| blob | ec136e44aef7f89067a3320878f49d94c108d8a5 |
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * raid10.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 2000-2004 Neil Brown
6 *
7 * RAID-10 support for md.
8 *
9 * Base on code in raid1.c. See raid1.c for further copyright information.
10 */
12 #include <linux/slab.h>
13 #include <linux/delay.h>
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/ratelimit.h>
18 #include <linux/kthread.h>
19 #include <linux/raid/md_p.h>
20 #include <trace/events/block.h>
26 /*
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
29 * chunk_size
30 * raid_disks
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
34 * use_far_sets (stored in bit 17 of layout )
35 * use_far_sets_bugfixed (stored in bit 18 of layout )
36 *
37 * The data to be stored is divided into chunks using chunksize. Each device
38 * is divided into far_copies sections. In each section, chunks are laid out
39 * in a style similar to raid0, but near_copies copies of each chunk is stored
40 * (each on a different drive). The starting device for each section is offset
41 * near_copies from the starting device of the previous section. Thus there
42 * are (near_copies * far_copies) of each chunk, and each is on a different
43 * drive. near_copies and far_copies must be at least one, and their product
44 * is at most raid_disks.
45 *
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of being very far
48 * apart on disk, there are adjacent stripes.
49 *
50 * The far and offset algorithms are handled slightly differently if
51 * 'use_far_sets' is true. In this case, the array's devices are grouped into
52 * sets that are (near_copies * far_copies) in size. The far copied stripes
53 * are still shifted by 'near_copies' devices, but this shifting stays confined
54 * to the set rather than the entire array. This is done to improve the number
55 * of device combinations that can fail without causing the array to fail.
56 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
57 * on a device):
58 * A B C D A B C D E
59 * ... ...
60 * D A B C E A B C D
61 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
62 * [A B] [C D] [A B] [C D E]
63 * |...| |...| |...| | ... |
64 * [B A] [D C] [B A] [E C D]
65 */
77 #define raid10_log(md, fmt, args...) \
82 /*
83 * for resync bio, r10bio pointer can be retrieved from the per-bio
84 * 'struct resync_pages'.
85 */
87 {
89 }
92 {
96 /* allocate a r10bio with room for raid_disks entries in the
97 * bios array */
99 }
101 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
102 /* amount of memory to reserve for resync requests */
103 #define RESYNC_WINDOW (1024*1024)
104 /* maximum number of concurrent requests, memory permitting */
105 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
106 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
107 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
109 /*
110 * When performing a resync, we need to read and compare, so
111 * we need as many pages are there are copies.
112 * When performing a recovery, we need 2 bios, one for read,
113 * one for write (we recover only one drive per r10buf)
114 *
115 */
117 {
132 else
135 /* allocate once for all bios */
138 else
144 /*
145 * Allocate bios.
146 */
158 }
159 /*
160 * Allocate RESYNC_PAGES data pages and attach them
161 * where needed.
162 */
180 }
187 }
188 }
192 out_free_pages:
197 out_free_bio:
203 }
205 out_free_r10bio:
208 }
211 {
224 }
229 }
231 /* resync pages array stored in the 1st bio's .bi_private */
235 }
238 {
250 }
251 }
254 {
259 }
262 {
268 }
271 {
281 /* wake up frozen array... */
285 }
287 /*
288 * raid_end_bio_io() is called when we have finished servicing a mirrored
289 * operation and are ready to return a success/failure code to the buffer
290 * cache layer.
291 */
293 {
301 /*
302 * Wake up any possible resync thread that waits for the device
303 * to go idle.
304 */
308 }
310 /*
311 * Update disk head position estimator based on IRQ completion info.
312 */
314 {
319 }
321 /*
322 * Find the disk number which triggered given bio
323 */
326 {
336 }
337 }
347 }
350 {
359 /*
360 * this branch is our 'one mirror IO has finished' event handler:
361 */
365 /*
366 * Set R10BIO_Uptodate in our master bio, so that
367 * we will return a good error code to the higher
368 * levels even if IO on some other mirrored buffer fails.
369 *
370 * The 'master' represents the composite IO operation to
371 * user-side. So if something waits for IO, then it will
372 * wait for the 'master' bio.
373 */
376 /* If all other devices that store this block have
377 * failed, we want to return the error upwards rather
378 * than fail the last device. Here we redefine
379 * "uptodate" to mean "Don't want to retry"
380 */
384 }
389 /*
390 * oops, read error - keep the refcount on the rdev
391 */
399 }
400 }
403 {
404 /* clear the bitmap if all writes complete successfully */
410 }
413 {
421 else
423 }
424 }
425 }
428 {
448 }
449 /*
450 * this branch is our 'one mirror IO has finished' event handler:
451 */
454 /* Never record new bad blocks to replacement,
455 * just fail it.
456 */
468 }
470 /*
471 * When the device is faulty, it is not necessary to
472 * handle write error.
473 * For failfast, this is the only remaining device,
474 * We need to retry the write without FailFast.
475 */
482 }
483 }
485 /*
486 * Set R10BIO_Uptodate in our master bio, so that
487 * we will return a good error code for to the higher
488 * levels even if IO on some other mirrored buffer fails.
489 *
490 * The 'master' represents the composite IO operation to
491 * user-side. So if something waits for IO, then it will
492 * wait for the 'master' bio.
493 */
494 sector_t first_bad;
497 /*
498 * Do not set R10BIO_Uptodate if the current device is
499 * rebuilding or Faulty. This is because we cannot use
500 * such device for properly reading the data back (we could
501 * potentially use it, if the current write would have felt
502 * before rdev->recovery_offset, but for simplicity we don't
503 * check this here.
504 */
509 /* Maybe we can clear some bad blocks. */
517 else
521 }
522 }
524 /*
525 *
526 * Let's see if all mirrored write operations have finished
527 * already.
528 */
534 }
536 /*
537 * RAID10 layout manager
538 * As well as the chunksize and raid_disks count, there are two
539 * parameters: near_copies and far_copies.
540 * near_copies * far_copies must be <= raid_disks.
541 * Normally one of these will be 1.
542 * If both are 1, we get raid0.
543 * If near_copies == raid_disks, we get raid1.
544 *
545 * Chunks are laid out in raid0 style with near_copies copies of the
546 * first chunk, followed by near_copies copies of the next chunk and
547 * so on.
548 * If far_copies > 1, then after 1/far_copies of the array has been assigned
549 * as described above, we start again with a device offset of near_copies.
550 * So we effectively have another copy of the whole array further down all
551 * the drives, but with blocks on different drives.
552 * With this layout, and block is never stored twice on the one device.
553 *
554 * raid10_find_phys finds the sector offset of a given virtual sector
555 * on each device that it is on.
556 *
557 * raid10_find_virt does the reverse mapping, from a device and a
558 * sector offset to a virtual address
559 */
562 {
564 sector_t sector;
565 sector_t chunk;
566 sector_t stripe;
577 /* now calculate first sector/dev */
589 /* and calculate all the others */
596 slot++;
610 }
614 slot++;
615 }
616 dev++;
620 }
621 }
622 }
625 {
637 }
640 {
642 /* Never use conf->prev as this is only called during resync
643 * or recovery, so reshape isn't happening
644 */
658 }
659 }
674 else
676 }
678 }
682 }
684 /*
685 * This routine returns the disk from which the requested read should
686 * be done. There is a per-array 'next expected sequential IO' sector
687 * number - if this matches on the next IO then we use the last disk.
688 * There is also a per-disk 'last know head position' sector that is
689 * maintained from IRQ contexts, both the normal and the resync IO
690 * completion handlers update this position correctly. If there is no
691 * perfect sequential match then we pick the disk whose head is closest.
692 *
693 * If there are 2 mirrors in the same 2 devices, performance degrades
694 * because position is mirror, not device based.
695 *
696 * The rdev for the device selected will have nr_pending incremented.
697 */
699 /*
700 * FIXME: possibly should rethink readbalancing and do it differently
701 * depending on near_copies / far_copies geometry.
702 */
706 {
729 /*
730 * Check if we can balance. We can balance on the whole
731 * device if no resync is going on (recovery is ok), or below
732 * the resync window. We take the first readable disk when
733 * above the resync window.
734 */
743 sector_t first_bad;
745 sector_t dev_sector;
767 /* Already have a better slot */
770 /* Cannot read here. If this is the
771 * 'primary' device, then we must not read
772 * beyond 'bad_sectors' from another device.
773 */
780 sector_t good_sectors =
786 }
788 /* Must read from here */
790 }
805 }
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
828 }
829 }
837 }
838 }
849 }
852 {
864 i++) {
870 }
871 }
874 }
877 {
878 /* Any writes that have been queued but are awaiting
879 * bitmap updates get flushed here.
880 */
891 /*
892 * As this is called in a wait_event() loop (see freeze_array),
893 * current->state might be TASK_UNINTERRUPTIBLE which will
894 * cause a warning when we prepare to wait again. As it is
895 * rare that this path is taken, it is perfectly safe to force
896 * us to go around the wait_event() loop again, so the warning
897 * is a false-positive. Silence the warning by resetting
898 * thread state
899 */
903 /* flush any pending bitmap writes to disk
904 * before proceeding w/ I/O */
917 /* Just ignore it */
919 else
922 }
926 }
928 /* Barriers....
929 * Sometimes we need to suspend IO while we do something else,
930 * either some resync/recovery, or reconfigure the array.
931 * To do this we raise a 'barrier'.
932 * The 'barrier' is a counter that can be raised multiple times
933 * to count how many activities are happening which preclude
934 * normal IO.
935 * We can only raise the barrier if there is no pending IO.
936 * i.e. if nr_pending == 0.
937 * We choose only to raise the barrier if no-one is waiting for the
938 * barrier to go down. This means that as soon as an IO request
939 * is ready, no other operations which require a barrier will start
940 * until the IO request has had a chance.
941 *
942 * So: regular IO calls 'wait_barrier'. When that returns there
943 * is no backgroup IO happening, It must arrange to call
944 * allow_barrier when it has finished its IO.
945 * backgroup IO calls must call raise_barrier. Once that returns
946 * there is no normal IO happeing. It must arrange to call
947 * lower_barrier when the particular background IO completes.
948 */
951 {
955 /* Wait until no block IO is waiting (unless 'force') */
959 /* block any new IO from starting */
962 /* Now wait for all pending IO to complete */
968 }
971 {
977 }
980 {
984 /* Wait for the barrier to drop.
985 * However if there are already pending
986 * requests (preventing the barrier from
987 * rising completely), and the
988 * pre-process bio queue isn't empty,
989 * then don't wait, as we need to empty
990 * that queue to get the nr_pending
991 * count down.
992 */
1004 }
1007 }
1010 {
1014 }
1017 {
1018 /* stop syncio and normal IO and wait for everything to
1019 * go quiet.
1020 * We increment barrier and nr_waiting, and then
1021 * wait until nr_pending match nr_queued+extra
1022 * This is called in the context of one normal IO request
1023 * that has failed. Thus any sync request that might be pending
1024 * will be blocked by nr_pending, and we need to wait for
1025 * pending IO requests to complete or be queued for re-try.
1026 * Thus the number queued (nr_queued) plus this request (extra)
1027 * must match the number of pending IOs (nr_pending) before
1028 * we continue.
1029 */
1041 }
1044 {
1045 /* reverse the effect of the freeze */
1051 }
1055 {
1059 else
1061 }
1067 };
1070 {
1072 cb);
1086 }
1088 /* we aren't scheduling, so we can do the write-out directly. */
1102 /* Just ignore it */
1104 else
1107 }
1109 }
1111 /*
1112 * 1. Register the new request and wait if the reconstruction thread has put
1113 * up a bar for new requests. Continue immediately if no resync is active
1114 * currently.
1115 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1116 */
1119 {
1129 sectors);
1131 }
1132 }
1136 {
1149 /*
1150 * This is an error retry, but we cannot
1151 * safely dereference the rdev in the r10_bio,
1152 * we must use the one in conf.
1153 * If it has already been disconnected (unlikely)
1154 * we lose the device name in error messages.
1155 */
1157 /*
1158 * As we are blocking raid10, it is a little safer to
1159 * use __GFP_HIGH.
1160 */
1170 /* This never gets dereferenced */
1172 }
1174 }
1183 }
1186 }
1202 }
1226 }
1231 {
1246 /* Replacement just got moved to main 'rdev' */
1249 }
1256 else
1274 /* flush_pending_writes() needs access to the rdev so...*/
1282 else
1293 }
1294 }
1298 {
1302 sector_t sectors;
1317 }
1319 }
1329 /* Need to update reshape_position in metadata */
1339 }
1346 }
1347 /* first select target devices under rcu_lock and
1348 * inc refcount on their rdev. Record them by setting
1349 * bios[x] to bio
1350 * If there are known/acknowledged bad blocks on any device
1351 * on which we have seen a write error, we want to avoid
1352 * writing to those blocks. This potentially requires several
1353 * writes to write around the bad blocks. Each set of writes
1354 * gets its own r10_bio with a set of bios attached.
1355 */
1359 retry_write:
1375 }
1380 }
1392 }
1394 sector_t first_bad;
1402 /* Mustn't write here until the bad block
1403 * is acknowledged
1404 */
1409 }
1411 /* Cannot write here at all */
1414 /* Mustn't write more than bad_sectors
1415 * to other devices yet
1416 */
1418 /* We don't set R10BIO_Degraded as that
1419 * only applies if the disk is missing,
1420 * so it might be re-added, and we want to
1421 * know to recover this chunk.
1422 * In this case the device is here, and the
1423 * fact that this chunk is not in-sync is
1424 * recorded in the bad block log.
1425 */
1427 }
1432 }
1433 }
1437 }
1441 }
1442 }
1446 /* Have to wait for this device to get unblocked, then retry */
1454 }
1460 /* Race with remove_disk */
1463 }
1465 }
1466 }
1472 }
1486 }
1496 }
1498 }
1501 {
1517 else
1519 }
1522 {
1535 /*
1536 * If this request crosses a chunk boundary, we need to split
1537 * it.
1538 */
1540 sectors > chunk_sects
1549 /* In case raid10d snuck in to freeze_array */
1552 }
1555 {
1566 else
1570 }
1577 }
1580 }
1582 /* check if there are enough drives for
1583 * every block to appear on atleast one.
1584 * Don't consider the device numbered 'ignore'
1585 * as we might be about to remove it.
1586 */
1588 {
1598 }
1610 cnt++;
1612 }
1618 out:
1621 }
1624 {
1625 /* when calling 'enough', both 'prev' and 'geo' must
1626 * be stable.
1627 * This is ensured if ->reconfig_mutex or ->device_lock
1628 * is held.
1629 */
1632 }
1635 {
1640 /*
1641 * If it is not operational, then we have already marked it as dead
1642 * else if it is the last working disks with "fail_last_dev == false",
1643 * ignore the error, let the next level up know.
1644 * else mark the drive as failed
1645 */
1649 /*
1650 * Don't fail the drive, just return an IO error.
1651 */
1654 }
1657 /*
1658 * If recovery is running, make sure it aborts.
1659 */
1670 }
1673 {
1681 }
1685 /* This is only called with ->reconfix_mutex held, so
1686 * rcu protection of rdev is not needed */
1695 }
1696 }
1699 {
1704 }
1707 {
1714 /*
1715 * Find all non-in_sync disks within the RAID10 configuration
1716 * and mark them in_sync
1717 */
1724 /* Replacement has just become active */
1727 count++;
1729 /* Replaced device not technically faulty,
1730 * but we need to be sure it gets removed
1731 * and never re-added.
1732 */
1736 }
1742 count++;
1744 }
1745 }
1752 }
1755 {
1763 /* only hot-add to in-sync arrays, as recovery is
1764 * very different from resync
1765 */
1780 else
1800 }
1814 }
1820 }
1823 {
1835 else
1842 }
1843 /* Only remove non-faulty devices if recovery
1844 * is not possible.
1845 */
1853 }
1858 /* lost the race, try later */
1862 }
1863 }
1865 /* We must have just cleared 'rdev' */
1869 * but will never see neither -- if they are careful.
1870 */
1872 }
1877 abort:
1881 }
1884 {
1889 else
1890 /* The write handler will notice the lack of
1891 * R10BIO_Uptodate and record any errors etc
1892 */
1896 /* for reconstruct, we always reschedule after a read.
1897 * for resync, only after all reads
1898 */
1902 /* we have read all the blocks,
1903 * do the comparison in process context in raid10d
1904 */
1906 }
1907 }
1910 {
1916 }
1919 {
1920 /* reshape read bio isn't allocated from r10buf_pool */
1924 }
1927 {
1932 /* the primary of several recovery bios */
1937 else
1946 else
1949 }
1950 }
1951 }
1954 {
1959 sector_t first_bad;
1968 else
1980 }
1990 }
1992 /*
1993 * Note: sync and recover and handled very differently for raid10
1994 * This code is for resync.
1995 * For resync, we read through virtual addresses and read all blocks.
1996 * If there is any error, we schedule a write. The lowest numbered
1997 * drive is authoritative.
1998 * However requests come for physical address, so we need to map.
1999 * For every physical address there are raid_disks/copies virtual addresses,
2000 * which is always are least one, but is not necessarly an integer.
2001 * This means that a physical address can span multiple chunks, so we may
2002 * have to submit multiple io requests for a single sync request.
2003 */
2004 /*
2005 * We check if all blocks are in-sync and only write to blocks that
2006 * aren't in sync
2007 */
2009 {
2018 /* find the first device with a block */
2033 /* now find blocks with errors */
2050 /* We know that the bi_io_vec layout is the same for
2051 * both 'first' and 'i', so we just compare them.
2052 * All vec entries are PAGE_SIZE;
2053 */
2061 len))
2064 }
2069 /* Don't fix anything. */
2072 /* Just give up on this device */
2075 }
2076 /* Ok, we need to write this bio, either to correct an
2077 * inconsistency or to correct an unreadable block.
2078 * First we need to fixup bv_offset, bv_len and
2079 * bi_vecs, as the read request might have corrupted these
2080 */
2103 }
2105 /* Now write out to any replacement devices
2106 * that are active
2107 */
2122 }
2124 done:
2128 }
2129 }
2131 /*
2132 * Now for the recovery code.
2133 * Recovery happens across physical sectors.
2134 * We recover all non-is_sync drives by finding the virtual address of
2135 * each, and then choose a working drive that also has that virt address.
2136 * There is a separate r10_bio for each non-in_sync drive.
2137 * Only the first two slots are in use. The first for reading,
2138 * The second for writing.
2139 *
2140 */
2142 {
2143 /* We got a read error during recovery.
2144 * We repeat the read in smaller page-sized sections.
2145 * If a read succeeds, write it to the new device or record
2146 * a bad block if we cannot.
2147 * If a read fails, record a bad block on both old and
2148 * new devices.
2149 */
2163 sector_t addr;
2172 addr,
2180 addr,
2190 }
2191 }
2193 /* We don't worry if we cannot set a bad block -
2194 * it really is bad so there is no loss in not
2195 * recording it yet
2196 */
2200 /* need bad block on destination too */
2205 /* just abort the recovery */
2214 }
2215 }
2216 }
2220 idx++;
2221 }
2222 }
2225 {
2234 }
2236 /*
2237 * share the pages with the first bio
2238 * and submit the write request
2239 */
2243 /* Need to test wbio2->bi_end_io before we call
2244 * generic_make_request as if the former is NULL,
2245 * the latter is free to free wbio2.
2246 */
2253 }
2259 }
2260 }
2262 /*
2263 * Used by fix_read_error() to decay the per rdev read_errors.
2264 * We halve the read error count for every hour that has elapsed
2265 * since the last recorded read error.
2266 *
2267 */
2269 {
2277 /* first time we've seen a read error */
2280 }
2287 /*
2288 * if hours_since_last is > the number of bits in read_errors
2289 * just set read errors to 0. We do this to avoid
2290 * overflowing the shift of read_errors by hours_since_last.
2291 */
2294 else
2296 }
2300 {
2301 sector_t first_bad;
2308 /* success */
2315 }
2316 /* need to record an error - either for the block or the device */
2320 }
2322 /*
2323 * This is a kernel thread which:
2324 *
2325 * 1. Retries failed read operations on working mirrors.
2326 * 2. Updates the raid superblock when problems encounter.
2327 * 3. Performs writes following reads for array synchronising.
2328 */
2331 {
2338 /* still own a reference to this rdev, so it cannot
2339 * have been cleared recently.
2340 */
2344 /* drive has already been failed, just ignore any
2345 more fix_read_error() attempts */
2362 }
2375 sector_t first_bad;
2389 sect,
2397 }
2398 sl++;
2405 /* Cannot read from anywhere, just mark the block
2406 * as bad on the first device to discourage future
2407 * reads.
2408 */
2413 rdev,
2420 }
2422 }
2425 /* write it back and re-read */
2432 sl--;
2444 sect,
2447 /* Well, this device is dead */
2451 sect +
2453 rdev)),
2458 }
2461 }
2468 sl--;
2480 sect,
2482 READ)) {
2484 /* Well, this device is dead */
2488 sect +
2499 sect +
2503 }
2507 }
2512 }
2513 }
2516 {
2521 /* bio has the data to be written to slot 'i' where
2522 * we just recently had a write error.
2523 * We repeatedly clone the bio and trim down to one block,
2524 * then try the write. Where the write fails we record
2525 * a bad block.
2526 * It is conceivable that the bio doesn't exactly align with
2527 * blocks. We must handle this.
2528 *
2529 * We currently own a reference to the rdev.
2530 */
2533 sector_t sector;
2550 sector_t wsector;
2553 /* Write at 'sector' for 'sectors' */
2563 /* Failure! */
2572 }
2574 }
2577 {
2583 /* we got a read error. Maybe the drive is bad. Maybe just
2584 * the block and we can fix it.
2585 * We freeze all other IO, and try reading the block from
2586 * other devices. When we find one, we re-write
2587 * and check it that fixes the read error.
2588 * This is all done synchronously while the array is
2589 * frozen.
2590 */
2608 }
2611 {
2612 /* Some sort of write request has finished and it
2613 * succeeded in writing where we thought there was a
2614 * bad block. So forget the bad block.
2615 * Or possibly if failed and we need to record
2616 * a bad block.
2617 */
2631 rdev,
2636 rdev,
2640 }
2648 rdev,
2653 rdev,
2657 }
2658 }
2668 rdev,
2678 }
2680 }
2685 rdev,
2689 }
2690 }
2696 /*
2697 * In case freeze_array() is waiting for condition
2698 * nr_pending == nr_queued + extra to be true.
2699 */
2707 }
2708 }
2709 }
2712 {
2730 }
2731 }
2735 retry_list);
2744 }
2745 }
2756 }
2775 else
2781 }
2783 }
2786 {
2801 }
2804 {
2814 else
2827 }
2828 }
2830 }
2832 /*
2833 * Set cluster_sync_high since we need other nodes to add the
2834 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2835 */
2837 {
2838 sector_t window_size;
2841 /*
2842 * First, here we define "stripe" as a unit which across
2843 * all member devices one time, so we get chunks by use
2844 * raid_disks / near_copies. Otherwise, if near_copies is
2845 * close to raid_disks, then resync window could increases
2846 * linearly with the increase of raid_disks, which means
2847 * we will suspend a really large IO window while it is not
2848 * necessary. If raid_disks is not divisible by near_copies,
2849 * an extra chunk is needed to ensure the whole "stripe" is
2850 * covered.
2851 */
2856 else
2860 /*
2861 * At least use a 32M window to align with raid1's resync window
2862 */
2867 }
2869 /*
2870 * perform a "sync" on one "block"
2871 *
2872 * We need to make sure that no normal I/O request - particularly write
2873 * requests - conflict with active sync requests.
2874 *
2875 * This is achieved by tracking pending requests and a 'barrier' concept
2876 * that can be installed to exclude normal IO requests.
2877 *
2878 * Resync and recovery are handled very differently.
2879 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2880 *
2881 * For resync, we iterate over virtual addresses, read all copies,
2882 * and update if there are differences. If only one copy is live,
2883 * skip it.
2884 * For recovery, we iterate over physical addresses, read a good
2885 * value for each non-in_sync drive, and over-write.
2886 *
2887 * So, for recovery we may have several outstanding complex requests for a
2888 * given address, one for each out-of-sync device. We model this by allocating
2889 * a number of r10_bio structures, one for each out-of-sync device.
2890 * As we setup these structures, we collect all bio's together into a list
2891 * which we then process collectively to add pages, and then process again
2892 * to pass to generic_make_request.
2893 *
2894 * The r10_bio structures are linked using a borrowed master_bio pointer.
2895 * This link is counted in ->remaining. When the r10_bio that points to NULL
2896 * has its remaining count decremented to 0, the whole complex operation
2897 * is complete.
2898 *
2899 */
2903 {
2910 sector_t sync_blocks;
2920 /*
2921 * Allow skipping a full rebuild for incremental assembly
2922 * of a clean array, like RAID1 does.
2923 */
2933 }
2935 skipped:
2944 /* If we aborted, we need to abort the
2945 * sync on the 'current' bitmap chucks (there can
2946 * be several when recovering multiple devices).
2947 * as we may have started syncing it but not finished.
2948 * We can find the current address in
2949 * mddev->curr_resync, but for recovery,
2950 * we need to convert that to several
2951 * virtual addresses.
2952 */
2957 }
2964 sector_t sect =
2968 }
2970 /* completed sync */
2974 /* Completed a full sync so the replacements
2975 * are now fully recovered.
2976 */
2983 }
2985 }
2987 }
2992 }
2998 /* if there has been nothing to do on any drive,
2999 * then there is nothing to do at all..
3000 */
3003 }
3008 /* make sure whole request will fit in a chunk - if chunks
3009 * are meaningful
3010 */
3015 /*
3016 * If there is non-resync activity waiting for a turn, then let it
3017 * though before starting on this new sync request.
3018 */
3022 /* Again, very different code for resync and recovery.
3023 * Both must result in an r10bio with a list of bios that
3024 * have bi_end_io, bi_sector, bi_disk set,
3025 * and bi_private set to the r10bio.
3026 * For recovery, we may actually create several r10bios
3027 * with 2 bios in each, that correspond to the bios in the main one.
3028 * In this case, the subordinate r10bios link back through a
3029 * borrowed master_bio pointer, and the counter in the master
3030 * includes a ref from each subordinate.
3031 */
3032 /* First, we decide what to do and set ->bi_end_io
3033 * To end_sync_read if we want to read, and
3034 * end_sync_write if we will want to write.
3035 */
3039 /* recovery... the complicated one */
3046 sector_t sect;
3069 }
3072 /* want to reconstruct this device */
3076 /* last stripe is not complete - don't
3077 * try to recover this sector.
3078 */
3081 }
3084 /* Unless we are doing a full sync, or a replacement
3085 * we only need to recover the block if it is set in
3086 * the bitmap
3087 */
3095 /* yep, skip the sync_blocks here, but don't assume
3096 * that there will never be anything to do here
3097 */
3101 }
3121 /* Need to check if the array will still be
3122 * degraded
3123 */
3131 }
3132 }
3149 /* This is where we read from */
3158 bad_sectors -= (sector
3163 }
3164 }
3177 /* and we write to 'i' (if not in_sync) */
3202 /* and maybe write to replacement */
3206 /* Note: if need_replace, then bio
3207 * cannot be NULL as r10buf_pool_alloc will
3208 * have allocated it.
3209 */
3221 }
3224 /* Cannot recover, so abort the recovery or
3225 * record a bad block */
3227 /* problem is that there are bad blocks
3228 * on other device(s)
3229 */
3237 mrdev,
3243 mreplace,
3247 }
3253 mirror->recovery_disabled
3255 }
3264 }
3269 /* Only want this if there is elsewhere to
3270 * read from. 'j' is currently the first
3271 * readable copy.
3272 */
3279 targets++;
3280 }
3284 }
3285 }
3292 }
3294 }
3296 /* resync. Schedule a read for every block at this virt offset */
3299 /*
3300 * Since curr_resync_completed could probably not update in
3301 * time, and we will set cluster_sync_low based on it.
3302 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3303 * safety reason, which ensures curr_resync_completed is
3304 * updated in bitmap_cond_end_sync.
3305 */
3314 /* We can skip this block */
3317 }
3350 }
3362 }
3363 }
3374 count++;
3380 }
3383 /* Need to set up for writing to the replacement */
3396 count++;
3398 }
3405 mddev);
3410 mddev);
3411 }
3415 }
3416 }
3431 /*
3432 * won't fail because the vec table is big enough
3433 * to hold all these pages
3434 */
3436 }
3444 /* It is resync not recovery */
3448 /* Send resync message */
3452 }
3454 /* This is recovery not resync */
3459 /*
3460 * sector_nr is a device address for recovery, so we
3461 * need translate it to array address before compare
3462 * with cluster_sync_high.
3463 */
3468 /*
3469 * curr_resync_completed is similar as
3470 * sector_nr, so make the translation too.
3471 */
3478 }
3479 }
3485 }
3486 }
3500 }
3501 }
3504 /* pretend they weren't skipped, it makes
3505 * no important difference in this case
3506 */
3510 giveup:
3511 /* There is nowhere to write, so all non-sync
3512 * drives must be failed or in resync, all drives
3513 * have a bad block, so try the next chunk...
3514 */
3519 chunks_skipped ++;
3522 }
3524 static sector_t
3526 {
3527 sector_t size;
3542 }
3545 {
3546 /* Calculate the number of sectors-per-device that will
3547 * actually be used, and set conf->dev_sectors and
3548 * conf->stride
3549 */
3555 /* 'size' is now the number of chunks in the array */
3556 /* calculate "used chunks per device" */
3559 /* We need to round up when dividing by raid_disks to
3560 * get the stride size.
3561 */
3571 }
3572 }
3576 {
3592 * updated yet. */
3597 }
3615 * actually using this, but leave code here just in case.*/
3625 }
3629 }
3632 {
3644 }
3650 }
3657 /* FIXME calc properly */
3660 GFP_KERNEL);
3687 }
3691 else
3692 /* far_copies must be 1 */
3694 }
3712 out:
3719 }
3721 }
3724 {
3729 sector_t size;
3742 }
3756 }
3757 }
3771 else
3774 }
3795 }
3813 }
3819 else
3822 }
3823 /* need to check that every block has at least one working mirror */
3828 }
3831 /* must ensure that shape change is supported */
3838 }
3844 i++) {
3849 /* The replacement is all we have - use it */
3853 }
3862 }
3868 }
3871 }
3879 /*
3880 * Ok, everything is just fine now
3881 */
3892 /* Calculate max read-ahead size.
3893 * We need to readahead at least twice a whole stripe....
3894 * maybe...
3895 */
3899 }
3913 /* This cannot work */
3916 }
3927 }
3931 out_free_conf:
3938 out:
3940 }
3943 {
3953 }
3956 {
3961 else
3963 }
3966 {
3967 /* Resize of 'far' arrays is not supported.
3968 * For 'near' and 'offset' arrays we can set the
3969 * number of sectors used to be an appropriate multiple
3970 * of the chunk size.
3971 * For 'offset', this is far_copies*chunksize.
3972 * For 'near' the multiplier is the LCM of
3973 * near_copies and raid_disks.
3974 * So if far_copies > 1 && !far_offset, fail.
3975 * Else find LCM(raid_disks, near_copy)*far_copies and
3976 * multiply by chunk_size. Then round to this number.
3977 * This is mostly done by raid10_size()
3978 */
3997 }
4003 }
4008 }
4011 {
4019 }
4022 /* Set new parameters */
4024 /* new layout: far_copies = 1, near_copies = 2 */
4029 /* make sure it will be not marked as dirty */
4039 }
4041 }
4044 }
4047 {
4050 /* raid10 can take over:
4051 * raid0 - providing it has only two drives
4052 */
4054 /* for raid0 takeover only one zone is supported */
4060 }
4064 }
4066 }
4069 {
4070 /* Called when there is a request to change
4071 * - layout (to ->new_layout)
4072 * - chunk size (to ->new_chunk_sectors)
4073 * - raid_disks (by delta_disks)
4074 * or when trying to restart a reshape that was ongoing.
4075 *
4076 * We need to validate the request and possibly allocate
4077 * space if that might be an issue later.
4078 *
4079 * Currently we reject any reshape of a 'far' mode array,
4080 * allow chunk size to change if new is generally acceptable,
4081 * allow raid_disks to increase, and allow
4082 * a switch between 'near' mode and 'offset' mode.
4083 */
4091 /* mustn't change number of copies */
4094 /* Cannot switch to 'far' mode */
4098 /* not factor of array size */
4107 /* allocate new 'mirrors' list */
4111 GFP_KERNEL);
4114 }
4116 }
4118 /*
4119 * Need to check if array has failed when deciding whether to:
4120 * - start an array
4121 * - remove non-faulty devices
4122 * - add a spare
4123 * - allow a reshape
4124 * This determination is simple when no reshape is happening.
4125 * However if there is a reshape, we need to carefully check
4126 * both the before and after sections.
4127 * This is because some failed devices may only affect one
4128 * of the two sections, and some non-in_sync devices may
4129 * be insync in the section most affected by failed devices.
4130 */
4132 {
4138 /* 'prev' section first */
4142 degraded++;
4144 /* When we can reduce the number of devices in
4145 * an array, this might not contribute to
4146 * 'degraded'. It does now.
4147 */
4148 degraded++;
4149 }
4158 degraded2++;
4160 /* If reshape is increasing the number of devices,
4161 * this section has already been recovered, so
4162 * it doesn't contribute to degraded.
4163 * else it does.
4164 */
4166 degraded2++;
4167 }
4168 }
4173 }
4176 {
4177 /* A 'reshape' has been requested. This commits
4178 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4179 * This also checks if there are enough spares and adds them
4180 * to the array.
4181 * We currently require enough spares to make the final
4182 * array non-degraded. We also require that the difference
4183 * between old and new data_offset - on each device - is
4184 * enough that we never risk over-writing.
4185 */
4210 spares++;
4221 }
4222 }
4240 }
4250 }
4269 else
4271 }
4277 }
4279 /*
4280 * some node is already performing reshape, and no need to
4281 * call md_bitmap_resize again since it should be called when
4282 * receiving BITMAP_RESIZE msg
4283 */
4296 }
4297 }
4298 out:
4307 else
4312 }
4315 /* This is a spare that was manually added */
4317 }
4318 }
4319 /* When a reshape changes the number of devices,
4320 * ->degraded is measured against the larger of the
4321 * pre and post numbers.
4322 */
4341 }
4347 abort:
4360 }
4362 /* Calculate the last device-address that could contain
4363 * any block from the chunk that includes the array-address 's'
4364 * and report the next address.
4365 * i.e. the address returned will be chunk-aligned and after
4366 * any data that is in the chunk containing 's'.
4367 */
4369 {
4377 }
4379 /* Calculate the first device-address that could contain
4380 * any block from the chunk that includes the array-address 's'.
4381 * This too will be the start of a chunk
4382 */
4384 {
4391 }
4395 {
4396 /* We simply copy at most one chunk (smallest of old and new)
4397 * at a time, possibly less if that exceeds RESYNC_PAGES,
4398 * or we hit a bad block or something.
4399 * This might mean we pause for normal IO in the middle of
4400 * a chunk, but that is not a problem as mddev->reshape_position
4401 * can record any location.
4402 *
4403 * If we will want to write to a location that isn't
4404 * yet recorded as 'safe' (i.e. in metadata on disk) then
4405 * we need to flush all reshape requests and update the metadata.
4406 *
4407 * When reshaping forwards (e.g. to more devices), we interpret
4408 * 'safe' as the earliest block which might not have been copied
4409 * down yet. We divide this by previous stripe size and multiply
4410 * by previous stripe length to get lowest device offset that we
4411 * cannot write to yet.
4412 * We interpret 'sector_nr' as an address that we want to write to.
4413 * From this we use last_device_address() to find where we might
4414 * write to, and first_device_address on the 'safe' position.
4415 * If this 'next' write position is after the 'safe' position,
4416 * we must update the metadata to increase the 'safe' position.
4417 *
4418 * When reshaping backwards, we round in the opposite direction
4419 * and perform the reverse test: next write position must not be
4420 * less than current safe position.
4421 *
4422 * In all this the minimum difference in data offsets
4423 * (conf->offset_diff - always positive) allows a bit of slack,
4424 * so next can be after 'safe', but not by more than offset_diff
4425 *
4426 * We need to prepare all the bios here before we start any IO
4427 * to ensure the size we choose is acceptable to all devices.
4428 * The means one for each copy for write-out and an extra one for
4429 * read-in.
4430 * We store the read-in bio in ->master_bio and the others in
4431 * ->devs[x].bio and ->devs[x].repl_bio.
4432 */
4447 /* If restarting in the middle, skip the initial sectors */
4460 }
4461 }
4463 /* We don't use sector_nr to track where we are up to
4464 * as that doesn't work well for ->reshape_backwards.
4465 * So just use ->reshape_progress.
4466 */
4468 /* 'next' is the earliest device address that we might
4469 * write to for this chunk in the new layout
4470 */
4474 /* 'safe' is the last device address that we might read from
4475 * in the old layout after a restart
4476 */
4489 /* 'next' is after the last device address that we
4490 * might write to for this chunk in the new layout
4491 */
4494 /* 'safe' is the earliest device address that we might
4495 * read from in the old layout after a restart
4496 */
4499 /* Need to update metadata if 'next' might be beyond 'safe'
4500 * as that would possibly corrupt data
4501 */
4511 }
4515 /* Need to update reshape_position in metadata */
4521 else
4531 }
4534 }
4537 read_more:
4538 /* Now schedule reads for blocks from sector_nr to last */
4551 /* Cannot read from here, so need to record bad blocks
4552 * on all the target devices.
4553 */
4554 // FIXME
4558 }
4575 /*
4576 * Broadcast RESYNC message to other nodes, so all nodes would not
4577 * write to the region to avoid conflict.
4578 */
4588 /*
4589 * Set cluster_sync_low again if next address for array
4590 * reshape is less than cluster_sync_low. Since we can't
4591 * update cluster_sync_low until it has finished reshape.
4592 */
4595 }
4599 }
4601 /* Now find the locations in the new layout */
4618 }
4629 }
4631 /* Now add as many pages as possible to all of these bios. */
4641 /*
4642 * won't fail because the vec table is big enough
4643 * to hold all these pages
4644 */
4646 }
4649 }
4653 /* Now submit the read */
4664 /* Now that we have done the whole section we can
4665 * update reshape_progress
4666 */
4669 else
4673 }
4679 {
4680 /* Reshape read completed. Hopefully we have a block
4681 * to write out.
4682 * If we got a read error then we do sync 1-page reads from
4683 * elsewhere until we find the data - or give up.
4684 */
4690 /* Reshape has been aborted */
4693 }
4695 /* We definitely have the data in the pages, schedule the
4696 * writes.
4697 */
4710 }
4714 }
4721 }
4723 }
4726 {
4738 /* read-ahead size must cover two whole stripes, which is
4739 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4740 */
4747 }
4749 }
4752 {
4760 else
4762 }
4766 {
4767 /* Use sync reads to get the blocks from somewhere else */
4779 }
4781 /* reshape IOs share pages from .devs[0].bio */
4799 sector_t addr;
4809 addr,
4817 failed:
4818 slot++;
4823 }
4826 /* couldn't read this block, must give up */
4831 }
4833 idx++;
4834 }
4837 }
4840 {
4855 }
4858 /* FIXME should record badblock */
4860 }
4864 }
4867 {
4873 }
4876 {
4886 }
4893 d++) {
4900 }
4902 }
4908 }
4911 {
4933 };
4936 {
4938 }
4941 {
4943 }