| blob | a95ada1cfac4ad88e4a3d6912d90a4162fd4fb21 |
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 }
930 }
933 {
944 else
946 }
954 }
957 {
961 /*
962 * If it is not operational, then we have already marked it as dead
963 * else if it is the last working disks, ignore the error, let the
964 * next level up know.
965 * else mark the drive as failed
966 */
969 /*
970 * Don't fail the drive, just return an IO error.
971 * The test should really be more sophisticated than
972 * "working_disks == 1", but it isn't critical, and
973 * can wait until we do more sophisticated "is the drive
974 * really dead" tests...
975 */
982 /*
983 * if recovery is running, make sure it aborts.
984 */
986 }
992 }
995 {
1003 }
1015 }
1016 }
1019 {
1025 }
1027 /* check if there are enough drives for
1028 * every block to appear on atleast one
1029 */
1031 {
1039 cnt++;
1041 }
1046 }
1049 {
1054 /*
1055 * Find all non-in_sync disks within the RAID10 configuration
1056 * and mark them in_sync
1057 */
1067 }
1068 }
1072 }
1076 {
1083 /* only hot-add to in-sync arrays, as recovery is
1084 * very different from resync
1085 */
1093 else
1100 /* as we don't honour merge_bvec_fn, we must never risk
1101 * violating it, so limit ->max_sector to one PAGE, as
1102 * a one page request is never in violation.
1103 */
1115 }
1119 }
1122 {
1135 }
1139 /* lost the race, try later */
1142 }
1143 }
1144 abort:
1148 }
1152 {
1175 }
1177 /* for reconstruct, we always reschedule after a read.
1178 * for resync, only after all reads
1179 */
1182 /* we have read all the blocks,
1183 * do the comparison in process context in raid10d
1184 */
1186 }
1189 }
1192 {
1213 /* the primary of several recovery bios */
1221 }
1222 }
1225 }
1227 /*
1228 * Note: sync and recover and handled very differently for raid10
1229 * This code is for resync.
1230 * For resync, we read through virtual addresses and read all blocks.
1231 * If there is any error, we schedule a write. The lowest numbered
1232 * drive is authoritative.
1233 * However requests come for physical address, so we need to map.
1234 * For every physical address there are raid_disks/copies virtual addresses,
1235 * which is always are least one, but is not necessarly an integer.
1236 * This means that a physical address can span multiple chunks, so we may
1237 * have to submit multiple io requests for a single sync request.
1238 */
1239 /*
1240 * We check if all blocks are in-sync and only write to blocks that
1241 * aren't in sync
1242 */
1244 {
1251 /* find the first device with a block */
1262 /* now find blocks with errors */
1274 /* We know that the bi_io_vec layout is the same for
1275 * both 'first' and 'i', so we just compare them.
1276 * All vec entries are PAGE_SIZE;
1277 */
1281 PAGE_SIZE))
1286 }
1288 /* Don't fix anything. */
1290 /* Ok, we need to write this bio
1291 * First we need to fixup bv_offset, bv_len and
1292 * bi_vecs, as the read request might have corrupted these
1293 */
1314 PAGE_SIZE);
1315 }
1326 }
1328 done:
1332 }
1333 }
1335 /*
1336 * Now for the recovery code.
1337 * Recovery happens across physical sectors.
1338 * We recover all non-is_sync drives by finding the virtual address of
1339 * each, and then choose a working drive that also has that virt address.
1340 * There is a separate r10_bio for each non-in_sync drive.
1341 * Only the first two slots are in use. The first for reading,
1342 * The second for writing.
1343 *
1344 */
1347 {
1353 /* move the pages across to the second bio
1354 * and submit the write request
1355 */
1362 }
1369 else
1371 }
1374 /*
1375 * This is a kernel thread which:
1376 *
1377 * 1. Retries failed read operations on working mirrors.
1378 * 2. Updates the raid superblock when problems encounter.
1379 * 3. Performs writes following reads for array synchronising.
1380 */
1383 {
1413 }
1414 sl++;
1421 /* Cannot read from anywhere -- bye bye array */
1425 }
1428 /* write it back and re-read */
1434 sl--;
1447 /* Well, this device is dead */
1451 }
1452 }
1458 sl--;
1470 /* Well, this device is dead */
1472 else
1474 "raid10:%s: read error corrected"
1483 }
1484 }
1489 }
1490 }
1493 {
1512 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1520 }
1524 }
1543 /* we got a read error. Maybe the drive is bad. Maybe just
1544 * the block and we can fix it.
1545 * We freeze all other IO, and try reading the block from
1546 * other devices. When we find one, we re-write
1547 * and check it that fixes the read error.
1548 * This is all done synchronously while the array is
1549 * frozen.
1550 */
1555 }
1586 }
1587 }
1588 }
1592 }
1596 {
1606 }
1608 /*
1609 * perform a "sync" on one "block"
1610 *
1611 * We need to make sure that no normal I/O request - particularly write
1612 * requests - conflict with active sync requests.
1613 *
1614 * This is achieved by tracking pending requests and a 'barrier' concept
1615 * that can be installed to exclude normal IO requests.
1616 *
1617 * Resync and recovery are handled very differently.
1618 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1619 *
1620 * For resync, we iterate over virtual addresses, read all copies,
1621 * and update if there are differences. If only one copy is live,
1622 * skip it.
1623 * For recovery, we iterate over physical addresses, read a good
1624 * value for each non-in_sync drive, and over-write.
1625 *
1626 * So, for recovery we may have several outstanding complex requests for a
1627 * given address, one for each out-of-sync device. We model this by allocating
1628 * a number of r10_bio structures, one for each out-of-sync device.
1629 * As we setup these structures, we collect all bio's together into a list
1630 * which we then process collectively to add pages, and then process again
1631 * to pass to generic_make_request.
1632 *
1633 * The r10_bio structures are linked using a borrowed master_bio pointer.
1634 * This link is counted in ->remaining. When the r10_bio that points to NULL
1635 * has its remaining count decremented to 0, the whole complex operation
1636 * is complete.
1637 *
1638 */
1641 {
1658 skipped:
1663 /* If we aborted, we need to abort the
1664 * sync on the 'current' bitmap chucks (there can
1665 * be several when recovering multiple devices).
1666 * as we may have started syncing it but not finished.
1667 * We can find the current address in
1668 * mddev->curr_resync, but for recovery,
1669 * we need to convert that to several
1670 * virtual addresses.
1671 */
1677 sector_t sect =
1681 }
1689 }
1691 /* if there has been nothing to do on any drive,
1692 * then there is nothing to do at all..
1693 */
1696 }
1698 /* make sure whole request will fit in a chunk - if chunks
1699 * are meaningful
1700 */
1704 /*
1705 * If there is non-resync activity waiting for us then
1706 * put in a delay to throttle resync.
1707 */
1711 /* Again, very different code for resync and recovery.
1712 * Both must result in an r10bio with a list of bios that
1713 * have bi_end_io, bi_sector, bi_bdev set,
1714 * and bi_private set to the r10bio.
1715 * For recovery, we may actually create several r10bios
1716 * with 2 bios in each, that correspond to the bios in the main one.
1717 * In this case, the subordinate r10bios link back through a
1718 * borrowed master_bio pointer, and the counter in the master
1719 * includes a ref from each subordinate.
1720 */
1721 /* First, we decide what to do and set ->bi_end_io
1722 * To end_sync_read if we want to read, and
1723 * end_sync_write if we will want to write.
1724 */
1728 /* recovery... the complicated one */
1736 /* want to reconstruct this device */
1740 /* Unless we are doing a full sync, we only need
1741 * to recover the block if it is set in the bitmap
1742 */
1749 /* yep, skip the sync_blocks here, but don't assume
1750 * that there will never be anything to do here
1751 */
1754 }
1768 /* Need to check if this section will still be
1769 * degraded
1770 */
1777 }
1778 }
1786 /* This is where we read from */
1798 /* and we write to 'i' */
1818 }
1819 }
1821 /* 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 */
2132 /* Calculate max read-ahead size.
2133 * We need to readahead at least twice a whole stripe....
2134 * maybe...
2135 */
2136 {
2141 }
2147 out_free_conf:
2154 out:
2156 }
2159 {
2171 }
2174 {
2184 }
2188 else
2191 }
2192 }
2195 {
2209 };
2212 {
2214 }
2217 {
2219 }