| blob | 6e8aa213f0d5208d917b8222a56980ab3582b4d6 |
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>
32 /*
33 * RAID10 provides a combination of RAID0 and RAID1 functionality.
34 * The layout of data is defined by
35 * chunk_size
36 * raid_disks
37 * near_copies (stored in low byte of layout)
38 * far_copies (stored in second byte of layout)
39 * far_offset (stored in bit 16 of layout )
40 *
41 * The data to be stored is divided into chunks using chunksize.
42 * Each device is divided into far_copies sections.
43 * In each section, chunks are laid out in a style similar to raid0, but
44 * near_copies copies of each chunk is stored (each on a different drive).
45 * The starting device for each section is offset near_copies from the starting
46 * device of the previous section.
47 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
48 * drive.
49 * near_copies and far_copies must be at least one, and their product is at most
50 * raid_disks.
51 *
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of be very far apart
54 * on disk, there are adjacent stripes.
55 */
57 /*
58 * Number of guaranteed r10bios in case of extreme VM load:
59 */
60 #define NR_RAID10_BIOS 256
62 /* When there are this many requests queue to be written by
63 * the raid10 thread, we become 'congested' to provide back-pressure
64 * for writeback.
65 */
72 {
76 /* allocate a r10bio with room for raid_disks entries in the
77 * bios array */
79 }
82 {
84 }
86 /* Maximum size of each resync request */
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
89 /* amount of memory to reserve for resync requests */
90 #define RESYNC_WINDOW (1024*1024)
91 /* maximum number of concurrent requests, memory permitting */
92 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
94 /*
95 * When performing a resync, we need to read and compare, so
96 * we need as many pages are there are copies.
97 * When performing a recovery, we need 2 bios, one for read,
98 * one for write (we recover only one drive per r10buf)
99 *
100 */
102 {
116 else
119 /*
120 * Allocate bios.
121 */
133 }
134 /*
135 * Allocate RESYNC_PAGES data pages and attach them
136 * where needed.
137 */
144 /* we can share bv_page's during recovery */
156 }
157 }
161 out_free_pages:
168 out_free_bio:
173 }
176 }
179 {
191 }
193 }
197 }
199 }
202 {
214 }
215 }
218 {
223 }
226 {
232 }
235 {
245 /* wake up frozen array... */
249 }
251 /*
252 * raid_end_bio_io() is called when we have finished servicing a mirrored
253 * operation and are ready to return a success/failure code to the buffer
254 * cache layer.
255 */
257 {
274 /*
275 * Wake up any possible resync thread that waits for the device
276 * to go idle.
277 */
279 }
281 }
283 /*
284 * Update disk head position estimator based on IRQ completion info.
285 */
287 {
292 }
294 /*
295 * Find the disk number which triggered given bio
296 */
299 {
309 }
310 }
320 }
323 {
334 /*
335 * this branch is our 'one mirror IO has finished' event handler:
336 */
340 /*
341 * Set R10BIO_Uptodate in our master bio, so that
342 * we will return a good error code to the higher
343 * levels even if IO on some other mirrored buffer fails.
344 *
345 * The 'master' represents the composite IO operation to
346 * user-side. So if something waits for IO, then it will
347 * wait for the 'master' bio.
348 */
353 /*
354 * oops, read error - keep the refcount on the rdev
355 */
364 }
365 }
368 {
369 /* clear the bitmap if all writes complete successfully */
375 }
378 {
386 else
388 }
389 }
390 }
393 {
410 }
411 /*
412 * this branch is our 'one mirror IO has finished' event handler:
413 */
416 /* Never record new bad blocks to replacement,
417 * just fail it.
418 */
427 }
429 /*
430 * Set R10BIO_Uptodate in our master bio, so that
431 * we will return a good error code for to the higher
432 * levels even if IO on some other mirrored buffer fails.
433 *
434 * The 'master' represents the composite IO operation to
435 * user-side. So if something waits for IO, then it will
436 * wait for the 'master' bio.
437 */
438 sector_t first_bad;
443 /* Maybe we can clear some bad blocks. */
451 else
455 }
456 }
458 /*
459 *
460 * Let's see if all mirrored write operations have finished
461 * already.
462 */
466 }
468 /*
469 * RAID10 layout manager
470 * As well as the chunksize and raid_disks count, there are two
471 * parameters: near_copies and far_copies.
472 * near_copies * far_copies must be <= raid_disks.
473 * Normally one of these will be 1.
474 * If both are 1, we get raid0.
475 * If near_copies == raid_disks, we get raid1.
476 *
477 * Chunks are laid out in raid0 style with near_copies copies of the
478 * first chunk, followed by near_copies copies of the next chunk and
479 * so on.
480 * If far_copies > 1, then after 1/far_copies of the array has been assigned
481 * as described above, we start again with a device offset of near_copies.
482 * So we effectively have another copy of the whole array further down all
483 * the drives, but with blocks on different drives.
484 * With this layout, and block is never stored twice on the one device.
485 *
486 * raid10_find_phys finds the sector offset of a given virtual sector
487 * on each device that it is on.
488 *
489 * raid10_find_virt does the reverse mapping, from a device and a
490 * sector offset to a virtual address
491 */
494 {
496 sector_t sector;
497 sector_t chunk;
498 sector_t stripe;
503 /* now calculate first sector/dev */
515 /* and calculate all the others */
521 slot++;
530 slot++;
531 }
532 dev++;
536 }
537 }
539 }
542 {
558 else
560 }
562 }
566 }
568 /**
569 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
570 * @q: request queue
571 * @bvm: properties of new bio
572 * @biovec: the request that could be merged to it.
573 *
574 * Return amount of bytes we can accept at this offset
575 * If near_copies == raid_disk, there are no striping issues,
576 * but in that case, the function isn't called at all.
577 */
581 {
592 else
594 }
596 /*
597 * This routine returns the disk from which the requested read should
598 * be done. There is a per-array 'next expected sequential IO' sector
599 * number - if this matches on the next IO then we use the last disk.
600 * There is also a per-disk 'last know head position' sector that is
601 * maintained from IRQ contexts, both the normal and the resync IO
602 * completion handlers update this position correctly. If there is no
603 * perfect sequential match then we pick the disk whose head is closest.
604 *
605 * If there are 2 mirrors in the same 2 devices, performance degrades
606 * because position is mirror, not device based.
607 *
608 * The rdev for the device selected will have nr_pending incremented.
609 */
611 /*
612 * FIXME: possibly should rethink readbalancing and do it differently
613 * depending on near_copies / far_copies geometry.
614 */
618 {
630 retry:
637 /*
638 * Check if we can balance. We can balance on the whole
639 * device if no resync is going on (recovery is ok), or below
640 * the resync window. We take the first readable disk when
641 * above the resync window.
642 */
648 sector_t first_bad;
650 sector_t dev_sector;
671 /* Already have a better slot */
674 /* Cannot read here. If this is the
675 * 'primary' device, then we must not read
676 * beyond 'bad_sectors' from another device.
677 */
684 sector_t good_sectors =
690 }
692 /* Must read from here */
694 }
702 /* This optimisation is debatable, and completely destroys
703 * sequential read speed for 'far copies' arrays. So only
704 * keep it for 'near' arrays, and review those later.
705 */
709 /* for far > 1 always use the lowest address */
712 else
719 }
720 }
724 }
729 /* Cannot risk returning a device that failed
730 * before we inc'ed nr_pending
731 */
734 }
742 }
745 {
763 }
764 }
767 }
770 {
771 /* Any writes that have been queued but are awaiting
772 * bitmap updates get flushed here.
773 */
781 /* flush any pending bitmap writes to disk
782 * before proceeding w/ I/O */
791 }
794 }
796 /* Barriers....
797 * Sometimes we need to suspend IO while we do something else,
798 * either some resync/recovery, or reconfigure the array.
799 * To do this we raise a 'barrier'.
800 * The 'barrier' is a counter that can be raised multiple times
801 * to count how many activities are happening which preclude
802 * normal IO.
803 * We can only raise the barrier if there is no pending IO.
804 * i.e. if nr_pending == 0.
805 * We choose only to raise the barrier if no-one is waiting for the
806 * barrier to go down. This means that as soon as an IO request
807 * is ready, no other operations which require a barrier will start
808 * until the IO request has had a chance.
809 *
810 * So: regular IO calls 'wait_barrier'. When that returns there
811 * is no backgroup IO happening, It must arrange to call
812 * allow_barrier when it has finished its IO.
813 * backgroup IO calls must call raise_barrier. Once that returns
814 * there is no normal IO happeing. It must arrange to call
815 * lower_barrier when the particular background IO completes.
816 */
819 {
823 /* Wait until no block IO is waiting (unless 'force') */
827 /* block any new IO from starting */
830 /* Now wait for all pending IO to complete */
836 }
839 {
845 }
848 {
854 );
856 }
859 }
862 {
868 }
871 {
872 /* stop syncio and normal IO and wait for everything to
873 * go quiet.
874 * We increment barrier and nr_waiting, and then
875 * wait until nr_pending match nr_queued+1
876 * This is called in the context of one normal IO request
877 * that has failed. Thus any sync request that might be pending
878 * will be blocked by nr_pending, and we need to wait for
879 * pending IO requests to complete or be queued for re-try.
880 * Thus the number queued (nr_queued) plus this request (1)
881 * must match the number of pending IOs (nr_pending) before
882 * we continue.
883 */
893 }
896 {
897 /* reverse the effect of the freeze */
903 }
906 {
924 }
926 /* If this request crosses a chunk boundary, we need to
927 * split it. This will only happen for 1 PAGE (or less) requests.
928 */
933 /* Sanity check -- queue functions should prevent this happening */
937 /* This is a one page bio that upper layers
938 * refuse to split for us, so we need to split it.
939 */
943 /* Each of these 'make_request' calls will call 'wait_barrier'.
944 * If the first succeeds but the second blocks due to the resync
945 * thread raising the barrier, we will deadlock because the
946 * IO to the underlying device will be queued in generic_make_request
947 * and will never complete, so will never reduce nr_pending.
948 * So increment nr_waiting here so no new raise_barriers will
949 * succeed, and so the second wait_barrier cannot block.
950 */
965 bad_map:
972 }
976 /*
977 * Register the new request and wait if the reconstruction
978 * thread has put up a bar for new requests.
979 * Continue immediately if no resync is active currently.
980 */
992 /* We might need to issue multiple reads to different
993 * devices if there are bad blocks around, so we keep
994 * track of the number of reads in bio->bi_phys_segments.
995 * If this is 0, there is only one r10_bio and no locking
996 * will be needed when the request completes. If it is
997 * non-zero, then it is the number of not-completed requests.
998 */
1003 /*
1004 * read balancing logic:
1005 */
1009 read_again:
1014 }
1019 max_sectors);
1032 /* Could not read all from this device, so we will
1033 * need another r10_bio.
1034 */
1041 else
1044 /* Cannot call generic_make_request directly
1045 * as that will be queued in __generic_make_request
1046 * and subsequent mempool_alloc might block
1047 * waiting for it. so hand bio over to raid10d.
1048 */
1063 }
1065 /*
1066 * WRITE:
1067 */
1072 }
1073 /* first select target devices under rcu_lock and
1074 * inc refcount on their rdev. Record them by setting
1075 * bios[x] to bio
1076 * If there are known/acknowledged bad blocks on any device
1077 * on which we have seen a write error, we want to avoid
1078 * writing to those blocks. This potentially requires several
1079 * writes to write around the bad blocks. Each set of writes
1080 * gets its own r10_bio with a set of bios attached. The number
1081 * of r10_bios is recored in bio->bi_phys_segments just as with
1082 * the read case.
1083 */
1088 retry_write:
1104 }
1109 }
1118 }
1120 sector_t first_bad;
1126 max_sectors,
1129 /* Mustn't write here until the bad block
1130 * is acknowledged
1131 */
1136 }
1138 /* Cannot write here at all */
1141 /* Mustn't write more than bad_sectors
1142 * to other devices yet
1143 */
1145 /* We don't set R10BIO_Degraded as that
1146 * only applies if the disk is missing,
1147 * so it might be re-added, and we want to
1148 * know to recover this chunk.
1149 * In this case the device is here, and the
1150 * fact that this chunk is not in-sync is
1151 * recorded in the bad block log.
1152 */
1154 }
1159 }
1160 }
1166 }
1167 }
1171 /* Have to wait for this device to get unblocked, then retry */
1179 }
1185 /* Race with remove_disk */
1188 }
1190 }
1191 }
1196 }
1199 /* We are splitting this into multiple parts, so
1200 * we need to prepare for allocating another r10_bio.
1201 */
1206 else
1209 }
1223 max_sectors);
1244 max_sectors);
1247 /* We are actively writing to the original device
1248 * so it cannot disappear, so the replacement cannot
1249 * become NULL here
1250 */
1263 }
1265 /* Don't remove the bias on 'remaining' (one_write_done) until
1266 * after checking if we need to go around again.
1267 */
1271 /* We need another r10_bio. It has already been counted
1272 * in bio->bi_phys_segments.
1273 */
1283 }
1286 /* In case raid10d snuck in to freeze_array */
1291 }
1294 {
1305 else
1307 }
1315 }
1317 /* check if there are enough drives for
1318 * every block to appear on atleast one.
1319 * Don't consider the device numbered 'ignore'
1320 * as we might be about to remove it.
1321 */
1323 {
1332 cnt++;
1334 }
1339 }
1342 {
1346 /*
1347 * If it is not operational, then we have already marked it as dead
1348 * else if it is the last working disks, ignore the error, let the
1349 * next level up know.
1350 * else mark the drive as failed
1351 */
1354 /*
1355 * Don't fail the drive, just return an IO error.
1356 */
1363 /*
1364 * if recovery is running, make sure it aborts.
1365 */
1367 }
1376 }
1379 {
1387 }
1399 }
1400 }
1403 {
1409 }
1412 {
1419 /*
1420 * Find all non-in_sync disks within the RAID10 configuration
1421 * and mark them in_sync
1422 */
1429 /* Replacement has just become active */
1432 count++;
1434 /* Replaced device not technically faulty,
1435 * but we need to be sure it gets removed
1436 * and never re-added.
1437 */
1441 }
1446 count++;
1448 }
1449 }
1456 }
1460 {
1468 /* only hot-add to in-sync arrays, as recovery is
1469 * very different from resync
1470 */
1481 else
1501 }
1505 }
1509 /* as we don't honour merge_bvec_fn, we must
1510 * never risk violating it, so limit
1511 * ->max_segments to one lying with a single
1512 * page, as a one page request is never in
1513 * violation.
1514 */
1519 }
1529 }
1534 }
1537 {
1549 else
1556 }
1557 /* Only remove faulty devices if recovery
1558 * is not possible.
1559 */
1566 }
1570 /* lost the race, try later */
1575 /* We must have just cleared 'rdev' */
1579 * but will never see neither -- if they are careful.
1580 */
1584 /* We might have just remove the Replacement as faulty
1585 * Clear the flag just in case
1586 */
1591 abort:
1595 }
1599 {
1608 else
1609 /* The write handler will notice the lack of
1610 * R10BIO_Uptodate and record any errors etc
1611 */
1615 /* for reconstruct, we always reschedule after a read.
1616 * for resync, only after all reads
1617 */
1621 /* we have read all the blocks,
1622 * do the comparison in process context in raid10d
1623 */
1625 }
1626 }
1629 {
1634 /* the primary of several recovery bios */
1639 else
1648 else
1651 }
1652 }
1653 }
1656 {
1662 sector_t first_bad;
1674 }
1685 }
1695 }
1697 /*
1698 * Note: sync and recover and handled very differently for raid10
1699 * This code is for resync.
1700 * For resync, we read through virtual addresses and read all blocks.
1701 * If there is any error, we schedule a write. The lowest numbered
1702 * drive is authoritative.
1703 * However requests come for physical address, so we need to map.
1704 * For every physical address there are raid_disks/copies virtual addresses,
1705 * which is always are least one, but is not necessarly an integer.
1706 * This means that a physical address can span multiple chunks, so we may
1707 * have to submit multiple io requests for a single sync request.
1708 */
1709 /*
1710 * We check if all blocks are in-sync and only write to blocks that
1711 * aren't in sync
1712 */
1714 {
1721 /* find the first device with a block */
1732 /* now find blocks with errors */
1744 /* We know that the bi_io_vec layout is the same for
1745 * both 'first' and 'i', so we just compare them.
1746 * All vec entries are PAGE_SIZE;
1747 */
1751 PAGE_SIZE))
1757 /* Don't fix anything. */
1759 }
1760 /* Ok, we need to write this bio, either to correct an
1761 * inconsistency or to correct an unreadable block.
1762 * First we need to fixup bv_offset, bv_len and
1763 * bi_vecs, as the read request might have corrupted these
1764 */
1782 PAGE_SIZE);
1783 }
1794 }
1796 /* Now write out to any replacement devices
1797 * that are active
1798 */
1811 PAGE_SIZE);
1817 }
1819 done:
1823 }
1824 }
1826 /*
1827 * Now for the recovery code.
1828 * Recovery happens across physical sectors.
1829 * We recover all non-is_sync drives by finding the virtual address of
1830 * each, and then choose a working drive that also has that virt address.
1831 * There is a separate r10_bio for each non-in_sync drive.
1832 * Only the first two slots are in use. The first for reading,
1833 * The second for writing.
1834 *
1835 */
1837 {
1838 /* We got a read error during recovery.
1839 * We repeat the read in smaller page-sized sections.
1840 * If a read succeeds, write it to the new device or record
1841 * a bad block if we cannot.
1842 * If a read fails, record a bad block on both old and
1843 * new devices.
1844 */
1857 sector_t addr;
1866 addr,
1874 addr,
1884 }
1885 }
1887 /* We don't worry if we cannot set a bad block -
1888 * it really is bad so there is no loss in not
1889 * recording it yet
1890 */
1894 /* need bad block on destination too */
1899 /* just abort the recovery */
1901 "md/raid10:%s: recovery aborted"
1910 }
1911 }
1912 }
1916 idx++;
1917 }
1918 }
1921 {
1930 }
1932 /*
1933 * share the pages with the first bio
1934 * and submit the write request
1935 */
1943 }
1949 }
1950 }
1953 /*
1954 * Used by fix_read_error() to decay the per rdev read_errors.
1955 * We halve the read error count for every hour that has elapsed
1956 * since the last recorded read error.
1957 *
1958 */
1960 {
1969 /* first time we've seen a read error */
1972 }
1979 /*
1980 * if hours_since_last is > the number of bits in read_errors
1981 * just set read errors to 0. We do this to avoid
1982 * overflowing the shift of read_errors by hours_since_last.
1983 */
1986 else
1988 }
1992 {
1993 sector_t first_bad;
2000 /* success */
2007 }
2008 /* need to record an error - either for the block or the device */
2012 }
2014 /*
2015 * This is a kernel thread which:
2016 *
2017 * 1. Retries failed read operations on working mirrors.
2018 * 2. Updates the raid superblock when problems encounter.
2019 * 3. Performs writes following reads for array synchronising.
2020 */
2023 {
2030 /* still own a reference to this rdev, so it cannot
2031 * have been cleared recently.
2032 */
2036 /* drive has already been failed, just ignore any
2037 more fix_read_error() attempts */
2047 "md/raid10:%s: %s: Raid device exceeded "
2056 }
2069 sector_t first_bad;
2082 sect,
2089 }
2090 sl++;
2097 /* Cannot read from anywhere, just mark the block
2098 * as bad on the first device to discourage future
2099 * reads.
2100 */
2105 rdev,
2111 }
2114 /* write it back and re-read */
2121 sl--;
2132 sect,
2135 /* Well, this device is dead */
2137 "md/raid10:%s: read correction "
2138 "write failed"
2148 }
2151 }
2158 sl--;
2169 sect,
2171 READ)) {
2173 /* Well, this device is dead */
2175 "md/raid10:%s: unable to read back "
2176 "corrected sectors"
2189 "md/raid10:%s: read error corrected"
2196 }
2200 }
2205 }
2206 }
2209 {
2211 }
2214 {
2225 }
2228 {
2233 /* bio has the data to be written to slot 'i' where
2234 * we just recently had a write error.
2235 * We repeatedly clone the bio and trim down to one block,
2236 * then try the write. Where the write fails we record
2237 * a bad block.
2238 * It is conceivable that the bio doesn't exactly align with
2239 * blocks. We must handle this.
2240 *
2241 * We currently own a reference to the rdev.
2242 */
2245 sector_t sector;
2263 /* Write at 'sector' for 'sectors' */
2271 /* Failure! */
2280 }
2282 }
2285 {
2294 /* we got a read error. Maybe the drive is bad. Maybe just
2295 * the block and we can fix it.
2296 * We freeze all other IO, and try reading the block from
2297 * other devices. When we find one, we re-write
2298 * and check it that fixes the read error.
2299 * This is all done synchronously while the array is
2300 * frozen.
2301 */
2306 }
2313 read_more:
2323 }
2330 KERN_ERR
2331 "md/raid10:%s: %s: redirecting"
2340 max_sectors);
2350 /* Drat - have to split this up more */
2359 else
2366 GFP_NOIO);
2380 }
2383 {
2384 /* Some sort of write request has finished and it
2385 * succeeded in writing where we thought there was a
2386 * bad block. So forget the bad block.
2387 * Or possibly if failed and we need to record
2388 * a bad block.
2389 */
2403 rdev,
2408 rdev,
2412 }
2419 rdev,
2424 rdev,
2428 }
2429 }
2438 rdev,
2448 }
2450 }
2455 rdev,
2459 }
2460 }
2465 }
2466 }
2469 {
2487 }
2505 /* just a partial read to be scheduled from a
2506 * separate context
2507 */
2510 }
2515 }
2517 }
2521 {
2536 }
2538 /*
2539 * perform a "sync" on one "block"
2540 *
2541 * We need to make sure that no normal I/O request - particularly write
2542 * requests - conflict with active sync requests.
2543 *
2544 * This is achieved by tracking pending requests and a 'barrier' concept
2545 * that can be installed to exclude normal IO requests.
2546 *
2547 * Resync and recovery are handled very differently.
2548 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2549 *
2550 * For resync, we iterate over virtual addresses, read all copies,
2551 * and update if there are differences. If only one copy is live,
2552 * skip it.
2553 * For recovery, we iterate over physical addresses, read a good
2554 * value for each non-in_sync drive, and over-write.
2555 *
2556 * So, for recovery we may have several outstanding complex requests for a
2557 * given address, one for each out-of-sync device. We model this by allocating
2558 * a number of r10_bio structures, one for each out-of-sync device.
2559 * As we setup these structures, we collect all bio's together into a list
2560 * which we then process collectively to add pages, and then process again
2561 * to pass to generic_make_request.
2562 *
2563 * The r10_bio structures are linked using a borrowed master_bio pointer.
2564 * This link is counted in ->remaining. When the r10_bio that points to NULL
2565 * has its remaining count decremented to 0, the whole complex operation
2566 * is complete.
2567 *
2568 */
2572 {
2579 sector_t sync_blocks;
2587 skipped:
2592 /* If we aborted, we need to abort the
2593 * sync on the 'current' bitmap chucks (there can
2594 * be several when recovering multiple devices).
2595 * as we may have started syncing it but not finished.
2596 * We can find the current address in
2597 * mddev->curr_resync, but for recovery,
2598 * we need to convert that to several
2599 * virtual addresses.
2600 */
2606 sector_t sect =
2610 }
2612 /* completed sync */
2616 /* Completed a full sync so the replacements
2617 * are now fully recovered.
2618 */
2622 ->recovery_offset
2624 }
2626 }
2631 }
2633 /* if there has been nothing to do on any drive,
2634 * then there is nothing to do at all..
2635 */
2638 }
2643 /* make sure whole request will fit in a chunk - if chunks
2644 * are meaningful
2645 */
2649 /*
2650 * If there is non-resync activity waiting for us then
2651 * put in a delay to throttle resync.
2652 */
2656 /* Again, very different code for resync and recovery.
2657 * Both must result in an r10bio with a list of bios that
2658 * have bi_end_io, bi_sector, bi_bdev set,
2659 * and bi_private set to the r10bio.
2660 * For recovery, we may actually create several r10bios
2661 * with 2 bios in each, that correspond to the bios in the main one.
2662 * In this case, the subordinate r10bios link back through a
2663 * borrowed master_bio pointer, and the counter in the master
2664 * includes a ref from each subordinate.
2665 */
2666 /* First, we decide what to do and set ->bi_end_io
2667 * To end_sync_read if we want to read, and
2668 * end_sync_write if we will want to write.
2669 */
2673 /* recovery... the complicated one */
2680 sector_t sect;
2687 &&
2694 /* want to reconstruct this device */
2697 /* Unless we are doing a full sync, or a replacement
2698 * we only need to recover the block if it is set in
2699 * the bitmap
2700 */
2708 /* yep, skip the sync_blocks here, but don't assume
2709 * that there will never be anything to do here
2710 */
2713 }
2728 /* Need to check if the array will still be
2729 * degraded
2730 */
2736 }
2752 /* This is where we read from */
2762 bad_sectors -= (sector
2767 }
2768 }
2779 /* and we write to 'i' (if not in_sync) */
2806 /* and maybe write to replacement */
2811 /* Note: if rdev != NULL, then bio
2812 * cannot be NULL as r10buf_pool_alloc will
2813 * have allocated it.
2814 * So the second test here is pointless.
2815 * But it keeps semantic-checkers happy, and
2816 * this comment keeps human reviewers
2817 * happy.
2818 */
2831 }
2833 /* Cannot recover, so abort the recovery or
2834 * record a bad block */
2840 /* problem is that there are bad blocks
2841 * on other device(s)
2842 */
2860 }
2867 mirror->recovery_disabled
2869 }
2871 }
2872 }
2879 }
2881 }
2883 /* resync. Schedule a read for every block at this virt offset */
2892 /* We can skip this block */
2895 }
2936 }
2937 }
2948 count++;
2955 /* Need to set up for writing to the replacement */
2969 count++;
2970 }
2977 mddev);
2982 mddev);
2983 }
2987 }
2988 }
2999 }
3017 /* stop here */
3022 /* remove last page from this bio */
3026 }
3028 }
3032 bio_full:
3046 }
3047 }
3050 /* pretend they weren't skipped, it makes
3051 * no important difference in this case
3052 */
3056 giveup:
3057 /* There is nowhere to write, so all non-sync
3058 * drives must be failed or in resync, all drives
3059 * have a bad block, so try the next chunk...
3060 */
3065 chunks_skipped ++;
3068 }
3070 static sector_t
3072 {
3073 sector_t size;
3087 }
3091 {
3103 }
3114 }
3122 GFP_KERNEL);
3148 /* 'size' is now the number of chunks in the array */
3149 /* calculate "used chunks per device" in 'stride' */
3152 /* We need to round up when dividing by raid_disks to
3153 * get the stride size.
3154 */
3162 else
3180 out:
3189 }
3191 }
3194 {
3199 sector_t size;
3201 /*
3202 * copy the already verified devices into our private RAID10
3203 * bookkeeping area. [whatever we allocate in run(),
3204 * should be freed in stop()]
3205 */
3212 }
3224 else
3244 }
3249 /* as we don't honour merge_bvec_fn, we must never risk
3250 * violating it, so limit max_segments to 1 lying
3251 * within a single page.
3252 */
3257 }
3260 }
3261 /* need to check that every block has at least one working mirror */
3266 }
3274 /* The replacement is all we have - use it */
3278 }
3286 }
3288 }
3298 /*
3299 * Ok, everything is just fine now
3300 */
3309 /* Calculate max read-ahead size.
3310 * We need to readahead at least twice a whole stripe....
3311 * maybe...
3312 */
3313 {
3319 }
3329 out_free_conf:
3337 out:
3339 }
3342 {
3356 }
3359 {
3369 }
3370 }
3373 {
3381 }
3383 /* Set new parameters */
3385 /* new layout: far_copies = 1, near_copies = 2 */
3390 /* make sure it will be not marked as dirty */
3399 }
3402 }
3405 {
3408 /* raid10 can take over:
3409 * raid0 - providing it has only two drives
3410 */
3412 /* for raid0 takeover only one zone is supported */
3419 }
3421 }
3423 }
3426 {
3442 };
3445 {
3447 }
3450 {
3452 }