| blob | 1440935414e63825adfb5278d2aad8d0be6e1822 |
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 *
33 * The data to be stored is divided into chunks using chunksize.
34 * Each device is divided into far_copies sections.
35 * In each section, chunks are laid out in a style similar to raid0, but
36 * near_copies copies of each chunk is stored (each on a different drive).
37 * The starting device for each section is offset near_copies from the starting
38 * device of the previous section.
39 * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
40 * drive.
41 * near_copies and far_copies must be at least one, and their product is at most
42 * raid_disks.
43 */
45 /*
46 * Number of guaranteed r10bios in case of extreme VM load:
47 */
48 #define NR_RAID10_BIOS 256
56 {
61 /* allocate a r10bio with room for raid_disks entries in the bios array */
67 }
70 {
72 }
74 #define RESYNC_BLOCK_SIZE (64*1024)
75 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
76 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
77 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
78 #define RESYNC_WINDOW (2048*1024)
80 /*
81 * When performing a resync, we need to read and compare, so
82 * we need as many pages are there are copies.
83 * When performing a recovery, we need 2 bios, one for read,
84 * one for write (we recover only one drive per r10buf)
85 *
86 */
88 {
100 }
104 else
107 /*
108 * Allocate bios.
109 */
115 }
116 /*
117 * Allocate RESYNC_PAGES data pages and attach them
118 * where needed.
119 */
128 }
129 }
133 out_free_pages:
140 out_free_bio:
145 }
148 {
160 }
162 }
163 }
165 }
168 {
176 }
177 }
180 {
183 /*
184 * Wake up any possible resync thread that waits for the device
185 * to go idle.
186 */
191 }
194 {
200 }
203 {
214 }
216 /*
217 * raid_end_bio_io() is called when we have finished servicing a mirrored
218 * operation and are ready to return a success/failure code to the buffer
219 * cache layer.
220 */
222 {
228 }
230 /*
231 * Update disk head position estimator based on IRQ completion info.
232 */
234 {
239 }
242 {
253 /*
254 * this branch is our 'one mirror IO has finished' event handler:
255 */
259 /*
260 * Set R10BIO_Uptodate in our master bio, so that
261 * we will return a good error code to the higher
262 * levels even if IO on some other mirrored buffer fails.
263 *
264 * The 'master' represents the composite IO operation to
265 * user-side. So if something waits for IO, then it will
266 * wait for the 'master' bio.
267 */
271 /*
272 * oops, read error:
273 */
279 }
283 }
286 {
300 /*
301 * this branch is our 'one mirror IO has finished' event handler:
302 */
305 /* an I/O failed, we can't clear the bitmap */
308 /*
309 * Set R10BIO_Uptodate in our master bio, so that
310 * we will return a good error code for to the higher
311 * levels even if IO on some other mirrored buffer fails.
312 *
313 * The 'master' represents the composite IO operation to
314 * user-side. So if something waits for IO, then it will
315 * wait for the 'master' bio.
316 */
321 /*
322 *
323 * Let's see if all mirrored write operations have finished
324 * already.
325 */
327 /* clear the bitmap if all writes complete successfully */
334 }
338 }
341 /*
342 * RAID10 layout manager
343 * Aswell as the chunksize and raid_disks count, there are two
344 * parameters: near_copies and far_copies.
345 * near_copies * far_copies must be <= raid_disks.
346 * Normally one of these will be 1.
347 * If both are 1, we get raid0.
348 * If near_copies == raid_disks, we get raid1.
349 *
350 * Chunks are layed out in raid0 style with near_copies copies of the
351 * first chunk, followed by near_copies copies of the next chunk and
352 * so on.
353 * If far_copies > 1, then after 1/far_copies of the array has been assigned
354 * as described above, we start again with a device offset of near_copies.
355 * So we effectively have another copy of the whole array further down all
356 * the drives, but with blocks on different drives.
357 * With this layout, and block is never stored twice on the one device.
358 *
359 * raid10_find_phys finds the sector offset of a given virtual sector
360 * on each device that it is on. If a block isn't on a device,
361 * that entry in the array is set to MaxSector.
362 *
363 * raid10_find_virt does the reverse mapping, from a device and a
364 * sector offset to a virtual address
365 */
368 {
370 sector_t sector;
371 sector_t chunk;
372 sector_t stripe;
377 /* now calculate first sector/dev */
387 /* and calculate all the others */
393 slot++;
402 slot++;
403 }
404 dev++;
408 }
409 }
411 }
414 {
421 else
423 }
430 }
432 /**
433 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
434 * @q: request queue
435 * @bio: the buffer head that's been built up so far
436 * @biovec: the request that could be merged to it.
437 *
438 * Return amount of bytes we can accept at this offset
439 * If near_copies == raid_disk, there are no striping issues,
440 * but in that case, the function isn't called at all.
441 */
444 {
455 else
457 }
459 /*
460 * This routine returns the disk from which the requested read should
461 * be done. There is a per-array 'next expected sequential IO' sector
462 * number - if this matches on the next IO then we use the last disk.
463 * There is also a per-disk 'last know head position' sector that is
464 * maintained from IRQ contexts, both the normal and the resync IO
465 * completion handlers update this position correctly. If there is no
466 * perfect sequential match then we pick the disk whose head is closest.
467 *
468 * If there are 2 mirrors in the same 2 devices, performance degrades
469 * because position is mirror, not device based.
470 *
471 * The rdev for the device selected will have nr_pending incremented.
472 */
474 /*
475 * FIXME: possibly should rethink readbalancing and do it differently
476 * depending on near_copies / far_copies geometry.
477 */
479 {
488 /*
489 * Check if we can balance. We can balance on the whole
490 * device if no resync is going on (recovery is ok), or below
491 * the resync window. We take the first readable disk when
492 * above the resync window.
493 */
496 /* make sure that disk is operational */
503 slot++;
508 }
510 }
512 }
515 /* make sure the disk is operational */
521 slot ++;
525 }
527 }
533 /* Find the disk whose head is closest */
544 /* This optimisation is debatable, and completely destroys
545 * sequential read speed for 'far copies' arrays. So only
546 * keep it for 'near' arrays, and review those later.
547 */
552 }
559 }
560 }
562 rb_out:
564 /* conf->next_seq_sect = this_sector + sectors;*/
568 else
573 }
576 {
594 }
595 }
597 }
600 {
605 }
609 {
627 error_sector);
630 }
631 }
632 }
635 }
637 /* Barriers....
638 * Sometimes we need to suspend IO while we do something else,
639 * either some resync/recovery, or reconfigure the array.
640 * To do this we raise a 'barrier'.
641 * The 'barrier' is a counter that can be raised multiple times
642 * to count how many activities are happening which preclude
643 * normal IO.
644 * We can only raise the barrier if there is no pending IO.
645 * i.e. if nr_pending == 0.
646 * We choose only to raise the barrier if no-one is waiting for the
647 * barrier to go down. This means that as soon as an IO request
648 * is ready, no other operations which require a barrier will start
649 * until the IO request has had a chance.
650 *
651 * So: regular IO calls 'wait_barrier'. When that returns there
652 * is no backgroup IO happening, It must arrange to call
653 * allow_barrier when it has finished its IO.
654 * backgroup IO calls must call raise_barrier. Once that returns
655 * there is no normal IO happeing. It must arrange to call
656 * lower_barrier when the particular background IO completes.
657 */
658 #define RESYNC_DEPTH 32
661 {
665 /* Wait until no block IO is waiting (unless 'force') */
670 /* block any new IO from starting */
673 /* No wait for all pending IO to complete */
680 }
683 {
689 }
692 {
700 }
703 }
706 {
712 }
715 {
716 /* stop syncio and normal IO and wait for everything to
717 * go quiet.
718 * We increment barrier and nr_waiting, and then
719 * wait until barrier+nr_pending match nr_queued+2
720 */
729 }
732 {
733 /* reverse the effect of the freeze */
739 }
742 {
757 }
759 /* If this request crosses a chunk boundary, we need to
760 * split it. This will only happen for 1 PAGE (or less) requests.
761 */
766 /* Sanity check -- queue functions should prevent this happening */
770 /* This is a one page bio that upper layers
771 * refuse to split for us, so we need to split it.
772 */
782 bad_map:
789 }
793 /*
794 * Register the new request and wait if the reconstruction
795 * thread has put up a bar for new requests.
796 * Continue immediately if no resync is active currently.
797 */
813 /*
814 * read balancing logic:
815 */
821 }
837 }
839 /*
840 * WRITE:
841 */
842 /* first select target devices under spinlock and
843 * inc refcount on their rdev. Record them by setting
844 * bios[x] to bio
845 */
858 }
859 }
883 }
892 }
895 {
913 }
916 {
920 /*
921 * If it is not operational, then we have already marked it as dead
922 * else if it is the last working disks, ignore the error, let the
923 * next level up know.
924 * else mark the drive as failed
925 */
928 /*
929 * Don't fail the drive, just return an IO error.
930 * The test should really be more sophisticated than
931 * "working_disks == 1", but it isn't critical, and
932 * can wait until we do more sophisticated "is the drive
933 * really dead" tests...
934 */
939 /*
940 * if recovery is running, make sure it aborts.
941 */
943 }
950 }
953 {
961 }
973 }
974 }
977 {
983 }
985 /* check if there are enough drives for
986 * every block to appear on atleast one
987 */
989 {
997 cnt++;
999 }
1004 }
1007 {
1012 /*
1013 * Find all non-in_sync disks within the RAID10 configuration
1014 * and mark them in_sync
1015 */
1024 }
1025 }
1029 }
1033 {
1040 /* only hot-add to in-sync arrays, as recovery is
1041 * very different from resync
1042 */
1050 else
1057 /* as we don't honour merge_bvec_fn, we must never risk
1058 * violating it, so limit ->max_sector to one PAGE, as
1059 * a one page request is never in violation.
1060 */
1072 }
1076 }
1079 {
1092 }
1096 /* lost the race, try later */
1099 }
1100 }
1101 abort:
1105 }
1109 {
1132 }
1134 /* for reconstruct, we always reschedule after a read.
1135 * for resync, only after all reads
1136 */
1139 /* we have read all the blocks,
1140 * do the comparison in process context in raid10d
1141 */
1143 }
1146 }
1149 {
1170 /* the primary of several recovery bios */
1178 }
1179 }
1182 }
1184 /*
1185 * Note: sync and recover and handled very differently for raid10
1186 * This code is for resync.
1187 * For resync, we read through virtual addresses and read all blocks.
1188 * If there is any error, we schedule a write. The lowest numbered
1189 * drive is authoritative.
1190 * However requests come for physical address, so we need to map.
1191 * For every physical address there are raid_disks/copies virtual addresses,
1192 * which is always are least one, but is not necessarly an integer.
1193 * This means that a physical address can span multiple chunks, so we may
1194 * have to submit multiple io requests for a single sync request.
1195 */
1196 /*
1197 * We check if all blocks are in-sync and only write to blocks that
1198 * aren't in sync
1199 */
1201 {
1208 /* find the first device with a block */
1219 /* now find blocks with errors */
1231 /* We know that the bi_io_vec layout is the same for
1232 * both 'first' and 'i', so we just compare them.
1233 * All vec entries are PAGE_SIZE;
1234 */
1238 PAGE_SIZE))
1243 }
1245 /* Don't fix anything. */
1247 /* Ok, we need to write this bio
1248 * First we need to fixup bv_offset, bv_len and
1249 * bi_vecs, as the read request might have corrupted these
1250 */
1271 PAGE_SIZE);
1272 }
1283 }
1285 done:
1289 }
1290 }
1292 /*
1293 * Now for the recovery code.
1294 * Recovery happens across physical sectors.
1295 * We recover all non-is_sync drives by finding the virtual address of
1296 * each, and then choose a working drive that also has that virt address.
1297 * There is a separate r10_bio for each non-in_sync drive.
1298 * Only the first two slots are in use. The first for reading,
1299 * The second for writing.
1300 *
1301 */
1304 {
1310 /* move the pages across to the second bio
1311 * and submit the write request
1312 */
1319 }
1326 else
1328 }
1331 /*
1332 * This is a kernel thread which:
1333 *
1334 * 1. Retries failed read operations on working mirrors.
1335 * 2. Updates the raid superblock when problems encounter.
1336 * 3. Performs writes following reads for array syncronising.
1337 */
1340 {
1359 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1368 }
1372 }
1391 /* we got a read error. Maybe the drive is bad. Maybe just
1392 * the block and we can fix it.
1393 * We freeze all other IO, and try reading the block from
1394 * other devices. When we find one, we re-write
1395 * and check it that fixes the read error.
1396 * This is all done synchronously while the array is
1397 * frozen.
1398 */
1427 }
1428 sl++;
1436 /* write it back and re-read */
1442 sl--;
1454 /* Well, this device is dead */
1458 }
1459 }
1465 sl--;
1476 /* Well, this device is dead */
1480 }
1481 }
1484 /* Cannot read from anywhere -- bye bye array */
1487 }
1490 }
1522 }
1523 }
1524 }
1528 }
1532 {
1542 }
1544 /*
1545 * perform a "sync" on one "block"
1546 *
1547 * We need to make sure that no normal I/O request - particularly write
1548 * requests - conflict with active sync requests.
1549 *
1550 * This is achieved by tracking pending requests and a 'barrier' concept
1551 * that can be installed to exclude normal IO requests.
1552 *
1553 * Resync and recovery are handled very differently.
1554 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1555 *
1556 * For resync, we iterate over virtual addresses, read all copies,
1557 * and update if there are differences. If only one copy is live,
1558 * skip it.
1559 * For recovery, we iterate over physical addresses, read a good
1560 * value for each non-in_sync drive, and over-write.
1561 *
1562 * So, for recovery we may have several outstanding complex requests for a
1563 * given address, one for each out-of-sync device. We model this by allocating
1564 * a number of r10_bio structures, one for each out-of-sync device.
1565 * As we setup these structures, we collect all bio's together into a list
1566 * which we then process collectively to add pages, and then process again
1567 * to pass to generic_make_request.
1568 *
1569 * The r10_bio structures are linked using a borrowed master_bio pointer.
1570 * This link is counted in ->remaining. When the r10_bio that points to NULL
1571 * has its remaining count decremented to 0, the whole complex operation
1572 * is complete.
1573 *
1574 */
1577 {
1594 skipped:
1599 /* If we aborted, we need to abort the
1600 * sync on the 'current' bitmap chucks (there can
1601 * be several when recovering multiple devices).
1602 * as we may have started syncing it but not finished.
1603 * We can find the current address in
1604 * mddev->curr_resync, but for recovery,
1605 * we need to convert that to several
1606 * virtual addresses.
1607 */
1613 sector_t sect =
1617 }
1625 }
1627 /* if there has been nothing to do on any drive,
1628 * then there is nothing to do at all..
1629 */
1632 }
1634 /* make sure whole request will fit in a chunk - if chunks
1635 * are meaningful
1636 */
1640 /*
1641 * If there is non-resync activity waiting for us then
1642 * put in a delay to throttle resync.
1643 */
1647 /* Again, very different code for resync and recovery.
1648 * Both must result in an r10bio with a list of bios that
1649 * have bi_end_io, bi_sector, bi_bdev set,
1650 * and bi_private set to the r10bio.
1651 * For recovery, we may actually create several r10bios
1652 * with 2 bios in each, that correspond to the bios in the main one.
1653 * In this case, the subordinate r10bios link back through a
1654 * borrowed master_bio pointer, and the counter in the master
1655 * includes a ref from each subordinate.
1656 */
1657 /* First, we decide what to do and set ->bi_end_io
1658 * To end_sync_read if we want to read, and
1659 * end_sync_write if we will want to write.
1660 */
1664 /* recovery... the complicated one */
1672 /* want to reconstruct this device */
1676 /* Unless we are doing a full sync, we only need
1677 * to recover the block if it is set in the bitmap
1678 */
1685 /* yep, skip the sync_blocks here, but don't assume
1686 * that there will never be anything to do here
1687 */
1690 }
1704 /* Need to check if this section will still be
1705 * degraded
1706 */
1713 }
1714 }
1722 /* This is where we read from */
1734 /* and we write to 'i' */
1753 }
1754 }
1756 /* Cannot recover, so abort the recovery */
1763 }
1764 }
1771 }
1773 }
1775 /* resync. Schedule a read for every block at this virt offset */
1781 /* We can skip this block */
1784 }
1817 count++;
1818 }
1825 }
1829 }
1830 }
1842 }
1858 /* stop here */
1862 /* remove last page from this bio */
1866 }
1868 }
1870 }
1874 bio_full:
1888 }
1889 }
1892 /* pretend they weren't skipped, it makes
1893 * no important difference in this case
1894 */
1898 giveup:
1899 /* There is nowhere to write, so all non-sync
1900 * drives must be failed, so try the next chunk...
1901 */
1902 {
1905 chunks_skipped ++;
1908 }
1909 }
1912 {
1924 }
1933 }
1934 /*
1935 * copy the already verified devices into our private RAID10
1936 * bookkeeping area. [whatever we allocate in run(),
1937 * should be freed in stop()]
1938 */
1945 }
1947 GFP_KERNEL);
1952 }
1973 }
1986 /* as we don't honour merge_bvec_fn, we must never risk
1987 * violating it, so limit ->max_sector to one PAGE, as
1988 * a one page request is never in violation.
1989 */
1997 }
2006 /* need to check that every block has at least one working mirror */
2011 }
2021 }
2022 }
2031 }
2037 /*
2038 * Ok, everything is just fine now
2039 */
2048 /* Calculate max read-ahead size.
2049 * We need to readahead at least twice a whole stripe....
2050 * maybe...
2051 */
2052 {
2057 }
2063 out_free_conf:
2070 out:
2072 }
2075 {
2087 }
2090 {
2100 }
2104 else
2107 }
2108 }
2111 {
2125 };
2128 {
2130 }
2133 {
2135 }