| blob | 9a08f621b27d49a497cbe2b62d678d3d212d122f |
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>
33 /*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 *
42 * The data to be stored is divided into chunks using chunksize.
43 * Each device is divided into far_copies sections.
44 * In each section, chunks are laid out in a style similar to raid0, but
45 * near_copies copies of each chunk is stored (each on a different drive).
46 * The starting device for each section is offset near_copies from the starting
47 * device of the previous section.
48 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
49 * drive.
50 * near_copies and far_copies must be at least one, and their product is at most
51 * raid_disks.
52 *
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of be very far apart
55 * on disk, there are adjacent stripes.
56 */
58 /*
59 * Number of guaranteed r10bios in case of extreme VM load:
60 */
61 #define NR_RAID10_BIOS 256
63 /* when we get a read error on a read-only array, we redirect to another
64 * device without failing the first device, or trying to over-write to
65 * correct the read error. To keep track of bad blocks on a per-bio
66 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
67 */
68 #define IO_BLOCKED ((struct bio *)1)
69 /* When we successfully write to a known bad-block, we need to remove the
70 * bad-block marking which must be done from process context. So we record
71 * the success by setting devs[n].bio to IO_MADE_GOOD
72 */
73 #define IO_MADE_GOOD ((struct bio *)2)
75 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
77 /* When there are this many requests queued to be written by
78 * the raid10 thread, we become 'congested' to provide back-pressure
79 * for writeback.
80 */
93 {
97 /* allocate a r10bio with room for raid_disks entries in the
98 * bios array */
100 }
103 {
105 }
107 /* Maximum size of each resync request */
108 #define RESYNC_BLOCK_SIZE (64*1024)
109 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
110 /* amount of memory to reserve for resync requests */
111 #define RESYNC_WINDOW (1024*1024)
112 /* maximum number of concurrent requests, memory permitting */
113 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
115 /*
116 * When performing a resync, we need to read and compare, so
117 * we need as many pages are there are copies.
118 * When performing a recovery, we need 2 bios, one for read,
119 * one for write (we recover only one drive per r10buf)
120 *
121 */
123 {
138 else
141 /*
142 * Allocate bios.
143 */
155 }
156 /*
157 * Allocate RESYNC_PAGES data pages and attach them
158 * where needed.
159 */
166 /* we can share bv_page's during recovery
167 * and reshape */
179 }
180 }
184 out_free_pages:
191 out_free_bio:
197 }
200 }
203 {
215 }
217 }
221 }
223 }
226 {
238 }
239 }
242 {
247 }
250 {
256 }
259 {
269 /* wake up frozen array... */
273 }
275 /*
276 * raid_end_bio_io() is called when we have finished servicing a mirrored
277 * operation and are ready to return a success/failure code to the buffer
278 * cache layer.
279 */
281 {
298 /*
299 * Wake up any possible resync thread that waits for the device
300 * to go idle.
301 */
303 }
305 }
307 /*
308 * Update disk head position estimator based on IRQ completion info.
309 */
311 {
316 }
318 /*
319 * Find the disk number which triggered given bio
320 */
323 {
333 }
334 }
344 }
347 {
358 /*
359 * this branch is our 'one mirror IO has finished' event handler:
360 */
364 /*
365 * Set R10BIO_Uptodate in our master bio, so that
366 * we will return a good error code to the higher
367 * levels even if IO on some other mirrored buffer fails.
368 *
369 * The 'master' represents the composite IO operation to
370 * user-side. So if something waits for IO, then it will
371 * wait for the 'master' bio.
372 */
375 /* If all other devices that store this block have
376 * failed, we want to return the error upwards rather
377 * than fail the last device. Here we redefine
378 * "uptodate" to mean "Don't want to retry"
379 */
385 }
390 /*
391 * oops, read error - keep the refcount on the rdev
392 */
401 }
402 }
405 {
406 /* clear the bitmap if all writes complete successfully */
412 }
415 {
423 else
425 }
426 }
427 }
430 {
447 }
448 /*
449 * this branch is our 'one mirror IO has finished' event handler:
450 */
453 /* Never record new bad blocks to replacement,
454 * just fail it.
455 */
464 }
466 /*
467 * Set R10BIO_Uptodate in our master bio, so that
468 * we will return a good error code for to the higher
469 * levels even if IO on some other mirrored buffer fails.
470 *
471 * The 'master' represents the composite IO operation to
472 * user-side. So if something waits for IO, then it will
473 * wait for the 'master' bio.
474 */
475 sector_t first_bad;
480 /* Maybe we can clear some bad blocks. */
488 else
492 }
493 }
495 /*
496 *
497 * Let's see if all mirrored write operations have finished
498 * already.
499 */
503 }
505 /*
506 * RAID10 layout manager
507 * As well as the chunksize and raid_disks count, there are two
508 * parameters: near_copies and far_copies.
509 * near_copies * far_copies must be <= raid_disks.
510 * Normally one of these will be 1.
511 * If both are 1, we get raid0.
512 * If near_copies == raid_disks, we get raid1.
513 *
514 * Chunks are laid out in raid0 style with near_copies copies of the
515 * first chunk, followed by near_copies copies of the next chunk and
516 * so on.
517 * If far_copies > 1, then after 1/far_copies of the array has been assigned
518 * as described above, we start again with a device offset of near_copies.
519 * So we effectively have another copy of the whole array further down all
520 * the drives, but with blocks on different drives.
521 * With this layout, and block is never stored twice on the one device.
522 *
523 * raid10_find_phys finds the sector offset of a given virtual sector
524 * on each device that it is on.
525 *
526 * raid10_find_virt does the reverse mapping, from a device and a
527 * sector offset to a virtual address
528 */
531 {
533 sector_t sector;
534 sector_t chunk;
535 sector_t stripe;
539 /* now calculate first sector/dev */
551 /* and calculate all the others */
557 slot++;
566 slot++;
567 }
568 dev++;
572 }
573 }
574 }
577 {
589 }
592 {
594 /* Never use conf->prev as this is only called during resync
595 * or recovery, so reshape isn't happening
596 */
612 else
614 }
616 }
620 }
622 /**
623 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
624 * @q: request queue
625 * @bvm: properties of new bio
626 * @biovec: the request that could be merged to it.
627 *
628 * Return amount of bytes we can accept at this offset
629 * This requires checking for end-of-chunk if near_copies != raid_disks,
630 * and for subordinate merge_bvec_fns if merge_check_needed.
631 */
635 {
654 /* bio_add cannot handle a negative return */
669 /* Cannot give any guidance during reshape */
673 }
690 }
691 }
702 }
703 }
704 }
706 }
708 }
710 /*
711 * This routine returns the disk from which the requested read should
712 * be done. There is a per-array 'next expected sequential IO' sector
713 * number - if this matches on the next IO then we use the last disk.
714 * There is also a per-disk 'last know head position' sector that is
715 * maintained from IRQ contexts, both the normal and the resync IO
716 * completion handlers update this position correctly. If there is no
717 * perfect sequential match then we pick the disk whose head is closest.
718 *
719 * If there are 2 mirrors in the same 2 devices, performance degrades
720 * because position is mirror, not device based.
721 *
722 * The rdev for the device selected will have nr_pending incremented.
723 */
725 /*
726 * FIXME: possibly should rethink readbalancing and do it differently
727 * depending on near_copies / far_copies geometry.
728 */
732 {
745 retry:
752 /*
753 * Check if we can balance. We can balance on the whole
754 * device if no resync is going on (recovery is ok), or below
755 * the resync window. We take the first readable disk when
756 * above the resync window.
757 */
763 sector_t first_bad;
765 sector_t dev_sector;
787 /* Already have a better slot */
790 /* Cannot read here. If this is the
791 * 'primary' device, then we must not read
792 * beyond 'bad_sectors' from another device.
793 */
800 sector_t good_sectors =
806 }
808 /* Must read from here */
810 }
818 /* This optimisation is debatable, and completely destroys
819 * sequential read speed for 'far copies' arrays. So only
820 * keep it for 'near' arrays, and review those later.
821 */
825 /* for far > 1 always use the lowest address */
828 else
835 }
836 }
840 }
845 /* Cannot risk returning a device that failed
846 * before we inc'ed nr_pending
847 */
850 }
858 }
861 {
873 i++) {
879 }
880 }
883 }
887 {
892 }
895 {
896 /* Any writes that have been queued but are awaiting
897 * bitmap updates get flushed here.
898 */
906 /* flush any pending bitmap writes to disk
907 * before proceeding w/ I/O */
916 /* Just ignore it */
918 else
921 }
924 }
926 /* Barriers....
927 * Sometimes we need to suspend IO while we do something else,
928 * either some resync/recovery, or reconfigure the array.
929 * To do this we raise a 'barrier'.
930 * The 'barrier' is a counter that can be raised multiple times
931 * to count how many activities are happening which preclude
932 * normal IO.
933 * We can only raise the barrier if there is no pending IO.
934 * i.e. if nr_pending == 0.
935 * We choose only to raise the barrier if no-one is waiting for the
936 * barrier to go down. This means that as soon as an IO request
937 * is ready, no other operations which require a barrier will start
938 * until the IO request has had a chance.
939 *
940 * So: regular IO calls 'wait_barrier'. When that returns there
941 * is no backgroup IO happening, It must arrange to call
942 * allow_barrier when it has finished its IO.
943 * backgroup IO calls must call raise_barrier. Once that returns
944 * there is no normal IO happeing. It must arrange to call
945 * lower_barrier when the particular background IO completes.
946 */
949 {
953 /* Wait until no block IO is waiting (unless 'force') */
957 /* block any new IO from starting */
960 /* Now wait for all pending IO to complete */
966 }
969 {
975 }
978 {
982 /* Wait for the barrier to drop.
983 * However if there are already pending
984 * requests (preventing the barrier from
985 * rising completely), and the
986 * pre-process bio queue isn't empty,
987 * then don't wait, as we need to empty
988 * that queue to get the nr_pending
989 * count down.
990 */
998 }
1001 }
1004 {
1010 }
1013 {
1014 /* stop syncio and normal IO and wait for everything to
1015 * go quiet.
1016 * We increment barrier and nr_waiting, and then
1017 * wait until nr_pending match nr_queued+1
1018 * This is called in the context of one normal IO request
1019 * that has failed. Thus any sync request that might be pending
1020 * will be blocked by nr_pending, and we need to wait for
1021 * pending IO requests to complete or be queued for re-try.
1022 * Thus the number queued (nr_queued) plus this request (1)
1023 * must match the number of pending IOs (nr_pending) before
1024 * we continue.
1025 */
1035 }
1038 {
1039 /* reverse the effect of the freeze */
1045 }
1049 {
1053 else
1055 }
1061 };
1064 {
1066 cb);
1079 }
1081 /* we aren't scheduling, so we can do the write-out directly. */
1091 }
1093 }
1096 {
1119 }
1121 /* If this request crosses a chunk boundary, we need to
1122 * split it. This will only happen for 1 PAGE (or less) requests.
1123 */
1125 > chunk_sects
1129 /* Sanity check -- queue functions should prevent this happening */
1133 /* This is a one page bio that upper layers
1134 * refuse to split for us, so we need to split it.
1135 */
1139 /* Each of these 'make_request' calls will call 'wait_barrier'.
1140 * If the first succeeds but the second blocks due to the resync
1141 * thread raising the barrier, we will deadlock because the
1142 * IO to the underlying device will be queued in generic_make_request
1143 * and will never complete, so will never reduce nr_pending.
1144 * So increment nr_waiting here so no new raise_barriers will
1145 * succeed, and so the second wait_barrier cannot block.
1146 */
1161 bad_map:
1168 }
1172 /*
1173 * Register the new request and wait if the reconstruction
1174 * thread has put up a bar for new requests.
1175 * Continue immediately if no resync is active currently.
1176 */
1183 /* IO spans the reshape position. Need to wait for
1184 * reshape to pass
1185 */
1191 }
1199 /* Need to update reshape_position in metadata */
1208 }
1219 /* We might need to issue multiple reads to different
1220 * devices if there are bad blocks around, so we keep
1221 * track of the number of reads in bio->bi_phys_segments.
1222 * If this is 0, there is only one r10_bio and no locking
1223 * will be needed when the request completes. If it is
1224 * non-zero, then it is the number of not-completed requests.
1225 */
1230 /*
1231 * read balancing logic:
1232 */
1236 read_again:
1241 }
1246 max_sectors);
1259 /* Could not read all from this device, so we will
1260 * need another r10_bio.
1261 */
1268 else
1271 /* Cannot call generic_make_request directly
1272 * as that will be queued in __generic_make_request
1273 * and subsequent mempool_alloc might block
1274 * waiting for it. so hand bio over to raid10d.
1275 */
1290 }
1292 /*
1293 * WRITE:
1294 */
1299 }
1300 /* first select target devices under rcu_lock and
1301 * inc refcount on their rdev. Record them by setting
1302 * bios[x] to bio
1303 * If there are known/acknowledged bad blocks on any device
1304 * on which we have seen a write error, we want to avoid
1305 * writing to those blocks. This potentially requires several
1306 * writes to write around the bad blocks. Each set of writes
1307 * gets its own r10_bio with a set of bios attached. The number
1308 * of r10_bios is recored in bio->bi_phys_segments just as with
1309 * the read case.
1310 */
1314 retry_write:
1330 }
1335 }
1346 }
1348 sector_t first_bad;
1354 max_sectors,
1357 /* Mustn't write here until the bad block
1358 * is acknowledged
1359 */
1364 }
1366 /* Cannot write here at all */
1369 /* Mustn't write more than bad_sectors
1370 * to other devices yet
1371 */
1373 /* We don't set R10BIO_Degraded as that
1374 * only applies if the disk is missing,
1375 * so it might be re-added, and we want to
1376 * know to recover this chunk.
1377 * In this case the device is here, and the
1378 * fact that this chunk is not in-sync is
1379 * recorded in the bad block log.
1380 */
1382 }
1387 }
1388 }
1394 }
1395 }
1399 /* Have to wait for this device to get unblocked, then retry */
1407 }
1413 /* Race with remove_disk */
1416 }
1418 }
1419 }
1424 }
1427 /* We are splitting this into multiple parts, so
1428 * we need to prepare for allocating another r10_bio.
1429 */
1434 else
1437 }
1451 max_sectors);
1467 else
1476 }
1486 max_sectors);
1489 /* We are actively writing to the original device
1490 * so it cannot disappear, so the replacement cannot
1491 * become NULL here
1492 */
1495 r10_bio,
1509 }
1511 /* Don't remove the bias on 'remaining' (one_write_done) until
1512 * after checking if we need to go around again.
1513 */
1517 /* We need another r10_bio. It has already been counted
1518 * in bio->bi_phys_segments.
1519 */
1529 }
1532 /* In case raid10d snuck in to freeze_array */
1534 }
1537 {
1548 else
1550 }
1558 }
1560 /* check if there are enough drives for
1561 * every block to appear on atleast one.
1562 * Don't consider the device numbered 'ignore'
1563 * as we might be about to remove it.
1564 */
1566 {
1576 cnt++;
1578 }
1584 }
1587 {
1590 }
1593 {
1597 /*
1598 * If it is not operational, then we have already marked it as dead
1599 * else if it is the last working disks, ignore the error, let the
1600 * next level up know.
1601 * else mark the drive as failed
1602 */
1605 /*
1606 * Don't fail the drive, just return an IO error.
1607 */
1614 /*
1615 * if recovery is running, make sure it aborts.
1616 */
1618 }
1627 }
1630 {
1638 }
1650 }
1651 }
1654 {
1660 }
1663 {
1670 /*
1671 * Find all non-in_sync disks within the RAID10 configuration
1672 * and mark them in_sync
1673 */
1680 /* Replacement has just become active */
1683 count++;
1685 /* Replaced device not technically faulty,
1686 * but we need to be sure it gets removed
1687 * and never re-added.
1688 */
1692 }
1697 count++;
1699 }
1700 }
1707 }
1711 {
1720 /* only hot-add to in-sync arrays, as recovery is
1721 * very different from resync
1722 */
1733 }
1738 else
1757 }
1770 }
1772 /* Some requests might not have seen this new
1773 * merge_bvec_fn. We must wait for them to complete
1774 * before merging the device fully.
1775 * First we make sure any code which has tested
1776 * our function has submitted the request, then
1777 * we wait for all outstanding requests to complete.
1778 */
1783 }
1790 }
1793 {
1805 else
1812 }
1813 /* Only remove faulty devices if recovery
1814 * is not possible.
1815 */
1823 }
1827 /* lost the race, try later */
1832 /* We must have just cleared 'rdev' */
1836 * but will never see neither -- if they are careful.
1837 */
1841 /* We might have just remove the Replacement as faulty
1842 * Clear the flag just in case
1843 */
1848 abort:
1852 }
1856 {
1862 /* this is a reshape read */
1869 else
1870 /* The write handler will notice the lack of
1871 * R10BIO_Uptodate and record any errors etc
1872 */
1876 /* for reconstruct, we always reschedule after a read.
1877 * for resync, only after all reads
1878 */
1882 /* we have read all the blocks,
1883 * do the comparison in process context in raid10d
1884 */
1886 }
1887 }
1890 {
1895 /* the primary of several recovery bios */
1900 else
1909 else
1912 }
1913 }
1914 }
1917 {
1923 sector_t first_bad;
1932 else
1944 }
1954 }
1956 /*
1957 * Note: sync and recover and handled very differently for raid10
1958 * This code is for resync.
1959 * For resync, we read through virtual addresses and read all blocks.
1960 * If there is any error, we schedule a write. The lowest numbered
1961 * drive is authoritative.
1962 * However requests come for physical address, so we need to map.
1963 * For every physical address there are raid_disks/copies virtual addresses,
1964 * which is always are least one, but is not necessarly an integer.
1965 * This means that a physical address can span multiple chunks, so we may
1966 * have to submit multiple io requests for a single sync request.
1967 */
1968 /*
1969 * We check if all blocks are in-sync and only write to blocks that
1970 * aren't in sync
1971 */
1973 {
1981 /* find the first device with a block */
1993 /* now find blocks with errors */
2004 /* We know that the bi_io_vec layout is the same for
2005 * both 'first' and 'i', so we just compare them.
2006 * All vec entries are PAGE_SIZE;
2007 */
2017 /* Don't fix anything. */
2019 }
2020 /* Ok, we need to write this bio, either to correct an
2021 * inconsistency or to correct an unreadable block.
2022 * First we need to fixup bv_offset, bv_len and
2023 * bi_vecs, as the read request might have corrupted these
2024 */
2042 PAGE_SIZE);
2043 }
2054 }
2056 /* Now write out to any replacement devices
2057 * that are active
2058 */
2070 PAGE_SIZE);
2076 }
2078 done:
2082 }
2083 }
2085 /*
2086 * Now for the recovery code.
2087 * Recovery happens across physical sectors.
2088 * We recover all non-is_sync drives by finding the virtual address of
2089 * each, and then choose a working drive that also has that virt address.
2090 * There is a separate r10_bio for each non-in_sync drive.
2091 * Only the first two slots are in use. The first for reading,
2092 * The second for writing.
2093 *
2094 */
2096 {
2097 /* We got a read error during recovery.
2098 * We repeat the read in smaller page-sized sections.
2099 * If a read succeeds, write it to the new device or record
2100 * a bad block if we cannot.
2101 * If a read fails, record a bad block on both old and
2102 * new devices.
2103 */
2116 sector_t addr;
2125 addr,
2133 addr,
2143 }
2144 }
2146 /* We don't worry if we cannot set a bad block -
2147 * it really is bad so there is no loss in not
2148 * recording it yet
2149 */
2153 /* need bad block on destination too */
2158 /* just abort the recovery */
2160 "md/raid10:%s: recovery aborted"
2169 }
2170 }
2171 }
2175 idx++;
2176 }
2177 }
2180 {
2189 }
2191 /*
2192 * share the pages with the first bio
2193 * and submit the write request
2194 */
2202 }
2208 }
2209 }
2212 /*
2213 * Used by fix_read_error() to decay the per rdev read_errors.
2214 * We halve the read error count for every hour that has elapsed
2215 * since the last recorded read error.
2216 *
2217 */
2219 {
2228 /* first time we've seen a read error */
2231 }
2238 /*
2239 * if hours_since_last is > the number of bits in read_errors
2240 * just set read errors to 0. We do this to avoid
2241 * overflowing the shift of read_errors by hours_since_last.
2242 */
2245 else
2247 }
2251 {
2252 sector_t first_bad;
2259 /* success */
2266 }
2267 /* need to record an error - either for the block or the device */
2271 }
2273 /*
2274 * This is a kernel thread which:
2275 *
2276 * 1. Retries failed read operations on working mirrors.
2277 * 2. Updates the raid superblock when problems encounter.
2278 * 3. Performs writes following reads for array synchronising.
2279 */
2282 {
2289 /* still own a reference to this rdev, so it cannot
2290 * have been cleared recently.
2291 */
2295 /* drive has already been failed, just ignore any
2296 more fix_read_error() attempts */
2306 "md/raid10:%s: %s: Raid device exceeded "
2316 }
2329 sector_t first_bad;
2343 sect,
2350 }
2351 sl++;
2358 /* Cannot read from anywhere, just mark the block
2359 * as bad on the first device to discourage future
2360 * reads.
2361 */
2366 rdev,
2373 }
2375 }
2378 /* write it back and re-read */
2385 sl--;
2397 sect,
2400 /* Well, this device is dead */
2402 "md/raid10:%s: read correction "
2403 "write failed"
2407 sect +
2409 rdev)),
2415 }
2418 }
2425 sl--;
2436 sect,
2438 READ)) {
2440 /* Well, this device is dead */
2442 "md/raid10:%s: unable to read back "
2443 "corrected sectors"
2447 sect +
2457 "md/raid10:%s: read error corrected"
2461 sect +
2465 }
2469 }
2474 }
2475 }
2478 {
2480 }
2483 {
2494 }
2497 {
2502 /* bio has the data to be written to slot 'i' where
2503 * we just recently had a write error.
2504 * We repeatedly clone the bio and trim down to one block,
2505 * then try the write. Where the write fails we record
2506 * a bad block.
2507 * It is conceivable that the bio doesn't exactly align with
2508 * blocks. We must handle this.
2509 *
2510 * We currently own a reference to the rdev.
2511 */
2514 sector_t sector;
2532 /* Write at 'sector' for 'sectors' */
2540 /* Failure! */
2549 }
2551 }
2554 {
2563 /* we got a read error. Maybe the drive is bad. Maybe just
2564 * the block and we can fix it.
2565 * We freeze all other IO, and try reading the block from
2566 * other devices. When we find one, we re-write
2567 * and check it that fixes the read error.
2568 * This is all done synchronously while the array is
2569 * frozen.
2570 */
2585 read_more:
2594 }
2599 KERN_ERR
2600 "md/raid10:%s: %s: redirecting "
2609 max_sectors);
2619 /* Drat - have to split this up more */
2628 else
2634 GFP_NOIO);
2648 }
2651 {
2652 /* Some sort of write request has finished and it
2653 * succeeded in writing where we thought there was a
2654 * bad block. So forget the bad block.
2655 * Or possibly if failed and we need to record
2656 * a bad block.
2657 */
2671 rdev,
2676 rdev,
2680 }
2687 rdev,
2692 rdev,
2696 }
2697 }
2706 rdev,
2716 }
2718 }
2723 rdev,
2727 }
2728 }
2733 }
2734 }
2737 {
2756 }
2776 /* just a partial read to be scheduled from a
2777 * separate context
2778 */
2781 }
2786 }
2788 }
2792 {
2807 }
2809 /*
2810 * perform a "sync" on one "block"
2811 *
2812 * We need to make sure that no normal I/O request - particularly write
2813 * requests - conflict with active sync requests.
2814 *
2815 * This is achieved by tracking pending requests and a 'barrier' concept
2816 * that can be installed to exclude normal IO requests.
2817 *
2818 * Resync and recovery are handled very differently.
2819 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2820 *
2821 * For resync, we iterate over virtual addresses, read all copies,
2822 * and update if there are differences. If only one copy is live,
2823 * skip it.
2824 * For recovery, we iterate over physical addresses, read a good
2825 * value for each non-in_sync drive, and over-write.
2826 *
2827 * So, for recovery we may have several outstanding complex requests for a
2828 * given address, one for each out-of-sync device. We model this by allocating
2829 * a number of r10_bio structures, one for each out-of-sync device.
2830 * As we setup these structures, we collect all bio's together into a list
2831 * which we then process collectively to add pages, and then process again
2832 * to pass to generic_make_request.
2833 *
2834 * The r10_bio structures are linked using a borrowed master_bio pointer.
2835 * This link is counted in ->remaining. When the r10_bio that points to NULL
2836 * has its remaining count decremented to 0, the whole complex operation
2837 * is complete.
2838 *
2839 */
2843 {
2850 sector_t sync_blocks;
2859 skipped:
2865 /* If we aborted, we need to abort the
2866 * sync on the 'current' bitmap chucks (there can
2867 * be several when recovering multiple devices).
2868 * as we may have started syncing it but not finished.
2869 * We can find the current address in
2870 * mddev->curr_resync, but for recovery,
2871 * we need to convert that to several
2872 * virtual addresses.
2873 */
2877 }
2884 sector_t sect =
2888 }
2890 /* completed sync */
2894 /* Completed a full sync so the replacements
2895 * are now fully recovered.
2896 */
2900 ->recovery_offset
2902 }
2904 }
2909 }
2915 /* if there has been nothing to do on any drive,
2916 * then there is nothing to do at all..
2917 */
2920 }
2925 /* make sure whole request will fit in a chunk - if chunks
2926 * are meaningful
2927 */
2931 /*
2932 * If there is non-resync activity waiting for us then
2933 * put in a delay to throttle resync.
2934 */
2938 /* Again, very different code for resync and recovery.
2939 * Both must result in an r10bio with a list of bios that
2940 * have bi_end_io, bi_sector, bi_bdev set,
2941 * and bi_private set to the r10bio.
2942 * For recovery, we may actually create several r10bios
2943 * with 2 bios in each, that correspond to the bios in the main one.
2944 * In this case, the subordinate r10bios link back through a
2945 * borrowed master_bio pointer, and the counter in the master
2946 * includes a ref from each subordinate.
2947 */
2948 /* First, we decide what to do and set ->bi_end_io
2949 * To end_sync_read if we want to read, and
2950 * end_sync_write if we will want to write.
2951 */
2955 /* recovery... the complicated one */
2962 sector_t sect;
2969 &&
2976 /* want to reconstruct this device */
2980 /* last stripe is not complete - don't
2981 * try to recover this sector.
2982 */
2984 }
2985 /* Unless we are doing a full sync, or a replacement
2986 * we only need to recover the block if it is set in
2987 * the bitmap
2988 */
2996 /* yep, skip the sync_blocks here, but don't assume
2997 * that there will never be anything to do here
2998 */
3001 }
3016 /* Need to check if the array will still be
3017 * degraded
3018 */
3024 }
3040 /* This is where we read from */
3050 bad_sectors -= (sector
3055 }
3056 }
3067 /* and we write to 'i' (if not in_sync) */
3094 /* and maybe write to replacement */
3099 /* Note: if rdev != NULL, then bio
3100 * cannot be NULL as r10buf_pool_alloc will
3101 * have allocated it.
3102 * So the second test here is pointless.
3103 * But it keeps semantic-checkers happy, and
3104 * this comment keeps human reviewers
3105 * happy.
3106 */
3119 }
3121 /* Cannot recover, so abort the recovery or
3122 * record a bad block */
3128 /* problem is that there are bad blocks
3129 * on other device(s)
3130 */
3148 }
3155 mirror->recovery_disabled
3157 }
3159 }
3160 }
3167 }
3169 }
3171 /* resync. Schedule a read for every block at this virt offset */
3180 /* We can skip this block */
3183 }
3224 }
3225 }
3236 count++;
3243 /* Need to set up for writing to the replacement */
3257 count++;
3258 }
3265 mddev);
3270 mddev);
3271 }
3275 }
3276 }
3287 }
3305 /* stop here */
3310 /* remove last page from this bio */
3314 }
3316 }
3320 bio_full:
3334 }
3335 }
3338 /* pretend they weren't skipped, it makes
3339 * no important difference in this case
3340 */
3344 giveup:
3345 /* There is nowhere to write, so all non-sync
3346 * drives must be failed or in resync, all drives
3347 * have a bad block, so try the next chunk...
3348 */
3353 chunks_skipped ++;
3356 }
3358 static sector_t
3360 {
3361 sector_t size;
3376 }
3379 {
3380 /* Calculate the number of sectors-per-device that will
3381 * actually be used, and set conf->dev_sectors and
3382 * conf->stride
3383 */
3389 /* 'size' is now the number of chunks in the array */
3390 /* calculate "used chunks per device" */
3393 /* We need to round up when dividing by raid_disks to
3394 * get the stride size.
3395 */
3405 }
3406 }
3410 {
3426 * updated yet. */
3431 }
3447 }
3450 {
3463 }
3469 }
3476 /* FIXME calc properly */
3479 GFP_KERNEL);
3502 }
3506 else
3507 /* far_copies must be 1 */
3509 }
3523 out:
3533 }
3535 }
3538 {
3543 sector_t size;
3553 }
3568 else
3571 }
3593 }
3613 }
3617 else
3620 /* need to check that every block has at least one working mirror */
3625 }
3628 /* must ensure that shape change is supported */
3635 }
3641 i++) {
3646 /* The replacement is all we have - use it */
3650 }
3658 }
3660 }
3670 /*
3671 * Ok, everything is just fine now
3672 */
3684 /* Calculate max read-ahead size.
3685 * We need to readahead at least twice a whole stripe....
3686 * maybe...
3687 */
3692 }
3707 /* This cannot work */
3710 }
3720 }
3724 out_free_conf:
3732 out:
3734 }
3737 {
3745 /* the unplug fn references 'conf'*/
3754 }
3757 {
3767 }
3768 }
3771 {
3772 /* Resize of 'far' arrays is not supported.
3773 * For 'near' and 'offset' arrays we can set the
3774 * number of sectors used to be an appropriate multiple
3775 * of the chunk size.
3776 * For 'offset', this is far_copies*chunksize.
3777 * For 'near' the multiplier is the LCM of
3778 * near_copies and raid_disks.
3779 * So if far_copies > 1 && !far_offset, fail.
3780 * Else find LCM(raid_disks, near_copy)*far_copies and
3781 * multiply by chunk_size. Then round to this number.
3782 * This is mostly done by raid10_size()
3783 */
3802 }
3810 }
3815 }
3818 {
3826 }
3828 /* Set new parameters */
3830 /* new layout: far_copies = 1, near_copies = 2 */
3835 /* make sure it will be not marked as dirty */
3844 }
3847 }
3850 {
3853 /* raid10 can take over:
3854 * raid0 - providing it has only two drives
3855 */
3857 /* for raid0 takeover only one zone is supported */
3864 }
3866 }
3868 }
3871 {
3872 /* Called when there is a request to change
3873 * - layout (to ->new_layout)
3874 * - chunk size (to ->new_chunk_sectors)
3875 * - raid_disks (by delta_disks)
3876 * or when trying to restart a reshape that was ongoing.
3877 *
3878 * We need to validate the request and possibly allocate
3879 * space if that might be an issue later.
3880 *
3881 * Currently we reject any reshape of a 'far' mode array,
3882 * allow chunk size to change if new is generally acceptable,
3883 * allow raid_disks to increase, and allow
3884 * a switch between 'near' mode and 'offset' mode.
3885 */
3893 /* mustn't change number of copies */
3896 /* Cannot switch to 'far' mode */
3900 /* not factor of array size */
3909 /* allocate new 'mirrors' list */
3914 GFP_KERNEL);
3917 }
3919 }
3921 /*
3922 * Need to check if array has failed when deciding whether to:
3923 * - start an array
3924 * - remove non-faulty devices
3925 * - add a spare
3926 * - allow a reshape
3927 * This determination is simple when no reshape is happening.
3928 * However if there is a reshape, we need to carefully check
3929 * both the before and after sections.
3930 * This is because some failed devices may only affect one
3931 * of the two sections, and some non-in_sync devices may
3932 * be insync in the section most affected by failed devices.
3933 */
3935 {
3941 /* 'prev' section first */
3945 degraded++;
3947 /* When we can reduce the number of devices in
3948 * an array, this might not contribute to
3949 * 'degraded'. It does now.
3950 */
3951 degraded++;
3952 }
3961 degraded2++;
3963 /* If reshape is increasing the number of devices,
3964 * this section has already been recovered, so
3965 * it doesn't contribute to degraded.
3966 * else it does.
3967 */
3969 degraded2++;
3970 }
3971 }
3976 }
3979 {
3980 /* A 'reshape' has been requested. This commits
3981 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3982 * This also checks if there are enough spares and adds them
3983 * to the array.
3984 * We currently require enough spares to make the final
3985 * array non-degraded. We also require that the difference
3986 * between old and new data_offset - on each device - is
3987 * enough that we never risk over-writing.
3988 */
4013 spares++;
4023 }
4024 }
4042 }
4052 }
4066 }
4075 else
4080 }
4083 /* This is a spare that was manually added */
4085 }
4086 }
4087 /* When a reshape changes the number of devices,
4088 * ->degraded is measured against the larger of the
4089 * pre and post numbers.
4090 */
4108 }
4114 abort:
4126 }
4128 /* Calculate the last device-address that could contain
4129 * any block from the chunk that includes the array-address 's'
4130 * and report the next address.
4131 * i.e. the address returned will be chunk-aligned and after
4132 * any data that is in the chunk containing 's'.
4133 */
4135 {
4143 }
4145 /* Calculate the first device-address that could contain
4146 * any block from the chunk that includes the array-address 's'.
4147 * This too will be the start of a chunk
4148 */
4150 {
4157 }
4161 {
4162 /* We simply copy at most one chunk (smallest of old and new)
4163 * at a time, possibly less if that exceeds RESYNC_PAGES,
4164 * or we hit a bad block or something.
4165 * This might mean we pause for normal IO in the middle of
4166 * a chunk, but that is not a problem was mddev->reshape_position
4167 * can record any location.
4168 *
4169 * If we will want to write to a location that isn't
4170 * yet recorded as 'safe' (i.e. in metadata on disk) then
4171 * we need to flush all reshape requests and update the metadata.
4172 *
4173 * When reshaping forwards (e.g. to more devices), we interpret
4174 * 'safe' as the earliest block which might not have been copied
4175 * down yet. We divide this by previous stripe size and multiply
4176 * by previous stripe length to get lowest device offset that we
4177 * cannot write to yet.
4178 * We interpret 'sector_nr' as an address that we want to write to.
4179 * From this we use last_device_address() to find where we might
4180 * write to, and first_device_address on the 'safe' position.
4181 * If this 'next' write position is after the 'safe' position,
4182 * we must update the metadata to increase the 'safe' position.
4183 *
4184 * When reshaping backwards, we round in the opposite direction
4185 * and perform the reverse test: next write position must not be
4186 * less than current safe position.
4187 *
4188 * In all this the minimum difference in data offsets
4189 * (conf->offset_diff - always positive) allows a bit of slack,
4190 * so next can be after 'safe', but not by more than offset_disk
4191 *
4192 * We need to prepare all the bios here before we start any IO
4193 * to ensure the size we choose is acceptable to all devices.
4194 * The means one for each copy for write-out and an extra one for
4195 * read-in.
4196 * We store the read-in bio in ->master_bio and the others in
4197 * ->devs[x].bio and ->devs[x].repl_bio.
4198 */
4212 /* If restarting in the middle, skip the initial sectors */
4225 }
4226 }
4228 /* We don't use sector_nr to track where we are up to
4229 * as that doesn't work well for ->reshape_backwards.
4230 * So just use ->reshape_progress.
4231 */
4233 /* 'next' is the earliest device address that we might
4234 * write to for this chunk in the new layout
4235 */
4239 /* 'safe' is the last device address that we might read from
4240 * in the old layout after a restart
4241 */
4254 /* 'next' is after the last device address that we
4255 * might write to for this chunk in the new layout
4256 */
4259 /* 'safe' is the earliest device address that we might
4260 * read from in the old layout after a restart
4261 */
4264 /* Need to update metadata if 'next' might be beyond 'safe'
4265 * as that would possibly corrupt data
4266 */
4276 }
4280 /* Need to update reshape_position in metadata */
4286 else
4295 }
4297 read_more:
4298 /* Now schedule reads for blocks from sector_nr to last */
4310 /* Cannot read from here, so need to record bad blocks
4311 * on all the target devices.
4312 */
4313 // FIXME
4316 }
4334 /* Now find the locations in the new layout */
4350 }
4365 }
4367 /* Now add as many pages as possible to all of these bios. */
4380 /* Didn't fit, must stop */
4384 /* Remove last page from this bio */
4388 }
4390 }
4393 }
4394 bio_full:
4397 /* Now submit the read */
4407 /* Now that we have done the whole section we can
4408 * update reshape_progress
4409 */
4412 else
4416 }
4422 {
4423 /* Reshape read completed. Hopefully we have a block
4424 * to write out.
4425 * If we got a read error then we do sync 1-page reads from
4426 * elsewhere until we find the data - or give up.
4427 */
4433 /* Reshape has been aborted */
4436 }
4438 /* We definitely have the data in the pages, schedule the
4439 * writes.
4440 */
4452 }
4460 }
4462 }
4465 {
4476 /* read-ahead size must cover two whole stripes, which is
4477 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4478 */
4485 }
4487 }
4492 {
4493 /* Use sync reads to get the blocks from somewhere else */
4519 sector_t addr;
4527 addr,
4533 failed:
4534 slot++;
4539 }
4541 /* couldn't read this block, must give up */
4545 }
4547 idx++;
4548 }
4550 }
4553 {
4569 }
4572 /* FIXME should record badblock */
4574 }
4578 }
4581 {
4587 }
4590 {
4602 }
4610 d++) {
4617 }
4618 }
4624 }
4627 {
4647 };
4650 {
4652 }
4655 {
4657 }