| blob | 6137d0041d4d49e88b60d88d78909377d3b2e41a |
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 */
637 }
638 }
649 }
650 }
651 }
653 }
655 }
657 /*
658 * This routine returns the disk from which the requested read should
659 * be done. There is a per-array 'next expected sequential IO' sector
660 * number - if this matches on the next IO then we use the last disk.
661 * There is also a per-disk 'last know head position' sector that is
662 * maintained from IRQ contexts, both the normal and the resync IO
663 * completion handlers update this position correctly. If there is no
664 * perfect sequential match then we pick the disk whose head is closest.
665 *
666 * If there are 2 mirrors in the same 2 devices, performance degrades
667 * because position is mirror, not device based.
668 *
669 * The rdev for the device selected will have nr_pending incremented.
670 */
672 /*
673 * FIXME: possibly should rethink readbalancing and do it differently
674 * depending on near_copies / far_copies geometry.
675 */
679 {
691 retry:
698 /*
699 * Check if we can balance. We can balance on the whole
700 * device if no resync is going on (recovery is ok), or below
701 * the resync window. We take the first readable disk when
702 * above the resync window.
703 */
709 sector_t first_bad;
711 sector_t dev_sector;
733 /* Already have a better slot */
736 /* Cannot read here. If this is the
737 * 'primary' device, then we must not read
738 * beyond 'bad_sectors' from another device.
739 */
746 sector_t good_sectors =
752 }
754 /* Must read from here */
756 }
764 /* This optimisation is debatable, and completely destroys
765 * sequential read speed for 'far copies' arrays. So only
766 * keep it for 'near' arrays, and review those later.
767 */
771 /* for far > 1 always use the lowest address */
774 else
781 }
782 }
786 }
791 /* Cannot risk returning a device that failed
792 * before we inc'ed nr_pending
793 */
796 }
804 }
807 {
825 }
826 }
829 }
832 {
833 /* Any writes that have been queued but are awaiting
834 * bitmap updates get flushed here.
835 */
843 /* flush any pending bitmap writes to disk
844 * before proceeding w/ I/O */
853 }
856 }
858 /* Barriers....
859 * Sometimes we need to suspend IO while we do something else,
860 * either some resync/recovery, or reconfigure the array.
861 * To do this we raise a 'barrier'.
862 * The 'barrier' is a counter that can be raised multiple times
863 * to count how many activities are happening which preclude
864 * normal IO.
865 * We can only raise the barrier if there is no pending IO.
866 * i.e. if nr_pending == 0.
867 * We choose only to raise the barrier if no-one is waiting for the
868 * barrier to go down. This means that as soon as an IO request
869 * is ready, no other operations which require a barrier will start
870 * until the IO request has had a chance.
871 *
872 * So: regular IO calls 'wait_barrier'. When that returns there
873 * is no backgroup IO happening, It must arrange to call
874 * allow_barrier when it has finished its IO.
875 * backgroup IO calls must call raise_barrier. Once that returns
876 * there is no normal IO happeing. It must arrange to call
877 * lower_barrier when the particular background IO completes.
878 */
881 {
885 /* Wait until no block IO is waiting (unless 'force') */
889 /* block any new IO from starting */
892 /* Now wait for all pending IO to complete */
898 }
901 {
907 }
910 {
914 /* Wait for the barrier to drop.
915 * However if there are already pending
916 * requests (preventing the barrier from
917 * rising completely), and the
918 * pre-process bio queue isn't empty,
919 * then don't wait, as we need to empty
920 * that queue to get the nr_pending
921 * count down.
922 */
929 );
931 }
934 }
937 {
943 }
946 {
947 /* stop syncio and normal IO and wait for everything to
948 * go quiet.
949 * We increment barrier and nr_waiting, and then
950 * wait until nr_pending match nr_queued+1
951 * This is called in the context of one normal IO request
952 * that has failed. Thus any sync request that might be pending
953 * will be blocked by nr_pending, and we need to wait for
954 * pending IO requests to complete or be queued for re-try.
955 * Thus the number queued (nr_queued) plus this request (1)
956 * must match the number of pending IOs (nr_pending) before
957 * we continue.
958 */
968 }
971 {
972 /* reverse the effect of the freeze */
978 }
981 {
999 }
1001 /* If this request crosses a chunk boundary, we need to
1002 * split it. This will only happen for 1 PAGE (or less) requests.
1003 */
1008 /* Sanity check -- queue functions should prevent this happening */
1012 /* This is a one page bio that upper layers
1013 * refuse to split for us, so we need to split it.
1014 */
1018 /* Each of these 'make_request' calls will call 'wait_barrier'.
1019 * If the first succeeds but the second blocks due to the resync
1020 * thread raising the barrier, we will deadlock because the
1021 * IO to the underlying device will be queued in generic_make_request
1022 * and will never complete, so will never reduce nr_pending.
1023 * So increment nr_waiting here so no new raise_barriers will
1024 * succeed, and so the second wait_barrier cannot block.
1025 */
1040 bad_map:
1047 }
1051 /*
1052 * Register the new request and wait if the reconstruction
1053 * thread has put up a bar for new requests.
1054 * Continue immediately if no resync is active currently.
1055 */
1067 /* We might need to issue multiple reads to different
1068 * devices if there are bad blocks around, so we keep
1069 * track of the number of reads in bio->bi_phys_segments.
1070 * If this is 0, there is only one r10_bio and no locking
1071 * will be needed when the request completes. If it is
1072 * non-zero, then it is the number of not-completed requests.
1073 */
1078 /*
1079 * read balancing logic:
1080 */
1084 read_again:
1089 }
1094 max_sectors);
1107 /* Could not read all from this device, so we will
1108 * need another r10_bio.
1109 */
1116 else
1119 /* Cannot call generic_make_request directly
1120 * as that will be queued in __generic_make_request
1121 * and subsequent mempool_alloc might block
1122 * waiting for it. so hand bio over to raid10d.
1123 */
1138 }
1140 /*
1141 * WRITE:
1142 */
1147 }
1148 /* first select target devices under rcu_lock and
1149 * inc refcount on their rdev. Record them by setting
1150 * bios[x] to bio
1151 * If there are known/acknowledged bad blocks on any device
1152 * on which we have seen a write error, we want to avoid
1153 * writing to those blocks. This potentially requires several
1154 * writes to write around the bad blocks. Each set of writes
1155 * gets its own r10_bio with a set of bios attached. The number
1156 * of r10_bios is recored in bio->bi_phys_segments just as with
1157 * the read case.
1158 */
1163 retry_write:
1179 }
1184 }
1198 }
1200 sector_t first_bad;
1206 max_sectors,
1209 /* Mustn't write here until the bad block
1210 * is acknowledged
1211 */
1216 }
1218 /* Cannot write here at all */
1221 /* Mustn't write more than bad_sectors
1222 * to other devices yet
1223 */
1225 /* We don't set R10BIO_Degraded as that
1226 * only applies if the disk is missing,
1227 * so it might be re-added, and we want to
1228 * know to recover this chunk.
1229 * In this case the device is here, and the
1230 * fact that this chunk is not in-sync is
1231 * recorded in the bad block log.
1232 */
1234 }
1239 }
1240 }
1244 }
1248 }
1249 }
1253 /* Have to wait for this device to get unblocked, then retry */
1261 }
1267 /* Race with remove_disk */
1270 }
1272 }
1273 }
1278 }
1281 /* We are splitting this into multiple parts, so
1282 * we need to prepare for allocating another r10_bio.
1283 */
1288 else
1291 }
1304 max_sectors);
1319 }
1324 /* Replacement just got moved to main 'rdev' */
1327 }
1330 max_sectors);
1345 }
1346 }
1348 /* Don't remove the bias on 'remaining' (one_write_done) until
1349 * after checking if we need to go around again.
1350 */
1354 /* We need another r10_bio. It has already been counted
1355 * in bio->bi_phys_segments.
1356 */
1366 }
1369 /* In case raid10d snuck in to freeze_array */
1374 }
1377 {
1388 else
1390 }
1398 }
1400 /* check if there are enough drives for
1401 * every block to appear on atleast one.
1402 * Don't consider the device numbered 'ignore'
1403 * as we might be about to remove it.
1404 */
1406 {
1415 cnt++;
1417 }
1422 }
1425 {
1429 /*
1430 * If it is not operational, then we have already marked it as dead
1431 * else if it is the last working disks, ignore the error, let the
1432 * next level up know.
1433 * else mark the drive as failed
1434 */
1437 /*
1438 * Don't fail the drive, just return an IO error.
1439 */
1446 /*
1447 * if recovery is running, make sure it aborts.
1448 */
1450 }
1459 }
1462 {
1470 }
1482 }
1483 }
1486 {
1492 }
1495 {
1502 /*
1503 * Find all non-in_sync disks within the RAID10 configuration
1504 * and mark them in_sync
1505 */
1512 /* Replacement has just become active */
1515 count++;
1517 /* Replaced device not technically faulty,
1518 * but we need to be sure it gets removed
1519 * and never re-added.
1520 */
1524 }
1529 count++;
1531 }
1532 }
1539 }
1543 {
1552 /* only hot-add to in-sync arrays, as recovery is
1553 * very different from resync
1554 */
1565 }
1570 else
1589 }
1602 }
1604 /* Some requests might not have seen this new
1605 * merge_bvec_fn. We must wait for them to complete
1606 * before merging the device fully.
1607 * First we make sure any code which has tested
1608 * our function has submitted the request, then
1609 * we wait for all outstanding requests to complete.
1610 */
1615 }
1619 }
1622 {
1634 else
1641 }
1642 /* Only remove faulty devices if recovery
1643 * is not possible.
1644 */
1651 }
1655 /* lost the race, try later */
1660 /* We must have just cleared 'rdev' */
1664 * but will never see neither -- if they are careful.
1665 */
1669 /* We might have just remove the Replacement as faulty
1670 * Clear the flag just in case
1671 */
1676 abort:
1680 }
1684 {
1693 else
1694 /* The write handler will notice the lack of
1695 * R10BIO_Uptodate and record any errors etc
1696 */
1700 /* for reconstruct, we always reschedule after a read.
1701 * for resync, only after all reads
1702 */
1706 /* we have read all the blocks,
1707 * do the comparison in process context in raid10d
1708 */
1710 }
1711 }
1714 {
1719 /* the primary of several recovery bios */
1724 else
1733 else
1736 }
1737 }
1738 }
1741 {
1747 sector_t first_bad;
1756 else
1768 }
1778 }
1780 /*
1781 * Note: sync and recover and handled very differently for raid10
1782 * This code is for resync.
1783 * For resync, we read through virtual addresses and read all blocks.
1784 * If there is any error, we schedule a write. The lowest numbered
1785 * drive is authoritative.
1786 * However requests come for physical address, so we need to map.
1787 * For every physical address there are raid_disks/copies virtual addresses,
1788 * which is always are least one, but is not necessarly an integer.
1789 * This means that a physical address can span multiple chunks, so we may
1790 * have to submit multiple io requests for a single sync request.
1791 */
1792 /*
1793 * We check if all blocks are in-sync and only write to blocks that
1794 * aren't in sync
1795 */
1797 {
1805 /* find the first device with a block */
1817 /* now find blocks with errors */
1828 /* We know that the bi_io_vec layout is the same for
1829 * both 'first' and 'i', so we just compare them.
1830 * All vec entries are PAGE_SIZE;
1831 */
1841 /* Don't fix anything. */
1843 }
1844 /* Ok, we need to write this bio, either to correct an
1845 * inconsistency or to correct an unreadable block.
1846 * First we need to fixup bv_offset, bv_len and
1847 * bi_vecs, as the read request might have corrupted these
1848 */
1866 PAGE_SIZE);
1867 }
1878 }
1880 /* Now write out to any replacement devices
1881 * that are active
1882 */
1894 PAGE_SIZE);
1900 }
1902 done:
1906 }
1907 }
1909 /*
1910 * Now for the recovery code.
1911 * Recovery happens across physical sectors.
1912 * We recover all non-is_sync drives by finding the virtual address of
1913 * each, and then choose a working drive that also has that virt address.
1914 * There is a separate r10_bio for each non-in_sync drive.
1915 * Only the first two slots are in use. The first for reading,
1916 * The second for writing.
1917 *
1918 */
1920 {
1921 /* We got a read error during recovery.
1922 * We repeat the read in smaller page-sized sections.
1923 * If a read succeeds, write it to the new device or record
1924 * a bad block if we cannot.
1925 * If a read fails, record a bad block on both old and
1926 * new devices.
1927 */
1940 sector_t addr;
1949 addr,
1957 addr,
1967 }
1968 }
1970 /* We don't worry if we cannot set a bad block -
1971 * it really is bad so there is no loss in not
1972 * recording it yet
1973 */
1977 /* need bad block on destination too */
1982 /* just abort the recovery */
1984 "md/raid10:%s: recovery aborted"
1993 }
1994 }
1995 }
1999 idx++;
2000 }
2001 }
2004 {
2013 }
2015 /*
2016 * share the pages with the first bio
2017 * and submit the write request
2018 */
2026 }
2032 }
2033 }
2036 /*
2037 * Used by fix_read_error() to decay the per rdev read_errors.
2038 * We halve the read error count for every hour that has elapsed
2039 * since the last recorded read error.
2040 *
2041 */
2043 {
2052 /* first time we've seen a read error */
2055 }
2062 /*
2063 * if hours_since_last is > the number of bits in read_errors
2064 * just set read errors to 0. We do this to avoid
2065 * overflowing the shift of read_errors by hours_since_last.
2066 */
2069 else
2071 }
2075 {
2076 sector_t first_bad;
2083 /* success */
2090 }
2091 /* need to record an error - either for the block or the device */
2095 }
2097 /*
2098 * This is a kernel thread which:
2099 *
2100 * 1. Retries failed read operations on working mirrors.
2101 * 2. Updates the raid superblock when problems encounter.
2102 * 3. Performs writes following reads for array synchronising.
2103 */
2106 {
2113 /* still own a reference to this rdev, so it cannot
2114 * have been cleared recently.
2115 */
2119 /* drive has already been failed, just ignore any
2120 more fix_read_error() attempts */
2130 "md/raid10:%s: %s: Raid device exceeded "
2140 }
2153 sector_t first_bad;
2167 sect,
2174 }
2175 sl++;
2182 /* Cannot read from anywhere, just mark the block
2183 * as bad on the first device to discourage future
2184 * reads.
2185 */
2190 rdev,
2197 }
2199 }
2202 /* write it back and re-read */
2209 sl--;
2221 sect,
2224 /* Well, this device is dead */
2226 "md/raid10:%s: read correction "
2227 "write failed"
2237 }
2240 }
2247 sl--;
2258 sect,
2260 READ)) {
2262 /* Well, this device is dead */
2264 "md/raid10:%s: unable to read back "
2265 "corrected sectors"
2278 "md/raid10:%s: read error corrected"
2285 }
2289 }
2294 }
2295 }
2298 {
2300 }
2303 {
2314 }
2317 {
2322 /* bio has the data to be written to slot 'i' where
2323 * we just recently had a write error.
2324 * We repeatedly clone the bio and trim down to one block,
2325 * then try the write. Where the write fails we record
2326 * a bad block.
2327 * It is conceivable that the bio doesn't exactly align with
2328 * blocks. We must handle this.
2329 *
2330 * We currently own a reference to the rdev.
2331 */
2334 sector_t sector;
2352 /* Write at 'sector' for 'sectors' */
2360 /* Failure! */
2369 }
2371 }
2374 {
2383 /* we got a read error. Maybe the drive is bad. Maybe just
2384 * the block and we can fix it.
2385 * We freeze all other IO, and try reading the block from
2386 * other devices. When we find one, we re-write
2387 * and check it that fixes the read error.
2388 * This is all done synchronously while the array is
2389 * frozen.
2390 */
2405 read_more:
2414 }
2419 KERN_ERR
2420 "md/raid10:%s: %s: redirecting "
2429 max_sectors);
2439 /* Drat - have to split this up more */
2448 else
2454 GFP_NOIO);
2468 }
2471 {
2472 /* Some sort of write request has finished and it
2473 * succeeded in writing where we thought there was a
2474 * bad block. So forget the bad block.
2475 * Or possibly if failed and we need to record
2476 * a bad block.
2477 */
2491 rdev,
2496 rdev,
2500 }
2507 rdev,
2512 rdev,
2516 }
2517 }
2526 rdev,
2536 }
2538 }
2543 rdev,
2547 }
2548 }
2553 }
2554 }
2557 {
2575 }
2593 /* just a partial read to be scheduled from a
2594 * separate context
2595 */
2598 }
2603 }
2605 }
2609 {
2624 }
2626 /*
2627 * perform a "sync" on one "block"
2628 *
2629 * We need to make sure that no normal I/O request - particularly write
2630 * requests - conflict with active sync requests.
2631 *
2632 * This is achieved by tracking pending requests and a 'barrier' concept
2633 * that can be installed to exclude normal IO requests.
2634 *
2635 * Resync and recovery are handled very differently.
2636 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2637 *
2638 * For resync, we iterate over virtual addresses, read all copies,
2639 * and update if there are differences. If only one copy is live,
2640 * skip it.
2641 * For recovery, we iterate over physical addresses, read a good
2642 * value for each non-in_sync drive, and over-write.
2643 *
2644 * So, for recovery we may have several outstanding complex requests for a
2645 * given address, one for each out-of-sync device. We model this by allocating
2646 * a number of r10_bio structures, one for each out-of-sync device.
2647 * As we setup these structures, we collect all bio's together into a list
2648 * which we then process collectively to add pages, and then process again
2649 * to pass to generic_make_request.
2650 *
2651 * The r10_bio structures are linked using a borrowed master_bio pointer.
2652 * This link is counted in ->remaining. When the r10_bio that points to NULL
2653 * has its remaining count decremented to 0, the whole complex operation
2654 * is complete.
2655 *
2656 */
2660 {
2667 sector_t sync_blocks;
2675 skipped:
2680 /* If we aborted, we need to abort the
2681 * sync on the 'current' bitmap chucks (there can
2682 * be several when recovering multiple devices).
2683 * as we may have started syncing it but not finished.
2684 * We can find the current address in
2685 * mddev->curr_resync, but for recovery,
2686 * we need to convert that to several
2687 * virtual addresses.
2688 */
2694 sector_t sect =
2698 }
2700 /* completed sync */
2704 /* Completed a full sync so the replacements
2705 * are now fully recovered.
2706 */
2710 ->recovery_offset
2712 }
2714 }
2719 }
2721 /* if there has been nothing to do on any drive,
2722 * then there is nothing to do at all..
2723 */
2726 }
2731 /* make sure whole request will fit in a chunk - if chunks
2732 * are meaningful
2733 */
2737 /*
2738 * If there is non-resync activity waiting for us then
2739 * put in a delay to throttle resync.
2740 */
2744 /* Again, very different code for resync and recovery.
2745 * Both must result in an r10bio with a list of bios that
2746 * have bi_end_io, bi_sector, bi_bdev set,
2747 * and bi_private set to the r10bio.
2748 * For recovery, we may actually create several r10bios
2749 * with 2 bios in each, that correspond to the bios in the main one.
2750 * In this case, the subordinate r10bios link back through a
2751 * borrowed master_bio pointer, and the counter in the master
2752 * includes a ref from each subordinate.
2753 */
2754 /* First, we decide what to do and set ->bi_end_io
2755 * To end_sync_read if we want to read, and
2756 * end_sync_write if we will want to write.
2757 */
2761 /* recovery... the complicated one */
2768 sector_t sect;
2775 &&
2782 /* want to reconstruct this device */
2786 /* last stripe is not complete - don't
2787 * try to recover this sector.
2788 */
2790 }
2791 /* Unless we are doing a full sync, or a replacement
2792 * we only need to recover the block if it is set in
2793 * the bitmap
2794 */
2802 /* yep, skip the sync_blocks here, but don't assume
2803 * that there will never be anything to do here
2804 */
2807 }
2822 /* Need to check if the array will still be
2823 * degraded
2824 */
2830 }
2846 /* This is where we read from */
2856 bad_sectors -= (sector
2861 }
2862 }
2873 /* and we write to 'i' (if not in_sync) */
2900 /* and maybe write to replacement */
2905 /* Note: if rdev != NULL, then bio
2906 * cannot be NULL as r10buf_pool_alloc will
2907 * have allocated it.
2908 * So the second test here is pointless.
2909 * But it keeps semantic-checkers happy, and
2910 * this comment keeps human reviewers
2911 * happy.
2912 */
2925 }
2927 /* Cannot recover, so abort the recovery or
2928 * record a bad block */
2934 /* problem is that there are bad blocks
2935 * on other device(s)
2936 */
2954 }
2961 mirror->recovery_disabled
2963 }
2965 }
2966 }
2973 }
2975 }
2977 /* resync. Schedule a read for every block at this virt offset */
2986 /* We can skip this block */
2989 }
3030 }
3031 }
3042 count++;
3049 /* Need to set up for writing to the replacement */
3063 count++;
3064 }
3071 mddev);
3076 mddev);
3077 }
3081 }
3082 }
3093 }
3111 /* stop here */
3116 /* remove last page from this bio */
3120 }
3122 }
3126 bio_full:
3140 }
3141 }
3144 /* pretend they weren't skipped, it makes
3145 * no important difference in this case
3146 */
3150 giveup:
3151 /* There is nowhere to write, so all non-sync
3152 * drives must be failed or in resync, all drives
3153 * have a bad block, so try the next chunk...
3154 */
3159 chunks_skipped ++;
3162 }
3164 static sector_t
3166 {
3167 sector_t size;
3181 }
3184 {
3185 /* Calculate the number of sectors-per-device that will
3186 * actually be used, and set conf->dev_sectors and
3187 * conf->stride
3188 */
3194 /* 'size' is now the number of chunks in the array */
3195 /* calculate "used chunks per device" */
3198 /* We need to round up when dividing by raid_disks to
3199 * get the stride size.
3200 */
3210 }
3211 }
3214 {
3225 }
3236 }
3244 GFP_KERNEL);
3281 out:
3290 }
3292 }
3295 {
3300 sector_t size;
3302 /*
3303 * copy the already verified devices into our private RAID10
3304 * bookkeeping area. [whatever we allocate in run(),
3305 * should be freed in stop()]
3306 */
3313 }
3325 else
3345 }
3354 }
3355 /* need to check that every block has at least one working mirror */
3360 }
3368 /* The replacement is all we have - use it */
3372 }
3380 }
3382 }
3392 /*
3393 * Ok, everything is just fine now
3394 */
3403 /* Calculate max read-ahead size.
3404 * We need to readahead at least twice a whole stripe....
3405 * maybe...
3406 */
3407 {
3413 }
3422 out_free_conf:
3430 out:
3432 }
3435 {
3449 }
3452 {
3462 }
3463 }
3466 {
3467 /* Resize of 'far' arrays is not supported.
3468 * For 'near' and 'offset' arrays we can set the
3469 * number of sectors used to be an appropriate multiple
3470 * of the chunk size.
3471 * For 'offset', this is far_copies*chunksize.
3472 * For 'near' the multiplier is the LCM of
3473 * near_copies and raid_disks.
3474 * So if far_copies > 1 && !far_offset, fail.
3475 * Else find LCM(raid_disks, near_copy)*far_copies and
3476 * multiply by chunk_size. Then round to this number.
3477 * This is mostly done by raid10_size()
3478 */
3496 }
3501 }
3504 {
3512 }
3514 /* Set new parameters */
3516 /* new layout: far_copies = 1, near_copies = 2 */
3521 /* make sure it will be not marked as dirty */
3530 }
3533 }
3536 {
3539 /* raid10 can take over:
3540 * raid0 - providing it has only two drives
3541 */
3543 /* for raid0 takeover only one zone is supported */
3550 }
3552 }
3554 }
3557 {
3574 };
3577 {
3579 }
3582 {
3584 }