| blob | 3540316886f2bca71a7bd943c4cf98de610ac1ae |
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;
457 /* Maybe we can clear some bad blocks. */
465 else
469 }
470 }
472 /*
473 *
474 * Let's see if all mirrored write operations have finished
475 * already.
476 */
480 }
482 /*
483 * RAID10 layout manager
484 * As well as the chunksize and raid_disks count, there are two
485 * parameters: near_copies and far_copies.
486 * near_copies * far_copies must be <= raid_disks.
487 * Normally one of these will be 1.
488 * If both are 1, we get raid0.
489 * If near_copies == raid_disks, we get raid1.
490 *
491 * Chunks are laid out in raid0 style with near_copies copies of the
492 * first chunk, followed by near_copies copies of the next chunk and
493 * so on.
494 * If far_copies > 1, then after 1/far_copies of the array has been assigned
495 * as described above, we start again with a device offset of near_copies.
496 * So we effectively have another copy of the whole array further down all
497 * the drives, but with blocks on different drives.
498 * With this layout, and block is never stored twice on the one device.
499 *
500 * raid10_find_phys finds the sector offset of a given virtual sector
501 * on each device that it is on.
502 *
503 * raid10_find_virt does the reverse mapping, from a device and a
504 * sector offset to a virtual address
505 */
508 {
510 sector_t sector;
511 sector_t chunk;
512 sector_t stripe;
517 /* now calculate first sector/dev */
529 /* and calculate all the others */
535 slot++;
544 slot++;
545 }
546 dev++;
550 }
551 }
553 }
556 {
572 else
574 }
576 }
580 }
582 /**
583 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
584 * @q: request queue
585 * @bvm: properties of new bio
586 * @biovec: the request that could be merged to it.
587 *
588 * Return amount of bytes we can accept at this offset
589 * This requires checking for end-of-chunk if near_copies != raid_disks,
590 * and for subordinate merge_bvec_fns if merge_check_needed.
591 */
595 {
607 /* bio_add cannot handle a negative return */
633 }
634 }
645 }
646 }
647 }
649 }
651 }
653 /*
654 * This routine returns the disk from which the requested read should
655 * be done. There is a per-array 'next expected sequential IO' sector
656 * number - if this matches on the next IO then we use the last disk.
657 * There is also a per-disk 'last know head position' sector that is
658 * maintained from IRQ contexts, both the normal and the resync IO
659 * completion handlers update this position correctly. If there is no
660 * perfect sequential match then we pick the disk whose head is closest.
661 *
662 * If there are 2 mirrors in the same 2 devices, performance degrades
663 * because position is mirror, not device based.
664 *
665 * The rdev for the device selected will have nr_pending incremented.
666 */
668 /*
669 * FIXME: possibly should rethink readbalancing and do it differently
670 * depending on near_copies / far_copies geometry.
671 */
675 {
687 retry:
694 /*
695 * Check if we can balance. We can balance on the whole
696 * device if no resync is going on (recovery is ok), or below
697 * the resync window. We take the first readable disk when
698 * above the resync window.
699 */
705 sector_t first_bad;
707 sector_t dev_sector;
729 /* Already have a better slot */
732 /* Cannot read here. If this is the
733 * 'primary' device, then we must not read
734 * beyond 'bad_sectors' from another device.
735 */
742 sector_t good_sectors =
748 }
750 /* Must read from here */
752 }
760 /* This optimisation is debatable, and completely destroys
761 * sequential read speed for 'far copies' arrays. So only
762 * keep it for 'near' arrays, and review those later.
763 */
767 /* for far > 1 always use the lowest address */
770 else
777 }
778 }
782 }
787 /* Cannot risk returning a device that failed
788 * before we inc'ed nr_pending
789 */
792 }
800 }
803 {
821 }
822 }
825 }
828 {
829 /* Any writes that have been queued but are awaiting
830 * bitmap updates get flushed here.
831 */
839 /* flush any pending bitmap writes to disk
840 * before proceeding w/ I/O */
849 }
852 }
854 /* Barriers....
855 * Sometimes we need to suspend IO while we do something else,
856 * either some resync/recovery, or reconfigure the array.
857 * To do this we raise a 'barrier'.
858 * The 'barrier' is a counter that can be raised multiple times
859 * to count how many activities are happening which preclude
860 * normal IO.
861 * We can only raise the barrier if there is no pending IO.
862 * i.e. if nr_pending == 0.
863 * We choose only to raise the barrier if no-one is waiting for the
864 * barrier to go down. This means that as soon as an IO request
865 * is ready, no other operations which require a barrier will start
866 * until the IO request has had a chance.
867 *
868 * So: regular IO calls 'wait_barrier'. When that returns there
869 * is no backgroup IO happening, It must arrange to call
870 * allow_barrier when it has finished its IO.
871 * backgroup IO calls must call raise_barrier. Once that returns
872 * there is no normal IO happeing. It must arrange to call
873 * lower_barrier when the particular background IO completes.
874 */
877 {
881 /* Wait until no block IO is waiting (unless 'force') */
885 /* block any new IO from starting */
888 /* Now wait for all pending IO to complete */
894 }
897 {
903 }
906 {
910 /* Wait for the barrier to drop.
911 * However if there are already pending
912 * requests (preventing the barrier from
913 * rising completely), and the
914 * pre-process bio queue isn't empty,
915 * then don't wait, as we need to empty
916 * that queue to get the nr_pending
917 * count down.
918 */
925 );
927 }
930 }
933 {
939 }
942 {
943 /* stop syncio and normal IO and wait for everything to
944 * go quiet.
945 * We increment barrier and nr_waiting, and then
946 * wait until nr_pending match nr_queued+1
947 * This is called in the context of one normal IO request
948 * that has failed. Thus any sync request that might be pending
949 * will be blocked by nr_pending, and we need to wait for
950 * pending IO requests to complete or be queued for re-try.
951 * Thus the number queued (nr_queued) plus this request (1)
952 * must match the number of pending IOs (nr_pending) before
953 * we continue.
954 */
964 }
967 {
968 /* reverse the effect of the freeze */
974 }
977 {
995 }
997 /* If this request crosses a chunk boundary, we need to
998 * split it. This will only happen for 1 PAGE (or less) requests.
999 */
1004 /* Sanity check -- queue functions should prevent this happening */
1008 /* This is a one page bio that upper layers
1009 * refuse to split for us, so we need to split it.
1010 */
1014 /* Each of these 'make_request' calls will call 'wait_barrier'.
1015 * If the first succeeds but the second blocks due to the resync
1016 * thread raising the barrier, we will deadlock because the
1017 * IO to the underlying device will be queued in generic_make_request
1018 * and will never complete, so will never reduce nr_pending.
1019 * So increment nr_waiting here so no new raise_barriers will
1020 * succeed, and so the second wait_barrier cannot block.
1021 */
1036 bad_map:
1043 }
1047 /*
1048 * Register the new request and wait if the reconstruction
1049 * thread has put up a bar for new requests.
1050 * Continue immediately if no resync is active currently.
1051 */
1063 /* We might need to issue multiple reads to different
1064 * devices if there are bad blocks around, so we keep
1065 * track of the number of reads in bio->bi_phys_segments.
1066 * If this is 0, there is only one r10_bio and no locking
1067 * will be needed when the request completes. If it is
1068 * non-zero, then it is the number of not-completed requests.
1069 */
1074 /*
1075 * read balancing logic:
1076 */
1080 read_again:
1085 }
1090 max_sectors);
1103 /* Could not read all from this device, so we will
1104 * need another r10_bio.
1105 */
1112 else
1115 /* Cannot call generic_make_request directly
1116 * as that will be queued in __generic_make_request
1117 * and subsequent mempool_alloc might block
1118 * waiting for it. so hand bio over to raid10d.
1119 */
1134 }
1136 /*
1137 * WRITE:
1138 */
1143 }
1144 /* first select target devices under rcu_lock and
1145 * inc refcount on their rdev. Record them by setting
1146 * bios[x] to bio
1147 * If there are known/acknowledged bad blocks on any device
1148 * on which we have seen a write error, we want to avoid
1149 * writing to those blocks. This potentially requires several
1150 * writes to write around the bad blocks. Each set of writes
1151 * gets its own r10_bio with a set of bios attached. The number
1152 * of r10_bios is recored in bio->bi_phys_segments just as with
1153 * the read case.
1154 */
1159 retry_write:
1175 }
1180 }
1191 }
1193 sector_t first_bad;
1199 max_sectors,
1202 /* Mustn't write here until the bad block
1203 * is acknowledged
1204 */
1209 }
1211 /* Cannot write here at all */
1214 /* Mustn't write more than bad_sectors
1215 * to other devices yet
1216 */
1218 /* We don't set R10BIO_Degraded as that
1219 * only applies if the disk is missing,
1220 * so it might be re-added, and we want to
1221 * know to recover this chunk.
1222 * In this case the device is here, and the
1223 * fact that this chunk is not in-sync is
1224 * recorded in the bad block log.
1225 */
1227 }
1232 }
1233 }
1239 }
1240 }
1244 /* Have to wait for this device to get unblocked, then retry */
1252 }
1258 /* Race with remove_disk */
1261 }
1263 }
1264 }
1269 }
1272 /* We are splitting this into multiple parts, so
1273 * we need to prepare for allocating another r10_bio.
1274 */
1279 else
1282 }
1296 max_sectors);
1317 max_sectors);
1320 /* We are actively writing to the original device
1321 * so it cannot disappear, so the replacement cannot
1322 * become NULL here
1323 */
1336 }
1338 /* Don't remove the bias on 'remaining' (one_write_done) until
1339 * after checking if we need to go around again.
1340 */
1344 /* We need another r10_bio. It has already been counted
1345 * in bio->bi_phys_segments.
1346 */
1356 }
1359 /* In case raid10d snuck in to freeze_array */
1364 }
1367 {
1378 else
1380 }
1388 }
1390 /* check if there are enough drives for
1391 * every block to appear on atleast one.
1392 * Don't consider the device numbered 'ignore'
1393 * as we might be about to remove it.
1394 */
1396 {
1405 cnt++;
1407 }
1412 }
1415 {
1419 /*
1420 * If it is not operational, then we have already marked it as dead
1421 * else if it is the last working disks, ignore the error, let the
1422 * next level up know.
1423 * else mark the drive as failed
1424 */
1427 /*
1428 * Don't fail the drive, just return an IO error.
1429 */
1436 /*
1437 * if recovery is running, make sure it aborts.
1438 */
1440 }
1449 }
1452 {
1460 }
1472 }
1473 }
1476 {
1482 }
1485 {
1492 /*
1493 * Find all non-in_sync disks within the RAID10 configuration
1494 * and mark them in_sync
1495 */
1502 /* Replacement has just become active */
1505 count++;
1507 /* Replaced device not technically faulty,
1508 * but we need to be sure it gets removed
1509 * and never re-added.
1510 */
1514 }
1519 count++;
1521 }
1522 }
1529 }
1533 {
1542 /* only hot-add to in-sync arrays, as recovery is
1543 * very different from resync
1544 */
1555 }
1560 else
1579 }
1592 }
1594 /* Some requests might not have seen this new
1595 * merge_bvec_fn. We must wait for them to complete
1596 * before merging the device fully.
1597 * First we make sure any code which has tested
1598 * our function has submitted the request, then
1599 * we wait for all outstanding requests to complete.
1600 */
1605 }
1609 }
1612 {
1624 else
1631 }
1632 /* Only remove faulty devices if recovery
1633 * is not possible.
1634 */
1641 }
1645 /* lost the race, try later */
1650 /* We must have just cleared 'rdev' */
1654 * but will never see neither -- if they are careful.
1655 */
1659 /* We might have just remove the Replacement as faulty
1660 * Clear the flag just in case
1661 */
1666 abort:
1670 }
1674 {
1683 else
1684 /* The write handler will notice the lack of
1685 * R10BIO_Uptodate and record any errors etc
1686 */
1690 /* for reconstruct, we always reschedule after a read.
1691 * for resync, only after all reads
1692 */
1696 /* we have read all the blocks,
1697 * do the comparison in process context in raid10d
1698 */
1700 }
1701 }
1704 {
1709 /* the primary of several recovery bios */
1714 else
1723 else
1726 }
1727 }
1728 }
1731 {
1737 sector_t first_bad;
1746 else
1758 }
1768 }
1770 /*
1771 * Note: sync and recover and handled very differently for raid10
1772 * This code is for resync.
1773 * For resync, we read through virtual addresses and read all blocks.
1774 * If there is any error, we schedule a write. The lowest numbered
1775 * drive is authoritative.
1776 * However requests come for physical address, so we need to map.
1777 * For every physical address there are raid_disks/copies virtual addresses,
1778 * which is always are least one, but is not necessarly an integer.
1779 * This means that a physical address can span multiple chunks, so we may
1780 * have to submit multiple io requests for a single sync request.
1781 */
1782 /*
1783 * We check if all blocks are in-sync and only write to blocks that
1784 * aren't in sync
1785 */
1787 {
1794 /* find the first device with a block */
1805 /* now find blocks with errors */
1817 /* We know that the bi_io_vec layout is the same for
1818 * both 'first' and 'i', so we just compare them.
1819 * All vec entries are PAGE_SIZE;
1820 */
1824 PAGE_SIZE))
1830 /* Don't fix anything. */
1832 }
1833 /* Ok, we need to write this bio, either to correct an
1834 * inconsistency or to correct an unreadable block.
1835 * First we need to fixup bv_offset, bv_len and
1836 * bi_vecs, as the read request might have corrupted these
1837 */
1855 PAGE_SIZE);
1856 }
1867 }
1869 /* Now write out to any replacement devices
1870 * that are active
1871 */
1884 PAGE_SIZE);
1890 }
1892 done:
1896 }
1897 }
1899 /*
1900 * Now for the recovery code.
1901 * Recovery happens across physical sectors.
1902 * We recover all non-is_sync drives by finding the virtual address of
1903 * each, and then choose a working drive that also has that virt address.
1904 * There is a separate r10_bio for each non-in_sync drive.
1905 * Only the first two slots are in use. The first for reading,
1906 * The second for writing.
1907 *
1908 */
1910 {
1911 /* We got a read error during recovery.
1912 * We repeat the read in smaller page-sized sections.
1913 * If a read succeeds, write it to the new device or record
1914 * a bad block if we cannot.
1915 * If a read fails, record a bad block on both old and
1916 * new devices.
1917 */
1930 sector_t addr;
1939 addr,
1947 addr,
1957 }
1958 }
1960 /* We don't worry if we cannot set a bad block -
1961 * it really is bad so there is no loss in not
1962 * recording it yet
1963 */
1967 /* need bad block on destination too */
1972 /* just abort the recovery */
1974 "md/raid10:%s: recovery aborted"
1983 }
1984 }
1985 }
1989 idx++;
1990 }
1991 }
1994 {
2003 }
2005 /*
2006 * share the pages with the first bio
2007 * and submit the write request
2008 */
2016 }
2022 }
2023 }
2026 /*
2027 * Used by fix_read_error() to decay the per rdev read_errors.
2028 * We halve the read error count for every hour that has elapsed
2029 * since the last recorded read error.
2030 *
2031 */
2033 {
2042 /* first time we've seen a read error */
2045 }
2052 /*
2053 * if hours_since_last is > the number of bits in read_errors
2054 * just set read errors to 0. We do this to avoid
2055 * overflowing the shift of read_errors by hours_since_last.
2056 */
2059 else
2061 }
2065 {
2066 sector_t first_bad;
2073 /* success */
2080 }
2081 /* need to record an error - either for the block or the device */
2085 }
2087 /*
2088 * This is a kernel thread which:
2089 *
2090 * 1. Retries failed read operations on working mirrors.
2091 * 2. Updates the raid superblock when problems encounter.
2092 * 3. Performs writes following reads for array synchronising.
2093 */
2096 {
2103 /* still own a reference to this rdev, so it cannot
2104 * have been cleared recently.
2105 */
2109 /* drive has already been failed, just ignore any
2110 more fix_read_error() attempts */
2120 "md/raid10:%s: %s: Raid device exceeded "
2130 }
2143 sector_t first_bad;
2157 sect,
2164 }
2165 sl++;
2172 /* Cannot read from anywhere, just mark the block
2173 * as bad on the first device to discourage future
2174 * reads.
2175 */
2180 rdev,
2187 }
2189 }
2192 /* write it back and re-read */
2199 sl--;
2211 sect,
2214 /* Well, this device is dead */
2216 "md/raid10:%s: read correction "
2217 "write failed"
2227 }
2230 }
2237 sl--;
2248 sect,
2250 READ)) {
2252 /* Well, this device is dead */
2254 "md/raid10:%s: unable to read back "
2255 "corrected sectors"
2268 "md/raid10:%s: read error corrected"
2275 }
2279 }
2284 }
2285 }
2288 {
2290 }
2293 {
2304 }
2307 {
2312 /* bio has the data to be written to slot 'i' where
2313 * we just recently had a write error.
2314 * We repeatedly clone the bio and trim down to one block,
2315 * then try the write. Where the write fails we record
2316 * a bad block.
2317 * It is conceivable that the bio doesn't exactly align with
2318 * blocks. We must handle this.
2319 *
2320 * We currently own a reference to the rdev.
2321 */
2324 sector_t sector;
2342 /* Write at 'sector' for 'sectors' */
2350 /* Failure! */
2359 }
2361 }
2364 {
2373 /* we got a read error. Maybe the drive is bad. Maybe just
2374 * the block and we can fix it.
2375 * We freeze all other IO, and try reading the block from
2376 * other devices. When we find one, we re-write
2377 * and check it that fixes the read error.
2378 * This is all done synchronously while the array is
2379 * frozen.
2380 */
2395 read_more:
2404 }
2409 KERN_ERR
2410 "md/raid10:%s: %s: redirecting"
2419 max_sectors);
2429 /* Drat - have to split this up more */
2438 else
2444 GFP_NOIO);
2458 }
2461 {
2462 /* Some sort of write request has finished and it
2463 * succeeded in writing where we thought there was a
2464 * bad block. So forget the bad block.
2465 * Or possibly if failed and we need to record
2466 * a bad block.
2467 */
2481 rdev,
2486 rdev,
2490 }
2497 rdev,
2502 rdev,
2506 }
2507 }
2516 rdev,
2526 }
2528 }
2533 rdev,
2537 }
2538 }
2543 }
2544 }
2547 {
2565 }
2583 /* just a partial read to be scheduled from a
2584 * separate context
2585 */
2588 }
2593 }
2595 }
2599 {
2614 }
2616 /*
2617 * perform a "sync" on one "block"
2618 *
2619 * We need to make sure that no normal I/O request - particularly write
2620 * requests - conflict with active sync requests.
2621 *
2622 * This is achieved by tracking pending requests and a 'barrier' concept
2623 * that can be installed to exclude normal IO requests.
2624 *
2625 * Resync and recovery are handled very differently.
2626 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2627 *
2628 * For resync, we iterate over virtual addresses, read all copies,
2629 * and update if there are differences. If only one copy is live,
2630 * skip it.
2631 * For recovery, we iterate over physical addresses, read a good
2632 * value for each non-in_sync drive, and over-write.
2633 *
2634 * So, for recovery we may have several outstanding complex requests for a
2635 * given address, one for each out-of-sync device. We model this by allocating
2636 * a number of r10_bio structures, one for each out-of-sync device.
2637 * As we setup these structures, we collect all bio's together into a list
2638 * which we then process collectively to add pages, and then process again
2639 * to pass to generic_make_request.
2640 *
2641 * The r10_bio structures are linked using a borrowed master_bio pointer.
2642 * This link is counted in ->remaining. When the r10_bio that points to NULL
2643 * has its remaining count decremented to 0, the whole complex operation
2644 * is complete.
2645 *
2646 */
2650 {
2657 sector_t sync_blocks;
2665 skipped:
2670 /* If we aborted, we need to abort the
2671 * sync on the 'current' bitmap chucks (there can
2672 * be several when recovering multiple devices).
2673 * as we may have started syncing it but not finished.
2674 * We can find the current address in
2675 * mddev->curr_resync, but for recovery,
2676 * we need to convert that to several
2677 * virtual addresses.
2678 */
2684 sector_t sect =
2688 }
2690 /* completed sync */
2694 /* Completed a full sync so the replacements
2695 * are now fully recovered.
2696 */
2700 ->recovery_offset
2702 }
2704 }
2709 }
2711 /* if there has been nothing to do on any drive,
2712 * then there is nothing to do at all..
2713 */
2716 }
2721 /* make sure whole request will fit in a chunk - if chunks
2722 * are meaningful
2723 */
2727 /*
2728 * If there is non-resync activity waiting for us then
2729 * put in a delay to throttle resync.
2730 */
2734 /* Again, very different code for resync and recovery.
2735 * Both must result in an r10bio with a list of bios that
2736 * have bi_end_io, bi_sector, bi_bdev set,
2737 * and bi_private set to the r10bio.
2738 * For recovery, we may actually create several r10bios
2739 * with 2 bios in each, that correspond to the bios in the main one.
2740 * In this case, the subordinate r10bios link back through a
2741 * borrowed master_bio pointer, and the counter in the master
2742 * includes a ref from each subordinate.
2743 */
2744 /* First, we decide what to do and set ->bi_end_io
2745 * To end_sync_read if we want to read, and
2746 * end_sync_write if we will want to write.
2747 */
2751 /* recovery... the complicated one */
2758 sector_t sect;
2765 &&
2772 /* want to reconstruct this device */
2775 /* Unless we are doing a full sync, or a replacement
2776 * we only need to recover the block if it is set in
2777 * the bitmap
2778 */
2786 /* yep, skip the sync_blocks here, but don't assume
2787 * that there will never be anything to do here
2788 */
2791 }
2806 /* Need to check if the array will still be
2807 * degraded
2808 */
2814 }
2830 /* This is where we read from */
2840 bad_sectors -= (sector
2845 }
2846 }
2857 /* and we write to 'i' (if not in_sync) */
2884 /* and maybe write to replacement */
2889 /* Note: if rdev != NULL, then bio
2890 * cannot be NULL as r10buf_pool_alloc will
2891 * have allocated it.
2892 * So the second test here is pointless.
2893 * But it keeps semantic-checkers happy, and
2894 * this comment keeps human reviewers
2895 * happy.
2896 */
2909 }
2911 /* Cannot recover, so abort the recovery or
2912 * record a bad block */
2918 /* problem is that there are bad blocks
2919 * on other device(s)
2920 */
2938 }
2945 mirror->recovery_disabled
2947 }
2949 }
2950 }
2957 }
2959 }
2961 /* resync. Schedule a read for every block at this virt offset */
2970 /* We can skip this block */
2973 }
3014 }
3015 }
3026 count++;
3033 /* Need to set up for writing to the replacement */
3047 count++;
3048 }
3055 mddev);
3060 mddev);
3061 }
3065 }
3066 }
3077 }
3095 /* stop here */
3100 /* remove last page from this bio */
3104 }
3106 }
3110 bio_full:
3124 }
3125 }
3128 /* pretend they weren't skipped, it makes
3129 * no important difference in this case
3130 */
3134 giveup:
3135 /* There is nowhere to write, so all non-sync
3136 * drives must be failed or in resync, all drives
3137 * have a bad block, so try the next chunk...
3138 */
3143 chunks_skipped ++;
3146 }
3148 static sector_t
3150 {
3151 sector_t size;
3165 }
3169 {
3181 }
3192 }
3200 GFP_KERNEL);
3226 /* 'size' is now the number of chunks in the array */
3227 /* calculate "used chunks per device" in 'stride' */
3230 /* We need to round up when dividing by raid_disks to
3231 * get the stride size.
3232 */
3240 else
3258 out:
3267 }
3269 }
3272 {
3277 sector_t size;
3279 /*
3280 * copy the already verified devices into our private RAID10
3281 * bookkeeping area. [whatever we allocate in run(),
3282 * should be freed in stop()]
3283 */
3290 }
3302 else
3322 }
3328 }
3329 /* need to check that every block has at least one working mirror */
3334 }
3342 /* The replacement is all we have - use it */
3346 }
3354 }
3356 }
3366 /*
3367 * Ok, everything is just fine now
3368 */
3377 /* Calculate max read-ahead size.
3378 * We need to readahead at least twice a whole stripe....
3379 * maybe...
3380 */
3381 {
3387 }
3396 out_free_conf:
3404 out:
3406 }
3409 {
3423 }
3426 {
3436 }
3437 }
3440 {
3441 /* Resize of 'far' arrays is not supported.
3442 * For 'near' and 'offset' arrays we can set the
3443 * number of sectors used to be an appropriate multiple
3444 * of the chunk size.
3445 * For 'offset', this is far_copies*chunksize.
3446 * For 'near' the multiplier is the LCM of
3447 * near_copies and raid_disks.
3448 * So if far_copies > 1 && !far_offset, fail.
3449 * Else find LCM(raid_disks, near_copy)*far_copies and
3450 * multiply by chunk_size. Then round to this number.
3451 * This is mostly done by raid10_size()
3452 */
3470 }
3474 }
3477 {
3485 }
3487 /* Set new parameters */
3489 /* new layout: far_copies = 1, near_copies = 2 */
3494 /* make sure it will be not marked as dirty */
3503 }
3506 }
3509 {
3512 /* raid10 can take over:
3513 * raid0 - providing it has only two drives
3514 */
3516 /* for raid0 takeover only one zone is supported */
3523 }
3525 }
3527 }
3530 {
3547 };
3550 {
3552 }
3555 {
3557 }