,
3159 }
3166 mirror->recovery_disabled
3168 }
3177 }
3181 }
3188 }
3190 }
3192 /* resync. Schedule a read for every block at this virt offset */
3201 /* We can skip this block */
3204 }
3238 }
3250 }
3251 }
3261 count++;
3267 }
3271 /* Need to set up for writing to the replacement */
3284 count++;
3285 }
3292 mddev);
3297 mddev);
3298 }
3302 }
3303 }
3321 /* stop here */
3326 /* remove last page from this bio */
3330 }
3332 }
3336 bio_full:
3351 }
3352 }
3355 /* pretend they weren't skipped, it makes
3356 * no important difference in this case
3357 */
3361 giveup:
3362 /* There is nowhere to write, so all non-sync
3363 * drives must be failed or in resync, all drives
3364 * have a bad block, so try the next chunk...
3365 */
3370 chunks_skipped ++;
3373 }
3375 static sector_t
3377 {
3378 sector_t size;
3393 }
3396 {
3397 /* Calculate the number of sectors-per-device that will
3398 * actually be used, and set conf->dev_sectors and
3399 * conf->stride
3400 */
3406 /* 'size' is now the number of chunks in the array */
3407 /* calculate "used chunks per device" */
3410 /* We need to round up when dividing by raid_disks to
3411 * get the stride size.
3412 */
3422 }
3423 }
3427 {
3443 * updated yet. */
3448 }
3466 * actually using this, but leave code here just in case.*/
3476 }
3480 }
3483 {
3496 }
3502 }
3509 /* FIXME calc properly */
3512 GFP_KERNEL);
3535 }
3539 else
3540 /* far_copies must be 1 */
3542 }
3559 out:
3568 }
3570 }
3573 {
3578 sector_t size;
3588 }
3604 else
3607 }
3629 }
3647 }
3653 else
3656 }
3657 /* need to check that every block has at least one working mirror */
3662 }
3665 /* must ensure that shape change is supported */
3672 }
3678 i++) {
3683 /* The replacement is all we have - use it */
3687 }
3696 }
3698 }
3708 /*
3709 * Ok, everything is just fine now
3710 */
3720 /* Calculate max read-ahead size.
3721 * We need to readahead at least twice a whole stripe....
3722 * maybe...
3723 */
3727 }
3741 /* This cannot work */
3744 }
3753 }
3757 out_free_conf:
3764 out:
3766 }
3769 {
3778 }
3781 {
3791 }
3792 }
3795 {
3796 /* Resize of 'far' arrays is not supported.
3797 * For 'near' and 'offset' arrays we can set the
3798 * number of sectors used to be an appropriate multiple
3799 * of the chunk size.
3800 * For 'offset', this is far_copies*chunksize.
3801 * For 'near' the multiplier is the LCM of
3802 * near_copies and raid_disks.
3803 * So if far_copies > 1 && !far_offset, fail.
3804 * Else find LCM(raid_disks, near_copy)*far_copies and
3805 * multiply by chunk_size. Then round to this number.
3806 * This is mostly done by raid10_size()
3807 */
3826 }
3831 }
3836 }
3841 }
3844 {
3852 }
3855 /* Set new parameters */
3857 /* new layout: far_copies = 1, near_copies = 2 */
3862 /* make sure it will be not marked as dirty */
3872 }
3874 }
3877 }
3880 {
3883 /* raid10 can take over:
3884 * raid0 - providing it has only two drives
3885 */
3887 /* for raid0 takeover only one zone is supported */
3894 }
3898 }
3900 }
3903 {
3904 /* Called when there is a request to change
3905 * - layout (to ->new_layout)
3906 * - chunk size (to ->new_chunk_sectors)
3907 * - raid_disks (by delta_disks)
3908 * or when trying to restart a reshape that was ongoing.
3909 *
3910 * We need to validate the request and possibly allocate
3911 * space if that might be an issue later.
3912 *
3913 * Currently we reject any reshape of a 'far' mode array,
3914 * allow chunk size to change if new is generally acceptable,
3915 * allow raid_disks to increase, and allow
3916 * a switch between 'near' mode and 'offset' mode.
3917 */
3925 /* mustn't change number of copies */
3928 /* Cannot switch to 'far' mode */
3932 /* not factor of array size */
3941 /* allocate new 'mirrors' list */
3946 GFP_KERNEL);
3949 }
3951 }
3953 /*
3954 * Need to check if array has failed when deciding whether to:
3955 * - start an array
3956 * - remove non-faulty devices
3957 * - add a spare
3958 * - allow a reshape
3959 * This determination is simple when no reshape is happening.
3960 * However if there is a reshape, we need to carefully check
3961 * both the before and after sections.
3962 * This is because some failed devices may only affect one
3963 * of the two sections, and some non-in_sync devices may
3964 * be insync in the section most affected by failed devices.
3965 */
3967 {
3973 /* 'prev' section first */
3977 degraded++;
3979 /* When we can reduce the number of devices in
3980 * an array, this might not contribute to
3981 * 'degraded'. It does now.
3982 */
3983 degraded++;
3984 }
3993 degraded2++;
3995 /* If reshape is increasing the number of devices,
3996 * this section has already been recovered, so
3997 * it doesn't contribute to degraded.
3998 * else it does.
3999 */
4001 degraded2++;
4002 }
4003 }
4008 }
4011 {
4012 /* A 'reshape' has been requested. This commits
4013 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4014 * This also checks if there are enough spares and adds them
4015 * to the array.
4016 * We currently require enough spares to make the final
4017 * array non-degraded. We also require that the difference
4018 * between old and new data_offset - on each device - is
4019 * enough that we never risk over-writing.
4020 */
4045 spares++;
4055 }
4056 }
4074 }
4084 }
4099 }
4108 else
4113 }
4116 /* This is a spare that was manually added */
4118 }
4119 }
4120 /* When a reshape changes the number of devices,
4121 * ->degraded is measured against the larger of the
4122 * pre and post numbers.
4123 */
4142 }
4148 abort:
4161 }
4163 /* Calculate the last device-address that could contain
4164 * any block from the chunk that includes the array-address 's'
4165 * and report the next address.
4166 * i.e. the address returned will be chunk-aligned and after
4167 * any data that is in the chunk containing 's'.
4168 */
4170 {
4178 }
4180 /* Calculate the first device-address that could contain
4181 * any block from the chunk that includes the array-address 's'.
4182 * This too will be the start of a chunk
4183 */
4185 {
4192 }
4196 {
4197 /* We simply copy at most one chunk (smallest of old and new)
4198 * at a time, possibly less if that exceeds RESYNC_PAGES,
4199 * or we hit a bad block or something.
4200 * This might mean we pause for normal IO in the middle of
4201 * a chunk, but that is not a problem as mddev->reshape_position
4202 * can record any location.
4203 *
4204 * If we will want to write to a location that isn't
4205 * yet recorded as 'safe' (i.e. in metadata on disk) then
4206 * we need to flush all reshape requests and update the metadata.
4207 *
4208 * When reshaping forwards (e.g. to more devices), we interpret
4209 * 'safe' as the earliest block which might not have been copied
4210 * down yet. We divide this by previous stripe size and multiply
4211 * by previous stripe length to get lowest device offset that we
4212 * cannot write to yet.
4213 * We interpret 'sector_nr' as an address that we want to write to.
4214 * From this we use last_device_address() to find where we might
4215 * write to, and first_device_address on the 'safe' position.
4216 * If this 'next' write position is after the 'safe' position,
4217 * we must update the metadata to increase the 'safe' position.
4218 *
4219 * When reshaping backwards, we round in the opposite direction
4220 * and perform the reverse test: next write position must not be
4221 * less than current safe position.
4222 *
4223 * In all this the minimum difference in data offsets
4224 * (conf->offset_diff - always positive) allows a bit of slack,
4225 * so next can be after 'safe', but not by more than offset_diff
4226 *
4227 * We need to prepare all the bios here before we start any IO
4228 * to ensure the size we choose is acceptable to all devices.
4229 * The means one for each copy for write-out and an extra one for
4230 * read-in.
4231 * We store the read-in bio in ->master_bio and the others in
4232 * ->devs[x].bio and ->devs[x].repl_bio.
4233 */
4247 /* If restarting in the middle, skip the initial sectors */
4260 }
4261 }
4263 /* We don't use sector_nr to track where we are up to
4264 * as that doesn't work well for ->reshape_backwards.
4265 * So just use ->reshape_progress.
4266 */
4268 /* 'next' is the earliest device address that we might
4269 * write to for this chunk in the new layout
4270 */
4274 /* 'safe' is the last device address that we might read from
4275 * in the old layout after a restart
4276 */
4289 /* 'next' is after the last device address that we
4290 * might write to for this chunk in the new layout
4291 */
4294 /* 'safe' is the earliest device address that we might
4295 * read from in the old layout after a restart
4296 */
4299 /* Need to update metadata if 'next' might be beyond 'safe'
4300 * as that would possibly corrupt data
4301 */
4311 }
4315 /* Need to update reshape_position in metadata */
4321 else
4331 }
4334 }
4336 read_more:
4337 /* Now schedule reads for blocks from sector_nr to last */
4350 /* Cannot read from here, so need to record bad blocks
4351 * on all the target devices.
4352 */
4353 // FIXME
4357 }
4374 /* Now find the locations in the new layout */
4391 }
4404 }
4406 /* Now add as many pages as possible to all of these bios. */
4419 /* Didn't fit, must stop */
4423 /* Remove last page from this bio */
4427 }
4429 }
4432 }
4433 bio_full:
4437 /* Now submit the read */
4447 /* Now that we have done the whole section we can
4448 * update reshape_progress
4449 */
4452 else
4456 }
4462 {
4463 /* Reshape read completed. Hopefully we have a block
4464 * to write out.
4465 * If we got a read error then we do sync 1-page reads from
4466 * elsewhere until we find the data - or give up.
4467 */
4473 /* Reshape has been aborted */
4476 }
4478 /* We definitely have the data in the pages, schedule the
4479 * writes.
4480 */
4493 }
4497 }
4504 }
4506 }
4509 {
4521 /* read-ahead size must cover two whole stripes, which is
4522 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4523 */
4530 }
4532 }
4536 {
4537 /* Use sync reads to get the blocks from somewhere else */
4564 sector_t addr;
4574 addr,
4582 failed:
4583 slot++;
4588 }
4591 /* couldn't read this block, must give up */
4595 }
4597 idx++;
4598 }
4600 }
4603 {
4618 }
4621 /* FIXME should record badblock */
4623 }
4627 }
4630 {
4636 }
4639 {
4651 }
4656 }
4662 d++) {
4669 }
4671 }
4677 }
4680 {
4701 };
4704 {
4706 }
4709 {
4711 }