| blob | 32389d2f18fcdfcadc87137936441e6d51560c05 |
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 {
256 /*
257 * this branch is our 'one mirror IO has finished' event handler:
258 */
262 /*
263 * Set R10BIO_Uptodate in our master bio, so that
264 * we will return a good error code to the higher
265 * levels even if IO on some other mirrored buffer fails.
266 *
267 * The 'master' represents the composite IO operation to
268 * user-side. So if something waits for IO, then it will
269 * wait for the 'master' bio.
270 */
274 /*
275 * oops, read error:
276 */
282 }
285 }
288 {
299 /*
300 * this branch is our 'one mirror IO has finished' event handler:
301 */
304 /* an I/O failed, we can't clear the bitmap */
307 /*
308 * Set R10BIO_Uptodate in our master bio, so that
309 * we will return a good error code for to the higher
310 * levels even if IO on some other mirrored buffer fails.
311 *
312 * The 'master' represents the composite IO operation to
313 * user-side. So if something waits for IO, then it will
314 * wait for the 'master' bio.
315 */
320 /*
321 *
322 * Let's see if all mirrored write operations have finished
323 * already.
324 */
326 /* clear the bitmap if all writes complete successfully */
333 }
336 }
339 /*
340 * RAID10 layout manager
341 * Aswell as the chunksize and raid_disks count, there are two
342 * parameters: near_copies and far_copies.
343 * near_copies * far_copies must be <= raid_disks.
344 * Normally one of these will be 1.
345 * If both are 1, we get raid0.
346 * If near_copies == raid_disks, we get raid1.
347 *
348 * Chunks are layed out in raid0 style with near_copies copies of the
349 * first chunk, followed by near_copies copies of the next chunk and
350 * so on.
351 * If far_copies > 1, then after 1/far_copies of the array has been assigned
352 * as described above, we start again with a device offset of near_copies.
353 * So we effectively have another copy of the whole array further down all
354 * the drives, but with blocks on different drives.
355 * With this layout, and block is never stored twice on the one device.
356 *
357 * raid10_find_phys finds the sector offset of a given virtual sector
358 * on each device that it is on.
359 *
360 * raid10_find_virt does the reverse mapping, from a device and a
361 * sector offset to a virtual address
362 */
365 {
367 sector_t sector;
368 sector_t chunk;
369 sector_t stripe;
374 /* now calculate first sector/dev */
386 /* and calculate all the others */
392 slot++;
401 slot++;
402 }
403 dev++;
407 }
408 }
410 }
413 {
429 else
431 }
433 }
437 }
439 /**
440 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
441 * @q: request queue
442 * @bio: the buffer head that's been built up so far
443 * @biovec: the request that could be merged to it.
444 *
445 * Return amount of bytes we can accept at this offset
446 * If near_copies == raid_disk, there are no striping issues,
447 * but in that case, the function isn't called at all.
448 */
451 {
462 else
464 }
466 /*
467 * This routine returns the disk from which the requested read should
468 * be done. There is a per-array 'next expected sequential IO' sector
469 * number - if this matches on the next IO then we use the last disk.
470 * There is also a per-disk 'last know head position' sector that is
471 * maintained from IRQ contexts, both the normal and the resync IO
472 * completion handlers update this position correctly. If there is no
473 * perfect sequential match then we pick the disk whose head is closest.
474 *
475 * If there are 2 mirrors in the same 2 devices, performance degrades
476 * because position is mirror, not device based.
477 *
478 * The rdev for the device selected will have nr_pending incremented.
479 */
481 /*
482 * FIXME: possibly should rethink readbalancing and do it differently
483 * depending on near_copies / far_copies geometry.
484 */
486 {
495 /*
496 * Check if we can balance. We can balance on the whole
497 * device if no resync is going on (recovery is ok), or below
498 * the resync window. We take the first readable disk when
499 * above the resync window.
500 */
503 /* make sure that disk is operational */
510 slot++;
515 }
517 }
519 }
522 /* make sure the disk is operational */
528 slot ++;
532 }
534 }
540 /* Find the disk whose head is closest,
541 * or - for far > 1 - find the closest to partition beginning */
552 /* This optimisation is debatable, and completely destroys
553 * sequential read speed for 'far copies' arrays. So only
554 * keep it for 'near' arrays, and review those later.
555 */
560 }
562 /* for far > 1 always use the lowest address */
565 else
572 }
573 }
575 rb_out:
577 /* conf->next_seq_sect = this_sector + sectors;*/
581 else
586 }
589 {
606 }
607 }
609 }
612 {
617 }
620 {
632 }
633 }
636 }
639 {
640 /* Any writes that have been queued but are awaiting
641 * bitmap updates get flushed here.
642 * We return 1 if any requests were actually submitted.
643 */
653 /* flush any pending bitmap writes to disk
654 * before proceeding w/ I/O */
662 }
667 }
668 /* Barriers....
669 * Sometimes we need to suspend IO while we do something else,
670 * either some resync/recovery, or reconfigure the array.
671 * To do this we raise a 'barrier'.
672 * The 'barrier' is a counter that can be raised multiple times
673 * to count how many activities are happening which preclude
674 * normal IO.
675 * We can only raise the barrier if there is no pending IO.
676 * i.e. if nr_pending == 0.
677 * We choose only to raise the barrier if no-one is waiting for the
678 * barrier to go down. This means that as soon as an IO request
679 * is ready, no other operations which require a barrier will start
680 * until the IO request has had a chance.
681 *
682 * So: regular IO calls 'wait_barrier'. When that returns there
683 * is no backgroup IO happening, It must arrange to call
684 * allow_barrier when it has finished its IO.
685 * backgroup IO calls must call raise_barrier. Once that returns
686 * there is no normal IO happeing. It must arrange to call
687 * lower_barrier when the particular background IO completes.
688 */
689 #define RESYNC_DEPTH 32
692 {
696 /* Wait until no block IO is waiting (unless 'force') */
701 /* block any new IO from starting */
704 /* No wait for all pending IO to complete */
711 }
714 {
720 }
723 {
731 }
734 }
737 {
743 }
746 {
747 /* stop syncio and normal IO and wait for everything to
748 * go quiet.
749 * We increment barrier and nr_waiting, and then
750 * wait until nr_pending match nr_queued+1
751 * This is called in the context of one normal IO request
752 * that has failed. Thus any sync request that might be pending
753 * will be blocked by nr_pending, and we need to wait for
754 * pending IO requests to complete or be queued for re-try.
755 * Thus the number queued (nr_queued) plus this request (1)
756 * must match the number of pending IOs (nr_pending) before
757 * we continue.
758 */
768 }
771 {
772 /* reverse the effect of the freeze */
778 }
781 {
797 }
799 /* If this request crosses a chunk boundary, we need to
800 * split it. This will only happen for 1 PAGE (or less) requests.
801 */
806 /* Sanity check -- queue functions should prevent this happening */
810 /* This is a one page bio that upper layers
811 * refuse to split for us, so we need to split it.
812 */
822 bad_map:
829 }
833 /*
834 * Register the new request and wait if the reconstruction
835 * thread has put up a bar for new requests.
836 * Continue immediately if no resync is active currently.
837 */
853 /*
854 * read balancing logic:
855 */
861 }
877 }
879 /*
880 * WRITE:
881 */
882 /* first select target devices under spinlock and
883 * inc refcount on their rdev. Record them by setting
884 * bios[x] to bio
885 */
898 }
899 }
923 }
926 /* the array is dead */
930 }
938 /* In case raid10d snuck in to freeze_array */
945 }
948 {
959 else
961 }
969 }
972 {
976 /*
977 * If it is not operational, then we have already marked it as dead
978 * else if it is the last working disks, ignore the error, let the
979 * next level up know.
980 * else mark the drive as failed
981 */
984 /*
985 * Don't fail the drive, just return an IO error.
986 * The test should really be more sophisticated than
987 * "working_disks == 1", but it isn't critical, and
988 * can wait until we do more sophisticated "is the drive
989 * really dead" tests...
990 */
997 /*
998 * if recovery is running, make sure it aborts.
999 */
1001 }
1007 }
1010 {
1018 }
1030 }
1031 }
1034 {
1040 }
1042 /* check if there are enough drives for
1043 * every block to appear on atleast one
1044 */
1046 {
1054 cnt++;
1056 }
1061 }
1064 {
1069 /*
1070 * Find all non-in_sync disks within the RAID10 configuration
1071 * and mark them in_sync
1072 */
1082 }
1083 }
1087 }
1091 {
1098 /* only hot-add to in-sync arrays, as recovery is
1099 * very different from resync
1100 */
1108 else
1115 /* as we don't honour merge_bvec_fn, we must never risk
1116 * violating it, so limit ->max_sector to one PAGE, as
1117 * a one page request is never in violation.
1118 */
1130 }
1134 }
1137 {
1150 }
1154 /* lost the race, try later */
1157 }
1158 }
1159 abort:
1163 }
1167 {
1187 }
1189 /* for reconstruct, we always reschedule after a read.
1190 * for resync, only after all reads
1191 */
1194 /* we have read all the blocks,
1195 * do the comparison in process context in raid10d
1196 */
1198 }
1200 }
1203 {
1221 /* the primary of several recovery bios */
1229 }
1230 }
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 {
1520 }
1536 /* we got a read error. Maybe the drive is bad. Maybe just
1537 * the block and we can fix it.
1538 * We freeze all other IO, and try reading the block from
1539 * other devices. When we find one, we re-write
1540 * and check it that fixes the read error.
1541 * This is all done synchronously while the array is
1542 * frozen.
1543 */
1548 }
1580 }
1581 }
1582 }
1585 }
1589 {
1599 }
1601 /*
1602 * perform a "sync" on one "block"
1603 *
1604 * We need to make sure that no normal I/O request - particularly write
1605 * requests - conflict with active sync requests.
1606 *
1607 * This is achieved by tracking pending requests and a 'barrier' concept
1608 * that can be installed to exclude normal IO requests.
1609 *
1610 * Resync and recovery are handled very differently.
1611 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1612 *
1613 * For resync, we iterate over virtual addresses, read all copies,
1614 * and update if there are differences. If only one copy is live,
1615 * skip it.
1616 * For recovery, we iterate over physical addresses, read a good
1617 * value for each non-in_sync drive, and over-write.
1618 *
1619 * So, for recovery we may have several outstanding complex requests for a
1620 * given address, one for each out-of-sync device. We model this by allocating
1621 * a number of r10_bio structures, one for each out-of-sync device.
1622 * As we setup these structures, we collect all bio's together into a list
1623 * which we then process collectively to add pages, and then process again
1624 * to pass to generic_make_request.
1625 *
1626 * The r10_bio structures are linked using a borrowed master_bio pointer.
1627 * This link is counted in ->remaining. When the r10_bio that points to NULL
1628 * has its remaining count decremented to 0, the whole complex operation
1629 * is complete.
1630 *
1631 */
1634 {
1651 skipped:
1656 /* If we aborted, we need to abort the
1657 * sync on the 'current' bitmap chucks (there can
1658 * be several when recovering multiple devices).
1659 * as we may have started syncing it but not finished.
1660 * We can find the current address in
1661 * mddev->curr_resync, but for recovery,
1662 * we need to convert that to several
1663 * virtual addresses.
1664 */
1670 sector_t sect =
1674 }
1682 }
1684 /* if there has been nothing to do on any drive,
1685 * then there is nothing to do at all..
1686 */
1689 }
1694 /* make sure whole request will fit in a chunk - if chunks
1695 * are meaningful
1696 */
1700 /*
1701 * If there is non-resync activity waiting for us then
1702 * put in a delay to throttle resync.
1703 */
1709 /* Again, very different code for resync and recovery.
1710 * Both must result in an r10bio with a list of bios that
1711 * have bi_end_io, bi_sector, bi_bdev set,
1712 * and bi_private set to the r10bio.
1713 * For recovery, we may actually create several r10bios
1714 * with 2 bios in each, that correspond to the bios in the main one.
1715 * In this case, the subordinate r10bios link back through a
1716 * borrowed master_bio pointer, and the counter in the master
1717 * includes a ref from each subordinate.
1718 */
1719 /* First, we decide what to do and set ->bi_end_io
1720 * To end_sync_read if we want to read, and
1721 * end_sync_write if we will want to write.
1722 */
1726 /* recovery... the complicated one */
1734 /* want to reconstruct this device */
1738 /* Unless we are doing a full sync, we only need
1739 * to recover the block if it is set in the bitmap
1740 */
1747 /* yep, skip the sync_blocks here, but don't assume
1748 * that there will never be anything to do here
1749 */
1752 }
1766 /* Need to check if this section will still be
1767 * degraded
1768 */
1775 }
1776 }
1784 /* This is where we read from */
1796 /* and we write to 'i' */
1816 }
1817 }
1819 /* Cannot recover, so abort the recovery */
1828 }
1829 }
1836 }
1838 }
1840 /* resync. Schedule a read for every block at this virt offset */
1846 /* We can skip this block */
1849 }
1883 count++;
1884 }
1891 }
1895 }
1896 }
1908 }
1924 /* stop here */
1928 /* remove last page from this bio */
1932 }
1934 }
1936 }
1940 bio_full:
1954 }
1955 }
1958 /* pretend they weren't skipped, it makes
1959 * no important difference in this case
1960 */
1964 giveup:
1965 /* There is nowhere to write, so all non-sync
1966 * drives must be failed, so try the next chunk...
1967 */
1968 {
1971 chunks_skipped ++;
1974 }
1975 }
1978 {
1990 }
2000 }
2001 /*
2002 * copy the already verified devices into our private RAID10
2003 * bookkeeping area. [whatever we allocate in run(),
2004 * should be freed in stop()]
2005 */
2012 }
2014 GFP_KERNEL);
2019 }
2037 /* 'size' is now the number of chunks in the array */
2038 /* calculate "used chunks per device" in 'stride' */
2041 /* We need to round up when dividing by raid_disks to
2042 * get the stride size.
2043 */
2050 else
2060 }
2073 /* as we don't honour merge_bvec_fn, we must never risk
2074 * violating it, so limit ->max_sector to one PAGE, as
2075 * a one page request is never in violation.
2076 */
2082 }
2089 /* need to check that every block has at least one working mirror */
2094 }
2105 }
2106 }
2115 }
2121 /*
2122 * Ok, everything is just fine now
2123 */
2131 /* Calculate max read-ahead size.
2132 * We need to readahead at least twice a whole stripe....
2133 * maybe...
2134 */
2135 {
2140 }
2146 out_free_conf:
2153 out:
2155 }
2158 {
2170 }
2173 {
2183 }
2187 else
2190 }
2191 }
2194 {
2208 };
2211 {
2213 }
2216 {
2218 }