| blob | 189838956aa992dab25079bf453ad80ef03ea399 |
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 * If near_copies == raid_disk, there are no striping issues,
590 * but in that case, the function isn't called at all.
591 */
595 {
606 else
608 }
610 /*
611 * This routine returns the disk from which the requested read should
612 * be done. There is a per-array 'next expected sequential IO' sector
613 * number - if this matches on the next IO then we use the last disk.
614 * There is also a per-disk 'last know head position' sector that is
615 * maintained from IRQ contexts, both the normal and the resync IO
616 * completion handlers update this position correctly. If there is no
617 * perfect sequential match then we pick the disk whose head is closest.
618 *
619 * If there are 2 mirrors in the same 2 devices, performance degrades
620 * because position is mirror, not device based.
621 *
622 * The rdev for the device selected will have nr_pending incremented.
623 */
625 /*
626 * FIXME: possibly should rethink readbalancing and do it differently
627 * depending on near_copies / far_copies geometry.
628 */
632 {
644 retry:
651 /*
652 * Check if we can balance. We can balance on the whole
653 * device if no resync is going on (recovery is ok), or below
654 * the resync window. We take the first readable disk when
655 * above the resync window.
656 */
662 sector_t first_bad;
664 sector_t dev_sector;
685 /* Already have a better slot */
688 /* Cannot read here. If this is the
689 * 'primary' device, then we must not read
690 * beyond 'bad_sectors' from another device.
691 */
698 sector_t good_sectors =
704 }
706 /* Must read from here */
708 }
716 /* This optimisation is debatable, and completely destroys
717 * sequential read speed for 'far copies' arrays. So only
718 * keep it for 'near' arrays, and review those later.
719 */
723 /* for far > 1 always use the lowest address */
726 else
733 }
734 }
738 }
743 /* Cannot risk returning a device that failed
744 * before we inc'ed nr_pending
745 */
748 }
756 }
759 {
777 }
778 }
781 }
784 {
785 /* Any writes that have been queued but are awaiting
786 * bitmap updates get flushed here.
787 */
795 /* flush any pending bitmap writes to disk
796 * before proceeding w/ I/O */
805 }
808 }
810 /* Barriers....
811 * Sometimes we need to suspend IO while we do something else,
812 * either some resync/recovery, or reconfigure the array.
813 * To do this we raise a 'barrier'.
814 * The 'barrier' is a counter that can be raised multiple times
815 * to count how many activities are happening which preclude
816 * normal IO.
817 * We can only raise the barrier if there is no pending IO.
818 * i.e. if nr_pending == 0.
819 * We choose only to raise the barrier if no-one is waiting for the
820 * barrier to go down. This means that as soon as an IO request
821 * is ready, no other operations which require a barrier will start
822 * until the IO request has had a chance.
823 *
824 * So: regular IO calls 'wait_barrier'. When that returns there
825 * is no backgroup IO happening, It must arrange to call
826 * allow_barrier when it has finished its IO.
827 * backgroup IO calls must call raise_barrier. Once that returns
828 * there is no normal IO happeing. It must arrange to call
829 * lower_barrier when the particular background IO completes.
830 */
833 {
837 /* Wait until no block IO is waiting (unless 'force') */
841 /* block any new IO from starting */
844 /* Now wait for all pending IO to complete */
850 }
853 {
859 }
862 {
866 /* Wait for the barrier to drop.
867 * However if there are already pending
868 * requests (preventing the barrier from
869 * rising completely), and the
870 * pre-process bio queue isn't empty,
871 * then don't wait, as we need to empty
872 * that queue to get the nr_pending
873 * count down.
874 */
881 );
883 }
886 }
889 {
895 }
898 {
899 /* stop syncio and normal IO and wait for everything to
900 * go quiet.
901 * We increment barrier and nr_waiting, and then
902 * wait until nr_pending match nr_queued+1
903 * This is called in the context of one normal IO request
904 * that has failed. Thus any sync request that might be pending
905 * will be blocked by nr_pending, and we need to wait for
906 * pending IO requests to complete or be queued for re-try.
907 * Thus the number queued (nr_queued) plus this request (1)
908 * must match the number of pending IOs (nr_pending) before
909 * we continue.
910 */
920 }
923 {
924 /* reverse the effect of the freeze */
930 }
933 {
951 }
953 /* If this request crosses a chunk boundary, we need to
954 * split it. This will only happen for 1 PAGE (or less) requests.
955 */
960 /* Sanity check -- queue functions should prevent this happening */
964 /* This is a one page bio that upper layers
965 * refuse to split for us, so we need to split it.
966 */
970 /* Each of these 'make_request' calls will call 'wait_barrier'.
971 * If the first succeeds but the second blocks due to the resync
972 * thread raising the barrier, we will deadlock because the
973 * IO to the underlying device will be queued in generic_make_request
974 * and will never complete, so will never reduce nr_pending.
975 * So increment nr_waiting here so no new raise_barriers will
976 * succeed, and so the second wait_barrier cannot block.
977 */
992 bad_map:
999 }
1003 /*
1004 * Register the new request and wait if the reconstruction
1005 * thread has put up a bar for new requests.
1006 * Continue immediately if no resync is active currently.
1007 */
1019 /* We might need to issue multiple reads to different
1020 * devices if there are bad blocks around, so we keep
1021 * track of the number of reads in bio->bi_phys_segments.
1022 * If this is 0, there is only one r10_bio and no locking
1023 * will be needed when the request completes. If it is
1024 * non-zero, then it is the number of not-completed requests.
1025 */
1030 /*
1031 * read balancing logic:
1032 */
1036 read_again:
1041 }
1046 max_sectors);
1059 /* Could not read all from this device, so we will
1060 * need another r10_bio.
1061 */
1068 else
1071 /* Cannot call generic_make_request directly
1072 * as that will be queued in __generic_make_request
1073 * and subsequent mempool_alloc might block
1074 * waiting for it. so hand bio over to raid10d.
1075 */
1090 }
1092 /*
1093 * WRITE:
1094 */
1099 }
1100 /* first select target devices under rcu_lock and
1101 * inc refcount on their rdev. Record them by setting
1102 * bios[x] to bio
1103 * If there are known/acknowledged bad blocks on any device
1104 * on which we have seen a write error, we want to avoid
1105 * writing to those blocks. This potentially requires several
1106 * writes to write around the bad blocks. Each set of writes
1107 * gets its own r10_bio with a set of bios attached. The number
1108 * of r10_bios is recored in bio->bi_phys_segments just as with
1109 * the read case.
1110 */
1115 retry_write:
1131 }
1136 }
1145 }
1147 sector_t first_bad;
1153 max_sectors,
1156 /* Mustn't write here until the bad block
1157 * is acknowledged
1158 */
1163 }
1165 /* Cannot write here at all */
1168 /* Mustn't write more than bad_sectors
1169 * to other devices yet
1170 */
1172 /* We don't set R10BIO_Degraded as that
1173 * only applies if the disk is missing,
1174 * so it might be re-added, and we want to
1175 * know to recover this chunk.
1176 * In this case the device is here, and the
1177 * fact that this chunk is not in-sync is
1178 * recorded in the bad block log.
1179 */
1181 }
1186 }
1187 }
1193 }
1194 }
1198 /* Have to wait for this device to get unblocked, then retry */
1206 }
1212 /* Race with remove_disk */
1215 }
1217 }
1218 }
1223 }
1226 /* We are splitting this into multiple parts, so
1227 * we need to prepare for allocating another r10_bio.
1228 */
1233 else
1236 }
1250 max_sectors);
1271 max_sectors);
1274 /* We are actively writing to the original device
1275 * so it cannot disappear, so the replacement cannot
1276 * become NULL here
1277 */
1290 }
1292 /* Don't remove the bias on 'remaining' (one_write_done) until
1293 * after checking if we need to go around again.
1294 */
1298 /* We need another r10_bio. It has already been counted
1299 * in bio->bi_phys_segments.
1300 */
1310 }
1313 /* In case raid10d snuck in to freeze_array */
1318 }
1321 {
1332 else
1334 }
1342 }
1344 /* check if there are enough drives for
1345 * every block to appear on atleast one.
1346 * Don't consider the device numbered 'ignore'
1347 * as we might be about to remove it.
1348 */
1350 {
1359 cnt++;
1361 }
1366 }
1369 {
1373 /*
1374 * If it is not operational, then we have already marked it as dead
1375 * else if it is the last working disks, ignore the error, let the
1376 * next level up know.
1377 * else mark the drive as failed
1378 */
1381 /*
1382 * Don't fail the drive, just return an IO error.
1383 */
1390 /*
1391 * if recovery is running, make sure it aborts.
1392 */
1394 }
1403 }
1406 {
1414 }
1426 }
1427 }
1430 {
1436 }
1439 {
1446 /*
1447 * Find all non-in_sync disks within the RAID10 configuration
1448 * and mark them in_sync
1449 */
1456 /* Replacement has just become active */
1459 count++;
1461 /* Replaced device not technically faulty,
1462 * but we need to be sure it gets removed
1463 * and never re-added.
1464 */
1468 }
1473 count++;
1475 }
1476 }
1483 }
1487 {
1495 /* only hot-add to in-sync arrays, as recovery is
1496 * very different from resync
1497 */
1508 else
1528 }
1532 }
1536 /* as we don't honour merge_bvec_fn, we must
1537 * never risk violating it, so limit
1538 * ->max_segments to one lying with a single
1539 * page, as a one page request is never in
1540 * violation.
1541 */
1546 }
1556 }
1561 }
1564 {
1576 else
1583 }
1584 /* Only remove faulty devices if recovery
1585 * is not possible.
1586 */
1593 }
1597 /* lost the race, try later */
1602 /* We must have just cleared 'rdev' */
1606 * but will never see neither -- if they are careful.
1607 */
1611 /* We might have just remove the Replacement as faulty
1612 * Clear the flag just in case
1613 */
1618 abort:
1622 }
1626 {
1635 else
1636 /* The write handler will notice the lack of
1637 * R10BIO_Uptodate and record any errors etc
1638 */
1642 /* for reconstruct, we always reschedule after a read.
1643 * for resync, only after all reads
1644 */
1648 /* we have read all the blocks,
1649 * do the comparison in process context in raid10d
1650 */
1652 }
1653 }
1656 {
1661 /* the primary of several recovery bios */
1666 else
1675 else
1678 }
1679 }
1680 }
1683 {
1689 sector_t first_bad;
1701 }
1712 }
1722 }
1724 /*
1725 * Note: sync and recover and handled very differently for raid10
1726 * This code is for resync.
1727 * For resync, we read through virtual addresses and read all blocks.
1728 * If there is any error, we schedule a write. The lowest numbered
1729 * drive is authoritative.
1730 * However requests come for physical address, so we need to map.
1731 * For every physical address there are raid_disks/copies virtual addresses,
1732 * which is always are least one, but is not necessarly an integer.
1733 * This means that a physical address can span multiple chunks, so we may
1734 * have to submit multiple io requests for a single sync request.
1735 */
1736 /*
1737 * We check if all blocks are in-sync and only write to blocks that
1738 * aren't in sync
1739 */
1741 {
1749 /* find the first device with a block */
1761 /* now find blocks with errors */
1772 /* We know that the bi_io_vec layout is the same for
1773 * both 'first' and 'i', so we just compare them.
1774 * All vec entries are PAGE_SIZE;
1775 */
1779 PAGE_SIZE))
1785 /* Don't fix anything. */
1787 }
1788 /* Ok, we need to write this bio, either to correct an
1789 * inconsistency or to correct an unreadable block.
1790 * First we need to fixup bv_offset, bv_len and
1791 * bi_vecs, as the read request might have corrupted these
1792 */
1810 PAGE_SIZE);
1811 }
1822 }
1824 /* Now write out to any replacement devices
1825 * that are active
1826 */
1838 PAGE_SIZE);
1844 }
1846 done:
1850 }
1851 }
1853 /*
1854 * Now for the recovery code.
1855 * Recovery happens across physical sectors.
1856 * We recover all non-is_sync drives by finding the virtual address of
1857 * each, and then choose a working drive that also has that virt address.
1858 * There is a separate r10_bio for each non-in_sync drive.
1859 * Only the first two slots are in use. The first for reading,
1860 * The second for writing.
1861 *
1862 */
1864 {
1865 /* We got a read error during recovery.
1866 * We repeat the read in smaller page-sized sections.
1867 * If a read succeeds, write it to the new device or record
1868 * a bad block if we cannot.
1869 * If a read fails, record a bad block on both old and
1870 * new devices.
1871 */
1884 sector_t addr;
1893 addr,
1901 addr,
1911 }
1912 }
1914 /* We don't worry if we cannot set a bad block -
1915 * it really is bad so there is no loss in not
1916 * recording it yet
1917 */
1921 /* need bad block on destination too */
1926 /* just abort the recovery */
1928 "md/raid10:%s: recovery aborted"
1937 }
1938 }
1939 }
1943 idx++;
1944 }
1945 }
1948 {
1957 }
1959 /*
1960 * share the pages with the first bio
1961 * and submit the write request
1962 */
1970 }
1976 }
1977 }
1980 /*
1981 * Used by fix_read_error() to decay the per rdev read_errors.
1982 * We halve the read error count for every hour that has elapsed
1983 * since the last recorded read error.
1984 *
1985 */
1987 {
1996 /* first time we've seen a read error */
1999 }
2006 /*
2007 * if hours_since_last is > the number of bits in read_errors
2008 * just set read errors to 0. We do this to avoid
2009 * overflowing the shift of read_errors by hours_since_last.
2010 */
2013 else
2015 }
2019 {
2020 sector_t first_bad;
2027 /* success */
2034 }
2035 /* need to record an error - either for the block or the device */
2039 }
2041 /*
2042 * This is a kernel thread which:
2043 *
2044 * 1. Retries failed read operations on working mirrors.
2045 * 2. Updates the raid superblock when problems encounter.
2046 * 3. Performs writes following reads for array synchronising.
2047 */
2050 {
2057 /* still own a reference to this rdev, so it cannot
2058 * have been cleared recently.
2059 */
2063 /* drive has already been failed, just ignore any
2064 more fix_read_error() attempts */
2074 "md/raid10:%s: %s: Raid device exceeded "
2084 }
2097 sector_t first_bad;
2110 sect,
2117 }
2118 sl++;
2125 /* Cannot read from anywhere, just mark the block
2126 * as bad on the first device to discourage future
2127 * reads.
2128 */
2133 rdev,
2140 }
2142 }
2145 /* write it back and re-read */
2152 sl--;
2163 sect,
2166 /* Well, this device is dead */
2168 "md/raid10:%s: read correction "
2169 "write failed"
2179 }
2182 }
2189 sl--;
2200 sect,
2202 READ)) {
2204 /* Well, this device is dead */
2206 "md/raid10:%s: unable to read back "
2207 "corrected sectors"
2220 "md/raid10:%s: read error corrected"
2227 }
2231 }
2236 }
2237 }
2240 {
2242 }
2245 {
2256 }
2259 {
2264 /* bio has the data to be written to slot 'i' where
2265 * we just recently had a write error.
2266 * We repeatedly clone the bio and trim down to one block,
2267 * then try the write. Where the write fails we record
2268 * a bad block.
2269 * It is conceivable that the bio doesn't exactly align with
2270 * blocks. We must handle this.
2271 *
2272 * We currently own a reference to the rdev.
2273 */
2276 sector_t sector;
2294 /* Write at 'sector' for 'sectors' */
2302 /* Failure! */
2311 }
2313 }
2316 {
2325 /* we got a read error. Maybe the drive is bad. Maybe just
2326 * the block and we can fix it.
2327 * We freeze all other IO, and try reading the block from
2328 * other devices. When we find one, we re-write
2329 * and check it that fixes the read error.
2330 * This is all done synchronously while the array is
2331 * frozen.
2332 */
2347 read_more:
2356 }
2361 KERN_ERR
2362 "md/raid10:%s: %s: redirecting"
2371 max_sectors);
2381 /* Drat - have to split this up more */
2390 else
2396 GFP_NOIO);
2410 }
2413 {
2414 /* Some sort of write request has finished and it
2415 * succeeded in writing where we thought there was a
2416 * bad block. So forget the bad block.
2417 * Or possibly if failed and we need to record
2418 * a bad block.
2419 */
2433 rdev,
2438 rdev,
2442 }
2449 rdev,
2454 rdev,
2458 }
2459 }
2468 rdev,
2478 }
2480 }
2485 rdev,
2489 }
2490 }
2495 }
2496 }
2499 {
2517 }
2535 /* just a partial read to be scheduled from a
2536 * separate context
2537 */
2540 }
2545 }
2547 }
2551 {
2566 }
2568 /*
2569 * perform a "sync" on one "block"
2570 *
2571 * We need to make sure that no normal I/O request - particularly write
2572 * requests - conflict with active sync requests.
2573 *
2574 * This is achieved by tracking pending requests and a 'barrier' concept
2575 * that can be installed to exclude normal IO requests.
2576 *
2577 * Resync and recovery are handled very differently.
2578 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2579 *
2580 * For resync, we iterate over virtual addresses, read all copies,
2581 * and update if there are differences. If only one copy is live,
2582 * skip it.
2583 * For recovery, we iterate over physical addresses, read a good
2584 * value for each non-in_sync drive, and over-write.
2585 *
2586 * So, for recovery we may have several outstanding complex requests for a
2587 * given address, one for each out-of-sync device. We model this by allocating
2588 * a number of r10_bio structures, one for each out-of-sync device.
2589 * As we setup these structures, we collect all bio's together into a list
2590 * which we then process collectively to add pages, and then process again
2591 * to pass to generic_make_request.
2592 *
2593 * The r10_bio structures are linked using a borrowed master_bio pointer.
2594 * This link is counted in ->remaining. When the r10_bio that points to NULL
2595 * has its remaining count decremented to 0, the whole complex operation
2596 * is complete.
2597 *
2598 */
2602 {
2609 sector_t sync_blocks;
2617 skipped:
2622 /* If we aborted, we need to abort the
2623 * sync on the 'current' bitmap chucks (there can
2624 * be several when recovering multiple devices).
2625 * as we may have started syncing it but not finished.
2626 * We can find the current address in
2627 * mddev->curr_resync, but for recovery,
2628 * we need to convert that to several
2629 * virtual addresses.
2630 */
2636 sector_t sect =
2640 }
2642 /* completed sync */
2646 /* Completed a full sync so the replacements
2647 * are now fully recovered.
2648 */
2652 ->recovery_offset
2654 }
2656 }
2661 }
2663 /* if there has been nothing to do on any drive,
2664 * then there is nothing to do at all..
2665 */
2668 }
2673 /* make sure whole request will fit in a chunk - if chunks
2674 * are meaningful
2675 */
2679 /*
2680 * If there is non-resync activity waiting for us then
2681 * put in a delay to throttle resync.
2682 */
2686 /* Again, very different code for resync and recovery.
2687 * Both must result in an r10bio with a list of bios that
2688 * have bi_end_io, bi_sector, bi_bdev set,
2689 * and bi_private set to the r10bio.
2690 * For recovery, we may actually create several r10bios
2691 * with 2 bios in each, that correspond to the bios in the main one.
2692 * In this case, the subordinate r10bios link back through a
2693 * borrowed master_bio pointer, and the counter in the master
2694 * includes a ref from each subordinate.
2695 */
2696 /* First, we decide what to do and set ->bi_end_io
2697 * To end_sync_read if we want to read, and
2698 * end_sync_write if we will want to write.
2699 */
2703 /* recovery... the complicated one */
2710 sector_t sect;
2717 &&
2724 /* want to reconstruct this device */
2727 /* Unless we are doing a full sync, or a replacement
2728 * we only need to recover the block if it is set in
2729 * the bitmap
2730 */
2738 /* yep, skip the sync_blocks here, but don't assume
2739 * that there will never be anything to do here
2740 */
2743 }
2758 /* Need to check if the array will still be
2759 * degraded
2760 */
2766 }
2782 /* This is where we read from */
2792 bad_sectors -= (sector
2797 }
2798 }
2809 /* and we write to 'i' (if not in_sync) */
2836 /* and maybe write to replacement */
2841 /* Note: if rdev != NULL, then bio
2842 * cannot be NULL as r10buf_pool_alloc will
2843 * have allocated it.
2844 * So the second test here is pointless.
2845 * But it keeps semantic-checkers happy, and
2846 * this comment keeps human reviewers
2847 * happy.
2848 */
2861 }
2863 /* Cannot recover, so abort the recovery or
2864 * record a bad block */
2870 /* problem is that there are bad blocks
2871 * on other device(s)
2872 */
2890 }
2897 mirror->recovery_disabled
2899 }
2901 }
2902 }
2909 }
2911 }
2913 /* resync. Schedule a read for every block at this virt offset */
2922 /* We can skip this block */
2925 }
2966 }
2967 }
2978 count++;
2985 /* Need to set up for writing to the replacement */
2999 count++;
3000 }
3007 mddev);
3012 mddev);
3013 }
3017 }
3018 }
3029 }
3047 /* stop here */
3052 /* remove last page from this bio */
3056 }
3058 }
3062 bio_full:
3076 }
3077 }
3080 /* pretend they weren't skipped, it makes
3081 * no important difference in this case
3082 */
3086 giveup:
3087 /* There is nowhere to write, so all non-sync
3088 * drives must be failed or in resync, all drives
3089 * have a bad block, so try the next chunk...
3090 */
3095 chunks_skipped ++;
3098 }
3100 static sector_t
3102 {
3103 sector_t size;
3117 }
3121 {
3133 }
3144 }
3152 GFP_KERNEL);
3178 /* 'size' is now the number of chunks in the array */
3179 /* calculate "used chunks per device" in 'stride' */
3182 /* We need to round up when dividing by raid_disks to
3183 * get the stride size.
3184 */
3192 else
3210 out:
3219 }
3221 }
3224 {
3229 sector_t size;
3231 /*
3232 * copy the already verified devices into our private RAID10
3233 * bookkeeping area. [whatever we allocate in run(),
3234 * should be freed in stop()]
3235 */
3242 }
3254 else
3274 }
3278 /* as we don't honour merge_bvec_fn, we must never risk
3279 * violating it, so limit max_segments to 1 lying
3280 * within a single page.
3281 */
3286 }
3289 }
3290 /* need to check that every block has at least one working mirror */
3295 }
3303 /* The replacement is all we have - use it */
3307 }
3315 }
3317 }
3327 /*
3328 * Ok, everything is just fine now
3329 */
3338 /* Calculate max read-ahead size.
3339 * We need to readahead at least twice a whole stripe....
3340 * maybe...
3341 */
3342 {
3348 }
3358 out_free_conf:
3366 out:
3368 }
3371 {
3385 }
3388 {
3398 }
3399 }
3402 {
3410 }
3412 /* Set new parameters */
3414 /* new layout: far_copies = 1, near_copies = 2 */
3419 /* make sure it will be not marked as dirty */
3428 }
3431 }
3434 {
3437 /* raid10 can take over:
3438 * raid0 - providing it has only two drives
3439 */
3441 /* for raid0 takeover only one zone is supported */
3448 }
3450 }
3452 }
3455 {
3471 };
3474 {
3476 }
3479 {
3481 }