| blob | cb882aae9e20d4f7032a400884f45f50de57528f |
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 {
1175 /*
1176 * Register the new request and wait if the reconstruction
1177 * thread has put up a bar for new requests.
1178 * Continue immediately if no resync is active currently.
1179 */
1186 /* IO spans the reshape position. Need to wait for
1187 * reshape to pass
1188 */
1193 sectors);
1195 }
1203 /* Need to update reshape_position in metadata */
1212 }
1223 /* We might need to issue multiple reads to different
1224 * devices if there are bad blocks around, so we keep
1225 * track of the number of reads in bio->bi_phys_segments.
1226 * If this is 0, there is only one r10_bio and no locking
1227 * will be needed when the request completes. If it is
1228 * non-zero, then it is the number of not-completed requests.
1229 */
1234 /*
1235 * read balancing logic:
1236 */
1240 read_again:
1245 }
1250 max_sectors);
1263 /* Could not read all from this device, so we will
1264 * need another r10_bio.
1265 */
1272 else
1275 /* Cannot call generic_make_request directly
1276 * as that will be queued in __generic_make_request
1277 * and subsequent mempool_alloc might block
1278 * waiting for it. so hand bio over to raid10d.
1279 */
1289 sectors_handled;
1294 }
1296 /*
1297 * WRITE:
1298 */
1303 }
1304 /* first select target devices under rcu_lock and
1305 * inc refcount on their rdev. Record them by setting
1306 * bios[x] to bio
1307 * If there are known/acknowledged bad blocks on any device
1308 * on which we have seen a write error, we want to avoid
1309 * writing to those blocks. This potentially requires several
1310 * writes to write around the bad blocks. Each set of writes
1311 * gets its own r10_bio with a set of bios attached. The number
1312 * of r10_bios is recored in bio->bi_phys_segments just as with
1313 * the read case.
1314 */
1318 retry_write:
1334 }
1339 }
1353 }
1355 sector_t first_bad;
1361 max_sectors,
1364 /* Mustn't write here until the bad block
1365 * is acknowledged
1366 */
1371 }
1373 /* Cannot write here at all */
1376 /* Mustn't write more than bad_sectors
1377 * to other devices yet
1378 */
1380 /* We don't set R10BIO_Degraded as that
1381 * only applies if the disk is missing,
1382 * so it might be re-added, and we want to
1383 * know to recover this chunk.
1384 * In this case the device is here, and the
1385 * fact that this chunk is not in-sync is
1386 * recorded in the bad block log.
1387 */
1389 }
1394 }
1395 }
1399 }
1403 }
1404 }
1408 /* Have to wait for this device to get unblocked, then retry */
1416 }
1422 /* Race with remove_disk */
1425 }
1427 }
1428 }
1433 }
1436 /* We are splitting this into multiple parts, so
1437 * we need to prepare for allocating another r10_bio.
1438 */
1443 else
1446 }
1460 max_sectors);
1465 rdev));
1478 cb);
1479 else
1488 }
1492 }
1497 /* Replacement just got moved to main 'rdev' */
1500 }
1503 max_sectors);
1522 }
1523 }
1525 /* Don't remove the bias on 'remaining' (one_write_done) until
1526 * after checking if we need to go around again.
1527 */
1531 /* We need another r10_bio. It has already been counted
1532 * in bio->bi_phys_segments.
1533 */
1543 }
1545 }
1548 {
1558 }
1565 /*
1566 * If this request crosses a chunk boundary, we need to split
1567 * it.
1568 */
1581 }
1586 /* In case raid10d snuck in to freeze_array */
1588 }
1591 {
1602 else
1604 }
1612 }
1614 /* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1618 */
1620 {
1630 }
1642 cnt++;
1644 }
1650 out:
1653 }
1656 {
1657 /* when calling 'enough', both 'prev' and 'geo' must
1658 * be stable.
1659 * This is ensured if ->reconfig_mutex or ->device_lock
1660 * is held.
1661 */
1664 }
1667 {
1672 /*
1673 * If it is not operational, then we have already marked it as dead
1674 * else if it is the last working disks, ignore the error, let the
1675 * next level up know.
1676 * else mark the drive as failed
1677 */
1681 /*
1682 * Don't fail the drive, just return an IO error.
1683 */
1686 }
1689 /*
1690 * if recovery is running, make sure it aborts.
1691 */
1693 }
1703 }
1706 {
1714 }
1726 }
1727 }
1730 {
1736 }
1739 {
1746 /*
1747 * Find all non-in_sync disks within the RAID10 configuration
1748 * and mark them in_sync
1749 */
1756 /* Replacement has just become active */
1759 count++;
1761 /* Replaced device not technically faulty,
1762 * but we need to be sure it gets removed
1763 * and never re-added.
1764 */
1768 }
1774 count++;
1776 }
1777 }
1784 }
1788 {
1797 /* only hot-add to in-sync arrays, as recovery is
1798 * very different from resync
1799 */
1810 }
1815 else
1835 }
1849 }
1851 /* Some requests might not have seen this new
1852 * merge_bvec_fn. We must wait for them to complete
1853 * before merging the device fully.
1854 * First we make sure any code which has tested
1855 * our function has submitted the request, then
1856 * we wait for all outstanding requests to complete.
1857 */
1862 }
1869 }
1872 {
1884 else
1891 }
1892 /* Only remove faulty devices if recovery
1893 * is not possible.
1894 */
1902 }
1906 /* lost the race, try later */
1911 /* We must have just cleared 'rdev' */
1915 * but will never see neither -- if they are careful.
1916 */
1920 /* We might have just remove the Replacement as faulty
1921 * Clear the flag just in case
1922 */
1927 abort:
1931 }
1935 {
1941 /* this is a reshape read */
1948 else
1949 /* The write handler will notice the lack of
1950 * R10BIO_Uptodate and record any errors etc
1951 */
1955 /* for reconstruct, we always reschedule after a read.
1956 * for resync, only after all reads
1957 */
1961 /* we have read all the blocks,
1962 * do the comparison in process context in raid10d
1963 */
1965 }
1966 }
1969 {
1974 /* the primary of several recovery bios */
1979 else
1988 else
1991 }
1992 }
1993 }
1996 {
2002 sector_t first_bad;
2011 else
2023 }
2033 }
2035 /*
2036 * Note: sync and recover and handled very differently for raid10
2037 * This code is for resync.
2038 * For resync, we read through virtual addresses and read all blocks.
2039 * If there is any error, we schedule a write. The lowest numbered
2040 * drive is authoritative.
2041 * However requests come for physical address, so we need to map.
2042 * For every physical address there are raid_disks/copies virtual addresses,
2043 * which is always are least one, but is not necessarly an integer.
2044 * This means that a physical address can span multiple chunks, so we may
2045 * have to submit multiple io requests for a single sync request.
2046 */
2047 /*
2048 * We check if all blocks are in-sync and only write to blocks that
2049 * aren't in sync
2050 */
2052 {
2060 /* find the first device with a block */
2072 /* now find blocks with errors */
2083 /* We know that the bi_io_vec layout is the same for
2084 * both 'first' and 'i', so we just compare them.
2085 * All vec entries are PAGE_SIZE;
2086 */
2094 len))
2097 }
2102 /* Don't fix anything. */
2104 }
2105 /* Ok, we need to write this bio, either to correct an
2106 * inconsistency or to correct an unreadable block.
2107 * First we need to fixup bv_offset, bv_len and
2108 * bi_vecs, as the read request might have corrupted these
2109 */
2124 PAGE_SIZE);
2125 }
2136 }
2138 /* Now write out to any replacement devices
2139 * that are active
2140 */
2152 PAGE_SIZE);
2158 }
2160 done:
2164 }
2165 }
2167 /*
2168 * Now for the recovery code.
2169 * Recovery happens across physical sectors.
2170 * We recover all non-is_sync drives by finding the virtual address of
2171 * each, and then choose a working drive that also has that virt address.
2172 * There is a separate r10_bio for each non-in_sync drive.
2173 * Only the first two slots are in use. The first for reading,
2174 * The second for writing.
2175 *
2176 */
2178 {
2179 /* We got a read error during recovery.
2180 * We repeat the read in smaller page-sized sections.
2181 * If a read succeeds, write it to the new device or record
2182 * a bad block if we cannot.
2183 * If a read fails, record a bad block on both old and
2184 * new devices.
2185 */
2198 sector_t addr;
2207 addr,
2215 addr,
2225 }
2226 }
2228 /* We don't worry if we cannot set a bad block -
2229 * it really is bad so there is no loss in not
2230 * recording it yet
2231 */
2235 /* need bad block on destination too */
2240 /* just abort the recovery */
2242 "md/raid10:%s: recovery aborted"
2251 }
2252 }
2253 }
2257 idx++;
2258 }
2259 }
2262 {
2271 }
2273 /*
2274 * share the pages with the first bio
2275 * and submit the write request
2276 */
2280 /* Need to test wbio2->bi_end_io before we call
2281 * generic_make_request as if the former is NULL,
2282 * the latter is free to free wbio2.
2283 */
2290 }
2296 }
2297 }
2300 /*
2301 * Used by fix_read_error() to decay the per rdev read_errors.
2302 * We halve the read error count for every hour that has elapsed
2303 * since the last recorded read error.
2304 *
2305 */
2307 {
2316 /* first time we've seen a read error */
2319 }
2326 /*
2327 * if hours_since_last is > the number of bits in read_errors
2328 * just set read errors to 0. We do this to avoid
2329 * overflowing the shift of read_errors by hours_since_last.
2330 */
2333 else
2335 }
2339 {
2340 sector_t first_bad;
2347 /* success */
2354 }
2355 /* need to record an error - either for the block or the device */
2359 }
2361 /*
2362 * This is a kernel thread which:
2363 *
2364 * 1. Retries failed read operations on working mirrors.
2365 * 2. Updates the raid superblock when problems encounter.
2366 * 3. Performs writes following reads for array synchronising.
2367 */
2370 {
2377 /* still own a reference to this rdev, so it cannot
2378 * have been cleared recently.
2379 */
2383 /* drive has already been failed, just ignore any
2384 more fix_read_error() attempts */
2394 "md/raid10:%s: %s: Raid device exceeded "
2404 }
2417 sector_t first_bad;
2431 sect,
2438 }
2439 sl++;
2446 /* Cannot read from anywhere, just mark the block
2447 * as bad on the first device to discourage future
2448 * reads.
2449 */
2454 rdev,
2461 }
2463 }
2466 /* write it back and re-read */
2473 sl--;
2485 sect,
2488 /* Well, this device is dead */
2490 "md/raid10:%s: read correction "
2491 "write failed"
2495 sect +
2497 rdev)),
2503 }
2506 }
2513 sl--;
2524 sect,
2526 READ)) {
2528 /* Well, this device is dead */
2530 "md/raid10:%s: unable to read back "
2531 "corrected sectors"
2535 sect +
2545 "md/raid10:%s: read error corrected"
2549 sect +
2553 }
2557 }
2562 }
2563 }
2566 {
2571 /* bio has the data to be written to slot 'i' where
2572 * we just recently had a write error.
2573 * We repeatedly clone the bio and trim down to one block,
2574 * then try the write. Where the write fails we record
2575 * a bad block.
2576 * It is conceivable that the bio doesn't exactly align with
2577 * blocks. We must handle this.
2578 *
2579 * We currently own a reference to the rdev.
2580 */
2583 sector_t sector;
2601 /* Write at 'sector' for 'sectors' */
2609 /* Failure! */
2618 }
2620 }
2623 {
2632 /* we got a read error. Maybe the drive is bad. Maybe just
2633 * the block and we can fix it.
2634 * We freeze all other IO, and try reading the block from
2635 * other devices. When we find one, we re-write
2636 * and check it that fixes the read error.
2637 * This is all done synchronously while the array is
2638 * frozen.
2639 */
2654 read_more:
2663 }
2668 KERN_ERR
2669 "md/raid10:%s: %s: redirecting "
2686 /* Drat - have to split this up more */
2695 else
2701 GFP_NOIO);
2714 }
2717 {
2718 /* Some sort of write request has finished and it
2719 * succeeded in writing where we thought there was a
2720 * bad block. So forget the bad block.
2721 * Or possibly if failed and we need to record
2722 * a bad block.
2723 */
2737 rdev,
2742 rdev,
2746 }
2753 rdev,
2758 rdev,
2762 }
2763 }
2772 rdev,
2782 }
2784 }
2789 rdev,
2793 }
2794 }
2799 }
2800 }
2803 {
2822 }
2842 /* just a partial read to be scheduled from a
2843 * separate context
2844 */
2847 }
2852 }
2854 }
2858 {
2873 }
2875 /*
2876 * perform a "sync" on one "block"
2877 *
2878 * We need to make sure that no normal I/O request - particularly write
2879 * requests - conflict with active sync requests.
2880 *
2881 * This is achieved by tracking pending requests and a 'barrier' concept
2882 * that can be installed to exclude normal IO requests.
2883 *
2884 * Resync and recovery are handled very differently.
2885 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2886 *
2887 * For resync, we iterate over virtual addresses, read all copies,
2888 * and update if there are differences. If only one copy is live,
2889 * skip it.
2890 * For recovery, we iterate over physical addresses, read a good
2891 * value for each non-in_sync drive, and over-write.
2892 *
2893 * So, for recovery we may have several outstanding complex requests for a
2894 * given address, one for each out-of-sync device. We model this by allocating
2895 * a number of r10_bio structures, one for each out-of-sync device.
2896 * As we setup these structures, we collect all bio's together into a list
2897 * which we then process collectively to add pages, and then process again
2898 * to pass to generic_make_request.
2899 *
2900 * The r10_bio structures are linked using a borrowed master_bio pointer.
2901 * This link is counted in ->remaining. When the r10_bio that points to NULL
2902 * has its remaining count decremented to 0, the whole complex operation
2903 * is complete.
2904 *
2905 */
2909 {
2916 sector_t sync_blocks;
2925 /*
2926 * Allow skipping a full rebuild for incremental assembly
2927 * of a clean array, like RAID1 does.
2928 */
2938 }
2940 skipped:
2946 /* If we aborted, we need to abort the
2947 * sync on the 'current' bitmap chucks (there can
2948 * be several when recovering multiple devices).
2949 * as we may have started syncing it but not finished.
2950 * We can find the current address in
2951 * mddev->curr_resync, but for recovery,
2952 * we need to convert that to several
2953 * virtual addresses.
2954 */
2958 }
2965 sector_t sect =
2969 }
2971 /* completed sync */
2975 /* Completed a full sync so the replacements
2976 * are now fully recovered.
2977 */
2981 ->recovery_offset
2983 }
2985 }
2990 }
2996 /* if there has been nothing to do on any drive,
2997 * then there is nothing to do at all..
2998 */
3001 }
3006 /* make sure whole request will fit in a chunk - if chunks
3007 * are meaningful
3008 */
3012 /*
3013 * If there is non-resync activity waiting for us then
3014 * put in a delay to throttle resync.
3015 */
3019 /* Again, very different code for resync and recovery.
3020 * Both must result in an r10bio with a list of bios that
3021 * have bi_end_io, bi_sector, bi_bdev set,
3022 * and bi_private set to the r10bio.
3023 * For recovery, we may actually create several r10bios
3024 * with 2 bios in each, that correspond to the bios in the main one.
3025 * In this case, the subordinate r10bios link back through a
3026 * borrowed master_bio pointer, and the counter in the master
3027 * includes a ref from each subordinate.
3028 */
3029 /* First, we decide what to do and set ->bi_end_io
3030 * To end_sync_read if we want to read, and
3031 * end_sync_write if we will want to write.
3032 */
3036 /* recovery... the complicated one */
3043 sector_t sect;
3050 &&
3057 /* want to reconstruct this device */
3061 /* last stripe is not complete - don't
3062 * try to recover this sector.
3063 */
3065 }
3066 /* Unless we are doing a full sync, or a replacement
3067 * we only need to recover the block if it is set in
3068 * the bitmap
3069 */
3077 /* yep, skip the sync_blocks here, but don't assume
3078 * that there will never be anything to do here
3079 */
3082 }
3097 /* Need to check if the array will still be
3098 * degraded
3099 */
3105 }
3121 /* This is where we read from */
3131 bad_sectors -= (sector
3136 }
3137 }
3150 /* and we write to 'i' (if not in_sync) */
3178 /* and maybe write to replacement */
3183 /* Note: if rdev != NULL, then bio
3184 * cannot be NULL as r10buf_pool_alloc will
3185 * have allocated it.
3186 * So the second test here is pointless.
3187 * But it keeps semantic-checkers happy, and
3188 * this comment keeps human reviewers
3189 * happy.
3190 */
3205 }
3207 /* Cannot recover, so abort the recovery or
3208 * record a bad block */
3210 /* problem is that there are bad blocks
3211 * on other device(s)
3212 */
3230 }
3237 mirror->recovery_disabled
3239 }
3245 }
3246 }
3253 }
3255 }
3257 /* resync. Schedule a read for every block at this virt offset */
3266 /* We can skip this block */
3269 }
3310 }
3311 }
3322 count++;
3329 /* Need to set up for writing to the replacement */
3344 count++;
3345 }
3352 mddev);
3357 mddev);
3358 }
3362 }
3363 }
3381 /* stop here */
3386 /* remove last page from this bio */
3390 }
3392 }
3396 bio_full:
3411 }
3412 }
3415 /* pretend they weren't skipped, it makes
3416 * no important difference in this case
3417 */
3421 giveup:
3422 /* There is nowhere to write, so all non-sync
3423 * drives must be failed or in resync, all drives
3424 * have a bad block, so try the next chunk...
3425 */
3430 chunks_skipped ++;
3433 }
3435 static sector_t
3437 {
3438 sector_t size;
3453 }
3456 {
3457 /* Calculate the number of sectors-per-device that will
3458 * actually be used, and set conf->dev_sectors and
3459 * conf->stride
3460 */
3466 /* 'size' is now the number of chunks in the array */
3467 /* calculate "used chunks per device" */
3470 /* We need to round up when dividing by raid_disks to
3471 * get the stride size.
3472 */
3482 }
3483 }
3487 {
3503 * updated yet. */
3508 }
3525 }
3528 {
3541 }
3547 }
3554 /* FIXME calc properly */
3557 GFP_KERNEL);
3580 }
3584 else
3585 /* far_copies must be 1 */
3587 }
3601 out:
3611 }
3613 }
3616 {
3621 sector_t size;
3631 }
3647 else
3650 }
3672 }
3692 }
3698 else
3701 }
3702 /* need to check that every block has at least one working mirror */
3707 }
3710 /* must ensure that shape change is supported */
3717 }
3723 i++) {
3728 /* The replacement is all we have - use it */
3732 }
3741 }
3743 }
3753 /*
3754 * Ok, everything is just fine now
3755 */
3767 /* Calculate max read-ahead size.
3768 * We need to readahead at least twice a whole stripe....
3769 * maybe...
3770 */
3775 }
3790 /* This cannot work */
3793 }
3803 }
3807 out_free_conf:
3815 out:
3817 }
3820 {
3828 /* the unplug fn references 'conf'*/
3838 }
3841 {
3851 }
3852 }
3855 {
3856 /* Resize of 'far' arrays is not supported.
3857 * For 'near' and 'offset' arrays we can set the
3858 * number of sectors used to be an appropriate multiple
3859 * of the chunk size.
3860 * For 'offset', this is far_copies*chunksize.
3861 * For 'near' the multiplier is the LCM of
3862 * near_copies and raid_disks.
3863 * So if far_copies > 1 && !far_offset, fail.
3864 * Else find LCM(raid_disks, near_copy)*far_copies and
3865 * multiply by chunk_size. Then round to this number.
3866 * This is mostly done by raid10_size()
3867 */
3886 }
3894 }
3899 }
3902 {
3910 }
3912 /* Set new parameters */
3914 /* new layout: far_copies = 1, near_copies = 2 */
3919 /* make sure it will be not marked as dirty */
3928 }
3931 }
3934 {
3937 /* raid10 can take over:
3938 * raid0 - providing it has only two drives
3939 */
3941 /* for raid0 takeover only one zone is supported */
3948 }
3950 }
3952 }
3955 {
3956 /* Called when there is a request to change
3957 * - layout (to ->new_layout)
3958 * - chunk size (to ->new_chunk_sectors)
3959 * - raid_disks (by delta_disks)
3960 * or when trying to restart a reshape that was ongoing.
3961 *
3962 * We need to validate the request and possibly allocate
3963 * space if that might be an issue later.
3964 *
3965 * Currently we reject any reshape of a 'far' mode array,
3966 * allow chunk size to change if new is generally acceptable,
3967 * allow raid_disks to increase, and allow
3968 * a switch between 'near' mode and 'offset' mode.
3969 */
3977 /* mustn't change number of copies */
3980 /* Cannot switch to 'far' mode */
3984 /* not factor of array size */
3993 /* allocate new 'mirrors' list */
3998 GFP_KERNEL);
4001 }
4003 }
4005 /*
4006 * Need to check if array has failed when deciding whether to:
4007 * - start an array
4008 * - remove non-faulty devices
4009 * - add a spare
4010 * - allow a reshape
4011 * This determination is simple when no reshape is happening.
4012 * However if there is a reshape, we need to carefully check
4013 * both the before and after sections.
4014 * This is because some failed devices may only affect one
4015 * of the two sections, and some non-in_sync devices may
4016 * be insync in the section most affected by failed devices.
4017 */
4019 {
4025 /* 'prev' section first */
4029 degraded++;
4031 /* When we can reduce the number of devices in
4032 * an array, this might not contribute to
4033 * 'degraded'. It does now.
4034 */
4035 degraded++;
4036 }
4045 degraded2++;
4047 /* If reshape is increasing the number of devices,
4048 * this section has already been recovered, so
4049 * it doesn't contribute to degraded.
4050 * else it does.
4051 */
4053 degraded2++;
4054 }
4055 }
4060 }
4063 {
4064 /* A 'reshape' has been requested. This commits
4065 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4066 * This also checks if there are enough spares and adds them
4067 * to the array.
4068 * We currently require enough spares to make the final
4069 * array non-degraded. We also require that the difference
4070 * between old and new data_offset - on each device - is
4071 * enough that we never risk over-writing.
4072 */
4097 spares++;
4107 }
4108 }
4126 }
4136 }
4150 }
4159 else
4164 }
4167 /* This is a spare that was manually added */
4169 }
4170 }
4171 /* When a reshape changes the number of devices,
4172 * ->degraded is measured against the larger of the
4173 * pre and post numbers.
4174 */
4192 }
4198 abort:
4210 }
4212 /* Calculate the last device-address that could contain
4213 * any block from the chunk that includes the array-address 's'
4214 * and report the next address.
4215 * i.e. the address returned will be chunk-aligned and after
4216 * any data that is in the chunk containing 's'.
4217 */
4219 {
4227 }
4229 /* Calculate the first device-address that could contain
4230 * any block from the chunk that includes the array-address 's'.
4231 * This too will be the start of a chunk
4232 */
4234 {
4241 }
4245 {
4246 /* We simply copy at most one chunk (smallest of old and new)
4247 * at a time, possibly less if that exceeds RESYNC_PAGES,
4248 * or we hit a bad block or something.
4249 * This might mean we pause for normal IO in the middle of
4250 * a chunk, but that is not a problem was mddev->reshape_position
4251 * can record any location.
4252 *
4253 * If we will want to write to a location that isn't
4254 * yet recorded as 'safe' (i.e. in metadata on disk) then
4255 * we need to flush all reshape requests and update the metadata.
4256 *
4257 * When reshaping forwards (e.g. to more devices), we interpret
4258 * 'safe' as the earliest block which might not have been copied
4259 * down yet. We divide this by previous stripe size and multiply
4260 * by previous stripe length to get lowest device offset that we
4261 * cannot write to yet.
4262 * We interpret 'sector_nr' as an address that we want to write to.
4263 * From this we use last_device_address() to find where we might
4264 * write to, and first_device_address on the 'safe' position.
4265 * If this 'next' write position is after the 'safe' position,
4266 * we must update the metadata to increase the 'safe' position.
4267 *
4268 * When reshaping backwards, we round in the opposite direction
4269 * and perform the reverse test: next write position must not be
4270 * less than current safe position.
4271 *
4272 * In all this the minimum difference in data offsets
4273 * (conf->offset_diff - always positive) allows a bit of slack,
4274 * so next can be after 'safe', but not by more than offset_disk
4275 *
4276 * We need to prepare all the bios here before we start any IO
4277 * to ensure the size we choose is acceptable to all devices.
4278 * The means one for each copy for write-out and an extra one for
4279 * read-in.
4280 * We store the read-in bio in ->master_bio and the others in
4281 * ->devs[x].bio and ->devs[x].repl_bio.
4282 */
4296 /* If restarting in the middle, skip the initial sectors */
4309 }
4310 }
4312 /* We don't use sector_nr to track where we are up to
4313 * as that doesn't work well for ->reshape_backwards.
4314 * So just use ->reshape_progress.
4315 */
4317 /* 'next' is the earliest device address that we might
4318 * write to for this chunk in the new layout
4319 */
4323 /* 'safe' is the last device address that we might read from
4324 * in the old layout after a restart
4325 */
4338 /* 'next' is after the last device address that we
4339 * might write to for this chunk in the new layout
4340 */
4343 /* 'safe' is the earliest device address that we might
4344 * read from in the old layout after a restart
4345 */
4348 /* Need to update metadata if 'next' might be beyond 'safe'
4349 * as that would possibly corrupt data
4350 */
4360 }
4364 /* Need to update reshape_position in metadata */
4370 else
4380 }
4383 }
4385 read_more:
4386 /* Now schedule reads for blocks from sector_nr to last */
4398 /* Cannot read from here, so need to record bad blocks
4399 * on all the target devices.
4400 */
4401 // FIXME
4404 }
4421 /* Now find the locations in the new layout */
4437 }
4450 }
4452 /* Now add as many pages as possible to all of these bios. */
4465 /* Didn't fit, must stop */
4469 /* Remove last page from this bio */
4473 }
4475 }
4478 }
4479 bio_full:
4482 /* Now submit the read */
4492 /* Now that we have done the whole section we can
4493 * update reshape_progress
4494 */
4497 else
4501 }
4507 {
4508 /* Reshape read completed. Hopefully we have a block
4509 * to write out.
4510 * If we got a read error then we do sync 1-page reads from
4511 * elsewhere until we find the data - or give up.
4512 */
4518 /* Reshape has been aborted */
4521 }
4523 /* We definitely have the data in the pages, schedule the
4524 * writes.
4525 */
4537 }
4545 }
4547 }
4550 {
4561 /* read-ahead size must cover two whole stripes, which is
4562 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4563 */
4570 }
4572 }
4577 {
4578 /* Use sync reads to get the blocks from somewhere else */
4604 sector_t addr;
4612 addr,
4618 failed:
4619 slot++;
4624 }
4626 /* couldn't read this block, must give up */
4630 }
4632 idx++;
4633 }
4635 }
4638 {
4654 }
4657 /* FIXME should record badblock */
4659 }
4663 }
4666 {
4672 }
4675 {
4687 }
4695 d++) {
4702 }
4703 }
4709 }
4712 {
4732 };
4735 {
4737 }
4740 {
4742 }