| blob | 9ccb107c982e5d42bae477234f10784e58180e25 |
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 * use_far_sets (stored in bit 17 of layout )
42 *
43 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most 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 being very far
54 * apart on disk, there are adjacent stripes.
55 *
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
63 * on a device):
64 * A B C D A B C D E
65 * ... ...
66 * D A B C E A B C D
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
71 */
73 /*
74 * Number of guaranteed r10bios in case of extreme VM load:
75 */
76 #define NR_RAID10_BIOS 256
78 /* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
82 */
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
87 */
88 #define IO_MADE_GOOD ((struct bio *)2)
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92 /* When there are this many requests queued to be written by
93 * the raid10 thread, we become 'congested' to provide back-pressure
94 * for writeback.
95 */
108 {
112 /* allocate a r10bio with room for raid_disks entries in the
113 * bios array */
115 }
118 {
120 }
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
130 /*
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
135 *
136 */
138 {
153 else
156 /*
157 * Allocate bios.
158 */
170 }
171 /*
172 * Allocate RESYNC_PAGES data pages and attach them
173 * where needed.
174 */
181 /* we can share bv_page's during recovery
182 * and reshape */
194 }
195 }
199 out_free_pages:
206 out_free_bio:
212 }
215 }
218 {
230 }
232 }
236 }
238 }
241 {
253 }
254 }
257 {
262 }
265 {
271 }
274 {
284 /* wake up frozen array... */
288 }
290 /*
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
293 * cache layer.
294 */
296 {
313 /*
314 * Wake up any possible resync thread that waits for the device
315 * to go idle.
316 */
318 }
320 }
322 /*
323 * Update disk head position estimator based on IRQ completion info.
324 */
326 {
331 }
333 /*
334 * Find the disk number which triggered given bio
335 */
338 {
348 }
349 }
359 }
362 {
373 /*
374 * this branch is our 'one mirror IO has finished' event handler:
375 */
379 /*
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
383 *
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
387 */
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
394 */
398 }
403 /*
404 * oops, read error - keep the refcount on the rdev
405 */
414 }
415 }
418 {
419 /* clear the bitmap if all writes complete successfully */
425 }
428 {
436 else
438 }
439 }
440 }
443 {
460 }
461 /*
462 * this branch is our 'one mirror IO has finished' event handler:
463 */
466 /* Never record new bad blocks to replacement,
467 * just fail it.
468 */
477 }
479 /*
480 * Set R10BIO_Uptodate in our master bio, so that
481 * we will return a good error code for to the higher
482 * levels even if IO on some other mirrored buffer fails.
483 *
484 * The 'master' represents the composite IO operation to
485 * user-side. So if something waits for IO, then it will
486 * wait for the 'master' bio.
487 */
488 sector_t first_bad;
491 /*
492 * Do not set R10BIO_Uptodate if the current device is
493 * rebuilding or Faulty. This is because we cannot use
494 * such device for properly reading the data back (we could
495 * potentially use it, if the current write would have felt
496 * before rdev->recovery_offset, but for simplicity we don't
497 * check this here.
498 */
503 /* Maybe we can clear some bad blocks. */
511 else
515 }
516 }
518 /*
519 *
520 * Let's see if all mirrored write operations have finished
521 * already.
522 */
526 }
528 /*
529 * RAID10 layout manager
530 * As well as the chunksize and raid_disks count, there are two
531 * parameters: near_copies and far_copies.
532 * near_copies * far_copies must be <= raid_disks.
533 * Normally one of these will be 1.
534 * If both are 1, we get raid0.
535 * If near_copies == raid_disks, we get raid1.
536 *
537 * Chunks are laid out in raid0 style with near_copies copies of the
538 * first chunk, followed by near_copies copies of the next chunk and
539 * so on.
540 * If far_copies > 1, then after 1/far_copies of the array has been assigned
541 * as described above, we start again with a device offset of near_copies.
542 * So we effectively have another copy of the whole array further down all
543 * the drives, but with blocks on different drives.
544 * With this layout, and block is never stored twice on the one device.
545 *
546 * raid10_find_phys finds the sector offset of a given virtual sector
547 * on each device that it is on.
548 *
549 * raid10_find_virt does the reverse mapping, from a device and a
550 * sector offset to a virtual address
551 */
554 {
556 sector_t sector;
557 sector_t chunk;
558 sector_t stripe;
569 /* now calculate first sector/dev */
581 /* and calculate all the others */
588 slot++;
602 }
606 slot++;
607 }
608 dev++;
612 }
613 }
614 }
617 {
629 }
632 {
634 /* Never use conf->prev as this is only called during resync
635 * or recovery, so reshape isn't happening
636 */
650 }
651 }
666 else
668 }
670 }
674 }
676 /**
677 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
678 * @q: request queue
679 * @bvm: properties of new bio
680 * @biovec: the request that could be merged to it.
681 *
682 * Return amount of bytes we can accept at this offset
683 * This requires checking for end-of-chunk if near_copies != raid_disks,
684 * and for subordinate merge_bvec_fns if merge_check_needed.
685 */
689 {
708 /* bio_add cannot handle a negative return */
723 /* Cannot give any guidance during reshape */
727 }
744 }
745 }
756 }
757 }
758 }
760 }
762 }
764 /*
765 * This routine returns the disk from which the requested read should
766 * be done. There is a per-array 'next expected sequential IO' sector
767 * number - if this matches on the next IO then we use the last disk.
768 * There is also a per-disk 'last know head position' sector that is
769 * maintained from IRQ contexts, both the normal and the resync IO
770 * completion handlers update this position correctly. If there is no
771 * perfect sequential match then we pick the disk whose head is closest.
772 *
773 * If there are 2 mirrors in the same 2 devices, performance degrades
774 * because position is mirror, not device based.
775 *
776 * The rdev for the device selected will have nr_pending incremented.
777 */
779 /*
780 * FIXME: possibly should rethink readbalancing and do it differently
781 * depending on near_copies / far_copies geometry.
782 */
786 {
799 retry:
806 /*
807 * Check if we can balance. We can balance on the whole
808 * device if no resync is going on (recovery is ok), or below
809 * the resync window. We take the first readable disk when
810 * above the resync window.
811 */
817 sector_t first_bad;
819 sector_t dev_sector;
841 /* Already have a better slot */
844 /* Cannot read here. If this is the
845 * 'primary' device, then we must not read
846 * beyond 'bad_sectors' from another device.
847 */
854 sector_t good_sectors =
860 }
862 /* Must read from here */
864 }
872 /* This optimisation is debatable, and completely destroys
873 * sequential read speed for 'far copies' arrays. So only
874 * keep it for 'near' arrays, and review those later.
875 */
879 /* for far > 1 always use the lowest address */
882 else
889 }
890 }
894 }
899 /* Cannot risk returning a device that failed
900 * before we inc'ed nr_pending
901 */
904 }
912 }
915 {
927 i++) {
933 }
934 }
937 }
941 {
946 }
949 {
950 /* Any writes that have been queued but are awaiting
951 * bitmap updates get flushed here.
952 */
960 /* flush any pending bitmap writes to disk
961 * before proceeding w/ I/O */
970 /* Just ignore it */
972 else
975 }
978 }
980 /* Barriers....
981 * Sometimes we need to suspend IO while we do something else,
982 * either some resync/recovery, or reconfigure the array.
983 * To do this we raise a 'barrier'.
984 * The 'barrier' is a counter that can be raised multiple times
985 * to count how many activities are happening which preclude
986 * normal IO.
987 * We can only raise the barrier if there is no pending IO.
988 * i.e. if nr_pending == 0.
989 * We choose only to raise the barrier if no-one is waiting for the
990 * barrier to go down. This means that as soon as an IO request
991 * is ready, no other operations which require a barrier will start
992 * until the IO request has had a chance.
993 *
994 * So: regular IO calls 'wait_barrier'. When that returns there
995 * is no backgroup IO happening, It must arrange to call
996 * allow_barrier when it has finished its IO.
997 * backgroup IO calls must call raise_barrier. Once that returns
998 * there is no normal IO happeing. It must arrange to call
999 * lower_barrier when the particular background IO completes.
1000 */
1003 {
1007 /* Wait until no block IO is waiting (unless 'force') */
1011 /* block any new IO from starting */
1014 /* Now wait for all pending IO to complete */
1020 }
1023 {
1029 }
1032 {
1036 /* Wait for the barrier to drop.
1037 * However if there are already pending
1038 * requests (preventing the barrier from
1039 * rising completely), and the
1040 * pre-process bio queue isn't empty,
1041 * then don't wait, as we need to empty
1042 * that queue to get the nr_pending
1043 * count down.
1044 */
1052 }
1055 }
1058 {
1064 }
1067 {
1068 /* stop syncio and normal IO and wait for everything to
1069 * go quiet.
1070 * We increment barrier and nr_waiting, and then
1071 * wait until nr_pending match nr_queued+extra
1072 * This is called in the context of one normal IO request
1073 * that has failed. Thus any sync request that might be pending
1074 * will be blocked by nr_pending, and we need to wait for
1075 * pending IO requests to complete or be queued for re-try.
1076 * Thus the number queued (nr_queued) plus this request (extra)
1077 * must match the number of pending IOs (nr_pending) before
1078 * we continue.
1079 */
1089 }
1092 {
1093 /* reverse the effect of the freeze */
1099 }
1103 {
1107 else
1109 }
1115 };
1118 {
1120 cb);
1134 }
1136 /* we aren't scheduling, so we can do the write-out directly. */
1146 /* Just ignore it */
1148 else
1151 }
1153 }
1156 {
1180 }
1182 /* If this request crosses a chunk boundary, we need to
1183 * split it. This will only happen for 1 PAGE (or less) requests.
1184 */
1186 > chunk_sects
1190 /* Sanity check -- queue functions should prevent this happening */
1193 /* This is a one page bio that upper layers
1194 * refuse to split for us, so we need to split it.
1195 */
1199 /* Each of these 'make_request' calls will call 'wait_barrier'.
1200 * If the first succeeds but the second blocks due to the resync
1201 * thread raising the barrier, we will deadlock because the
1202 * IO to the underlying device will be queued in generic_make_request
1203 * and will never complete, so will never reduce nr_pending.
1204 * So increment nr_waiting here so no new raise_barriers will
1205 * succeed, and so the second wait_barrier cannot block.
1206 */
1221 bad_map:
1228 }
1232 /*
1233 * Register the new request and wait if the reconstruction
1234 * thread has put up a bar for new requests.
1235 * Continue immediately if no resync is active currently.
1236 */
1243 /* IO spans the reshape position. Need to wait for
1244 * reshape to pass
1245 */
1251 }
1259 /* Need to update reshape_position in metadata */
1268 }
1279 /* We might need to issue multiple reads to different
1280 * devices if there are bad blocks around, so we keep
1281 * track of the number of reads in bio->bi_phys_segments.
1282 * If this is 0, there is only one r10_bio and no locking
1283 * will be needed when the request completes. If it is
1284 * non-zero, then it is the number of not-completed requests.
1285 */
1290 /*
1291 * read balancing logic:
1292 */
1296 read_again:
1301 }
1306 max_sectors);
1319 /* Could not read all from this device, so we will
1320 * need another r10_bio.
1321 */
1328 else
1331 /* Cannot call generic_make_request directly
1332 * as that will be queued in __generic_make_request
1333 * and subsequent mempool_alloc might block
1334 * waiting for it. so hand bio over to raid10d.
1335 */
1349 }
1351 /*
1352 * WRITE:
1353 */
1358 }
1359 /* first select target devices under rcu_lock and
1360 * inc refcount on their rdev. Record them by setting
1361 * bios[x] to bio
1362 * If there are known/acknowledged bad blocks on any device
1363 * on which we have seen a write error, we want to avoid
1364 * writing to those blocks. This potentially requires several
1365 * writes to write around the bad blocks. Each set of writes
1366 * gets its own r10_bio with a set of bios attached. The number
1367 * of r10_bios is recored in bio->bi_phys_segments just as with
1368 * the read case.
1369 */
1373 retry_write:
1389 }
1394 }
1408 }
1410 sector_t first_bad;
1416 max_sectors,
1419 /* Mustn't write here until the bad block
1420 * is acknowledged
1421 */
1426 }
1428 /* Cannot write here at all */
1431 /* Mustn't write more than bad_sectors
1432 * to other devices yet
1433 */
1435 /* We don't set R10BIO_Degraded as that
1436 * only applies if the disk is missing,
1437 * so it might be re-added, and we want to
1438 * know to recover this chunk.
1439 * In this case the device is here, and the
1440 * fact that this chunk is not in-sync is
1441 * recorded in the bad block log.
1442 */
1444 }
1449 }
1450 }
1454 }
1458 }
1459 }
1463 /* Have to wait for this device to get unblocked, then retry */
1471 }
1477 /* Race with remove_disk */
1480 }
1482 }
1483 }
1488 }
1491 /* We are splitting this into multiple parts, so
1492 * we need to prepare for allocating another r10_bio.
1493 */
1498 else
1501 }
1514 max_sectors);
1519 rdev));
1532 cb);
1533 else
1542 }
1546 }
1551 /* Replacement just got moved to main 'rdev' */
1554 }
1557 max_sectors);
1576 }
1577 }
1579 /* Don't remove the bias on 'remaining' (one_write_done) until
1580 * after checking if we need to go around again.
1581 */
1585 /* We need another r10_bio. It has already been counted
1586 * in bio->bi_phys_segments.
1587 */
1597 }
1600 /* In case raid10d snuck in to freeze_array */
1602 }
1605 {
1616 else
1618 }
1626 }
1628 /* check if there are enough drives for
1629 * every block to appear on atleast one.
1630 * Don't consider the device numbered 'ignore'
1631 * as we might be about to remove it.
1632 */
1634 {
1644 }
1656 cnt++;
1658 }
1664 out:
1667 }
1670 {
1671 /* when calling 'enough', both 'prev' and 'geo' must
1672 * be stable.
1673 * This is ensured if ->reconfig_mutex or ->device_lock
1674 * is held.
1675 */
1678 }
1681 {
1686 /*
1687 * If it is not operational, then we have already marked it as dead
1688 * else if it is the last working disks, ignore the error, let the
1689 * next level up know.
1690 * else mark the drive as failed
1691 */
1695 /*
1696 * Don't fail the drive, just return an IO error.
1697 */
1700 }
1703 /*
1704 * If recovery is running, make sure it aborts.
1705 */
1716 }
1719 {
1727 }
1739 }
1740 }
1743 {
1749 }
1752 {
1759 /*
1760 * Find all non-in_sync disks within the RAID10 configuration
1761 * and mark them in_sync
1762 */
1769 /* Replacement has just become active */
1772 count++;
1774 /* Replaced device not technically faulty,
1775 * but we need to be sure it gets removed
1776 * and never re-added.
1777 */
1781 }
1787 count++;
1789 }
1790 }
1797 }
1801 {
1810 /* only hot-add to in-sync arrays, as recovery is
1811 * very different from resync
1812 */
1823 }
1828 else
1848 }
1862 }
1864 /* Some requests might not have seen this new
1865 * merge_bvec_fn. We must wait for them to complete
1866 * before merging the device fully.
1867 * First we make sure any code which has tested
1868 * our function has submitted the request, then
1869 * we wait for all outstanding requests to complete.
1870 */
1875 }
1882 }
1885 {
1897 else
1904 }
1905 /* Only remove faulty devices if recovery
1906 * is not possible.
1907 */
1915 }
1919 /* lost the race, try later */
1924 /* We must have just cleared 'rdev' */
1928 * but will never see neither -- if they are careful.
1929 */
1933 /* We might have just remove the Replacement as faulty
1934 * Clear the flag just in case
1935 */
1940 abort:
1944 }
1948 {
1954 /* this is a reshape read */
1961 else
1962 /* The write handler will notice the lack of
1963 * R10BIO_Uptodate and record any errors etc
1964 */
1968 /* for reconstruct, we always reschedule after a read.
1969 * for resync, only after all reads
1970 */
1974 /* we have read all the blocks,
1975 * do the comparison in process context in raid10d
1976 */
1978 }
1979 }
1982 {
1987 /* the primary of several recovery bios */
1992 else
2001 else
2004 }
2005 }
2006 }
2009 {
2015 sector_t first_bad;
2024 else
2036 }
2046 }
2048 /*
2049 * Note: sync and recover and handled very differently for raid10
2050 * This code is for resync.
2051 * For resync, we read through virtual addresses and read all blocks.
2052 * If there is any error, we schedule a write. The lowest numbered
2053 * drive is authoritative.
2054 * However requests come for physical address, so we need to map.
2055 * For every physical address there are raid_disks/copies virtual addresses,
2056 * which is always are least one, but is not necessarly an integer.
2057 * This means that a physical address can span multiple chunks, so we may
2058 * have to submit multiple io requests for a single sync request.
2059 */
2060 /*
2061 * We check if all blocks are in-sync and only write to blocks that
2062 * aren't in sync
2063 */
2065 {
2073 /* find the first device with a block */
2085 /* now find blocks with errors */
2096 /* We know that the bi_io_vec layout is the same for
2097 * both 'first' and 'i', so we just compare them.
2098 * All vec entries are PAGE_SIZE;
2099 */
2107 len))
2110 }
2115 /* Don't fix anything. */
2117 }
2118 /* Ok, we need to write this bio, either to correct an
2119 * inconsistency or to correct an unreadable block.
2120 * First we need to fixup bv_offset, bv_len and
2121 * bi_vecs, as the read request might have corrupted these
2122 */
2137 PAGE_SIZE);
2138 }
2149 }
2151 /* Now write out to any replacement devices
2152 * that are active
2153 */
2165 PAGE_SIZE);
2171 }
2173 done:
2177 }
2178 }
2180 /*
2181 * Now for the recovery code.
2182 * Recovery happens across physical sectors.
2183 * We recover all non-is_sync drives by finding the virtual address of
2184 * each, and then choose a working drive that also has that virt address.
2185 * There is a separate r10_bio for each non-in_sync drive.
2186 * Only the first two slots are in use. The first for reading,
2187 * The second for writing.
2188 *
2189 */
2191 {
2192 /* We got a read error during recovery.
2193 * We repeat the read in smaller page-sized sections.
2194 * If a read succeeds, write it to the new device or record
2195 * a bad block if we cannot.
2196 * If a read fails, record a bad block on both old and
2197 * new devices.
2198 */
2211 sector_t addr;
2220 addr,
2228 addr,
2238 }
2239 }
2241 /* We don't worry if we cannot set a bad block -
2242 * it really is bad so there is no loss in not
2243 * recording it yet
2244 */
2248 /* need bad block on destination too */
2253 /* just abort the recovery */
2255 "md/raid10:%s: recovery aborted"
2264 }
2265 }
2266 }
2270 idx++;
2271 }
2272 }
2275 {
2284 }
2286 /*
2287 * share the pages with the first bio
2288 * and submit the write request
2289 */
2293 /* Need to test wbio2->bi_end_io before we call
2294 * generic_make_request as if the former is NULL,
2295 * the latter is free to free wbio2.
2296 */
2303 }
2309 }
2310 }
2313 /*
2314 * Used by fix_read_error() to decay the per rdev read_errors.
2315 * We halve the read error count for every hour that has elapsed
2316 * since the last recorded read error.
2317 *
2318 */
2320 {
2329 /* first time we've seen a read error */
2332 }
2339 /*
2340 * if hours_since_last is > the number of bits in read_errors
2341 * just set read errors to 0. We do this to avoid
2342 * overflowing the shift of read_errors by hours_since_last.
2343 */
2346 else
2348 }
2352 {
2353 sector_t first_bad;
2360 /* success */
2367 }
2368 /* need to record an error - either for the block or the device */
2372 }
2374 /*
2375 * This is a kernel thread which:
2376 *
2377 * 1. Retries failed read operations on working mirrors.
2378 * 2. Updates the raid superblock when problems encounter.
2379 * 3. Performs writes following reads for array synchronising.
2380 */
2383 {
2390 /* still own a reference to this rdev, so it cannot
2391 * have been cleared recently.
2392 */
2396 /* drive has already been failed, just ignore any
2397 more fix_read_error() attempts */
2407 "md/raid10:%s: %s: Raid device exceeded "
2417 }
2430 sector_t first_bad;
2444 sect,
2451 }
2452 sl++;
2459 /* Cannot read from anywhere, just mark the block
2460 * as bad on the first device to discourage future
2461 * reads.
2462 */
2467 rdev,
2474 }
2476 }
2479 /* write it back and re-read */
2486 sl--;
2498 sect,
2501 /* Well, this device is dead */
2503 "md/raid10:%s: read correction "
2504 "write failed"
2508 sect +
2510 rdev)),
2516 }
2519 }
2526 sl--;
2537 sect,
2539 READ)) {
2541 /* Well, this device is dead */
2543 "md/raid10:%s: unable to read back "
2544 "corrected sectors"
2548 sect +
2558 "md/raid10:%s: read error corrected"
2562 sect +
2566 }
2570 }
2575 }
2576 }
2579 {
2584 /* bio has the data to be written to slot 'i' where
2585 * we just recently had a write error.
2586 * We repeatedly clone the bio and trim down to one block,
2587 * then try the write. Where the write fails we record
2588 * a bad block.
2589 * It is conceivable that the bio doesn't exactly align with
2590 * blocks. We must handle this.
2591 *
2592 * We currently own a reference to the rdev.
2593 */
2596 sector_t sector;
2614 /* Write at 'sector' for 'sectors' */
2622 /* Failure! */
2631 }
2633 }
2636 {
2645 /* we got a read error. Maybe the drive is bad. Maybe just
2646 * the block and we can fix it.
2647 * We freeze all other IO, and try reading the block from
2648 * other devices. When we find one, we re-write
2649 * and check it that fixes the read error.
2650 * This is all done synchronously while the array is
2651 * frozen.
2652 */
2667 read_more:
2676 }
2681 KERN_ERR
2682 "md/raid10:%s: %s: redirecting "
2691 max_sectors);
2701 /* Drat - have to split this up more */
2710 else
2716 GFP_NOIO);
2729 }
2732 {
2733 /* Some sort of write request has finished and it
2734 * succeeded in writing where we thought there was a
2735 * bad block. So forget the bad block.
2736 * Or possibly if failed and we need to record
2737 * a bad block.
2738 */
2752 rdev,
2757 rdev,
2761 }
2768 rdev,
2773 rdev,
2777 }
2778 }
2787 rdev,
2797 }
2799 }
2804 rdev,
2808 }
2809 }
2814 }
2815 }
2818 {
2837 }
2857 /* just a partial read to be scheduled from a
2858 * separate context
2859 */
2862 }
2867 }
2869 }
2873 {
2888 }
2890 /*
2891 * perform a "sync" on one "block"
2892 *
2893 * We need to make sure that no normal I/O request - particularly write
2894 * requests - conflict with active sync requests.
2895 *
2896 * This is achieved by tracking pending requests and a 'barrier' concept
2897 * that can be installed to exclude normal IO requests.
2898 *
2899 * Resync and recovery are handled very differently.
2900 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2901 *
2902 * For resync, we iterate over virtual addresses, read all copies,
2903 * and update if there are differences. If only one copy is live,
2904 * skip it.
2905 * For recovery, we iterate over physical addresses, read a good
2906 * value for each non-in_sync drive, and over-write.
2907 *
2908 * So, for recovery we may have several outstanding complex requests for a
2909 * given address, one for each out-of-sync device. We model this by allocating
2910 * a number of r10_bio structures, one for each out-of-sync device.
2911 * As we setup these structures, we collect all bio's together into a list
2912 * which we then process collectively to add pages, and then process again
2913 * to pass to generic_make_request.
2914 *
2915 * The r10_bio structures are linked using a borrowed master_bio pointer.
2916 * This link is counted in ->remaining. When the r10_bio that points to NULL
2917 * has its remaining count decremented to 0, the whole complex operation
2918 * is complete.
2919 *
2920 */
2924 {
2931 sector_t sync_blocks;
2940 /*
2941 * Allow skipping a full rebuild for incremental assembly
2942 * of a clean array, like RAID1 does.
2943 */
2953 }
2955 skipped:
2961 /* If we aborted, we need to abort the
2962 * sync on the 'current' bitmap chucks (there can
2963 * be several when recovering multiple devices).
2964 * as we may have started syncing it but not finished.
2965 * We can find the current address in
2966 * mddev->curr_resync, but for recovery,
2967 * we need to convert that to several
2968 * virtual addresses.
2969 */
2974 }
2981 sector_t sect =
2985 }
2987 /* completed sync */
2991 /* Completed a full sync so the replacements
2992 * are now fully recovered.
2993 */
2997 ->recovery_offset
2999 }
3001 }
3006 }
3012 /* if there has been nothing to do on any drive,
3013 * then there is nothing to do at all..
3014 */
3017 }
3022 /* make sure whole request will fit in a chunk - if chunks
3023 * are meaningful
3024 */
3028 /*
3029 * If there is non-resync activity waiting for us then
3030 * put in a delay to throttle resync.
3031 */
3035 /* Again, very different code for resync and recovery.
3036 * Both must result in an r10bio with a list of bios that
3037 * have bi_end_io, bi_sector, bi_bdev set,
3038 * and bi_private set to the r10bio.
3039 * For recovery, we may actually create several r10bios
3040 * with 2 bios in each, that correspond to the bios in the main one.
3041 * In this case, the subordinate r10bios link back through a
3042 * borrowed master_bio pointer, and the counter in the master
3043 * includes a ref from each subordinate.
3044 */
3045 /* First, we decide what to do and set ->bi_end_io
3046 * To end_sync_read if we want to read, and
3047 * end_sync_write if we will want to write.
3048 */
3052 /* recovery... the complicated one */
3059 sector_t sect;
3066 &&
3073 /* want to reconstruct this device */
3077 /* last stripe is not complete - don't
3078 * try to recover this sector.
3079 */
3081 }
3082 /* Unless we are doing a full sync, or a replacement
3083 * we only need to recover the block if it is set in
3084 * the bitmap
3085 */
3093 /* yep, skip the sync_blocks here, but don't assume
3094 * that there will never be anything to do here
3095 */
3098 }
3113 /* Need to check if the array will still be
3114 * degraded
3115 */
3121 }
3137 /* This is where we read from */
3147 bad_sectors -= (sector
3152 }
3153 }
3165 /* and we write to 'i' (if not in_sync) */
3193 /* and maybe write to replacement */
3198 /* Note: if rdev != NULL, then bio
3199 * cannot be NULL as r10buf_pool_alloc will
3200 * have allocated it.
3201 * So the second test here is pointless.
3202 * But it keeps semantic-checkers happy, and
3203 * this comment keeps human reviewers
3204 * happy.
3205 */
3219 }
3221 /* Cannot recover, so abort the recovery or
3222 * record a bad block */
3224 /* problem is that there are bad blocks
3225 * on other device(s)
3226 */
3244 }
3251 mirror->recovery_disabled
3253 }
3259 }
3260 }
3267 }
3269 }
3271 /* resync. Schedule a read for every block at this virt offset */
3280 /* We can skip this block */
3283 }
3324 }
3325 }
3336 count++;
3343 /* Need to set up for writing to the replacement */
3358 count++;
3359 }
3366 mddev);
3371 mddev);
3372 }
3376 }
3377 }
3395 /* stop here */
3400 /* remove last page from this bio */
3404 }
3406 }
3410 bio_full:
3425 }
3426 }
3429 /* pretend they weren't skipped, it makes
3430 * no important difference in this case
3431 */
3435 giveup:
3436 /* There is nowhere to write, so all non-sync
3437 * drives must be failed or in resync, all drives
3438 * have a bad block, so try the next chunk...
3439 */
3444 chunks_skipped ++;
3447 }
3449 static sector_t
3451 {
3452 sector_t size;
3467 }
3470 {
3471 /* Calculate the number of sectors-per-device that will
3472 * actually be used, and set conf->dev_sectors and
3473 * conf->stride
3474 */
3480 /* 'size' is now the number of chunks in the array */
3481 /* calculate "used chunks per device" */
3484 /* We need to round up when dividing by raid_disks to
3485 * get the stride size.
3486 */
3496 }
3497 }
3501 {
3517 * updated yet. */
3522 }
3539 }
3542 {
3555 }
3561 }
3568 /* FIXME calc properly */
3571 GFP_KERNEL);
3594 }
3598 else
3599 /* far_copies must be 1 */
3601 }
3615 out:
3625 }
3627 }
3630 {
3635 sector_t size;
3645 }
3661 else
3664 }
3686 }
3706 }
3712 else
3715 }
3716 /* need to check that every block has at least one working mirror */
3721 }
3724 /* must ensure that shape change is supported */
3731 }
3737 i++) {
3742 /* The replacement is all we have - use it */
3746 }
3754 }
3756 }
3766 /*
3767 * Ok, everything is just fine now
3768 */
3780 /* Calculate max read-ahead size.
3781 * We need to readahead at least twice a whole stripe....
3782 * maybe...
3783 */
3788 }
3803 /* This cannot work */
3806 }
3816 }
3820 out_free_conf:
3828 out:
3830 }
3833 {
3841 /* the unplug fn references 'conf'*/
3851 }
3854 {
3864 }
3865 }
3868 {
3869 /* Resize of 'far' arrays is not supported.
3870 * For 'near' and 'offset' arrays we can set the
3871 * number of sectors used to be an appropriate multiple
3872 * of the chunk size.
3873 * For 'offset', this is far_copies*chunksize.
3874 * For 'near' the multiplier is the LCM of
3875 * near_copies and raid_disks.
3876 * So if far_copies > 1 && !far_offset, fail.
3877 * Else find LCM(raid_disks, near_copy)*far_copies and
3878 * multiply by chunk_size. Then round to this number.
3879 * This is mostly done by raid10_size()
3880 */
3899 }
3907 }
3912 }
3915 {
3923 }
3925 /* Set new parameters */
3927 /* new layout: far_copies = 1, near_copies = 2 */
3932 /* make sure it will be not marked as dirty */
3941 }
3944 }
3947 {
3950 /* raid10 can take over:
3951 * raid0 - providing it has only two drives
3952 */
3954 /* for raid0 takeover only one zone is supported */
3961 }
3963 }
3965 }
3968 {
3969 /* Called when there is a request to change
3970 * - layout (to ->new_layout)
3971 * - chunk size (to ->new_chunk_sectors)
3972 * - raid_disks (by delta_disks)
3973 * or when trying to restart a reshape that was ongoing.
3974 *
3975 * We need to validate the request and possibly allocate
3976 * space if that might be an issue later.
3977 *
3978 * Currently we reject any reshape of a 'far' mode array,
3979 * allow chunk size to change if new is generally acceptable,
3980 * allow raid_disks to increase, and allow
3981 * a switch between 'near' mode and 'offset' mode.
3982 */
3990 /* mustn't change number of copies */
3993 /* Cannot switch to 'far' mode */
3997 /* not factor of array size */
4006 /* allocate new 'mirrors' list */
4011 GFP_KERNEL);
4014 }
4016 }
4018 /*
4019 * Need to check if array has failed when deciding whether to:
4020 * - start an array
4021 * - remove non-faulty devices
4022 * - add a spare
4023 * - allow a reshape
4024 * This determination is simple when no reshape is happening.
4025 * However if there is a reshape, we need to carefully check
4026 * both the before and after sections.
4027 * This is because some failed devices may only affect one
4028 * of the two sections, and some non-in_sync devices may
4029 * be insync in the section most affected by failed devices.
4030 */
4032 {
4038 /* 'prev' section first */
4042 degraded++;
4044 /* When we can reduce the number of devices in
4045 * an array, this might not contribute to
4046 * 'degraded'. It does now.
4047 */
4048 degraded++;
4049 }
4058 degraded2++;
4060 /* If reshape is increasing the number of devices,
4061 * this section has already been recovered, so
4062 * it doesn't contribute to degraded.
4063 * else it does.
4064 */
4066 degraded2++;
4067 }
4068 }
4073 }
4076 {
4077 /* A 'reshape' has been requested. This commits
4078 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4079 * This also checks if there are enough spares and adds them
4080 * to the array.
4081 * We currently require enough spares to make the final
4082 * array non-degraded. We also require that the difference
4083 * between old and new data_offset - on each device - is
4084 * enough that we never risk over-writing.
4085 */
4110 spares++;
4120 }
4121 }
4139 }
4149 }
4163 }
4172 else
4177 }
4180 /* This is a spare that was manually added */
4182 }
4183 }
4184 /* When a reshape changes the number of devices,
4185 * ->degraded is measured against the larger of the
4186 * pre and post numbers.
4187 */
4205 }
4211 abort:
4223 }
4225 /* Calculate the last device-address that could contain
4226 * any block from the chunk that includes the array-address 's'
4227 * and report the next address.
4228 * i.e. the address returned will be chunk-aligned and after
4229 * any data that is in the chunk containing 's'.
4230 */
4232 {
4240 }
4242 /* Calculate the first device-address that could contain
4243 * any block from the chunk that includes the array-address 's'.
4244 * This too will be the start of a chunk
4245 */
4247 {
4254 }
4258 {
4259 /* We simply copy at most one chunk (smallest of old and new)
4260 * at a time, possibly less if that exceeds RESYNC_PAGES,
4261 * or we hit a bad block or something.
4262 * This might mean we pause for normal IO in the middle of
4263 * a chunk, but that is not a problem was mddev->reshape_position
4264 * can record any location.
4265 *
4266 * If we will want to write to a location that isn't
4267 * yet recorded as 'safe' (i.e. in metadata on disk) then
4268 * we need to flush all reshape requests and update the metadata.
4269 *
4270 * When reshaping forwards (e.g. to more devices), we interpret
4271 * 'safe' as the earliest block which might not have been copied
4272 * down yet. We divide this by previous stripe size and multiply
4273 * by previous stripe length to get lowest device offset that we
4274 * cannot write to yet.
4275 * We interpret 'sector_nr' as an address that we want to write to.
4276 * From this we use last_device_address() to find where we might
4277 * write to, and first_device_address on the 'safe' position.
4278 * If this 'next' write position is after the 'safe' position,
4279 * we must update the metadata to increase the 'safe' position.
4280 *
4281 * When reshaping backwards, we round in the opposite direction
4282 * and perform the reverse test: next write position must not be
4283 * less than current safe position.
4284 *
4285 * In all this the minimum difference in data offsets
4286 * (conf->offset_diff - always positive) allows a bit of slack,
4287 * so next can be after 'safe', but not by more than offset_disk
4288 *
4289 * We need to prepare all the bios here before we start any IO
4290 * to ensure the size we choose is acceptable to all devices.
4291 * The means one for each copy for write-out and an extra one for
4292 * read-in.
4293 * We store the read-in bio in ->master_bio and the others in
4294 * ->devs[x].bio and ->devs[x].repl_bio.
4295 */
4309 /* If restarting in the middle, skip the initial sectors */
4322 }
4323 }
4325 /* We don't use sector_nr to track where we are up to
4326 * as that doesn't work well for ->reshape_backwards.
4327 * So just use ->reshape_progress.
4328 */
4330 /* 'next' is the earliest device address that we might
4331 * write to for this chunk in the new layout
4332 */
4336 /* 'safe' is the last device address that we might read from
4337 * in the old layout after a restart
4338 */
4351 /* 'next' is after the last device address that we
4352 * might write to for this chunk in the new layout
4353 */
4356 /* 'safe' is the earliest device address that we might
4357 * read from in the old layout after a restart
4358 */
4361 /* Need to update metadata if 'next' might be beyond 'safe'
4362 * as that would possibly corrupt data
4363 */
4373 }
4377 /* Need to update reshape_position in metadata */
4383 else
4392 }
4394 read_more:
4395 /* Now schedule reads for blocks from sector_nr to last */
4407 /* Cannot read from here, so need to record bad blocks
4408 * on all the target devices.
4409 */
4410 // FIXME
4413 }
4430 /* Now find the locations in the new layout */
4446 }
4458 }
4460 /* Now add as many pages as possible to all of these bios. */
4473 /* Didn't fit, must stop */
4477 /* Remove last page from this bio */
4481 }
4483 }
4486 }
4487 bio_full:
4490 /* Now submit the read */
4500 /* Now that we have done the whole section we can
4501 * update reshape_progress
4502 */
4505 else
4509 }
4515 {
4516 /* Reshape read completed. Hopefully we have a block
4517 * to write out.
4518 * If we got a read error then we do sync 1-page reads from
4519 * elsewhere until we find the data - or give up.
4520 */
4526 /* Reshape has been aborted */
4529 }
4531 /* We definitely have the data in the pages, schedule the
4532 * writes.
4533 */
4545 }
4553 }
4555 }
4558 {
4569 /* read-ahead size must cover two whole stripes, which is
4570 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4571 */
4578 }
4580 }
4585 {
4586 /* Use sync reads to get the blocks from somewhere else */
4612 sector_t addr;
4620 addr,
4626 failed:
4627 slot++;
4632 }
4634 /* couldn't read this block, must give up */
4638 }
4640 idx++;
4641 }
4643 }
4646 {
4662 }
4665 /* FIXME should record badblock */
4667 }
4671 }
4674 {
4680 }
4683 {
4695 }
4703 d++) {
4710 }
4711 }
4717 }
4720 {
4740 };
4743 {
4745 }
4748 {
4750 }