| blob | a48c215f1262e74246ea49ef25aa6f5892997034 |
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 }
919 }
921 /* Barriers....
922 * Sometimes we need to suspend IO while we do something else,
923 * either some resync/recovery, or reconfigure the array.
924 * To do this we raise a 'barrier'.
925 * The 'barrier' is a counter that can be raised multiple times
926 * to count how many activities are happening which preclude
927 * normal IO.
928 * We can only raise the barrier if there is no pending IO.
929 * i.e. if nr_pending == 0.
930 * We choose only to raise the barrier if no-one is waiting for the
931 * barrier to go down. This means that as soon as an IO request
932 * is ready, no other operations which require a barrier will start
933 * until the IO request has had a chance.
934 *
935 * So: regular IO calls 'wait_barrier'. When that returns there
936 * is no backgroup IO happening, It must arrange to call
937 * allow_barrier when it has finished its IO.
938 * backgroup IO calls must call raise_barrier. Once that returns
939 * there is no normal IO happeing. It must arrange to call
940 * lower_barrier when the particular background IO completes.
941 */
944 {
948 /* Wait until no block IO is waiting (unless 'force') */
952 /* block any new IO from starting */
955 /* Now wait for all pending IO to complete */
961 }
964 {
970 }
973 {
977 /* Wait for the barrier to drop.
978 * However if there are already pending
979 * requests (preventing the barrier from
980 * rising completely), and the
981 * pre-process bio queue isn't empty,
982 * then don't wait, as we need to empty
983 * that queue to get the nr_pending
984 * count down.
985 */
992 );
994 }
997 }
1000 {
1006 }
1009 {
1010 /* stop syncio and normal IO and wait for everything to
1011 * go quiet.
1012 * We increment barrier and nr_waiting, and then
1013 * wait until nr_pending match nr_queued+1
1014 * This is called in the context of one normal IO request
1015 * that has failed. Thus any sync request that might be pending
1016 * will be blocked by nr_pending, and we need to wait for
1017 * pending IO requests to complete or be queued for re-try.
1018 * Thus the number queued (nr_queued) plus this request (1)
1019 * must match the number of pending IOs (nr_pending) before
1020 * we continue.
1021 */
1031 }
1034 {
1035 /* reverse the effect of the freeze */
1041 }
1045 {
1049 else
1051 }
1054 {
1073 }
1075 /* If this request crosses a chunk boundary, we need to
1076 * split it. This will only happen for 1 PAGE (or less) requests.
1077 */
1079 > chunk_sects
1083 /* Sanity check -- queue functions should prevent this happening */
1087 /* This is a one page bio that upper layers
1088 * refuse to split for us, so we need to split it.
1089 */
1093 /* Each of these 'make_request' calls will call 'wait_barrier'.
1094 * If the first succeeds but the second blocks due to the resync
1095 * thread raising the barrier, we will deadlock because the
1096 * IO to the underlying device will be queued in generic_make_request
1097 * and will never complete, so will never reduce nr_pending.
1098 * So increment nr_waiting here so no new raise_barriers will
1099 * succeed, and so the second wait_barrier cannot block.
1100 */
1115 bad_map:
1122 }
1126 /*
1127 * Register the new request and wait if the reconstruction
1128 * thread has put up a bar for new requests.
1129 * Continue immediately if no resync is active currently.
1130 */
1137 /* IO spans the reshape position. Need to wait for
1138 * reshape to pass
1139 */
1145 }
1153 /* Need to update reshape_position in metadata */
1162 }
1173 /* We might need to issue multiple reads to different
1174 * devices if there are bad blocks around, so we keep
1175 * track of the number of reads in bio->bi_phys_segments.
1176 * If this is 0, there is only one r10_bio and no locking
1177 * will be needed when the request completes. If it is
1178 * non-zero, then it is the number of not-completed requests.
1179 */
1184 /*
1185 * read balancing logic:
1186 */
1190 read_again:
1195 }
1200 max_sectors);
1213 /* Could not read all from this device, so we will
1214 * need another r10_bio.
1215 */
1222 else
1225 /* Cannot call generic_make_request directly
1226 * as that will be queued in __generic_make_request
1227 * and subsequent mempool_alloc might block
1228 * waiting for it. so hand bio over to raid10d.
1229 */
1244 }
1246 /*
1247 * WRITE:
1248 */
1253 }
1254 /* first select target devices under rcu_lock and
1255 * inc refcount on their rdev. Record them by setting
1256 * bios[x] to bio
1257 * If there are known/acknowledged bad blocks on any device
1258 * on which we have seen a write error, we want to avoid
1259 * writing to those blocks. This potentially requires several
1260 * writes to write around the bad blocks. Each set of writes
1261 * gets its own r10_bio with a set of bios attached. The number
1262 * of r10_bios is recored in bio->bi_phys_segments just as with
1263 * the read case.
1264 */
1268 retry_write:
1284 }
1289 }
1300 }
1302 sector_t first_bad;
1308 max_sectors,
1311 /* Mustn't write here until the bad block
1312 * is acknowledged
1313 */
1318 }
1320 /* Cannot write here at all */
1323 /* Mustn't write more than bad_sectors
1324 * to other devices yet
1325 */
1327 /* We don't set R10BIO_Degraded as that
1328 * only applies if the disk is missing,
1329 * so it might be re-added, and we want to
1330 * know to recover this chunk.
1331 * In this case the device is here, and the
1332 * fact that this chunk is not in-sync is
1333 * recorded in the bad block log.
1334 */
1336 }
1341 }
1342 }
1348 }
1349 }
1353 /* Have to wait for this device to get unblocked, then retry */
1361 }
1367 /* Race with remove_disk */
1370 }
1372 }
1373 }
1378 }
1381 /* We are splitting this into multiple parts, so
1382 * we need to prepare for allocating another r10_bio.
1383 */
1388 else
1391 }
1405 max_sectors);
1429 max_sectors);
1432 /* We are actively writing to the original device
1433 * so it cannot disappear, so the replacement cannot
1434 * become NULL here
1435 */
1438 r10_bio,
1452 }
1454 /* Don't remove the bias on 'remaining' (one_write_done) until
1455 * after checking if we need to go around again.
1456 */
1460 /* We need another r10_bio. It has already been counted
1461 * in bio->bi_phys_segments.
1462 */
1472 }
1475 /* In case raid10d snuck in to freeze_array */
1477 }
1480 {
1491 else
1493 }
1501 }
1503 /* check if there are enough drives for
1504 * every block to appear on atleast one.
1505 * Don't consider the device numbered 'ignore'
1506 * as we might be about to remove it.
1507 */
1509 {
1519 cnt++;
1521 }
1527 }
1530 {
1533 }
1536 {
1540 /*
1541 * If it is not operational, then we have already marked it as dead
1542 * else if it is the last working disks, ignore the error, let the
1543 * next level up know.
1544 * else mark the drive as failed
1545 */
1548 /*
1549 * Don't fail the drive, just return an IO error.
1550 */
1557 /*
1558 * if recovery is running, make sure it aborts.
1559 */
1561 }
1570 }
1573 {
1581 }
1593 }
1594 }
1597 {
1603 }
1606 {
1613 /*
1614 * Find all non-in_sync disks within the RAID10 configuration
1615 * and mark them in_sync
1616 */
1623 /* Replacement has just become active */
1626 count++;
1628 /* Replaced device not technically faulty,
1629 * but we need to be sure it gets removed
1630 * and never re-added.
1631 */
1635 }
1640 count++;
1642 }
1643 }
1650 }
1654 {
1663 /* only hot-add to in-sync arrays, as recovery is
1664 * very different from resync
1665 */
1676 }
1681 else
1700 }
1713 }
1715 /* Some requests might not have seen this new
1716 * merge_bvec_fn. We must wait for them to complete
1717 * before merging the device fully.
1718 * First we make sure any code which has tested
1719 * our function has submitted the request, then
1720 * we wait for all outstanding requests to complete.
1721 */
1726 }
1730 }
1733 {
1745 else
1752 }
1753 /* Only remove faulty devices if recovery
1754 * is not possible.
1755 */
1763 }
1767 /* lost the race, try later */
1772 /* We must have just cleared 'rdev' */
1776 * but will never see neither -- if they are careful.
1777 */
1781 /* We might have just remove the Replacement as faulty
1782 * Clear the flag just in case
1783 */
1788 abort:
1792 }
1796 {
1802 /* this is a reshape read */
1809 else
1810 /* The write handler will notice the lack of
1811 * R10BIO_Uptodate and record any errors etc
1812 */
1816 /* for reconstruct, we always reschedule after a read.
1817 * for resync, only after all reads
1818 */
1822 /* we have read all the blocks,
1823 * do the comparison in process context in raid10d
1824 */
1826 }
1827 }
1830 {
1835 /* the primary of several recovery bios */
1840 else
1849 else
1852 }
1853 }
1854 }
1857 {
1863 sector_t first_bad;
1872 else
1884 }
1894 }
1896 /*
1897 * Note: sync and recover and handled very differently for raid10
1898 * This code is for resync.
1899 * For resync, we read through virtual addresses and read all blocks.
1900 * If there is any error, we schedule a write. The lowest numbered
1901 * drive is authoritative.
1902 * However requests come for physical address, so we need to map.
1903 * For every physical address there are raid_disks/copies virtual addresses,
1904 * which is always are least one, but is not necessarly an integer.
1905 * This means that a physical address can span multiple chunks, so we may
1906 * have to submit multiple io requests for a single sync request.
1907 */
1908 /*
1909 * We check if all blocks are in-sync and only write to blocks that
1910 * aren't in sync
1911 */
1913 {
1921 /* find the first device with a block */
1933 /* now find blocks with errors */
1944 /* We know that the bi_io_vec layout is the same for
1945 * both 'first' and 'i', so we just compare them.
1946 * All vec entries are PAGE_SIZE;
1947 */
1957 /* Don't fix anything. */
1959 }
1960 /* Ok, we need to write this bio, either to correct an
1961 * inconsistency or to correct an unreadable block.
1962 * First we need to fixup bv_offset, bv_len and
1963 * bi_vecs, as the read request might have corrupted these
1964 */
1982 PAGE_SIZE);
1983 }
1994 }
1996 /* Now write out to any replacement devices
1997 * that are active
1998 */
2010 PAGE_SIZE);
2016 }
2018 done:
2022 }
2023 }
2025 /*
2026 * Now for the recovery code.
2027 * Recovery happens across physical sectors.
2028 * We recover all non-is_sync drives by finding the virtual address of
2029 * each, and then choose a working drive that also has that virt address.
2030 * There is a separate r10_bio for each non-in_sync drive.
2031 * Only the first two slots are in use. The first for reading,
2032 * The second for writing.
2033 *
2034 */
2036 {
2037 /* We got a read error during recovery.
2038 * We repeat the read in smaller page-sized sections.
2039 * If a read succeeds, write it to the new device or record
2040 * a bad block if we cannot.
2041 * If a read fails, record a bad block on both old and
2042 * new devices.
2043 */
2056 sector_t addr;
2065 addr,
2073 addr,
2083 }
2084 }
2086 /* We don't worry if we cannot set a bad block -
2087 * it really is bad so there is no loss in not
2088 * recording it yet
2089 */
2093 /* need bad block on destination too */
2098 /* just abort the recovery */
2100 "md/raid10:%s: recovery aborted"
2109 }
2110 }
2111 }
2115 idx++;
2116 }
2117 }
2120 {
2129 }
2131 /*
2132 * share the pages with the first bio
2133 * and submit the write request
2134 */
2142 }
2148 }
2149 }
2152 /*
2153 * Used by fix_read_error() to decay the per rdev read_errors.
2154 * We halve the read error count for every hour that has elapsed
2155 * since the last recorded read error.
2156 *
2157 */
2159 {
2168 /* first time we've seen a read error */
2171 }
2178 /*
2179 * if hours_since_last is > the number of bits in read_errors
2180 * just set read errors to 0. We do this to avoid
2181 * overflowing the shift of read_errors by hours_since_last.
2182 */
2185 else
2187 }
2191 {
2192 sector_t first_bad;
2199 /* success */
2206 }
2207 /* need to record an error - either for the block or the device */
2211 }
2213 /*
2214 * This is a kernel thread which:
2215 *
2216 * 1. Retries failed read operations on working mirrors.
2217 * 2. Updates the raid superblock when problems encounter.
2218 * 3. Performs writes following reads for array synchronising.
2219 */
2222 {
2229 /* still own a reference to this rdev, so it cannot
2230 * have been cleared recently.
2231 */
2235 /* drive has already been failed, just ignore any
2236 more fix_read_error() attempts */
2246 "md/raid10:%s: %s: Raid device exceeded "
2256 }
2269 sector_t first_bad;
2283 sect,
2290 }
2291 sl++;
2298 /* Cannot read from anywhere, just mark the block
2299 * as bad on the first device to discourage future
2300 * reads.
2301 */
2306 rdev,
2313 }
2315 }
2318 /* write it back and re-read */
2325 sl--;
2337 sect,
2340 /* Well, this device is dead */
2342 "md/raid10:%s: read correction "
2343 "write failed"
2347 sect +
2349 rdev)),
2355 }
2358 }
2365 sl--;
2376 sect,
2378 READ)) {
2380 /* Well, this device is dead */
2382 "md/raid10:%s: unable to read back "
2383 "corrected sectors"
2387 sect +
2397 "md/raid10:%s: read error corrected"
2401 sect +
2405 }
2409 }
2414 }
2415 }
2418 {
2420 }
2423 {
2434 }
2437 {
2442 /* bio has the data to be written to slot 'i' where
2443 * we just recently had a write error.
2444 * We repeatedly clone the bio and trim down to one block,
2445 * then try the write. Where the write fails we record
2446 * a bad block.
2447 * It is conceivable that the bio doesn't exactly align with
2448 * blocks. We must handle this.
2449 *
2450 * We currently own a reference to the rdev.
2451 */
2454 sector_t sector;
2472 /* Write at 'sector' for 'sectors' */
2480 /* Failure! */
2489 }
2491 }
2494 {
2503 /* we got a read error. Maybe the drive is bad. Maybe just
2504 * the block and we can fix it.
2505 * We freeze all other IO, and try reading the block from
2506 * other devices. When we find one, we re-write
2507 * and check it that fixes the read error.
2508 * This is all done synchronously while the array is
2509 * frozen.
2510 */
2525 read_more:
2534 }
2539 KERN_ERR
2540 "md/raid10:%s: %s: redirecting "
2549 max_sectors);
2559 /* Drat - have to split this up more */
2568 else
2574 GFP_NOIO);
2588 }
2591 {
2592 /* Some sort of write request has finished and it
2593 * succeeded in writing where we thought there was a
2594 * bad block. So forget the bad block.
2595 * Or possibly if failed and we need to record
2596 * a bad block.
2597 */
2611 rdev,
2616 rdev,
2620 }
2627 rdev,
2632 rdev,
2636 }
2637 }
2646 rdev,
2656 }
2658 }
2663 rdev,
2667 }
2668 }
2673 }
2674 }
2677 {
2695 }
2715 /* just a partial read to be scheduled from a
2716 * separate context
2717 */
2720 }
2725 }
2727 }
2731 {
2746 }
2748 /*
2749 * perform a "sync" on one "block"
2750 *
2751 * We need to make sure that no normal I/O request - particularly write
2752 * requests - conflict with active sync requests.
2753 *
2754 * This is achieved by tracking pending requests and a 'barrier' concept
2755 * that can be installed to exclude normal IO requests.
2756 *
2757 * Resync and recovery are handled very differently.
2758 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2759 *
2760 * For resync, we iterate over virtual addresses, read all copies,
2761 * and update if there are differences. If only one copy is live,
2762 * skip it.
2763 * For recovery, we iterate over physical addresses, read a good
2764 * value for each non-in_sync drive, and over-write.
2765 *
2766 * So, for recovery we may have several outstanding complex requests for a
2767 * given address, one for each out-of-sync device. We model this by allocating
2768 * a number of r10_bio structures, one for each out-of-sync device.
2769 * As we setup these structures, we collect all bio's together into a list
2770 * which we then process collectively to add pages, and then process again
2771 * to pass to generic_make_request.
2772 *
2773 * The r10_bio structures are linked using a borrowed master_bio pointer.
2774 * This link is counted in ->remaining. When the r10_bio that points to NULL
2775 * has its remaining count decremented to 0, the whole complex operation
2776 * is complete.
2777 *
2778 */
2782 {
2789 sector_t sync_blocks;
2798 skipped:
2804 /* If we aborted, we need to abort the
2805 * sync on the 'current' bitmap chucks (there can
2806 * be several when recovering multiple devices).
2807 * as we may have started syncing it but not finished.
2808 * We can find the current address in
2809 * mddev->curr_resync, but for recovery,
2810 * we need to convert that to several
2811 * virtual addresses.
2812 */
2816 }
2823 sector_t sect =
2827 }
2829 /* completed sync */
2833 /* Completed a full sync so the replacements
2834 * are now fully recovered.
2835 */
2839 ->recovery_offset
2841 }
2843 }
2848 }
2854 /* if there has been nothing to do on any drive,
2855 * then there is nothing to do at all..
2856 */
2859 }
2864 /* make sure whole request will fit in a chunk - if chunks
2865 * are meaningful
2866 */
2870 /*
2871 * If there is non-resync activity waiting for us then
2872 * put in a delay to throttle resync.
2873 */
2877 /* Again, very different code for resync and recovery.
2878 * Both must result in an r10bio with a list of bios that
2879 * have bi_end_io, bi_sector, bi_bdev set,
2880 * and bi_private set to the r10bio.
2881 * For recovery, we may actually create several r10bios
2882 * with 2 bios in each, that correspond to the bios in the main one.
2883 * In this case, the subordinate r10bios link back through a
2884 * borrowed master_bio pointer, and the counter in the master
2885 * includes a ref from each subordinate.
2886 */
2887 /* First, we decide what to do and set ->bi_end_io
2888 * To end_sync_read if we want to read, and
2889 * end_sync_write if we will want to write.
2890 */
2894 /* recovery... the complicated one */
2901 sector_t sect;
2908 &&
2915 /* want to reconstruct this device */
2919 /* last stripe is not complete - don't
2920 * try to recover this sector.
2921 */
2923 }
2924 /* Unless we are doing a full sync, or a replacement
2925 * we only need to recover the block if it is set in
2926 * the bitmap
2927 */
2935 /* yep, skip the sync_blocks here, but don't assume
2936 * that there will never be anything to do here
2937 */
2940 }
2955 /* Need to check if the array will still be
2956 * degraded
2957 */
2963 }
2979 /* This is where we read from */
2989 bad_sectors -= (sector
2994 }
2995 }
3006 /* and we write to 'i' (if not in_sync) */
3033 /* and maybe write to replacement */
3038 /* Note: if rdev != NULL, then bio
3039 * cannot be NULL as r10buf_pool_alloc will
3040 * have allocated it.
3041 * So the second test here is pointless.
3042 * But it keeps semantic-checkers happy, and
3043 * this comment keeps human reviewers
3044 * happy.
3045 */
3058 }
3060 /* Cannot recover, so abort the recovery or
3061 * record a bad block */
3067 /* problem is that there are bad blocks
3068 * on other device(s)
3069 */
3087 }
3094 mirror->recovery_disabled
3096 }
3098 }
3099 }
3106 }
3108 }
3110 /* resync. Schedule a read for every block at this virt offset */
3119 /* We can skip this block */
3122 }
3163 }
3164 }
3175 count++;
3182 /* Need to set up for writing to the replacement */
3196 count++;
3197 }
3204 mddev);
3209 mddev);
3210 }
3214 }
3215 }
3226 }
3244 /* stop here */
3249 /* remove last page from this bio */
3253 }
3255 }
3259 bio_full:
3273 }
3274 }
3277 /* pretend they weren't skipped, it makes
3278 * no important difference in this case
3279 */
3283 giveup:
3284 /* There is nowhere to write, so all non-sync
3285 * drives must be failed or in resync, all drives
3286 * have a bad block, so try the next chunk...
3287 */
3292 chunks_skipped ++;
3295 }
3297 static sector_t
3299 {
3300 sector_t size;
3315 }
3318 {
3319 /* Calculate the number of sectors-per-device that will
3320 * actually be used, and set conf->dev_sectors and
3321 * conf->stride
3322 */
3328 /* 'size' is now the number of chunks in the array */
3329 /* calculate "used chunks per device" */
3332 /* We need to round up when dividing by raid_disks to
3333 * get the stride size.
3334 */
3344 }
3345 }
3349 {
3365 * updated yet. */
3370 }
3386 }
3389 {
3402 }
3408 }
3415 /* FIXME calc properly */
3418 GFP_KERNEL);
3441 }
3445 else
3446 /* far_copies must be 1 */
3448 }
3462 out:
3472 }
3474 }
3477 {
3482 sector_t size;
3491 }
3504 else
3507 }
3529 }
3546 }
3548 /* need to check that every block has at least one working mirror */
3553 }
3556 /* must ensure that shape change is supported */
3563 }
3569 i++) {
3574 /* The replacement is all we have - use it */
3578 }
3586 }
3588 }
3598 /*
3599 * Ok, everything is just fine now
3600 */
3612 /* Calculate max read-ahead size.
3613 * We need to readahead at least twice a whole stripe....
3614 * maybe...
3615 */
3620 }
3635 /* This cannot work */
3638 }
3648 }
3652 out_free_conf:
3660 out:
3662 }
3665 {
3673 /* the unplug fn references 'conf'*/
3682 }
3685 {
3695 }
3696 }
3699 {
3700 /* Resize of 'far' arrays is not supported.
3701 * For 'near' and 'offset' arrays we can set the
3702 * number of sectors used to be an appropriate multiple
3703 * of the chunk size.
3704 * For 'offset', this is far_copies*chunksize.
3705 * For 'near' the multiplier is the LCM of
3706 * near_copies and raid_disks.
3707 * So if far_copies > 1 && !far_offset, fail.
3708 * Else find LCM(raid_disks, near_copy)*far_copies and
3709 * multiply by chunk_size. Then round to this number.
3710 * This is mostly done by raid10_size()
3711 */
3730 }
3738 }
3743 }
3746 {
3754 }
3756 /* Set new parameters */
3758 /* new layout: far_copies = 1, near_copies = 2 */
3763 /* make sure it will be not marked as dirty */
3772 }
3775 }
3778 {
3781 /* raid10 can take over:
3782 * raid0 - providing it has only two drives
3783 */
3785 /* for raid0 takeover only one zone is supported */
3792 }
3794 }
3796 }
3799 {
3800 /* Called when there is a request to change
3801 * - layout (to ->new_layout)
3802 * - chunk size (to ->new_chunk_sectors)
3803 * - raid_disks (by delta_disks)
3804 * or when trying to restart a reshape that was ongoing.
3805 *
3806 * We need to validate the request and possibly allocate
3807 * space if that might be an issue later.
3808 *
3809 * Currently we reject any reshape of a 'far' mode array,
3810 * allow chunk size to change if new is generally acceptable,
3811 * allow raid_disks to increase, and allow
3812 * a switch between 'near' mode and 'offset' mode.
3813 */
3821 /* mustn't change number of copies */
3824 /* Cannot switch to 'far' mode */
3828 /* not factor of array size */
3837 /* allocate new 'mirrors' list */
3842 GFP_KERNEL);
3845 }
3847 }
3849 /*
3850 * Need to check if array has failed when deciding whether to:
3851 * - start an array
3852 * - remove non-faulty devices
3853 * - add a spare
3854 * - allow a reshape
3855 * This determination is simple when no reshape is happening.
3856 * However if there is a reshape, we need to carefully check
3857 * both the before and after sections.
3858 * This is because some failed devices may only affect one
3859 * of the two sections, and some non-in_sync devices may
3860 * be insync in the section most affected by failed devices.
3861 */
3863 {
3869 /* 'prev' section first */
3873 degraded++;
3875 /* When we can reduce the number of devices in
3876 * an array, this might not contribute to
3877 * 'degraded'. It does now.
3878 */
3879 degraded++;
3880 }
3889 degraded2++;
3891 /* If reshape is increasing the number of devices,
3892 * this section has already been recovered, so
3893 * it doesn't contribute to degraded.
3894 * else it does.
3895 */
3897 degraded2++;
3898 }
3899 }
3904 }
3907 {
3908 /* A 'reshape' has been requested. This commits
3909 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3910 * This also checks if there are enough spares and adds them
3911 * to the array.
3912 * We currently require enough spares to make the final
3913 * array non-degraded. We also require that the difference
3914 * between old and new data_offset - on each device - is
3915 * enough that we never risk over-writing.
3916 */
3941 spares++;
3951 }
3952 }
3970 }
3980 }
3994 }
4003 else
4008 }
4011 /* This is a spare that was manually added */
4013 }
4014 }
4015 /* When a reshape changes the number of devices,
4016 * ->degraded is measured against the larger of the
4017 * pre and post numbers.
4018 */
4036 }
4042 abort:
4054 }
4056 /* Calculate the last device-address that could contain
4057 * any block from the chunk that includes the array-address 's'
4058 * and report the next address.
4059 * i.e. the address returned will be chunk-aligned and after
4060 * any data that is in the chunk containing 's'.
4061 */
4063 {
4071 }
4073 /* Calculate the first device-address that could contain
4074 * any block from the chunk that includes the array-address 's'.
4075 * This too will be the start of a chunk
4076 */
4078 {
4085 }
4089 {
4090 /* We simply copy at most one chunk (smallest of old and new)
4091 * at a time, possibly less if that exceeds RESYNC_PAGES,
4092 * or we hit a bad block or something.
4093 * This might mean we pause for normal IO in the middle of
4094 * a chunk, but that is not a problem was mddev->reshape_position
4095 * can record any location.
4096 *
4097 * If we will want to write to a location that isn't
4098 * yet recorded as 'safe' (i.e. in metadata on disk) then
4099 * we need to flush all reshape requests and update the metadata.
4100 *
4101 * When reshaping forwards (e.g. to more devices), we interpret
4102 * 'safe' as the earliest block which might not have been copied
4103 * down yet. We divide this by previous stripe size and multiply
4104 * by previous stripe length to get lowest device offset that we
4105 * cannot write to yet.
4106 * We interpret 'sector_nr' as an address that we want to write to.
4107 * From this we use last_device_address() to find where we might
4108 * write to, and first_device_address on the 'safe' position.
4109 * If this 'next' write position is after the 'safe' position,
4110 * we must update the metadata to increase the 'safe' position.
4111 *
4112 * When reshaping backwards, we round in the opposite direction
4113 * and perform the reverse test: next write position must not be
4114 * less than current safe position.
4115 *
4116 * In all this the minimum difference in data offsets
4117 * (conf->offset_diff - always positive) allows a bit of slack,
4118 * so next can be after 'safe', but not by more than offset_disk
4119 *
4120 * We need to prepare all the bios here before we start any IO
4121 * to ensure the size we choose is acceptable to all devices.
4122 * The means one for each copy for write-out and an extra one for
4123 * read-in.
4124 * We store the read-in bio in ->master_bio and the others in
4125 * ->devs[x].bio and ->devs[x].repl_bio.
4126 */
4140 /* If restarting in the middle, skip the initial sectors */
4153 }
4154 }
4156 /* We don't use sector_nr to track where we are up to
4157 * as that doesn't work well for ->reshape_backwards.
4158 * So just use ->reshape_progress.
4159 */
4161 /* 'next' is the earliest device address that we might
4162 * write to for this chunk in the new layout
4163 */
4167 /* 'safe' is the last device address that we might read from
4168 * in the old layout after a restart
4169 */
4182 /* 'next' is after the last device address that we
4183 * might write to for this chunk in the new layout
4184 */
4187 /* 'safe' is the earliest device address that we might
4188 * read from in the old layout after a restart
4189 */
4192 /* Need to update metadata if 'next' might be beyond 'safe'
4193 * as that would possibly corrupt data
4194 */
4204 }
4208 /* Need to update reshape_position in metadata */
4214 else
4223 }
4225 read_more:
4226 /* Now schedule reads for blocks from sector_nr to last */
4238 /* Cannot read from here, so need to record bad blocks
4239 * on all the target devices.
4240 */
4241 // FIXME
4244 }
4262 /* Now find the locations in the new layout */
4278 }
4293 }
4295 /* Now add as many pages as possible to all of these bios. */
4308 /* Didn't fit, must stop */
4312 /* Remove last page from this bio */
4316 }
4318 }
4321 }
4322 bio_full:
4325 /* Now submit the read */
4335 /* Now that we have done the whole section we can
4336 * update reshape_progress
4337 */
4340 else
4344 }
4350 {
4351 /* Reshape read completed. Hopefully we have a block
4352 * to write out.
4353 * If we got a read error then we do sync 1-page reads from
4354 * elsewhere until we find the data - or give up.
4355 */
4361 /* Reshape has been aborted */
4364 }
4366 /* We definitely have the data in the pages, schedule the
4367 * writes.
4368 */
4380 }
4388 }
4390 }
4393 {
4404 /* read-ahead size must cover two whole stripes, which is
4405 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4406 */
4413 }
4415 }
4420 {
4421 /* Use sync reads to get the blocks from somewhere else */
4447 sector_t addr;
4455 addr,
4461 failed:
4462 slot++;
4467 }
4469 /* couldn't read this block, must give up */
4473 }
4475 idx++;
4476 }
4478 }
4481 {
4497 }
4500 /* FIXME should record badblock */
4502 }
4506 }
4509 {
4515 }
4518 {
4530 }
4538 d++) {
4545 }
4546 }
4552 }
4555 {
4575 };
4578 {
4580 }
4583 {
4585 }