| blob | 33fc408e5eacef0a1dce55fd5c0d578fc244b663 |
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 {
1179 /* IO spans the reshape position. Need to wait for
1180 * reshape to pass
1181 */
1186 sectors);
1188 }
1196 /* Need to update reshape_position in metadata */
1205 }
1216 /* We might need to issue multiple reads to different
1217 * devices if there are bad blocks around, so we keep
1218 * track of the number of reads in bio->bi_phys_segments.
1219 * If this is 0, there is only one r10_bio and no locking
1220 * will be needed when the request completes. If it is
1221 * non-zero, then it is the number of not-completed requests.
1222 */
1227 /*
1228 * read balancing logic:
1229 */
1233 read_again:
1238 }
1243 max_sectors);
1256 /* Could not read all from this device, so we will
1257 * need another r10_bio.
1258 */
1265 else
1268 /* Cannot call generic_make_request directly
1269 * as that will be queued in __generic_make_request
1270 * and subsequent mempool_alloc might block
1271 * waiting for it. so hand bio over to raid10d.
1272 */
1282 sectors_handled;
1287 }
1289 /*
1290 * WRITE:
1291 */
1296 }
1297 /* first select target devices under rcu_lock and
1298 * inc refcount on their rdev. Record them by setting
1299 * bios[x] to bio
1300 * If there are known/acknowledged bad blocks on any device
1301 * on which we have seen a write error, we want to avoid
1302 * writing to those blocks. This potentially requires several
1303 * writes to write around the bad blocks. Each set of writes
1304 * gets its own r10_bio with a set of bios attached. The number
1305 * of r10_bios is recored in bio->bi_phys_segments just as with
1306 * the read case.
1307 */
1311 retry_write:
1327 }
1332 }
1346 }
1348 sector_t first_bad;
1354 max_sectors,
1357 /* Mustn't write here until the bad block
1358 * is acknowledged
1359 */
1364 }
1366 /* Cannot write here at all */
1369 /* Mustn't write more than bad_sectors
1370 * to other devices yet
1371 */
1373 /* We don't set R10BIO_Degraded as that
1374 * only applies if the disk is missing,
1375 * so it might be re-added, and we want to
1376 * know to recover this chunk.
1377 * In this case the device is here, and the
1378 * fact that this chunk is not in-sync is
1379 * recorded in the bad block log.
1380 */
1382 }
1387 }
1388 }
1392 }
1396 }
1397 }
1401 /* Have to wait for this device to get unblocked, then retry */
1409 }
1415 /* Race with remove_disk */
1418 }
1420 }
1421 }
1426 }
1429 /* We are splitting this into multiple parts, so
1430 * we need to prepare for allocating another r10_bio.
1431 */
1436 else
1439 }
1453 max_sectors);
1458 rdev));
1471 cb);
1472 else
1481 }
1485 }
1490 /* Replacement just got moved to main 'rdev' */
1493 }
1496 max_sectors);
1515 }
1516 }
1518 /* Don't remove the bias on 'remaining' (one_write_done) until
1519 * after checking if we need to go around again.
1520 */
1524 /* We need another r10_bio. It has already been counted
1525 * in bio->bi_phys_segments.
1526 */
1536 }
1538 }
1541 {
1551 }
1555 /*
1556 * Register the new request and wait if the reconstruction
1557 * thread has put up a bar for new requests.
1558 * Continue immediately if no resync is active currently.
1559 */
1564 /*
1565 * If this request crosses a chunk boundary, we need to split
1566 * it.
1567 */
1580 }
1585 /* In case raid10d snuck in to freeze_array */
1587 }
1590 {
1601 else
1603 }
1611 }
1613 /* check if there are enough drives for
1614 * every block to appear on atleast one.
1615 * Don't consider the device numbered 'ignore'
1616 * as we might be about to remove it.
1617 */
1619 {
1629 }
1641 cnt++;
1643 }
1649 out:
1652 }
1655 {
1656 /* when calling 'enough', both 'prev' and 'geo' must
1657 * be stable.
1658 * This is ensured if ->reconfig_mutex or ->device_lock
1659 * is held.
1660 */
1663 }
1666 {
1671 /*
1672 * If it is not operational, then we have already marked it as dead
1673 * else if it is the last working disks, ignore the error, let the
1674 * next level up know.
1675 * else mark the drive as failed
1676 */
1680 /*
1681 * Don't fail the drive, just return an IO error.
1682 */
1685 }
1688 /*
1689 * if recovery is running, make sure it aborts.
1690 */
1692 }
1702 }
1705 {
1713 }
1725 }
1726 }
1729 {
1735 }
1738 {
1745 /*
1746 * Find all non-in_sync disks within the RAID10 configuration
1747 * and mark them in_sync
1748 */
1755 /* Replacement has just become active */
1758 count++;
1760 /* Replaced device not technically faulty,
1761 * but we need to be sure it gets removed
1762 * and never re-added.
1763 */
1767 }
1773 count++;
1775 }
1776 }
1783 }
1787 {
1796 /* only hot-add to in-sync arrays, as recovery is
1797 * very different from resync
1798 */
1809 }
1814 else
1834 }
1848 }
1850 /* Some requests might not have seen this new
1851 * merge_bvec_fn. We must wait for them to complete
1852 * before merging the device fully.
1853 * First we make sure any code which has tested
1854 * our function has submitted the request, then
1855 * we wait for all outstanding requests to complete.
1856 */
1861 }
1868 }
1871 {
1883 else
1890 }
1891 /* Only remove faulty devices if recovery
1892 * is not possible.
1893 */
1901 }
1905 /* lost the race, try later */
1910 /* We must have just cleared 'rdev' */
1914 * but will never see neither -- if they are careful.
1915 */
1919 /* We might have just remove the Replacement as faulty
1920 * Clear the flag just in case
1921 */
1926 abort:
1930 }
1934 {
1940 /* this is a reshape read */
1947 else
1948 /* The write handler will notice the lack of
1949 * R10BIO_Uptodate and record any errors etc
1950 */
1954 /* for reconstruct, we always reschedule after a read.
1955 * for resync, only after all reads
1956 */
1960 /* we have read all the blocks,
1961 * do the comparison in process context in raid10d
1962 */
1964 }
1965 }
1968 {
1973 /* the primary of several recovery bios */
1978 else
1987 else
1990 }
1991 }
1992 }
1995 {
2001 sector_t first_bad;
2010 else
2022 }
2032 }
2034 /*
2035 * Note: sync and recover and handled very differently for raid10
2036 * This code is for resync.
2037 * For resync, we read through virtual addresses and read all blocks.
2038 * If there is any error, we schedule a write. The lowest numbered
2039 * drive is authoritative.
2040 * However requests come for physical address, so we need to map.
2041 * For every physical address there are raid_disks/copies virtual addresses,
2042 * which is always are least one, but is not necessarly an integer.
2043 * This means that a physical address can span multiple chunks, so we may
2044 * have to submit multiple io requests for a single sync request.
2045 */
2046 /*
2047 * We check if all blocks are in-sync and only write to blocks that
2048 * aren't in sync
2049 */
2051 {
2059 /* find the first device with a block */
2071 /* now find blocks with errors */
2082 /* We know that the bi_io_vec layout is the same for
2083 * both 'first' and 'i', so we just compare them.
2084 * All vec entries are PAGE_SIZE;
2085 */
2093 len))
2096 }
2101 /* Don't fix anything. */
2103 }
2104 /* Ok, we need to write this bio, either to correct an
2105 * inconsistency or to correct an unreadable block.
2106 * First we need to fixup bv_offset, bv_len and
2107 * bi_vecs, as the read request might have corrupted these
2108 */
2123 PAGE_SIZE);
2124 }
2135 }
2137 /* Now write out to any replacement devices
2138 * that are active
2139 */
2151 PAGE_SIZE);
2157 }
2159 done:
2163 }
2164 }
2166 /*
2167 * Now for the recovery code.
2168 * Recovery happens across physical sectors.
2169 * We recover all non-is_sync drives by finding the virtual address of
2170 * each, and then choose a working drive that also has that virt address.
2171 * There is a separate r10_bio for each non-in_sync drive.
2172 * Only the first two slots are in use. The first for reading,
2173 * The second for writing.
2174 *
2175 */
2177 {
2178 /* We got a read error during recovery.
2179 * We repeat the read in smaller page-sized sections.
2180 * If a read succeeds, write it to the new device or record
2181 * a bad block if we cannot.
2182 * If a read fails, record a bad block on both old and
2183 * new devices.
2184 */
2197 sector_t addr;
2206 addr,
2214 addr,
2224 }
2225 }
2227 /* We don't worry if we cannot set a bad block -
2228 * it really is bad so there is no loss in not
2229 * recording it yet
2230 */
2234 /* need bad block on destination too */
2239 /* just abort the recovery */
2241 "md/raid10:%s: recovery aborted"
2250 }
2251 }
2252 }
2256 idx++;
2257 }
2258 }
2261 {
2270 }
2272 /*
2273 * share the pages with the first bio
2274 * and submit the write request
2275 */
2279 /* Need to test wbio2->bi_end_io before we call
2280 * generic_make_request as if the former is NULL,
2281 * the latter is free to free wbio2.
2282 */
2289 }
2295 }
2296 }
2299 /*
2300 * Used by fix_read_error() to decay the per rdev read_errors.
2301 * We halve the read error count for every hour that has elapsed
2302 * since the last recorded read error.
2303 *
2304 */
2306 {
2315 /* first time we've seen a read error */
2318 }
2325 /*
2326 * if hours_since_last is > the number of bits in read_errors
2327 * just set read errors to 0. We do this to avoid
2328 * overflowing the shift of read_errors by hours_since_last.
2329 */
2332 else
2334 }
2338 {
2339 sector_t first_bad;
2346 /* success */
2353 }
2354 /* need to record an error - either for the block or the device */
2358 }
2360 /*
2361 * This is a kernel thread which:
2362 *
2363 * 1. Retries failed read operations on working mirrors.
2364 * 2. Updates the raid superblock when problems encounter.
2365 * 3. Performs writes following reads for array synchronising.
2366 */
2369 {
2376 /* still own a reference to this rdev, so it cannot
2377 * have been cleared recently.
2378 */
2382 /* drive has already been failed, just ignore any
2383 more fix_read_error() attempts */
2393 "md/raid10:%s: %s: Raid device exceeded "
2403 }
2416 sector_t first_bad;
2430 sect,
2437 }
2438 sl++;
2445 /* Cannot read from anywhere, just mark the block
2446 * as bad on the first device to discourage future
2447 * reads.
2448 */
2453 rdev,
2460 }
2462 }
2465 /* write it back and re-read */
2472 sl--;
2484 sect,
2487 /* Well, this device is dead */
2489 "md/raid10:%s: read correction "
2490 "write failed"
2494 sect +
2496 rdev)),
2502 }
2505 }
2512 sl--;
2523 sect,
2525 READ)) {
2527 /* Well, this device is dead */
2529 "md/raid10:%s: unable to read back "
2530 "corrected sectors"
2534 sect +
2544 "md/raid10:%s: read error corrected"
2548 sect +
2552 }
2556 }
2561 }
2562 }
2565 {
2570 /* bio has the data to be written to slot 'i' where
2571 * we just recently had a write error.
2572 * We repeatedly clone the bio and trim down to one block,
2573 * then try the write. Where the write fails we record
2574 * a bad block.
2575 * It is conceivable that the bio doesn't exactly align with
2576 * blocks. We must handle this.
2577 *
2578 * We currently own a reference to the rdev.
2579 */
2582 sector_t sector;
2600 /* Write at 'sector' for 'sectors' */
2608 /* Failure! */
2617 }
2619 }
2622 {
2631 /* we got a read error. Maybe the drive is bad. Maybe just
2632 * the block and we can fix it.
2633 * We freeze all other IO, and try reading the block from
2634 * other devices. When we find one, we re-write
2635 * and check it that fixes the read error.
2636 * This is all done synchronously while the array is
2637 * frozen.
2638 */
2653 read_more:
2662 }
2667 KERN_ERR
2668 "md/raid10:%s: %s: redirecting "
2685 /* Drat - have to split this up more */
2694 else
2700 GFP_NOIO);
2713 }
2716 {
2717 /* Some sort of write request has finished and it
2718 * succeeded in writing where we thought there was a
2719 * bad block. So forget the bad block.
2720 * Or possibly if failed and we need to record
2721 * a bad block.
2722 */
2736 rdev,
2741 rdev,
2745 }
2752 rdev,
2757 rdev,
2761 }
2762 }
2771 rdev,
2781 }
2783 }
2788 rdev,
2792 }
2793 }
2798 }
2799 }
2802 {
2821 }
2841 /* just a partial read to be scheduled from a
2842 * separate context
2843 */
2846 }
2851 }
2853 }
2857 {
2872 }
2874 /*
2875 * perform a "sync" on one "block"
2876 *
2877 * We need to make sure that no normal I/O request - particularly write
2878 * requests - conflict with active sync requests.
2879 *
2880 * This is achieved by tracking pending requests and a 'barrier' concept
2881 * that can be installed to exclude normal IO requests.
2882 *
2883 * Resync and recovery are handled very differently.
2884 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2885 *
2886 * For resync, we iterate over virtual addresses, read all copies,
2887 * and update if there are differences. If only one copy is live,
2888 * skip it.
2889 * For recovery, we iterate over physical addresses, read a good
2890 * value for each non-in_sync drive, and over-write.
2891 *
2892 * So, for recovery we may have several outstanding complex requests for a
2893 * given address, one for each out-of-sync device. We model this by allocating
2894 * a number of r10_bio structures, one for each out-of-sync device.
2895 * As we setup these structures, we collect all bio's together into a list
2896 * which we then process collectively to add pages, and then process again
2897 * to pass to generic_make_request.
2898 *
2899 * The r10_bio structures are linked using a borrowed master_bio pointer.
2900 * This link is counted in ->remaining. When the r10_bio that points to NULL
2901 * has its remaining count decremented to 0, the whole complex operation
2902 * is complete.
2903 *
2904 */
2908 {
2915 sector_t sync_blocks;
2924 /*
2925 * Allow skipping a full rebuild for incremental assembly
2926 * of a clean array, like RAID1 does.
2927 */
2937 }
2939 skipped:
2945 /* If we aborted, we need to abort the
2946 * sync on the 'current' bitmap chucks (there can
2947 * be several when recovering multiple devices).
2948 * as we may have started syncing it but not finished.
2949 * We can find the current address in
2950 * mddev->curr_resync, but for recovery,
2951 * we need to convert that to several
2952 * virtual addresses.
2953 */
2957 }
2964 sector_t sect =
2968 }
2970 /* completed sync */
2974 /* Completed a full sync so the replacements
2975 * are now fully recovered.
2976 */
2980 ->recovery_offset
2982 }
2984 }
2989 }
2995 /* if there has been nothing to do on any drive,
2996 * then there is nothing to do at all..
2997 */
3000 }
3005 /* make sure whole request will fit in a chunk - if chunks
3006 * are meaningful
3007 */
3011 /*
3012 * If there is non-resync activity waiting for us then
3013 * put in a delay to throttle resync.
3014 */
3018 /* Again, very different code for resync and recovery.
3019 * Both must result in an r10bio with a list of bios that
3020 * have bi_end_io, bi_sector, bi_bdev set,
3021 * and bi_private set to the r10bio.
3022 * For recovery, we may actually create several r10bios
3023 * with 2 bios in each, that correspond to the bios in the main one.
3024 * In this case, the subordinate r10bios link back through a
3025 * borrowed master_bio pointer, and the counter in the master
3026 * includes a ref from each subordinate.
3027 */
3028 /* First, we decide what to do and set ->bi_end_io
3029 * To end_sync_read if we want to read, and
3030 * end_sync_write if we will want to write.
3031 */
3035 /* recovery... the complicated one */
3042 sector_t sect;
3049 &&
3056 /* want to reconstruct this device */
3060 /* last stripe is not complete - don't
3061 * try to recover this sector.
3062 */
3064 }
3065 /* Unless we are doing a full sync, or a replacement
3066 * we only need to recover the block if it is set in
3067 * the bitmap
3068 */
3076 /* yep, skip the sync_blocks here, but don't assume
3077 * that there will never be anything to do here
3078 */
3081 }
3096 /* Need to check if the array will still be
3097 * degraded
3098 */
3104 }
3120 /* This is where we read from */
3130 bad_sectors -= (sector
3135 }
3136 }
3149 /* and we write to 'i' (if not in_sync) */
3177 /* and maybe write to replacement */
3182 /* Note: if rdev != NULL, then bio
3183 * cannot be NULL as r10buf_pool_alloc will
3184 * have allocated it.
3185 * So the second test here is pointless.
3186 * But it keeps semantic-checkers happy, and
3187 * this comment keeps human reviewers
3188 * happy.
3189 */
3204 }
3206 /* Cannot recover, so abort the recovery or
3207 * record a bad block */
3209 /* problem is that there are bad blocks
3210 * on other device(s)
3211 */
3229 }
3236 mirror->recovery_disabled
3238 }
3244 }
3245 }
3252 }
3254 }
3256 /* resync. Schedule a read for every block at this virt offset */
3265 /* We can skip this block */
3268 }
3309 }
3310 }
3321 count++;
3328 /* Need to set up for writing to the replacement */
3343 count++;
3344 }
3351 mddev);
3356 mddev);
3357 }
3361 }
3362 }
3380 /* stop here */
3385 /* remove last page from this bio */
3389 }
3391 }
3395 bio_full:
3410 }
3411 }
3414 /* pretend they weren't skipped, it makes
3415 * no important difference in this case
3416 */
3420 giveup:
3421 /* There is nowhere to write, so all non-sync
3422 * drives must be failed or in resync, all drives
3423 * have a bad block, so try the next chunk...
3424 */
3429 chunks_skipped ++;
3432 }
3434 static sector_t
3436 {
3437 sector_t size;
3452 }
3455 {
3456 /* Calculate the number of sectors-per-device that will
3457 * actually be used, and set conf->dev_sectors and
3458 * conf->stride
3459 */
3465 /* 'size' is now the number of chunks in the array */
3466 /* calculate "used chunks per device" */
3469 /* We need to round up when dividing by raid_disks to
3470 * get the stride size.
3471 */
3481 }
3482 }
3486 {
3502 * updated yet. */
3507 }
3524 }
3527 {
3540 }
3546 }
3553 /* FIXME calc properly */
3556 GFP_KERNEL);
3579 }
3583 else
3584 /* far_copies must be 1 */
3586 }
3600 out:
3610 }
3612 }
3615 {
3620 sector_t size;
3630 }
3646 else
3649 }
3671 }
3691 }
3697 else
3700 }
3701 /* need to check that every block has at least one working mirror */
3706 }
3709 /* must ensure that shape change is supported */
3716 }
3722 i++) {
3727 /* The replacement is all we have - use it */
3731 }
3740 }
3742 }
3752 /*
3753 * Ok, everything is just fine now
3754 */
3766 /* Calculate max read-ahead size.
3767 * We need to readahead at least twice a whole stripe....
3768 * maybe...
3769 */
3774 }
3789 /* This cannot work */
3792 }
3802 }
3806 out_free_conf:
3814 out:
3816 }
3819 {
3827 /* the unplug fn references 'conf'*/
3837 }
3840 {
3850 }
3851 }
3854 {
3855 /* Resize of 'far' arrays is not supported.
3856 * For 'near' and 'offset' arrays we can set the
3857 * number of sectors used to be an appropriate multiple
3858 * of the chunk size.
3859 * For 'offset', this is far_copies*chunksize.
3860 * For 'near' the multiplier is the LCM of
3861 * near_copies and raid_disks.
3862 * So if far_copies > 1 && !far_offset, fail.
3863 * Else find LCM(raid_disks, near_copy)*far_copies and
3864 * multiply by chunk_size. Then round to this number.
3865 * This is mostly done by raid10_size()
3866 */
3885 }
3893 }
3898 }
3901 {
3909 }
3911 /* Set new parameters */
3913 /* new layout: far_copies = 1, near_copies = 2 */
3918 /* make sure it will be not marked as dirty */
3927 }
3930 }
3933 {
3936 /* raid10 can take over:
3937 * raid0 - providing it has only two drives
3938 */
3940 /* for raid0 takeover only one zone is supported */
3947 }
3949 }
3951 }
3954 {
3955 /* Called when there is a request to change
3956 * - layout (to ->new_layout)
3957 * - chunk size (to ->new_chunk_sectors)
3958 * - raid_disks (by delta_disks)
3959 * or when trying to restart a reshape that was ongoing.
3960 *
3961 * We need to validate the request and possibly allocate
3962 * space if that might be an issue later.
3963 *
3964 * Currently we reject any reshape of a 'far' mode array,
3965 * allow chunk size to change if new is generally acceptable,
3966 * allow raid_disks to increase, and allow
3967 * a switch between 'near' mode and 'offset' mode.
3968 */
3976 /* mustn't change number of copies */
3979 /* Cannot switch to 'far' mode */
3983 /* not factor of array size */
3992 /* allocate new 'mirrors' list */
3997 GFP_KERNEL);
4000 }
4002 }
4004 /*
4005 * Need to check if array has failed when deciding whether to:
4006 * - start an array
4007 * - remove non-faulty devices
4008 * - add a spare
4009 * - allow a reshape
4010 * This determination is simple when no reshape is happening.
4011 * However if there is a reshape, we need to carefully check
4012 * both the before and after sections.
4013 * This is because some failed devices may only affect one
4014 * of the two sections, and some non-in_sync devices may
4015 * be insync in the section most affected by failed devices.
4016 */
4018 {
4024 /* 'prev' section first */
4028 degraded++;
4030 /* When we can reduce the number of devices in
4031 * an array, this might not contribute to
4032 * 'degraded'. It does now.
4033 */
4034 degraded++;
4035 }
4044 degraded2++;
4046 /* If reshape is increasing the number of devices,
4047 * this section has already been recovered, so
4048 * it doesn't contribute to degraded.
4049 * else it does.
4050 */
4052 degraded2++;
4053 }
4054 }
4059 }
4062 {
4063 /* A 'reshape' has been requested. This commits
4064 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4065 * This also checks if there are enough spares and adds them
4066 * to the array.
4067 * We currently require enough spares to make the final
4068 * array non-degraded. We also require that the difference
4069 * between old and new data_offset - on each device - is
4070 * enough that we never risk over-writing.
4071 */
4096 spares++;
4106 }
4107 }
4125 }
4135 }
4149 }
4158 else
4163 }
4166 /* This is a spare that was manually added */
4168 }
4169 }
4170 /* When a reshape changes the number of devices,
4171 * ->degraded is measured against the larger of the
4172 * pre and post numbers.
4173 */
4191 }
4197 abort:
4209 }
4211 /* Calculate the last device-address that could contain
4212 * any block from the chunk that includes the array-address 's'
4213 * and report the next address.
4214 * i.e. the address returned will be chunk-aligned and after
4215 * any data that is in the chunk containing 's'.
4216 */
4218 {
4226 }
4228 /* Calculate the first device-address that could contain
4229 * any block from the chunk that includes the array-address 's'.
4230 * This too will be the start of a chunk
4231 */
4233 {
4240 }
4244 {
4245 /* We simply copy at most one chunk (smallest of old and new)
4246 * at a time, possibly less if that exceeds RESYNC_PAGES,
4247 * or we hit a bad block or something.
4248 * This might mean we pause for normal IO in the middle of
4249 * a chunk, but that is not a problem was mddev->reshape_position
4250 * can record any location.
4251 *
4252 * If we will want to write to a location that isn't
4253 * yet recorded as 'safe' (i.e. in metadata on disk) then
4254 * we need to flush all reshape requests and update the metadata.
4255 *
4256 * When reshaping forwards (e.g. to more devices), we interpret
4257 * 'safe' as the earliest block which might not have been copied
4258 * down yet. We divide this by previous stripe size and multiply
4259 * by previous stripe length to get lowest device offset that we
4260 * cannot write to yet.
4261 * We interpret 'sector_nr' as an address that we want to write to.
4262 * From this we use last_device_address() to find where we might
4263 * write to, and first_device_address on the 'safe' position.
4264 * If this 'next' write position is after the 'safe' position,
4265 * we must update the metadata to increase the 'safe' position.
4266 *
4267 * When reshaping backwards, we round in the opposite direction
4268 * and perform the reverse test: next write position must not be
4269 * less than current safe position.
4270 *
4271 * In all this the minimum difference in data offsets
4272 * (conf->offset_diff - always positive) allows a bit of slack,
4273 * so next can be after 'safe', but not by more than offset_disk
4274 *
4275 * We need to prepare all the bios here before we start any IO
4276 * to ensure the size we choose is acceptable to all devices.
4277 * The means one for each copy for write-out and an extra one for
4278 * read-in.
4279 * We store the read-in bio in ->master_bio and the others in
4280 * ->devs[x].bio and ->devs[x].repl_bio.
4281 */
4295 /* If restarting in the middle, skip the initial sectors */
4308 }
4309 }
4311 /* We don't use sector_nr to track where we are up to
4312 * as that doesn't work well for ->reshape_backwards.
4313 * So just use ->reshape_progress.
4314 */
4316 /* 'next' is the earliest device address that we might
4317 * write to for this chunk in the new layout
4318 */
4322 /* 'safe' is the last device address that we might read from
4323 * in the old layout after a restart
4324 */
4337 /* 'next' is after the last device address that we
4338 * might write to for this chunk in the new layout
4339 */
4342 /* 'safe' is the earliest device address that we might
4343 * read from in the old layout after a restart
4344 */
4347 /* Need to update metadata if 'next' might be beyond 'safe'
4348 * as that would possibly corrupt data
4349 */
4359 }
4363 /* Need to update reshape_position in metadata */
4369 else
4379 }
4382 }
4384 read_more:
4385 /* Now schedule reads for blocks from sector_nr to last */
4397 /* Cannot read from here, so need to record bad blocks
4398 * on all the target devices.
4399 */
4400 // FIXME
4403 }
4420 /* Now find the locations in the new layout */
4436 }
4449 }
4451 /* Now add as many pages as possible to all of these bios. */
4464 /* Didn't fit, must stop */
4468 /* Remove last page from this bio */
4472 }
4474 }
4477 }
4478 bio_full:
4481 /* Now submit the read */
4491 /* Now that we have done the whole section we can
4492 * update reshape_progress
4493 */
4496 else
4500 }
4506 {
4507 /* Reshape read completed. Hopefully we have a block
4508 * to write out.
4509 * If we got a read error then we do sync 1-page reads from
4510 * elsewhere until we find the data - or give up.
4511 */
4517 /* Reshape has been aborted */
4520 }
4522 /* We definitely have the data in the pages, schedule the
4523 * writes.
4524 */
4536 }
4544 }
4546 }
4549 {
4560 /* read-ahead size must cover two whole stripes, which is
4561 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4562 */
4569 }
4571 }
4576 {
4577 /* Use sync reads to get the blocks from somewhere else */
4603 sector_t addr;
4611 addr,
4617 failed:
4618 slot++;
4623 }
4625 /* couldn't read this block, must give up */
4629 }
4631 idx++;
4632 }
4634 }
4637 {
4653 }
4656 /* FIXME should record badblock */
4658 }
4662 }
4665 {
4671 }
4674 {
4686 }
4694 d++) {
4701 }
4702 }
4708 }
4711 {
4731 };
4734 {
4736 }
4739 {
4741 }