| blob | 5cdcc938620050b39fb30593565847859d6ad1a4 |
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 */
551 /* This optimisation is debatable, and completely destroys
552 * sequential read speed for 'far copies' arrays. So only
553 * keep it for 'near' arrays, and review those later.
554 */
559 }
566 }
567 }
569 rb_out:
571 /* conf->next_seq_sect = this_sector + sectors;*/
575 else
580 }
583 {
600 }
601 }
603 }
606 {
611 }
614 {
626 }
627 }
630 }
633 /* Barriers....
634 * Sometimes we need to suspend IO while we do something else,
635 * either some resync/recovery, or reconfigure the array.
636 * To do this we raise a 'barrier'.
637 * The 'barrier' is a counter that can be raised multiple times
638 * to count how many activities are happening which preclude
639 * normal IO.
640 * We can only raise the barrier if there is no pending IO.
641 * i.e. if nr_pending == 0.
642 * We choose only to raise the barrier if no-one is waiting for the
643 * barrier to go down. This means that as soon as an IO request
644 * is ready, no other operations which require a barrier will start
645 * until the IO request has had a chance.
646 *
647 * So: regular IO calls 'wait_barrier'. When that returns there
648 * is no backgroup IO happening, It must arrange to call
649 * allow_barrier when it has finished its IO.
650 * backgroup IO calls must call raise_barrier. Once that returns
651 * there is no normal IO happeing. It must arrange to call
652 * lower_barrier when the particular background IO completes.
653 */
654 #define RESYNC_DEPTH 32
657 {
661 /* Wait until no block IO is waiting (unless 'force') */
666 /* block any new IO from starting */
669 /* No wait for all pending IO to complete */
676 }
679 {
685 }
688 {
696 }
699 }
702 {
708 }
711 {
712 /* stop syncio and normal IO and wait for everything to
713 * go quiet.
714 * We increment barrier and nr_waiting, and then
715 * wait until barrier+nr_pending match nr_queued+2
716 */
725 }
728 {
729 /* reverse the effect of the freeze */
735 }
738 {
754 }
756 /* If this request crosses a chunk boundary, we need to
757 * split it. This will only happen for 1 PAGE (or less) requests.
758 */
763 /* Sanity check -- queue functions should prevent this happening */
767 /* This is a one page bio that upper layers
768 * refuse to split for us, so we need to split it.
769 */
779 bad_map:
786 }
790 /*
791 * Register the new request and wait if the reconstruction
792 * thread has put up a bar for new requests.
793 * Continue immediately if no resync is active currently.
794 */
810 /*
811 * read balancing logic:
812 */
818 }
834 }
836 /*
837 * WRITE:
838 */
839 /* first select target devices under spinlock and
840 * inc refcount on their rdev. Record them by setting
841 * bios[x] to bio
842 */
855 }
856 }
880 }
883 /* the array is dead */
887 }
899 }
902 {
913 else
915 }
923 }
926 {
930 /*
931 * If it is not operational, then we have already marked it as dead
932 * else if it is the last working disks, ignore the error, let the
933 * next level up know.
934 * else mark the drive as failed
935 */
938 /*
939 * Don't fail the drive, just return an IO error.
940 * The test should really be more sophisticated than
941 * "working_disks == 1", but it isn't critical, and
942 * can wait until we do more sophisticated "is the drive
943 * really dead" tests...
944 */
951 /*
952 * if recovery is running, make sure it aborts.
953 */
955 }
961 }
964 {
972 }
984 }
985 }
988 {
994 }
996 /* check if there are enough drives for
997 * every block to appear on atleast one
998 */
1000 {
1008 cnt++;
1010 }
1015 }
1018 {
1023 /*
1024 * Find all non-in_sync disks within the RAID10 configuration
1025 * and mark them in_sync
1026 */
1036 }
1037 }
1041 }
1045 {
1052 /* only hot-add to in-sync arrays, as recovery is
1053 * very different from resync
1054 */
1062 else
1069 /* as we don't honour merge_bvec_fn, we must never risk
1070 * violating it, so limit ->max_sector to one PAGE, as
1071 * a one page request is never in violation.
1072 */
1084 }
1088 }
1091 {
1104 }
1108 /* lost the race, try later */
1111 }
1112 }
1113 abort:
1117 }
1121 {
1141 }
1143 /* for reconstruct, we always reschedule after a read.
1144 * for resync, only after all reads
1145 */
1148 /* we have read all the blocks,
1149 * do the comparison in process context in raid10d
1150 */
1152 }
1154 }
1157 {
1175 /* the primary of several recovery bios */
1183 }
1184 }
1186 }
1188 /*
1189 * Note: sync and recover and handled very differently for raid10
1190 * This code is for resync.
1191 * For resync, we read through virtual addresses and read all blocks.
1192 * If there is any error, we schedule a write. The lowest numbered
1193 * drive is authoritative.
1194 * However requests come for physical address, so we need to map.
1195 * For every physical address there are raid_disks/copies virtual addresses,
1196 * which is always are least one, but is not necessarly an integer.
1197 * This means that a physical address can span multiple chunks, so we may
1198 * have to submit multiple io requests for a single sync request.
1199 */
1200 /*
1201 * We check if all blocks are in-sync and only write to blocks that
1202 * aren't in sync
1203 */
1205 {
1212 /* find the first device with a block */
1223 /* now find blocks with errors */
1235 /* We know that the bi_io_vec layout is the same for
1236 * both 'first' and 'i', so we just compare them.
1237 * All vec entries are PAGE_SIZE;
1238 */
1242 PAGE_SIZE))
1247 }
1249 /* Don't fix anything. */
1251 /* Ok, we need to write this bio
1252 * First we need to fixup bv_offset, bv_len and
1253 * bi_vecs, as the read request might have corrupted these
1254 */
1275 PAGE_SIZE);
1276 }
1287 }
1289 done:
1293 }
1294 }
1296 /*
1297 * Now for the recovery code.
1298 * Recovery happens across physical sectors.
1299 * We recover all non-is_sync drives by finding the virtual address of
1300 * each, and then choose a working drive that also has that virt address.
1301 * There is a separate r10_bio for each non-in_sync drive.
1302 * Only the first two slots are in use. The first for reading,
1303 * The second for writing.
1304 *
1305 */
1308 {
1314 /* move the pages across to the second bio
1315 * and submit the write request
1316 */
1323 }
1330 else
1332 }
1335 /*
1336 * This is a kernel thread which:
1337 *
1338 * 1. Retries failed read operations on working mirrors.
1339 * 2. Updates the raid superblock when problems encounter.
1340 * 3. Performs writes following reads for array synchronising.
1341 */
1344 {
1374 }
1375 sl++;
1382 /* Cannot read from anywhere -- bye bye array */
1386 }
1389 /* write it back and re-read */
1395 sl--;
1408 /* Well, this device is dead */
1412 }
1413 }
1419 sl--;
1431 /* Well, this device is dead */
1433 else
1435 "raid10:%s: read error corrected"
1444 }
1445 }
1450 }
1451 }
1454 {
1473 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1481 }
1485 }
1504 /* we got a read error. Maybe the drive is bad. Maybe just
1505 * the block and we can fix it.
1506 * We freeze all other IO, and try reading the block from
1507 * other devices. When we find one, we re-write
1508 * and check it that fixes the read error.
1509 * This is all done synchronously while the array is
1510 * frozen.
1511 */
1516 }
1548 }
1549 }
1550 }
1554 }
1558 {
1568 }
1570 /*
1571 * perform a "sync" on one "block"
1572 *
1573 * We need to make sure that no normal I/O request - particularly write
1574 * requests - conflict with active sync requests.
1575 *
1576 * This is achieved by tracking pending requests and a 'barrier' concept
1577 * that can be installed to exclude normal IO requests.
1578 *
1579 * Resync and recovery are handled very differently.
1580 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1581 *
1582 * For resync, we iterate over virtual addresses, read all copies,
1583 * and update if there are differences. If only one copy is live,
1584 * skip it.
1585 * For recovery, we iterate over physical addresses, read a good
1586 * value for each non-in_sync drive, and over-write.
1587 *
1588 * So, for recovery we may have several outstanding complex requests for a
1589 * given address, one for each out-of-sync device. We model this by allocating
1590 * a number of r10_bio structures, one for each out-of-sync device.
1591 * As we setup these structures, we collect all bio's together into a list
1592 * which we then process collectively to add pages, and then process again
1593 * to pass to generic_make_request.
1594 *
1595 * The r10_bio structures are linked using a borrowed master_bio pointer.
1596 * This link is counted in ->remaining. When the r10_bio that points to NULL
1597 * has its remaining count decremented to 0, the whole complex operation
1598 * is complete.
1599 *
1600 */
1603 {
1620 skipped:
1625 /* If we aborted, we need to abort the
1626 * sync on the 'current' bitmap chucks (there can
1627 * be several when recovering multiple devices).
1628 * as we may have started syncing it but not finished.
1629 * We can find the current address in
1630 * mddev->curr_resync, but for recovery,
1631 * we need to convert that to several
1632 * virtual addresses.
1633 */
1639 sector_t sect =
1643 }
1651 }
1653 /* if there has been nothing to do on any drive,
1654 * then there is nothing to do at all..
1655 */
1658 }
1660 /* make sure whole request will fit in a chunk - if chunks
1661 * are meaningful
1662 */
1666 /*
1667 * If there is non-resync activity waiting for us then
1668 * put in a delay to throttle resync.
1669 */
1673 /* Again, very different code for resync and recovery.
1674 * Both must result in an r10bio with a list of bios that
1675 * have bi_end_io, bi_sector, bi_bdev set,
1676 * and bi_private set to the r10bio.
1677 * For recovery, we may actually create several r10bios
1678 * with 2 bios in each, that correspond to the bios in the main one.
1679 * In this case, the subordinate r10bios link back through a
1680 * borrowed master_bio pointer, and the counter in the master
1681 * includes a ref from each subordinate.
1682 */
1683 /* First, we decide what to do and set ->bi_end_io
1684 * To end_sync_read if we want to read, and
1685 * end_sync_write if we will want to write.
1686 */
1690 /* recovery... the complicated one */
1698 /* want to reconstruct this device */
1702 /* Unless we are doing a full sync, we only need
1703 * to recover the block if it is set in the bitmap
1704 */
1711 /* yep, skip the sync_blocks here, but don't assume
1712 * that there will never be anything to do here
1713 */
1716 }
1730 /* Need to check if this section will still be
1731 * degraded
1732 */
1739 }
1740 }
1748 /* This is where we read from */
1760 /* and we write to 'i' */
1780 }
1781 }
1783 /* Cannot recover, so abort the recovery */
1790 }
1791 }
1798 }
1800 }
1802 /* resync. Schedule a read for every block at this virt offset */
1808 /* We can skip this block */
1811 }
1845 count++;
1846 }
1853 }
1857 }
1858 }
1870 }
1886 /* stop here */
1890 /* remove last page from this bio */
1894 }
1896 }
1898 }
1902 bio_full:
1916 }
1917 }
1920 /* pretend they weren't skipped, it makes
1921 * no important difference in this case
1922 */
1926 giveup:
1927 /* There is nowhere to write, so all non-sync
1928 * drives must be failed, so try the next chunk...
1929 */
1930 {
1933 chunks_skipped ++;
1936 }
1937 }
1940 {
1952 }
1962 }
1963 /*
1964 * copy the already verified devices into our private RAID10
1965 * bookkeeping area. [whatever we allocate in run(),
1966 * should be freed in stop()]
1967 */
1974 }
1976 GFP_KERNEL);
1981 }
1999 /* 'size' is now the number of chunks in the array */
2000 /* calculate "used chunks per device" in 'stride' */
2003 /* We need to round up when dividing by raid_disks to
2004 * get the stride size.
2005 */
2012 else
2022 }
2035 /* as we don't honour merge_bvec_fn, we must never risk
2036 * violating it, so limit ->max_sector to one PAGE, as
2037 * a one page request is never in violation.
2038 */
2044 }
2051 /* need to check that every block has at least one working mirror */
2056 }
2067 }
2068 }
2077 }
2083 /*
2084 * Ok, everything is just fine now
2085 */
2093 /* Calculate max read-ahead size.
2094 * We need to readahead at least twice a whole stripe....
2095 * maybe...
2096 */
2097 {
2102 }
2108 out_free_conf:
2115 out:
2117 }
2120 {
2132 }
2135 {
2145 }
2149 else
2152 }
2153 }
2156 {
2170 };
2173 {
2175 }
2178 {
2180 }