| blob | 2ce079a0b0bdbc286bfe94e5dd927dbbe1ee30a9 |
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 /* amount of memory to reserve for resync requests */
140 #define RESYNC_WINDOW (1024*1024)
141 /* maximum number of concurrent requests, memory permitting */
142 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144 /*
145 * When performing a resync, we need to read and compare, so
146 * we need as many pages are there are copies.
147 * When performing a recovery, we need 2 bios, one for read,
148 * one for write (we recover only one drive per r10buf)
149 *
150 */
152 {
167 else
170 /* allocate once for all bios */
173 else
179 /*
180 * Allocate bios.
181 */
193 }
194 /*
195 * Allocate RESYNC_PAGES data pages and attach them
196 * where needed.
197 */
215 }
222 }
223 }
227 out_free_pages:
232 out_free_bio:
238 }
240 out_free_r10bio:
243 }
246 {
262 }
264 /* resync pages array stored in the 1st bio's .bi_private */
268 }
271 {
283 }
284 }
287 {
292 }
295 {
301 }
304 {
314 /* wake up frozen array... */
318 }
320 /*
321 * raid_end_bio_io() is called when we have finished servicing a mirrored
322 * operation and are ready to return a success/failure code to the buffer
323 * cache layer.
324 */
326 {
334 /*
335 * Wake up any possible resync thread that waits for the device
336 * to go idle.
337 */
341 }
343 /*
344 * Update disk head position estimator based on IRQ completion info.
345 */
347 {
352 }
354 /*
355 * Find the disk number which triggered given bio
356 */
359 {
369 }
370 }
380 }
383 {
393 /*
394 * this branch is our 'one mirror IO has finished' event handler:
395 */
399 /*
400 * Set R10BIO_Uptodate in our master bio, so that
401 * we will return a good error code to the higher
402 * levels even if IO on some other mirrored buffer fails.
403 *
404 * The 'master' represents the composite IO operation to
405 * user-side. So if something waits for IO, then it will
406 * wait for the 'master' bio.
407 */
410 /* If all other devices that store this block have
411 * failed, we want to return the error upwards rather
412 * than fail the last device. Here we redefine
413 * "uptodate" to mean "Don't want to retry"
414 */
418 }
423 /*
424 * oops, read error - keep the refcount on the rdev
425 */
433 }
434 }
437 {
438 /* clear the bitmap if all writes complete successfully */
444 }
447 {
455 else
457 }
458 }
459 }
462 {
482 }
483 /*
484 * this branch is our 'one mirror IO has finished' event handler:
485 */
488 /* Never record new bad blocks to replacement,
489 * just fail it.
490 */
503 /* This is the only remaining device,
504 * We need to retry the write without
505 * FailFast
506 */
512 }
515 }
517 /*
518 * Set R10BIO_Uptodate in our master bio, so that
519 * we will return a good error code for to the higher
520 * levels even if IO on some other mirrored buffer fails.
521 *
522 * The 'master' represents the composite IO operation to
523 * user-side. So if something waits for IO, then it will
524 * wait for the 'master' bio.
525 */
526 sector_t first_bad;
529 /*
530 * Do not set R10BIO_Uptodate if the current device is
531 * rebuilding or Faulty. This is because we cannot use
532 * such device for properly reading the data back (we could
533 * potentially use it, if the current write would have felt
534 * before rdev->recovery_offset, but for simplicity we don't
535 * check this here.
536 */
541 /* Maybe we can clear some bad blocks. */
549 else
553 }
554 }
556 /*
557 *
558 * Let's see if all mirrored write operations have finished
559 * already.
560 */
566 }
568 /*
569 * RAID10 layout manager
570 * As well as the chunksize and raid_disks count, there are two
571 * parameters: near_copies and far_copies.
572 * near_copies * far_copies must be <= raid_disks.
573 * Normally one of these will be 1.
574 * If both are 1, we get raid0.
575 * If near_copies == raid_disks, we get raid1.
576 *
577 * Chunks are laid out in raid0 style with near_copies copies of the
578 * first chunk, followed by near_copies copies of the next chunk and
579 * so on.
580 * If far_copies > 1, then after 1/far_copies of the array has been assigned
581 * as described above, we start again with a device offset of near_copies.
582 * So we effectively have another copy of the whole array further down all
583 * the drives, but with blocks on different drives.
584 * With this layout, and block is never stored twice on the one device.
585 *
586 * raid10_find_phys finds the sector offset of a given virtual sector
587 * on each device that it is on.
588 *
589 * raid10_find_virt does the reverse mapping, from a device and a
590 * sector offset to a virtual address
591 */
594 {
596 sector_t sector;
597 sector_t chunk;
598 sector_t stripe;
609 /* now calculate first sector/dev */
621 /* and calculate all the others */
628 slot++;
642 }
646 slot++;
647 }
648 dev++;
652 }
653 }
654 }
657 {
669 }
672 {
674 /* Never use conf->prev as this is only called during resync
675 * or recovery, so reshape isn't happening
676 */
690 }
691 }
706 else
708 }
710 }
714 }
716 /*
717 * This routine returns the disk from which the requested read should
718 * be done. There is a per-array 'next expected sequential IO' sector
719 * number - if this matches on the next IO then we use the last disk.
720 * There is also a per-disk 'last know head position' sector that is
721 * maintained from IRQ contexts, both the normal and the resync IO
722 * completion handlers update this position correctly. If there is no
723 * perfect sequential match then we pick the disk whose head is closest.
724 *
725 * If there are 2 mirrors in the same 2 devices, performance degrades
726 * because position is mirror, not device based.
727 *
728 * The rdev for the device selected will have nr_pending incremented.
729 */
731 /*
732 * FIXME: possibly should rethink readbalancing and do it differently
733 * depending on near_copies / far_copies geometry.
734 */
738 {
758 /*
759 * Check if we can balance. We can balance on the whole
760 * device if no resync is going on (recovery is ok), or below
761 * the resync window. We take the first readable disk when
762 * above the resync window.
763 */
769 sector_t first_bad;
771 sector_t dev_sector;
791 /* Already have a better slot */
794 /* Cannot read here. If this is the
795 * 'primary' device, then we must not read
796 * beyond 'bad_sectors' from another device.
797 */
804 sector_t good_sectors =
810 }
812 /* Must read from here */
814 }
823 /* At least 2 disks to choose from so failfast is OK */
825 /* This optimisation is debatable, and completely destroys
826 * sequential read speed for 'far copies' arrays. So only
827 * keep it for 'near' arrays, and review those later.
828 */
832 /* for far > 1 always use the lowest address */
835 else
842 }
843 }
847 }
858 }
861 {
873 i++) {
879 }
880 }
883 }
886 {
887 /* Any writes that have been queued but are awaiting
888 * bitmap updates get flushed here.
889 */
900 /* flush any pending bitmap writes to disk
901 * before proceeding w/ I/O */
914 /* Just ignore it */
916 else
919 }
923 }
925 /* Barriers....
926 * Sometimes we need to suspend IO while we do something else,
927 * either some resync/recovery, or reconfigure the array.
928 * To do this we raise a 'barrier'.
929 * The 'barrier' is a counter that can be raised multiple times
930 * to count how many activities are happening which preclude
931 * normal IO.
932 * We can only raise the barrier if there is no pending IO.
933 * i.e. if nr_pending == 0.
934 * We choose only to raise the barrier if no-one is waiting for the
935 * barrier to go down. This means that as soon as an IO request
936 * is ready, no other operations which require a barrier will start
937 * until the IO request has had a chance.
938 *
939 * So: regular IO calls 'wait_barrier'. When that returns there
940 * is no backgroup IO happening, It must arrange to call
941 * allow_barrier when it has finished its IO.
942 * backgroup IO calls must call raise_barrier. Once that returns
943 * there is no normal IO happeing. It must arrange to call
944 * lower_barrier when the particular background IO completes.
945 */
948 {
952 /* Wait until no block IO is waiting (unless 'force') */
956 /* block any new IO from starting */
959 /* Now wait for all pending IO to complete */
965 }
968 {
974 }
977 {
981 /* Wait for the barrier to drop.
982 * However if there are already pending
983 * requests (preventing the barrier from
984 * rising completely), and the
985 * pre-process bio queue isn't empty,
986 * then don't wait, as we need to empty
987 * that queue to get the nr_pending
988 * count down.
989 */
1001 }
1004 }
1007 {
1011 }
1014 {
1015 /* stop syncio and normal IO and wait for everything to
1016 * go quiet.
1017 * We increment barrier and nr_waiting, and then
1018 * wait until nr_pending match nr_queued+extra
1019 * This is called in the context of one normal IO request
1020 * that has failed. Thus any sync request that might be pending
1021 * will be blocked by nr_pending, and we need to wait for
1022 * pending IO requests to complete or be queued for re-try.
1023 * Thus the number queued (nr_queued) plus this request (extra)
1024 * must match the number of pending IOs (nr_pending) before
1025 * we continue.
1026 */
1038 }
1041 {
1042 /* reverse the effect of the freeze */
1048 }
1052 {
1056 else
1058 }
1064 };
1067 {
1069 cb);
1083 }
1085 /* we aren't scheduling, so we can do the write-out directly. */
1099 /* Just ignore it */
1101 else
1104 }
1106 }
1110 {
1116 sector_t sectors;
1124 /*
1125 * This is an error retry, but we cannot
1126 * safely dereference the rdev in the r10_bio,
1127 * we must use the one in conf.
1128 * If it has already been disconnected (unlikely)
1129 * we lose the device name in error messages.
1130 */
1132 /*
1133 * As we are blocking raid10, it is a little safer to
1134 * use __GFP_HIGH.
1135 */
1145 /* This never gets dereferenced */
1147 }
1149 }
1150 /*
1151 * Register the new request and wait if the reconstruction
1152 * thread has put up a bar for new requests.
1153 * Continue immediately if no resync is active currently.
1154 */
1161 /*
1162 * IO spans the reshape position. Need to wait for reshape to
1163 * pass
1164 */
1170 sectors);
1172 }
1180 }
1183 }
1199 }
1223 }
1228 {
1243 /* Replacement just got moved to main 'rdev' */
1246 }
1253 else
1271 /* flush_pending_writes() needs access to the rdev so...*/
1279 else
1290 }
1291 }
1295 {
1299 sector_t sectors;
1302 /*
1303 * Register the new request and wait if the reconstruction
1304 * thread has put up a bar for new requests.
1305 * Continue immediately if no resync is active currently.
1306 */
1313 /*
1314 * IO spans the reshape position. Need to wait for reshape to
1315 * pass
1316 */
1322 sectors);
1324 }
1332 /* Need to update reshape_position in metadata */
1342 }
1349 }
1350 /* first select target devices under rcu_lock and
1351 * inc refcount on their rdev. Record them by setting
1352 * bios[x] to bio
1353 * If there are known/acknowledged bad blocks on any device
1354 * on which we have seen a write error, we want to avoid
1355 * writing to those blocks. This potentially requires several
1356 * writes to write around the bad blocks. Each set of writes
1357 * gets its own r10_bio with a set of bios attached.
1358 */
1362 retry_write:
1378 }
1383 }
1395 }
1397 sector_t first_bad;
1405 /* Mustn't write here until the bad block
1406 * is acknowledged
1407 */
1412 }
1414 /* Cannot write here at all */
1417 /* Mustn't write more than bad_sectors
1418 * to other devices yet
1419 */
1421 /* We don't set R10BIO_Degraded as that
1422 * only applies if the disk is missing,
1423 * so it might be re-added, and we want to
1424 * know to recover this chunk.
1425 * In this case the device is here, and the
1426 * fact that this chunk is not in-sync is
1427 * recorded in the bad block log.
1428 */
1430 }
1435 }
1436 }
1440 }
1444 }
1445 }
1449 /* Have to wait for this device to get unblocked, then retry */
1457 }
1463 /* Race with remove_disk */
1466 }
1468 }
1469 }
1475 }
1489 }
1499 }
1501 }
1504 {
1520 else
1522 }
1525 {
1534 }
1539 /*
1540 * If this request crosses a chunk boundary, we need to split
1541 * it.
1542 */
1544 sectors > chunk_sects
1553 /* In case raid10d snuck in to freeze_array */
1556 }
1559 {
1570 else
1574 }
1581 }
1584 }
1586 /* check if there are enough drives for
1587 * every block to appear on atleast one.
1588 * Don't consider the device numbered 'ignore'
1589 * as we might be about to remove it.
1590 */
1592 {
1602 }
1614 cnt++;
1616 }
1622 out:
1625 }
1628 {
1629 /* when calling 'enough', both 'prev' and 'geo' must
1630 * be stable.
1631 * This is ensured if ->reconfig_mutex or ->device_lock
1632 * is held.
1633 */
1636 }
1639 {
1644 /*
1645 * If it is not operational, then we have already marked it as dead
1646 * else if it is the last working disks, ignore the error, let the
1647 * next level up know.
1648 * else mark the drive as failed
1649 */
1653 /*
1654 * Don't fail the drive, just return an IO error.
1655 */
1658 }
1661 /*
1662 * If recovery is running, make sure it aborts.
1663 */
1674 }
1677 {
1685 }
1689 /* This is only called with ->reconfix_mutex held, so
1690 * rcu protection of rdev is not needed */
1699 }
1700 }
1703 {
1709 }
1712 {
1719 /*
1720 * Find all non-in_sync disks within the RAID10 configuration
1721 * and mark them in_sync
1722 */
1729 /* Replacement has just become active */
1732 count++;
1734 /* Replaced device not technically faulty,
1735 * but we need to be sure it gets removed
1736 * and never re-added.
1737 */
1741 }
1747 count++;
1749 }
1750 }
1757 }
1760 {
1768 /* only hot-add to in-sync arrays, as recovery is
1769 * very different from resync
1770 */
1785 else
1805 }
1819 }
1825 }
1828 {
1840 else
1847 }
1848 /* Only remove non-faulty devices if recovery
1849 * is not possible.
1850 */
1858 }
1863 /* lost the race, try later */
1867 }
1868 }
1870 /* We must have just cleared 'rdev' */
1874 * but will never see neither -- if they are careful.
1875 */
1877 }
1882 abort:
1886 }
1889 {
1894 else
1895 /* The write handler will notice the lack of
1896 * R10BIO_Uptodate and record any errors etc
1897 */
1901 /* for reconstruct, we always reschedule after a read.
1902 * for resync, only after all reads
1903 */
1907 /* we have read all the blocks,
1908 * do the comparison in process context in raid10d
1909 */
1911 }
1912 }
1915 {
1921 }
1924 {
1925 /* reshape read bio isn't allocated from r10buf_pool */
1929 }
1932 {
1937 /* the primary of several recovery bios */
1942 else
1951 else
1954 }
1955 }
1956 }
1959 {
1964 sector_t first_bad;
1973 else
1985 }
1995 }
1997 /*
1998 * Note: sync and recover and handled very differently for raid10
1999 * This code is for resync.
2000 * For resync, we read through virtual addresses and read all blocks.
2001 * If there is any error, we schedule a write. The lowest numbered
2002 * drive is authoritative.
2003 * However requests come for physical address, so we need to map.
2004 * For every physical address there are raid_disks/copies virtual addresses,
2005 * which is always are least one, but is not necessarly an integer.
2006 * This means that a physical address can span multiple chunks, so we may
2007 * have to submit multiple io requests for a single sync request.
2008 */
2009 /*
2010 * We check if all blocks are in-sync and only write to blocks that
2011 * aren't in sync
2012 */
2014 {
2023 /* find the first device with a block */
2038 /* now find blocks with errors */
2055 /* We know that the bi_io_vec layout is the same for
2056 * both 'first' and 'i', so we just compare them.
2057 * All vec entries are PAGE_SIZE;
2058 */
2066 len))
2069 }
2074 /* Don't fix anything. */
2077 /* Just give up on this device */
2080 }
2081 /* Ok, we need to write this bio, either to correct an
2082 * inconsistency or to correct an unreadable block.
2083 * First we need to fixup bv_offset, bv_len and
2084 * bi_vecs, as the read request might have corrupted these
2085 */
2108 }
2110 /* Now write out to any replacement devices
2111 * that are active
2112 */
2127 }
2129 done:
2133 }
2134 }
2136 /*
2137 * Now for the recovery code.
2138 * Recovery happens across physical sectors.
2139 * We recover all non-is_sync drives by finding the virtual address of
2140 * each, and then choose a working drive that also has that virt address.
2141 * There is a separate r10_bio for each non-in_sync drive.
2142 * Only the first two slots are in use. The first for reading,
2143 * The second for writing.
2144 *
2145 */
2147 {
2148 /* We got a read error during recovery.
2149 * We repeat the read in smaller page-sized sections.
2150 * If a read succeeds, write it to the new device or record
2151 * a bad block if we cannot.
2152 * If a read fails, record a bad block on both old and
2153 * new devices.
2154 */
2168 sector_t addr;
2177 addr,
2185 addr,
2195 }
2196 }
2198 /* We don't worry if we cannot set a bad block -
2199 * it really is bad so there is no loss in not
2200 * recording it yet
2201 */
2205 /* need bad block on destination too */
2210 /* just abort the recovery */
2219 }
2220 }
2221 }
2225 idx++;
2226 }
2227 }
2230 {
2239 }
2241 /*
2242 * share the pages with the first bio
2243 * and submit the write request
2244 */
2248 /* Need to test wbio2->bi_end_io before we call
2249 * generic_make_request as if the former is NULL,
2250 * the latter is free to free wbio2.
2251 */
2258 }
2264 }
2265 }
2267 /*
2268 * Used by fix_read_error() to decay the per rdev read_errors.
2269 * We halve the read error count for every hour that has elapsed
2270 * since the last recorded read error.
2271 *
2272 */
2274 {
2282 /* first time we've seen a read error */
2285 }
2292 /*
2293 * if hours_since_last is > the number of bits in read_errors
2294 * just set read errors to 0. We do this to avoid
2295 * overflowing the shift of read_errors by hours_since_last.
2296 */
2299 else
2301 }
2305 {
2306 sector_t first_bad;
2313 /* success */
2320 }
2321 /* need to record an error - either for the block or the device */
2325 }
2327 /*
2328 * This is a kernel thread which:
2329 *
2330 * 1. Retries failed read operations on working mirrors.
2331 * 2. Updates the raid superblock when problems encounter.
2332 * 3. Performs writes following reads for array synchronising.
2333 */
2336 {
2343 /* still own a reference to this rdev, so it cannot
2344 * have been cleared recently.
2345 */
2349 /* drive has already been failed, just ignore any
2350 more fix_read_error() attempts */
2367 }
2380 sector_t first_bad;
2394 sect,
2402 }
2403 sl++;
2410 /* Cannot read from anywhere, just mark the block
2411 * as bad on the first device to discourage future
2412 * reads.
2413 */
2418 rdev,
2425 }
2427 }
2430 /* write it back and re-read */
2437 sl--;
2449 sect,
2452 /* Well, this device is dead */
2456 sect +
2458 rdev)),
2463 }
2466 }
2473 sl--;
2485 sect,
2487 READ)) {
2489 /* Well, this device is dead */
2493 sect +
2504 sect +
2508 }
2512 }
2517 }
2518 }
2521 {
2526 /* bio has the data to be written to slot 'i' where
2527 * we just recently had a write error.
2528 * We repeatedly clone the bio and trim down to one block,
2529 * then try the write. Where the write fails we record
2530 * a bad block.
2531 * It is conceivable that the bio doesn't exactly align with
2532 * blocks. We must handle this.
2533 *
2534 * We currently own a reference to the rdev.
2535 */
2538 sector_t sector;
2555 sector_t wsector;
2558 /* Write at 'sector' for 'sectors' */
2568 /* Failure! */
2577 }
2579 }
2582 {
2587 sector_t bio_last_sector;
2589 /* we got a read error. Maybe the drive is bad. Maybe just
2590 * the block and we can fix it.
2591 * We freeze all other IO, and try reading the block from
2592 * other devices. When we find one, we re-write
2593 * and check it that fixes the read error.
2594 * This is all done synchronously while the array is
2595 * frozen.
2596 */
2615 }
2618 {
2619 /* Some sort of write request has finished and it
2620 * succeeded in writing where we thought there was a
2621 * bad block. So forget the bad block.
2622 * Or possibly if failed and we need to record
2623 * a bad block.
2624 */
2638 rdev,
2643 rdev,
2647 }
2655 rdev,
2660 rdev,
2664 }
2665 }
2675 rdev,
2685 }
2687 }
2692 rdev,
2696 }
2697 }
2703 /*
2704 * In case freeze_array() is waiting for condition
2705 * nr_pending == nr_queued + extra to be true.
2706 */
2714 }
2715 }
2716 }
2719 {
2737 }
2738 }
2742 retry_list);
2751 }
2752 }
2763 }
2782 else
2788 }
2790 }
2793 {
2808 }
2811 {
2821 else
2834 }
2835 }
2837 }
2839 /*
2840 * perform a "sync" on one "block"
2841 *
2842 * We need to make sure that no normal I/O request - particularly write
2843 * requests - conflict with active sync requests.
2844 *
2845 * This is achieved by tracking pending requests and a 'barrier' concept
2846 * that can be installed to exclude normal IO requests.
2847 *
2848 * Resync and recovery are handled very differently.
2849 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2850 *
2851 * For resync, we iterate over virtual addresses, read all copies,
2852 * and update if there are differences. If only one copy is live,
2853 * skip it.
2854 * For recovery, we iterate over physical addresses, read a good
2855 * value for each non-in_sync drive, and over-write.
2856 *
2857 * So, for recovery we may have several outstanding complex requests for a
2858 * given address, one for each out-of-sync device. We model this by allocating
2859 * a number of r10_bio structures, one for each out-of-sync device.
2860 * As we setup these structures, we collect all bio's together into a list
2861 * which we then process collectively to add pages, and then process again
2862 * to pass to generic_make_request.
2863 *
2864 * The r10_bio structures are linked using a borrowed master_bio pointer.
2865 * This link is counted in ->remaining. When the r10_bio that points to NULL
2866 * has its remaining count decremented to 0, the whole complex operation
2867 * is complete.
2868 *
2869 */
2873 {
2880 sector_t sync_blocks;
2890 /*
2891 * Allow skipping a full rebuild for incremental assembly
2892 * of a clean array, like RAID1 does.
2893 */
2903 }
2905 skipped:
2911 /* If we aborted, we need to abort the
2912 * sync on the 'current' bitmap chucks (there can
2913 * be several when recovering multiple devices).
2914 * as we may have started syncing it but not finished.
2915 * We can find the current address in
2916 * mddev->curr_resync, but for recovery,
2917 * we need to convert that to several
2918 * virtual addresses.
2919 */
2924 }
2931 sector_t sect =
2935 }
2937 /* completed sync */
2941 /* Completed a full sync so the replacements
2942 * are now fully recovered.
2943 */
2950 }
2952 }
2954 }
2959 }
2965 /* if there has been nothing to do on any drive,
2966 * then there is nothing to do at all..
2967 */
2970 }
2975 /* make sure whole request will fit in a chunk - if chunks
2976 * are meaningful
2977 */
2982 /*
2983 * If there is non-resync activity waiting for a turn, then let it
2984 * though before starting on this new sync request.
2985 */
2989 /* Again, very different code for resync and recovery.
2990 * Both must result in an r10bio with a list of bios that
2991 * have bi_end_io, bi_sector, bi_disk set,
2992 * and bi_private set to the r10bio.
2993 * For recovery, we may actually create several r10bios
2994 * with 2 bios in each, that correspond to the bios in the main one.
2995 * In this case, the subordinate r10bios link back through a
2996 * borrowed master_bio pointer, and the counter in the master
2997 * includes a ref from each subordinate.
2998 */
2999 /* First, we decide what to do and set ->bi_end_io
3000 * To end_sync_read if we want to read, and
3001 * end_sync_write if we will want to write.
3002 */
3006 /* recovery... the complicated one */
3013 sector_t sect;
3030 }
3033 /* want to reconstruct this device */
3037 /* last stripe is not complete - don't
3038 * try to recover this sector.
3039 */
3042 }
3045 /* Unless we are doing a full sync, or a replacement
3046 * we only need to recover the block if it is set in
3047 * the bitmap
3048 */
3056 /* yep, skip the sync_blocks here, but don't assume
3057 * that there will never be anything to do here
3058 */
3062 }
3082 /* Need to check if the array will still be
3083 * degraded
3084 */
3092 }
3093 }
3110 /* This is where we read from */
3119 bad_sectors -= (sector
3124 }
3125 }
3138 /* and we write to 'i' (if not in_sync) */
3163 /* and maybe write to replacement */
3167 /* Note: if mreplace != NULL, then bio
3168 * cannot be NULL as r10buf_pool_alloc will
3169 * have allocated it.
3170 * So the second test here is pointless.
3171 * But it keeps semantic-checkers happy, and
3172 * this comment keeps human reviewers
3173 * happy.
3174 */
3187 }
3190 /* Cannot recover, so abort the recovery or
3191 * record a bad block */
3193 /* problem is that there are bad blocks
3194 * on other device(s)
3195 */
3203 mrdev,
3209 mreplace,
3213 }
3219 mirror->recovery_disabled
3221 }
3230 }
3235 /* Only want this if there is elsewhere to
3236 * read from. 'j' is currently the first
3237 * readable copy.
3238 */
3245 targets++;
3246 }
3250 }
3251 }
3258 }
3260 }
3262 /* resync. Schedule a read for every block at this virt offset */
3271 /* We can skip this block */
3274 }
3307 }
3319 }
3320 }
3331 count++;
3337 }
3340 /* Need to set up for writing to the replacement */
3353 count++;
3355 }
3362 mddev);
3367 mddev);
3368 }
3372 }
3373 }
3388 /*
3389 * won't fail because the vec table is big enough
3390 * to hold all these pages
3391 */
3393 }
3411 }
3412 }
3415 /* pretend they weren't skipped, it makes
3416 * no important difference in this case
3417 */
3421 giveup:
3422 /* There is nowhere to write, so all non-sync
3423 * drives must be failed or in resync, all drives
3424 * have a bad block, so try the next chunk...
3425 */
3430 chunks_skipped ++;
3433 }
3435 static sector_t
3437 {
3438 sector_t size;
3453 }
3456 {
3457 /* Calculate the number of sectors-per-device that will
3458 * actually be used, and set conf->dev_sectors and
3459 * conf->stride
3460 */
3466 /* 'size' is now the number of chunks in the array */
3467 /* calculate "used chunks per device" */
3470 /* We need to round up when dividing by raid_disks to
3471 * get the stride size.
3472 */
3482 }
3483 }
3487 {
3503 * updated yet. */
3508 }
3526 * actually using this, but leave code here just in case.*/
3536 }
3540 }
3543 {
3555 }
3561 }
3568 /* FIXME calc properly */
3571 GFP_KERNEL);
3598 }
3602 else
3603 /* far_copies must be 1 */
3605 }
3622 out:
3630 }
3632 }
3635 {
3640 sector_t size;
3653 }
3670 else
3673 }
3694 }
3712 }
3718 else
3721 }
3722 /* need to check that every block has at least one working mirror */
3727 }
3730 /* must ensure that shape change is supported */
3737 }
3743 i++) {
3748 /* The replacement is all we have - use it */
3752 }
3761 }
3767 }
3770 }
3778 /*
3779 * Ok, everything is just fine now
3780 */
3791 /* Calculate max read-ahead size.
3792 * We need to readahead at least twice a whole stripe....
3793 * maybe...
3794 */
3798 }
3812 /* This cannot work */
3815 }
3824 }
3828 out_free_conf:
3835 out:
3837 }
3840 {
3851 }
3854 {
3859 else
3861 }
3864 {
3865 /* Resize of 'far' arrays is not supported.
3866 * For 'near' and 'offset' arrays we can set the
3867 * number of sectors used to be an appropriate multiple
3868 * of the chunk size.
3869 * For 'offset', this is far_copies*chunksize.
3870 * For 'near' the multiplier is the LCM of
3871 * near_copies and raid_disks.
3872 * So if far_copies > 1 && !far_offset, fail.
3873 * Else find LCM(raid_disks, near_copy)*far_copies and
3874 * multiply by chunk_size. Then round to this number.
3875 * This is mostly done by raid10_size()
3876 */
3895 }
3901 }
3906 }
3909 {
3917 }
3920 /* Set new parameters */
3922 /* new layout: far_copies = 1, near_copies = 2 */
3927 /* make sure it will be not marked as dirty */
3937 }
3939 }
3942 }
3945 {
3948 /* raid10 can take over:
3949 * raid0 - providing it has only two drives
3950 */
3952 /* for raid0 takeover only one zone is supported */
3958 }
3962 }
3964 }
3967 {
3968 /* Called when there is a request to change
3969 * - layout (to ->new_layout)
3970 * - chunk size (to ->new_chunk_sectors)
3971 * - raid_disks (by delta_disks)
3972 * or when trying to restart a reshape that was ongoing.
3973 *
3974 * We need to validate the request and possibly allocate
3975 * space if that might be an issue later.
3976 *
3977 * Currently we reject any reshape of a 'far' mode array,
3978 * allow chunk size to change if new is generally acceptable,
3979 * allow raid_disks to increase, and allow
3980 * a switch between 'near' mode and 'offset' mode.
3981 */
3989 /* mustn't change number of copies */
3992 /* Cannot switch to 'far' mode */
3996 /* not factor of array size */
4005 /* allocate new 'mirrors' list */
4010 GFP_KERNEL);
4013 }
4015 }
4017 /*
4018 * Need to check if array has failed when deciding whether to:
4019 * - start an array
4020 * - remove non-faulty devices
4021 * - add a spare
4022 * - allow a reshape
4023 * This determination is simple when no reshape is happening.
4024 * However if there is a reshape, we need to carefully check
4025 * both the before and after sections.
4026 * This is because some failed devices may only affect one
4027 * of the two sections, and some non-in_sync devices may
4028 * be insync in the section most affected by failed devices.
4029 */
4031 {
4037 /* 'prev' section first */
4041 degraded++;
4043 /* When we can reduce the number of devices in
4044 * an array, this might not contribute to
4045 * 'degraded'. It does now.
4046 */
4047 degraded++;
4048 }
4057 degraded2++;
4059 /* If reshape is increasing the number of devices,
4060 * this section has already been recovered, so
4061 * it doesn't contribute to degraded.
4062 * else it does.
4063 */
4065 degraded2++;
4066 }
4067 }
4072 }
4075 {
4076 /* A 'reshape' has been requested. This commits
4077 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4078 * This also checks if there are enough spares and adds them
4079 * to the array.
4080 * We currently require enough spares to make the final
4081 * array non-degraded. We also require that the difference
4082 * between old and new data_offset - on each device - is
4083 * enough that we never risk over-writing.
4084 */
4109 spares++;
4120 }
4121 }
4139 }
4149 }
4164 }
4173 else
4178 }
4181 /* This is a spare that was manually added */
4183 }
4184 }
4185 /* When a reshape changes the number of devices,
4186 * ->degraded is measured against the larger of the
4187 * pre and post numbers.
4188 */
4207 }
4213 abort:
4226 }
4228 /* Calculate the last device-address that could contain
4229 * any block from the chunk that includes the array-address 's'
4230 * and report the next address.
4231 * i.e. the address returned will be chunk-aligned and after
4232 * any data that is in the chunk containing 's'.
4233 */
4235 {
4243 }
4245 /* Calculate the first device-address that could contain
4246 * any block from the chunk that includes the array-address 's'.
4247 * This too will be the start of a chunk
4248 */
4250 {
4257 }
4261 {
4262 /* We simply copy at most one chunk (smallest of old and new)
4263 * at a time, possibly less if that exceeds RESYNC_PAGES,
4264 * or we hit a bad block or something.
4265 * This might mean we pause for normal IO in the middle of
4266 * a chunk, but that is not a problem as mddev->reshape_position
4267 * can record any location.
4268 *
4269 * If we will want to write to a location that isn't
4270 * yet recorded as 'safe' (i.e. in metadata on disk) then
4271 * we need to flush all reshape requests and update the metadata.
4272 *
4273 * When reshaping forwards (e.g. to more devices), we interpret
4274 * 'safe' as the earliest block which might not have been copied
4275 * down yet. We divide this by previous stripe size and multiply
4276 * by previous stripe length to get lowest device offset that we
4277 * cannot write to yet.
4278 * We interpret 'sector_nr' as an address that we want to write to.
4279 * From this we use last_device_address() to find where we might
4280 * write to, and first_device_address on the 'safe' position.
4281 * If this 'next' write position is after the 'safe' position,
4282 * we must update the metadata to increase the 'safe' position.
4283 *
4284 * When reshaping backwards, we round in the opposite direction
4285 * and perform the reverse test: next write position must not be
4286 * less than current safe position.
4287 *
4288 * In all this the minimum difference in data offsets
4289 * (conf->offset_diff - always positive) allows a bit of slack,
4290 * so next can be after 'safe', but not by more than offset_diff
4291 *
4292 * We need to prepare all the bios here before we start any IO
4293 * to ensure the size we choose is acceptable to all devices.
4294 * The means one for each copy for write-out and an extra one for
4295 * read-in.
4296 * We store the read-in bio in ->master_bio and the others in
4297 * ->devs[x].bio and ->devs[x].repl_bio.
4298 */
4313 /* If restarting in the middle, skip the initial sectors */
4326 }
4327 }
4329 /* We don't use sector_nr to track where we are up to
4330 * as that doesn't work well for ->reshape_backwards.
4331 * So just use ->reshape_progress.
4332 */
4334 /* 'next' is the earliest device address that we might
4335 * write to for this chunk in the new layout
4336 */
4340 /* 'safe' is the last device address that we might read from
4341 * in the old layout after a restart
4342 */
4355 /* 'next' is after the last device address that we
4356 * might write to for this chunk in the new layout
4357 */
4360 /* 'safe' is the earliest device address that we might
4361 * read from in the old layout after a restart
4362 */
4365 /* Need to update metadata if 'next' might be beyond 'safe'
4366 * as that would possibly corrupt data
4367 */
4377 }
4381 /* Need to update reshape_position in metadata */
4387 else
4397 }
4400 }
4403 read_more:
4404 /* Now schedule reads for blocks from sector_nr to last */
4417 /* Cannot read from here, so need to record bad blocks
4418 * on all the target devices.
4419 */
4420 // FIXME
4424 }
4441 /* Now find the locations in the new layout */
4458 }
4469 }
4471 /* Now add as many pages as possible to all of these bios. */
4481 /*
4482 * won't fail because the vec table is big enough
4483 * to hold all these pages
4484 */
4486 }
4489 }
4493 /* Now submit the read */
4505 /* Now that we have done the whole section we can
4506 * update reshape_progress
4507 */
4510 else
4514 }
4520 {
4521 /* Reshape read completed. Hopefully we have a block
4522 * to write out.
4523 * If we got a read error then we do sync 1-page reads from
4524 * elsewhere until we find the data - or give up.
4525 */
4531 /* Reshape has been aborted */
4534 }
4536 /* We definitely have the data in the pages, schedule the
4537 * writes.
4538 */
4551 }
4555 }
4562 }
4564 }
4567 {
4579 /* read-ahead size must cover two whole stripes, which is
4580 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4581 */
4588 }
4590 }
4594 {
4595 /* Use sync reads to get the blocks from somewhere else */
4608 }
4610 /* reshape IOs share pages from .devs[0].bio */
4628 sector_t addr;
4638 addr,
4646 failed:
4647 slot++;
4652 }
4655 /* couldn't read this block, must give up */
4660 }
4662 idx++;
4663 }
4666 }
4669 {
4684 }
4687 /* FIXME should record badblock */
4689 }
4693 }
4696 {
4702 }
4705 {
4715 }
4722 d++) {
4729 }
4731 }
4737 }
4740 {
4761 };
4764 {
4766 }
4769 {
4771 }