| blob | d1295aff41739eea048f0e43513dfba6ade4c8f1 |
1 /*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
33 /*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 *
42 * The data to be stored is divided into chunks using chunksize.
43 * Each device is divided into far_copies sections.
44 * In each section, chunks are laid out in a style similar to raid0, but
45 * near_copies copies of each chunk is stored (each on a different drive).
46 * The starting device for each section is offset near_copies from the starting
47 * device of the previous section.
48 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
49 * drive.
50 * near_copies and far_copies must be at least one, and their product is at most
51 * raid_disks.
52 *
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of be very far apart
55 * on disk, there are adjacent stripes.
56 */
58 /*
59 * Number of guaranteed r10bios in case of extreme VM load:
60 */
61 #define NR_RAID10_BIOS 256
63 /* when we get a read error on a read-only array, we redirect to another
64 * device without failing the first device, or trying to over-write to
65 * correct the read error. To keep track of bad blocks on a per-bio
66 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
67 */
68 #define IO_BLOCKED ((struct bio *)1)
69 /* When we successfully write to a known bad-block, we need to remove the
70 * bad-block marking which must be done from process context. So we record
71 * the success by setting devs[n].bio to IO_MADE_GOOD
72 */
73 #define IO_MADE_GOOD ((struct bio *)2)
75 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
77 /* When there are this many requests queued to be written by
78 * the raid10 thread, we become 'congested' to provide back-pressure
79 * for writeback.
80 */
93 {
97 /* allocate a r10bio with room for raid_disks entries in the
98 * bios array */
100 }
103 {
105 }
107 /* Maximum size of each resync request */
108 #define RESYNC_BLOCK_SIZE (64*1024)
109 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
110 /* amount of memory to reserve for resync requests */
111 #define RESYNC_WINDOW (1024*1024)
112 /* maximum number of concurrent requests, memory permitting */
113 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
115 /*
116 * When performing a resync, we need to read and compare, so
117 * we need as many pages are there are copies.
118 * When performing a recovery, we need 2 bios, one for read,
119 * one for write (we recover only one drive per r10buf)
120 *
121 */
123 {
138 else
141 /*
142 * Allocate bios.
143 */
155 }
156 /*
157 * Allocate RESYNC_PAGES data pages and attach them
158 * where needed.
159 */
166 /* we can share bv_page's during recovery
167 * and reshape */
179 }
180 }
184 out_free_pages:
191 out_free_bio:
197 }
200 }
203 {
215 }
217 }
221 }
223 }
226 {
238 }
239 }
242 {
247 }
250 {
256 }
259 {
269 /* wake up frozen array... */
273 }
275 /*
276 * raid_end_bio_io() is called when we have finished servicing a mirrored
277 * operation and are ready to return a success/failure code to the buffer
278 * cache layer.
279 */
281 {
298 /*
299 * Wake up any possible resync thread that waits for the device
300 * to go idle.
301 */
303 }
305 }
307 /*
308 * Update disk head position estimator based on IRQ completion info.
309 */
311 {
316 }
318 /*
319 * Find the disk number which triggered given bio
320 */
323 {
333 }
334 }
344 }
347 {
358 /*
359 * this branch is our 'one mirror IO has finished' event handler:
360 */
364 /*
365 * Set R10BIO_Uptodate in our master bio, so that
366 * we will return a good error code to the higher
367 * levels even if IO on some other mirrored buffer fails.
368 *
369 * The 'master' represents the composite IO operation to
370 * user-side. So if something waits for IO, then it will
371 * wait for the 'master' bio.
372 */
375 /* If all other devices that store this block have
376 * failed, we want to return the error upwards rather
377 * than fail the last device. Here we redefine
378 * "uptodate" to mean "Don't want to retry"
379 */
385 }
390 /*
391 * oops, read error - keep the refcount on the rdev
392 */
401 }
402 }
405 {
406 /* clear the bitmap if all writes complete successfully */
412 }
415 {
423 else
425 }
426 }
427 }
430 {
447 }
448 /*
449 * this branch is our 'one mirror IO has finished' event handler:
450 */
453 /* Never record new bad blocks to replacement,
454 * just fail it.
455 */
464 }
466 /*
467 * Set R10BIO_Uptodate in our master bio, so that
468 * we will return a good error code for to the higher
469 * levels even if IO on some other mirrored buffer fails.
470 *
471 * The 'master' represents the composite IO operation to
472 * user-side. So if something waits for IO, then it will
473 * wait for the 'master' bio.
474 */
475 sector_t first_bad;
480 /* Maybe we can clear some bad blocks. */
488 else
492 }
493 }
495 /*
496 *
497 * Let's see if all mirrored write operations have finished
498 * already.
499 */
503 }
505 /*
506 * RAID10 layout manager
507 * As well as the chunksize and raid_disks count, there are two
508 * parameters: near_copies and far_copies.
509 * near_copies * far_copies must be <= raid_disks.
510 * Normally one of these will be 1.
511 * If both are 1, we get raid0.
512 * If near_copies == raid_disks, we get raid1.
513 *
514 * Chunks are laid out in raid0 style with near_copies copies of the
515 * first chunk, followed by near_copies copies of the next chunk and
516 * so on.
517 * If far_copies > 1, then after 1/far_copies of the array has been assigned
518 * as described above, we start again with a device offset of near_copies.
519 * So we effectively have another copy of the whole array further down all
520 * the drives, but with blocks on different drives.
521 * With this layout, and block is never stored twice on the one device.
522 *
523 * raid10_find_phys finds the sector offset of a given virtual sector
524 * on each device that it is on.
525 *
526 * raid10_find_virt does the reverse mapping, from a device and a
527 * sector offset to a virtual address
528 */
531 {
533 sector_t sector;
534 sector_t chunk;
535 sector_t stripe;
539 /* now calculate first sector/dev */
551 /* and calculate all the others */
557 slot++;
566 slot++;
567 }
568 dev++;
572 }
573 }
574 }
577 {
589 }
592 {
594 /* Never use conf->prev as this is only called during resync
595 * or recovery, so reshape isn't happening
596 */
612 else
614 }
616 }
620 }
622 /**
623 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
624 * @q: request queue
625 * @bvm: properties of new bio
626 * @biovec: the request that could be merged to it.
627 *
628 * Return amount of bytes we can accept at this offset
629 * This requires checking for end-of-chunk if near_copies != raid_disks,
630 * and for subordinate merge_bvec_fns if merge_check_needed.
631 */
635 {
654 /* bio_add cannot handle a negative return */
669 /* Cannot give any guidance during reshape */
673 }
690 }
691 }
702 }
703 }
704 }
706 }
708 }
710 /*
711 * This routine returns the disk from which the requested read should
712 * be done. There is a per-array 'next expected sequential IO' sector
713 * number - if this matches on the next IO then we use the last disk.
714 * There is also a per-disk 'last know head position' sector that is
715 * maintained from IRQ contexts, both the normal and the resync IO
716 * completion handlers update this position correctly. If there is no
717 * perfect sequential match then we pick the disk whose head is closest.
718 *
719 * If there are 2 mirrors in the same 2 devices, performance degrades
720 * because position is mirror, not device based.
721 *
722 * The rdev for the device selected will have nr_pending incremented.
723 */
725 /*
726 * FIXME: possibly should rethink readbalancing and do it differently
727 * depending on near_copies / far_copies geometry.
728 */
732 {
745 retry:
752 /*
753 * Check if we can balance. We can balance on the whole
754 * device if no resync is going on (recovery is ok), or below
755 * the resync window. We take the first readable disk when
756 * above the resync window.
757 */
763 sector_t first_bad;
765 sector_t dev_sector;
787 /* Already have a better slot */
790 /* Cannot read here. If this is the
791 * 'primary' device, then we must not read
792 * beyond 'bad_sectors' from another device.
793 */
800 sector_t good_sectors =
806 }
808 /* Must read from here */
810 }
818 /* This optimisation is debatable, and completely destroys
819 * sequential read speed for 'far copies' arrays. So only
820 * keep it for 'near' arrays, and review those later.
821 */
825 /* for far > 1 always use the lowest address */
828 else
835 }
836 }
840 }
845 /* Cannot risk returning a device that failed
846 * before we inc'ed nr_pending
847 */
850 }
858 }
861 {
873 i++) {
879 }
880 }
883 }
887 {
892 }
895 {
896 /* Any writes that have been queued but are awaiting
897 * bitmap updates get flushed here.
898 */
906 /* flush any pending bitmap writes to disk
907 * before proceeding w/ I/O */
916 /* Just ignore it */
918 else
921 }
924 }
926 /* Barriers....
927 * Sometimes we need to suspend IO while we do something else,
928 * either some resync/recovery, or reconfigure the array.
929 * To do this we raise a 'barrier'.
930 * The 'barrier' is a counter that can be raised multiple times
931 * to count how many activities are happening which preclude
932 * normal IO.
933 * We can only raise the barrier if there is no pending IO.
934 * i.e. if nr_pending == 0.
935 * We choose only to raise the barrier if no-one is waiting for the
936 * barrier to go down. This means that as soon as an IO request
937 * is ready, no other operations which require a barrier will start
938 * until the IO request has had a chance.
939 *
940 * So: regular IO calls 'wait_barrier'. When that returns there
941 * is no backgroup IO happening, It must arrange to call
942 * allow_barrier when it has finished its IO.
943 * backgroup IO calls must call raise_barrier. Once that returns
944 * there is no normal IO happeing. It must arrange to call
945 * lower_barrier when the particular background IO completes.
946 */
949 {
953 /* Wait until no block IO is waiting (unless 'force') */
957 /* block any new IO from starting */
960 /* Now wait for all pending IO to complete */
966 }
969 {
975 }
978 {
982 /* Wait for the barrier to drop.
983 * However if there are already pending
984 * requests (preventing the barrier from
985 * rising completely), and the
986 * pre-process bio queue isn't empty,
987 * then don't wait, as we need to empty
988 * that queue to get the nr_pending
989 * count down.
990 */
997 );
999 }
1002 }
1005 {
1011 }
1014 {
1015 /* stop syncio and normal IO and wait for everything to
1016 * go quiet.
1017 * We increment barrier and nr_waiting, and then
1018 * wait until nr_pending match nr_queued+1
1019 * This is called in the context of one normal IO request
1020 * that has failed. Thus any sync request that might be pending
1021 * will be blocked by nr_pending, and we need to wait for
1022 * pending IO requests to complete or be queued for re-try.
1023 * Thus the number queued (nr_queued) plus this request (1)
1024 * must match the number of pending IOs (nr_pending) before
1025 * we continue.
1026 */
1036 }
1039 {
1040 /* reverse the effect of the freeze */
1046 }
1050 {
1054 else
1056 }
1062 };
1065 {
1067 cb);
1080 }
1082 /* we aren't scheduling, so we can do the write-out directly. */
1092 }
1094 }
1097 {
1120 }
1122 /* If this request crosses a chunk boundary, we need to
1123 * split it. This will only happen for 1 PAGE (or less) requests.
1124 */
1126 > chunk_sects
1130 /* Sanity check -- queue functions should prevent this happening */
1134 /* This is a one page bio that upper layers
1135 * refuse to split for us, so we need to split it.
1136 */
1140 /* Each of these 'make_request' calls will call 'wait_barrier'.
1141 * If the first succeeds but the second blocks due to the resync
1142 * thread raising the barrier, we will deadlock because the
1143 * IO to the underlying device will be queued in generic_make_request
1144 * and will never complete, so will never reduce nr_pending.
1145 * So increment nr_waiting here so no new raise_barriers will
1146 * succeed, and so the second wait_barrier cannot block.
1147 */
1162 bad_map:
1169 }
1173 /*
1174 * Register the new request and wait if the reconstruction
1175 * thread has put up a bar for new requests.
1176 * Continue immediately if no resync is active currently.
1177 */
1184 /* IO spans the reshape position. Need to wait for
1185 * reshape to pass
1186 */
1192 }
1200 /* Need to update reshape_position in metadata */
1209 }
1220 /* We might need to issue multiple reads to different
1221 * devices if there are bad blocks around, so we keep
1222 * track of the number of reads in bio->bi_phys_segments.
1223 * If this is 0, there is only one r10_bio and no locking
1224 * will be needed when the request completes. If it is
1225 * non-zero, then it is the number of not-completed requests.
1226 */
1231 /*
1232 * read balancing logic:
1233 */
1237 read_again:
1242 }
1247 max_sectors);
1260 /* Could not read all from this device, so we will
1261 * need another r10_bio.
1262 */
1269 else
1272 /* Cannot call generic_make_request directly
1273 * as that will be queued in __generic_make_request
1274 * and subsequent mempool_alloc might block
1275 * waiting for it. so hand bio over to raid10d.
1276 */
1291 }
1293 /*
1294 * WRITE:
1295 */
1300 }
1301 /* first select target devices under rcu_lock and
1302 * inc refcount on their rdev. Record them by setting
1303 * bios[x] to bio
1304 * If there are known/acknowledged bad blocks on any device
1305 * on which we have seen a write error, we want to avoid
1306 * writing to those blocks. This potentially requires several
1307 * writes to write around the bad blocks. Each set of writes
1308 * gets its own r10_bio with a set of bios attached. The number
1309 * of r10_bios is recored in bio->bi_phys_segments just as with
1310 * the read case.
1311 */
1315 retry_write:
1331 }
1336 }
1347 }
1349 sector_t first_bad;
1355 max_sectors,
1358 /* Mustn't write here until the bad block
1359 * is acknowledged
1360 */
1365 }
1367 /* Cannot write here at all */
1370 /* Mustn't write more than bad_sectors
1371 * to other devices yet
1372 */
1374 /* We don't set R10BIO_Degraded as that
1375 * only applies if the disk is missing,
1376 * so it might be re-added, and we want to
1377 * know to recover this chunk.
1378 * In this case the device is here, and the
1379 * fact that this chunk is not in-sync is
1380 * recorded in the bad block log.
1381 */
1383 }
1388 }
1389 }
1395 }
1396 }
1400 /* Have to wait for this device to get unblocked, then retry */
1408 }
1414 /* Race with remove_disk */
1417 }
1419 }
1420 }
1425 }
1428 /* We are splitting this into multiple parts, so
1429 * we need to prepare for allocating another r10_bio.
1430 */
1435 else
1438 }
1452 max_sectors);
1468 else
1477 }
1487 max_sectors);
1490 /* We are actively writing to the original device
1491 * so it cannot disappear, so the replacement cannot
1492 * become NULL here
1493 */
1496 r10_bio,
1510 }
1512 /* Don't remove the bias on 'remaining' (one_write_done) until
1513 * after checking if we need to go around again.
1514 */
1518 /* We need another r10_bio. It has already been counted
1519 * in bio->bi_phys_segments.
1520 */
1530 }
1533 /* In case raid10d snuck in to freeze_array */
1535 }
1538 {
1549 else
1551 }
1559 }
1561 /* check if there are enough drives for
1562 * every block to appear on atleast one.
1563 * Don't consider the device numbered 'ignore'
1564 * as we might be about to remove it.
1565 */
1567 {
1577 cnt++;
1579 }
1585 }
1588 {
1591 }
1594 {
1598 /*
1599 * If it is not operational, then we have already marked it as dead
1600 * else if it is the last working disks, ignore the error, let the
1601 * next level up know.
1602 * else mark the drive as failed
1603 */
1606 /*
1607 * Don't fail the drive, just return an IO error.
1608 */
1615 /*
1616 * if recovery is running, make sure it aborts.
1617 */
1619 }
1628 }
1631 {
1639 }
1651 }
1652 }
1655 {
1661 }
1664 {
1671 /*
1672 * Find all non-in_sync disks within the RAID10 configuration
1673 * and mark them in_sync
1674 */
1681 /* Replacement has just become active */
1684 count++;
1686 /* Replaced device not technically faulty,
1687 * but we need to be sure it gets removed
1688 * and never re-added.
1689 */
1693 }
1698 count++;
1700 }
1701 }
1708 }
1712 {
1721 /* only hot-add to in-sync arrays, as recovery is
1722 * very different from resync
1723 */
1734 }
1739 else
1758 }
1771 }
1773 /* Some requests might not have seen this new
1774 * merge_bvec_fn. We must wait for them to complete
1775 * before merging the device fully.
1776 * First we make sure any code which has tested
1777 * our function has submitted the request, then
1778 * we wait for all outstanding requests to complete.
1779 */
1784 }
1791 }
1794 {
1806 else
1813 }
1814 /* Only remove faulty devices if recovery
1815 * is not possible.
1816 */
1824 }
1828 /* lost the race, try later */
1833 /* We must have just cleared 'rdev' */
1837 * but will never see neither -- if they are careful.
1838 */
1842 /* We might have just remove the Replacement as faulty
1843 * Clear the flag just in case
1844 */
1849 abort:
1853 }
1857 {
1863 /* this is a reshape read */
1870 else
1871 /* The write handler will notice the lack of
1872 * R10BIO_Uptodate and record any errors etc
1873 */
1877 /* for reconstruct, we always reschedule after a read.
1878 * for resync, only after all reads
1879 */
1883 /* we have read all the blocks,
1884 * do the comparison in process context in raid10d
1885 */
1887 }
1888 }
1891 {
1896 /* the primary of several recovery bios */
1901 else
1910 else
1913 }
1914 }
1915 }
1918 {
1924 sector_t first_bad;
1933 else
1945 }
1955 }
1957 /*
1958 * Note: sync and recover and handled very differently for raid10
1959 * This code is for resync.
1960 * For resync, we read through virtual addresses and read all blocks.
1961 * If there is any error, we schedule a write. The lowest numbered
1962 * drive is authoritative.
1963 * However requests come for physical address, so we need to map.
1964 * For every physical address there are raid_disks/copies virtual addresses,
1965 * which is always are least one, but is not necessarly an integer.
1966 * This means that a physical address can span multiple chunks, so we may
1967 * have to submit multiple io requests for a single sync request.
1968 */
1969 /*
1970 * We check if all blocks are in-sync and only write to blocks that
1971 * aren't in sync
1972 */
1974 {
1982 /* find the first device with a block */
1994 /* now find blocks with errors */
2005 /* We know that the bi_io_vec layout is the same for
2006 * both 'first' and 'i', so we just compare them.
2007 * All vec entries are PAGE_SIZE;
2008 */
2018 /* Don't fix anything. */
2020 }
2021 /* Ok, we need to write this bio, either to correct an
2022 * inconsistency or to correct an unreadable block.
2023 * First we need to fixup bv_offset, bv_len and
2024 * bi_vecs, as the read request might have corrupted these
2025 */
2043 PAGE_SIZE);
2044 }
2055 }
2057 /* Now write out to any replacement devices
2058 * that are active
2059 */
2071 PAGE_SIZE);
2077 }
2079 done:
2083 }
2084 }
2086 /*
2087 * Now for the recovery code.
2088 * Recovery happens across physical sectors.
2089 * We recover all non-is_sync drives by finding the virtual address of
2090 * each, and then choose a working drive that also has that virt address.
2091 * There is a separate r10_bio for each non-in_sync drive.
2092 * Only the first two slots are in use. The first for reading,
2093 * The second for writing.
2094 *
2095 */
2097 {
2098 /* We got a read error during recovery.
2099 * We repeat the read in smaller page-sized sections.
2100 * If a read succeeds, write it to the new device or record
2101 * a bad block if we cannot.
2102 * If a read fails, record a bad block on both old and
2103 * new devices.
2104 */
2117 sector_t addr;
2126 addr,
2134 addr,
2144 }
2145 }
2147 /* We don't worry if we cannot set a bad block -
2148 * it really is bad so there is no loss in not
2149 * recording it yet
2150 */
2154 /* need bad block on destination too */
2159 /* just abort the recovery */
2161 "md/raid10:%s: recovery aborted"
2170 }
2171 }
2172 }
2176 idx++;
2177 }
2178 }
2181 {
2190 }
2192 /*
2193 * share the pages with the first bio
2194 * and submit the write request
2195 */
2203 }
2209 }
2210 }
2213 /*
2214 * Used by fix_read_error() to decay the per rdev read_errors.
2215 * We halve the read error count for every hour that has elapsed
2216 * since the last recorded read error.
2217 *
2218 */
2220 {
2229 /* first time we've seen a read error */
2232 }
2239 /*
2240 * if hours_since_last is > the number of bits in read_errors
2241 * just set read errors to 0. We do this to avoid
2242 * overflowing the shift of read_errors by hours_since_last.
2243 */
2246 else
2248 }
2252 {
2253 sector_t first_bad;
2260 /* success */
2267 }
2268 /* need to record an error - either for the block or the device */
2272 }
2274 /*
2275 * This is a kernel thread which:
2276 *
2277 * 1. Retries failed read operations on working mirrors.
2278 * 2. Updates the raid superblock when problems encounter.
2279 * 3. Performs writes following reads for array synchronising.
2280 */
2283 {
2290 /* still own a reference to this rdev, so it cannot
2291 * have been cleared recently.
2292 */
2296 /* drive has already been failed, just ignore any
2297 more fix_read_error() attempts */
2307 "md/raid10:%s: %s: Raid device exceeded "
2317 }
2330 sector_t first_bad;
2344 sect,
2351 }
2352 sl++;
2359 /* Cannot read from anywhere, just mark the block
2360 * as bad on the first device to discourage future
2361 * reads.
2362 */
2367 rdev,
2374 }
2376 }
2379 /* write it back and re-read */
2386 sl--;
2398 sect,
2401 /* Well, this device is dead */
2403 "md/raid10:%s: read correction "
2404 "write failed"
2408 sect +
2410 rdev)),
2416 }
2419 }
2426 sl--;
2437 sect,
2439 READ)) {
2441 /* Well, this device is dead */
2443 "md/raid10:%s: unable to read back "
2444 "corrected sectors"
2448 sect +
2458 "md/raid10:%s: read error corrected"
2462 sect +
2466 }
2470 }
2475 }
2476 }
2479 {
2481 }
2484 {
2495 }
2498 {
2503 /* bio has the data to be written to slot 'i' where
2504 * we just recently had a write error.
2505 * We repeatedly clone the bio and trim down to one block,
2506 * then try the write. Where the write fails we record
2507 * a bad block.
2508 * It is conceivable that the bio doesn't exactly align with
2509 * blocks. We must handle this.
2510 *
2511 * We currently own a reference to the rdev.
2512 */
2515 sector_t sector;
2533 /* Write at 'sector' for 'sectors' */
2541 /* Failure! */
2550 }
2552 }
2555 {
2564 /* we got a read error. Maybe the drive is bad. Maybe just
2565 * the block and we can fix it.
2566 * We freeze all other IO, and try reading the block from
2567 * other devices. When we find one, we re-write
2568 * and check it that fixes the read error.
2569 * This is all done synchronously while the array is
2570 * frozen.
2571 */
2586 read_more:
2595 }
2600 KERN_ERR
2601 "md/raid10:%s: %s: redirecting "
2610 max_sectors);
2620 /* Drat - have to split this up more */
2629 else
2635 GFP_NOIO);
2649 }
2652 {
2653 /* Some sort of write request has finished and it
2654 * succeeded in writing where we thought there was a
2655 * bad block. So forget the bad block.
2656 * Or possibly if failed and we need to record
2657 * a bad block.
2658 */
2672 rdev,
2677 rdev,
2681 }
2688 rdev,
2693 rdev,
2697 }
2698 }
2707 rdev,
2717 }
2719 }
2724 rdev,
2728 }
2729 }
2734 }
2735 }
2738 {
2757 }
2777 /* just a partial read to be scheduled from a
2778 * separate context
2779 */
2782 }
2787 }
2789 }
2793 {
2808 }
2810 /*
2811 * perform a "sync" on one "block"
2812 *
2813 * We need to make sure that no normal I/O request - particularly write
2814 * requests - conflict with active sync requests.
2815 *
2816 * This is achieved by tracking pending requests and a 'barrier' concept
2817 * that can be installed to exclude normal IO requests.
2818 *
2819 * Resync and recovery are handled very differently.
2820 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2821 *
2822 * For resync, we iterate over virtual addresses, read all copies,
2823 * and update if there are differences. If only one copy is live,
2824 * skip it.
2825 * For recovery, we iterate over physical addresses, read a good
2826 * value for each non-in_sync drive, and over-write.
2827 *
2828 * So, for recovery we may have several outstanding complex requests for a
2829 * given address, one for each out-of-sync device. We model this by allocating
2830 * a number of r10_bio structures, one for each out-of-sync device.
2831 * As we setup these structures, we collect all bio's together into a list
2832 * which we then process collectively to add pages, and then process again
2833 * to pass to generic_make_request.
2834 *
2835 * The r10_bio structures are linked using a borrowed master_bio pointer.
2836 * This link is counted in ->remaining. When the r10_bio that points to NULL
2837 * has its remaining count decremented to 0, the whole complex operation
2838 * is complete.
2839 *
2840 */
2844 {
2851 sector_t sync_blocks;
2860 skipped:
2866 /* If we aborted, we need to abort the
2867 * sync on the 'current' bitmap chucks (there can
2868 * be several when recovering multiple devices).
2869 * as we may have started syncing it but not finished.
2870 * We can find the current address in
2871 * mddev->curr_resync, but for recovery,
2872 * we need to convert that to several
2873 * virtual addresses.
2874 */
2878 }
2885 sector_t sect =
2889 }
2891 /* completed sync */
2895 /* Completed a full sync so the replacements
2896 * are now fully recovered.
2897 */
2901 ->recovery_offset
2903 }
2905 }
2910 }
2916 /* if there has been nothing to do on any drive,
2917 * then there is nothing to do at all..
2918 */
2921 }
2926 /* make sure whole request will fit in a chunk - if chunks
2927 * are meaningful
2928 */
2932 /*
2933 * If there is non-resync activity waiting for us then
2934 * put in a delay to throttle resync.
2935 */
2939 /* Again, very different code for resync and recovery.
2940 * Both must result in an r10bio with a list of bios that
2941 * have bi_end_io, bi_sector, bi_bdev set,
2942 * and bi_private set to the r10bio.
2943 * For recovery, we may actually create several r10bios
2944 * with 2 bios in each, that correspond to the bios in the main one.
2945 * In this case, the subordinate r10bios link back through a
2946 * borrowed master_bio pointer, and the counter in the master
2947 * includes a ref from each subordinate.
2948 */
2949 /* First, we decide what to do and set ->bi_end_io
2950 * To end_sync_read if we want to read, and
2951 * end_sync_write if we will want to write.
2952 */
2956 /* recovery... the complicated one */
2963 sector_t sect;
2970 &&
2977 /* want to reconstruct this device */
2981 /* last stripe is not complete - don't
2982 * try to recover this sector.
2983 */
2985 }
2986 /* Unless we are doing a full sync, or a replacement
2987 * we only need to recover the block if it is set in
2988 * the bitmap
2989 */
2997 /* yep, skip the sync_blocks here, but don't assume
2998 * that there will never be anything to do here
2999 */
3002 }
3017 /* Need to check if the array will still be
3018 * degraded
3019 */
3025 }
3041 /* This is where we read from */
3051 bad_sectors -= (sector
3056 }
3057 }
3068 /* and we write to 'i' (if not in_sync) */
3095 /* and maybe write to replacement */
3100 /* Note: if rdev != NULL, then bio
3101 * cannot be NULL as r10buf_pool_alloc will
3102 * have allocated it.
3103 * So the second test here is pointless.
3104 * But it keeps semantic-checkers happy, and
3105 * this comment keeps human reviewers
3106 * happy.
3107 */
3120 }
3122 /* Cannot recover, so abort the recovery or
3123 * record a bad block */
3129 /* problem is that there are bad blocks
3130 * on other device(s)
3131 */
3149 }
3156 mirror->recovery_disabled
3158 }
3160 }
3161 }
3168 }
3170 }
3172 /* resync. Schedule a read for every block at this virt offset */
3181 /* We can skip this block */
3184 }
3225 }
3226 }
3237 count++;
3244 /* Need to set up for writing to the replacement */
3258 count++;
3259 }
3266 mddev);
3271 mddev);
3272 }
3276 }
3277 }
3288 }
3306 /* stop here */
3311 /* remove last page from this bio */
3315 }
3317 }
3321 bio_full:
3335 }
3336 }
3339 /* pretend they weren't skipped, it makes
3340 * no important difference in this case
3341 */
3345 giveup:
3346 /* There is nowhere to write, so all non-sync
3347 * drives must be failed or in resync, all drives
3348 * have a bad block, so try the next chunk...
3349 */
3354 chunks_skipped ++;
3357 }
3359 static sector_t
3361 {
3362 sector_t size;
3377 }
3380 {
3381 /* Calculate the number of sectors-per-device that will
3382 * actually be used, and set conf->dev_sectors and
3383 * conf->stride
3384 */
3390 /* 'size' is now the number of chunks in the array */
3391 /* calculate "used chunks per device" */
3394 /* We need to round up when dividing by raid_disks to
3395 * get the stride size.
3396 */
3406 }
3407 }
3411 {
3427 * updated yet. */
3432 }
3448 }
3451 {
3464 }
3470 }
3477 /* FIXME calc properly */
3480 GFP_KERNEL);
3503 }
3507 else
3508 /* far_copies must be 1 */
3510 }
3524 out:
3534 }
3536 }
3539 {
3544 sector_t size;
3554 }
3569 else
3572 }
3594 }
3614 }
3620 else
3623 }
3624 /* need to check that every block has at least one working mirror */
3629 }
3632 /* must ensure that shape change is supported */
3639 }
3645 i++) {
3650 /* The replacement is all we have - use it */
3654 }
3662 }
3664 }
3674 /*
3675 * Ok, everything is just fine now
3676 */
3688 /* Calculate max read-ahead size.
3689 * We need to readahead at least twice a whole stripe....
3690 * maybe...
3691 */
3696 }
3711 /* This cannot work */
3714 }
3724 }
3728 out_free_conf:
3736 out:
3738 }
3741 {
3749 /* the unplug fn references 'conf'*/
3758 }
3761 {
3771 }
3772 }
3775 {
3776 /* Resize of 'far' arrays is not supported.
3777 * For 'near' and 'offset' arrays we can set the
3778 * number of sectors used to be an appropriate multiple
3779 * of the chunk size.
3780 * For 'offset', this is far_copies*chunksize.
3781 * For 'near' the multiplier is the LCM of
3782 * near_copies and raid_disks.
3783 * So if far_copies > 1 && !far_offset, fail.
3784 * Else find LCM(raid_disks, near_copy)*far_copies and
3785 * multiply by chunk_size. Then round to this number.
3786 * This is mostly done by raid10_size()
3787 */
3806 }
3814 }
3819 }
3822 {
3830 }
3832 /* Set new parameters */
3834 /* new layout: far_copies = 1, near_copies = 2 */
3839 /* make sure it will be not marked as dirty */
3848 }
3851 }
3854 {
3857 /* raid10 can take over:
3858 * raid0 - providing it has only two drives
3859 */
3861 /* for raid0 takeover only one zone is supported */
3868 }
3870 }
3872 }
3875 {
3876 /* Called when there is a request to change
3877 * - layout (to ->new_layout)
3878 * - chunk size (to ->new_chunk_sectors)
3879 * - raid_disks (by delta_disks)
3880 * or when trying to restart a reshape that was ongoing.
3881 *
3882 * We need to validate the request and possibly allocate
3883 * space if that might be an issue later.
3884 *
3885 * Currently we reject any reshape of a 'far' mode array,
3886 * allow chunk size to change if new is generally acceptable,
3887 * allow raid_disks to increase, and allow
3888 * a switch between 'near' mode and 'offset' mode.
3889 */
3897 /* mustn't change number of copies */
3900 /* Cannot switch to 'far' mode */
3904 /* not factor of array size */
3913 /* allocate new 'mirrors' list */
3918 GFP_KERNEL);
3921 }
3923 }
3925 /*
3926 * Need to check if array has failed when deciding whether to:
3927 * - start an array
3928 * - remove non-faulty devices
3929 * - add a spare
3930 * - allow a reshape
3931 * This determination is simple when no reshape is happening.
3932 * However if there is a reshape, we need to carefully check
3933 * both the before and after sections.
3934 * This is because some failed devices may only affect one
3935 * of the two sections, and some non-in_sync devices may
3936 * be insync in the section most affected by failed devices.
3937 */
3939 {
3945 /* 'prev' section first */
3949 degraded++;
3951 /* When we can reduce the number of devices in
3952 * an array, this might not contribute to
3953 * 'degraded'. It does now.
3954 */
3955 degraded++;
3956 }
3965 degraded2++;
3967 /* If reshape is increasing the number of devices,
3968 * this section has already been recovered, so
3969 * it doesn't contribute to degraded.
3970 * else it does.
3971 */
3973 degraded2++;
3974 }
3975 }
3980 }
3983 {
3984 /* A 'reshape' has been requested. This commits
3985 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3986 * This also checks if there are enough spares and adds them
3987 * to the array.
3988 * We currently require enough spares to make the final
3989 * array non-degraded. We also require that the difference
3990 * between old and new data_offset - on each device - is
3991 * enough that we never risk over-writing.
3992 */
4017 spares++;
4027 }
4028 }
4046 }
4056 }
4070 }
4079 else
4084 }
4087 /* This is a spare that was manually added */
4089 }
4090 }
4091 /* When a reshape changes the number of devices,
4092 * ->degraded is measured against the larger of the
4093 * pre and post numbers.
4094 */
4112 }
4118 abort:
4130 }
4132 /* Calculate the last device-address that could contain
4133 * any block from the chunk that includes the array-address 's'
4134 * and report the next address.
4135 * i.e. the address returned will be chunk-aligned and after
4136 * any data that is in the chunk containing 's'.
4137 */
4139 {
4147 }
4149 /* Calculate the first device-address that could contain
4150 * any block from the chunk that includes the array-address 's'.
4151 * This too will be the start of a chunk
4152 */
4154 {
4161 }
4165 {
4166 /* We simply copy at most one chunk (smallest of old and new)
4167 * at a time, possibly less if that exceeds RESYNC_PAGES,
4168 * or we hit a bad block or something.
4169 * This might mean we pause for normal IO in the middle of
4170 * a chunk, but that is not a problem was mddev->reshape_position
4171 * can record any location.
4172 *
4173 * If we will want to write to a location that isn't
4174 * yet recorded as 'safe' (i.e. in metadata on disk) then
4175 * we need to flush all reshape requests and update the metadata.
4176 *
4177 * When reshaping forwards (e.g. to more devices), we interpret
4178 * 'safe' as the earliest block which might not have been copied
4179 * down yet. We divide this by previous stripe size and multiply
4180 * by previous stripe length to get lowest device offset that we
4181 * cannot write to yet.
4182 * We interpret 'sector_nr' as an address that we want to write to.
4183 * From this we use last_device_address() to find where we might
4184 * write to, and first_device_address on the 'safe' position.
4185 * If this 'next' write position is after the 'safe' position,
4186 * we must update the metadata to increase the 'safe' position.
4187 *
4188 * When reshaping backwards, we round in the opposite direction
4189 * and perform the reverse test: next write position must not be
4190 * less than current safe position.
4191 *
4192 * In all this the minimum difference in data offsets
4193 * (conf->offset_diff - always positive) allows a bit of slack,
4194 * so next can be after 'safe', but not by more than offset_disk
4195 *
4196 * We need to prepare all the bios here before we start any IO
4197 * to ensure the size we choose is acceptable to all devices.
4198 * The means one for each copy for write-out and an extra one for
4199 * read-in.
4200 * We store the read-in bio in ->master_bio and the others in
4201 * ->devs[x].bio and ->devs[x].repl_bio.
4202 */
4216 /* If restarting in the middle, skip the initial sectors */
4229 }
4230 }
4232 /* We don't use sector_nr to track where we are up to
4233 * as that doesn't work well for ->reshape_backwards.
4234 * So just use ->reshape_progress.
4235 */
4237 /* 'next' is the earliest device address that we might
4238 * write to for this chunk in the new layout
4239 */
4243 /* 'safe' is the last device address that we might read from
4244 * in the old layout after a restart
4245 */
4258 /* 'next' is after the last device address that we
4259 * might write to for this chunk in the new layout
4260 */
4263 /* 'safe' is the earliest device address that we might
4264 * read from in the old layout after a restart
4265 */
4268 /* Need to update metadata if 'next' might be beyond 'safe'
4269 * as that would possibly corrupt data
4270 */
4280 }
4284 /* Need to update reshape_position in metadata */
4290 else
4299 }
4301 read_more:
4302 /* Now schedule reads for blocks from sector_nr to last */
4314 /* Cannot read from here, so need to record bad blocks
4315 * on all the target devices.
4316 */
4317 // FIXME
4320 }
4338 /* Now find the locations in the new layout */
4354 }
4369 }
4371 /* Now add as many pages as possible to all of these bios. */
4384 /* Didn't fit, must stop */
4388 /* Remove last page from this bio */
4392 }
4394 }
4397 }
4398 bio_full:
4401 /* Now submit the read */
4411 /* Now that we have done the whole section we can
4412 * update reshape_progress
4413 */
4416 else
4420 }
4426 {
4427 /* Reshape read completed. Hopefully we have a block
4428 * to write out.
4429 * If we got a read error then we do sync 1-page reads from
4430 * elsewhere until we find the data - or give up.
4431 */
4437 /* Reshape has been aborted */
4440 }
4442 /* We definitely have the data in the pages, schedule the
4443 * writes.
4444 */
4456 }
4464 }
4466 }
4469 {
4480 /* read-ahead size must cover two whole stripes, which is
4481 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4482 */
4489 }
4491 }
4496 {
4497 /* Use sync reads to get the blocks from somewhere else */
4523 sector_t addr;
4531 addr,
4537 failed:
4538 slot++;
4543 }
4545 /* couldn't read this block, must give up */
4549 }
4551 idx++;
4552 }
4554 }
4557 {
4573 }
4576 /* FIXME should record badblock */
4578 }
4582 }
4585 {
4591 }
4594 {
4606 }
4614 d++) {
4621 }
4622 }
4628 }
4631 {
4651 };
4654 {
4656 }
4659 {
4661 }