| blob | c5d88ef6a45c75084c0cf6bbf7f5c8f97246988e |
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 {
853 /* Any writes that have been queued but are awaiting
854 * bitmap updates get flushed here.
855 */
866 /*
867 * As this is called in a wait_event() loop (see freeze_array),
868 * current->state might be TASK_UNINTERRUPTIBLE which will
869 * cause a warning when we prepare to wait again. As it is
870 * rare that this path is taken, it is perfectly safe to force
871 * us to go around the wait_event() loop again, so the warning
872 * is a false-positive. Silence the warning by resetting
873 * thread state
874 */
878 /* flush any pending bitmap writes to disk
879 * before proceeding w/ I/O */
892 /* Just ignore it */
894 else
897 }
901 }
903 /* Barriers....
904 * Sometimes we need to suspend IO while we do something else,
905 * either some resync/recovery, or reconfigure the array.
906 * To do this we raise a 'barrier'.
907 * The 'barrier' is a counter that can be raised multiple times
908 * to count how many activities are happening which preclude
909 * normal IO.
910 * We can only raise the barrier if there is no pending IO.
911 * i.e. if nr_pending == 0.
912 * We choose only to raise the barrier if no-one is waiting for the
913 * barrier to go down. This means that as soon as an IO request
914 * is ready, no other operations which require a barrier will start
915 * until the IO request has had a chance.
916 *
917 * So: regular IO calls 'wait_barrier'. When that returns there
918 * is no backgroup IO happening, It must arrange to call
919 * allow_barrier when it has finished its IO.
920 * backgroup IO calls must call raise_barrier. Once that returns
921 * there is no normal IO happeing. It must arrange to call
922 * lower_barrier when the particular background IO completes.
923 */
926 {
930 /* Wait until no block IO is waiting (unless 'force') */
934 /* block any new IO from starting */
937 /* Now wait for all pending IO to complete */
943 }
946 {
952 }
955 {
960 /* Wait for the barrier to drop.
961 * However if there are already pending
962 * requests (preventing the barrier from
963 * rising completely), and the
964 * pre-process bio queue isn't empty,
965 * then don't wait, as we need to empty
966 * that queue to get the nr_pending
967 * count down.
968 */
973 bio_list &&
976 /* move on if recovery thread is
977 * blocked by us
978 */
987 }
990 }
993 {
997 }
1000 {
1001 /* stop syncio and normal IO and wait for everything to
1002 * go quiet.
1003 * We increment barrier and nr_waiting, and then
1004 * wait until nr_pending match nr_queued+extra
1005 * This is called in the context of one normal IO request
1006 * that has failed. Thus any sync request that might be pending
1007 * will be blocked by nr_pending, and we need to wait for
1008 * pending IO requests to complete or be queued for re-try.
1009 * Thus the number queued (nr_queued) plus this request (extra)
1010 * must match the number of pending IOs (nr_pending) before
1011 * we continue.
1012 */
1024 }
1027 {
1028 /* reverse the effect of the freeze */
1034 }
1038 {
1042 else
1044 }
1050 };
1053 {
1055 cb);
1069 }
1071 /* we aren't scheduling, so we can do the write-out directly. */
1085 /* Just ignore it */
1087 else
1090 }
1092 }
1094 /*
1095 * 1. Register the new request and wait if the reconstruction thread has put
1096 * up a bar for new requests. Continue immediately if no resync is active
1097 * currently.
1098 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1099 */
1102 {
1112 sectors);
1114 }
1115 }
1119 {
1132 /*
1133 * This is an error retry, but we cannot
1134 * safely dereference the rdev in the r10_bio,
1135 * we must use the one in conf.
1136 * If it has already been disconnected (unlikely)
1137 * we lose the device name in error messages.
1138 */
1140 /*
1141 * As we are blocking raid10, it is a little safer to
1142 * use __GFP_HIGH.
1143 */
1153 /* This never gets dereferenced */
1155 }
1157 }
1166 }
1169 }
1185 }
1208 }
1213 {
1228 /* Replacement just got moved to main 'rdev' */
1231 }
1238 else
1255 /* flush_pending_writes() needs access to the rdev so...*/
1263 else
1274 }
1275 }
1279 {
1283 sector_t sectors;
1298 }
1300 }
1310 /* Need to update reshape_position in metadata */
1320 }
1327 }
1328 /* first select target devices under rcu_lock and
1329 * inc refcount on their rdev. Record them by setting
1330 * bios[x] to bio
1331 * If there are known/acknowledged bad blocks on any device
1332 * on which we have seen a write error, we want to avoid
1333 * writing to those blocks. This potentially requires several
1334 * writes to write around the bad blocks. Each set of writes
1335 * gets its own r10_bio with a set of bios attached.
1336 */
1340 retry_write:
1356 }
1361 }
1373 }
1375 sector_t first_bad;
1383 /* Mustn't write here until the bad block
1384 * is acknowledged
1385 */
1390 }
1392 /* Cannot write here at all */
1395 /* Mustn't write more than bad_sectors
1396 * to other devices yet
1397 */
1399 /* We don't set R10BIO_Degraded as that
1400 * only applies if the disk is missing,
1401 * so it might be re-added, and we want to
1402 * know to recover this chunk.
1403 * In this case the device is here, and the
1404 * fact that this chunk is not in-sync is
1405 * recorded in the bad block log.
1406 */
1408 }
1413 }
1414 }
1418 }
1422 }
1423 }
1427 /* Have to wait for this device to get unblocked, then retry */
1435 }
1441 /* Race with remove_disk */
1444 }
1446 }
1447 }
1453 }
1467 }
1477 }
1479 }
1482 {
1499 else
1501 }
1504 {
1517 /*
1518 * If this request crosses a chunk boundary, we need to split
1519 * it.
1520 */
1522 sectors > chunk_sects
1531 /* In case raid10d snuck in to freeze_array */
1534 }
1537 {
1548 else
1552 }
1559 }
1562 }
1564 /* check if there are enough drives for
1565 * every block to appear on atleast one.
1566 * Don't consider the device numbered 'ignore'
1567 * as we might be about to remove it.
1568 */
1570 {
1580 }
1592 cnt++;
1594 }
1600 out:
1603 }
1606 {
1607 /* when calling 'enough', both 'prev' and 'geo' must
1608 * be stable.
1609 * This is ensured if ->reconfig_mutex or ->device_lock
1610 * is held.
1611 */
1614 }
1617 {
1622 /*
1623 * If it is not operational, then we have already marked it as dead
1624 * else if it is the last working disks with "fail_last_dev == false",
1625 * ignore the error, let the next level up know.
1626 * else mark the drive as failed
1627 */
1631 /*
1632 * Don't fail the drive, just return an IO error.
1633 */
1636 }
1639 /*
1640 * If recovery is running, make sure it aborts.
1641 */
1652 }
1655 {
1663 }
1667 /* This is only called with ->reconfix_mutex held, so
1668 * rcu protection of rdev is not needed */
1677 }
1678 }
1681 {
1686 }
1689 {
1696 /*
1697 * Find all non-in_sync disks within the RAID10 configuration
1698 * and mark them in_sync
1699 */
1706 /* Replacement has just become active */
1709 count++;
1711 /* Replaced device not technically faulty,
1712 * but we need to be sure it gets removed
1713 * and never re-added.
1714 */
1718 }
1724 count++;
1726 }
1727 }
1734 }
1737 {
1745 /* only hot-add to in-sync arrays, as recovery is
1746 * very different from resync
1747 */
1762 else
1782 }
1796 }
1802 }
1805 {
1817 else
1824 }
1825 /* Only remove non-faulty devices if recovery
1826 * is not possible.
1827 */
1835 }
1840 /* lost the race, try later */
1844 }
1845 }
1847 /* We must have just cleared 'rdev' */
1851 * but will never see neither -- if they are careful.
1852 */
1854 }
1859 abort:
1863 }
1866 {
1871 else
1872 /* The write handler will notice the lack of
1873 * R10BIO_Uptodate and record any errors etc
1874 */
1878 /* for reconstruct, we always reschedule after a read.
1879 * for resync, only after all reads
1880 */
1884 /* we have read all the blocks,
1885 * do the comparison in process context in raid10d
1886 */
1888 }
1889 }
1892 {
1898 }
1901 {
1902 /* reshape read bio isn't allocated from r10buf_pool */
1906 }
1909 {
1914 /* the primary of several recovery bios */
1919 else
1928 else
1931 }
1932 }
1933 }
1936 {
1941 sector_t first_bad;
1950 else
1962 }
1972 }
1974 /*
1975 * Note: sync and recover and handled very differently for raid10
1976 * This code is for resync.
1977 * For resync, we read through virtual addresses and read all blocks.
1978 * If there is any error, we schedule a write. The lowest numbered
1979 * drive is authoritative.
1980 * However requests come for physical address, so we need to map.
1981 * For every physical address there are raid_disks/copies virtual addresses,
1982 * which is always are least one, but is not necessarly an integer.
1983 * This means that a physical address can span multiple chunks, so we may
1984 * have to submit multiple io requests for a single sync request.
1985 */
1986 /*
1987 * We check if all blocks are in-sync and only write to blocks that
1988 * aren't in sync
1989 */
1991 {
2000 /* find the first device with a block */
2015 /* 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 */
2085 }
2087 /* Now write out to any replacement devices
2088 * that are active
2089 */
2104 }
2106 done:
2110 }
2111 }
2113 /*
2114 * Now for the recovery code.
2115 * Recovery happens across physical sectors.
2116 * We recover all non-is_sync drives by finding the virtual address of
2117 * each, and then choose a working drive that also has that virt address.
2118 * There is a separate r10_bio for each non-in_sync drive.
2119 * Only the first two slots are in use. The first for reading,
2120 * The second for writing.
2121 *
2122 */
2124 {
2125 /* We got a read error during recovery.
2126 * We repeat the read in smaller page-sized sections.
2127 * If a read succeeds, write it to the new device or record
2128 * a bad block if we cannot.
2129 * If a read fails, record a bad block on both old and
2130 * new devices.
2131 */
2145 sector_t addr;
2154 addr,
2162 addr,
2172 }
2173 }
2175 /* We don't worry if we cannot set a bad block -
2176 * it really is bad so there is no loss in not
2177 * recording it yet
2178 */
2182 /* need bad block on destination too */
2187 /* just abort the recovery */
2196 }
2197 }
2198 }
2202 idx++;
2203 }
2204 }
2207 {
2216 }
2218 /*
2219 * share the pages with the first bio
2220 * and submit the write request
2221 */
2225 /* Need to test wbio2->bi_end_io before we call
2226 * submit_bio_noacct as if the former is NULL,
2227 * the latter is free to free wbio2.
2228 */
2235 }
2241 }
2242 }
2244 /*
2245 * Used by fix_read_error() to decay the per rdev read_errors.
2246 * We halve the read error count for every hour that has elapsed
2247 * since the last recorded read error.
2248 *
2249 */
2251 {
2259 /* first time we've seen a read error */
2262 }
2269 /*
2270 * if hours_since_last is > the number of bits in read_errors
2271 * just set read errors to 0. We do this to avoid
2272 * overflowing the shift of read_errors by hours_since_last.
2273 */
2276 else
2278 }
2282 {
2283 sector_t first_bad;
2290 /* success */
2297 }
2298 /* need to record an error - either for the block or the device */
2302 }
2304 /*
2305 * This is a kernel thread which:
2306 *
2307 * 1. Retries failed read operations on working mirrors.
2308 * 2. Updates the raid superblock when problems encounter.
2309 * 3. Performs writes following reads for array synchronising.
2310 */
2313 {
2320 /* still own a reference to this rdev, so it cannot
2321 * have been cleared recently.
2322 */
2326 /* drive has already been failed, just ignore any
2327 more fix_read_error() attempts */
2344 }
2357 sector_t first_bad;
2371 sect,
2379 }
2380 sl++;
2387 /* Cannot read from anywhere, just mark the block
2388 * as bad on the first device to discourage future
2389 * reads.
2390 */
2395 rdev,
2402 }
2404 }
2407 /* write it back and re-read */
2414 sl--;
2426 sect,
2429 /* Well, this device is dead */
2433 sect +
2435 rdev)),
2440 }
2443 }
2450 sl--;
2462 sect,
2464 READ)) {
2466 /* Well, this device is dead */
2470 sect +
2481 sect +
2485 }
2489 }
2494 }
2495 }
2498 {
2503 /* bio has the data to be written to slot 'i' where
2504 * we just recently had a write error.
2505 * We repeatedly clone the bio and trim down to one block,
2506 * then try the write. Where the write fails we record
2507 * a bad block.
2508 * It is conceivable that the bio doesn't exactly align with
2509 * blocks. We must handle this.
2510 *
2511 * We currently own a reference to the rdev.
2512 */
2515 sector_t sector;
2532 sector_t wsector;
2535 /* Write at 'sector' for 'sectors' */
2545 /* Failure! */
2554 }
2556 }
2559 {
2565 /* we got a read error. Maybe the drive is bad. Maybe just
2566 * the block and we can fix it.
2567 * We freeze all other IO, and try reading the block from
2568 * other devices. When we find one, we re-write
2569 * and check it that fixes the read error.
2570 * This is all done synchronously while the array is
2571 * frozen.
2572 */
2590 }
2593 {
2594 /* Some sort of write request has finished and it
2595 * succeeded in writing where we thought there was a
2596 * bad block. So forget the bad block.
2597 * Or possibly if failed and we need to record
2598 * a bad block.
2599 */
2613 rdev,
2618 rdev,
2622 }
2630 rdev,
2635 rdev,
2639 }
2640 }
2650 rdev,
2660 }
2662 }
2667 rdev,
2671 }
2672 }
2678 /*
2679 * In case freeze_array() is waiting for condition
2680 * nr_pending == nr_queued + extra to be true.
2681 */
2689 }
2690 }
2691 }
2694 {
2712 }
2713 }
2717 retry_list);
2726 }
2727 }
2738 }
2757 else
2763 }
2765 }
2768 {
2783 }
2786 {
2796 else
2809 }
2810 }
2812 }
2814 /*
2815 * Set cluster_sync_high since we need other nodes to add the
2816 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2817 */
2819 {
2820 sector_t window_size;
2823 /*
2824 * First, here we define "stripe" as a unit which across
2825 * all member devices one time, so we get chunks by use
2826 * raid_disks / near_copies. Otherwise, if near_copies is
2827 * close to raid_disks, then resync window could increases
2828 * linearly with the increase of raid_disks, which means
2829 * we will suspend a really large IO window while it is not
2830 * necessary. If raid_disks is not divisible by near_copies,
2831 * an extra chunk is needed to ensure the whole "stripe" is
2832 * covered.
2833 */
2838 else
2842 /*
2843 * At least use a 32M window to align with raid1's resync window
2844 */
2849 }
2851 /*
2852 * perform a "sync" on one "block"
2853 *
2854 * We need to make sure that no normal I/O request - particularly write
2855 * requests - conflict with active sync requests.
2856 *
2857 * This is achieved by tracking pending requests and a 'barrier' concept
2858 * that can be installed to exclude normal IO requests.
2859 *
2860 * Resync and recovery are handled very differently.
2861 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2862 *
2863 * For resync, we iterate over virtual addresses, read all copies,
2864 * and update if there are differences. If only one copy is live,
2865 * skip it.
2866 * For recovery, we iterate over physical addresses, read a good
2867 * value for each non-in_sync drive, and over-write.
2868 *
2869 * So, for recovery we may have several outstanding complex requests for a
2870 * given address, one for each out-of-sync device. We model this by allocating
2871 * a number of r10_bio structures, one for each out-of-sync device.
2872 * As we setup these structures, we collect all bio's together into a list
2873 * which we then process collectively to add pages, and then process again
2874 * to pass to submit_bio_noacct.
2875 *
2876 * The r10_bio structures are linked using a borrowed master_bio pointer.
2877 * This link is counted in ->remaining. When the r10_bio that points to NULL
2878 * has its remaining count decremented to 0, the whole complex operation
2879 * is complete.
2880 *
2881 */
2885 {
2892 sector_t sync_blocks;
2902 /*
2903 * Allow skipping a full rebuild for incremental assembly
2904 * of a clean array, like RAID1 does.
2905 */
2915 }
2917 skipped:
2926 /* If we aborted, we need to abort the
2927 * sync on the 'current' bitmap chucks (there can
2928 * be several when recovering multiple devices).
2929 * as we may have started syncing it but not finished.
2930 * We can find the current address in
2931 * mddev->curr_resync, but for recovery,
2932 * we need to convert that to several
2933 * virtual addresses.
2934 */
2939 }
2946 sector_t sect =
2950 }
2952 /* completed sync */
2956 /* Completed a full sync so the replacements
2957 * are now fully recovered.
2958 */
2965 }
2967 }
2969 }
2974 }
2980 /* if there has been nothing to do on any drive,
2981 * then there is nothing to do at all..
2982 */
2985 }
2990 /* make sure whole request will fit in a chunk - if chunks
2991 * are meaningful
2992 */
2997 /*
2998 * If there is non-resync activity waiting for a turn, then let it
2999 * though before starting on this new sync request.
3000 */
3004 /* Again, very different code for resync and recovery.
3005 * Both must result in an r10bio with a list of bios that
3006 * have bi_end_io, bi_sector, bi_disk set,
3007 * and bi_private set to the r10bio.
3008 * For recovery, we may actually create several r10bios
3009 * with 2 bios in each, that correspond to the bios in the main one.
3010 * In this case, the subordinate r10bios link back through a
3011 * borrowed master_bio pointer, and the counter in the master
3012 * includes a ref from each subordinate.
3013 */
3014 /* First, we decide what to do and set ->bi_end_io
3015 * To end_sync_read if we want to read, and
3016 * end_sync_write if we will want to write.
3017 */
3021 /* recovery... the complicated one */
3028 sector_t sect;
3051 }
3054 /* want to reconstruct this device */
3058 /* last stripe is not complete - don't
3059 * try to recover this sector.
3060 */
3063 }
3066 /* Unless we are doing a full sync, or a replacement
3067 * we only need to recover the block if it is set in
3068 * the bitmap
3069 */
3077 /* yep, skip the sync_blocks here, but don't assume
3078 * that there will never be anything to do here
3079 */
3083 }
3103 /* Need to check if the array will still be
3104 * degraded
3105 */
3113 }
3114 }
3131 /* This is where we read from */
3140 bad_sectors -= (sector
3145 }
3146 }
3159 /* and we write to 'i' (if not in_sync) */
3184 /* and maybe write to replacement */
3188 /* Note: if need_replace, then bio
3189 * cannot be NULL as r10buf_pool_alloc will
3190 * have allocated it.
3191 */
3203 }
3206 /* Cannot recover, so abort the recovery or
3207 * record a bad block */
3209 /* problem is that there are bad blocks
3210 * on other device(s)
3211 */
3219 mrdev,
3225 mreplace,
3229 }
3235 mirror->recovery_disabled
3237 }
3246 }
3251 /* Only want this if there is elsewhere to
3252 * read from. 'j' is currently the first
3253 * readable copy.
3254 */
3261 targets++;
3262 }
3266 }
3267 }
3274 }
3276 }
3278 /* resync. Schedule a read for every block at this virt offset */
3281 /*
3282 * Since curr_resync_completed could probably not update in
3283 * time, and we will set cluster_sync_low based on it.
3284 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3285 * safety reason, which ensures curr_resync_completed is
3286 * updated in bitmap_cond_end_sync.
3287 */
3296 /* We can skip this block */
3299 }
3332 }
3344 }
3345 }
3356 count++;
3362 }
3365 /* Need to set up for writing to the replacement */
3378 count++;
3380 }
3387 mddev);
3392 mddev);
3393 }
3397 }
3398 }
3413 /*
3414 * won't fail because the vec table is big enough
3415 * to hold all these pages
3416 */
3418 }
3426 /* It is resync not recovery */
3430 /* Send resync message */
3434 }
3436 /* This is recovery not resync */
3441 /*
3442 * sector_nr is a device address for recovery, so we
3443 * need translate it to array address before compare
3444 * with cluster_sync_high.
3445 */
3450 /*
3451 * curr_resync_completed is similar as
3452 * sector_nr, so make the translation too.
3453 */
3460 }
3461 }
3467 }
3468 }
3482 }
3483 }
3486 /* pretend they weren't skipped, it makes
3487 * no important difference in this case
3488 */
3492 giveup:
3493 /* There is nowhere to write, so all non-sync
3494 * drives must be failed or in resync, all drives
3495 * have a bad block, so try the next chunk...
3496 */
3501 chunks_skipped ++;
3504 }
3506 static sector_t
3508 {
3509 sector_t size;
3524 }
3527 {
3528 /* Calculate the number of sectors-per-device that will
3529 * actually be used, and set conf->dev_sectors and
3530 * conf->stride
3531 */
3537 /* 'size' is now the number of chunks in the array */
3538 /* calculate "used chunks per device" */
3541 /* We need to round up when dividing by raid_disks to
3542 * get the stride size.
3543 */
3553 }
3554 }
3558 {
3574 * updated yet. */
3579 }
3597 * actually using this, but leave code here just in case.*/
3607 }
3611 }
3614 {
3626 }
3632 }
3639 /* FIXME calc properly */
3642 GFP_KERNEL);
3669 }
3673 else
3674 /* far_copies must be 1 */
3676 }
3694 out:
3701 }
3703 }
3706 {
3712 raid_disks);
3713 }
3716 {
3721 sector_t size;
3734 }
3748 }
3749 }
3761 }
3782 }
3800 }
3806 else
3809 }
3810 /* need to check that every block has at least one working mirror */
3815 }
3818 /* must ensure that shape change is supported */
3825 }
3831 i++) {
3836 /* The replacement is all we have - use it */
3840 }
3849 }
3855 }
3858 }
3866 /*
3867 * Ok, everything is just fine now
3868 */
3887 /* This cannot work */
3890 }
3901 }
3905 out_free_conf:
3912 out:
3914 }
3917 {
3927 }
3930 {
3935 else
3937 }
3940 {
3941 /* Resize of 'far' arrays is not supported.
3942 * For 'near' and 'offset' arrays we can set the
3943 * number of sectors used to be an appropriate multiple
3944 * of the chunk size.
3945 * For 'offset', this is far_copies*chunksize.
3946 * For 'near' the multiplier is the LCM of
3947 * near_copies and raid_disks.
3948 * So if far_copies > 1 && !far_offset, fail.
3949 * Else find LCM(raid_disks, near_copy)*far_copies and
3950 * multiply by chunk_size. Then round to this number.
3951 * This is mostly done by raid10_size()
3952 */
3971 }
3977 }
3982 }
3985 {
3993 }
3996 /* Set new parameters */
3998 /* new layout: far_copies = 1, near_copies = 2 */
4003 /* make sure it will be not marked as dirty */
4013 }
4015 }
4018 }
4021 {
4024 /* raid10 can take over:
4025 * raid0 - providing it has only two drives
4026 */
4028 /* for raid0 takeover only one zone is supported */
4034 }
4038 }
4040 }
4043 {
4044 /* Called when there is a request to change
4045 * - layout (to ->new_layout)
4046 * - chunk size (to ->new_chunk_sectors)
4047 * - raid_disks (by delta_disks)
4048 * or when trying to restart a reshape that was ongoing.
4049 *
4050 * We need to validate the request and possibly allocate
4051 * space if that might be an issue later.
4052 *
4053 * Currently we reject any reshape of a 'far' mode array,
4054 * allow chunk size to change if new is generally acceptable,
4055 * allow raid_disks to increase, and allow
4056 * a switch between 'near' mode and 'offset' mode.
4057 */
4065 /* mustn't change number of copies */
4068 /* Cannot switch to 'far' mode */
4072 /* not factor of array size */
4081 /* allocate new 'mirrors' list */
4085 GFP_KERNEL);
4088 }
4090 }
4092 /*
4093 * Need to check if array has failed when deciding whether to:
4094 * - start an array
4095 * - remove non-faulty devices
4096 * - add a spare
4097 * - allow a reshape
4098 * This determination is simple when no reshape is happening.
4099 * However if there is a reshape, we need to carefully check
4100 * both the before and after sections.
4101 * This is because some failed devices may only affect one
4102 * of the two sections, and some non-in_sync devices may
4103 * be insync in the section most affected by failed devices.
4104 */
4106 {
4112 /* 'prev' section first */
4116 degraded++;
4118 /* When we can reduce the number of devices in
4119 * an array, this might not contribute to
4120 * 'degraded'. It does now.
4121 */
4122 degraded++;
4123 }
4132 degraded2++;
4134 /* If reshape is increasing the number of devices,
4135 * this section has already been recovered, so
4136 * it doesn't contribute to degraded.
4137 * else it does.
4138 */
4140 degraded2++;
4141 }
4142 }
4147 }
4150 {
4151 /* A 'reshape' has been requested. This commits
4152 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4153 * This also checks if there are enough spares and adds them
4154 * to the array.
4155 * We currently require enough spares to make the final
4156 * array non-degraded. We also require that the difference
4157 * between old and new data_offset - on each device - is
4158 * enough that we never risk over-writing.
4159 */
4184 spares++;
4195 }
4196 }
4214 }
4224 }
4243 else
4245 }
4251 }
4253 /*
4254 * some node is already performing reshape, and no need to
4255 * call md_bitmap_resize again since it should be called when
4256 * receiving BITMAP_RESIZE msg
4257 */
4270 }
4271 }
4272 out:
4281 else
4284 /* Failure here is OK */
4286 }
4289 /* This is a spare that was manually added */
4291 }
4292 }
4293 /* When a reshape changes the number of devices,
4294 * ->degraded is measured against the larger of the
4295 * pre and post numbers.
4296 */
4315 }
4321 abort:
4334 }
4336 /* Calculate the last device-address that could contain
4337 * any block from the chunk that includes the array-address 's'
4338 * and report the next address.
4339 * i.e. the address returned will be chunk-aligned and after
4340 * any data that is in the chunk containing 's'.
4341 */
4343 {
4351 }
4353 /* Calculate the first device-address that could contain
4354 * any block from the chunk that includes the array-address 's'.
4355 * This too will be the start of a chunk
4356 */
4358 {
4365 }
4369 {
4370 /* We simply copy at most one chunk (smallest of old and new)
4371 * at a time, possibly less if that exceeds RESYNC_PAGES,
4372 * or we hit a bad block or something.
4373 * This might mean we pause for normal IO in the middle of
4374 * a chunk, but that is not a problem as mddev->reshape_position
4375 * can record any location.
4376 *
4377 * If we will want to write to a location that isn't
4378 * yet recorded as 'safe' (i.e. in metadata on disk) then
4379 * we need to flush all reshape requests and update the metadata.
4380 *
4381 * When reshaping forwards (e.g. to more devices), we interpret
4382 * 'safe' as the earliest block which might not have been copied
4383 * down yet. We divide this by previous stripe size and multiply
4384 * by previous stripe length to get lowest device offset that we
4385 * cannot write to yet.
4386 * We interpret 'sector_nr' as an address that we want to write to.
4387 * From this we use last_device_address() to find where we might
4388 * write to, and first_device_address on the 'safe' position.
4389 * If this 'next' write position is after the 'safe' position,
4390 * we must update the metadata to increase the 'safe' position.
4391 *
4392 * When reshaping backwards, we round in the opposite direction
4393 * and perform the reverse test: next write position must not be
4394 * less than current safe position.
4395 *
4396 * In all this the minimum difference in data offsets
4397 * (conf->offset_diff - always positive) allows a bit of slack,
4398 * so next can be after 'safe', but not by more than offset_diff
4399 *
4400 * We need to prepare all the bios here before we start any IO
4401 * to ensure the size we choose is acceptable to all devices.
4402 * The means one for each copy for write-out and an extra one for
4403 * read-in.
4404 * We store the read-in bio in ->master_bio and the others in
4405 * ->devs[x].bio and ->devs[x].repl_bio.
4406 */
4421 /* If restarting in the middle, skip the initial sectors */
4434 }
4435 }
4437 /* We don't use sector_nr to track where we are up to
4438 * as that doesn't work well for ->reshape_backwards.
4439 * So just use ->reshape_progress.
4440 */
4442 /* 'next' is the earliest device address that we might
4443 * write to for this chunk in the new layout
4444 */
4448 /* 'safe' is the last device address that we might read from
4449 * in the old layout after a restart
4450 */
4463 /* 'next' is after the last device address that we
4464 * might write to for this chunk in the new layout
4465 */
4468 /* 'safe' is the earliest device address that we might
4469 * read from in the old layout after a restart
4470 */
4473 /* Need to update metadata if 'next' might be beyond 'safe'
4474 * as that would possibly corrupt data
4475 */
4485 }
4489 /* Need to update reshape_position in metadata */
4495 else
4505 }
4508 }
4511 read_more:
4512 /* Now schedule reads for blocks from sector_nr to last */
4525 /* Cannot read from here, so need to record bad blocks
4526 * on all the target devices.
4527 */
4528 // FIXME
4532 }
4549 /*
4550 * Broadcast RESYNC message to other nodes, so all nodes would not
4551 * write to the region to avoid conflict.
4552 */
4562 /*
4563 * Set cluster_sync_low again if next address for array
4564 * reshape is less than cluster_sync_low. Since we can't
4565 * update cluster_sync_low until it has finished reshape.
4566 */
4569 }
4573 }
4575 /* Now find the locations in the new layout */
4592 }
4603 }
4605 /* Now add as many pages as possible to all of these bios. */
4615 /*
4616 * won't fail because the vec table is big enough
4617 * to hold all these pages
4618 */
4620 }
4623 }
4627 /* Now submit the read */
4638 /* Now that we have done the whole section we can
4639 * update reshape_progress
4640 */
4643 else
4647 }
4653 {
4654 /* Reshape read completed. Hopefully we have a block
4655 * to write out.
4656 * If we got a read error then we do sync 1-page reads from
4657 * elsewhere until we find the data - or give up.
4658 */
4664 /* Reshape has been aborted */
4667 }
4669 /* We definitely have the data in the pages, schedule the
4670 * writes.
4671 */
4684 }
4688 }
4695 }
4697 }
4700 {
4715 }
4718 {
4726 else
4728 }
4732 {
4733 /* Use sync reads to get the blocks from somewhere else */
4745 }
4747 /* reshape IOs share pages from .devs[0].bio */
4765 sector_t addr;
4775 addr,
4783 failed:
4784 slot++;
4789 }
4792 /* couldn't read this block, must give up */
4797 }
4799 idx++;
4800 }
4803 }
4806 {
4821 }
4824 /* FIXME should record badblock */
4826 }
4830 }
4833 {
4839 }
4842 {
4852 }
4859 d++) {
4866 }
4868 }
4874 }
4877 {
4898 };
4901 {
4903 }
4906 {
4908 }