| blob | 89111d455b71bae079715fb7c6c22782bee1ece9 |
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 */
953 }
956 }
959 {
965 }
968 {
969 /* stop syncio and normal IO and wait for everything to
970 * go quiet.
971 * We increment barrier and nr_waiting, and then
972 * wait until nr_pending match nr_queued+extra
973 * This is called in the context of one normal IO request
974 * that has failed. Thus any sync request that might be pending
975 * will be blocked by nr_pending, and we need to wait for
976 * pending IO requests to complete or be queued for re-try.
977 * Thus the number queued (nr_queued) plus this request (extra)
978 * must match the number of pending IOs (nr_pending) before
979 * we continue.
980 */
990 }
993 {
994 /* reverse the effect of the freeze */
1000 }
1004 {
1008 else
1010 }
1016 };
1019 {
1021 cb);
1035 }
1037 /* we aren't scheduling, so we can do the write-out directly. */
1047 /* Just ignore it */
1049 else
1052 }
1054 }
1057 {
1078 /*
1079 * Register the new request and wait if the reconstruction
1080 * thread has put up a bar for new requests.
1081 * Continue immediately if no resync is active currently.
1082 */
1089 /* IO spans the reshape position. Need to wait for
1090 * reshape to pass
1091 */
1096 sectors);
1098 }
1106 /* Need to update reshape_position in metadata */
1115 }
1126 /* We might need to issue multiple reads to different
1127 * devices if there are bad blocks around, so we keep
1128 * track of the number of reads in bio->bi_phys_segments.
1129 * If this is 0, there is only one r10_bio and no locking
1130 * will be needed when the request completes. If it is
1131 * non-zero, then it is the number of not-completed requests.
1132 */
1137 /*
1138 * read balancing logic:
1139 */
1143 read_again:
1148 }
1153 max_sectors);
1166 /* Could not read all from this device, so we will
1167 * need another r10_bio.
1168 */
1175 else
1178 /* Cannot call generic_make_request directly
1179 * as that will be queued in __generic_make_request
1180 * and subsequent mempool_alloc might block
1181 * waiting for it. so hand bio over to raid10d.
1182 */
1192 sectors_handled;
1197 }
1199 /*
1200 * WRITE:
1201 */
1206 }
1207 /* first select target devices under rcu_lock and
1208 * inc refcount on their rdev. Record them by setting
1209 * bios[x] to bio
1210 * If there are known/acknowledged bad blocks on any device
1211 * on which we have seen a write error, we want to avoid
1212 * writing to those blocks. This potentially requires several
1213 * writes to write around the bad blocks. Each set of writes
1214 * gets its own r10_bio with a set of bios attached. The number
1215 * of r10_bios is recored in bio->bi_phys_segments just as with
1216 * the read case.
1217 */
1221 retry_write:
1237 }
1242 }
1254 }
1256 sector_t first_bad;
1262 max_sectors,
1265 /* Mustn't write here until the bad block
1266 * is acknowledged
1267 */
1272 }
1274 /* Cannot write here at all */
1277 /* Mustn't write more than bad_sectors
1278 * to other devices yet
1279 */
1281 /* We don't set R10BIO_Degraded as that
1282 * only applies if the disk is missing,
1283 * so it might be re-added, and we want to
1284 * know to recover this chunk.
1285 * In this case the device is here, and the
1286 * fact that this chunk is not in-sync is
1287 * recorded in the bad block log.
1288 */
1290 }
1295 }
1296 }
1300 }
1304 }
1305 }
1309 /* Have to wait for this device to get unblocked, then retry */
1317 }
1323 /* Race with remove_disk */
1326 }
1328 }
1329 }
1334 }
1337 /* We are splitting this into multiple parts, so
1338 * we need to prepare for allocating another r10_bio.
1339 */
1344 else
1347 }
1361 max_sectors);
1366 rdev));
1379 cb);
1380 else
1389 }
1393 }
1398 /* Replacement just got moved to main 'rdev' */
1401 }
1404 max_sectors);
1422 cb);
1423 else
1432 }
1436 }
1437 }
1439 /* Don't remove the bias on 'remaining' (one_write_done) until
1440 * after checking if we need to go around again.
1441 */
1445 /* We need another r10_bio. It has already been counted
1446 * in bio->bi_phys_segments.
1447 */
1457 }
1459 }
1462 {
1472 }
1476 /*
1477 * If this request crosses a chunk boundary, we need to split
1478 * it.
1479 */
1492 }
1494 /*
1495 * If a bio is splitted, the first part of bio will pass
1496 * barrier but the bio is queued in current->bio_list (see
1497 * generic_make_request). If there is a raise_barrier() called
1498 * here, the second part of bio can't pass barrier. But since
1499 * the first part bio isn't dispatched to underlaying disks
1500 * yet, the barrier is never released, hence raise_barrier will
1501 * alays wait. We have a deadlock.
1502 * Note, this only happens in read path. For write path, the
1503 * first part of bio is dispatched in a schedule() call
1504 * (because of blk plug) or offloaded to raid10d.
1505 * Quitting from the function immediately can change the bio
1506 * order queued in bio_list and avoid the deadlock.
1507 */
1512 }
1515 /* In case raid10d snuck in to freeze_array */
1517 }
1520 {
1531 else
1535 }
1543 }
1545 /* check if there are enough drives for
1546 * every block to appear on atleast one.
1547 * Don't consider the device numbered 'ignore'
1548 * as we might be about to remove it.
1549 */
1551 {
1561 }
1573 cnt++;
1575 }
1581 out:
1584 }
1587 {
1588 /* when calling 'enough', both 'prev' and 'geo' must
1589 * be stable.
1590 * This is ensured if ->reconfig_mutex or ->device_lock
1591 * is held.
1592 */
1595 }
1598 {
1603 /*
1604 * If it is not operational, then we have already marked it as dead
1605 * else if it is the last working disks, ignore the error, let the
1606 * next level up know.
1607 * else mark the drive as failed
1608 */
1612 /*
1613 * Don't fail the drive, just return an IO error.
1614 */
1617 }
1620 /*
1621 * If recovery is running, make sure it aborts.
1622 */
1634 }
1637 {
1645 }
1657 }
1658 }
1661 {
1667 }
1670 {
1677 /*
1678 * Find all non-in_sync disks within the RAID10 configuration
1679 * and mark them in_sync
1680 */
1687 /* Replacement has just become active */
1690 count++;
1692 /* Replaced device not technically faulty,
1693 * but we need to be sure it gets removed
1694 * and never re-added.
1695 */
1699 }
1705 count++;
1707 }
1708 }
1715 }
1718 {
1726 /* only hot-add to in-sync arrays, as recovery is
1727 * very different from resync
1728 */
1742 else
1762 }
1776 }
1782 }
1785 {
1797 else
1804 }
1805 /* Only remove faulty devices if recovery
1806 * is not possible.
1807 */
1815 }
1819 /* lost the race, try later */
1824 /* We must have just cleared 'rdev' */
1828 * but will never see neither -- if they are careful.
1829 */
1833 /* We might have just remove the Replacement as faulty
1834 * Clear the flag just in case
1835 */
1840 abort:
1844 }
1847 {
1853 /* this is a reshape read */
1860 else
1861 /* The write handler will notice the lack of
1862 * R10BIO_Uptodate and record any errors etc
1863 */
1867 /* for reconstruct, we always reschedule after a read.
1868 * for resync, only after all reads
1869 */
1873 /* we have read all the blocks,
1874 * do the comparison in process context in raid10d
1875 */
1877 }
1878 }
1881 {
1886 /* the primary of several recovery bios */
1891 else
1900 else
1903 }
1904 }
1905 }
1908 {
1913 sector_t first_bad;
1922 else
1934 }
1944 }
1946 /*
1947 * Note: sync and recover and handled very differently for raid10
1948 * This code is for resync.
1949 * For resync, we read through virtual addresses and read all blocks.
1950 * If there is any error, we schedule a write. The lowest numbered
1951 * drive is authoritative.
1952 * However requests come for physical address, so we need to map.
1953 * For every physical address there are raid_disks/copies virtual addresses,
1954 * which is always are least one, but is not necessarly an integer.
1955 * This means that a physical address can span multiple chunks, so we may
1956 * have to submit multiple io requests for a single sync request.
1957 */
1958 /*
1959 * We check if all blocks are in-sync and only write to blocks that
1960 * aren't in sync
1961 */
1963 {
1971 /* find the first device with a block */
1985 /* now find blocks with errors */
1996 /* We know that the bi_io_vec layout is the same for
1997 * both 'first' and 'i', so we just compare them.
1998 * All vec entries are PAGE_SIZE;
1999 */
2007 len))
2010 }
2015 /* Don't fix anything. */
2017 }
2018 /* Ok, we need to write this bio, either to correct an
2019 * inconsistency or to correct an unreadable block.
2020 * First we need to fixup bv_offset, bv_len and
2021 * bi_vecs, as the read request might have corrupted these
2022 */
2042 }
2044 /* Now write out to any replacement devices
2045 * that are active
2046 */
2061 }
2063 done:
2067 }
2068 }
2070 /*
2071 * Now for the recovery code.
2072 * Recovery happens across physical sectors.
2073 * We recover all non-is_sync drives by finding the virtual address of
2074 * each, and then choose a working drive that also has that virt address.
2075 * There is a separate r10_bio for each non-in_sync drive.
2076 * Only the first two slots are in use. The first for reading,
2077 * The second for writing.
2078 *
2079 */
2081 {
2082 /* We got a read error during recovery.
2083 * We repeat the read in smaller page-sized sections.
2084 * If a read succeeds, write it to the new device or record
2085 * a bad block if we cannot.
2086 * If a read fails, record a bad block on both old and
2087 * new devices.
2088 */
2101 sector_t addr;
2110 addr,
2118 addr,
2128 }
2129 }
2131 /* We don't worry if we cannot set a bad block -
2132 * it really is bad so there is no loss in not
2133 * recording it yet
2134 */
2138 /* need bad block on destination too */
2143 /* just abort the recovery */
2145 "md/raid10:%s: recovery aborted"
2154 }
2155 }
2156 }
2160 idx++;
2161 }
2162 }
2165 {
2174 }
2176 /*
2177 * share the pages with the first bio
2178 * and submit the write request
2179 */
2183 /* Need to test wbio2->bi_end_io before we call
2184 * generic_make_request as if the former is NULL,
2185 * the latter is free to free wbio2.
2186 */
2193 }
2199 }
2200 }
2202 /*
2203 * Used by fix_read_error() to decay the per rdev read_errors.
2204 * We halve the read error count for every hour that has elapsed
2205 * since the last recorded read error.
2206 *
2207 */
2209 {
2218 /* first time we've seen a read error */
2221 }
2228 /*
2229 * if hours_since_last is > the number of bits in read_errors
2230 * just set read errors to 0. We do this to avoid
2231 * overflowing the shift of read_errors by hours_since_last.
2232 */
2235 else
2237 }
2241 {
2242 sector_t first_bad;
2249 /* success */
2256 }
2257 /* need to record an error - either for the block or the device */
2261 }
2263 /*
2264 * This is a kernel thread which:
2265 *
2266 * 1. Retries failed read operations on working mirrors.
2267 * 2. Updates the raid superblock when problems encounter.
2268 * 3. Performs writes following reads for array synchronising.
2269 */
2272 {
2279 /* still own a reference to this rdev, so it cannot
2280 * have been cleared recently.
2281 */
2285 /* drive has already been failed, just ignore any
2286 more fix_read_error() attempts */
2296 "md/raid10:%s: %s: Raid device exceeded "
2306 }
2319 sector_t first_bad;
2332 sect,
2339 }
2340 sl++;
2347 /* Cannot read from anywhere, just mark the block
2348 * as bad on the first device to discourage future
2349 * reads.
2350 */
2355 rdev,
2362 }
2364 }
2367 /* write it back and re-read */
2374 sl--;
2385 sect,
2388 /* Well, this device is dead */
2390 "md/raid10:%s: read correction "
2391 "write failed"
2395 sect +
2397 rdev)),
2403 }
2406 }
2413 sl--;
2424 sect,
2426 READ)) {
2428 /* Well, this device is dead */
2430 "md/raid10:%s: unable to read back "
2431 "corrected sectors"
2435 sect +
2445 "md/raid10:%s: read error corrected"
2449 sect +
2453 }
2457 }
2462 }
2463 }
2466 {
2471 /* bio has the data to be written to slot 'i' where
2472 * we just recently had a write error.
2473 * We repeatedly clone the bio and trim down to one block,
2474 * then try the write. Where the write fails we record
2475 * a bad block.
2476 * It is conceivable that the bio doesn't exactly align with
2477 * blocks. We must handle this.
2478 *
2479 * We currently own a reference to the rdev.
2480 */
2483 sector_t sector;
2502 /* Write at 'sector' for 'sectors' */
2510 /* Failure! */
2519 }
2521 }
2524 {
2533 /* we got a read error. Maybe the drive is bad. Maybe just
2534 * the block and we can fix it.
2535 * We freeze all other IO, and try reading the block from
2536 * other devices. When we find one, we re-write
2537 * and check it that fixes the read error.
2538 * This is all done synchronously while the array is
2539 * frozen.
2540 */
2555 read_more:
2564 }
2569 KERN_ERR
2570 "md/raid10:%s: %s: redirecting "
2587 /* Drat - have to split this up more */
2596 else
2602 GFP_NOIO);
2615 }
2618 {
2619 /* Some sort of write request has finished and it
2620 * succeeded in writing where we thought there was a
2621 * bad block. So forget the bad block.
2622 * Or possibly if failed and we need to record
2623 * a bad block.
2624 */
2638 rdev,
2643 rdev,
2647 }
2655 rdev,
2660 rdev,
2664 }
2665 }
2675 rdev,
2685 }
2687 }
2692 rdev,
2696 }
2697 }
2703 /*
2704 * In case freeze_array() is waiting for condition
2705 * nr_pending == nr_queued + extra to be true.
2706 */
2714 }
2715 }
2716 }
2719 {
2737 }
2738 }
2742 retry_list);
2751 }
2752 }
2763 }
2783 /* just a partial read to be scheduled from a
2784 * separate context
2785 */
2788 }
2793 }
2795 }
2798 {
2813 }
2815 /*
2816 * perform a "sync" on one "block"
2817 *
2818 * We need to make sure that no normal I/O request - particularly write
2819 * requests - conflict with active sync requests.
2820 *
2821 * This is achieved by tracking pending requests and a 'barrier' concept
2822 * that can be installed to exclude normal IO requests.
2823 *
2824 * Resync and recovery are handled very differently.
2825 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2826 *
2827 * For resync, we iterate over virtual addresses, read all copies,
2828 * and update if there are differences. If only one copy is live,
2829 * skip it.
2830 * For recovery, we iterate over physical addresses, read a good
2831 * value for each non-in_sync drive, and over-write.
2832 *
2833 * So, for recovery we may have several outstanding complex requests for a
2834 * given address, one for each out-of-sync device. We model this by allocating
2835 * a number of r10_bio structures, one for each out-of-sync device.
2836 * As we setup these structures, we collect all bio's together into a list
2837 * which we then process collectively to add pages, and then process again
2838 * to pass to generic_make_request.
2839 *
2840 * The r10_bio structures are linked using a borrowed master_bio pointer.
2841 * This link is counted in ->remaining. When the r10_bio that points to NULL
2842 * has its remaining count decremented to 0, the whole complex operation
2843 * is complete.
2844 *
2845 */
2849 {
2856 sector_t sync_blocks;
2865 /*
2866 * Allow skipping a full rebuild for incremental assembly
2867 * of a clean array, like RAID1 does.
2868 */
2878 }
2880 skipped:
2886 /* If we aborted, we need to abort the
2887 * sync on the 'current' bitmap chucks (there can
2888 * be several when recovering multiple devices).
2889 * as we may have started syncing it but not finished.
2890 * We can find the current address in
2891 * mddev->curr_resync, but for recovery,
2892 * we need to convert that to several
2893 * virtual addresses.
2894 */
2899 }
2906 sector_t sect =
2910 }
2912 /* completed sync */
2916 /* Completed a full sync so the replacements
2917 * are now fully recovered.
2918 */
2922 ->recovery_offset
2924 }
2926 }
2931 }
2937 /* if there has been nothing to do on any drive,
2938 * then there is nothing to do at all..
2939 */
2942 }
2947 /* make sure whole request will fit in a chunk - if chunks
2948 * are meaningful
2949 */
2954 /* Again, very different code for resync and recovery.
2955 * Both must result in an r10bio with a list of bios that
2956 * have bi_end_io, bi_sector, bi_bdev set,
2957 * and bi_private set to the r10bio.
2958 * For recovery, we may actually create several r10bios
2959 * with 2 bios in each, that correspond to the bios in the main one.
2960 * In this case, the subordinate r10bios link back through a
2961 * borrowed master_bio pointer, and the counter in the master
2962 * includes a ref from each subordinate.
2963 */
2964 /* First, we decide what to do and set ->bi_end_io
2965 * To end_sync_read if we want to read, and
2966 * end_sync_write if we will want to write.
2967 */
2971 /* recovery... the complicated one */
2978 sector_t sect;
2985 &&
2992 /* want to reconstruct this device */
2996 /* last stripe is not complete - don't
2997 * try to recover this sector.
2998 */
3000 }
3001 /* Unless we are doing a full sync, or a replacement
3002 * we only need to recover the block if it is set in
3003 * the bitmap
3004 */
3012 /* yep, skip the sync_blocks here, but don't assume
3013 * that there will never be anything to do here
3014 */
3017 }
3033 /* Need to check if the array will still be
3034 * degraded
3035 */
3041 }
3057 /* This is where we read from */
3067 bad_sectors -= (sector
3072 }
3073 }
3086 /* and we write to 'i' (if not in_sync) */
3114 /* and maybe write to replacement */
3119 /* Note: if rdev != NULL, then bio
3120 * cannot be NULL as r10buf_pool_alloc will
3121 * have allocated it.
3122 * So the second test here is pointless.
3123 * But it keeps semantic-checkers happy, and
3124 * this comment keeps human reviewers
3125 * happy.
3126 */
3141 }
3143 /* Cannot recover, so abort the recovery or
3144 * record a bad block */
3146 /* problem is that there are bad blocks
3147 * on other device(s)
3148 */
3166 }
3173 mirror->recovery_disabled
3175 }
3181 }
3182 }
3189 }
3191 }
3193 /* resync. Schedule a read for every block at this virt offset */
3202 /* We can skip this block */
3205 }
3247 }
3248 }
3259 count++;
3266 /* 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 }
3555 out:
3564 }
3566 }
3569 {
3574 sector_t size;
3584 }
3600 else
3603 }
3625 }
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 }
3699 }
3702 }
3712 /*
3713 * Ok, everything is just fine now
3714 */
3724 /* Calculate max read-ahead size.
3725 * We need to readahead at least twice a whole stripe....
3726 * maybe...
3727 */
3731 }
3745 /* This cannot work */
3748 }
3757 }
3761 out_free_conf:
3768 out:
3770 }
3773 {
3782 }
3785 {
3795 }
3796 }
3799 {
3800 /* Resize of 'far' arrays is not supported.
3801 * For 'near' and 'offset' arrays we can set the
3802 * number of sectors used to be an appropriate multiple
3803 * of the chunk size.
3804 * For 'offset', this is far_copies*chunksize.
3805 * For 'near' the multiplier is the LCM of
3806 * near_copies and raid_disks.
3807 * So if far_copies > 1 && !far_offset, fail.
3808 * Else find LCM(raid_disks, near_copy)*far_copies and
3809 * multiply by chunk_size. Then round to this number.
3810 * This is mostly done by raid10_size()
3811 */
3830 }
3838 }
3843 }
3846 {
3854 }
3857 /* Set new parameters */
3859 /* new layout: far_copies = 1, near_copies = 2 */
3864 /* make sure it will be not marked as dirty */
3874 }
3876 }
3879 }
3882 {
3885 /* raid10 can take over:
3886 * raid0 - providing it has only two drives
3887 */
3889 /* for raid0 takeover only one zone is supported */
3896 }
3900 }
3902 }
3905 {
3906 /* Called when there is a request to change
3907 * - layout (to ->new_layout)
3908 * - chunk size (to ->new_chunk_sectors)
3909 * - raid_disks (by delta_disks)
3910 * or when trying to restart a reshape that was ongoing.
3911 *
3912 * We need to validate the request and possibly allocate
3913 * space if that might be an issue later.
3914 *
3915 * Currently we reject any reshape of a 'far' mode array,
3916 * allow chunk size to change if new is generally acceptable,
3917 * allow raid_disks to increase, and allow
3918 * a switch between 'near' mode and 'offset' mode.
3919 */
3927 /* mustn't change number of copies */
3930 /* Cannot switch to 'far' mode */
3934 /* not factor of array size */
3943 /* allocate new 'mirrors' list */
3948 GFP_KERNEL);
3951 }
3953 }
3955 /*
3956 * Need to check if array has failed when deciding whether to:
3957 * - start an array
3958 * - remove non-faulty devices
3959 * - add a spare
3960 * - allow a reshape
3961 * This determination is simple when no reshape is happening.
3962 * However if there is a reshape, we need to carefully check
3963 * both the before and after sections.
3964 * This is because some failed devices may only affect one
3965 * of the two sections, and some non-in_sync devices may
3966 * be insync in the section most affected by failed devices.
3967 */
3969 {
3975 /* 'prev' section first */
3979 degraded++;
3981 /* When we can reduce the number of devices in
3982 * an array, this might not contribute to
3983 * 'degraded'. It does now.
3984 */
3985 degraded++;
3986 }
3995 degraded2++;
3997 /* If reshape is increasing the number of devices,
3998 * this section has already been recovered, so
3999 * it doesn't contribute to degraded.
4000 * else it does.
4001 */
4003 degraded2++;
4004 }
4005 }
4010 }
4013 {
4014 /* A 'reshape' has been requested. This commits
4015 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4016 * This also checks if there are enough spares and adds them
4017 * to the array.
4018 * We currently require enough spares to make the final
4019 * array non-degraded. We also require that the difference
4020 * between old and new data_offset - on each device - is
4021 * enough that we never risk over-writing.
4022 */
4047 spares++;
4058 }
4059 }
4077 }
4087 }
4102 }
4111 else
4116 }
4119 /* This is a spare that was manually added */
4121 }
4122 }
4123 /* When a reshape changes the number of devices,
4124 * ->degraded is measured against the larger of the
4125 * pre and post numbers.
4126 */
4145 }
4151 abort:
4164 }
4166 /* Calculate the last device-address that could contain
4167 * any block from the chunk that includes the array-address 's'
4168 * and report the next address.
4169 * i.e. the address returned will be chunk-aligned and after
4170 * any data that is in the chunk containing 's'.
4171 */
4173 {
4181 }
4183 /* Calculate the first device-address that could contain
4184 * any block from the chunk that includes the array-address 's'.
4185 * This too will be the start of a chunk
4186 */
4188 {
4195 }
4199 {
4200 /* We simply copy at most one chunk (smallest of old and new)
4201 * at a time, possibly less if that exceeds RESYNC_PAGES,
4202 * or we hit a bad block or something.
4203 * This might mean we pause for normal IO in the middle of
4204 * a chunk, but that is not a problem as mddev->reshape_position
4205 * can record any location.
4206 *
4207 * If we will want to write to a location that isn't
4208 * yet recorded as 'safe' (i.e. in metadata on disk) then
4209 * we need to flush all reshape requests and update the metadata.
4210 *
4211 * When reshaping forwards (e.g. to more devices), we interpret
4212 * 'safe' as the earliest block which might not have been copied
4213 * down yet. We divide this by previous stripe size and multiply
4214 * by previous stripe length to get lowest device offset that we
4215 * cannot write to yet.
4216 * We interpret 'sector_nr' as an address that we want to write to.
4217 * From this we use last_device_address() to find where we might
4218 * write to, and first_device_address on the 'safe' position.
4219 * If this 'next' write position is after the 'safe' position,
4220 * we must update the metadata to increase the 'safe' position.
4221 *
4222 * When reshaping backwards, we round in the opposite direction
4223 * and perform the reverse test: next write position must not be
4224 * less than current safe position.
4225 *
4226 * In all this the minimum difference in data offsets
4227 * (conf->offset_diff - always positive) allows a bit of slack,
4228 * so next can be after 'safe', but not by more than offset_diff
4229 *
4230 * We need to prepare all the bios here before we start any IO
4231 * to ensure the size we choose is acceptable to all devices.
4232 * The means one for each copy for write-out and an extra one for
4233 * read-in.
4234 * We store the read-in bio in ->master_bio and the others in
4235 * ->devs[x].bio and ->devs[x].repl_bio.
4236 */
4250 /* If restarting in the middle, skip the initial sectors */
4263 }
4264 }
4266 /* We don't use sector_nr to track where we are up to
4267 * as that doesn't work well for ->reshape_backwards.
4268 * So just use ->reshape_progress.
4269 */
4271 /* 'next' is the earliest device address that we might
4272 * write to for this chunk in the new layout
4273 */
4277 /* 'safe' is the last device address that we might read from
4278 * in the old layout after a restart
4279 */
4292 /* 'next' is after the last device address that we
4293 * might write to for this chunk in the new layout
4294 */
4297 /* 'safe' is the earliest device address that we might
4298 * read from in the old layout after a restart
4299 */
4302 /* Need to update metadata if 'next' might be beyond 'safe'
4303 * as that would possibly corrupt data
4304 */
4314 }
4318 /* Need to update reshape_position in metadata */
4324 else
4334 }
4337 }
4340 read_more:
4341 /* Now schedule reads for blocks from sector_nr to last */
4354 /* Cannot read from here, so need to record bad blocks
4355 * on all the target devices.
4356 */
4357 // FIXME
4361 }
4378 /* Now find the locations in the new layout */
4394 }
4407 }
4409 /* Now add as many pages as possible to all of these bios. */
4422 /* Didn't fit, must stop */
4426 /* Remove last page from this bio */
4430 }
4432 }
4435 }
4436 bio_full:
4439 /* Now submit the read */
4451 /* Now that we have done the whole section we can
4452 * update reshape_progress
4453 */
4456 else
4460 }
4466 {
4467 /* Reshape read completed. Hopefully we have a block
4468 * to write out.
4469 * If we got a read error then we do sync 1-page reads from
4470 * elsewhere until we find the data - or give up.
4471 */
4477 /* Reshape has been aborted */
4480 }
4482 /* We definitely have the data in the pages, schedule the
4483 * writes.
4484 */
4496 }
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 */
4563 sector_t addr;
4571 addr,
4577 failed:
4578 slot++;
4583 }
4585 /* couldn't read this block, must give up */
4589 }
4591 idx++;
4592 }
4594 }
4597 {
4612 }
4615 /* FIXME should record badblock */
4617 }
4621 }
4624 {
4630 }
4633 {
4645 }
4653 d++) {
4660 }
4661 }
4667 }
4670 {
4691 };
4694 {
4696 }
4699 {
4701 }