| blob | 4e53792aa520b8c5fc0aa18df1266dae15d998bf |
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 */
22 #include <linux/raid/raid10.h>
23 #include <linux/raid/bitmap.h>
25 /*
26 * RAID10 provides a combination of RAID0 and RAID1 functionality.
27 * The layout of data is defined by
28 * chunk_size
29 * raid_disks
30 * near_copies (stored in low byte of layout)
31 * far_copies (stored in second byte of layout)
32 * far_offset (stored in bit 16 of layout )
33 *
34 * The data to be stored is divided into chunks using chunksize.
35 * Each device is divided into far_copies sections.
36 * In each section, chunks are laid out in a style similar to raid0, but
37 * near_copies copies of each chunk is stored (each on a different drive).
38 * The starting device for each section is offset near_copies from the starting
39 * device of the previous section.
40 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
41 * drive.
42 * near_copies and far_copies must be at least one, and their product is at most
43 * raid_disks.
44 *
45 * If far_offset is true, then the far_copies are handled a bit differently.
46 * The copies are still in different stripes, but instead of be very far apart
47 * on disk, there are adjacent stripes.
48 */
50 /*
51 * Number of guaranteed r10bios in case of extreme VM load:
52 */
53 #define NR_RAID10_BIOS 256
61 {
66 /* allocate a r10bio with room for raid_disks entries in the bios array */
72 }
75 {
77 }
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
85 /*
86 * When performing a resync, we need to read and compare, so
87 * we need as many pages are there are copies.
88 * When performing a recovery, we need 2 bios, one for read,
89 * one for write (we recover only one drive per r10buf)
90 *
91 */
93 {
105 }
109 else
112 /*
113 * Allocate bios.
114 */
120 }
121 /*
122 * Allocate RESYNC_PAGES data pages and attach them
123 * where needed.
124 */
133 }
134 }
138 out_free_pages:
145 out_free_bio:
150 }
153 {
165 }
167 }
168 }
170 }
173 {
181 }
182 }
185 {
188 /*
189 * Wake up any possible resync thread that waits for the device
190 * to go idle.
191 */
196 }
199 {
205 }
208 {
219 }
221 /*
222 * raid_end_bio_io() is called when we have finished servicing a mirrored
223 * operation and are ready to return a success/failure code to the buffer
224 * cache layer.
225 */
227 {
233 }
235 /*
236 * Update disk head position estimator based on IRQ completion info.
237 */
239 {
244 }
247 {
258 /*
259 * this branch is our 'one mirror IO has finished' event handler:
260 */
264 /*
265 * Set R10BIO_Uptodate in our master bio, so that
266 * we will return a good error code to the higher
267 * levels even if IO on some other mirrored buffer fails.
268 *
269 * The 'master' represents the composite IO operation to
270 * user-side. So if something waits for IO, then it will
271 * wait for the 'master' bio.
272 */
276 /*
277 * oops, read error:
278 */
284 }
288 }
291 {
305 /*
306 * this branch is our 'one mirror IO has finished' event handler:
307 */
310 /* an I/O failed, we can't clear the bitmap */
313 /*
314 * Set R10BIO_Uptodate in our master bio, so that
315 * we will return a good error code for to the higher
316 * levels even if IO on some other mirrored buffer fails.
317 *
318 * The 'master' represents the composite IO operation to
319 * user-side. So if something waits for IO, then it will
320 * wait for the 'master' bio.
321 */
326 /*
327 *
328 * Let's see if all mirrored write operations have finished
329 * already.
330 */
332 /* clear the bitmap if all writes complete successfully */
339 }
343 }
346 /*
347 * RAID10 layout manager
348 * Aswell as the chunksize and raid_disks count, there are two
349 * parameters: near_copies and far_copies.
350 * near_copies * far_copies must be <= raid_disks.
351 * Normally one of these will be 1.
352 * If both are 1, we get raid0.
353 * If near_copies == raid_disks, we get raid1.
354 *
355 * Chunks are layed out in raid0 style with near_copies copies of the
356 * first chunk, followed by near_copies copies of the next chunk and
357 * so on.
358 * If far_copies > 1, then after 1/far_copies of the array has been assigned
359 * as described above, we start again with a device offset of near_copies.
360 * So we effectively have another copy of the whole array further down all
361 * the drives, but with blocks on different drives.
362 * With this layout, and block is never stored twice on the one device.
363 *
364 * raid10_find_phys finds the sector offset of a given virtual sector
365 * on each device that it is on.
366 *
367 * raid10_find_virt does the reverse mapping, from a device and a
368 * sector offset to a virtual address
369 */
372 {
374 sector_t sector;
375 sector_t chunk;
376 sector_t stripe;
381 /* now calculate first sector/dev */
393 /* and calculate all the others */
399 slot++;
408 slot++;
409 }
410 dev++;
414 }
415 }
417 }
420 {
436 else
438 }
440 }
444 }
446 /**
447 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
448 * @q: request queue
449 * @bio: the buffer head that's been built up so far
450 * @biovec: the request that could be merged to it.
451 *
452 * Return amount of bytes we can accept at this offset
453 * If near_copies == raid_disk, there are no striping issues,
454 * but in that case, the function isn't called at all.
455 */
458 {
469 else
471 }
473 /*
474 * This routine returns the disk from which the requested read should
475 * be done. There is a per-array 'next expected sequential IO' sector
476 * number - if this matches on the next IO then we use the last disk.
477 * There is also a per-disk 'last know head position' sector that is
478 * maintained from IRQ contexts, both the normal and the resync IO
479 * completion handlers update this position correctly. If there is no
480 * perfect sequential match then we pick the disk whose head is closest.
481 *
482 * If there are 2 mirrors in the same 2 devices, performance degrades
483 * because position is mirror, not device based.
484 *
485 * The rdev for the device selected will have nr_pending incremented.
486 */
488 /*
489 * FIXME: possibly should rethink readbalancing and do it differently
490 * depending on near_copies / far_copies geometry.
491 */
493 {
502 /*
503 * Check if we can balance. We can balance on the whole
504 * device if no resync is going on (recovery is ok), or below
505 * the resync window. We take the first readable disk when
506 * above the resync window.
507 */
510 /* make sure that disk is operational */
517 slot++;
522 }
524 }
526 }
529 /* make sure the disk is operational */
535 slot ++;
539 }
541 }
547 /* Find the disk whose head is closest */
558 /* This optimisation is debatable, and completely destroys
559 * sequential read speed for 'far copies' arrays. So only
560 * keep it for 'near' arrays, and review those later.
561 */
566 }
573 }
574 }
576 rb_out:
578 /* conf->next_seq_sect = this_sector + sectors;*/
582 else
587 }
590 {
608 }
609 }
611 }
614 {
619 }
623 {
641 error_sector);
644 }
645 }
646 }
649 }
652 {
664 }
665 }
668 }
671 /* Barriers....
672 * Sometimes we need to suspend IO while we do something else,
673 * either some resync/recovery, or reconfigure the array.
674 * To do this we raise a 'barrier'.
675 * The 'barrier' is a counter that can be raised multiple times
676 * to count how many activities are happening which preclude
677 * normal IO.
678 * We can only raise the barrier if there is no pending IO.
679 * i.e. if nr_pending == 0.
680 * We choose only to raise the barrier if no-one is waiting for the
681 * barrier to go down. This means that as soon as an IO request
682 * is ready, no other operations which require a barrier will start
683 * until the IO request has had a chance.
684 *
685 * So: regular IO calls 'wait_barrier'. When that returns there
686 * is no backgroup IO happening, It must arrange to call
687 * allow_barrier when it has finished its IO.
688 * backgroup IO calls must call raise_barrier. Once that returns
689 * there is no normal IO happeing. It must arrange to call
690 * lower_barrier when the particular background IO completes.
691 */
692 #define RESYNC_DEPTH 32
695 {
699 /* Wait until no block IO is waiting (unless 'force') */
704 /* block any new IO from starting */
707 /* No wait for all pending IO to complete */
714 }
717 {
723 }
726 {
734 }
737 }
740 {
746 }
749 {
750 /* stop syncio and normal IO and wait for everything to
751 * go quiet.
752 * We increment barrier and nr_waiting, and then
753 * wait until barrier+nr_pending match nr_queued+2
754 */
763 }
766 {
767 /* reverse the effect of the freeze */
773 }
776 {
792 }
794 /* If this request crosses a chunk boundary, we need to
795 * split it. This will only happen for 1 PAGE (or less) requests.
796 */
801 /* Sanity check -- queue functions should prevent this happening */
805 /* This is a one page bio that upper layers
806 * refuse to split for us, so we need to split it.
807 */
817 bad_map:
824 }
828 /*
829 * Register the new request and wait if the reconstruction
830 * thread has put up a bar for new requests.
831 * Continue immediately if no resync is active currently.
832 */
848 /*
849 * read balancing logic:
850 */
856 }
872 }
874 /*
875 * WRITE:
876 */
877 /* first select target devices under spinlock and
878 * inc refcount on their rdev. Record them by setting
879 * bios[x] to bio
880 */
893 }
894 }
918 }
921 /* the array is dead */
925 }
937 }
940 {
951 else
953 }
961 }
964 {
968 /*
969 * If it is not operational, then we have already marked it as dead
970 * else if it is the last working disks, ignore the error, let the
971 * next level up know.
972 * else mark the drive as failed
973 */
976 /*
977 * Don't fail the drive, just return an IO error.
978 * The test should really be more sophisticated than
979 * "working_disks == 1", but it isn't critical, and
980 * can wait until we do more sophisticated "is the drive
981 * really dead" tests...
982 */
989 /*
990 * if recovery is running, make sure it aborts.
991 */
993 }
999 }
1002 {
1010 }
1022 }
1023 }
1026 {
1032 }
1034 /* check if there are enough drives for
1035 * every block to appear on atleast one
1036 */
1038 {
1046 cnt++;
1048 }
1053 }
1056 {
1061 /*
1062 * Find all non-in_sync disks within the RAID10 configuration
1063 * and mark them in_sync
1064 */
1074 }
1075 }
1079 }
1083 {
1090 /* only hot-add to in-sync arrays, as recovery is
1091 * very different from resync
1092 */
1100 else
1107 /* as we don't honour merge_bvec_fn, we must never risk
1108 * violating it, so limit ->max_sector to one PAGE, as
1109 * a one page request is never in violation.
1110 */
1122 }
1126 }
1129 {
1142 }
1146 /* lost the race, try later */
1149 }
1150 }
1151 abort:
1155 }
1159 {
1182 }
1184 /* for reconstruct, we always reschedule after a read.
1185 * for resync, only after all reads
1186 */
1189 /* we have read all the blocks,
1190 * do the comparison in process context in raid10d
1191 */
1193 }
1196 }
1199 {
1220 /* the primary of several recovery bios */
1228 }
1229 }
1232 }
1234 /*
1235 * Note: sync and recover and handled very differently for raid10
1236 * This code is for resync.
1237 * For resync, we read through virtual addresses and read all blocks.
1238 * If there is any error, we schedule a write. The lowest numbered
1239 * drive is authoritative.
1240 * However requests come for physical address, so we need to map.
1241 * For every physical address there are raid_disks/copies virtual addresses,
1242 * which is always are least one, but is not necessarly an integer.
1243 * This means that a physical address can span multiple chunks, so we may
1244 * have to submit multiple io requests for a single sync request.
1245 */
1246 /*
1247 * We check if all blocks are in-sync and only write to blocks that
1248 * aren't in sync
1249 */
1251 {
1258 /* find the first device with a block */
1269 /* now find blocks with errors */
1281 /* We know that the bi_io_vec layout is the same for
1282 * both 'first' and 'i', so we just compare them.
1283 * All vec entries are PAGE_SIZE;
1284 */
1288 PAGE_SIZE))
1293 }
1295 /* Don't fix anything. */
1297 /* Ok, we need to write this bio
1298 * First we need to fixup bv_offset, bv_len and
1299 * bi_vecs, as the read request might have corrupted these
1300 */
1321 PAGE_SIZE);
1322 }
1333 }
1335 done:
1339 }
1340 }
1342 /*
1343 * Now for the recovery code.
1344 * Recovery happens across physical sectors.
1345 * We recover all non-is_sync drives by finding the virtual address of
1346 * each, and then choose a working drive that also has that virt address.
1347 * There is a separate r10_bio for each non-in_sync drive.
1348 * Only the first two slots are in use. The first for reading,
1349 * The second for writing.
1350 *
1351 */
1354 {
1360 /* move the pages across to the second bio
1361 * and submit the write request
1362 */
1369 }
1376 else
1378 }
1381 /*
1382 * This is a kernel thread which:
1383 *
1384 * 1. Retries failed read operations on working mirrors.
1385 * 2. Updates the raid superblock when problems encounter.
1386 * 3. Performs writes following reads for array synchronising.
1387 */
1390 {
1420 }
1421 sl++;
1428 /* Cannot read from anywhere -- bye bye array */
1432 }
1435 /* write it back and re-read */
1441 sl--;
1454 /* Well, this device is dead */
1458 }
1459 }
1465 sl--;
1477 /* Well, this device is dead */
1479 else
1481 "raid10:%s: read error corrected"
1490 }
1491 }
1496 }
1497 }
1500 {
1519 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1527 }
1531 }
1550 /* we got a read error. Maybe the drive is bad. Maybe just
1551 * the block and we can fix it.
1552 * We freeze all other IO, and try reading the block from
1553 * other devices. When we find one, we re-write
1554 * and check it that fixes the read error.
1555 * This is all done synchronously while the array is
1556 * frozen.
1557 */
1562 }
1594 }
1595 }
1596 }
1600 }
1604 {
1614 }
1616 /*
1617 * perform a "sync" on one "block"
1618 *
1619 * We need to make sure that no normal I/O request - particularly write
1620 * requests - conflict with active sync requests.
1621 *
1622 * This is achieved by tracking pending requests and a 'barrier' concept
1623 * that can be installed to exclude normal IO requests.
1624 *
1625 * Resync and recovery are handled very differently.
1626 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1627 *
1628 * For resync, we iterate over virtual addresses, read all copies,
1629 * and update if there are differences. If only one copy is live,
1630 * skip it.
1631 * For recovery, we iterate over physical addresses, read a good
1632 * value for each non-in_sync drive, and over-write.
1633 *
1634 * So, for recovery we may have several outstanding complex requests for a
1635 * given address, one for each out-of-sync device. We model this by allocating
1636 * a number of r10_bio structures, one for each out-of-sync device.
1637 * As we setup these structures, we collect all bio's together into a list
1638 * which we then process collectively to add pages, and then process again
1639 * to pass to generic_make_request.
1640 *
1641 * The r10_bio structures are linked using a borrowed master_bio pointer.
1642 * This link is counted in ->remaining. When the r10_bio that points to NULL
1643 * has its remaining count decremented to 0, the whole complex operation
1644 * is complete.
1645 *
1646 */
1649 {
1666 skipped:
1671 /* If we aborted, we need to abort the
1672 * sync on the 'current' bitmap chucks (there can
1673 * be several when recovering multiple devices).
1674 * as we may have started syncing it but not finished.
1675 * We can find the current address in
1676 * mddev->curr_resync, but for recovery,
1677 * we need to convert that to several
1678 * virtual addresses.
1679 */
1685 sector_t sect =
1689 }
1697 }
1699 /* if there has been nothing to do on any drive,
1700 * then there is nothing to do at all..
1701 */
1704 }
1706 /* make sure whole request will fit in a chunk - if chunks
1707 * are meaningful
1708 */
1712 /*
1713 * If there is non-resync activity waiting for us then
1714 * put in a delay to throttle resync.
1715 */
1719 /* Again, very different code for resync and recovery.
1720 * Both must result in an r10bio with a list of bios that
1721 * have bi_end_io, bi_sector, bi_bdev set,
1722 * and bi_private set to the r10bio.
1723 * For recovery, we may actually create several r10bios
1724 * with 2 bios in each, that correspond to the bios in the main one.
1725 * In this case, the subordinate r10bios link back through a
1726 * borrowed master_bio pointer, and the counter in the master
1727 * includes a ref from each subordinate.
1728 */
1729 /* First, we decide what to do and set ->bi_end_io
1730 * To end_sync_read if we want to read, and
1731 * end_sync_write if we will want to write.
1732 */
1736 /* recovery... the complicated one */
1744 /* want to reconstruct this device */
1748 /* Unless we are doing a full sync, we only need
1749 * to recover the block if it is set in the bitmap
1750 */
1757 /* yep, skip the sync_blocks here, but don't assume
1758 * that there will never be anything to do here
1759 */
1762 }
1776 /* Need to check if this section will still be
1777 * degraded
1778 */
1785 }
1786 }
1794 /* This is where we read from */
1806 /* and we write to 'i' */
1826 }
1827 }
1829 /* Cannot recover, so abort the recovery */
1836 }
1837 }
1844 }
1846 }
1848 /* resync. Schedule a read for every block at this virt offset */
1854 /* We can skip this block */
1857 }
1891 count++;
1892 }
1899 }
1903 }
1904 }
1916 }
1932 /* stop here */
1936 /* remove last page from this bio */
1940 }
1942 }
1944 }
1948 bio_full:
1962 }
1963 }
1966 /* pretend they weren't skipped, it makes
1967 * no important difference in this case
1968 */
1972 giveup:
1973 /* There is nowhere to write, so all non-sync
1974 * drives must be failed, so try the next chunk...
1975 */
1976 {
1979 chunks_skipped ++;
1982 }
1983 }
1986 {
1998 }
2008 }
2009 /*
2010 * copy the already verified devices into our private RAID10
2011 * bookkeeping area. [whatever we allocate in run(),
2012 * should be freed in stop()]
2013 */
2020 }
2022 GFP_KERNEL);
2027 }
2045 /* 'size' is now the number of chunks in the array */
2046 /* calculate "used chunks per device" in 'stride' */
2049 /* We need to round up when dividing by raid_disks to
2050 * get the stride size.
2051 */
2058 else
2068 }
2081 /* as we don't honour merge_bvec_fn, we must never risk
2082 * violating it, so limit ->max_sector to one PAGE, as
2083 * a one page request is never in violation.
2084 */
2090 }
2097 /* need to check that every block has at least one working mirror */
2102 }
2113 }
2114 }
2123 }
2129 /*
2130 * Ok, everything is just fine now
2131 */
2140 /* Calculate max read-ahead size.
2141 * We need to readahead at least twice a whole stripe....
2142 * maybe...
2143 */
2144 {
2149 }
2155 out_free_conf:
2162 out:
2164 }
2167 {
2179 }
2182 {
2192 }
2196 else
2199 }
2200 }
2203 {
2217 };
2220 {
2222 }
2225 {
2227 }