| blob | e2766d8251a184a6c28a67430ca34c7e68d0df67 |
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 futher 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/seq_file.h>
29 /*
30 * RAID10 provides a combination of RAID0 and RAID1 functionality.
31 * The layout of data is defined by
32 * chunk_size
33 * raid_disks
34 * near_copies (stored in low byte of layout)
35 * far_copies (stored in second byte of layout)
36 * far_offset (stored in bit 16 of layout )
37 *
38 * The data to be stored is divided into chunks using chunksize.
39 * Each device is divided into far_copies sections.
40 * In each section, chunks are laid out in a style similar to raid0, but
41 * near_copies copies of each chunk is stored (each on a different drive).
42 * The starting device for each section is offset near_copies from the starting
43 * device of the previous section.
44 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
45 * drive.
46 * near_copies and far_copies must be at least one, and their product is at most
47 * raid_disks.
48 *
49 * If far_offset is true, then the far_copies are handled a bit differently.
50 * The copies are still in different stripes, but instead of be very far apart
51 * on disk, there are adjacent stripes.
52 */
54 /*
55 * Number of guaranteed r10bios in case of extreme VM load:
56 */
57 #define NR_RAID10_BIOS 256
65 {
70 /* allocate a r10bio with room for raid_disks entries in the bios array */
76 }
79 {
81 }
83 /* Maximum size of each resync request */
84 #define RESYNC_BLOCK_SIZE (64*1024)
85 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
86 /* amount of memory to reserve for resync requests */
87 #define RESYNC_WINDOW (1024*1024)
88 /* maximum number of concurrent requests, memory permitting */
89 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
91 /*
92 * When performing a resync, we need to read and compare, so
93 * we need as many pages are there are copies.
94 * When performing a recovery, we need 2 bios, one for read,
95 * one for write (we recover only one drive per r10buf)
96 *
97 */
99 {
111 }
115 else
118 /*
119 * Allocate bios.
120 */
126 }
127 /*
128 * Allocate RESYNC_PAGES data pages and attach them
129 * where needed.
130 */
139 }
140 }
144 out_free_pages:
151 out_free_bio:
156 }
159 {
171 }
173 }
174 }
176 }
179 {
187 }
188 }
191 {
194 /*
195 * Wake up any possible resync thread that waits for the device
196 * to go idle.
197 */
202 }
205 {
211 }
214 {
224 /* wake up frozen array... */
228 }
230 /*
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
233 * cache layer.
234 */
236 {
242 }
244 /*
245 * Update disk head position estimator based on IRQ completion info.
246 */
248 {
253 }
256 {
265 /*
266 * this branch is our 'one mirror IO has finished' event handler:
267 */
271 /*
272 * Set R10BIO_Uptodate in our master bio, so that
273 * we will return a good error code to the higher
274 * levels even if IO on some other mirrored buffer fails.
275 *
276 * The 'master' represents the composite IO operation to
277 * user-side. So if something waits for IO, then it will
278 * wait for the 'master' bio.
279 */
283 /*
284 * oops, read error:
285 */
291 }
294 }
297 {
308 /*
309 * this branch is our 'one mirror IO has finished' event handler:
310 */
313 /* an I/O failed, we can't clear the bitmap */
316 /*
317 * Set R10BIO_Uptodate in our master bio, so that
318 * we will return a good error code for to the higher
319 * levels even if IO on some other mirrored buffer fails.
320 *
321 * The 'master' represents the composite IO operation to
322 * user-side. So if something waits for IO, then it will
323 * wait for the 'master' bio.
324 */
329 /*
330 *
331 * Let's see if all mirrored write operations have finished
332 * already.
333 */
335 /* clear the bitmap if all writes complete successfully */
342 }
345 }
348 /*
349 * RAID10 layout manager
350 * Aswell as the chunksize and raid_disks count, there are two
351 * parameters: near_copies and far_copies.
352 * near_copies * far_copies must be <= raid_disks.
353 * Normally one of these will be 1.
354 * If both are 1, we get raid0.
355 * If near_copies == raid_disks, we get raid1.
356 *
357 * Chunks are layed out in raid0 style with near_copies copies of the
358 * first chunk, followed by near_copies copies of the next chunk and
359 * so on.
360 * If far_copies > 1, then after 1/far_copies of the array has been assigned
361 * as described above, we start again with a device offset of near_copies.
362 * So we effectively have another copy of the whole array further down all
363 * the drives, but with blocks on different drives.
364 * With this layout, and block is never stored twice on the one device.
365 *
366 * raid10_find_phys finds the sector offset of a given virtual sector
367 * on each device that it is on.
368 *
369 * raid10_find_virt does the reverse mapping, from a device and a
370 * sector offset to a virtual address
371 */
374 {
376 sector_t sector;
377 sector_t chunk;
378 sector_t stripe;
383 /* now calculate first sector/dev */
395 /* and calculate all the others */
401 slot++;
410 slot++;
411 }
412 dev++;
416 }
417 }
419 }
422 {
438 else
440 }
442 }
446 }
448 /**
449 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
450 * @q: request queue
451 * @bvm: properties of new bio
452 * @biovec: the request that could be merged to it.
453 *
454 * Return amount of bytes we can accept at this offset
455 * If near_copies == raid_disk, there are no striping issues,
456 * but in that case, the function isn't called at all.
457 */
461 {
472 else
474 }
476 /*
477 * This routine returns the disk from which the requested read should
478 * be done. There is a per-array 'next expected sequential IO' sector
479 * number - if this matches on the next IO then we use the last disk.
480 * There is also a per-disk 'last know head position' sector that is
481 * maintained from IRQ contexts, both the normal and the resync IO
482 * completion handlers update this position correctly. If there is no
483 * perfect sequential match then we pick the disk whose head is closest.
484 *
485 * If there are 2 mirrors in the same 2 devices, performance degrades
486 * because position is mirror, not device based.
487 *
488 * The rdev for the device selected will have nr_pending incremented.
489 */
491 /*
492 * FIXME: possibly should rethink readbalancing and do it differently
493 * depending on near_copies / far_copies geometry.
494 */
496 {
505 /*
506 * Check if we can balance. We can balance on the whole
507 * device if no resync is going on (recovery is ok), or below
508 * the resync window. We take the first readable disk when
509 * above the resync window.
510 */
513 /* make sure that disk is operational */
520 slot++;
525 }
527 }
529 }
532 /* make sure the disk is operational */
538 slot ++;
542 }
544 }
550 /* Find the disk whose head is closest,
551 * or - for far > 1 - find the closest to partition beginning */
562 /* This optimisation is debatable, and completely destroys
563 * sequential read speed for 'far copies' arrays. So only
564 * keep it for 'near' arrays, and review those later.
565 */
570 }
572 /* for far > 1 always use the lowest address */
575 else
582 }
583 }
585 rb_out:
587 /* conf->next_seq_sect = this_sector + sectors;*/
591 else
596 }
599 {
616 }
617 }
619 }
622 {
627 }
630 {
644 }
645 }
648 }
651 {
652 /* Any writes that have been queued but are awaiting
653 * bitmap updates get flushed here.
654 * We return 1 if any requests were actually submitted.
655 */
665 /* flush any pending bitmap writes to disk
666 * before proceeding w/ I/O */
674 }
679 }
680 /* Barriers....
681 * Sometimes we need to suspend IO while we do something else,
682 * either some resync/recovery, or reconfigure the array.
683 * To do this we raise a 'barrier'.
684 * The 'barrier' is a counter that can be raised multiple times
685 * to count how many activities are happening which preclude
686 * normal IO.
687 * We can only raise the barrier if there is no pending IO.
688 * i.e. if nr_pending == 0.
689 * We choose only to raise the barrier if no-one is waiting for the
690 * barrier to go down. This means that as soon as an IO request
691 * is ready, no other operations which require a barrier will start
692 * until the IO request has had a chance.
693 *
694 * So: regular IO calls 'wait_barrier'. When that returns there
695 * is no backgroup IO happening, It must arrange to call
696 * allow_barrier when it has finished its IO.
697 * backgroup IO calls must call raise_barrier. Once that returns
698 * there is no normal IO happeing. It must arrange to call
699 * lower_barrier when the particular background IO completes.
700 */
703 {
707 /* Wait until no block IO is waiting (unless 'force') */
712 /* block any new IO from starting */
715 /* No wait for all pending IO to complete */
722 }
725 {
731 }
734 {
742 }
745 }
748 {
754 }
757 {
758 /* stop syncio and normal IO and wait for everything to
759 * go quiet.
760 * We increment barrier and nr_waiting, and then
761 * wait until nr_pending match nr_queued+1
762 * This is called in the context of one normal IO request
763 * that has failed. Thus any sync request that might be pending
764 * will be blocked by nr_pending, and we need to wait for
765 * pending IO requests to complete or be queued for re-try.
766 * Thus the number queued (nr_queued) plus this request (1)
767 * must match the number of pending IOs (nr_pending) before
768 * we continue.
769 */
779 }
782 {
783 /* reverse the effect of the freeze */
789 }
792 {
810 }
812 /* If this request crosses a chunk boundary, we need to
813 * split it. This will only happen for 1 PAGE (or less) requests.
814 */
819 /* Sanity check -- queue functions should prevent this happening */
823 /* This is a one page bio that upper layers
824 * refuse to split for us, so we need to split it.
825 */
835 bad_map:
842 }
846 /*
847 * Register the new request and wait if the reconstruction
848 * thread has put up a bar for new requests.
849 * Continue immediately if no resync is active currently.
850 */
869 /*
870 * read balancing logic:
871 */
877 }
893 }
895 /*
896 * WRITE:
897 */
898 /* first select target devices under rcu_lock and
899 * inc refcount on their rdev. Record them by setting
900 * bios[x] to bio
901 */
903 retry_write:
913 }
920 }
921 }
925 /* Have to wait for this device to get unblocked, then retry */
933 }
938 }
961 }
964 /* the array is dead */
968 }
976 /* In case raid10d snuck in to freeze_array */
983 }
986 {
997 else
999 }
1007 }
1010 {
1014 /*
1015 * If it is not operational, then we have already marked it as dead
1016 * else if it is the last working disks, ignore the error, let the
1017 * next level up know.
1018 * else mark the drive as failed
1019 */
1022 /*
1023 * Don't fail the drive, just return an IO error.
1024 * The test should really be more sophisticated than
1025 * "working_disks == 1", but it isn't critical, and
1026 * can wait until we do more sophisticated "is the drive
1027 * really dead" tests...
1028 */
1035 /*
1036 * if recovery is running, make sure it aborts.
1037 */
1039 }
1045 }
1048 {
1056 }
1068 }
1069 }
1072 {
1078 }
1080 /* check if there are enough drives for
1081 * every block to appear on atleast one
1082 */
1084 {
1092 cnt++;
1094 }
1099 }
1102 {
1107 /*
1108 * Find all non-in_sync disks within the RAID10 configuration
1109 * and mark them in_sync
1110 */
1120 }
1121 }
1125 }
1129 {
1138 /* only hot-add to in-sync arrays, as recovery is
1139 * very different from resync
1140 */
1152 else
1159 /* as we don't honour merge_bvec_fn, we must
1160 * never risk violating it, so limit
1161 * ->max_segments to one lying with a single
1162 * page, as a one page request is never in
1163 * violation.
1164 */
1169 }
1178 }
1183 }
1186 {
1199 }
1200 /* Only remove faulty devices in recovery
1201 * is not possible.
1202 */
1207 }
1211 /* lost the race, try later */
1215 }
1217 }
1218 abort:
1222 }
1226 {
1246 }
1248 /* for reconstruct, we always reschedule after a read.
1249 * for resync, only after all reads
1250 */
1254 /* we have read all the blocks,
1255 * do the comparison in process context in raid10d
1256 */
1258 }
1259 }
1262 {
1282 /* the primary of several recovery bios */
1291 }
1292 }
1293 }
1295 /*
1296 * Note: sync and recover and handled very differently for raid10
1297 * This code is for resync.
1298 * For resync, we read through virtual addresses and read all blocks.
1299 * If there is any error, we schedule a write. The lowest numbered
1300 * drive is authoritative.
1301 * However requests come for physical address, so we need to map.
1302 * For every physical address there are raid_disks/copies virtual addresses,
1303 * which is always are least one, but is not necessarly an integer.
1304 * This means that a physical address can span multiple chunks, so we may
1305 * have to submit multiple io requests for a single sync request.
1306 */
1307 /*
1308 * We check if all blocks are in-sync and only write to blocks that
1309 * aren't in sync
1310 */
1312 {
1319 /* find the first device with a block */
1330 /* now find blocks with errors */
1342 /* We know that the bi_io_vec layout is the same for
1343 * both 'first' and 'i', so we just compare them.
1344 * All vec entries are PAGE_SIZE;
1345 */
1349 PAGE_SIZE))
1354 }
1356 /* Don't fix anything. */
1358 /* Ok, we need to write this bio
1359 * First we need to fixup bv_offset, bv_len and
1360 * bi_vecs, as the read request might have corrupted these
1361 */
1379 PAGE_SIZE);
1380 }
1391 }
1393 done:
1397 }
1398 }
1400 /*
1401 * Now for the recovery code.
1402 * Recovery happens across physical sectors.
1403 * We recover all non-is_sync drives by finding the virtual address of
1404 * each, and then choose a working drive that also has that virt address.
1405 * There is a separate r10_bio for each non-in_sync drive.
1406 * Only the first two slots are in use. The first for reading,
1407 * The second for writing.
1408 *
1409 */
1412 {
1418 /* move the pages across to the second bio
1419 * and submit the write request
1420 */
1427 }
1434 else
1436 }
1439 /*
1440 * Used by fix_read_error() to decay the per rdev read_errors.
1441 * We halve the read error count for every hour that has elapsed
1442 * since the last recorded read error.
1443 *
1444 */
1446 {
1455 /* first time we've seen a read error */
1458 }
1465 /*
1466 * if hours_since_last is > the number of bits in read_errors
1467 * just set read errors to 0. We do this to avoid
1468 * overflowing the shift of read_errors by hours_since_last.
1469 */
1472 else
1474 }
1476 /*
1477 * This is a kernel thread which:
1478 *
1479 * 1. Retries failed read operations on working mirrors.
1480 * 2. Updates the raid superblock when problems encounter.
1481 * 3. Performs writes following reads for array synchronising.
1482 */
1485 {
1492 {
1502 /* drive has already been failed, just ignore any
1503 more fix_read_error() attempts */
1505 }
1513 "raid10: %s: Raid device exceeded "
1514 "read_error threshold "
1518 "raid10: %s: Failing raid "
1522 }
1523 }
1552 }
1553 sl++;
1560 /* Cannot read from anywhere -- bye bye array */
1564 }
1567 /* write it back and re-read */
1574 sl--;
1587 /* Well, this device is dead */
1589 "raid10:%s: read correction "
1590 "write failed"
1600 }
1603 }
1604 }
1610 sl--;
1623 /* Well, this device is dead */
1625 "raid10:%s: unable to read back "
1626 "corrected sectors"
1638 "raid10:%s: read error corrected"
1644 }
1648 }
1649 }
1654 }
1655 }
1658 {
1678 }
1694 /* we got a read error. Maybe the drive is bad. Maybe just
1695 * the block and we can fix it.
1696 * We freeze all other IO, and try reading the block from
1697 * other devices. When we find one, we re-write
1698 * and check it that fixes the read error.
1699 * This is all done synchronously while the array is
1700 * frozen.
1701 */
1706 }
1738 }
1739 }
1741 }
1744 }
1748 {
1758 }
1760 /*
1761 * perform a "sync" on one "block"
1762 *
1763 * We need to make sure that no normal I/O request - particularly write
1764 * requests - conflict with active sync requests.
1765 *
1766 * This is achieved by tracking pending requests and a 'barrier' concept
1767 * that can be installed to exclude normal IO requests.
1768 *
1769 * Resync and recovery are handled very differently.
1770 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1771 *
1772 * For resync, we iterate over virtual addresses, read all copies,
1773 * and update if there are differences. If only one copy is live,
1774 * skip it.
1775 * For recovery, we iterate over physical addresses, read a good
1776 * value for each non-in_sync drive, and over-write.
1777 *
1778 * So, for recovery we may have several outstanding complex requests for a
1779 * given address, one for each out-of-sync device. We model this by allocating
1780 * a number of r10_bio structures, one for each out-of-sync device.
1781 * As we setup these structures, we collect all bio's together into a list
1782 * which we then process collectively to add pages, and then process again
1783 * to pass to generic_make_request.
1784 *
1785 * The r10_bio structures are linked using a borrowed master_bio pointer.
1786 * This link is counted in ->remaining. When the r10_bio that points to NULL
1787 * has its remaining count decremented to 0, the whole complex operation
1788 * is complete.
1789 *
1790 */
1793 {
1810 skipped:
1815 /* If we aborted, we need to abort the
1816 * sync on the 'current' bitmap chucks (there can
1817 * be several when recovering multiple devices).
1818 * as we may have started syncing it but not finished.
1819 * We can find the current address in
1820 * mddev->curr_resync, but for recovery,
1821 * we need to convert that to several
1822 * virtual addresses.
1823 */
1829 sector_t sect =
1833 }
1841 }
1843 /* if there has been nothing to do on any drive,
1844 * then there is nothing to do at all..
1845 */
1848 }
1853 /* make sure whole request will fit in a chunk - if chunks
1854 * are meaningful
1855 */
1859 /*
1860 * If there is non-resync activity waiting for us then
1861 * put in a delay to throttle resync.
1862 */
1866 /* Again, very different code for resync and recovery.
1867 * Both must result in an r10bio with a list of bios that
1868 * have bi_end_io, bi_sector, bi_bdev set,
1869 * and bi_private set to the r10bio.
1870 * For recovery, we may actually create several r10bios
1871 * with 2 bios in each, that correspond to the bios in the main one.
1872 * In this case, the subordinate r10bios link back through a
1873 * borrowed master_bio pointer, and the counter in the master
1874 * includes a ref from each subordinate.
1875 */
1876 /* First, we decide what to do and set ->bi_end_io
1877 * To end_sync_read if we want to read, and
1878 * end_sync_write if we will want to write.
1879 */
1883 /* recovery... the complicated one */
1891 /* want to reconstruct this device */
1895 /* Unless we are doing a full sync, we only need
1896 * to recover the block if it is set in the bitmap
1897 */
1904 /* yep, skip the sync_blocks here, but don't assume
1905 * that there will never be anything to do here
1906 */
1909 }
1924 /* Need to check if the array will still be
1925 * degraded
1926 */
1932 }
1941 /* This is where we read from */
1953 /* and we write to 'i' */
1973 }
1974 }
1976 /* Cannot recover, so abort the recovery */
1986 }
1987 }
1994 }
1996 }
1998 /* resync. Schedule a read for every block at this virt offset */
2006 /* We can skip this block */
2009 }
2043 count++;
2044 }
2051 }
2055 }
2056 }
2067 }
2083 /* stop here */
2087 /* remove last page from this bio */
2091 }
2093 }
2095 }
2099 bio_full:
2113 }
2114 }
2117 /* pretend they weren't skipped, it makes
2118 * no important difference in this case
2119 */
2123 giveup:
2124 /* There is nowhere to write, so all non-sync
2125 * drives must be failed, so try the next chunk...
2126 */
2131 chunks_skipped ++;
2134 }
2136 static sector_t
2138 {
2139 sector_t size;
2153 }
2156 {
2169 }
2179 }
2180 /*
2181 * copy the already verified devices into our private RAID10
2182 * bookkeeping area. [whatever we allocate in run(),
2183 * should be freed in stop()]
2184 */
2191 }
2193 GFP_KERNEL);
2198 }
2215 /* 'size' is now the number of chunks in the array */
2216 /* calculate "used chunks per device" in 'stride' */
2219 /* We need to round up when dividing by raid_disks to
2220 * get the stride size.
2221 */
2228 else
2238 }
2248 else
2262 /* as we don't honour merge_bvec_fn, we must never risk
2263 * violating it, so limit max_segments to 1 lying
2264 * within a single page.
2265 */
2270 }
2273 }
2279 /* need to check that every block has at least one working mirror */
2284 }
2297 }
2298 }
2307 }
2317 /*
2318 * Ok, everything is just fine now
2319 */
2327 /* Calculate max read-ahead size.
2328 * We need to readahead at least twice a whole stripe....
2329 * maybe...
2330 */
2331 {
2337 }
2344 out_free_conf:
2351 out:
2353 }
2356 {
2371 }
2374 {
2384 }
2385 }
2388 {
2403 };
2406 {
2408 }
2411 {
2413 }