| blob | ab5e86209322fc0d0ed3e2db18a468c6f5054f13 |
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>
28 #include <trace/events/block.h>
34 /*
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
37 * chunk_size
38 * raid_disks
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
44 *
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
53 *
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
57 *
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
65 * on a device):
66 * A B C D A B C D E
67 * ... ...
68 * D A B C E A B C D
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
73 */
75 /*
76 * Number of guaranteed r10bios in case of extreme VM load:
77 */
78 #define NR_RAID10_BIOS 256
80 /* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84 */
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
89 */
90 #define IO_MADE_GOOD ((struct bio *)2)
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
94 /* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
96 * for writeback.
97 */
110 #define raid10_log(md, fmt, args...) \
114 {
118 /* allocate a r10bio with room for raid_disks entries in the
119 * bios array */
121 }
124 {
126 }
128 /* Maximum size of each resync request */
129 #define RESYNC_BLOCK_SIZE (64*1024)
130 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
131 /* amount of memory to reserve for resync requests */
132 #define RESYNC_WINDOW (1024*1024)
133 /* maximum number of concurrent requests, memory permitting */
134 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
136 /*
137 * When performing a resync, we need to read and compare, so
138 * we need as many pages are there are copies.
139 * When performing a recovery, we need 2 bios, one for read,
140 * one for write (we recover only one drive per r10buf)
141 *
142 */
144 {
159 else
162 /*
163 * Allocate bios.
164 */
176 }
177 /*
178 * Allocate RESYNC_PAGES data pages and attach them
179 * where needed.
180 */
187 /* we can share bv_page's during recovery
188 * and reshape */
200 }
201 }
205 out_free_pages:
212 out_free_bio:
218 }
221 }
224 {
236 }
238 }
242 }
244 }
247 {
259 }
260 }
263 {
268 }
271 {
277 }
280 {
290 /* wake up frozen array... */
294 }
296 /*
297 * raid_end_bio_io() is called when we have finished servicing a mirrored
298 * operation and are ready to return a success/failure code to the buffer
299 * cache layer.
300 */
302 {
319 /*
320 * Wake up any possible resync thread that waits for the device
321 * to go idle.
322 */
324 }
326 }
328 /*
329 * Update disk head position estimator based on IRQ completion info.
330 */
332 {
337 }
339 /*
340 * Find the disk number which triggered given bio
341 */
344 {
354 }
355 }
365 }
368 {
378 /*
379 * this branch is our 'one mirror IO has finished' event handler:
380 */
384 /*
385 * Set R10BIO_Uptodate in our master bio, so that
386 * we will return a good error code to the higher
387 * levels even if IO on some other mirrored buffer fails.
388 *
389 * The 'master' represents the composite IO operation to
390 * user-side. So if something waits for IO, then it will
391 * wait for the 'master' bio.
392 */
395 /* If all other devices that store this block have
396 * failed, we want to return the error upwards rather
397 * than fail the last device. Here we redefine
398 * "uptodate" to mean "Don't want to retry"
399 */
403 }
408 /*
409 * oops, read error - keep the refcount on the rdev
410 */
418 }
419 }
422 {
423 /* clear the bitmap if all writes complete successfully */
429 }
432 {
440 else
442 }
443 }
444 }
447 {
467 }
468 /*
469 * this branch is our 'one mirror IO has finished' event handler:
470 */
473 /* Never record new bad blocks to replacement,
474 * just fail it.
475 */
488 /* This is the only remaining device,
489 * We need to retry the write without
490 * FailFast
491 */
497 }
500 }
502 /*
503 * Set R10BIO_Uptodate in our master bio, so that
504 * we will return a good error code for to the higher
505 * levels even if IO on some other mirrored buffer fails.
506 *
507 * The 'master' represents the composite IO operation to
508 * user-side. So if something waits for IO, then it will
509 * wait for the 'master' bio.
510 */
511 sector_t first_bad;
514 /*
515 * Do not set R10BIO_Uptodate if the current device is
516 * rebuilding or Faulty. This is because we cannot use
517 * such device for properly reading the data back (we could
518 * potentially use it, if the current write would have felt
519 * before rdev->recovery_offset, but for simplicity we don't
520 * check this here.
521 */
526 /* Maybe we can clear some bad blocks. */
534 else
538 }
539 }
541 /*
542 *
543 * Let's see if all mirrored write operations have finished
544 * already.
545 */
551 }
553 /*
554 * RAID10 layout manager
555 * As well as the chunksize and raid_disks count, there are two
556 * parameters: near_copies and far_copies.
557 * near_copies * far_copies must be <= raid_disks.
558 * Normally one of these will be 1.
559 * If both are 1, we get raid0.
560 * If near_copies == raid_disks, we get raid1.
561 *
562 * Chunks are laid out in raid0 style with near_copies copies of the
563 * first chunk, followed by near_copies copies of the next chunk and
564 * so on.
565 * If far_copies > 1, then after 1/far_copies of the array has been assigned
566 * as described above, we start again with a device offset of near_copies.
567 * So we effectively have another copy of the whole array further down all
568 * the drives, but with blocks on different drives.
569 * With this layout, and block is never stored twice on the one device.
570 *
571 * raid10_find_phys finds the sector offset of a given virtual sector
572 * on each device that it is on.
573 *
574 * raid10_find_virt does the reverse mapping, from a device and a
575 * sector offset to a virtual address
576 */
579 {
581 sector_t sector;
582 sector_t chunk;
583 sector_t stripe;
594 /* now calculate first sector/dev */
606 /* and calculate all the others */
613 slot++;
627 }
631 slot++;
632 }
633 dev++;
637 }
638 }
639 }
642 {
654 }
657 {
659 /* Never use conf->prev as this is only called during resync
660 * or recovery, so reshape isn't happening
661 */
675 }
676 }
691 else
693 }
695 }
699 }
701 /*
702 * This routine returns the disk from which the requested read should
703 * be done. There is a per-array 'next expected sequential IO' sector
704 * number - if this matches on the next IO then we use the last disk.
705 * There is also a per-disk 'last know head position' sector that is
706 * maintained from IRQ contexts, both the normal and the resync IO
707 * completion handlers update this position correctly. If there is no
708 * perfect sequential match then we pick the disk whose head is closest.
709 *
710 * If there are 2 mirrors in the same 2 devices, performance degrades
711 * because position is mirror, not device based.
712 *
713 * The rdev for the device selected will have nr_pending incremented.
714 */
716 /*
717 * FIXME: possibly should rethink readbalancing and do it differently
718 * depending on near_copies / far_copies geometry.
719 */
723 {
743 /*
744 * Check if we can balance. We can balance on the whole
745 * device if no resync is going on (recovery is ok), or below
746 * the resync window. We take the first readable disk when
747 * above the resync window.
748 */
754 sector_t first_bad;
756 sector_t dev_sector;
776 /* Already have a better slot */
779 /* Cannot read here. If this is the
780 * 'primary' device, then we must not read
781 * beyond 'bad_sectors' from another device.
782 */
789 sector_t good_sectors =
795 }
797 /* Must read from here */
799 }
808 /* At least 2 disks to choose from so failfast is OK */
810 /* This optimisation is debatable, and completely destroys
811 * sequential read speed for 'far copies' arrays. So only
812 * keep it for 'near' arrays, and review those later.
813 */
817 /* for far > 1 always use the lowest address */
820 else
827 }
828 }
832 }
843 }
846 {
858 i++) {
864 }
865 }
868 }
871 {
872 /* Any writes that have been queued but are awaiting
873 * bitmap updates get flushed here.
874 */
882 /* flush any pending bitmap writes to disk
883 * before proceeding w/ I/O */
897 /* Just ignore it */
899 else
902 }
905 }
907 /* Barriers....
908 * Sometimes we need to suspend IO while we do something else,
909 * either some resync/recovery, or reconfigure the array.
910 * To do this we raise a 'barrier'.
911 * The 'barrier' is a counter that can be raised multiple times
912 * to count how many activities are happening which preclude
913 * normal IO.
914 * We can only raise the barrier if there is no pending IO.
915 * i.e. if nr_pending == 0.
916 * We choose only to raise the barrier if no-one is waiting for the
917 * barrier to go down. This means that as soon as an IO request
918 * is ready, no other operations which require a barrier will start
919 * until the IO request has had a chance.
920 *
921 * So: regular IO calls 'wait_barrier'. When that returns there
922 * is no backgroup IO happening, It must arrange to call
923 * allow_barrier when it has finished its IO.
924 * backgroup IO calls must call raise_barrier. Once that returns
925 * there is no normal IO happeing. It must arrange to call
926 * lower_barrier when the particular background IO completes.
927 */
930 {
934 /* Wait until no block IO is waiting (unless 'force') */
938 /* block any new IO from starting */
941 /* Now wait for all pending IO to complete */
947 }
950 {
956 }
959 {
963 /* Wait for the barrier to drop.
964 * However if there are already pending
965 * requests (preventing the barrier from
966 * rising completely), and the
967 * pre-process bio queue isn't empty,
968 * then don't wait, as we need to empty
969 * that queue to get the nr_pending
970 * count down.
971 */
982 }
985 }
988 {
992 }
995 {
996 /* stop syncio and normal IO and wait for everything to
997 * go quiet.
998 * We increment barrier and nr_waiting, and then
999 * wait until nr_pending match nr_queued+extra
1000 * This is called in the context of one normal IO request
1001 * that has failed. Thus any sync request that might be pending
1002 * will be blocked by nr_pending, and we need to wait for
1003 * pending IO requests to complete or be queued for re-try.
1004 * Thus the number queued (nr_queued) plus this request (extra)
1005 * must match the number of pending IOs (nr_pending) before
1006 * we continue.
1007 */
1019 }
1022 {
1023 /* reverse the effect of the freeze */
1029 }
1033 {
1037 else
1039 }
1045 };
1048 {
1050 cb);
1064 }
1066 /* we aren't scheduling, so we can do the write-out directly. */
1081 /* Just ignore it */
1083 else
1086 }
1088 }
1091 {
1110 /*
1111 * Register the new request and wait if the reconstruction
1112 * thread has put up a bar for new requests.
1113 * Continue immediately if no resync is active currently.
1114 */
1121 /* IO spans the reshape position. Need to wait for
1122 * reshape to pass
1123 */
1129 sectors);
1131 }
1139 /* Need to update reshape_position in metadata */
1149 }
1160 /* We might need to issue multiple reads to different
1161 * devices if there are bad blocks around, so we keep
1162 * track of the number of reads in bio->bi_phys_segments.
1163 * If this is 0, there is only one r10_bio and no locking
1164 * will be needed when the request completes. If it is
1165 * non-zero, then it is the number of not-completed requests.
1166 */
1171 /*
1172 * read balancing logic:
1173 */
1177 read_again:
1182 }
1187 max_sectors);
1207 /* Could not read all from this device, so we will
1208 * need another r10_bio.
1209 */
1216 else
1219 /* Cannot call generic_make_request directly
1220 * as that will be queued in __generic_make_request
1221 * and subsequent mempool_alloc might block
1222 * waiting for it. so hand bio over to raid10d.
1223 */
1233 sectors_handled;
1238 }
1240 /*
1241 * WRITE:
1242 */
1248 }
1249 /* first select target devices under rcu_lock and
1250 * inc refcount on their rdev. Record them by setting
1251 * bios[x] to bio
1252 * If there are known/acknowledged bad blocks on any device
1253 * on which we have seen a write error, we want to avoid
1254 * writing to those blocks. This potentially requires several
1255 * writes to write around the bad blocks. Each set of writes
1256 * gets its own r10_bio with a set of bios attached. The number
1257 * of r10_bios is recored in bio->bi_phys_segments just as with
1258 * the read case.
1259 */
1263 retry_write:
1279 }
1284 }
1296 }
1298 sector_t first_bad;
1304 max_sectors,
1307 /* Mustn't write here until the bad block
1308 * is acknowledged
1309 */
1314 }
1316 /* Cannot write here at all */
1319 /* Mustn't write more than bad_sectors
1320 * to other devices yet
1321 */
1323 /* We don't set R10BIO_Degraded as that
1324 * only applies if the disk is missing,
1325 * so it might be re-added, and we want to
1326 * know to recover this chunk.
1327 * In this case the device is here, and the
1328 * fact that this chunk is not in-sync is
1329 * recorded in the bad block log.
1330 */
1332 }
1337 }
1338 }
1342 }
1346 }
1347 }
1351 /* Have to wait for this device to get unblocked, then retry */
1359 }
1365 /* Race with remove_disk */
1368 }
1370 }
1371 }
1377 }
1380 /* We are splitting this into multiple parts, so
1381 * we need to prepare for allocating another r10_bio.
1382 */
1387 else
1390 }
1404 max_sectors);
1409 rdev));
1422 /* flush_pending_writes() needs access to the rdev so...*/
1431 cb);
1432 else
1441 }
1445 }
1450 /* Replacement just got moved to main 'rdev' */
1453 }
1456 max_sectors);
1471 /* flush_pending_writes() needs access to the rdev so...*/
1481 }
1482 }
1484 /* Don't remove the bias on 'remaining' (one_write_done) until
1485 * after checking if we need to go around again.
1486 */
1490 /* We need another r10_bio. It has already been counted
1491 * in bio->bi_phys_segments.
1492 */
1502 }
1504 }
1507 {
1517 }
1521 /*
1522 * If this request crosses a chunk boundary, we need to split
1523 * it.
1524 */
1537 }
1542 /* In case raid10d snuck in to freeze_array */
1544 }
1547 {
1558 else
1562 }
1569 }
1572 }
1574 /* check if there are enough drives for
1575 * every block to appear on atleast one.
1576 * Don't consider the device numbered 'ignore'
1577 * as we might be about to remove it.
1578 */
1580 {
1590 }
1602 cnt++;
1604 }
1610 out:
1613 }
1616 {
1617 /* when calling 'enough', both 'prev' and 'geo' must
1618 * be stable.
1619 * This is ensured if ->reconfig_mutex or ->device_lock
1620 * is held.
1621 */
1624 }
1627 {
1632 /*
1633 * If it is not operational, then we have already marked it as dead
1634 * else if it is the last working disks, ignore the error, let the
1635 * next level up know.
1636 * else mark the drive as failed
1637 */
1641 /*
1642 * Don't fail the drive, just return an IO error.
1643 */
1646 }
1649 /*
1650 * If recovery is running, make sure it aborts.
1651 */
1662 }
1665 {
1673 }
1677 /* This is only called with ->reconfix_mutex held, so
1678 * rcu protection of rdev is not needed */
1687 }
1688 }
1691 {
1697 }
1700 {
1707 /*
1708 * Find all non-in_sync disks within the RAID10 configuration
1709 * and mark them in_sync
1710 */
1717 /* Replacement has just become active */
1720 count++;
1722 /* Replaced device not technically faulty,
1723 * but we need to be sure it gets removed
1724 * and never re-added.
1725 */
1729 }
1735 count++;
1737 }
1738 }
1745 }
1748 {
1756 /* only hot-add to in-sync arrays, as recovery is
1757 * very different from resync
1758 */
1772 else
1792 }
1806 }
1812 }
1815 {
1827 else
1834 }
1835 /* Only remove non-faulty devices if recovery
1836 * is not possible.
1837 */
1845 }
1850 /* lost the race, try later */
1854 }
1855 }
1857 /* We must have just cleared 'rdev' */
1861 * but will never see neither -- if they are careful.
1862 */
1866 /* We might have just remove the Replacement as faulty
1867 * Clear the flag just in case
1868 */
1873 abort:
1877 }
1880 {
1886 /* this is a reshape read */
1893 else
1894 /* The write handler will notice the lack of
1895 * R10BIO_Uptodate and record any errors etc
1896 */
1900 /* for reconstruct, we always reschedule after a read.
1901 * for resync, only after all reads
1902 */
1906 /* we have read all the blocks,
1907 * do the comparison in process context in raid10d
1908 */
1910 }
1911 }
1914 {
1919 /* the primary of several recovery bios */
1924 else
1933 else
1936 }
1937 }
1938 }
1941 {
1946 sector_t first_bad;
1955 else
1967 }
1977 }
1979 /*
1980 * Note: sync and recover and handled very differently for raid10
1981 * This code is for resync.
1982 * For resync, we read through virtual addresses and read all blocks.
1983 * If there is any error, we schedule a write. The lowest numbered
1984 * drive is authoritative.
1985 * However requests come for physical address, so we need to map.
1986 * For every physical address there are raid_disks/copies virtual addresses,
1987 * which is always are least one, but is not necessarly an integer.
1988 * This means that a physical address can span multiple chunks, so we may
1989 * have to submit multiple io requests for a single sync request.
1990 */
1991 /*
1992 * We check if all blocks are in-sync and only write to blocks that
1993 * aren't in sync
1994 */
1996 {
2004 /* find the first device with a block */
2018 /* now find blocks with errors */
2032 /* We know that the bi_io_vec layout is the same for
2033 * both 'first' and 'i', so we just compare them.
2034 * All vec entries are PAGE_SIZE;
2035 */
2043 len))
2046 }
2051 /* Don't fix anything. */
2054 /* Just give up on this device */
2057 }
2058 /* Ok, we need to write this bio, either to correct an
2059 * inconsistency or to correct an unreadable block.
2060 * First we need to fixup bv_offset, bv_len and
2061 * bi_vecs, as the read request might have corrupted these
2062 */
2083 }
2085 /* Now write out to any replacement devices
2086 * that are active
2087 */
2102 }
2104 done:
2108 }
2109 }
2111 /*
2112 * Now for the recovery code.
2113 * Recovery happens across physical sectors.
2114 * We recover all non-is_sync drives by finding the virtual address of
2115 * each, and then choose a working drive that also has that virt address.
2116 * There is a separate r10_bio for each non-in_sync drive.
2117 * Only the first two slots are in use. The first for reading,
2118 * The second for writing.
2119 *
2120 */
2122 {
2123 /* We got a read error during recovery.
2124 * We repeat the read in smaller page-sized sections.
2125 * If a read succeeds, write it to the new device or record
2126 * a bad block if we cannot.
2127 * If a read fails, record a bad block on both old and
2128 * new devices.
2129 */
2142 sector_t addr;
2151 addr,
2159 addr,
2169 }
2170 }
2172 /* We don't worry if we cannot set a bad block -
2173 * it really is bad so there is no loss in not
2174 * recording it yet
2175 */
2179 /* need bad block on destination too */
2184 /* just abort the recovery */
2193 }
2194 }
2195 }
2199 idx++;
2200 }
2201 }
2204 {
2213 }
2215 /*
2216 * share the pages with the first bio
2217 * and submit the write request
2218 */
2222 /* Need to test wbio2->bi_end_io before we call
2223 * generic_make_request as if the former is NULL,
2224 * the latter is free to free wbio2.
2225 */
2232 }
2238 }
2239 }
2241 /*
2242 * Used by fix_read_error() to decay the per rdev read_errors.
2243 * We halve the read error count for every hour that has elapsed
2244 * since the last recorded read error.
2245 *
2246 */
2248 {
2256 /* first time we've seen a read error */
2259 }
2266 /*
2267 * if hours_since_last is > the number of bits in read_errors
2268 * just set read errors to 0. We do this to avoid
2269 * overflowing the shift of read_errors by hours_since_last.
2270 */
2273 else
2275 }
2279 {
2280 sector_t first_bad;
2287 /* success */
2294 }
2295 /* need to record an error - either for the block or the device */
2299 }
2301 /*
2302 * This is a kernel thread which:
2303 *
2304 * 1. Retries failed read operations on working mirrors.
2305 * 2. Updates the raid superblock when problems encounter.
2306 * 3. Performs writes following reads for array synchronising.
2307 */
2310 {
2317 /* still own a reference to this rdev, so it cannot
2318 * have been cleared recently.
2319 */
2323 /* drive has already been failed, just ignore any
2324 more fix_read_error() attempts */
2341 }
2354 sector_t first_bad;
2368 sect,
2376 }
2377 sl++;
2384 /* Cannot read from anywhere, just mark the block
2385 * as bad on the first device to discourage future
2386 * reads.
2387 */
2392 rdev,
2399 }
2401 }
2404 /* write it back and re-read */
2411 sl--;
2423 sect,
2426 /* Well, this device is dead */
2430 sect +
2432 rdev)),
2437 }
2440 }
2447 sl--;
2459 sect,
2461 READ)) {
2463 /* Well, this device is dead */
2467 sect +
2478 sect +
2482 }
2486 }
2491 }
2492 }
2495 {
2500 /* bio has the data to be written to slot 'i' where
2501 * we just recently had a write error.
2502 * We repeatedly clone the bio and trim down to one block,
2503 * then try the write. Where the write fails we record
2504 * a bad block.
2505 * It is conceivable that the bio doesn't exactly align with
2506 * blocks. We must handle this.
2507 *
2508 * We currently own a reference to the rdev.
2509 */
2512 sector_t sector;
2529 sector_t wsector;
2532 /* Write at 'sector' for 'sectors' */
2542 /* Failure! */
2551 }
2553 }
2556 {
2564 dev_t bio_dev;
2565 sector_t bio_last_sector;
2567 /* we got a read error. Maybe the drive is bad. Maybe just
2568 * the block and we can fix it.
2569 * We freeze all other IO, and try reading the block from
2570 * other devices. When we find one, we re-write
2571 * and check it that fixes the read error.
2572 * This is all done synchronously while the array is
2573 * frozen.
2574 */
2593 read_more:
2601 }
2628 /* Drat - have to split this up more */
2637 else
2643 GFP_NOIO);
2656 }
2659 {
2660 /* Some sort of write request has finished and it
2661 * succeeded in writing where we thought there was a
2662 * bad block. So forget the bad block.
2663 * Or possibly if failed and we need to record
2664 * a bad block.
2665 */
2678 rdev,
2683 rdev,
2687 }
2694 rdev,
2699 rdev,
2703 }
2704 }
2714 rdev,
2724 }
2726 }
2731 rdev,
2735 }
2736 }
2748 }
2749 }
2750 }
2753 {
2771 }
2772 }
2776 retry_list);
2785 }
2786 }
2797 }
2817 /* just a partial read to be scheduled from a
2818 * separate context
2819 */
2822 }
2827 }
2829 }
2832 {
2847 }
2849 /*
2850 * perform a "sync" on one "block"
2851 *
2852 * We need to make sure that no normal I/O request - particularly write
2853 * requests - conflict with active sync requests.
2854 *
2855 * This is achieved by tracking pending requests and a 'barrier' concept
2856 * that can be installed to exclude normal IO requests.
2857 *
2858 * Resync and recovery are handled very differently.
2859 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2860 *
2861 * For resync, we iterate over virtual addresses, read all copies,
2862 * and update if there are differences. If only one copy is live,
2863 * skip it.
2864 * For recovery, we iterate over physical addresses, read a good
2865 * value for each non-in_sync drive, and over-write.
2866 *
2867 * So, for recovery we may have several outstanding complex requests for a
2868 * given address, one for each out-of-sync device. We model this by allocating
2869 * a number of r10_bio structures, one for each out-of-sync device.
2870 * As we setup these structures, we collect all bio's together into a list
2871 * which we then process collectively to add pages, and then process again
2872 * to pass to generic_make_request.
2873 *
2874 * The r10_bio structures are linked using a borrowed master_bio pointer.
2875 * This link is counted in ->remaining. When the r10_bio that points to NULL
2876 * has its remaining count decremented to 0, the whole complex operation
2877 * is complete.
2878 *
2879 */
2883 {
2890 sector_t sync_blocks;
2899 /*
2900 * Allow skipping a full rebuild for incremental assembly
2901 * of a clean array, like RAID1 does.
2902 */
2912 }
2914 skipped:
2920 /* If we aborted, we need to abort the
2921 * sync on the 'current' bitmap chucks (there can
2922 * be several when recovering multiple devices).
2923 * as we may have started syncing it but not finished.
2924 * We can find the current address in
2925 * mddev->curr_resync, but for recovery,
2926 * we need to convert that to several
2927 * virtual addresses.
2928 */
2933 }
2940 sector_t sect =
2944 }
2946 /* completed sync */
2950 /* Completed a full sync so the replacements
2951 * are now fully recovered.
2952 */
2959 }
2961 }
2963 }
2968 }
2974 /* if there has been nothing to do on any drive,
2975 * then there is nothing to do at all..
2976 */
2979 }
2984 /* make sure whole request will fit in a chunk - if chunks
2985 * are meaningful
2986 */
2991 /*
2992 * If there is non-resync activity waiting for a turn, then let it
2993 * though before starting on this new sync request.
2994 */
2998 /* Again, very different code for resync and recovery.
2999 * Both must result in an r10bio with a list of bios that
3000 * have bi_end_io, bi_sector, bi_bdev set,
3001 * and bi_private set to the r10bio.
3002 * For recovery, we may actually create several r10bios
3003 * with 2 bios in each, that correspond to the bios in the main one.
3004 * In this case, the subordinate r10bios link back through a
3005 * borrowed master_bio pointer, and the counter in the master
3006 * includes a ref from each subordinate.
3007 */
3008 /* First, we decide what to do and set ->bi_end_io
3009 * To end_sync_read if we want to read, and
3010 * end_sync_write if we will want to write.
3011 */
3015 /* recovery... the complicated one */
3022 sector_t sect;
3039 }
3042 /* want to reconstruct this device */
3046 /* last stripe is not complete - don't
3047 * try to recover this sector.
3048 */
3051 }
3054 /* Unless we are doing a full sync, or a replacement
3055 * we only need to recover the block if it is set in
3056 * the bitmap
3057 */
3065 /* yep, skip the sync_blocks here, but don't assume
3066 * that there will never be anything to do here
3067 */
3071 }
3091 /* Need to check if the array will still be
3092 * degraded
3093 */
3101 }
3102 }
3119 /* This is where we read from */
3128 bad_sectors -= (sector
3133 }
3134 }
3149 /* and we write to 'i' (if not in_sync) */
3176 /* and maybe write to replacement */
3180 /* Note: if mreplace != NULL, then bio
3181 * cannot be NULL as r10buf_pool_alloc will
3182 * have allocated it.
3183 * So the second test here is pointless.
3184 * But it keeps semantic-checkers happy, and
3185 * this comment keeps human reviewers
3186 * happy.
3187 */
3202 }
3205 /* Cannot recover, so abort the recovery or
3206 * record a bad block */
3208 /* problem is that there are bad blocks
3209 * on other device(s)
3210 */
3218 mrdev,
3224 mreplace,
3228 }
3234 mirror->recovery_disabled
3236 }
3245 }
3250 /* Only want this if there is elsewhere to
3251 * read from. 'j' is currently the first
3252 * readable copy.
3253 */
3260 targets++;
3261 }
3265 }
3266 }
3273 }
3275 }
3277 /* resync. Schedule a read for every block at this virt offset */
3286 /* We can skip this block */
3289 }
3323 }
3335 }
3336 }
3348 count++;
3354 }
3358 /* Need to set up for writing to the replacement */
3373 count++;
3374 }
3381 mddev);
3386 mddev);
3387 }
3391 }
3392 }
3410 /* stop here */
3415 /* remove last page from this bio */
3419 }
3421 }
3425 bio_full:
3440 }
3441 }
3444 /* pretend they weren't skipped, it makes
3445 * no important difference in this case
3446 */
3450 giveup:
3451 /* There is nowhere to write, so all non-sync
3452 * drives must be failed or in resync, all drives
3453 * have a bad block, so try the next chunk...
3454 */
3459 chunks_skipped ++;
3462 }
3464 static sector_t
3466 {
3467 sector_t size;
3482 }
3485 {
3486 /* Calculate the number of sectors-per-device that will
3487 * actually be used, and set conf->dev_sectors and
3488 * conf->stride
3489 */
3495 /* 'size' is now the number of chunks in the array */
3496 /* calculate "used chunks per device" */
3499 /* We need to round up when dividing by raid_disks to
3500 * get the stride size.
3501 */
3511 }
3512 }
3516 {
3532 * updated yet. */
3537 }
3555 * actually using this, but leave code here just in case.*/
3565 }
3569 }
3572 {
3584 }
3590 }
3597 /* FIXME calc properly */
3600 GFP_KERNEL);
3623 }
3627 else
3628 /* far_copies must be 1 */
3630 }
3647 out:
3653 }
3655 }
3658 {
3663 sector_t size;
3673 }
3689 else
3692 }
3714 }
3732 }
3738 else
3741 }
3742 /* need to check that every block has at least one working mirror */
3747 }
3750 /* must ensure that shape change is supported */
3757 }
3763 i++) {
3768 /* The replacement is all we have - use it */
3772 }
3781 }
3783 }
3791 /*
3792 * Ok, everything is just fine now
3793 */
3804 /* Calculate max read-ahead size.
3805 * We need to readahead at least twice a whole stripe....
3806 * maybe...
3807 */
3811 }
3825 /* This cannot work */
3828 }
3837 }
3841 out_free_conf:
3848 out:
3850 }
3853 {
3862 }
3865 {
3875 }
3876 }
3879 {
3880 /* Resize of 'far' arrays is not supported.
3881 * For 'near' and 'offset' arrays we can set the
3882 * number of sectors used to be an appropriate multiple
3883 * of the chunk size.
3884 * For 'offset', this is far_copies*chunksize.
3885 * For 'near' the multiplier is the LCM of
3886 * near_copies and raid_disks.
3887 * So if far_copies > 1 && !far_offset, fail.
3888 * Else find LCM(raid_disks, near_copy)*far_copies and
3889 * multiply by chunk_size. Then round to this number.
3890 * This is mostly done by raid10_size()
3891 */
3910 }
3915 }
3920 }
3925 }
3928 {
3936 }
3939 /* Set new parameters */
3941 /* new layout: far_copies = 1, near_copies = 2 */
3946 /* make sure it will be not marked as dirty */
3956 }
3958 }
3961 }
3964 {
3967 /* raid10 can take over:
3968 * raid0 - providing it has only two drives
3969 */
3971 /* for raid0 takeover only one zone is supported */
3977 }
3981 }
3983 }
3986 {
3987 /* Called when there is a request to change
3988 * - layout (to ->new_layout)
3989 * - chunk size (to ->new_chunk_sectors)
3990 * - raid_disks (by delta_disks)
3991 * or when trying to restart a reshape that was ongoing.
3992 *
3993 * We need to validate the request and possibly allocate
3994 * space if that might be an issue later.
3995 *
3996 * Currently we reject any reshape of a 'far' mode array,
3997 * allow chunk size to change if new is generally acceptable,
3998 * allow raid_disks to increase, and allow
3999 * a switch between 'near' mode and 'offset' mode.
4000 */
4008 /* mustn't change number of copies */
4011 /* Cannot switch to 'far' mode */
4015 /* not factor of array size */
4024 /* allocate new 'mirrors' list */
4029 GFP_KERNEL);
4032 }
4034 }
4036 /*
4037 * Need to check if array has failed when deciding whether to:
4038 * - start an array
4039 * - remove non-faulty devices
4040 * - add a spare
4041 * - allow a reshape
4042 * This determination is simple when no reshape is happening.
4043 * However if there is a reshape, we need to carefully check
4044 * both the before and after sections.
4045 * This is because some failed devices may only affect one
4046 * of the two sections, and some non-in_sync devices may
4047 * be insync in the section most affected by failed devices.
4048 */
4050 {
4056 /* 'prev' section first */
4060 degraded++;
4062 /* When we can reduce the number of devices in
4063 * an array, this might not contribute to
4064 * 'degraded'. It does now.
4065 */
4066 degraded++;
4067 }
4076 degraded2++;
4078 /* If reshape is increasing the number of devices,
4079 * this section has already been recovered, so
4080 * it doesn't contribute to degraded.
4081 * else it does.
4082 */
4084 degraded2++;
4085 }
4086 }
4091 }
4094 {
4095 /* A 'reshape' has been requested. This commits
4096 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4097 * This also checks if there are enough spares and adds them
4098 * to the array.
4099 * We currently require enough spares to make the final
4100 * array non-degraded. We also require that the difference
4101 * between old and new data_offset - on each device - is
4102 * enough that we never risk over-writing.
4103 */
4128 spares++;
4138 }
4139 }
4157 }
4167 }
4182 }
4191 else
4196 }
4199 /* This is a spare that was manually added */
4201 }
4202 }
4203 /* When a reshape changes the number of devices,
4204 * ->degraded is measured against the larger of the
4205 * pre and post numbers.
4206 */
4225 }
4231 abort:
4244 }
4246 /* Calculate the last device-address that could contain
4247 * any block from the chunk that includes the array-address 's'
4248 * and report the next address.
4249 * i.e. the address returned will be chunk-aligned and after
4250 * any data that is in the chunk containing 's'.
4251 */
4253 {
4261 }
4263 /* Calculate the first device-address that could contain
4264 * any block from the chunk that includes the array-address 's'.
4265 * This too will be the start of a chunk
4266 */
4268 {
4275 }
4279 {
4280 /* We simply copy at most one chunk (smallest of old and new)
4281 * at a time, possibly less if that exceeds RESYNC_PAGES,
4282 * or we hit a bad block or something.
4283 * This might mean we pause for normal IO in the middle of
4284 * a chunk, but that is not a problem as mddev->reshape_position
4285 * can record any location.
4286 *
4287 * If we will want to write to a location that isn't
4288 * yet recorded as 'safe' (i.e. in metadata on disk) then
4289 * we need to flush all reshape requests and update the metadata.
4290 *
4291 * When reshaping forwards (e.g. to more devices), we interpret
4292 * 'safe' as the earliest block which might not have been copied
4293 * down yet. We divide this by previous stripe size and multiply
4294 * by previous stripe length to get lowest device offset that we
4295 * cannot write to yet.
4296 * We interpret 'sector_nr' as an address that we want to write to.
4297 * From this we use last_device_address() to find where we might
4298 * write to, and first_device_address on the 'safe' position.
4299 * If this 'next' write position is after the 'safe' position,
4300 * we must update the metadata to increase the 'safe' position.
4301 *
4302 * When reshaping backwards, we round in the opposite direction
4303 * and perform the reverse test: next write position must not be
4304 * less than current safe position.
4305 *
4306 * In all this the minimum difference in data offsets
4307 * (conf->offset_diff - always positive) allows a bit of slack,
4308 * so next can be after 'safe', but not by more than offset_diff
4309 *
4310 * We need to prepare all the bios here before we start any IO
4311 * to ensure the size we choose is acceptable to all devices.
4312 * The means one for each copy for write-out and an extra one for
4313 * read-in.
4314 * We store the read-in bio in ->master_bio and the others in
4315 * ->devs[x].bio and ->devs[x].repl_bio.
4316 */
4330 /* If restarting in the middle, skip the initial sectors */
4343 }
4344 }
4346 /* We don't use sector_nr to track where we are up to
4347 * as that doesn't work well for ->reshape_backwards.
4348 * So just use ->reshape_progress.
4349 */
4351 /* 'next' is the earliest device address that we might
4352 * write to for this chunk in the new layout
4353 */
4357 /* 'safe' is the last device address that we might read from
4358 * in the old layout after a restart
4359 */
4372 /* 'next' is after the last device address that we
4373 * might write to for this chunk in the new layout
4374 */
4377 /* 'safe' is the earliest device address that we might
4378 * read from in the old layout after a restart
4379 */
4382 /* Need to update metadata if 'next' might be beyond 'safe'
4383 * as that would possibly corrupt data
4384 */
4394 }
4398 /* Need to update reshape_position in metadata */
4404 else
4414 }
4417 }
4419 read_more:
4420 /* Now schedule reads for blocks from sector_nr to last */
4433 /* Cannot read from here, so need to record bad blocks
4434 * on all the target devices.
4435 */
4436 // FIXME
4440 }
4457 /* Now find the locations in the new layout */
4474 }
4487 }
4489 /* Now add as many pages as possible to all of these bios. */
4502 /* Didn't fit, must stop */
4506 /* Remove last page from this bio */
4510 }
4512 }
4515 }
4516 bio_full:
4520 /* Now submit the read */
4530 /* Now that we have done the whole section we can
4531 * update reshape_progress
4532 */
4535 else
4539 }
4545 {
4546 /* Reshape read completed. Hopefully we have a block
4547 * to write out.
4548 * If we got a read error then we do sync 1-page reads from
4549 * elsewhere until we find the data - or give up.
4550 */
4556 /* Reshape has been aborted */
4559 }
4561 /* We definitely have the data in the pages, schedule the
4562 * writes.
4563 */
4576 }
4580 }
4587 }
4589 }
4592 {
4604 /* read-ahead size must cover two whole stripes, which is
4605 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4606 */
4613 }
4615 }
4619 {
4620 /* Use sync reads to get the blocks from somewhere else */
4647 sector_t addr;
4657 addr,
4665 failed:
4666 slot++;
4671 }
4674 /* couldn't read this block, must give up */
4678 }
4680 idx++;
4681 }
4683 }
4686 {
4701 }
4704 /* FIXME should record badblock */
4706 }
4710 }
4713 {
4719 }
4722 {
4734 }
4739 }
4745 d++) {
4752 }
4754 }
4760 }
4763 {
4784 };
4787 {
4789 }
4792 {
4794 }