| blob | 84e597e1c4890cd911243c511d1a65e39232e98e |
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 {
710 retry:
717 /*
718 * Check if we can balance. We can balance on the whole
719 * device if no resync is going on (recovery is ok), or below
720 * the resync window. We take the first readable disk when
721 * above the resync window.
722 */
728 sector_t first_bad;
730 sector_t dev_sector;
750 /* Already have a better slot */
753 /* Cannot read here. If this is the
754 * 'primary' device, then we must not read
755 * beyond 'bad_sectors' from another device.
756 */
763 sector_t good_sectors =
769 }
771 /* Must read from here */
773 }
781 /* This optimisation is debatable, and completely destroys
782 * sequential read speed for 'far copies' arrays. So only
783 * keep it for 'near' arrays, and review those later.
784 */
788 /* for far > 1 always use the lowest address */
791 else
798 }
799 }
803 }
808 /* Cannot risk returning a device that failed
809 * before we inc'ed nr_pending
810 */
813 }
821 }
824 {
836 i++) {
842 }
843 }
846 }
849 {
850 /* Any writes that have been queued but are awaiting
851 * bitmap updates get flushed here.
852 */
860 /* flush any pending bitmap writes to disk
861 * before proceeding w/ I/O */
870 /* Just ignore it */
872 else
875 }
878 }
880 /* Barriers....
881 * Sometimes we need to suspend IO while we do something else,
882 * either some resync/recovery, or reconfigure the array.
883 * To do this we raise a 'barrier'.
884 * The 'barrier' is a counter that can be raised multiple times
885 * to count how many activities are happening which preclude
886 * normal IO.
887 * We can only raise the barrier if there is no pending IO.
888 * i.e. if nr_pending == 0.
889 * We choose only to raise the barrier if no-one is waiting for the
890 * barrier to go down. This means that as soon as an IO request
891 * is ready, no other operations which require a barrier will start
892 * until the IO request has had a chance.
893 *
894 * So: regular IO calls 'wait_barrier'. When that returns there
895 * is no backgroup IO happening, It must arrange to call
896 * allow_barrier when it has finished its IO.
897 * backgroup IO calls must call raise_barrier. Once that returns
898 * there is no normal IO happeing. It must arrange to call
899 * lower_barrier when the particular background IO completes.
900 */
903 {
907 /* Wait until no block IO is waiting (unless 'force') */
911 /* block any new IO from starting */
914 /* Now wait for all pending IO to complete */
920 }
923 {
929 }
932 {
936 /* Wait for the barrier to drop.
937 * However if there are already pending
938 * requests (preventing the barrier from
939 * rising completely), and the
940 * pre-process bio queue isn't empty,
941 * then don't wait, as we need to empty
942 * that queue to get the nr_pending
943 * count down.
944 */
952 }
955 }
958 {
964 }
967 {
968 /* stop syncio and normal IO and wait for everything to
969 * go quiet.
970 * We increment barrier and nr_waiting, and then
971 * wait until nr_pending match nr_queued+extra
972 * This is called in the context of one normal IO request
973 * that has failed. Thus any sync request that might be pending
974 * will be blocked by nr_pending, and we need to wait for
975 * pending IO requests to complete or be queued for re-try.
976 * Thus the number queued (nr_queued) plus this request (extra)
977 * must match the number of pending IOs (nr_pending) before
978 * we continue.
979 */
989 }
992 {
993 /* reverse the effect of the freeze */
999 }
1003 {
1007 else
1009 }
1015 };
1018 {
1020 cb);
1034 }
1036 /* we aren't scheduling, so we can do the write-out directly. */
1046 /* Just ignore it */
1048 else
1051 }
1053 }
1056 {
1075 /*
1076 * Register the new request and wait if the reconstruction
1077 * thread has put up a bar for new requests.
1078 * Continue immediately if no resync is active currently.
1079 */
1086 /* IO spans the reshape position. Need to wait for
1087 * reshape to pass
1088 */
1093 sectors);
1095 }
1103 /* Need to update reshape_position in metadata */
1112 }
1123 /* We might need to issue multiple reads to different
1124 * devices if there are bad blocks around, so we keep
1125 * track of the number of reads in bio->bi_phys_segments.
1126 * If this is 0, there is only one r10_bio and no locking
1127 * will be needed when the request completes. If it is
1128 * non-zero, then it is the number of not-completed requests.
1129 */
1134 /*
1135 * read balancing logic:
1136 */
1140 read_again:
1145 }
1150 max_sectors);
1163 /* Could not read all from this device, so we will
1164 * need another r10_bio.
1165 */
1172 else
1175 /* Cannot call generic_make_request directly
1176 * as that will be queued in __generic_make_request
1177 * and subsequent mempool_alloc might block
1178 * waiting for it. so hand bio over to raid10d.
1179 */
1189 sectors_handled;
1194 }
1196 /*
1197 * WRITE:
1198 */
1203 }
1204 /* first select target devices under rcu_lock and
1205 * inc refcount on their rdev. Record them by setting
1206 * bios[x] to bio
1207 * If there are known/acknowledged bad blocks on any device
1208 * on which we have seen a write error, we want to avoid
1209 * writing to those blocks. This potentially requires several
1210 * writes to write around the bad blocks. Each set of writes
1211 * gets its own r10_bio with a set of bios attached. The number
1212 * of r10_bios is recored in bio->bi_phys_segments just as with
1213 * the read case.
1214 */
1218 retry_write:
1234 }
1239 }
1251 }
1253 sector_t first_bad;
1259 max_sectors,
1262 /* Mustn't write here until the bad block
1263 * is acknowledged
1264 */
1269 }
1271 /* Cannot write here at all */
1274 /* Mustn't write more than bad_sectors
1275 * to other devices yet
1276 */
1278 /* We don't set R10BIO_Degraded as that
1279 * only applies if the disk is missing,
1280 * so it might be re-added, and we want to
1281 * know to recover this chunk.
1282 * In this case the device is here, and the
1283 * fact that this chunk is not in-sync is
1284 * recorded in the bad block log.
1285 */
1287 }
1292 }
1293 }
1297 }
1301 }
1302 }
1306 /* Have to wait for this device to get unblocked, then retry */
1314 }
1320 /* Race with remove_disk */
1323 }
1325 }
1326 }
1331 }
1334 /* We are splitting this into multiple parts, so
1335 * we need to prepare for allocating another r10_bio.
1336 */
1341 else
1344 }
1358 max_sectors);
1363 rdev));
1376 cb);
1377 else
1386 }
1390 }
1395 /* Replacement just got moved to main 'rdev' */
1398 }
1401 max_sectors);
1420 }
1421 }
1423 /* Don't remove the bias on 'remaining' (one_write_done) until
1424 * after checking if we need to go around again.
1425 */
1429 /* We need another r10_bio. It has already been counted
1430 * in bio->bi_phys_segments.
1431 */
1441 }
1443 }
1446 {
1456 }
1462 /*
1463 * If this request crosses a chunk boundary, we need to split
1464 * it.
1465 */
1478 }
1483 /* In case raid10d snuck in to freeze_array */
1485 }
1488 {
1499 else
1503 }
1511 }
1513 /* check if there are enough drives for
1514 * every block to appear on atleast one.
1515 * Don't consider the device numbered 'ignore'
1516 * as we might be about to remove it.
1517 */
1519 {
1529 }
1541 cnt++;
1543 }
1549 out:
1552 }
1555 {
1556 /* when calling 'enough', both 'prev' and 'geo' must
1557 * be stable.
1558 * This is ensured if ->reconfig_mutex or ->device_lock
1559 * is held.
1560 */
1563 }
1566 {
1571 /*
1572 * If it is not operational, then we have already marked it as dead
1573 * else if it is the last working disks, ignore the error, let the
1574 * next level up know.
1575 * else mark the drive as failed
1576 */
1580 /*
1581 * Don't fail the drive, just return an IO error.
1582 */
1585 }
1588 /*
1589 * If recovery is running, make sure it aborts.
1590 */
1602 }
1605 {
1613 }
1625 }
1626 }
1629 {
1635 }
1638 {
1645 /*
1646 * Find all non-in_sync disks within the RAID10 configuration
1647 * and mark them in_sync
1648 */
1655 /* Replacement has just become active */
1658 count++;
1660 /* Replaced device not technically faulty,
1661 * but we need to be sure it gets removed
1662 * and never re-added.
1663 */
1667 }
1673 count++;
1675 }
1676 }
1683 }
1686 {
1694 /* only hot-add to in-sync arrays, as recovery is
1695 * very different from resync
1696 */
1707 else
1727 }
1741 }
1750 }
1753 {
1765 else
1772 }
1773 /* Only remove faulty devices if recovery
1774 * is not possible.
1775 */
1783 }
1787 /* lost the race, try later */
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 {
2186 /* first time we've seen a read error */
2189 }
2196 /*
2197 * if hours_since_last is > the number of bits in read_errors
2198 * just set read errors to 0. We do this to avoid
2199 * overflowing the shift of read_errors by hours_since_last.
2200 */
2203 else
2205 }
2209 {
2210 sector_t first_bad;
2217 /* success */
2224 }
2225 /* need to record an error - either for the block or the device */
2229 }
2231 /*
2232 * This is a kernel thread which:
2233 *
2234 * 1. Retries failed read operations on working mirrors.
2235 * 2. Updates the raid superblock when problems encounter.
2236 * 3. Performs writes following reads for array synchronising.
2237 */
2240 {
2247 /* still own a reference to this rdev, so it cannot
2248 * have been cleared recently.
2249 */
2253 /* drive has already been failed, just ignore any
2254 more fix_read_error() attempts */
2264 "md/raid10:%s: %s: Raid device exceeded "
2274 }
2287 sector_t first_bad;
2300 sect,
2307 }
2308 sl++;
2315 /* Cannot read from anywhere, just mark the block
2316 * as bad on the first device to discourage future
2317 * reads.
2318 */
2323 rdev,
2330 }
2332 }
2335 /* write it back and re-read */
2342 sl--;
2353 sect,
2356 /* Well, this device is dead */
2358 "md/raid10:%s: read correction "
2359 "write failed"
2363 sect +
2365 rdev)),
2371 }
2374 }
2381 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;
2470 /* Write at 'sector' for 'sectors' */
2478 /* Failure! */
2487 }
2489 }
2492 {
2501 /* we got a read error. Maybe the drive is bad. Maybe just
2502 * the block and we can fix it.
2503 * We freeze all other IO, and try reading the block from
2504 * other devices. When we find one, we re-write
2505 * and check it that fixes the read error.
2506 * This is all done synchronously while the array is
2507 * frozen.
2508 */
2523 read_more:
2532 }
2537 KERN_ERR
2538 "md/raid10:%s: %s: redirecting "
2555 /* Drat - have to split this up more */
2564 else
2570 GFP_NOIO);
2583 }
2586 {
2587 /* Some sort of write request has finished and it
2588 * succeeded in writing where we thought there was a
2589 * bad block. So forget the bad block.
2590 * Or possibly if failed and we need to record
2591 * a bad block.
2592 */
2605 rdev,
2610 rdev,
2614 }
2621 rdev,
2626 rdev,
2630 }
2631 }
2641 rdev,
2651 }
2653 }
2658 rdev,
2662 }
2663 }
2674 }
2675 }
2676 }
2679 {
2696 }
2700 retry_list);
2709 }
2710 }
2721 }
2741 /* just a partial read to be scheduled from a
2742 * separate context
2743 */
2746 }
2751 }
2753 }
2756 {
2771 }
2773 /*
2774 * perform a "sync" on one "block"
2775 *
2776 * We need to make sure that no normal I/O request - particularly write
2777 * requests - conflict with active sync requests.
2778 *
2779 * This is achieved by tracking pending requests and a 'barrier' concept
2780 * that can be installed to exclude normal IO requests.
2781 *
2782 * Resync and recovery are handled very differently.
2783 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2784 *
2785 * For resync, we iterate over virtual addresses, read all copies,
2786 * and update if there are differences. If only one copy is live,
2787 * skip it.
2788 * For recovery, we iterate over physical addresses, read a good
2789 * value for each non-in_sync drive, and over-write.
2790 *
2791 * So, for recovery we may have several outstanding complex requests for a
2792 * given address, one for each out-of-sync device. We model this by allocating
2793 * a number of r10_bio structures, one for each out-of-sync device.
2794 * As we setup these structures, we collect all bio's together into a list
2795 * which we then process collectively to add pages, and then process again
2796 * to pass to generic_make_request.
2797 *
2798 * The r10_bio structures are linked using a borrowed master_bio pointer.
2799 * This link is counted in ->remaining. When the r10_bio that points to NULL
2800 * has its remaining count decremented to 0, the whole complex operation
2801 * is complete.
2802 *
2803 */
2807 {
2814 sector_t sync_blocks;
2823 /*
2824 * Allow skipping a full rebuild for incremental assembly
2825 * of a clean array, like RAID1 does.
2826 */
2836 }
2838 skipped:
2844 /* If we aborted, we need to abort the
2845 * sync on the 'current' bitmap chucks (there can
2846 * be several when recovering multiple devices).
2847 * as we may have started syncing it but not finished.
2848 * We can find the current address in
2849 * mddev->curr_resync, but for recovery,
2850 * we need to convert that to several
2851 * virtual addresses.
2852 */
2857 }
2864 sector_t sect =
2868 }
2870 /* completed sync */
2874 /* Completed a full sync so the replacements
2875 * are now fully recovered.
2876 */
2880 ->recovery_offset
2882 }
2884 }
2889 }
2895 /* if there has been nothing to do on any drive,
2896 * then there is nothing to do at all..
2897 */
2900 }
2905 /* make sure whole request will fit in a chunk - if chunks
2906 * are meaningful
2907 */
2912 /* Again, very different code for resync and recovery.
2913 * Both must result in an r10bio with a list of bios that
2914 * have bi_end_io, bi_sector, bi_bdev set,
2915 * and bi_private set to the r10bio.
2916 * For recovery, we may actually create several r10bios
2917 * with 2 bios in each, that correspond to the bios in the main one.
2918 * In this case, the subordinate r10bios link back through a
2919 * borrowed master_bio pointer, and the counter in the master
2920 * includes a ref from each subordinate.
2921 */
2922 /* First, we decide what to do and set ->bi_end_io
2923 * To end_sync_read if we want to read, and
2924 * end_sync_write if we will want to write.
2925 */
2929 /* recovery... the complicated one */
2936 sector_t sect;
2943 &&
2950 /* want to reconstruct this device */
2954 /* last stripe is not complete - don't
2955 * try to recover this sector.
2956 */
2958 }
2959 /* Unless we are doing a full sync, or a replacement
2960 * we only need to recover the block if it is set in
2961 * the bitmap
2962 */
2970 /* yep, skip the sync_blocks here, but don't assume
2971 * that there will never be anything to do here
2972 */
2975 }
2991 /* Need to check if the array will still be
2992 * degraded
2993 */
2999 }
3015 /* This is where we read from */
3025 bad_sectors -= (sector
3030 }
3031 }
3044 /* and we write to 'i' (if not in_sync) */
3072 /* and maybe write to replacement */
3077 /* Note: if rdev != NULL, then bio
3078 * cannot be NULL as r10buf_pool_alloc will
3079 * have allocated it.
3080 * So the second test here is pointless.
3081 * But it keeps semantic-checkers happy, and
3082 * this comment keeps human reviewers
3083 * happy.
3084 */
3099 }
3101 /* Cannot recover, so abort the recovery or
3102 * record a bad block */
3104 /* problem is that there are bad blocks
3105 * on other device(s)
3106 */
3124 }
3131 mirror->recovery_disabled
3133 }
3139 }
3140 }
3147 }
3149 }
3151 /* resync. Schedule a read for every block at this virt offset */
3160 /* We can skip this block */
3163 }
3205 }
3206 }
3217 count++;
3224 /* Need to set up for writing to the replacement */
3239 count++;
3240 }
3247 mddev);
3252 mddev);
3253 }
3257 }
3258 }
3276 /* stop here */
3281 /* remove last page from this bio */
3285 }
3287 }
3291 bio_full:
3306 }
3307 }
3310 /* pretend they weren't skipped, it makes
3311 * no important difference in this case
3312 */
3316 giveup:
3317 /* There is nowhere to write, so all non-sync
3318 * drives must be failed or in resync, all drives
3319 * have a bad block, so try the next chunk...
3320 */
3325 chunks_skipped ++;
3328 }
3330 static sector_t
3332 {
3333 sector_t size;
3348 }
3351 {
3352 /* Calculate the number of sectors-per-device that will
3353 * actually be used, and set conf->dev_sectors and
3354 * conf->stride
3355 */
3361 /* 'size' is now the number of chunks in the array */
3362 /* calculate "used chunks per device" */
3365 /* We need to round up when dividing by raid_disks to
3366 * get the stride size.
3367 */
3377 }
3378 }
3382 {
3398 * updated yet. */
3403 }
3421 * actually using this, but leave code here just in case.*/
3431 }
3435 }
3438 {
3451 }
3457 }
3464 /* FIXME calc properly */
3467 GFP_KERNEL);
3490 }
3494 else
3495 /* far_copies must be 1 */
3497 }
3513 out:
3522 }
3524 }
3527 {
3532 sector_t size;
3542 }
3558 else
3561 }
3583 }
3601 }
3607 else
3610 }
3611 /* need to check that every block has at least one working mirror */
3616 }
3619 /* must ensure that shape change is supported */
3626 }
3632 i++) {
3637 /* The replacement is all we have - use it */
3641 }
3650 }
3652 }
3662 /*
3663 * Ok, everything is just fine now
3664 */
3674 /* Calculate max read-ahead size.
3675 * We need to readahead at least twice a whole stripe....
3676 * maybe...
3677 */
3681 }
3695 /* This cannot work */
3698 }
3707 }
3711 out_free_conf:
3718 out:
3720 }
3723 {
3732 }
3735 {
3745 }
3746 }
3749 {
3750 /* Resize of 'far' arrays is not supported.
3751 * For 'near' and 'offset' arrays we can set the
3752 * number of sectors used to be an appropriate multiple
3753 * of the chunk size.
3754 * For 'offset', this is far_copies*chunksize.
3755 * For 'near' the multiplier is the LCM of
3756 * near_copies and raid_disks.
3757 * So if far_copies > 1 && !far_offset, fail.
3758 * Else find LCM(raid_disks, near_copy)*far_copies and
3759 * multiply by chunk_size. Then round to this number.
3760 * This is mostly done by raid10_size()
3761 */
3780 }
3788 }
3793 }
3796 {
3804 }
3807 /* Set new parameters */
3809 /* new layout: far_copies = 1, near_copies = 2 */
3814 /* make sure it will be not marked as dirty */
3824 }
3826 }
3829 }
3832 {
3835 /* raid10 can take over:
3836 * raid0 - providing it has only two drives
3837 */
3839 /* for raid0 takeover only one zone is supported */
3846 }
3850 }
3852 }
3855 {
3856 /* Called when there is a request to change
3857 * - layout (to ->new_layout)
3858 * - chunk size (to ->new_chunk_sectors)
3859 * - raid_disks (by delta_disks)
3860 * or when trying to restart a reshape that was ongoing.
3861 *
3862 * We need to validate the request and possibly allocate
3863 * space if that might be an issue later.
3864 *
3865 * Currently we reject any reshape of a 'far' mode array,
3866 * allow chunk size to change if new is generally acceptable,
3867 * allow raid_disks to increase, and allow
3868 * a switch between 'near' mode and 'offset' mode.
3869 */
3877 /* mustn't change number of copies */
3880 /* Cannot switch to 'far' mode */
3884 /* not factor of array size */
3893 /* allocate new 'mirrors' list */
3898 GFP_KERNEL);
3901 }
3903 }
3905 /*
3906 * Need to check if array has failed when deciding whether to:
3907 * - start an array
3908 * - remove non-faulty devices
3909 * - add a spare
3910 * - allow a reshape
3911 * This determination is simple when no reshape is happening.
3912 * However if there is a reshape, we need to carefully check
3913 * both the before and after sections.
3914 * This is because some failed devices may only affect one
3915 * of the two sections, and some non-in_sync devices may
3916 * be insync in the section most affected by failed devices.
3917 */
3919 {
3925 /* 'prev' section first */
3929 degraded++;
3931 /* When we can reduce the number of devices in
3932 * an array, this might not contribute to
3933 * 'degraded'. It does now.
3934 */
3935 degraded++;
3936 }
3945 degraded2++;
3947 /* If reshape is increasing the number of devices,
3948 * this section has already been recovered, so
3949 * it doesn't contribute to degraded.
3950 * else it does.
3951 */
3953 degraded2++;
3954 }
3955 }
3960 }
3963 {
3964 /* A 'reshape' has been requested. This commits
3965 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3966 * This also checks if there are enough spares and adds them
3967 * to the array.
3968 * We currently require enough spares to make the final
3969 * array non-degraded. We also require that the difference
3970 * between old and new data_offset - on each device - is
3971 * enough that we never risk over-writing.
3972 */
3997 spares++;
4007 }
4008 }
4026 }
4036 }
4051 }
4060 else
4065 }
4068 /* This is a spare that was manually added */
4070 }
4071 }
4072 /* When a reshape changes the number of devices,
4073 * ->degraded is measured against the larger of the
4074 * pre and post numbers.
4075 */
4094 }
4100 abort:
4113 }
4115 /* Calculate the last device-address that could contain
4116 * any block from the chunk that includes the array-address 's'
4117 * and report the next address.
4118 * i.e. the address returned will be chunk-aligned and after
4119 * any data that is in the chunk containing 's'.
4120 */
4122 {
4130 }
4132 /* Calculate the first device-address that could contain
4133 * any block from the chunk that includes the array-address 's'.
4134 * This too will be the start of a chunk
4135 */
4137 {
4144 }
4148 {
4149 /* We simply copy at most one chunk (smallest of old and new)
4150 * at a time, possibly less if that exceeds RESYNC_PAGES,
4151 * or we hit a bad block or something.
4152 * This might mean we pause for normal IO in the middle of
4153 * a chunk, but that is not a problem as mddev->reshape_position
4154 * can record any location.
4155 *
4156 * If we will want to write to a location that isn't
4157 * yet recorded as 'safe' (i.e. in metadata on disk) then
4158 * we need to flush all reshape requests and update the metadata.
4159 *
4160 * When reshaping forwards (e.g. to more devices), we interpret
4161 * 'safe' as the earliest block which might not have been copied
4162 * down yet. We divide this by previous stripe size and multiply
4163 * by previous stripe length to get lowest device offset that we
4164 * cannot write to yet.
4165 * We interpret 'sector_nr' as an address that we want to write to.
4166 * From this we use last_device_address() to find where we might
4167 * write to, and first_device_address on the 'safe' position.
4168 * If this 'next' write position is after the 'safe' position,
4169 * we must update the metadata to increase the 'safe' position.
4170 *
4171 * When reshaping backwards, we round in the opposite direction
4172 * and perform the reverse test: next write position must not be
4173 * less than current safe position.
4174 *
4175 * In all this the minimum difference in data offsets
4176 * (conf->offset_diff - always positive) allows a bit of slack,
4177 * so next can be after 'safe', but not by more than offset_diff
4178 *
4179 * We need to prepare all the bios here before we start any IO
4180 * to ensure the size we choose is acceptable to all devices.
4181 * The means one for each copy for write-out and an extra one for
4182 * read-in.
4183 * We store the read-in bio in ->master_bio and the others in
4184 * ->devs[x].bio and ->devs[x].repl_bio.
4185 */
4199 /* If restarting in the middle, skip the initial sectors */
4212 }
4213 }
4215 /* We don't use sector_nr to track where we are up to
4216 * as that doesn't work well for ->reshape_backwards.
4217 * So just use ->reshape_progress.
4218 */
4220 /* 'next' is the earliest device address that we might
4221 * write to for this chunk in the new layout
4222 */
4226 /* 'safe' is the last device address that we might read from
4227 * in the old layout after a restart
4228 */
4241 /* 'next' is after the last device address that we
4242 * might write to for this chunk in the new layout
4243 */
4246 /* 'safe' is the earliest device address that we might
4247 * read from in the old layout after a restart
4248 */
4251 /* Need to update metadata if 'next' might be beyond 'safe'
4252 * as that would possibly corrupt data
4253 */
4263 }
4267 /* Need to update reshape_position in metadata */
4273 else
4283 }
4286 }
4288 read_more:
4289 /* Now schedule reads for blocks from sector_nr to last */
4302 /* Cannot read from here, so need to record bad blocks
4303 * on all the target devices.
4304 */
4305 // FIXME
4309 }
4326 /* Now find the locations in the new layout */
4342 }
4355 }
4357 /* Now add as many pages as possible to all of these bios. */
4370 /* Didn't fit, must stop */
4374 /* Remove last page from this bio */
4378 }
4380 }
4383 }
4384 bio_full:
4387 /* Now submit the read */
4397 /* Now that we have done the whole section we can
4398 * update reshape_progress
4399 */
4402 else
4406 }
4412 {
4413 /* Reshape read completed. Hopefully we have a block
4414 * to write out.
4415 * If we got a read error then we do sync 1-page reads from
4416 * elsewhere until we find the data - or give up.
4417 */
4423 /* Reshape has been aborted */
4426 }
4428 /* We definitely have the data in the pages, schedule the
4429 * writes.
4430 */
4442 }
4450 }
4452 }
4455 {
4467 /* read-ahead size must cover two whole stripes, which is
4468 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4469 */
4476 }
4478 }
4482 {
4483 /* Use sync reads to get the blocks from somewhere else */
4509 sector_t addr;
4517 addr,
4523 failed:
4524 slot++;
4529 }
4531 /* couldn't read this block, must give up */
4535 }
4537 idx++;
4538 }
4540 }
4543 {
4558 }
4561 /* FIXME should record badblock */
4563 }
4567 }
4570 {
4576 }
4579 {
4591 }
4599 d++) {
4606 }
4607 }
4613 }
4616 {
4637 };
4640 {
4642 }
4645 {
4647 }