| blob | 149426cd1e8430e4a3332aeb36818cb8c25aa170 |
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>
32 /*
33 * RAID10 provides a combination of RAID0 and RAID1 functionality.
34 * The layout of data is defined by
35 * chunk_size
36 * raid_disks
37 * near_copies (stored in low byte of layout)
38 * far_copies (stored in second byte of layout)
39 * far_offset (stored in bit 16 of layout )
40 *
41 * The data to be stored is divided into chunks using chunksize.
42 * Each device is divided into far_copies sections.
43 * In each section, chunks are laid out in a style similar to raid0, but
44 * near_copies copies of each chunk is stored (each on a different drive).
45 * The starting device for each section is offset near_copies from the starting
46 * device of the previous section.
47 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
48 * drive.
49 * near_copies and far_copies must be at least one, and their product is at most
50 * raid_disks.
51 *
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of be very far apart
54 * on disk, there are adjacent stripes.
55 */
57 /*
58 * Number of guaranteed r10bios in case of extreme VM load:
59 */
60 #define NR_RAID10_BIOS 256
62 /* When there are this many requests queue to be written by
63 * the raid10 thread, we become 'congested' to provide back-pressure
64 * for writeback.
65 */
73 {
77 /* allocate a r10bio with room for raid_disks entries in the
78 * bios array */
80 }
83 {
85 }
87 /* Maximum size of each resync request */
88 #define RESYNC_BLOCK_SIZE (64*1024)
89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
90 /* amount of memory to reserve for resync requests */
91 #define RESYNC_WINDOW (1024*1024)
92 /* maximum number of concurrent requests, memory permitting */
93 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
95 /*
96 * When performing a resync, we need to read and compare, so
97 * we need as many pages are there are copies.
98 * When performing a recovery, we need 2 bios, one for read,
99 * one for write (we recover only one drive per r10buf)
100 *
101 */
103 {
117 else
120 /*
121 * Allocate bios.
122 */
134 }
135 /*
136 * Allocate RESYNC_PAGES data pages and attach them
137 * where needed.
138 */
145 /* we can share bv_page's during recovery */
157 }
158 }
162 out_free_pages:
169 out_free_bio:
174 }
177 }
180 {
192 }
194 }
198 }
200 }
203 {
215 }
216 }
219 {
224 }
227 {
233 }
236 {
246 /* wake up frozen array... */
250 }
252 /*
253 * raid_end_bio_io() is called when we have finished servicing a mirrored
254 * operation and are ready to return a success/failure code to the buffer
255 * cache layer.
256 */
258 {
275 /*
276 * Wake up any possible resync thread that waits for the device
277 * to go idle.
278 */
280 }
282 }
284 /*
285 * Update disk head position estimator based on IRQ completion info.
286 */
288 {
293 }
295 /*
296 * Find the disk number which triggered given bio
297 */
300 {
310 }
311 }
321 }
324 {
335 /*
336 * this branch is our 'one mirror IO has finished' event handler:
337 */
341 /*
342 * Set R10BIO_Uptodate in our master bio, so that
343 * we will return a good error code to the higher
344 * levels even if IO on some other mirrored buffer fails.
345 *
346 * The 'master' represents the composite IO operation to
347 * user-side. So if something waits for IO, then it will
348 * wait for the 'master' bio.
349 */
352 /* If all other devices that store this block have
353 * failed, we want to return the error upwards rather
354 * than fail the last device. Here we redefine
355 * "uptodate" to mean "Don't want to retry"
356 */
362 }
367 /*
368 * oops, read error - keep the refcount on the rdev
369 */
378 }
379 }
382 {
383 /* clear the bitmap if all writes complete successfully */
389 }
392 {
400 else
402 }
403 }
404 }
407 {
424 }
425 /*
426 * this branch is our 'one mirror IO has finished' event handler:
427 */
430 /* Never record new bad blocks to replacement,
431 * just fail it.
432 */
441 }
443 /*
444 * Set R10BIO_Uptodate in our master bio, so that
445 * we will return a good error code for to the higher
446 * levels even if IO on some other mirrored buffer fails.
447 *
448 * The 'master' represents the composite IO operation to
449 * user-side. So if something waits for IO, then it will
450 * wait for the 'master' bio.
451 */
452 sector_t first_bad;
455 /*
456 * Do not set R10BIO_Uptodate if the current device is
457 * rebuilding or Faulty. This is because we cannot use
458 * such device for properly reading the data back (we could
459 * potentially use it, if the current write would have felt
460 * before rdev->recovery_offset, but for simplicity we don't
461 * check this here.
462 */
467 /* Maybe we can clear some bad blocks. */
475 else
479 }
480 }
482 /*
483 *
484 * Let's see if all mirrored write operations have finished
485 * already.
486 */
490 }
492 /*
493 * RAID10 layout manager
494 * As well as the chunksize and raid_disks count, there are two
495 * parameters: near_copies and far_copies.
496 * near_copies * far_copies must be <= raid_disks.
497 * Normally one of these will be 1.
498 * If both are 1, we get raid0.
499 * If near_copies == raid_disks, we get raid1.
500 *
501 * Chunks are laid out in raid0 style with near_copies copies of the
502 * first chunk, followed by near_copies copies of the next chunk and
503 * so on.
504 * If far_copies > 1, then after 1/far_copies of the array has been assigned
505 * as described above, we start again with a device offset of near_copies.
506 * So we effectively have another copy of the whole array further down all
507 * the drives, but with blocks on different drives.
508 * With this layout, and block is never stored twice on the one device.
509 *
510 * raid10_find_phys finds the sector offset of a given virtual sector
511 * on each device that it is on.
512 *
513 * raid10_find_virt does the reverse mapping, from a device and a
514 * sector offset to a virtual address
515 */
518 {
520 sector_t sector;
521 sector_t chunk;
522 sector_t stripe;
527 /* now calculate first sector/dev */
539 /* and calculate all the others */
545 slot++;
554 slot++;
555 }
556 dev++;
560 }
561 }
563 }
566 {
582 else
584 }
586 }
590 }
592 /**
593 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
594 * @q: request queue
595 * @bvm: properties of new bio
596 * @biovec: the request that could be merged to it.
597 *
598 * Return amount of bytes we can accept at this offset
599 * This requires checking for end-of-chunk if near_copies != raid_disks,
600 * and for subordinate merge_bvec_fns if merge_check_needed.
601 */
605 {
617 /* bio_add cannot handle a negative return */
647 }
648 }
659 }
660 }
661 }
663 }
665 }
667 /*
668 * This routine returns the disk from which the requested read should
669 * be done. There is a per-array 'next expected sequential IO' sector
670 * number - if this matches on the next IO then we use the last disk.
671 * There is also a per-disk 'last know head position' sector that is
672 * maintained from IRQ contexts, both the normal and the resync IO
673 * completion handlers update this position correctly. If there is no
674 * perfect sequential match then we pick the disk whose head is closest.
675 *
676 * If there are 2 mirrors in the same 2 devices, performance degrades
677 * because position is mirror, not device based.
678 *
679 * The rdev for the device selected will have nr_pending incremented.
680 */
682 /*
683 * FIXME: possibly should rethink readbalancing and do it differently
684 * depending on near_copies / far_copies geometry.
685 */
689 {
701 retry:
708 /*
709 * Check if we can balance. We can balance on the whole
710 * device if no resync is going on (recovery is ok), or below
711 * the resync window. We take the first readable disk when
712 * above the resync window.
713 */
719 sector_t first_bad;
721 sector_t dev_sector;
743 /* Already have a better slot */
746 /* Cannot read here. If this is the
747 * 'primary' device, then we must not read
748 * beyond 'bad_sectors' from another device.
749 */
756 sector_t good_sectors =
762 }
764 /* Must read from here */
766 }
774 /* This optimisation is debatable, and completely destroys
775 * sequential read speed for 'far copies' arrays. So only
776 * keep it for 'near' arrays, and review those later.
777 */
781 /* for far > 1 always use the lowest address */
784 else
791 }
792 }
796 }
801 /* Cannot risk returning a device that failed
802 * before we inc'ed nr_pending
803 */
806 }
814 }
817 {
835 }
836 }
839 }
842 {
843 /* Any writes that have been queued but are awaiting
844 * bitmap updates get flushed here.
845 */
853 /* flush any pending bitmap writes to disk
854 * before proceeding w/ I/O */
863 }
866 }
868 /* Barriers....
869 * Sometimes we need to suspend IO while we do something else,
870 * either some resync/recovery, or reconfigure the array.
871 * To do this we raise a 'barrier'.
872 * The 'barrier' is a counter that can be raised multiple times
873 * to count how many activities are happening which preclude
874 * normal IO.
875 * We can only raise the barrier if there is no pending IO.
876 * i.e. if nr_pending == 0.
877 * We choose only to raise the barrier if no-one is waiting for the
878 * barrier to go down. This means that as soon as an IO request
879 * is ready, no other operations which require a barrier will start
880 * until the IO request has had a chance.
881 *
882 * So: regular IO calls 'wait_barrier'. When that returns there
883 * is no backgroup IO happening, It must arrange to call
884 * allow_barrier when it has finished its IO.
885 * backgroup IO calls must call raise_barrier. Once that returns
886 * there is no normal IO happeing. It must arrange to call
887 * lower_barrier when the particular background IO completes.
888 */
891 {
895 /* Wait until no block IO is waiting (unless 'force') */
899 /* block any new IO from starting */
902 /* Now wait for all pending IO to complete */
908 }
911 {
917 }
920 {
924 /* Wait for the barrier to drop.
925 * However if there are already pending
926 * requests (preventing the barrier from
927 * rising completely), and the
928 * pre-process bio queue isn't empty,
929 * then don't wait, as we need to empty
930 * that queue to get the nr_pending
931 * count down.
932 */
939 );
941 }
944 }
947 {
953 }
956 {
957 /* stop syncio and normal IO and wait for everything to
958 * go quiet.
959 * We increment barrier and nr_waiting, and then
960 * wait until nr_pending match nr_queued+extra
961 * This is called in the context of one normal IO request
962 * that has failed. Thus any sync request that might be pending
963 * will be blocked by nr_pending, and we need to wait for
964 * pending IO requests to complete or be queued for re-try.
965 * Thus the number queued (nr_queued) plus this request (extra)
966 * must match the number of pending IOs (nr_pending) before
967 * we continue.
968 */
978 }
981 {
982 /* reverse the effect of the freeze */
988 }
991 {
1009 }
1011 /* If this request crosses a chunk boundary, we need to
1012 * split it. This will only happen for 1 PAGE (or less) requests.
1013 */
1018 /* Sanity check -- queue functions should prevent this happening */
1022 /* This is a one page bio that upper layers
1023 * refuse to split for us, so we need to split it.
1024 */
1028 /* Each of these 'make_request' calls will call 'wait_barrier'.
1029 * If the first succeeds but the second blocks due to the resync
1030 * thread raising the barrier, we will deadlock because the
1031 * IO to the underlying device will be queued in generic_make_request
1032 * and will never complete, so will never reduce nr_pending.
1033 * So increment nr_waiting here so no new raise_barriers will
1034 * succeed, and so the second wait_barrier cannot block.
1035 */
1050 bad_map:
1057 }
1061 /*
1062 * Register the new request and wait if the reconstruction
1063 * thread has put up a bar for new requests.
1064 * Continue immediately if no resync is active currently.
1065 */
1077 /* We might need to issue multiple reads to different
1078 * devices if there are bad blocks around, so we keep
1079 * track of the number of reads in bio->bi_phys_segments.
1080 * If this is 0, there is only one r10_bio and no locking
1081 * will be needed when the request completes. If it is
1082 * non-zero, then it is the number of not-completed requests.
1083 */
1088 /*
1089 * read balancing logic:
1090 */
1094 read_again:
1099 }
1104 max_sectors);
1117 /* Could not read all from this device, so we will
1118 * need another r10_bio.
1119 */
1126 else
1129 /* Cannot call generic_make_request directly
1130 * as that will be queued in __generic_make_request
1131 * and subsequent mempool_alloc might block
1132 * waiting for it. so hand bio over to raid10d.
1133 */
1148 }
1150 /*
1151 * WRITE:
1152 */
1157 }
1158 /* first select target devices under rcu_lock and
1159 * inc refcount on their rdev. Record them by setting
1160 * bios[x] to bio
1161 * If there are known/acknowledged bad blocks on any device
1162 * on which we have seen a write error, we want to avoid
1163 * writing to those blocks. This potentially requires several
1164 * writes to write around the bad blocks. Each set of writes
1165 * gets its own r10_bio with a set of bios attached. The number
1166 * of r10_bios is recored in bio->bi_phys_segments just as with
1167 * the read case.
1168 */
1173 retry_write:
1189 }
1194 }
1208 }
1210 sector_t first_bad;
1216 max_sectors,
1219 /* Mustn't write here until the bad block
1220 * is acknowledged
1221 */
1226 }
1228 /* Cannot write here at all */
1231 /* Mustn't write more than bad_sectors
1232 * to other devices yet
1233 */
1235 /* We don't set R10BIO_Degraded as that
1236 * only applies if the disk is missing,
1237 * so it might be re-added, and we want to
1238 * know to recover this chunk.
1239 * In this case the device is here, and the
1240 * fact that this chunk is not in-sync is
1241 * recorded in the bad block log.
1242 */
1244 }
1249 }
1250 }
1254 }
1258 }
1259 }
1263 /* Have to wait for this device to get unblocked, then retry */
1271 }
1277 /* Race with remove_disk */
1280 }
1282 }
1283 }
1288 }
1291 /* We are splitting this into multiple parts, so
1292 * we need to prepare for allocating another r10_bio.
1293 */
1298 else
1301 }
1314 max_sectors);
1329 }
1334 /* Replacement just got moved to main 'rdev' */
1337 }
1340 max_sectors);
1355 }
1356 }
1358 /* Don't remove the bias on 'remaining' (one_write_done) until
1359 * after checking if we need to go around again.
1360 */
1364 /* We need another r10_bio. It has already been counted
1365 * in bio->bi_phys_segments.
1366 */
1376 }
1379 /* In case raid10d snuck in to freeze_array */
1384 }
1387 {
1398 else
1400 }
1408 }
1410 /* check if there are enough drives for
1411 * every block to appear on atleast one.
1412 * Don't consider the device numbered 'ignore'
1413 * as we might be about to remove it.
1414 */
1416 {
1426 cnt++;
1428 }
1434 }
1437 {
1441 /*
1442 * If it is not operational, then we have already marked it as dead
1443 * else if it is the last working disks, ignore the error, let the
1444 * next level up know.
1445 * else mark the drive as failed
1446 */
1449 /*
1450 * Don't fail the drive, just return an IO error.
1451 */
1458 /*
1459 * if recovery is running, make sure it aborts.
1460 */
1462 }
1471 }
1474 {
1482 }
1494 }
1495 }
1498 {
1504 }
1507 {
1514 /*
1515 * Find all non-in_sync disks within the RAID10 configuration
1516 * and mark them in_sync
1517 */
1524 /* Replacement has just become active */
1527 count++;
1529 /* Replaced device not technically faulty,
1530 * but we need to be sure it gets removed
1531 * and never re-added.
1532 */
1536 }
1542 count++;
1544 }
1545 }
1552 }
1556 {
1565 /* only hot-add to in-sync arrays, as recovery is
1566 * very different from resync
1567 */
1578 }
1583 else
1602 }
1615 }
1617 /* Some requests might not have seen this new
1618 * merge_bvec_fn. We must wait for them to complete
1619 * before merging the device fully.
1620 * First we make sure any code which has tested
1621 * our function has submitted the request, then
1622 * we wait for all outstanding requests to complete.
1623 */
1628 }
1632 }
1635 {
1647 else
1654 }
1655 /* Only remove faulty devices if recovery
1656 * is not possible.
1657 */
1664 }
1668 /* lost the race, try later */
1673 /* We must have just cleared 'rdev' */
1677 * but will never see neither -- if they are careful.
1678 */
1682 /* We might have just remove the Replacement as faulty
1683 * Clear the flag just in case
1684 */
1689 abort:
1693 }
1697 {
1706 else
1707 /* The write handler will notice the lack of
1708 * R10BIO_Uptodate and record any errors etc
1709 */
1713 /* for reconstruct, we always reschedule after a read.
1714 * for resync, only after all reads
1715 */
1719 /* we have read all the blocks,
1720 * do the comparison in process context in raid10d
1721 */
1723 }
1724 }
1727 {
1732 /* the primary of several recovery bios */
1737 else
1746 else
1749 }
1750 }
1751 }
1754 {
1760 sector_t first_bad;
1769 else
1781 }
1791 }
1793 /*
1794 * Note: sync and recover and handled very differently for raid10
1795 * This code is for resync.
1796 * For resync, we read through virtual addresses and read all blocks.
1797 * If there is any error, we schedule a write. The lowest numbered
1798 * drive is authoritative.
1799 * However requests come for physical address, so we need to map.
1800 * For every physical address there are raid_disks/copies virtual addresses,
1801 * which is always are least one, but is not necessarly an integer.
1802 * This means that a physical address can span multiple chunks, so we may
1803 * have to submit multiple io requests for a single sync request.
1804 */
1805 /*
1806 * We check if all blocks are in-sync and only write to blocks that
1807 * aren't in sync
1808 */
1810 {
1818 /* find the first device with a block */
1830 /* now find blocks with errors */
1841 /* We know that the bi_io_vec layout is the same for
1842 * both 'first' and 'i', so we just compare them.
1843 * All vec entries are PAGE_SIZE;
1844 */
1854 /* Don't fix anything. */
1856 }
1857 /* Ok, we need to write this bio, either to correct an
1858 * inconsistency or to correct an unreadable block.
1859 * First we need to fixup bv_offset, bv_len and
1860 * bi_vecs, as the read request might have corrupted these
1861 */
1879 PAGE_SIZE);
1880 }
1891 }
1893 /* Now write out to any replacement devices
1894 * that are active
1895 */
1907 PAGE_SIZE);
1913 }
1915 done:
1919 }
1920 }
1922 /*
1923 * Now for the recovery code.
1924 * Recovery happens across physical sectors.
1925 * We recover all non-is_sync drives by finding the virtual address of
1926 * each, and then choose a working drive that also has that virt address.
1927 * There is a separate r10_bio for each non-in_sync drive.
1928 * Only the first two slots are in use. The first for reading,
1929 * The second for writing.
1930 *
1931 */
1933 {
1934 /* We got a read error during recovery.
1935 * We repeat the read in smaller page-sized sections.
1936 * If a read succeeds, write it to the new device or record
1937 * a bad block if we cannot.
1938 * If a read fails, record a bad block on both old and
1939 * new devices.
1940 */
1953 sector_t addr;
1962 addr,
1970 addr,
1980 }
1981 }
1983 /* We don't worry if we cannot set a bad block -
1984 * it really is bad so there is no loss in not
1985 * recording it yet
1986 */
1990 /* need bad block on destination too */
1995 /* just abort the recovery */
1997 "md/raid10:%s: recovery aborted"
2006 }
2007 }
2008 }
2012 idx++;
2013 }
2014 }
2017 {
2026 }
2028 /*
2029 * share the pages with the first bio
2030 * and submit the write request
2031 */
2035 /* Need to test wbio2->bi_end_io before we call
2036 * generic_make_request as if the former is NULL,
2037 * the latter is free to free wbio2.
2038 */
2045 }
2051 }
2052 }
2055 /*
2056 * Used by fix_read_error() to decay the per rdev read_errors.
2057 * We halve the read error count for every hour that has elapsed
2058 * since the last recorded read error.
2059 *
2060 */
2062 {
2071 /* first time we've seen a read error */
2074 }
2081 /*
2082 * if hours_since_last is > the number of bits in read_errors
2083 * just set read errors to 0. We do this to avoid
2084 * overflowing the shift of read_errors by hours_since_last.
2085 */
2088 else
2090 }
2094 {
2095 sector_t first_bad;
2102 /* success */
2109 }
2110 /* need to record an error - either for the block or the device */
2114 }
2116 /*
2117 * This is a kernel thread which:
2118 *
2119 * 1. Retries failed read operations on working mirrors.
2120 * 2. Updates the raid superblock when problems encounter.
2121 * 3. Performs writes following reads for array synchronising.
2122 */
2125 {
2132 /* still own a reference to this rdev, so it cannot
2133 * have been cleared recently.
2134 */
2138 /* drive has already been failed, just ignore any
2139 more fix_read_error() attempts */
2149 "md/raid10:%s: %s: Raid device exceeded "
2159 }
2172 sector_t first_bad;
2186 sect,
2193 }
2194 sl++;
2201 /* Cannot read from anywhere, just mark the block
2202 * as bad on the first device to discourage future
2203 * reads.
2204 */
2209 rdev,
2216 }
2218 }
2221 /* write it back and re-read */
2228 sl--;
2240 sect,
2243 /* Well, this device is dead */
2245 "md/raid10:%s: read correction "
2246 "write failed"
2256 }
2259 }
2266 sl--;
2277 sect,
2279 READ)) {
2281 /* Well, this device is dead */
2283 "md/raid10:%s: unable to read back "
2284 "corrected sectors"
2297 "md/raid10:%s: read error corrected"
2304 }
2308 }
2313 }
2314 }
2317 {
2319 }
2322 {
2333 }
2336 {
2341 /* bio has the data to be written to slot 'i' where
2342 * we just recently had a write error.
2343 * We repeatedly clone the bio and trim down to one block,
2344 * then try the write. Where the write fails we record
2345 * a bad block.
2346 * It is conceivable that the bio doesn't exactly align with
2347 * blocks. We must handle this.
2348 *
2349 * We currently own a reference to the rdev.
2350 */
2353 sector_t sector;
2371 /* Write at 'sector' for 'sectors' */
2379 /* Failure! */
2388 }
2390 }
2393 {
2402 /* we got a read error. Maybe the drive is bad. Maybe just
2403 * the block and we can fix it.
2404 * We freeze all other IO, and try reading the block from
2405 * other devices. When we find one, we re-write
2406 * and check it that fixes the read error.
2407 * This is all done synchronously while the array is
2408 * frozen.
2409 */
2424 read_more:
2433 }
2438 KERN_ERR
2439 "md/raid10:%s: %s: redirecting "
2448 max_sectors);
2458 /* Drat - have to split this up more */
2467 else
2473 GFP_NOIO);
2487 }
2490 {
2491 /* Some sort of write request has finished and it
2492 * succeeded in writing where we thought there was a
2493 * bad block. So forget the bad block.
2494 * Or possibly if failed and we need to record
2495 * a bad block.
2496 */
2510 rdev,
2515 rdev,
2519 }
2526 rdev,
2531 rdev,
2535 }
2536 }
2545 rdev,
2555 }
2557 }
2562 rdev,
2566 }
2567 }
2572 }
2573 }
2576 {
2594 }
2612 /* just a partial read to be scheduled from a
2613 * separate context
2614 */
2617 }
2622 }
2624 }
2628 {
2643 }
2645 /*
2646 * perform a "sync" on one "block"
2647 *
2648 * We need to make sure that no normal I/O request - particularly write
2649 * requests - conflict with active sync requests.
2650 *
2651 * This is achieved by tracking pending requests and a 'barrier' concept
2652 * that can be installed to exclude normal IO requests.
2653 *
2654 * Resync and recovery are handled very differently.
2655 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2656 *
2657 * For resync, we iterate over virtual addresses, read all copies,
2658 * and update if there are differences. If only one copy is live,
2659 * skip it.
2660 * For recovery, we iterate over physical addresses, read a good
2661 * value for each non-in_sync drive, and over-write.
2662 *
2663 * So, for recovery we may have several outstanding complex requests for a
2664 * given address, one for each out-of-sync device. We model this by allocating
2665 * a number of r10_bio structures, one for each out-of-sync device.
2666 * As we setup these structures, we collect all bio's together into a list
2667 * which we then process collectively to add pages, and then process again
2668 * to pass to generic_make_request.
2669 *
2670 * The r10_bio structures are linked using a borrowed master_bio pointer.
2671 * This link is counted in ->remaining. When the r10_bio that points to NULL
2672 * has its remaining count decremented to 0, the whole complex operation
2673 * is complete.
2674 *
2675 */
2679 {
2686 sector_t sync_blocks;
2694 skipped:
2699 /* If we aborted, we need to abort the
2700 * sync on the 'current' bitmap chucks (there can
2701 * be several when recovering multiple devices).
2702 * as we may have started syncing it but not finished.
2703 * We can find the current address in
2704 * mddev->curr_resync, but for recovery,
2705 * we need to convert that to several
2706 * virtual addresses.
2707 */
2713 sector_t sect =
2717 }
2719 /* completed sync */
2723 /* Completed a full sync so the replacements
2724 * are now fully recovered.
2725 */
2729 ->recovery_offset
2731 }
2733 }
2738 }
2740 /* if there has been nothing to do on any drive,
2741 * then there is nothing to do at all..
2742 */
2745 }
2750 /* make sure whole request will fit in a chunk - if chunks
2751 * are meaningful
2752 */
2756 /*
2757 * If there is non-resync activity waiting for us then
2758 * put in a delay to throttle resync.
2759 */
2763 /* Again, very different code for resync and recovery.
2764 * Both must result in an r10bio with a list of bios that
2765 * have bi_end_io, bi_sector, bi_bdev set,
2766 * and bi_private set to the r10bio.
2767 * For recovery, we may actually create several r10bios
2768 * with 2 bios in each, that correspond to the bios in the main one.
2769 * In this case, the subordinate r10bios link back through a
2770 * borrowed master_bio pointer, and the counter in the master
2771 * includes a ref from each subordinate.
2772 */
2773 /* First, we decide what to do and set ->bi_end_io
2774 * To end_sync_read if we want to read, and
2775 * end_sync_write if we will want to write.
2776 */
2780 /* recovery... the complicated one */
2787 sector_t sect;
2794 &&
2801 /* want to reconstruct this device */
2805 /* last stripe is not complete - don't
2806 * try to recover this sector.
2807 */
2809 }
2810 /* Unless we are doing a full sync, or a replacement
2811 * we only need to recover the block if it is set in
2812 * the bitmap
2813 */
2821 /* yep, skip the sync_blocks here, but don't assume
2822 * that there will never be anything to do here
2823 */
2826 }
2841 /* Need to check if the array will still be
2842 * degraded
2843 */
2849 }
2865 /* This is where we read from */
2875 bad_sectors -= (sector
2880 }
2881 }
2892 /* and we write to 'i' (if not in_sync) */
2919 /* and maybe write to replacement */
2924 /* Note: if rdev != NULL, then bio
2925 * cannot be NULL as r10buf_pool_alloc will
2926 * have allocated it.
2927 * So the second test here is pointless.
2928 * But it keeps semantic-checkers happy, and
2929 * this comment keeps human reviewers
2930 * happy.
2931 */
2944 }
2946 /* Cannot recover, so abort the recovery or
2947 * record a bad block */
2949 /* problem is that there are bad blocks
2950 * on other device(s)
2951 */
2969 }
2976 mirror->recovery_disabled
2978 }
2984 }
2985 }
2992 }
2994 }
2996 /* resync. Schedule a read for every block at this virt offset */
3005 /* We can skip this block */
3008 }
3049 }
3050 }
3061 count++;
3068 /* Need to set up for writing to the replacement */
3082 count++;
3083 }
3090 mddev);
3095 mddev);
3096 }
3100 }
3101 }
3112 }
3130 /* stop here */
3135 /* remove last page from this bio */
3139 }
3141 }
3145 bio_full:
3159 }
3160 }
3163 /* pretend they weren't skipped, it makes
3164 * no important difference in this case
3165 */
3169 giveup:
3170 /* There is nowhere to write, so all non-sync
3171 * drives must be failed or in resync, all drives
3172 * have a bad block, so try the next chunk...
3173 */
3178 chunks_skipped ++;
3181 }
3183 static sector_t
3185 {
3186 sector_t size;
3200 }
3203 {
3204 /* Calculate the number of sectors-per-device that will
3205 * actually be used, and set conf->dev_sectors and
3206 * conf->stride
3207 */
3213 /* 'size' is now the number of chunks in the array */
3214 /* calculate "used chunks per device" */
3217 /* We need to round up when dividing by raid_disks to
3218 * get the stride size.
3219 */
3229 }
3230 }
3233 {
3244 }
3255 }
3263 GFP_KERNEL);
3300 out:
3309 }
3311 }
3314 {
3319 sector_t size;
3321 /*
3322 * copy the already verified devices into our private RAID10
3323 * bookkeeping area. [whatever we allocate in run(),
3324 * should be freed in stop()]
3325 */
3332 }
3344 else
3364 }
3373 }
3374 /* need to check that every block has at least one working mirror */
3379 }
3387 /* The replacement is all we have - use it */
3391 }
3399 }
3401 }
3411 /*
3412 * Ok, everything is just fine now
3413 */
3422 /* Calculate max read-ahead size.
3423 * We need to readahead at least twice a whole stripe....
3424 * maybe...
3425 */
3426 {
3432 }
3441 out_free_conf:
3449 out:
3451 }
3454 {
3468 }
3471 {
3481 }
3482 }
3485 {
3486 /* Resize of 'far' arrays is not supported.
3487 * For 'near' and 'offset' arrays we can set the
3488 * number of sectors used to be an appropriate multiple
3489 * of the chunk size.
3490 * For 'offset', this is far_copies*chunksize.
3491 * For 'near' the multiplier is the LCM of
3492 * near_copies and raid_disks.
3493 * So if far_copies > 1 && !far_offset, fail.
3494 * Else find LCM(raid_disks, near_copy)*far_copies and
3495 * multiply by chunk_size. Then round to this number.
3496 * This is mostly done by raid10_size()
3497 */
3515 }
3520 }
3523 {
3531 }
3533 /* Set new parameters */
3535 /* new layout: far_copies = 1, near_copies = 2 */
3540 /* make sure it will be not marked as dirty */
3549 }
3552 }
3555 {
3558 /* raid10 can take over:
3559 * raid0 - providing it has only two drives
3560 */
3562 /* for raid0 takeover only one zone is supported */
3569 }
3571 }
3573 }
3576 {
3593 };
3596 {
3598 }
3601 {
3603 }