| blob | b98e746e7fc4fd05fb8c0eaf2118f9f6a4f778d3 |
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 <linux/raid/md_p.h>
29 #include <trace/events/block.h>
35 /*
36 * RAID10 provides a combination of RAID0 and RAID1 functionality.
37 * The layout of data is defined by
38 * chunk_size
39 * raid_disks
40 * near_copies (stored in low byte of layout)
41 * far_copies (stored in second byte of layout)
42 * far_offset (stored in bit 16 of layout )
43 * use_far_sets (stored in bit 17 of layout )
44 * use_far_sets_bugfixed (stored in bit 18 of layout )
45 *
46 * The data to be stored is divided into chunks using chunksize. Each device
47 * is divided into far_copies sections. In each section, chunks are laid out
48 * in a style similar to raid0, but near_copies copies of each chunk is stored
49 * (each on a different drive). The starting device for each section is offset
50 * near_copies from the starting device of the previous section. Thus there
51 * are (near_copies * far_copies) of each chunk, and each is on a different
52 * drive. near_copies and far_copies must be at least one, and their product
53 * is at most raid_disks.
54 *
55 * If far_offset is true, then the far_copies are handled a bit differently.
56 * The copies are still in different stripes, but instead of being very far
57 * apart on disk, there are adjacent stripes.
58 *
59 * The far and offset algorithms are handled slightly differently if
60 * 'use_far_sets' is true. In this case, the array's devices are grouped into
61 * sets that are (near_copies * far_copies) in size. The far copied stripes
62 * are still shifted by 'near_copies' devices, but this shifting stays confined
63 * to the set rather than the entire array. This is done to improve the number
64 * of device combinations that can fail without causing the array to fail.
65 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
66 * on a device):
67 * A B C D A B C D E
68 * ... ...
69 * D A B C E A B C D
70 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
71 * [A B] [C D] [A B] [C D E]
72 * |...| |...| |...| | ... |
73 * [B A] [D C] [B A] [E C D]
74 */
76 /*
77 * Number of guaranteed r10bios in case of extreme VM load:
78 */
79 #define NR_RAID10_BIOS 256
81 /* when we get a read error on a read-only array, we redirect to another
82 * device without failing the first device, or trying to over-write to
83 * correct the read error. To keep track of bad blocks on a per-bio
84 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
85 */
86 #define IO_BLOCKED ((struct bio *)1)
87 /* When we successfully write to a known bad-block, we need to remove the
88 * bad-block marking which must be done from process context. So we record
89 * the success by setting devs[n].bio to IO_MADE_GOOD
90 */
91 #define IO_MADE_GOOD ((struct bio *)2)
93 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
95 /* When there are this many requests queued to be written by
96 * the raid10 thread, we become 'congested' to provide back-pressure
97 * for writeback.
98 */
111 #define raid10_log(md, fmt, args...) \
116 /*
117 * for resync bio, r10bio pointer can be retrieved from the per-bio
118 * 'struct resync_pages'.
119 */
121 {
123 }
126 {
130 /* allocate a r10bio with room for raid_disks entries in the
131 * bios array */
133 }
136 {
138 }
140 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
141 /* amount of memory to reserve for resync requests */
142 #define RESYNC_WINDOW (1024*1024)
143 /* maximum number of concurrent requests, memory permitting */
144 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
145 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
146 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
148 /*
149 * When performing a resync, we need to read and compare, so
150 * we need as many pages are there are copies.
151 * When performing a recovery, we need 2 bios, one for read,
152 * one for write (we recover only one drive per r10buf)
153 *
154 */
156 {
171 else
174 /* allocate once for all bios */
177 else
183 /*
184 * Allocate bios.
185 */
197 }
198 /*
199 * Allocate RESYNC_PAGES data pages and attach them
200 * where needed.
201 */
219 }
226 }
227 }
231 out_free_pages:
236 out_free_bio:
242 }
244 out_free_r10bio:
247 }
250 {
263 }
268 }
270 /* resync pages array stored in the 1st bio's .bi_private */
274 }
277 {
289 }
290 }
293 {
298 }
301 {
307 }
310 {
320 /* wake up frozen array... */
324 }
326 /*
327 * raid_end_bio_io() is called when we have finished servicing a mirrored
328 * operation and are ready to return a success/failure code to the buffer
329 * cache layer.
330 */
332 {
340 /*
341 * Wake up any possible resync thread that waits for the device
342 * to go idle.
343 */
347 }
349 /*
350 * Update disk head position estimator based on IRQ completion info.
351 */
353 {
358 }
360 /*
361 * Find the disk number which triggered given bio
362 */
365 {
375 }
376 }
386 }
389 {
398 /*
399 * this branch is our 'one mirror IO has finished' event handler:
400 */
404 /*
405 * Set R10BIO_Uptodate in our master bio, so that
406 * we will return a good error code to the higher
407 * levels even if IO on some other mirrored buffer fails.
408 *
409 * The 'master' represents the composite IO operation to
410 * user-side. So if something waits for IO, then it will
411 * wait for the 'master' bio.
412 */
415 /* If all other devices that store this block have
416 * failed, we want to return the error upwards rather
417 * than fail the last device. Here we redefine
418 * "uptodate" to mean "Don't want to retry"
419 */
423 }
428 /*
429 * oops, read error - keep the refcount on the rdev
430 */
438 }
439 }
442 {
443 /* clear the bitmap if all writes complete successfully */
449 }
452 {
460 else
462 }
463 }
464 }
467 {
487 }
488 /*
489 * this branch is our 'one mirror IO has finished' event handler:
490 */
493 /* Never record new bad blocks to replacement,
494 * just fail it.
495 */
508 /* This is the only remaining device,
509 * We need to retry the write without
510 * FailFast
511 */
517 }
520 }
522 /*
523 * Set R10BIO_Uptodate in our master bio, so that
524 * we will return a good error code for to the higher
525 * levels even if IO on some other mirrored buffer fails.
526 *
527 * The 'master' represents the composite IO operation to
528 * user-side. So if something waits for IO, then it will
529 * wait for the 'master' bio.
530 */
531 sector_t first_bad;
534 /*
535 * Do not set R10BIO_Uptodate if the current device is
536 * rebuilding or Faulty. This is because we cannot use
537 * such device for properly reading the data back (we could
538 * potentially use it, if the current write would have felt
539 * before rdev->recovery_offset, but for simplicity we don't
540 * check this here.
541 */
546 /* Maybe we can clear some bad blocks. */
554 else
558 }
559 }
561 /*
562 *
563 * Let's see if all mirrored write operations have finished
564 * already.
565 */
571 }
573 /*
574 * RAID10 layout manager
575 * As well as the chunksize and raid_disks count, there are two
576 * parameters: near_copies and far_copies.
577 * near_copies * far_copies must be <= raid_disks.
578 * Normally one of these will be 1.
579 * If both are 1, we get raid0.
580 * If near_copies == raid_disks, we get raid1.
581 *
582 * Chunks are laid out in raid0 style with near_copies copies of the
583 * first chunk, followed by near_copies copies of the next chunk and
584 * so on.
585 * If far_copies > 1, then after 1/far_copies of the array has been assigned
586 * as described above, we start again with a device offset of near_copies.
587 * So we effectively have another copy of the whole array further down all
588 * the drives, but with blocks on different drives.
589 * With this layout, and block is never stored twice on the one device.
590 *
591 * raid10_find_phys finds the sector offset of a given virtual sector
592 * on each device that it is on.
593 *
594 * raid10_find_virt does the reverse mapping, from a device and a
595 * sector offset to a virtual address
596 */
599 {
601 sector_t sector;
602 sector_t chunk;
603 sector_t stripe;
614 /* now calculate first sector/dev */
626 /* and calculate all the others */
633 slot++;
647 }
651 slot++;
652 }
653 dev++;
657 }
658 }
659 }
662 {
674 }
677 {
679 /* Never use conf->prev as this is only called during resync
680 * or recovery, so reshape isn't happening
681 */
695 }
696 }
711 else
713 }
715 }
719 }
721 /*
722 * This routine returns the disk from which the requested read should
723 * be done. There is a per-array 'next expected sequential IO' sector
724 * number - if this matches on the next IO then we use the last disk.
725 * There is also a per-disk 'last know head position' sector that is
726 * maintained from IRQ contexts, both the normal and the resync IO
727 * completion handlers update this position correctly. If there is no
728 * perfect sequential match then we pick the disk whose head is closest.
729 *
730 * If there are 2 mirrors in the same 2 devices, performance degrades
731 * because position is mirror, not device based.
732 *
733 * The rdev for the device selected will have nr_pending incremented.
734 */
736 /*
737 * FIXME: possibly should rethink readbalancing and do it differently
738 * depending on near_copies / far_copies geometry.
739 */
743 {
762 /*
763 * Check if we can balance. We can balance on the whole
764 * device if no resync is going on (recovery is ok), or below
765 * the resync window. We take the first readable disk when
766 * above the resync window.
767 */
776 sector_t first_bad;
778 sector_t dev_sector;
798 /* Already have a better slot */
801 /* Cannot read here. If this is the
802 * 'primary' device, then we must not read
803 * beyond 'bad_sectors' from another device.
804 */
811 sector_t good_sectors =
817 }
819 /* Must read from here */
821 }
830 /* At least 2 disks to choose from so failfast is OK */
832 /* This optimisation is debatable, and completely destroys
833 * sequential read speed for 'far copies' arrays. So only
834 * keep it for 'near' arrays, and review those later.
835 */
839 /* for far > 1 always use the lowest address */
842 else
849 }
850 }
854 }
865 }
868 {
880 i++) {
886 }
887 }
890 }
893 {
894 /* Any writes that have been queued but are awaiting
895 * bitmap updates get flushed here.
896 */
907 /*
908 * As this is called in a wait_event() loop (see freeze_array),
909 * current->state might be TASK_UNINTERRUPTIBLE which will
910 * cause a warning when we prepare to wait again. As it is
911 * rare that this path is taken, it is perfectly safe to force
912 * us to go around the wait_event() loop again, so the warning
913 * is a false-positive. Silence the warning by resetting
914 * thread state
915 */
919 /* flush any pending bitmap writes to disk
920 * before proceeding w/ I/O */
933 /* Just ignore it */
935 else
938 }
942 }
944 /* Barriers....
945 * Sometimes we need to suspend IO while we do something else,
946 * either some resync/recovery, or reconfigure the array.
947 * To do this we raise a 'barrier'.
948 * The 'barrier' is a counter that can be raised multiple times
949 * to count how many activities are happening which preclude
950 * normal IO.
951 * We can only raise the barrier if there is no pending IO.
952 * i.e. if nr_pending == 0.
953 * We choose only to raise the barrier if no-one is waiting for the
954 * barrier to go down. This means that as soon as an IO request
955 * is ready, no other operations which require a barrier will start
956 * until the IO request has had a chance.
957 *
958 * So: regular IO calls 'wait_barrier'. When that returns there
959 * is no backgroup IO happening, It must arrange to call
960 * allow_barrier when it has finished its IO.
961 * backgroup IO calls must call raise_barrier. Once that returns
962 * there is no normal IO happeing. It must arrange to call
963 * lower_barrier when the particular background IO completes.
964 */
967 {
971 /* Wait until no block IO is waiting (unless 'force') */
975 /* block any new IO from starting */
978 /* Now wait for all pending IO to complete */
984 }
987 {
993 }
996 {
1000 /* Wait for the barrier to drop.
1001 * However if there are already pending
1002 * requests (preventing the barrier from
1003 * rising completely), and the
1004 * pre-process bio queue isn't empty,
1005 * then don't wait, as we need to empty
1006 * that queue to get the nr_pending
1007 * count down.
1008 */
1020 }
1023 }
1026 {
1030 }
1033 {
1034 /* stop syncio and normal IO and wait for everything to
1035 * go quiet.
1036 * We increment barrier and nr_waiting, and then
1037 * wait until nr_pending match nr_queued+extra
1038 * This is called in the context of one normal IO request
1039 * that has failed. Thus any sync request that might be pending
1040 * will be blocked by nr_pending, and we need to wait for
1041 * pending IO requests to complete or be queued for re-try.
1042 * Thus the number queued (nr_queued) plus this request (extra)
1043 * must match the number of pending IOs (nr_pending) before
1044 * we continue.
1045 */
1057 }
1060 {
1061 /* reverse the effect of the freeze */
1067 }
1071 {
1075 else
1077 }
1083 };
1086 {
1088 cb);
1102 }
1104 /* we aren't scheduling, so we can do the write-out directly. */
1118 /* Just ignore it */
1120 else
1123 }
1125 }
1129 {
1135 sector_t sectors;
1143 /*
1144 * This is an error retry, but we cannot
1145 * safely dereference the rdev in the r10_bio,
1146 * we must use the one in conf.
1147 * If it has already been disconnected (unlikely)
1148 * we lose the device name in error messages.
1149 */
1151 /*
1152 * As we are blocking raid10, it is a little safer to
1153 * use __GFP_HIGH.
1154 */
1164 /* This never gets dereferenced */
1166 }
1168 }
1169 /*
1170 * Register the new request and wait if the reconstruction
1171 * thread has put up a bar for new requests.
1172 * Continue immediately if no resync is active currently.
1173 */
1180 /*
1181 * IO spans the reshape position. Need to wait for reshape to
1182 * pass
1183 */
1189 sectors);
1191 }
1199 }
1202 }
1216 }
1240 }
1245 {
1260 /* Replacement just got moved to main 'rdev' */
1263 }
1270 else
1288 /* flush_pending_writes() needs access to the rdev so...*/
1296 else
1307 }
1308 }
1312 {
1316 sector_t sectors;
1331 }
1333 }
1335 /*
1336 * Register the new request and wait if the reconstruction
1337 * thread has put up a bar for new requests.
1338 * Continue immediately if no resync is active currently.
1339 */
1346 /*
1347 * IO spans the reshape position. Need to wait for reshape to
1348 * pass
1349 */
1355 sectors);
1357 }
1365 /* Need to update reshape_position in metadata */
1375 }
1382 }
1383 /* first select target devices under rcu_lock and
1384 * inc refcount on their rdev. Record them by setting
1385 * bios[x] to bio
1386 * If there are known/acknowledged bad blocks on any device
1387 * on which we have seen a write error, we want to avoid
1388 * writing to those blocks. This potentially requires several
1389 * writes to write around the bad blocks. Each set of writes
1390 * gets its own r10_bio with a set of bios attached.
1391 */
1395 retry_write:
1411 }
1416 }
1428 }
1430 sector_t first_bad;
1438 /* Mustn't write here until the bad block
1439 * is acknowledged
1440 */
1445 }
1447 /* Cannot write here at all */
1450 /* Mustn't write more than bad_sectors
1451 * to other devices yet
1452 */
1454 /* We don't set R10BIO_Degraded as that
1455 * only applies if the disk is missing,
1456 * so it might be re-added, and we want to
1457 * know to recover this chunk.
1458 * In this case the device is here, and the
1459 * fact that this chunk is not in-sync is
1460 * recorded in the bad block log.
1461 */
1463 }
1468 }
1469 }
1473 }
1477 }
1478 }
1482 /* Have to wait for this device to get unblocked, then retry */
1490 }
1496 /* Race with remove_disk */
1499 }
1501 }
1502 }
1508 }
1520 }
1530 }
1532 }
1535 {
1551 else
1553 }
1556 {
1565 }
1570 /*
1571 * If this request crosses a chunk boundary, we need to split
1572 * it.
1573 */
1575 sectors > chunk_sects
1584 /* In case raid10d snuck in to freeze_array */
1587 }
1590 {
1601 else
1605 }
1612 }
1615 }
1617 /* check if there are enough drives for
1618 * every block to appear on atleast one.
1619 * Don't consider the device numbered 'ignore'
1620 * as we might be about to remove it.
1621 */
1623 {
1633 }
1645 cnt++;
1647 }
1653 out:
1656 }
1659 {
1660 /* when calling 'enough', both 'prev' and 'geo' must
1661 * be stable.
1662 * This is ensured if ->reconfig_mutex or ->device_lock
1663 * is held.
1664 */
1667 }
1670 {
1675 /*
1676 * If it is not operational, then we have already marked it as dead
1677 * else if it is the last working disks, ignore the error, let the
1678 * next level up know.
1679 * else mark the drive as failed
1680 */
1684 /*
1685 * Don't fail the drive, just return an IO error.
1686 */
1689 }
1692 /*
1693 * If recovery is running, make sure it aborts.
1694 */
1705 }
1708 {
1716 }
1720 /* This is only called with ->reconfix_mutex held, so
1721 * rcu protection of rdev is not needed */
1730 }
1731 }
1734 {
1739 }
1742 {
1749 /*
1750 * Find all non-in_sync disks within the RAID10 configuration
1751 * and mark them in_sync
1752 */
1759 /* Replacement has just become active */
1762 count++;
1764 /* Replaced device not technically faulty,
1765 * but we need to be sure it gets removed
1766 * and never re-added.
1767 */
1771 }
1777 count++;
1779 }
1780 }
1787 }
1790 {
1798 /* only hot-add to in-sync arrays, as recovery is
1799 * very different from resync
1800 */
1815 else
1835 }
1849 }
1855 }
1858 {
1870 else
1877 }
1878 /* Only remove non-faulty devices if recovery
1879 * is not possible.
1880 */
1888 }
1893 /* lost the race, try later */
1897 }
1898 }
1900 /* We must have just cleared 'rdev' */
1904 * but will never see neither -- if they are careful.
1905 */
1907 }
1912 abort:
1916 }
1919 {
1924 else
1925 /* The write handler will notice the lack of
1926 * R10BIO_Uptodate and record any errors etc
1927 */
1931 /* for reconstruct, we always reschedule after a read.
1932 * for resync, only after all reads
1933 */
1937 /* we have read all the blocks,
1938 * do the comparison in process context in raid10d
1939 */
1941 }
1942 }
1945 {
1951 }
1954 {
1955 /* reshape read bio isn't allocated from r10buf_pool */
1959 }
1962 {
1967 /* the primary of several recovery bios */
1972 else
1981 else
1984 }
1985 }
1986 }
1989 {
1994 sector_t first_bad;
2003 else
2015 }
2025 }
2027 /*
2028 * Note: sync and recover and handled very differently for raid10
2029 * This code is for resync.
2030 * For resync, we read through virtual addresses and read all blocks.
2031 * If there is any error, we schedule a write. The lowest numbered
2032 * drive is authoritative.
2033 * However requests come for physical address, so we need to map.
2034 * For every physical address there are raid_disks/copies virtual addresses,
2035 * which is always are least one, but is not necessarly an integer.
2036 * This means that a physical address can span multiple chunks, so we may
2037 * have to submit multiple io requests for a single sync request.
2038 */
2039 /*
2040 * We check if all blocks are in-sync and only write to blocks that
2041 * aren't in sync
2042 */
2044 {
2053 /* find the first device with a block */
2068 /* now find blocks with errors */
2085 /* We know that the bi_io_vec layout is the same for
2086 * both 'first' and 'i', so we just compare them.
2087 * All vec entries are PAGE_SIZE;
2088 */
2096 len))
2099 }
2104 /* Don't fix anything. */
2107 /* Just give up on this device */
2110 }
2111 /* Ok, we need to write this bio, either to correct an
2112 * inconsistency or to correct an unreadable block.
2113 * First we need to fixup bv_offset, bv_len and
2114 * bi_vecs, as the read request might have corrupted these
2115 */
2138 }
2140 /* Now write out to any replacement devices
2141 * that are active
2142 */
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 */
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 */
2249 }
2250 }
2251 }
2255 idx++;
2256 }
2257 }
2260 {
2269 }
2271 /*
2272 * share the pages with the first bio
2273 * and submit the write request
2274 */
2278 /* Need to test wbio2->bi_end_io before we call
2279 * generic_make_request as if the former is NULL,
2280 * the latter is free to free wbio2.
2281 */
2288 }
2294 }
2295 }
2297 /*
2298 * Used by fix_read_error() to decay the per rdev read_errors.
2299 * We halve the read error count for every hour that has elapsed
2300 * since the last recorded read error.
2301 *
2302 */
2304 {
2312 /* first time we've seen a read error */
2315 }
2322 /*
2323 * if hours_since_last is > the number of bits in read_errors
2324 * just set read errors to 0. We do this to avoid
2325 * overflowing the shift of read_errors by hours_since_last.
2326 */
2329 else
2331 }
2335 {
2336 sector_t first_bad;
2343 /* success */
2350 }
2351 /* need to record an error - either for the block or the device */
2355 }
2357 /*
2358 * This is a kernel thread which:
2359 *
2360 * 1. Retries failed read operations on working mirrors.
2361 * 2. Updates the raid superblock when problems encounter.
2362 * 3. Performs writes following reads for array synchronising.
2363 */
2366 {
2373 /* still own a reference to this rdev, so it cannot
2374 * have been cleared recently.
2375 */
2379 /* drive has already been failed, just ignore any
2380 more fix_read_error() attempts */
2397 }
2410 sector_t first_bad;
2424 sect,
2432 }
2433 sl++;
2440 /* Cannot read from anywhere, just mark the block
2441 * as bad on the first device to discourage future
2442 * reads.
2443 */
2448 rdev,
2455 }
2457 }
2460 /* write it back and re-read */
2467 sl--;
2479 sect,
2482 /* Well, this device is dead */
2486 sect +
2488 rdev)),
2493 }
2496 }
2503 sl--;
2515 sect,
2517 READ)) {
2519 /* Well, this device is dead */
2523 sect +
2534 sect +
2538 }
2542 }
2547 }
2548 }
2551 {
2556 /* bio has the data to be written to slot 'i' where
2557 * we just recently had a write error.
2558 * We repeatedly clone the bio and trim down to one block,
2559 * then try the write. Where the write fails we record
2560 * a bad block.
2561 * It is conceivable that the bio doesn't exactly align with
2562 * blocks. We must handle this.
2563 *
2564 * We currently own a reference to the rdev.
2565 */
2568 sector_t sector;
2585 sector_t wsector;
2588 /* Write at 'sector' for 'sectors' */
2598 /* Failure! */
2607 }
2609 }
2612 {
2618 /* we got a read error. Maybe the drive is bad. Maybe just
2619 * the block and we can fix it.
2620 * We freeze all other IO, and try reading the block from
2621 * other devices. When we find one, we re-write
2622 * and check it that fixes the read error.
2623 * This is all done synchronously while the array is
2624 * frozen.
2625 */
2643 }
2646 {
2647 /* Some sort of write request has finished and it
2648 * succeeded in writing where we thought there was a
2649 * bad block. So forget the bad block.
2650 * Or possibly if failed and we need to record
2651 * a bad block.
2652 */
2666 rdev,
2671 rdev,
2675 }
2683 rdev,
2688 rdev,
2692 }
2693 }
2703 rdev,
2713 }
2715 }
2720 rdev,
2724 }
2725 }
2731 /*
2732 * In case freeze_array() is waiting for condition
2733 * nr_pending == nr_queued + extra to be true.
2734 */
2742 }
2743 }
2744 }
2747 {
2765 }
2766 }
2770 retry_list);
2779 }
2780 }
2791 }
2810 else
2816 }
2818 }
2821 {
2836 }
2839 {
2849 else
2862 }
2863 }
2865 }
2867 /*
2868 * Set cluster_sync_high since we need other nodes to add the
2869 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2870 */
2872 {
2873 sector_t window_size;
2876 /*
2877 * First, here we define "stripe" as a unit which across
2878 * all member devices one time, so we get chunks by use
2879 * raid_disks / near_copies. Otherwise, if near_copies is
2880 * close to raid_disks, then resync window could increases
2881 * linearly with the increase of raid_disks, which means
2882 * we will suspend a really large IO window while it is not
2883 * necessary. If raid_disks is not divisible by near_copies,
2884 * an extra chunk is needed to ensure the whole "stripe" is
2885 * covered.
2886 */
2891 else
2895 /*
2896 * At least use a 32M window to align with raid1's resync window
2897 */
2902 }
2904 /*
2905 * perform a "sync" on one "block"
2906 *
2907 * We need to make sure that no normal I/O request - particularly write
2908 * requests - conflict with active sync requests.
2909 *
2910 * This is achieved by tracking pending requests and a 'barrier' concept
2911 * that can be installed to exclude normal IO requests.
2912 *
2913 * Resync and recovery are handled very differently.
2914 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2915 *
2916 * For resync, we iterate over virtual addresses, read all copies,
2917 * and update if there are differences. If only one copy is live,
2918 * skip it.
2919 * For recovery, we iterate over physical addresses, read a good
2920 * value for each non-in_sync drive, and over-write.
2921 *
2922 * So, for recovery we may have several outstanding complex requests for a
2923 * given address, one for each out-of-sync device. We model this by allocating
2924 * a number of r10_bio structures, one for each out-of-sync device.
2925 * As we setup these structures, we collect all bio's together into a list
2926 * which we then process collectively to add pages, and then process again
2927 * to pass to generic_make_request.
2928 *
2929 * The r10_bio structures are linked using a borrowed master_bio pointer.
2930 * This link is counted in ->remaining. When the r10_bio that points to NULL
2931 * has its remaining count decremented to 0, the whole complex operation
2932 * is complete.
2933 *
2934 */
2938 {
2945 sector_t sync_blocks;
2955 /*
2956 * Allow skipping a full rebuild for incremental assembly
2957 * of a clean array, like RAID1 does.
2958 */
2968 }
2970 skipped:
2979 /* If we aborted, we need to abort the
2980 * sync on the 'current' bitmap chucks (there can
2981 * be several when recovering multiple devices).
2982 * as we may have started syncing it but not finished.
2983 * We can find the current address in
2984 * mddev->curr_resync, but for recovery,
2985 * we need to convert that to several
2986 * virtual addresses.
2987 */
2992 }
2999 sector_t sect =
3003 }
3005 /* completed sync */
3009 /* Completed a full sync so the replacements
3010 * are now fully recovered.
3011 */
3018 }
3020 }
3022 }
3027 }
3033 /* if there has been nothing to do on any drive,
3034 * then there is nothing to do at all..
3035 */
3038 }
3043 /* make sure whole request will fit in a chunk - if chunks
3044 * are meaningful
3045 */
3050 /*
3051 * If there is non-resync activity waiting for a turn, then let it
3052 * though before starting on this new sync request.
3053 */
3057 /* Again, very different code for resync and recovery.
3058 * Both must result in an r10bio with a list of bios that
3059 * have bi_end_io, bi_sector, bi_disk set,
3060 * and bi_private set to the r10bio.
3061 * For recovery, we may actually create several r10bios
3062 * with 2 bios in each, that correspond to the bios in the main one.
3063 * In this case, the subordinate r10bios link back through a
3064 * borrowed master_bio pointer, and the counter in the master
3065 * includes a ref from each subordinate.
3066 */
3067 /* First, we decide what to do and set ->bi_end_io
3068 * To end_sync_read if we want to read, and
3069 * end_sync_write if we will want to write.
3070 */
3074 /* recovery... the complicated one */
3081 sector_t sect;
3104 }
3107 /* want to reconstruct this device */
3111 /* last stripe is not complete - don't
3112 * try to recover this sector.
3113 */
3116 }
3119 /* Unless we are doing a full sync, or a replacement
3120 * we only need to recover the block if it is set in
3121 * the bitmap
3122 */
3130 /* yep, skip the sync_blocks here, but don't assume
3131 * that there will never be anything to do here
3132 */
3136 }
3156 /* Need to check if the array will still be
3157 * degraded
3158 */
3166 }
3167 }
3184 /* This is where we read from */
3193 bad_sectors -= (sector
3198 }
3199 }
3212 /* and we write to 'i' (if not in_sync) */
3237 /* and maybe write to replacement */
3241 /* Note: if need_replace, then bio
3242 * cannot be NULL as r10buf_pool_alloc will
3243 * have allocated it.
3244 */
3256 }
3259 /* Cannot recover, so abort the recovery or
3260 * record a bad block */
3262 /* problem is that there are bad blocks
3263 * on other device(s)
3264 */
3272 mrdev,
3278 mreplace,
3282 }
3288 mirror->recovery_disabled
3290 }
3299 }
3304 /* Only want this if there is elsewhere to
3305 * read from. 'j' is currently the first
3306 * readable copy.
3307 */
3314 targets++;
3315 }
3319 }
3320 }
3327 }
3329 }
3331 /* resync. Schedule a read for every block at this virt offset */
3334 /*
3335 * Since curr_resync_completed could probably not update in
3336 * time, and we will set cluster_sync_low based on it.
3337 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3338 * safety reason, which ensures curr_resync_completed is
3339 * updated in bitmap_cond_end_sync.
3340 */
3349 /* We can skip this block */
3352 }
3385 }
3397 }
3398 }
3409 count++;
3415 }
3418 /* Need to set up for writing to the replacement */
3431 count++;
3433 }
3440 mddev);
3445 mddev);
3446 }
3450 }
3451 }
3466 /*
3467 * won't fail because the vec table is big enough
3468 * to hold all these pages
3469 */
3471 }
3479 /* It is resync not recovery */
3483 /* Send resync message */
3487 }
3489 /* This is recovery not resync */
3494 /*
3495 * sector_nr is a device address for recovery, so we
3496 * need translate it to array address before compare
3497 * with cluster_sync_high.
3498 */
3503 /*
3504 * curr_resync_completed is similar as
3505 * sector_nr, so make the translation too.
3506 */
3513 }
3514 }
3520 }
3521 }
3535 }
3536 }
3539 /* pretend they weren't skipped, it makes
3540 * no important difference in this case
3541 */
3545 giveup:
3546 /* There is nowhere to write, so all non-sync
3547 * drives must be failed or in resync, all drives
3548 * have a bad block, so try the next chunk...
3549 */
3554 chunks_skipped ++;
3557 }
3559 static sector_t
3561 {
3562 sector_t size;
3577 }
3580 {
3581 /* Calculate the number of sectors-per-device that will
3582 * actually be used, and set conf->dev_sectors and
3583 * conf->stride
3584 */
3590 /* 'size' is now the number of chunks in the array */
3591 /* calculate "used chunks per device" */
3594 /* We need to round up when dividing by raid_disks to
3595 * get the stride size.
3596 */
3606 }
3607 }
3611 {
3627 * updated yet. */
3632 }
3650 * actually using this, but leave code here just in case.*/
3660 }
3664 }
3667 {
3679 }
3685 }
3692 /* FIXME calc properly */
3695 GFP_KERNEL);
3722 }
3726 else
3727 /* far_copies must be 1 */
3729 }
3747 out:
3754 }
3756 }
3759 {
3764 sector_t size;
3777 }
3791 }
3792 }
3806 else
3809 }
3830 }
3848 }
3854 else
3857 }
3858 /* need to check that every block has at least one working mirror */
3863 }
3866 /* must ensure that shape change is supported */
3873 }
3879 i++) {
3884 /* The replacement is all we have - use it */
3888 }
3897 }
3903 }
3906 }
3914 /*
3915 * Ok, everything is just fine now
3916 */
3927 /* Calculate max read-ahead size.
3928 * We need to readahead at least twice a whole stripe....
3929 * maybe...
3930 */
3934 }
3948 /* This cannot work */
3951 }
3960 }
3964 out_free_conf:
3971 out:
3973 }
3976 {
3986 }
3989 {
3994 else
3996 }
3999 {
4000 /* Resize of 'far' arrays is not supported.
4001 * For 'near' and 'offset' arrays we can set the
4002 * number of sectors used to be an appropriate multiple
4003 * of the chunk size.
4004 * For 'offset', this is far_copies*chunksize.
4005 * For 'near' the multiplier is the LCM of
4006 * near_copies and raid_disks.
4007 * So if far_copies > 1 && !far_offset, fail.
4008 * Else find LCM(raid_disks, near_copy)*far_copies and
4009 * multiply by chunk_size. Then round to this number.
4010 * This is mostly done by raid10_size()
4011 */
4030 }
4036 }
4041 }
4044 {
4052 }
4055 /* Set new parameters */
4057 /* new layout: far_copies = 1, near_copies = 2 */
4062 /* make sure it will be not marked as dirty */
4072 }
4074 }
4077 }
4080 {
4083 /* raid10 can take over:
4084 * raid0 - providing it has only two drives
4085 */
4087 /* for raid0 takeover only one zone is supported */
4093 }
4097 }
4099 }
4102 {
4103 /* Called when there is a request to change
4104 * - layout (to ->new_layout)
4105 * - chunk size (to ->new_chunk_sectors)
4106 * - raid_disks (by delta_disks)
4107 * or when trying to restart a reshape that was ongoing.
4108 *
4109 * We need to validate the request and possibly allocate
4110 * space if that might be an issue later.
4111 *
4112 * Currently we reject any reshape of a 'far' mode array,
4113 * allow chunk size to change if new is generally acceptable,
4114 * allow raid_disks to increase, and allow
4115 * a switch between 'near' mode and 'offset' mode.
4116 */
4124 /* mustn't change number of copies */
4127 /* Cannot switch to 'far' mode */
4131 /* not factor of array size */
4140 /* allocate new 'mirrors' list */
4144 GFP_KERNEL);
4147 }
4149 }
4151 /*
4152 * Need to check if array has failed when deciding whether to:
4153 * - start an array
4154 * - remove non-faulty devices
4155 * - add a spare
4156 * - allow a reshape
4157 * This determination is simple when no reshape is happening.
4158 * However if there is a reshape, we need to carefully check
4159 * both the before and after sections.
4160 * This is because some failed devices may only affect one
4161 * of the two sections, and some non-in_sync devices may
4162 * be insync in the section most affected by failed devices.
4163 */
4165 {
4171 /* 'prev' section first */
4175 degraded++;
4177 /* When we can reduce the number of devices in
4178 * an array, this might not contribute to
4179 * 'degraded'. It does now.
4180 */
4181 degraded++;
4182 }
4191 degraded2++;
4193 /* If reshape is increasing the number of devices,
4194 * this section has already been recovered, so
4195 * it doesn't contribute to degraded.
4196 * else it does.
4197 */
4199 degraded2++;
4200 }
4201 }
4206 }
4209 {
4210 /* A 'reshape' has been requested. This commits
4211 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4212 * This also checks if there are enough spares and adds them
4213 * to the array.
4214 * We currently require enough spares to make the final
4215 * array non-degraded. We also require that the difference
4216 * between old and new data_offset - on each device - is
4217 * enough that we never risk over-writing.
4218 */
4243 spares++;
4254 }
4255 }
4273 }
4283 }
4302 else
4304 }
4310 }
4312 /*
4313 * some node is already performing reshape, and no need to
4314 * call md_bitmap_resize again since it should be called when
4315 * receiving BITMAP_RESIZE msg
4316 */
4329 }
4330 }
4331 out:
4340 else
4345 }
4348 /* This is a spare that was manually added */
4350 }
4351 }
4352 /* When a reshape changes the number of devices,
4353 * ->degraded is measured against the larger of the
4354 * pre and post numbers.
4355 */
4374 }
4380 abort:
4393 }
4395 /* Calculate the last device-address that could contain
4396 * any block from the chunk that includes the array-address 's'
4397 * and report the next address.
4398 * i.e. the address returned will be chunk-aligned and after
4399 * any data that is in the chunk containing 's'.
4400 */
4402 {
4410 }
4412 /* Calculate the first device-address that could contain
4413 * any block from the chunk that includes the array-address 's'.
4414 * This too will be the start of a chunk
4415 */
4417 {
4424 }
4428 {
4429 /* We simply copy at most one chunk (smallest of old and new)
4430 * at a time, possibly less if that exceeds RESYNC_PAGES,
4431 * or we hit a bad block or something.
4432 * This might mean we pause for normal IO in the middle of
4433 * a chunk, but that is not a problem as mddev->reshape_position
4434 * can record any location.
4435 *
4436 * If we will want to write to a location that isn't
4437 * yet recorded as 'safe' (i.e. in metadata on disk) then
4438 * we need to flush all reshape requests and update the metadata.
4439 *
4440 * When reshaping forwards (e.g. to more devices), we interpret
4441 * 'safe' as the earliest block which might not have been copied
4442 * down yet. We divide this by previous stripe size and multiply
4443 * by previous stripe length to get lowest device offset that we
4444 * cannot write to yet.
4445 * We interpret 'sector_nr' as an address that we want to write to.
4446 * From this we use last_device_address() to find where we might
4447 * write to, and first_device_address on the 'safe' position.
4448 * If this 'next' write position is after the 'safe' position,
4449 * we must update the metadata to increase the 'safe' position.
4450 *
4451 * When reshaping backwards, we round in the opposite direction
4452 * and perform the reverse test: next write position must not be
4453 * less than current safe position.
4454 *
4455 * In all this the minimum difference in data offsets
4456 * (conf->offset_diff - always positive) allows a bit of slack,
4457 * so next can be after 'safe', but not by more than offset_diff
4458 *
4459 * We need to prepare all the bios here before we start any IO
4460 * to ensure the size we choose is acceptable to all devices.
4461 * The means one for each copy for write-out and an extra one for
4462 * read-in.
4463 * We store the read-in bio in ->master_bio and the others in
4464 * ->devs[x].bio and ->devs[x].repl_bio.
4465 */
4480 /* If restarting in the middle, skip the initial sectors */
4493 }
4494 }
4496 /* We don't use sector_nr to track where we are up to
4497 * as that doesn't work well for ->reshape_backwards.
4498 * So just use ->reshape_progress.
4499 */
4501 /* 'next' is the earliest device address that we might
4502 * write to for this chunk in the new layout
4503 */
4507 /* 'safe' is the last device address that we might read from
4508 * in the old layout after a restart
4509 */
4522 /* 'next' is after the last device address that we
4523 * might write to for this chunk in the new layout
4524 */
4527 /* 'safe' is the earliest device address that we might
4528 * read from in the old layout after a restart
4529 */
4532 /* Need to update metadata if 'next' might be beyond 'safe'
4533 * as that would possibly corrupt data
4534 */
4544 }
4548 /* Need to update reshape_position in metadata */
4554 else
4564 }
4567 }
4570 read_more:
4571 /* Now schedule reads for blocks from sector_nr to last */
4584 /* Cannot read from here, so need to record bad blocks
4585 * on all the target devices.
4586 */
4587 // FIXME
4591 }
4608 /*
4609 * Broadcast RESYNC message to other nodes, so all nodes would not
4610 * write to the region to avoid conflict.
4611 */
4621 /*
4622 * Set cluster_sync_low again if next address for array
4623 * reshape is less than cluster_sync_low. Since we can't
4624 * update cluster_sync_low until it has finished reshape.
4625 */
4628 }
4632 }
4634 /* Now find the locations in the new layout */
4651 }
4662 }
4664 /* Now add as many pages as possible to all of these bios. */
4674 /*
4675 * won't fail because the vec table is big enough
4676 * to hold all these pages
4677 */
4679 }
4682 }
4686 /* Now submit the read */
4698 /* Now that we have done the whole section we can
4699 * update reshape_progress
4700 */
4703 else
4707 }
4713 {
4714 /* Reshape read completed. Hopefully we have a block
4715 * to write out.
4716 * If we got a read error then we do sync 1-page reads from
4717 * elsewhere until we find the data - or give up.
4718 */
4724 /* Reshape has been aborted */
4727 }
4729 /* We definitely have the data in the pages, schedule the
4730 * writes.
4731 */
4744 }
4748 }
4755 }
4757 }
4760 {
4772 /* read-ahead size must cover two whole stripes, which is
4773 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4774 */
4781 }
4783 }
4786 {
4794 else
4796 }
4800 {
4801 /* Use sync reads to get the blocks from somewhere else */
4814 }
4816 /* reshape IOs share pages from .devs[0].bio */
4834 sector_t addr;
4844 addr,
4852 failed:
4853 slot++;
4858 }
4861 /* couldn't read this block, must give up */
4866 }
4868 idx++;
4869 }
4872 }
4875 {
4890 }
4893 /* FIXME should record badblock */
4895 }
4899 }
4902 {
4908 }
4911 {
4921 }
4928 d++) {
4935 }
4937 }
4943 }
4946 {
4968 };
4971 {
4973 }
4976 {
4978 }