| blob | 8da6282254c3e822a27702c469b9afce720bda43 |
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 there are this many requests queue to be written by
64 * the raid10 thread, we become 'congested' to provide back-pressure
65 * for writeback.
66 */
79 {
83 /* allocate a r10bio with room for raid_disks entries in the
84 * bios array */
86 }
89 {
91 }
93 /* Maximum size of each resync request */
94 #define RESYNC_BLOCK_SIZE (64*1024)
95 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
96 /* amount of memory to reserve for resync requests */
97 #define RESYNC_WINDOW (1024*1024)
98 /* maximum number of concurrent requests, memory permitting */
99 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
101 /*
102 * When performing a resync, we need to read and compare, so
103 * we need as many pages are there are copies.
104 * When performing a recovery, we need 2 bios, one for read,
105 * one for write (we recover only one drive per r10buf)
106 *
107 */
109 {
124 else
127 /*
128 * Allocate bios.
129 */
141 }
142 /*
143 * Allocate RESYNC_PAGES data pages and attach them
144 * where needed.
145 */
152 /* we can share bv_page's during recovery
153 * and reshape */
165 }
166 }
170 out_free_pages:
177 out_free_bio:
183 }
186 }
189 {
201 }
203 }
207 }
209 }
212 {
224 }
225 }
228 {
233 }
236 {
242 }
245 {
255 /* wake up frozen array... */
259 }
261 /*
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
264 * cache layer.
265 */
267 {
284 /*
285 * Wake up any possible resync thread that waits for the device
286 * to go idle.
287 */
289 }
291 }
293 /*
294 * Update disk head position estimator based on IRQ completion info.
295 */
297 {
302 }
304 /*
305 * Find the disk number which triggered given bio
306 */
309 {
319 }
320 }
330 }
333 {
344 /*
345 * this branch is our 'one mirror IO has finished' event handler:
346 */
350 /*
351 * Set R10BIO_Uptodate in our master bio, so that
352 * we will return a good error code to the higher
353 * levels even if IO on some other mirrored buffer fails.
354 *
355 * The 'master' represents the composite IO operation to
356 * user-side. So if something waits for IO, then it will
357 * wait for the 'master' bio.
358 */
361 /* If all other devices that store this block have
362 * failed, we want to return the error upwards rather
363 * than fail the last device. Here we redefine
364 * "uptodate" to mean "Don't want to retry"
365 */
371 }
376 /*
377 * oops, read error - keep the refcount on the rdev
378 */
387 }
388 }
391 {
392 /* clear the bitmap if all writes complete successfully */
398 }
401 {
409 else
411 }
412 }
413 }
416 {
433 }
434 /*
435 * this branch is our 'one mirror IO has finished' event handler:
436 */
439 /* Never record new bad blocks to replacement,
440 * just fail it.
441 */
450 }
452 /*
453 * Set R10BIO_Uptodate in our master bio, so that
454 * we will return a good error code for to the higher
455 * levels even if IO on some other mirrored buffer fails.
456 *
457 * The 'master' represents the composite IO operation to
458 * user-side. So if something waits for IO, then it will
459 * wait for the 'master' bio.
460 */
461 sector_t first_bad;
466 /* Maybe we can clear some bad blocks. */
474 else
478 }
479 }
481 /*
482 *
483 * Let's see if all mirrored write operations have finished
484 * already.
485 */
489 }
491 /*
492 * RAID10 layout manager
493 * As well as the chunksize and raid_disks count, there are two
494 * parameters: near_copies and far_copies.
495 * near_copies * far_copies must be <= raid_disks.
496 * Normally one of these will be 1.
497 * If both are 1, we get raid0.
498 * If near_copies == raid_disks, we get raid1.
499 *
500 * Chunks are laid out in raid0 style with near_copies copies of the
501 * first chunk, followed by near_copies copies of the next chunk and
502 * so on.
503 * If far_copies > 1, then after 1/far_copies of the array has been assigned
504 * as described above, we start again with a device offset of near_copies.
505 * So we effectively have another copy of the whole array further down all
506 * the drives, but with blocks on different drives.
507 * With this layout, and block is never stored twice on the one device.
508 *
509 * raid10_find_phys finds the sector offset of a given virtual sector
510 * on each device that it is on.
511 *
512 * raid10_find_virt does the reverse mapping, from a device and a
513 * sector offset to a virtual address
514 */
517 {
519 sector_t sector;
520 sector_t chunk;
521 sector_t stripe;
525 /* now calculate first sector/dev */
537 /* and calculate all the others */
543 slot++;
552 slot++;
553 }
554 dev++;
558 }
559 }
560 }
563 {
575 }
578 {
580 /* Never use conf->prev as this is only called during resync
581 * or recovery, so reshape isn't happening
582 */
598 else
600 }
602 }
606 }
608 /**
609 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
610 * @q: request queue
611 * @bvm: properties of new bio
612 * @biovec: the request that could be merged to it.
613 *
614 * Return amount of bytes we can accept at this offset
615 * This requires checking for end-of-chunk if near_copies != raid_disks,
616 * and for subordinate merge_bvec_fns if merge_check_needed.
617 */
621 {
640 /* bio_add cannot handle a negative return */
651 /* Cannot give any guidance during reshape */
655 }
672 }
673 }
684 }
685 }
686 }
688 }
690 }
692 /*
693 * This routine returns the disk from which the requested read should
694 * be done. There is a per-array 'next expected sequential IO' sector
695 * number - if this matches on the next IO then we use the last disk.
696 * There is also a per-disk 'last know head position' sector that is
697 * maintained from IRQ contexts, both the normal and the resync IO
698 * completion handlers update this position correctly. If there is no
699 * perfect sequential match then we pick the disk whose head is closest.
700 *
701 * If there are 2 mirrors in the same 2 devices, performance degrades
702 * because position is mirror, not device based.
703 *
704 * The rdev for the device selected will have nr_pending incremented.
705 */
707 /*
708 * FIXME: possibly should rethink readbalancing and do it differently
709 * depending on near_copies / far_copies geometry.
710 */
714 {
727 retry:
734 /*
735 * Check if we can balance. We can balance on the whole
736 * device if no resync is going on (recovery is ok), or below
737 * the resync window. We take the first readable disk when
738 * above the resync window.
739 */
745 sector_t first_bad;
747 sector_t dev_sector;
769 /* Already have a better slot */
772 /* Cannot read here. If this is the
773 * 'primary' device, then we must not read
774 * beyond 'bad_sectors' from another device.
775 */
782 sector_t good_sectors =
788 }
790 /* Must read from here */
792 }
800 /* This optimisation is debatable, and completely destroys
801 * sequential read speed for 'far copies' arrays. So only
802 * keep it for 'near' arrays, and review those later.
803 */
807 /* for far > 1 always use the lowest address */
810 else
817 }
818 }
822 }
827 /* Cannot risk returning a device that failed
828 * before we inc'ed nr_pending
829 */
832 }
840 }
843 {
858 i++) {
864 }
865 }
868 }
871 {
872 /* Any writes that have been queued but are awaiting
873 * bitmap updates get flushed here.
874 */
882 /* flush any pending bitmap writes to disk
883 * before proceeding w/ I/O */
892 }
895 }
897 /* Barriers....
898 * Sometimes we need to suspend IO while we do something else,
899 * either some resync/recovery, or reconfigure the array.
900 * To do this we raise a 'barrier'.
901 * The 'barrier' is a counter that can be raised multiple times
902 * to count how many activities are happening which preclude
903 * normal IO.
904 * We can only raise the barrier if there is no pending IO.
905 * i.e. if nr_pending == 0.
906 * We choose only to raise the barrier if no-one is waiting for the
907 * barrier to go down. This means that as soon as an IO request
908 * is ready, no other operations which require a barrier will start
909 * until the IO request has had a chance.
910 *
911 * So: regular IO calls 'wait_barrier'. When that returns there
912 * is no backgroup IO happening, It must arrange to call
913 * allow_barrier when it has finished its IO.
914 * backgroup IO calls must call raise_barrier. Once that returns
915 * there is no normal IO happeing. It must arrange to call
916 * lower_barrier when the particular background IO completes.
917 */
920 {
924 /* Wait until no block IO is waiting (unless 'force') */
928 /* block any new IO from starting */
931 /* Now wait for all pending IO to complete */
937 }
940 {
946 }
949 {
953 /* Wait for the barrier to drop.
954 * However if there are already pending
955 * requests (preventing the barrier from
956 * rising completely), and the
957 * pre-process bio queue isn't empty,
958 * then don't wait, as we need to empty
959 * that queue to get the nr_pending
960 * count down.
961 */
968 );
970 }
973 }
976 {
982 }
985 {
986 /* stop syncio and normal IO and wait for everything to
987 * go quiet.
988 * We increment barrier and nr_waiting, and then
989 * wait until nr_pending match nr_queued+1
990 * This is called in the context of one normal IO request
991 * that has failed. Thus any sync request that might be pending
992 * will be blocked by nr_pending, and we need to wait for
993 * pending IO requests to complete or be queued for re-try.
994 * Thus the number queued (nr_queued) plus this request (1)
995 * must match the number of pending IOs (nr_pending) before
996 * we continue.
997 */
1007 }
1010 {
1011 /* reverse the effect of the freeze */
1017 }
1021 {
1025 else
1027 }
1030 {
1049 }
1051 /* If this request crosses a chunk boundary, we need to
1052 * split it. This will only happen for 1 PAGE (or less) requests.
1053 */
1055 > chunk_sects
1059 /* Sanity check -- queue functions should prevent this happening */
1063 /* This is a one page bio that upper layers
1064 * refuse to split for us, so we need to split it.
1065 */
1069 /* Each of these 'make_request' calls will call 'wait_barrier'.
1070 * If the first succeeds but the second blocks due to the resync
1071 * thread raising the barrier, we will deadlock because the
1072 * IO to the underlying device will be queued in generic_make_request
1073 * and will never complete, so will never reduce nr_pending.
1074 * So increment nr_waiting here so no new raise_barriers will
1075 * succeed, and so the second wait_barrier cannot block.
1076 */
1091 bad_map:
1098 }
1102 /*
1103 * Register the new request and wait if the reconstruction
1104 * thread has put up a bar for new requests.
1105 * Continue immediately if no resync is active currently.
1106 */
1113 /* IO spans the reshape position. Need to wait for
1114 * reshape to pass
1115 */
1121 }
1129 /* Need to update reshape_position in metadata */
1138 }
1149 /* We might need to issue multiple reads to different
1150 * devices if there are bad blocks around, so we keep
1151 * track of the number of reads in bio->bi_phys_segments.
1152 * If this is 0, there is only one r10_bio and no locking
1153 * will be needed when the request completes. If it is
1154 * non-zero, then it is the number of not-completed requests.
1155 */
1160 /*
1161 * read balancing logic:
1162 */
1166 read_again:
1171 }
1176 max_sectors);
1189 /* Could not read all from this device, so we will
1190 * need another r10_bio.
1191 */
1198 else
1201 /* Cannot call generic_make_request directly
1202 * as that will be queued in __generic_make_request
1203 * and subsequent mempool_alloc might block
1204 * waiting for it. so hand bio over to raid10d.
1205 */
1220 }
1222 /*
1223 * WRITE:
1224 */
1229 }
1230 /* first select target devices under rcu_lock and
1231 * inc refcount on their rdev. Record them by setting
1232 * bios[x] to bio
1233 * If there are known/acknowledged bad blocks on any device
1234 * on which we have seen a write error, we want to avoid
1235 * writing to those blocks. This potentially requires several
1236 * writes to write around the bad blocks. Each set of writes
1237 * gets its own r10_bio with a set of bios attached. The number
1238 * of r10_bios is recored in bio->bi_phys_segments just as with
1239 * the read case.
1240 */
1244 retry_write:
1260 }
1265 }
1276 }
1278 sector_t first_bad;
1284 max_sectors,
1287 /* Mustn't write here until the bad block
1288 * is acknowledged
1289 */
1294 }
1296 /* Cannot write here at all */
1299 /* Mustn't write more than bad_sectors
1300 * to other devices yet
1301 */
1303 /* We don't set R10BIO_Degraded as that
1304 * only applies if the disk is missing,
1305 * so it might be re-added, and we want to
1306 * know to recover this chunk.
1307 * In this case the device is here, and the
1308 * fact that this chunk is not in-sync is
1309 * recorded in the bad block log.
1310 */
1312 }
1317 }
1318 }
1324 }
1325 }
1329 /* Have to wait for this device to get unblocked, then retry */
1337 }
1343 /* Race with remove_disk */
1346 }
1348 }
1349 }
1354 }
1357 /* We are splitting this into multiple parts, so
1358 * we need to prepare for allocating another r10_bio.
1359 */
1364 else
1367 }
1381 max_sectors);
1405 max_sectors);
1408 /* We are actively writing to the original device
1409 * so it cannot disappear, so the replacement cannot
1410 * become NULL here
1411 */
1414 r10_bio,
1428 }
1430 /* Don't remove the bias on 'remaining' (one_write_done) until
1431 * after checking if we need to go around again.
1432 */
1436 /* We need another r10_bio. It has already been counted
1437 * in bio->bi_phys_segments.
1438 */
1448 }
1451 /* In case raid10d snuck in to freeze_array */
1453 }
1456 {
1467 else
1469 }
1477 }
1479 /* check if there are enough drives for
1480 * every block to appear on atleast one.
1481 * Don't consider the device numbered 'ignore'
1482 * as we might be about to remove it.
1483 */
1485 {
1494 cnt++;
1496 }
1501 }
1504 {
1507 }
1510 {
1514 /*
1515 * If it is not operational, then we have already marked it as dead
1516 * else if it is the last working disks, ignore the error, let the
1517 * next level up know.
1518 * else mark the drive as failed
1519 */
1522 /*
1523 * Don't fail the drive, just return an IO error.
1524 */
1531 /*
1532 * if recovery is running, make sure it aborts.
1533 */
1535 }
1544 }
1547 {
1555 }
1567 }
1568 }
1571 {
1577 }
1580 {
1587 /*
1588 * Find all non-in_sync disks within the RAID10 configuration
1589 * and mark them in_sync
1590 */
1597 /* Replacement has just become active */
1600 count++;
1602 /* Replaced device not technically faulty,
1603 * but we need to be sure it gets removed
1604 * and never re-added.
1605 */
1609 }
1614 count++;
1616 }
1617 }
1624 }
1628 {
1637 /* only hot-add to in-sync arrays, as recovery is
1638 * very different from resync
1639 */
1650 }
1655 else
1674 }
1687 }
1689 /* Some requests might not have seen this new
1690 * merge_bvec_fn. We must wait for them to complete
1691 * before merging the device fully.
1692 * First we make sure any code which has tested
1693 * our function has submitted the request, then
1694 * we wait for all outstanding requests to complete.
1695 */
1700 }
1704 }
1707 {
1719 else
1726 }
1727 /* Only remove faulty devices if recovery
1728 * is not possible.
1729 */
1737 }
1741 /* lost the race, try later */
1746 /* We must have just cleared 'rdev' */
1750 * but will never see neither -- if they are careful.
1751 */
1755 /* We might have just remove the Replacement as faulty
1756 * Clear the flag just in case
1757 */
1762 abort:
1766 }
1770 {
1776 /* this is a reshape read */
1783 else
1784 /* The write handler will notice the lack of
1785 * R10BIO_Uptodate and record any errors etc
1786 */
1790 /* for reconstruct, we always reschedule after a read.
1791 * for resync, only after all reads
1792 */
1796 /* we have read all the blocks,
1797 * do the comparison in process context in raid10d
1798 */
1800 }
1801 }
1804 {
1809 /* the primary of several recovery bios */
1814 else
1823 else
1826 }
1827 }
1828 }
1831 {
1837 sector_t first_bad;
1846 else
1858 }
1868 }
1870 /*
1871 * Note: sync and recover and handled very differently for raid10
1872 * This code is for resync.
1873 * For resync, we read through virtual addresses and read all blocks.
1874 * If there is any error, we schedule a write. The lowest numbered
1875 * drive is authoritative.
1876 * However requests come for physical address, so we need to map.
1877 * For every physical address there are raid_disks/copies virtual addresses,
1878 * which is always are least one, but is not necessarly an integer.
1879 * This means that a physical address can span multiple chunks, so we may
1880 * have to submit multiple io requests for a single sync request.
1881 */
1882 /*
1883 * We check if all blocks are in-sync and only write to blocks that
1884 * aren't in sync
1885 */
1887 {
1895 /* find the first device with a block */
1907 /* now find blocks with errors */
1918 /* We know that the bi_io_vec layout is the same for
1919 * both 'first' and 'i', so we just compare them.
1920 * All vec entries are PAGE_SIZE;
1921 */
1931 /* Don't fix anything. */
1933 }
1934 /* Ok, we need to write this bio, either to correct an
1935 * inconsistency or to correct an unreadable block.
1936 * First we need to fixup bv_offset, bv_len and
1937 * bi_vecs, as the read request might have corrupted these
1938 */
1956 PAGE_SIZE);
1957 }
1968 }
1970 /* Now write out to any replacement devices
1971 * that are active
1972 */
1984 PAGE_SIZE);
1990 }
1992 done:
1996 }
1997 }
1999 /*
2000 * Now for the recovery code.
2001 * Recovery happens across physical sectors.
2002 * We recover all non-is_sync drives by finding the virtual address of
2003 * each, and then choose a working drive that also has that virt address.
2004 * There is a separate r10_bio for each non-in_sync drive.
2005 * Only the first two slots are in use. The first for reading,
2006 * The second for writing.
2007 *
2008 */
2010 {
2011 /* We got a read error during recovery.
2012 * We repeat the read in smaller page-sized sections.
2013 * If a read succeeds, write it to the new device or record
2014 * a bad block if we cannot.
2015 * If a read fails, record a bad block on both old and
2016 * new devices.
2017 */
2030 sector_t addr;
2039 addr,
2047 addr,
2057 }
2058 }
2060 /* We don't worry if we cannot set a bad block -
2061 * it really is bad so there is no loss in not
2062 * recording it yet
2063 */
2067 /* need bad block on destination too */
2072 /* just abort the recovery */
2074 "md/raid10:%s: recovery aborted"
2083 }
2084 }
2085 }
2089 idx++;
2090 }
2091 }
2094 {
2103 }
2105 /*
2106 * share the pages with the first bio
2107 * and submit the write request
2108 */
2116 }
2122 }
2123 }
2126 /*
2127 * Used by fix_read_error() to decay the per rdev read_errors.
2128 * We halve the read error count for every hour that has elapsed
2129 * since the last recorded read error.
2130 *
2131 */
2133 {
2142 /* first time we've seen a read error */
2145 }
2152 /*
2153 * if hours_since_last is > the number of bits in read_errors
2154 * just set read errors to 0. We do this to avoid
2155 * overflowing the shift of read_errors by hours_since_last.
2156 */
2159 else
2161 }
2165 {
2166 sector_t first_bad;
2173 /* success */
2180 }
2181 /* need to record an error - either for the block or the device */
2185 }
2187 /*
2188 * This is a kernel thread which:
2189 *
2190 * 1. Retries failed read operations on working mirrors.
2191 * 2. Updates the raid superblock when problems encounter.
2192 * 3. Performs writes following reads for array synchronising.
2193 */
2196 {
2203 /* still own a reference to this rdev, so it cannot
2204 * have been cleared recently.
2205 */
2209 /* drive has already been failed, just ignore any
2210 more fix_read_error() attempts */
2220 "md/raid10:%s: %s: Raid device exceeded "
2230 }
2243 sector_t first_bad;
2257 sect,
2264 }
2265 sl++;
2272 /* Cannot read from anywhere, just mark the block
2273 * as bad on the first device to discourage future
2274 * reads.
2275 */
2280 rdev,
2287 }
2289 }
2292 /* write it back and re-read */
2299 sl--;
2311 sect,
2314 /* Well, this device is dead */
2316 "md/raid10:%s: read correction "
2317 "write failed"
2321 sect +
2323 rdev)),
2329 }
2332 }
2339 sl--;
2350 sect,
2352 READ)) {
2354 /* Well, this device is dead */
2356 "md/raid10:%s: unable to read back "
2357 "corrected sectors"
2361 sect +
2371 "md/raid10:%s: read error corrected"
2375 sect +
2379 }
2383 }
2388 }
2389 }
2392 {
2394 }
2397 {
2408 }
2411 {
2416 /* bio has the data to be written to slot 'i' where
2417 * we just recently had a write error.
2418 * We repeatedly clone the bio and trim down to one block,
2419 * then try the write. Where the write fails we record
2420 * a bad block.
2421 * It is conceivable that the bio doesn't exactly align with
2422 * blocks. We must handle this.
2423 *
2424 * We currently own a reference to the rdev.
2425 */
2428 sector_t sector;
2446 /* Write at 'sector' for 'sectors' */
2454 /* Failure! */
2463 }
2465 }
2468 {
2477 /* we got a read error. Maybe the drive is bad. Maybe just
2478 * the block and we can fix it.
2479 * We freeze all other IO, and try reading the block from
2480 * other devices. When we find one, we re-write
2481 * and check it that fixes the read error.
2482 * This is all done synchronously while the array is
2483 * frozen.
2484 */
2499 read_more:
2508 }
2513 KERN_ERR
2514 "md/raid10:%s: %s: redirecting "
2523 max_sectors);
2533 /* Drat - have to split this up more */
2542 else
2548 GFP_NOIO);
2562 }
2565 {
2566 /* Some sort of write request has finished and it
2567 * succeeded in writing where we thought there was a
2568 * bad block. So forget the bad block.
2569 * Or possibly if failed and we need to record
2570 * a bad block.
2571 */
2585 rdev,
2590 rdev,
2594 }
2601 rdev,
2606 rdev,
2610 }
2611 }
2620 rdev,
2630 }
2632 }
2637 rdev,
2641 }
2642 }
2647 }
2648 }
2651 {
2670 }
2690 /* just a partial read to be scheduled from a
2691 * separate context
2692 */
2695 }
2700 }
2702 }
2706 {
2721 }
2723 /*
2724 * perform a "sync" on one "block"
2725 *
2726 * We need to make sure that no normal I/O request - particularly write
2727 * requests - conflict with active sync requests.
2728 *
2729 * This is achieved by tracking pending requests and a 'barrier' concept
2730 * that can be installed to exclude normal IO requests.
2731 *
2732 * Resync and recovery are handled very differently.
2733 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2734 *
2735 * For resync, we iterate over virtual addresses, read all copies,
2736 * and update if there are differences. If only one copy is live,
2737 * skip it.
2738 * For recovery, we iterate over physical addresses, read a good
2739 * value for each non-in_sync drive, and over-write.
2740 *
2741 * So, for recovery we may have several outstanding complex requests for a
2742 * given address, one for each out-of-sync device. We model this by allocating
2743 * a number of r10_bio structures, one for each out-of-sync device.
2744 * As we setup these structures, we collect all bio's together into a list
2745 * which we then process collectively to add pages, and then process again
2746 * to pass to generic_make_request.
2747 *
2748 * The r10_bio structures are linked using a borrowed master_bio pointer.
2749 * This link is counted in ->remaining. When the r10_bio that points to NULL
2750 * has its remaining count decremented to 0, the whole complex operation
2751 * is complete.
2752 *
2753 */
2757 {
2764 sector_t sync_blocks;
2773 skipped:
2779 /* If we aborted, we need to abort the
2780 * sync on the 'current' bitmap chucks (there can
2781 * be several when recovering multiple devices).
2782 * as we may have started syncing it but not finished.
2783 * We can find the current address in
2784 * mddev->curr_resync, but for recovery,
2785 * we need to convert that to several
2786 * virtual addresses.
2787 */
2791 }
2798 sector_t sect =
2802 }
2804 /* completed sync */
2808 /* Completed a full sync so the replacements
2809 * are now fully recovered.
2810 */
2814 ->recovery_offset
2816 }
2818 }
2823 }
2829 /* if there has been nothing to do on any drive,
2830 * then there is nothing to do at all..
2831 */
2834 }
2839 /* make sure whole request will fit in a chunk - if chunks
2840 * are meaningful
2841 */
2845 /*
2846 * If there is non-resync activity waiting for us then
2847 * put in a delay to throttle resync.
2848 */
2852 /* Again, very different code for resync and recovery.
2853 * Both must result in an r10bio with a list of bios that
2854 * have bi_end_io, bi_sector, bi_bdev set,
2855 * and bi_private set to the r10bio.
2856 * For recovery, we may actually create several r10bios
2857 * with 2 bios in each, that correspond to the bios in the main one.
2858 * In this case, the subordinate r10bios link back through a
2859 * borrowed master_bio pointer, and the counter in the master
2860 * includes a ref from each subordinate.
2861 */
2862 /* First, we decide what to do and set ->bi_end_io
2863 * To end_sync_read if we want to read, and
2864 * end_sync_write if we will want to write.
2865 */
2869 /* recovery... the complicated one */
2876 sector_t sect;
2883 &&
2890 /* want to reconstruct this device */
2894 /* last stripe is not complete - don't
2895 * try to recover this sector.
2896 */
2898 }
2899 /* Unless we are doing a full sync, or a replacement
2900 * we only need to recover the block if it is set in
2901 * the bitmap
2902 */
2910 /* yep, skip the sync_blocks here, but don't assume
2911 * that there will never be anything to do here
2912 */
2915 }
2930 /* Need to check if the array will still be
2931 * degraded
2932 */
2938 }
2954 /* This is where we read from */
2964 bad_sectors -= (sector
2969 }
2970 }
2981 /* and we write to 'i' (if not in_sync) */
3008 /* and maybe write to replacement */
3013 /* Note: if rdev != NULL, then bio
3014 * cannot be NULL as r10buf_pool_alloc will
3015 * have allocated it.
3016 * So the second test here is pointless.
3017 * But it keeps semantic-checkers happy, and
3018 * this comment keeps human reviewers
3019 * happy.
3020 */
3033 }
3035 /* Cannot recover, so abort the recovery or
3036 * record a bad block */
3042 /* problem is that there are bad blocks
3043 * on other device(s)
3044 */
3062 }
3069 mirror->recovery_disabled
3071 }
3073 }
3074 }
3081 }
3083 }
3085 /* resync. Schedule a read for every block at this virt offset */
3094 /* We can skip this block */
3097 }
3138 }
3139 }
3150 count++;
3157 /* Need to set up for writing to the replacement */
3171 count++;
3172 }
3179 mddev);
3184 mddev);
3185 }
3189 }
3190 }
3201 }
3219 /* stop here */
3224 /* remove last page from this bio */
3228 }
3230 }
3234 bio_full:
3248 }
3249 }
3252 /* pretend they weren't skipped, it makes
3253 * no important difference in this case
3254 */
3258 giveup:
3259 /* There is nowhere to write, so all non-sync
3260 * drives must be failed or in resync, all drives
3261 * have a bad block, so try the next chunk...
3262 */
3267 chunks_skipped ++;
3270 }
3272 static sector_t
3274 {
3275 sector_t size;
3290 }
3293 {
3294 /* Calculate the number of sectors-per-device that will
3295 * actually be used, and set conf->dev_sectors and
3296 * conf->stride
3297 */
3303 /* 'size' is now the number of chunks in the array */
3304 /* calculate "used chunks per device" */
3307 /* We need to round up when dividing by raid_disks to
3308 * get the stride size.
3309 */
3319 }
3320 }
3324 {
3340 * updated yet. */
3345 }
3361 }
3364 {
3377 }
3383 }
3390 /* FIXME calc properly */
3393 GFP_KERNEL);
3416 }
3420 else
3421 /* far_copies must be 1 */
3423 }
3437 out:
3447 }
3449 }
3452 {
3457 sector_t size;
3466 }
3478 else
3502 }
3518 }
3520 /* need to check that every block has at least one working mirror */
3525 }
3528 /* must ensure that shape change is supported */
3535 }
3541 i++) {
3546 /* The replacement is all we have - use it */
3550 }
3558 }
3560 }
3570 /*
3571 * Ok, everything is just fine now
3572 */
3581 /* Calculate max read-ahead size.
3582 * We need to readahead at least twice a whole stripe....
3583 * maybe...
3584 */
3585 {
3591 }
3607 /* This cannot work */
3610 }
3620 }
3624 out_free_conf:
3632 out:
3634 }
3637 {
3651 }
3654 {
3664 }
3665 }
3668 {
3669 /* Resize of 'far' arrays is not supported.
3670 * For 'near' and 'offset' arrays we can set the
3671 * number of sectors used to be an appropriate multiple
3672 * of the chunk size.
3673 * For 'offset', this is far_copies*chunksize.
3674 * For 'near' the multiplier is the LCM of
3675 * near_copies and raid_disks.
3676 * So if far_copies > 1 && !far_offset, fail.
3677 * Else find LCM(raid_disks, near_copy)*far_copies and
3678 * multiply by chunk_size. Then round to this number.
3679 * This is mostly done by raid10_size()
3680 */
3699 }
3707 }
3712 }
3715 {
3723 }
3725 /* Set new parameters */
3727 /* new layout: far_copies = 1, near_copies = 2 */
3732 /* make sure it will be not marked as dirty */
3741 }
3744 }
3747 {
3750 /* raid10 can take over:
3751 * raid0 - providing it has only two drives
3752 */
3754 /* for raid0 takeover only one zone is supported */
3761 }
3763 }
3765 }
3768 {
3769 /* Called when there is a request to change
3770 * - layout (to ->new_layout)
3771 * - chunk size (to ->new_chunk_sectors)
3772 * - raid_disks (by delta_disks)
3773 * or when trying to restart a reshape that was ongoing.
3774 *
3775 * We need to validate the request and possibly allocate
3776 * space if that might be an issue later.
3777 *
3778 * Currently we reject any reshape of a 'far' mode array,
3779 * allow chunk size to change if new is generally acceptable,
3780 * allow raid_disks to increase, and allow
3781 * a switch between 'near' mode and 'offset' mode.
3782 */
3790 /* mustn't change number of copies */
3793 /* Cannot switch to 'far' mode */
3797 /* not factor of array size */
3806 /* allocate new 'mirrors' list */
3811 GFP_KERNEL);
3814 }
3816 }
3818 /*
3819 * Need to check if array has failed when deciding whether to:
3820 * - start an array
3821 * - remove non-faulty devices
3822 * - add a spare
3823 * - allow a reshape
3824 * This determination is simple when no reshape is happening.
3825 * However if there is a reshape, we need to carefully check
3826 * both the before and after sections.
3827 * This is because some failed devices may only affect one
3828 * of the two sections, and some non-in_sync devices may
3829 * be insync in the section most affected by failed devices.
3830 */
3832 {
3838 /* 'prev' section first */
3842 degraded++;
3844 /* When we can reduce the number of devices in
3845 * an array, this might not contribute to
3846 * 'degraded'. It does now.
3847 */
3848 degraded++;
3849 }
3858 degraded2++;
3860 /* If reshape is increasing the number of devices,
3861 * this section has already been recovered, so
3862 * it doesn't contribute to degraded.
3863 * else it does.
3864 */
3866 degraded2++;
3867 }
3868 }
3873 }
3876 {
3877 /* A 'reshape' has been requested. This commits
3878 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3879 * This also checks if there are enough spares and adds them
3880 * to the array.
3881 * We currently require enough spares to make the final
3882 * array non-degraded. We also require that the difference
3883 * between old and new data_offset - on each device - is
3884 * enough that we never risk over-writing.
3885 */
3910 spares++;
3920 }
3921 }
3939 }
3949 }
3963 }
3972 else
3977 }
3980 /* This is a spare that was manually added */
3982 }
3983 }
3984 /* When a reshape changes the number of devices,
3985 * ->degraded is measured against the larger of the
3986 * pre and post numbers.
3987 */
4005 }
4011 abort:
4023 }
4025 /* Calculate the last device-address that could contain
4026 * any block from the chunk that includes the array-address 's'
4027 * and report the next address.
4028 * i.e. the address returned will be chunk-aligned and after
4029 * any data that is in the chunk containing 's'.
4030 */
4032 {
4040 }
4042 /* Calculate the first device-address that could contain
4043 * any block from the chunk that includes the array-address 's'.
4044 * This too will be the start of a chunk
4045 */
4047 {
4054 }
4058 {
4059 /* We simply copy at most one chunk (smallest of old and new)
4060 * at a time, possibly less if that exceeds RESYNC_PAGES,
4061 * or we hit a bad block or something.
4062 * This might mean we pause for normal IO in the middle of
4063 * a chunk, but that is not a problem was mddev->reshape_position
4064 * can record any location.
4065 *
4066 * If we will want to write to a location that isn't
4067 * yet recorded as 'safe' (i.e. in metadata on disk) then
4068 * we need to flush all reshape requests and update the metadata.
4069 *
4070 * When reshaping forwards (e.g. to more devices), we interpret
4071 * 'safe' as the earliest block which might not have been copied
4072 * down yet. We divide this by previous stripe size and multiply
4073 * by previous stripe length to get lowest device offset that we
4074 * cannot write to yet.
4075 * We interpret 'sector_nr' as an address that we want to write to.
4076 * From this we use last_device_address() to find where we might
4077 * write to, and first_device_address on the 'safe' position.
4078 * If this 'next' write position is after the 'safe' position,
4079 * we must update the metadata to increase the 'safe' position.
4080 *
4081 * When reshaping backwards, we round in the opposite direction
4082 * and perform the reverse test: next write position must not be
4083 * less than current safe position.
4084 *
4085 * In all this the minimum difference in data offsets
4086 * (conf->offset_diff - always positive) allows a bit of slack,
4087 * so next can be after 'safe', but not by more than offset_disk
4088 *
4089 * We need to prepare all the bios here before we start any IO
4090 * to ensure the size we choose is acceptable to all devices.
4091 * The means one for each copy for write-out and an extra one for
4092 * read-in.
4093 * We store the read-in bio in ->master_bio and the others in
4094 * ->devs[x].bio and ->devs[x].repl_bio.
4095 */
4109 /* If restarting in the middle, skip the initial sectors */
4122 }
4123 }
4125 /* We don't use sector_nr to track where we are up to
4126 * as that doesn't work well for ->reshape_backwards.
4127 * So just use ->reshape_progress.
4128 */
4130 /* 'next' is the earliest device address that we might
4131 * write to for this chunk in the new layout
4132 */
4136 /* 'safe' is the last device address that we might read from
4137 * in the old layout after a restart
4138 */
4151 /* 'next' is after the last device address that we
4152 * might write to for this chunk in the new layout
4153 */
4156 /* 'safe' is the earliest device address that we might
4157 * read from in the old layout after a restart
4158 */
4161 /* Need to update metadata if 'next' might be beyond 'safe'
4162 * as that would possibly corrupt data
4163 */
4173 }
4177 /* Need to update reshape_position in metadata */
4183 else
4192 }
4194 read_more:
4195 /* Now schedule reads for blocks from sector_nr to last */
4207 /* Cannot read from here, so need to record bad blocks
4208 * on all the target devices.
4209 */
4210 // FIXME
4213 }
4231 /* Now find the locations in the new layout */
4247 }
4262 }
4264 /* Now add as many pages as possible to all of these bios. */
4277 /* Didn't fit, must stop */
4281 /* Remove last page from this bio */
4285 }
4287 }
4290 }
4291 bio_full:
4294 /* Now submit the read */
4304 /* Now that we have done the whole section we can
4305 * update reshape_progress
4306 */
4309 else
4313 }
4319 {
4320 /* Reshape read completed. Hopefully we have a block
4321 * to write out.
4322 * If we got a read error then we do sync 1-page reads from
4323 * elsewhere until we find the data - or give up.
4324 */
4330 /* Reshape has been aborted */
4333 }
4335 /* We definitely have the data in the pages, schedule the
4336 * writes.
4337 */
4349 }
4357 }
4359 }
4362 {
4373 /* read-ahead size must cover two whole stripes, which is
4374 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4375 */
4382 }
4384 }
4389 {
4390 /* Use sync reads to get the blocks from somewhere else */
4412 sector_t addr;
4420 addr,
4426 failed:
4427 slot++;
4432 }
4434 /* couldn't read this block, must give up */
4438 }
4440 idx++;
4441 }
4443 }
4446 {
4462 }
4465 /* FIXME should record badblock */
4467 }
4471 }
4474 {
4480 }
4483 {
4495 }
4503 d++) {
4510 }
4511 }
4517 }
4520 {
4540 };
4543 {
4545 }
4548 {
4550 }