| blob | 7f636283a1ba3dfcb78aa52f925fbc65fc9525a7 |
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 }
651 /* Barriers....
652 * Sometimes we need to suspend IO while we do something else,
653 * either some resync/recovery, or reconfigure the array.
654 * To do this we raise a 'barrier'.
655 * The 'barrier' is a counter that can be raised multiple times
656 * to count how many activities are happening which preclude
657 * normal IO.
658 * We can only raise the barrier if there is no pending IO.
659 * i.e. if nr_pending == 0.
660 * We choose only to raise the barrier if no-one is waiting for the
661 * barrier to go down. This means that as soon as an IO request
662 * is ready, no other operations which require a barrier will start
663 * until the IO request has had a chance.
664 *
665 * So: regular IO calls 'wait_barrier'. When that returns there
666 * is no backgroup IO happening, It must arrange to call
667 * allow_barrier when it has finished its IO.
668 * backgroup IO calls must call raise_barrier. Once that returns
669 * there is no normal IO happeing. It must arrange to call
670 * lower_barrier when the particular background IO completes.
671 */
672 #define RESYNC_DEPTH 32
675 {
679 /* Wait until no block IO is waiting (unless 'force') */
684 /* block any new IO from starting */
687 /* No wait for all pending IO to complete */
694 }
697 {
703 }
706 {
714 }
717 }
720 {
726 }
729 {
730 /* stop syncio and normal IO and wait for everything to
731 * go quiet.
732 * We increment barrier and nr_waiting, and then
733 * wait until barrier+nr_pending match nr_queued+2
734 */
743 }
746 {
747 /* reverse the effect of the freeze */
753 }
756 {
771 }
773 /* If this request crosses a chunk boundary, we need to
774 * split it. This will only happen for 1 PAGE (or less) requests.
775 */
780 /* Sanity check -- queue functions should prevent this happening */
784 /* This is a one page bio that upper layers
785 * refuse to split for us, so we need to split it.
786 */
796 bad_map:
803 }
807 /*
808 * Register the new request and wait if the reconstruction
809 * thread has put up a bar for new requests.
810 * Continue immediately if no resync is active currently.
811 */
827 /*
828 * read balancing logic:
829 */
835 }
851 }
853 /*
854 * WRITE:
855 */
856 /* first select target devices under spinlock and
857 * inc refcount on their rdev. Record them by setting
858 * bios[x] to bio
859 */
872 }
873 }
897 }
906 }
909 {
920 else
922 }
930 }
933 {
937 /*
938 * If it is not operational, then we have already marked it as dead
939 * else if it is the last working disks, ignore the error, let the
940 * next level up know.
941 * else mark the drive as failed
942 */
945 /*
946 * Don't fail the drive, just return an IO error.
947 * The test should really be more sophisticated than
948 * "working_disks == 1", but it isn't critical, and
949 * can wait until we do more sophisticated "is the drive
950 * really dead" tests...
951 */
956 /*
957 * if recovery is running, make sure it aborts.
958 */
960 }
967 }
970 {
978 }
990 }
991 }
994 {
1000 }
1002 /* check if there are enough drives for
1003 * every block to appear on atleast one
1004 */
1006 {
1014 cnt++;
1016 }
1021 }
1024 {
1029 /*
1030 * Find all non-in_sync disks within the RAID10 configuration
1031 * and mark them in_sync
1032 */
1041 }
1042 }
1046 }
1050 {
1057 /* only hot-add to in-sync arrays, as recovery is
1058 * very different from resync
1059 */
1067 else
1074 /* as we don't honour merge_bvec_fn, we must never risk
1075 * violating it, so limit ->max_sector to one PAGE, as
1076 * a one page request is never in violation.
1077 */
1089 }
1093 }
1096 {
1109 }
1113 /* lost the race, try later */
1116 }
1117 }
1118 abort:
1122 }
1126 {
1149 }
1151 /* for reconstruct, we always reschedule after a read.
1152 * for resync, only after all reads
1153 */
1156 /* we have read all the blocks,
1157 * do the comparison in process context in raid10d
1158 */
1160 }
1163 }
1166 {
1187 /* the primary of several recovery bios */
1195 }
1196 }
1199 }
1201 /*
1202 * Note: sync and recover and handled very differently for raid10
1203 * This code is for resync.
1204 * For resync, we read through virtual addresses and read all blocks.
1205 * If there is any error, we schedule a write. The lowest numbered
1206 * drive is authoritative.
1207 * However requests come for physical address, so we need to map.
1208 * For every physical address there are raid_disks/copies virtual addresses,
1209 * which is always are least one, but is not necessarly an integer.
1210 * This means that a physical address can span multiple chunks, so we may
1211 * have to submit multiple io requests for a single sync request.
1212 */
1213 /*
1214 * We check if all blocks are in-sync and only write to blocks that
1215 * aren't in sync
1216 */
1218 {
1225 /* find the first device with a block */
1236 /* now find blocks with errors */
1248 /* We know that the bi_io_vec layout is the same for
1249 * both 'first' and 'i', so we just compare them.
1250 * All vec entries are PAGE_SIZE;
1251 */
1255 PAGE_SIZE))
1260 }
1262 /* Don't fix anything. */
1264 /* Ok, we need to write this bio
1265 * First we need to fixup bv_offset, bv_len and
1266 * bi_vecs, as the read request might have corrupted these
1267 */
1288 PAGE_SIZE);
1289 }
1300 }
1302 done:
1306 }
1307 }
1309 /*
1310 * Now for the recovery code.
1311 * Recovery happens across physical sectors.
1312 * We recover all non-is_sync drives by finding the virtual address of
1313 * each, and then choose a working drive that also has that virt address.
1314 * There is a separate r10_bio for each non-in_sync drive.
1315 * Only the first two slots are in use. The first for reading,
1316 * The second for writing.
1317 *
1318 */
1321 {
1327 /* move the pages across to the second bio
1328 * and submit the write request
1329 */
1336 }
1343 else
1345 }
1348 /*
1349 * This is a kernel thread which:
1350 *
1351 * 1. Retries failed read operations on working mirrors.
1352 * 2. Updates the raid superblock when problems encounter.
1353 * 3. Performs writes following reads for array syncronising.
1354 */
1357 {
1376 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1385 }
1389 }
1408 /* we got a read error. Maybe the drive is bad. Maybe just
1409 * the block and we can fix it.
1410 * We freeze all other IO, and try reading the block from
1411 * other devices. When we find one, we re-write
1412 * and check it that fixes the read error.
1413 * This is all done synchronously while the array is
1414 * frozen.
1415 */
1444 }
1445 sl++;
1453 /* write it back and re-read */
1459 sl--;
1471 /* Well, this device is dead */
1475 }
1476 }
1482 sl--;
1493 /* Well, this device is dead */
1497 }
1498 }
1501 /* Cannot read from anywhere -- bye bye array */
1504 }
1507 }
1539 }
1540 }
1541 }
1545 }
1549 {
1559 }
1561 /*
1562 * perform a "sync" on one "block"
1563 *
1564 * We need to make sure that no normal I/O request - particularly write
1565 * requests - conflict with active sync requests.
1566 *
1567 * This is achieved by tracking pending requests and a 'barrier' concept
1568 * that can be installed to exclude normal IO requests.
1569 *
1570 * Resync and recovery are handled very differently.
1571 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1572 *
1573 * For resync, we iterate over virtual addresses, read all copies,
1574 * and update if there are differences. If only one copy is live,
1575 * skip it.
1576 * For recovery, we iterate over physical addresses, read a good
1577 * value for each non-in_sync drive, and over-write.
1578 *
1579 * So, for recovery we may have several outstanding complex requests for a
1580 * given address, one for each out-of-sync device. We model this by allocating
1581 * a number of r10_bio structures, one for each out-of-sync device.
1582 * As we setup these structures, we collect all bio's together into a list
1583 * which we then process collectively to add pages, and then process again
1584 * to pass to generic_make_request.
1585 *
1586 * The r10_bio structures are linked using a borrowed master_bio pointer.
1587 * This link is counted in ->remaining. When the r10_bio that points to NULL
1588 * has its remaining count decremented to 0, the whole complex operation
1589 * is complete.
1590 *
1591 */
1594 {
1611 skipped:
1616 /* If we aborted, we need to abort the
1617 * sync on the 'current' bitmap chucks (there can
1618 * be several when recovering multiple devices).
1619 * as we may have started syncing it but not finished.
1620 * We can find the current address in
1621 * mddev->curr_resync, but for recovery,
1622 * we need to convert that to several
1623 * virtual addresses.
1624 */
1630 sector_t sect =
1634 }
1642 }
1644 /* if there has been nothing to do on any drive,
1645 * then there is nothing to do at all..
1646 */
1649 }
1651 /* make sure whole request will fit in a chunk - if chunks
1652 * are meaningful
1653 */
1657 /*
1658 * If there is non-resync activity waiting for us then
1659 * put in a delay to throttle resync.
1660 */
1664 /* Again, very different code for resync and recovery.
1665 * Both must result in an r10bio with a list of bios that
1666 * have bi_end_io, bi_sector, bi_bdev set,
1667 * and bi_private set to the r10bio.
1668 * For recovery, we may actually create several r10bios
1669 * with 2 bios in each, that correspond to the bios in the main one.
1670 * In this case, the subordinate r10bios link back through a
1671 * borrowed master_bio pointer, and the counter in the master
1672 * includes a ref from each subordinate.
1673 */
1674 /* First, we decide what to do and set ->bi_end_io
1675 * To end_sync_read if we want to read, and
1676 * end_sync_write if we will want to write.
1677 */
1681 /* recovery... the complicated one */
1689 /* want to reconstruct this device */
1693 /* Unless we are doing a full sync, we only need
1694 * to recover the block if it is set in the bitmap
1695 */
1702 /* yep, skip the sync_blocks here, but don't assume
1703 * that there will never be anything to do here
1704 */
1707 }
1721 /* Need to check if this section will still be
1722 * degraded
1723 */
1730 }
1731 }
1739 /* This is where we read from */
1751 /* and we write to 'i' */
1770 }
1771 }
1773 /* Cannot recover, so abort the recovery */
1780 }
1781 }
1788 }
1790 }
1792 /* resync. Schedule a read for every block at this virt offset */
1798 /* We can skip this block */
1801 }
1834 count++;
1835 }
1842 }
1846 }
1847 }
1859 }
1875 /* stop here */
1879 /* remove last page from this bio */
1883 }
1885 }
1887 }
1891 bio_full:
1905 }
1906 }
1909 /* pretend they weren't skipped, it makes
1910 * no important difference in this case
1911 */
1915 giveup:
1916 /* There is nowhere to write, so all non-sync
1917 * drives must be failed, so try the next chunk...
1918 */
1919 {
1922 chunks_skipped ++;
1925 }
1926 }
1929 {
1941 }
1951 }
1952 /*
1953 * copy the already verified devices into our private RAID10
1954 * bookkeeping area. [whatever we allocate in run(),
1955 * should be freed in stop()]
1956 */
1963 }
1965 GFP_KERNEL);
1970 }
1988 }
1995 }
2008 /* as we don't honour merge_bvec_fn, we must never risk
2009 * violating it, so limit ->max_sector to one PAGE, as
2010 * a one page request is never in violation.
2011 */
2019 }
2028 /* need to check that every block has at least one working mirror */
2033 }
2044 }
2045 }
2054 }
2060 /*
2061 * Ok, everything is just fine now
2062 */
2077 /* Calculate max read-ahead size.
2078 * We need to readahead at least twice a whole stripe....
2079 * maybe...
2080 */
2081 {
2086 }
2092 out_free_conf:
2099 out:
2101 }
2104 {
2116 }
2119 {
2129 }
2133 else
2136 }
2137 }
2140 {
2154 };
2157 {
2159 }
2162 {
2164 }