| blob | 02c5e390f89f32ca26095dd015d4b09f01843c59 |
1 /*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include <trace/events/block.h>
34 /*
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
37 * chunk_size
38 * raid_disks
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
44 *
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
53 *
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
57 *
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
65 * on a device):
66 * A B C D A B C D E
67 * ... ...
68 * D A B C E A B C D
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
73 */
75 /*
76 * Number of guaranteed r10bios in case of extreme VM load:
77 */
78 #define NR_RAID10_BIOS 256
80 /* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84 */
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
89 */
90 #define IO_MADE_GOOD ((struct bio *)2)
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
94 /* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
96 * for writeback.
97 */
110 #define raid10_log(md, fmt, args...) \
115 /*
116 * for resync bio, r10bio pointer can be retrieved from the per-bio
117 * 'struct resync_pages'.
118 */
120 {
122 }
125 {
129 /* allocate a r10bio with room for raid_disks entries in the
130 * bios array */
132 }
135 {
137 }
139 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
140 /* amount of memory to reserve for resync requests */
141 #define RESYNC_WINDOW (1024*1024)
142 /* maximum number of concurrent requests, memory permitting */
143 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
145 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
147 /*
148 * When performing a resync, we need to read and compare, so
149 * we need as many pages are there are copies.
150 * When performing a recovery, we need 2 bios, one for read,
151 * one for write (we recover only one drive per r10buf)
152 *
153 */
155 {
170 else
173 /* allocate once for all bios */
176 else
182 /*
183 * Allocate bios.
184 */
196 }
197 /*
198 * Allocate RESYNC_PAGES data pages and attach them
199 * where needed.
200 */
218 }
225 }
226 }
230 out_free_pages:
235 out_free_bio:
241 }
243 out_free_r10bio:
246 }
249 {
262 }
267 }
269 /* resync pages array stored in the 1st bio's .bi_private */
273 }
276 {
288 }
289 }
292 {
297 }
300 {
306 }
309 {
319 /* wake up frozen array... */
323 }
325 /*
326 * raid_end_bio_io() is called when we have finished servicing a mirrored
327 * operation and are ready to return a success/failure code to the buffer
328 * cache layer.
329 */
331 {
339 /*
340 * Wake up any possible resync thread that waits for the device
341 * to go idle.
342 */
346 }
348 /*
349 * Update disk head position estimator based on IRQ completion info.
350 */
352 {
357 }
359 /*
360 * Find the disk number which triggered given bio
361 */
364 {
374 }
375 }
385 }
388 {
397 /*
398 * this branch is our 'one mirror IO has finished' event handler:
399 */
403 /*
404 * Set R10BIO_Uptodate in our master bio, so that
405 * we will return a good error code to the higher
406 * levels even if IO on some other mirrored buffer fails.
407 *
408 * The 'master' represents the composite IO operation to
409 * user-side. So if something waits for IO, then it will
410 * wait for the 'master' bio.
411 */
414 /* If all other devices that store this block have
415 * failed, we want to return the error upwards rather
416 * than fail the last device. Here we redefine
417 * "uptodate" to mean "Don't want to retry"
418 */
422 }
427 /*
428 * oops, read error - keep the refcount on the rdev
429 */
437 }
438 }
441 {
442 /* clear the bitmap if all writes complete successfully */
448 }
451 {
459 else
461 }
462 }
463 }
466 {
486 }
487 /*
488 * this branch is our 'one mirror IO has finished' event handler:
489 */
492 /* Never record new bad blocks to replacement,
493 * just fail it.
494 */
507 /* This is the only remaining device,
508 * We need to retry the write without
509 * FailFast
510 */
516 }
519 }
521 /*
522 * Set R10BIO_Uptodate in our master bio, so that
523 * we will return a good error code for to the higher
524 * levels even if IO on some other mirrored buffer fails.
525 *
526 * The 'master' represents the composite IO operation to
527 * user-side. So if something waits for IO, then it will
528 * wait for the 'master' bio.
529 */
530 sector_t first_bad;
533 /*
534 * Do not set R10BIO_Uptodate if the current device is
535 * rebuilding or Faulty. This is because we cannot use
536 * such device for properly reading the data back (we could
537 * potentially use it, if the current write would have felt
538 * before rdev->recovery_offset, but for simplicity we don't
539 * check this here.
540 */
545 /* Maybe we can clear some bad blocks. */
553 else
557 }
558 }
560 /*
561 *
562 * Let's see if all mirrored write operations have finished
563 * already.
564 */
570 }
572 /*
573 * RAID10 layout manager
574 * As well as the chunksize and raid_disks count, there are two
575 * parameters: near_copies and far_copies.
576 * near_copies * far_copies must be <= raid_disks.
577 * Normally one of these will be 1.
578 * If both are 1, we get raid0.
579 * If near_copies == raid_disks, we get raid1.
580 *
581 * Chunks are laid out in raid0 style with near_copies copies of the
582 * first chunk, followed by near_copies copies of the next chunk and
583 * so on.
584 * If far_copies > 1, then after 1/far_copies of the array has been assigned
585 * as described above, we start again with a device offset of near_copies.
586 * So we effectively have another copy of the whole array further down all
587 * the drives, but with blocks on different drives.
588 * With this layout, and block is never stored twice on the one device.
589 *
590 * raid10_find_phys finds the sector offset of a given virtual sector
591 * on each device that it is on.
592 *
593 * raid10_find_virt does the reverse mapping, from a device and a
594 * sector offset to a virtual address
595 */
598 {
600 sector_t sector;
601 sector_t chunk;
602 sector_t stripe;
613 /* now calculate first sector/dev */
625 /* and calculate all the others */
632 slot++;
646 }
650 slot++;
651 }
652 dev++;
656 }
657 }
658 }
661 {
673 }
676 {
678 /* Never use conf->prev as this is only called during resync
679 * or recovery, so reshape isn't happening
680 */
694 }
695 }
710 else
712 }
714 }
718 }
720 /*
721 * This routine returns the disk from which the requested read should
722 * be done. There is a per-array 'next expected sequential IO' sector
723 * number - if this matches on the next IO then we use the last disk.
724 * There is also a per-disk 'last know head position' sector that is
725 * maintained from IRQ contexts, both the normal and the resync IO
726 * completion handlers update this position correctly. If there is no
727 * perfect sequential match then we pick the disk whose head is closest.
728 *
729 * If there are 2 mirrors in the same 2 devices, performance degrades
730 * because position is mirror, not device based.
731 *
732 * The rdev for the device selected will have nr_pending incremented.
733 */
735 /*
736 * FIXME: possibly should rethink readbalancing and do it differently
737 * depending on near_copies / far_copies geometry.
738 */
742 {
761 /*
762 * Check if we can balance. We can balance on the whole
763 * device if no resync is going on (recovery is ok), or below
764 * the resync window. We take the first readable disk when
765 * above the resync window.
766 */
775 sector_t first_bad;
777 sector_t dev_sector;
797 /* Already have a better slot */
800 /* Cannot read here. If this is the
801 * 'primary' device, then we must not read
802 * beyond 'bad_sectors' from another device.
803 */
810 sector_t good_sectors =
816 }
818 /* Must read from here */
820 }
829 /* At least 2 disks to choose from so failfast is OK */
831 /* This optimisation is debatable, and completely destroys
832 * sequential read speed for 'far copies' arrays. So only
833 * keep it for 'near' arrays, and review those later.
834 */
838 /* for far > 1 always use the lowest address */
841 else
848 }
849 }
853 }
864 }
867 {
879 i++) {
885 }
886 }
889 }
892 {
893 /* Any writes that have been queued but are awaiting
894 * bitmap updates get flushed here.
895 */
906 /*
907 * As this is called in a wait_event() loop (see freeze_array),
908 * current->state might be TASK_UNINTERRUPTIBLE which will
909 * cause a warning when we prepare to wait again. As it is
910 * rare that this path is taken, it is perfectly safe to force
911 * us to go around the wait_event() loop again, so the warning
912 * is a false-positive. Silence the warning by resetting
913 * thread state
914 */
918 /* flush any pending bitmap writes to disk
919 * before proceeding w/ I/O */
932 /* Just ignore it */
934 else
937 }
941 }
943 /* Barriers....
944 * Sometimes we need to suspend IO while we do something else,
945 * either some resync/recovery, or reconfigure the array.
946 * To do this we raise a 'barrier'.
947 * The 'barrier' is a counter that can be raised multiple times
948 * to count how many activities are happening which preclude
949 * normal IO.
950 * We can only raise the barrier if there is no pending IO.
951 * i.e. if nr_pending == 0.
952 * We choose only to raise the barrier if no-one is waiting for the
953 * barrier to go down. This means that as soon as an IO request
954 * is ready, no other operations which require a barrier will start
955 * until the IO request has had a chance.
956 *
957 * So: regular IO calls 'wait_barrier'. When that returns there
958 * is no backgroup IO happening, It must arrange to call
959 * allow_barrier when it has finished its IO.
960 * backgroup IO calls must call raise_barrier. Once that returns
961 * there is no normal IO happeing. It must arrange to call
962 * lower_barrier when the particular background IO completes.
963 */
966 {
970 /* Wait until no block IO is waiting (unless 'force') */
974 /* block any new IO from starting */
977 /* Now wait for all pending IO to complete */
983 }
986 {
992 }
995 {
999 /* Wait for the barrier to drop.
1000 * However if there are already pending
1001 * requests (preventing the barrier from
1002 * rising completely), and the
1003 * pre-process bio queue isn't empty,
1004 * then don't wait, as we need to empty
1005 * that queue to get the nr_pending
1006 * count down.
1007 */
1019 }
1022 }
1025 {
1029 }
1032 {
1033 /* stop syncio and normal IO and wait for everything to
1034 * go quiet.
1035 * We increment barrier and nr_waiting, and then
1036 * wait until nr_pending match nr_queued+extra
1037 * This is called in the context of one normal IO request
1038 * that has failed. Thus any sync request that might be pending
1039 * will be blocked by nr_pending, and we need to wait for
1040 * pending IO requests to complete or be queued for re-try.
1041 * Thus the number queued (nr_queued) plus this request (extra)
1042 * must match the number of pending IOs (nr_pending) before
1043 * we continue.
1044 */
1056 }
1059 {
1060 /* reverse the effect of the freeze */
1066 }
1070 {
1074 else
1076 }
1082 };
1085 {
1087 cb);
1101 }
1103 /* we aren't scheduling, so we can do the write-out directly. */
1117 /* Just ignore it */
1119 else
1122 }
1124 }
1128 {
1134 sector_t sectors;
1142 /*
1143 * This is an error retry, but we cannot
1144 * safely dereference the rdev in the r10_bio,
1145 * we must use the one in conf.
1146 * If it has already been disconnected (unlikely)
1147 * we lose the device name in error messages.
1148 */
1150 /*
1151 * As we are blocking raid10, it is a little safer to
1152 * use __GFP_HIGH.
1153 */
1163 /* This never gets dereferenced */
1165 }
1167 }
1168 /*
1169 * Register the new request and wait if the reconstruction
1170 * thread has put up a bar for new requests.
1171 * Continue immediately if no resync is active currently.
1172 */
1179 /*
1180 * IO spans the reshape position. Need to wait for reshape to
1181 * pass
1182 */
1188 sectors);
1190 }
1198 }
1201 }
1217 }
1241 }
1246 {
1261 /* Replacement just got moved to main 'rdev' */
1264 }
1271 else
1289 /* flush_pending_writes() needs access to the rdev so...*/
1297 else
1308 }
1309 }
1313 {
1317 sector_t sectors;
1332 }
1334 }
1336 /*
1337 * Register the new request and wait if the reconstruction
1338 * thread has put up a bar for new requests.
1339 * Continue immediately if no resync is active currently.
1340 */
1347 /*
1348 * IO spans the reshape position. Need to wait for reshape to
1349 * pass
1350 */
1356 sectors);
1358 }
1366 /* Need to update reshape_position in metadata */
1376 }
1383 }
1384 /* first select target devices under rcu_lock and
1385 * inc refcount on their rdev. Record them by setting
1386 * bios[x] to bio
1387 * If there are known/acknowledged bad blocks on any device
1388 * on which we have seen a write error, we want to avoid
1389 * writing to those blocks. This potentially requires several
1390 * writes to write around the bad blocks. Each set of writes
1391 * gets its own r10_bio with a set of bios attached.
1392 */
1396 retry_write:
1412 }
1417 }
1429 }
1431 sector_t first_bad;
1439 /* Mustn't write here until the bad block
1440 * is acknowledged
1441 */
1446 }
1448 /* Cannot write here at all */
1451 /* Mustn't write more than bad_sectors
1452 * to other devices yet
1453 */
1455 /* We don't set R10BIO_Degraded as that
1456 * only applies if the disk is missing,
1457 * so it might be re-added, and we want to
1458 * know to recover this chunk.
1459 * In this case the device is here, and the
1460 * fact that this chunk is not in-sync is
1461 * recorded in the bad block log.
1462 */
1464 }
1469 }
1470 }
1474 }
1478 }
1479 }
1483 /* Have to wait for this device to get unblocked, then retry */
1491 }
1497 /* Race with remove_disk */
1500 }
1502 }
1503 }
1509 }
1523 }
1533 }
1535 }
1538 {
1554 else
1556 }
1559 {
1572 /*
1573 * If this request crosses a chunk boundary, we need to split
1574 * it.
1575 */
1577 sectors > chunk_sects
1586 /* In case raid10d snuck in to freeze_array */
1589 }
1592 {
1603 else
1607 }
1614 }
1617 }
1619 /* check if there are enough drives for
1620 * every block to appear on atleast one.
1621 * Don't consider the device numbered 'ignore'
1622 * as we might be about to remove it.
1623 */
1625 {
1635 }
1647 cnt++;
1649 }
1655 out:
1658 }
1661 {
1662 /* when calling 'enough', both 'prev' and 'geo' must
1663 * be stable.
1664 * This is ensured if ->reconfig_mutex or ->device_lock
1665 * is held.
1666 */
1669 }
1672 {
1677 /*
1678 * If it is not operational, then we have already marked it as dead
1679 * else if it is the last working disks, ignore the error, let the
1680 * next level up know.
1681 * else mark the drive as failed
1682 */
1686 /*
1687 * Don't fail the drive, just return an IO error.
1688 */
1691 }
1694 /*
1695 * If recovery is running, make sure it aborts.
1696 */
1707 }
1710 {
1718 }
1722 /* This is only called with ->reconfix_mutex held, so
1723 * rcu protection of rdev is not needed */
1732 }
1733 }
1736 {
1741 }
1744 {
1751 /*
1752 * Find all non-in_sync disks within the RAID10 configuration
1753 * and mark them in_sync
1754 */
1761 /* Replacement has just become active */
1764 count++;
1766 /* Replaced device not technically faulty,
1767 * but we need to be sure it gets removed
1768 * and never re-added.
1769 */
1773 }
1779 count++;
1781 }
1782 }
1789 }
1792 {
1800 /* only hot-add to in-sync arrays, as recovery is
1801 * very different from resync
1802 */
1817 else
1837 }
1851 }
1857 }
1860 {
1872 else
1879 }
1880 /* Only remove non-faulty devices if recovery
1881 * is not possible.
1882 */
1890 }
1895 /* lost the race, try later */
1899 }
1900 }
1902 /* We must have just cleared 'rdev' */
1906 * but will never see neither -- if they are careful.
1907 */
1909 }
1914 abort:
1918 }
1921 {
1926 else
1927 /* The write handler will notice the lack of
1928 * R10BIO_Uptodate and record any errors etc
1929 */
1933 /* for reconstruct, we always reschedule after a read.
1934 * for resync, only after all reads
1935 */
1939 /* we have read all the blocks,
1940 * do the comparison in process context in raid10d
1941 */
1943 }
1944 }
1947 {
1953 }
1956 {
1957 /* reshape read bio isn't allocated from r10buf_pool */
1961 }
1964 {
1969 /* the primary of several recovery bios */
1974 else
1983 else
1986 }
1987 }
1988 }
1991 {
1996 sector_t first_bad;
2005 else
2017 }
2027 }
2029 /*
2030 * Note: sync and recover and handled very differently for raid10
2031 * This code is for resync.
2032 * For resync, we read through virtual addresses and read all blocks.
2033 * If there is any error, we schedule a write. The lowest numbered
2034 * drive is authoritative.
2035 * However requests come for physical address, so we need to map.
2036 * For every physical address there are raid_disks/copies virtual addresses,
2037 * which is always are least one, but is not necessarly an integer.
2038 * This means that a physical address can span multiple chunks, so we may
2039 * have to submit multiple io requests for a single sync request.
2040 */
2041 /*
2042 * We check if all blocks are in-sync and only write to blocks that
2043 * aren't in sync
2044 */
2046 {
2055 /* find the first device with a block */
2070 /* now find blocks with errors */
2087 /* We know that the bi_io_vec layout is the same for
2088 * both 'first' and 'i', so we just compare them.
2089 * All vec entries are PAGE_SIZE;
2090 */
2098 len))
2101 }
2106 /* Don't fix anything. */
2109 /* Just give up on this device */
2112 }
2113 /* Ok, we need to write this bio, either to correct an
2114 * inconsistency or to correct an unreadable block.
2115 * First we need to fixup bv_offset, bv_len and
2116 * bi_vecs, as the read request might have corrupted these
2117 */
2140 }
2142 /* Now write out to any replacement devices
2143 * that are active
2144 */
2159 }
2161 done:
2165 }
2166 }
2168 /*
2169 * Now for the recovery code.
2170 * Recovery happens across physical sectors.
2171 * We recover all non-is_sync drives by finding the virtual address of
2172 * each, and then choose a working drive that also has that virt address.
2173 * There is a separate r10_bio for each non-in_sync drive.
2174 * Only the first two slots are in use. The first for reading,
2175 * The second for writing.
2176 *
2177 */
2179 {
2180 /* We got a read error during recovery.
2181 * We repeat the read in smaller page-sized sections.
2182 * If a read succeeds, write it to the new device or record
2183 * a bad block if we cannot.
2184 * If a read fails, record a bad block on both old and
2185 * new devices.
2186 */
2200 sector_t addr;
2209 addr,
2217 addr,
2227 }
2228 }
2230 /* We don't worry if we cannot set a bad block -
2231 * it really is bad so there is no loss in not
2232 * recording it yet
2233 */
2237 /* need bad block on destination too */
2242 /* just abort the recovery */
2251 }
2252 }
2253 }
2257 idx++;
2258 }
2259 }
2262 {
2271 }
2273 /*
2274 * share the pages with the first bio
2275 * and submit the write request
2276 */
2280 /* Need to test wbio2->bi_end_io before we call
2281 * generic_make_request as if the former is NULL,
2282 * the latter is free to free wbio2.
2283 */
2290 }
2296 }
2297 }
2299 /*
2300 * Used by fix_read_error() to decay the per rdev read_errors.
2301 * We halve the read error count for every hour that has elapsed
2302 * since the last recorded read error.
2303 *
2304 */
2306 {
2314 /* first time we've seen a read error */
2317 }
2324 /*
2325 * if hours_since_last is > the number of bits in read_errors
2326 * just set read errors to 0. We do this to avoid
2327 * overflowing the shift of read_errors by hours_since_last.
2328 */
2331 else
2333 }
2337 {
2338 sector_t first_bad;
2345 /* success */
2352 }
2353 /* need to record an error - either for the block or the device */
2357 }
2359 /*
2360 * This is a kernel thread which:
2361 *
2362 * 1. Retries failed read operations on working mirrors.
2363 * 2. Updates the raid superblock when problems encounter.
2364 * 3. Performs writes following reads for array synchronising.
2365 */
2368 {
2375 /* still own a reference to this rdev, so it cannot
2376 * have been cleared recently.
2377 */
2381 /* drive has already been failed, just ignore any
2382 more fix_read_error() attempts */
2399 }
2412 sector_t first_bad;
2426 sect,
2434 }
2435 sl++;
2442 /* Cannot read from anywhere, just mark the block
2443 * as bad on the first device to discourage future
2444 * reads.
2445 */
2450 rdev,
2457 }
2459 }
2462 /* write it back and re-read */
2469 sl--;
2481 sect,
2484 /* Well, this device is dead */
2488 sect +
2490 rdev)),
2495 }
2498 }
2505 sl--;
2517 sect,
2519 READ)) {
2521 /* Well, this device is dead */
2525 sect +
2536 sect +
2540 }
2544 }
2549 }
2550 }
2553 {
2558 /* bio has the data to be written to slot 'i' where
2559 * we just recently had a write error.
2560 * We repeatedly clone the bio and trim down to one block,
2561 * then try the write. Where the write fails we record
2562 * a bad block.
2563 * It is conceivable that the bio doesn't exactly align with
2564 * blocks. We must handle this.
2565 *
2566 * We currently own a reference to the rdev.
2567 */
2570 sector_t sector;
2587 sector_t wsector;
2590 /* Write at 'sector' for 'sectors' */
2600 /* Failure! */
2609 }
2611 }
2614 {
2620 /* we got a read error. Maybe the drive is bad. Maybe just
2621 * the block and we can fix it.
2622 * We freeze all other IO, and try reading the block from
2623 * other devices. When we find one, we re-write
2624 * and check it that fixes the read error.
2625 * This is all done synchronously while the array is
2626 * frozen.
2627 */
2645 }
2648 {
2649 /* Some sort of write request has finished and it
2650 * succeeded in writing where we thought there was a
2651 * bad block. So forget the bad block.
2652 * Or possibly if failed and we need to record
2653 * a bad block.
2654 */
2668 rdev,
2673 rdev,
2677 }
2685 rdev,
2690 rdev,
2694 }
2695 }
2705 rdev,
2715 }
2717 }
2722 rdev,
2726 }
2727 }
2733 /*
2734 * In case freeze_array() is waiting for condition
2735 * nr_pending == nr_queued + extra to be true.
2736 */
2744 }
2745 }
2746 }
2749 {
2767 }
2768 }
2772 retry_list);
2781 }
2782 }
2793 }
2812 else
2818 }
2820 }
2823 {
2838 }
2841 {
2851 else
2864 }
2865 }
2867 }
2869 /*
2870 * Set cluster_sync_high since we need other nodes to add the
2871 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2872 */
2874 {
2875 sector_t window_size;
2878 /*
2879 * First, here we define "stripe" as a unit which across
2880 * all member devices one time, so we get chunks by use
2881 * raid_disks / near_copies. Otherwise, if near_copies is
2882 * close to raid_disks, then resync window could increases
2883 * linearly with the increase of raid_disks, which means
2884 * we will suspend a really large IO window while it is not
2885 * necessary. If raid_disks is not divisible by near_copies,
2886 * an extra chunk is needed to ensure the whole "stripe" is
2887 * covered.
2888 */
2893 else
2897 /*
2898 * At least use a 32M window to align with raid1's resync window
2899 */
2904 }
2906 /*
2907 * perform a "sync" on one "block"
2908 *
2909 * We need to make sure that no normal I/O request - particularly write
2910 * requests - conflict with active sync requests.
2911 *
2912 * This is achieved by tracking pending requests and a 'barrier' concept
2913 * that can be installed to exclude normal IO requests.
2914 *
2915 * Resync and recovery are handled very differently.
2916 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2917 *
2918 * For resync, we iterate over virtual addresses, read all copies,
2919 * and update if there are differences. If only one copy is live,
2920 * skip it.
2921 * For recovery, we iterate over physical addresses, read a good
2922 * value for each non-in_sync drive, and over-write.
2923 *
2924 * So, for recovery we may have several outstanding complex requests for a
2925 * given address, one for each out-of-sync device. We model this by allocating
2926 * a number of r10_bio structures, one for each out-of-sync device.
2927 * As we setup these structures, we collect all bio's together into a list
2928 * which we then process collectively to add pages, and then process again
2929 * to pass to generic_make_request.
2930 *
2931 * The r10_bio structures are linked using a borrowed master_bio pointer.
2932 * This link is counted in ->remaining. When the r10_bio that points to NULL
2933 * has its remaining count decremented to 0, the whole complex operation
2934 * is complete.
2935 *
2936 */
2940 {
2947 sector_t sync_blocks;
2957 /*
2958 * Allow skipping a full rebuild for incremental assembly
2959 * of a clean array, like RAID1 does.
2960 */
2970 }
2972 skipped:
2981 /* If we aborted, we need to abort the
2982 * sync on the 'current' bitmap chucks (there can
2983 * be several when recovering multiple devices).
2984 * as we may have started syncing it but not finished.
2985 * We can find the current address in
2986 * mddev->curr_resync, but for recovery,
2987 * we need to convert that to several
2988 * virtual addresses.
2989 */
2994 }
3001 sector_t sect =
3005 }
3007 /* completed sync */
3011 /* Completed a full sync so the replacements
3012 * are now fully recovered.
3013 */
3020 }
3022 }
3024 }
3029 }
3035 /* if there has been nothing to do on any drive,
3036 * then there is nothing to do at all..
3037 */
3040 }
3045 /* make sure whole request will fit in a chunk - if chunks
3046 * are meaningful
3047 */
3052 /*
3053 * If there is non-resync activity waiting for a turn, then let it
3054 * though before starting on this new sync request.
3055 */
3059 /* Again, very different code for resync and recovery.
3060 * Both must result in an r10bio with a list of bios that
3061 * have bi_end_io, bi_sector, bi_disk set,
3062 * and bi_private set to the r10bio.
3063 * For recovery, we may actually create several r10bios
3064 * with 2 bios in each, that correspond to the bios in the main one.
3065 * In this case, the subordinate r10bios link back through a
3066 * borrowed master_bio pointer, and the counter in the master
3067 * includes a ref from each subordinate.
3068 */
3069 /* First, we decide what to do and set ->bi_end_io
3070 * To end_sync_read if we want to read, and
3071 * end_sync_write if we will want to write.
3072 */
3076 /* recovery... the complicated one */
3083 sector_t sect;
3100 }
3103 /* want to reconstruct this device */
3107 /* last stripe is not complete - don't
3108 * try to recover this sector.
3109 */
3112 }
3115 /* Unless we are doing a full sync, or a replacement
3116 * we only need to recover the block if it is set in
3117 * the bitmap
3118 */
3126 /* yep, skip the sync_blocks here, but don't assume
3127 * that there will never be anything to do here
3128 */
3132 }
3152 /* Need to check if the array will still be
3153 * degraded
3154 */
3162 }
3163 }
3180 /* This is where we read from */
3189 bad_sectors -= (sector
3194 }
3195 }
3208 /* and we write to 'i' (if not in_sync) */
3233 /* and maybe write to replacement */
3237 /* Note: if mreplace != NULL, then bio
3238 * cannot be NULL as r10buf_pool_alloc will
3239 * have allocated it.
3240 * So the second test here is pointless.
3241 * But it keeps semantic-checkers happy, and
3242 * this comment keeps human reviewers
3243 * happy.
3244 */
3257 }
3260 /* Cannot recover, so abort the recovery or
3261 * record a bad block */
3263 /* problem is that there are bad blocks
3264 * on other device(s)
3265 */
3273 mrdev,
3279 mreplace,
3283 }
3289 mirror->recovery_disabled
3291 }
3300 }
3305 /* Only want this if there is elsewhere to
3306 * read from. 'j' is currently the first
3307 * readable copy.
3308 */
3315 targets++;
3316 }
3320 }
3321 }
3328 }
3330 }
3332 /* resync. Schedule a read for every block at this virt offset */
3335 /*
3336 * Since curr_resync_completed could probably not update in
3337 * time, and we will set cluster_sync_low based on it.
3338 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3339 * safety reason, which ensures curr_resync_completed is
3340 * updated in bitmap_cond_end_sync.
3341 */
3350 /* We can skip this block */
3353 }
3386 }
3398 }
3399 }
3410 count++;
3416 }
3419 /* Need to set up for writing to the replacement */
3432 count++;
3434 }
3441 mddev);
3446 mddev);
3447 }
3451 }
3452 }
3467 /*
3468 * won't fail because the vec table is big enough
3469 * to hold all these pages
3470 */
3472 }
3480 /* It is resync not recovery */
3484 /* Send resync message */
3488 }
3490 /* This is recovery not resync */
3495 /*
3496 * sector_nr is a device address for recovery, so we
3497 * need translate it to array address before compare
3498 * with cluster_sync_high.
3499 */
3504 /*
3505 * curr_resync_completed is similar as
3506 * sector_nr, so make the translation too.
3507 */
3514 }
3515 }
3521 }
3522 }
3536 }
3537 }
3540 /* pretend they weren't skipped, it makes
3541 * no important difference in this case
3542 */
3546 giveup:
3547 /* There is nowhere to write, so all non-sync
3548 * drives must be failed or in resync, all drives
3549 * have a bad block, so try the next chunk...
3550 */
3555 chunks_skipped ++;
3558 }
3560 static sector_t
3562 {
3563 sector_t size;
3578 }
3581 {
3582 /* Calculate the number of sectors-per-device that will
3583 * actually be used, and set conf->dev_sectors and
3584 * conf->stride
3585 */
3591 /* 'size' is now the number of chunks in the array */
3592 /* calculate "used chunks per device" */
3595 /* We need to round up when dividing by raid_disks to
3596 * get the stride size.
3597 */
3607 }
3608 }
3612 {
3628 * updated yet. */
3633 }
3651 * actually using this, but leave code here just in case.*/
3661 }
3665 }
3668 {
3680 }
3686 }
3693 /* FIXME calc properly */
3696 GFP_KERNEL);
3723 }
3727 else
3728 /* far_copies must be 1 */
3730 }
3748 out:
3755 }
3757 }
3760 {
3765 sector_t size;
3778 }
3792 }
3793 }
3807 else
3810 }
3831 }
3849 }
3855 else
3858 }
3859 /* need to check that every block has at least one working mirror */
3864 }
3867 /* must ensure that shape change is supported */
3874 }
3880 i++) {
3885 /* The replacement is all we have - use it */
3889 }
3898 }
3904 }
3907 }
3915 /*
3916 * Ok, everything is just fine now
3917 */
3928 /* Calculate max read-ahead size.
3929 * We need to readahead at least twice a whole stripe....
3930 * maybe...
3931 */
3935 }
3949 /* This cannot work */
3952 }
3963 }
3967 out_free_conf:
3974 out:
3976 }
3979 {
3989 }
3992 {
3997 else
3999 }
4002 {
4003 /* Resize of 'far' arrays is not supported.
4004 * For 'near' and 'offset' arrays we can set the
4005 * number of sectors used to be an appropriate multiple
4006 * of the chunk size.
4007 * For 'offset', this is far_copies*chunksize.
4008 * For 'near' the multiplier is the LCM of
4009 * near_copies and raid_disks.
4010 * So if far_copies > 1 && !far_offset, fail.
4011 * Else find LCM(raid_disks, near_copy)*far_copies and
4012 * multiply by chunk_size. Then round to this number.
4013 * This is mostly done by raid10_size()
4014 */
4033 }
4039 }
4044 }
4047 {
4055 }
4058 /* Set new parameters */
4060 /* new layout: far_copies = 1, near_copies = 2 */
4065 /* make sure it will be not marked as dirty */
4075 }
4077 }
4080 }
4083 {
4086 /* raid10 can take over:
4087 * raid0 - providing it has only two drives
4088 */
4090 /* for raid0 takeover only one zone is supported */
4096 }
4100 }
4102 }
4105 {
4106 /* Called when there is a request to change
4107 * - layout (to ->new_layout)
4108 * - chunk size (to ->new_chunk_sectors)
4109 * - raid_disks (by delta_disks)
4110 * or when trying to restart a reshape that was ongoing.
4111 *
4112 * We need to validate the request and possibly allocate
4113 * space if that might be an issue later.
4114 *
4115 * Currently we reject any reshape of a 'far' mode array,
4116 * allow chunk size to change if new is generally acceptable,
4117 * allow raid_disks to increase, and allow
4118 * a switch between 'near' mode and 'offset' mode.
4119 */
4127 /* mustn't change number of copies */
4130 /* Cannot switch to 'far' mode */
4134 /* not factor of array size */
4143 /* allocate new 'mirrors' list */
4147 GFP_KERNEL);
4150 }
4152 }
4154 /*
4155 * Need to check if array has failed when deciding whether to:
4156 * - start an array
4157 * - remove non-faulty devices
4158 * - add a spare
4159 * - allow a reshape
4160 * This determination is simple when no reshape is happening.
4161 * However if there is a reshape, we need to carefully check
4162 * both the before and after sections.
4163 * This is because some failed devices may only affect one
4164 * of the two sections, and some non-in_sync devices may
4165 * be insync in the section most affected by failed devices.
4166 */
4168 {
4174 /* 'prev' section first */
4178 degraded++;
4180 /* When we can reduce the number of devices in
4181 * an array, this might not contribute to
4182 * 'degraded'. It does now.
4183 */
4184 degraded++;
4185 }
4194 degraded2++;
4196 /* If reshape is increasing the number of devices,
4197 * this section has already been recovered, so
4198 * it doesn't contribute to degraded.
4199 * else it does.
4200 */
4202 degraded2++;
4203 }
4204 }
4209 }
4212 {
4213 /* A 'reshape' has been requested. This commits
4214 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4215 * This also checks if there are enough spares and adds them
4216 * to the array.
4217 * We currently require enough spares to make the final
4218 * array non-degraded. We also require that the difference
4219 * between old and new data_offset - on each device - is
4220 * enough that we never risk over-writing.
4221 */
4246 spares++;
4257 }
4258 }
4276 }
4286 }
4300 }
4309 else
4314 }
4317 /* This is a spare that was manually added */
4319 }
4320 }
4321 /* When a reshape changes the number of devices,
4322 * ->degraded is measured against the larger of the
4323 * pre and post numbers.
4324 */
4343 }
4349 abort:
4362 }
4364 /* Calculate the last device-address that could contain
4365 * any block from the chunk that includes the array-address 's'
4366 * and report the next address.
4367 * i.e. the address returned will be chunk-aligned and after
4368 * any data that is in the chunk containing 's'.
4369 */
4371 {
4379 }
4381 /* Calculate the first device-address that could contain
4382 * any block from the chunk that includes the array-address 's'.
4383 * This too will be the start of a chunk
4384 */
4386 {
4393 }
4397 {
4398 /* We simply copy at most one chunk (smallest of old and new)
4399 * at a time, possibly less if that exceeds RESYNC_PAGES,
4400 * or we hit a bad block or something.
4401 * This might mean we pause for normal IO in the middle of
4402 * a chunk, but that is not a problem as mddev->reshape_position
4403 * can record any location.
4404 *
4405 * If we will want to write to a location that isn't
4406 * yet recorded as 'safe' (i.e. in metadata on disk) then
4407 * we need to flush all reshape requests and update the metadata.
4408 *
4409 * When reshaping forwards (e.g. to more devices), we interpret
4410 * 'safe' as the earliest block which might not have been copied
4411 * down yet. We divide this by previous stripe size and multiply
4412 * by previous stripe length to get lowest device offset that we
4413 * cannot write to yet.
4414 * We interpret 'sector_nr' as an address that we want to write to.
4415 * From this we use last_device_address() to find where we might
4416 * write to, and first_device_address on the 'safe' position.
4417 * If this 'next' write position is after the 'safe' position,
4418 * we must update the metadata to increase the 'safe' position.
4419 *
4420 * When reshaping backwards, we round in the opposite direction
4421 * and perform the reverse test: next write position must not be
4422 * less than current safe position.
4423 *
4424 * In all this the minimum difference in data offsets
4425 * (conf->offset_diff - always positive) allows a bit of slack,
4426 * so next can be after 'safe', but not by more than offset_diff
4427 *
4428 * We need to prepare all the bios here before we start any IO
4429 * to ensure the size we choose is acceptable to all devices.
4430 * The means one for each copy for write-out and an extra one for
4431 * read-in.
4432 * We store the read-in bio in ->master_bio and the others in
4433 * ->devs[x].bio and ->devs[x].repl_bio.
4434 */
4449 /* If restarting in the middle, skip the initial sectors */
4462 }
4463 }
4465 /* We don't use sector_nr to track where we are up to
4466 * as that doesn't work well for ->reshape_backwards.
4467 * So just use ->reshape_progress.
4468 */
4470 /* 'next' is the earliest device address that we might
4471 * write to for this chunk in the new layout
4472 */
4476 /* 'safe' is the last device address that we might read from
4477 * in the old layout after a restart
4478 */
4491 /* 'next' is after the last device address that we
4492 * might write to for this chunk in the new layout
4493 */
4496 /* 'safe' is the earliest device address that we might
4497 * read from in the old layout after a restart
4498 */
4501 /* Need to update metadata if 'next' might be beyond 'safe'
4502 * as that would possibly corrupt data
4503 */
4513 }
4517 /* Need to update reshape_position in metadata */
4523 else
4533 }
4536 }
4539 read_more:
4540 /* Now schedule reads for blocks from sector_nr to last */
4553 /* Cannot read from here, so need to record bad blocks
4554 * on all the target devices.
4555 */
4556 // FIXME
4560 }
4577 /* Now find the locations in the new layout */
4594 }
4605 }
4607 /* Now add as many pages as possible to all of these bios. */
4617 /*
4618 * won't fail because the vec table is big enough
4619 * to hold all these pages
4620 */
4622 }
4625 }
4629 /* Now submit the read */
4640 /* Now that we have done the whole section we can
4641 * update reshape_progress
4642 */
4645 else
4649 }
4655 {
4656 /* Reshape read completed. Hopefully we have a block
4657 * to write out.
4658 * If we got a read error then we do sync 1-page reads from
4659 * elsewhere until we find the data - or give up.
4660 */
4666 /* Reshape has been aborted */
4669 }
4671 /* We definitely have the data in the pages, schedule the
4672 * writes.
4673 */
4686 }
4690 }
4697 }
4699 }
4702 {
4714 /* read-ahead size must cover two whole stripes, which is
4715 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4716 */
4723 }
4725 }
4729 {
4730 /* Use sync reads to get the blocks from somewhere else */
4743 }
4745 /* reshape IOs share pages from .devs[0].bio */
4763 sector_t addr;
4773 addr,
4781 failed:
4782 slot++;
4787 }
4790 /* couldn't read this block, must give up */
4795 }
4797 idx++;
4798 }
4801 }
4804 {
4819 }
4822 /* FIXME should record badblock */
4824 }
4828 }
4831 {
4837 }
4840 {
4850 }
4857 d++) {
4864 }
4866 }
4872 }
4875 {
4896 };
4899 {
4901 }
4904 {
4906 }