| blob | 17cb6ab62308f449c81419bc380db57cb821d7e3 |
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 further 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 */
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/seq_file.h>
30 /*
31 * RAID10 provides a combination of RAID0 and RAID1 functionality.
32 * The layout of data is defined by
33 * chunk_size
34 * raid_disks
35 * near_copies (stored in low byte of layout)
36 * far_copies (stored in second byte of layout)
37 * far_offset (stored in bit 16 of layout )
38 *
39 * The data to be stored is divided into chunks using chunksize.
40 * Each device is divided into far_copies sections.
41 * In each section, chunks are laid out in a style similar to raid0, but
42 * near_copies copies of each chunk is stored (each on a different drive).
43 * The starting device for each section is offset near_copies from the starting
44 * device of the previous section.
45 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
46 * drive.
47 * near_copies and far_copies must be at least one, and their product is at most
48 * raid_disks.
49 *
50 * If far_offset is true, then the far_copies are handled a bit differently.
51 * The copies are still in different stripes, but instead of be very far apart
52 * on disk, there are adjacent stripes.
53 */
55 /*
56 * Number of guaranteed r10bios in case of extreme VM load:
57 */
58 #define NR_RAID10_BIOS 256
64 {
68 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 }
73 {
75 }
77 /* Maximum size of each resync request */
78 #define RESYNC_BLOCK_SIZE (64*1024)
79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
80 /* amount of memory to reserve for resync requests */
81 #define RESYNC_WINDOW (1024*1024)
82 /* maximum number of concurrent requests, memory permitting */
83 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
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 {
107 else
110 /*
111 * Allocate bios.
112 */
118 }
119 /*
120 * Allocate RESYNC_PAGES data pages and attach them
121 * where needed.
122 */
131 }
132 }
136 out_free_pages:
143 out_free_bio:
148 }
151 {
163 }
165 }
166 }
168 }
171 {
179 }
180 }
183 {
186 /*
187 * Wake up any possible resync thread that waits for the device
188 * to go idle.
189 */
194 }
197 {
203 }
206 {
216 /* wake up frozen array... */
220 }
222 /*
223 * raid_end_bio_io() is called when we have finished servicing a mirrored
224 * operation and are ready to return a success/failure code to the buffer
225 * cache layer.
226 */
228 {
234 }
236 /*
237 * Update disk head position estimator based on IRQ completion info.
238 */
240 {
245 }
248 {
257 /*
258 * this branch is our 'one mirror IO has finished' event handler:
259 */
263 /*
264 * Set R10BIO_Uptodate in our master bio, so that
265 * we will return a good error code to the higher
266 * levels even if IO on some other mirrored buffer fails.
267 *
268 * The 'master' represents the composite IO operation to
269 * user-side. So if something waits for IO, then it will
270 * wait for the 'master' bio.
271 */
276 /*
277 * oops, read error - keep the refcount on the rdev
278 */
285 }
286 }
289 {
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 }
337 }
340 /*
341 * RAID10 layout manager
342 * As well as the chunksize and raid_disks count, there are two
343 * parameters: near_copies and far_copies.
344 * near_copies * far_copies must be <= raid_disks.
345 * Normally one of these will be 1.
346 * If both are 1, we get raid0.
347 * If near_copies == raid_disks, we get raid1.
348 *
349 * Chunks are laid out in raid0 style with near_copies copies of the
350 * first chunk, followed by near_copies copies of the next chunk and
351 * so on.
352 * If far_copies > 1, then after 1/far_copies of the array has been assigned
353 * as described above, we start again with a device offset of near_copies.
354 * So we effectively have another copy of the whole array further down all
355 * the drives, but with blocks on different drives.
356 * With this layout, and block is never stored twice on the one device.
357 *
358 * raid10_find_phys finds the sector offset of a given virtual sector
359 * on each device that it is on.
360 *
361 * raid10_find_virt does the reverse mapping, from a device and a
362 * sector offset to a virtual address
363 */
366 {
368 sector_t sector;
369 sector_t chunk;
370 sector_t stripe;
375 /* now calculate first sector/dev */
387 /* and calculate all the others */
393 slot++;
402 slot++;
403 }
404 dev++;
408 }
409 }
411 }
414 {
430 else
432 }
434 }
438 }
440 /**
441 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
442 * @q: request queue
443 * @bvm: properties of new bio
444 * @biovec: the request that could be merged to it.
445 *
446 * Return amount of bytes we can accept at this offset
447 * If near_copies == raid_disk, there are no striping issues,
448 * but in that case, the function isn't called at all.
449 */
453 {
464 else
466 }
468 /*
469 * This routine returns the disk from which the requested read should
470 * be done. There is a per-array 'next expected sequential IO' sector
471 * number - if this matches on the next IO then we use the last disk.
472 * There is also a per-disk 'last know head position' sector that is
473 * maintained from IRQ contexts, both the normal and the resync IO
474 * completion handlers update this position correctly. If there is no
475 * perfect sequential match then we pick the disk whose head is closest.
476 *
477 * If there are 2 mirrors in the same 2 devices, performance degrades
478 * because position is mirror, not device based.
479 *
480 * The rdev for the device selected will have nr_pending incremented.
481 */
483 /*
484 * FIXME: possibly should rethink readbalancing and do it differently
485 * depending on near_copies / far_copies geometry.
486 */
488 {
499 retry:
503 /*
504 * Check if we can balance. We can balance on the whole
505 * device if no resync is going on (recovery is ok), or below
506 * the resync window. We take the first readable disk when
507 * above the resync window.
508 */
526 /* This optimisation is debatable, and completely destroys
527 * sequential read speed for 'far copies' arrays. So only
528 * keep it for 'near' arrays, and review those later.
529 */
533 /* for far > 1 always use the lowest address */
536 else
542 }
543 }
554 /* Cannot risk returning a device that failed
555 * before we inc'ed nr_pending
556 */
559 }
566 }
569 {
583 }
584 }
587 }
590 {
591 /* Any writes that have been queued but are awaiting
592 * bitmap updates get flushed here.
593 */
600 /* flush any pending bitmap writes to disk
601 * before proceeding w/ I/O */
609 }
612 }
614 /* Barriers....
615 * Sometimes we need to suspend IO while we do something else,
616 * either some resync/recovery, or reconfigure the array.
617 * To do this we raise a 'barrier'.
618 * The 'barrier' is a counter that can be raised multiple times
619 * to count how many activities are happening which preclude
620 * normal IO.
621 * We can only raise the barrier if there is no pending IO.
622 * i.e. if nr_pending == 0.
623 * We choose only to raise the barrier if no-one is waiting for the
624 * barrier to go down. This means that as soon as an IO request
625 * is ready, no other operations which require a barrier will start
626 * until the IO request has had a chance.
627 *
628 * So: regular IO calls 'wait_barrier'. When that returns there
629 * is no backgroup IO happening, It must arrange to call
630 * allow_barrier when it has finished its IO.
631 * backgroup IO calls must call raise_barrier. Once that returns
632 * there is no normal IO happeing. It must arrange to call
633 * lower_barrier when the particular background IO completes.
634 */
637 {
641 /* Wait until no block IO is waiting (unless 'force') */
645 /* block any new IO from starting */
648 /* Now wait for all pending IO to complete */
654 }
657 {
663 }
666 {
672 );
674 }
677 }
680 {
686 }
689 {
690 /* stop syncio and normal IO and wait for everything to
691 * go quiet.
692 * We increment barrier and nr_waiting, and then
693 * wait until nr_pending match nr_queued+1
694 * This is called in the context of one normal IO request
695 * that has failed. Thus any sync request that might be pending
696 * will be blocked by nr_pending, and we need to wait for
697 * pending IO requests to complete or be queued for re-try.
698 * Thus the number queued (nr_queued) plus this request (1)
699 * must match the number of pending IOs (nr_pending) before
700 * we continue.
701 */
711 }
714 {
715 /* reverse the effect of the freeze */
721 }
724 {
741 }
743 /* If this request crosses a chunk boundary, we need to
744 * split it. This will only happen for 1 PAGE (or less) requests.
745 */
750 /* Sanity check -- queue functions should prevent this happening */
754 /* This is a one page bio that upper layers
755 * refuse to split for us, so we need to split it.
756 */
760 /* Each of these 'make_request' calls will call 'wait_barrier'.
761 * If the first succeeds but the second blocks due to the resync
762 * thread raising the barrier, we will deadlock because the
763 * IO to the underlying device will be queued in generic_make_request
764 * and will never complete, so will never reduce nr_pending.
765 * So increment nr_waiting here so no new raise_barriers will
766 * succeed, and so the second wait_barrier cannot block.
767 */
784 bad_map:
791 }
795 /*
796 * Register the new request and wait if the reconstruction
797 * thread has put up a bar for new requests.
798 * Continue immediately if no resync is active currently.
799 */
812 /*
813 * read balancing logic:
814 */
820 }
836 }
838 /*
839 * WRITE:
840 */
841 /* first select target devices under rcu_lock and
842 * inc refcount on their rdev. Record them by setting
843 * bios[x] to bio
844 */
848 retry_write:
858 }
865 }
866 }
870 /* Have to wait for this device to get unblocked, then retry */
878 }
883 }
908 }
911 /* This matches the end of raid10_end_write_request() */
918 }
920 /* In case raid10d snuck in to freeze_array */
926 }
929 {
940 else
942 }
950 }
953 {
957 /*
958 * If it is not operational, then we have already marked it as dead
959 * else if it is the last working disks, ignore the error, let the
960 * next level up know.
961 * else mark the drive as failed
962 */
965 /*
966 * Don't fail the drive, just return an IO error.
967 * The test should really be more sophisticated than
968 * "working_disks == 1", but it isn't critical, and
969 * can wait until we do more sophisticated "is the drive
970 * really dead" tests...
971 */
978 /*
979 * if recovery is running, make sure it aborts.
980 */
982 }
990 }
993 {
1001 }
1013 }
1014 }
1017 {
1023 }
1025 /* check if there are enough drives for
1026 * every block to appear on atleast one
1027 */
1029 {
1037 cnt++;
1039 }
1044 }
1047 {
1054 /*
1055 * Find all non-in_sync disks within the RAID10 configuration
1056 * and mark them in_sync
1057 */
1063 count++;
1065 }
1066 }
1073 }
1077 {
1086 /* only hot-add to in-sync arrays, as recovery is
1087 * very different from resync
1088 */
1100 else
1107 /* as we don't honour merge_bvec_fn, we must
1108 * never risk violating it, so limit
1109 * ->max_segments to one lying with a single
1110 * page, as a one page request is never in
1111 * violation.
1112 */
1117 }
1126 }
1131 }
1134 {
1147 }
1148 /* Only remove faulty devices in recovery
1149 * is not possible.
1150 */
1155 }
1159 /* lost the race, try later */
1163 }
1165 }
1166 abort:
1170 }
1174 {
1194 }
1196 /* for reconstruct, we always reschedule after a read.
1197 * for resync, only after all reads
1198 */
1202 /* we have read all the blocks,
1203 * do the comparison in process context in raid10d
1204 */
1206 }
1207 }
1210 {
1230 /* the primary of several recovery bios */
1239 }
1240 }
1241 }
1243 /*
1244 * Note: sync and recover and handled very differently for raid10
1245 * This code is for resync.
1246 * For resync, we read through virtual addresses and read all blocks.
1247 * If there is any error, we schedule a write. The lowest numbered
1248 * drive is authoritative.
1249 * However requests come for physical address, so we need to map.
1250 * For every physical address there are raid_disks/copies virtual addresses,
1251 * which is always are least one, but is not necessarly an integer.
1252 * This means that a physical address can span multiple chunks, so we may
1253 * have to submit multiple io requests for a single sync request.
1254 */
1255 /*
1256 * We check if all blocks are in-sync and only write to blocks that
1257 * aren't in sync
1258 */
1260 {
1267 /* find the first device with a block */
1278 /* now find blocks with errors */
1290 /* We know that the bi_io_vec layout is the same for
1291 * both 'first' and 'i', so we just compare them.
1292 * All vec entries are PAGE_SIZE;
1293 */
1297 PAGE_SIZE))
1302 }
1304 /* Don't fix anything. */
1306 /* Ok, we need to write this bio
1307 * First we need to fixup bv_offset, bv_len and
1308 * bi_vecs, as the read request might have corrupted these
1309 */
1327 PAGE_SIZE);
1328 }
1339 }
1341 done:
1345 }
1346 }
1348 /*
1349 * Now for the recovery code.
1350 * Recovery happens across physical sectors.
1351 * We recover all non-is_sync drives by finding the virtual address of
1352 * each, and then choose a working drive that also has that virt address.
1353 * There is a separate r10_bio for each non-in_sync drive.
1354 * Only the first two slots are in use. The first for reading,
1355 * The second for writing.
1356 *
1357 */
1360 {
1366 /* move the pages across to the second bio
1367 * and submit the write request
1368 */
1375 }
1382 else
1384 }
1387 /*
1388 * Used by fix_read_error() to decay the per rdev read_errors.
1389 * We halve the read error count for every hour that has elapsed
1390 * since the last recorded read error.
1391 *
1392 */
1394 {
1403 /* first time we've seen a read error */
1406 }
1413 /*
1414 * if hours_since_last is > the number of bits in read_errors
1415 * just set read errors to 0. We do this to avoid
1416 * overflowing the shift of read_errors by hours_since_last.
1417 */
1420 else
1422 }
1424 /*
1425 * This is a kernel thread which:
1426 *
1427 * 1. Retries failed read operations on working mirrors.
1428 * 2. Updates the raid superblock when problems encounter.
1429 * 3. Performs writes following reads for array synchronising.
1430 */
1433 {
1440 /* still own a reference to this rdev, so it cannot
1441 * have been cleared recently.
1442 */
1446 /* drive has already been failed, just ignore any
1447 more fix_read_error() attempts */
1457 "md/raid10:%s: %s: Raid device exceeded "
1466 }
1487 sect,
1494 }
1495 sl++;
1502 /* Cannot read from anywhere -- bye bye array */
1506 }
1509 /* write it back and re-read */
1516 sl--;
1526 sect,
1529 /* Well, this device is dead */
1531 "md/raid10:%s: read correction "
1532 "write failed"
1543 }
1546 }
1547 }
1553 sl--;
1563 sect,
1566 /* Well, this device is dead */
1568 "md/raid10:%s: unable to read back "
1569 "corrected sectors"
1582 "md/raid10:%s: read error corrected"
1588 }
1592 }
1593 }
1598 }
1599 }
1602 {
1623 }
1638 /* we got a read error. Maybe the drive is bad. Maybe just
1639 * the block and we can fix it.
1640 * We freeze all other IO, and try reading the block from
1641 * other devices. When we find one, we re-write
1642 * and check it that fixes the read error.
1643 * This is all done synchronously while the array is
1644 * frozen.
1645 */
1650 }
1686 }
1687 }
1689 }
1691 }
1695 {
1705 }
1707 /*
1708 * perform a "sync" on one "block"
1709 *
1710 * We need to make sure that no normal I/O request - particularly write
1711 * requests - conflict with active sync requests.
1712 *
1713 * This is achieved by tracking pending requests and a 'barrier' concept
1714 * that can be installed to exclude normal IO requests.
1715 *
1716 * Resync and recovery are handled very differently.
1717 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1718 *
1719 * For resync, we iterate over virtual addresses, read all copies,
1720 * and update if there are differences. If only one copy is live,
1721 * skip it.
1722 * For recovery, we iterate over physical addresses, read a good
1723 * value for each non-in_sync drive, and over-write.
1724 *
1725 * So, for recovery we may have several outstanding complex requests for a
1726 * given address, one for each out-of-sync device. We model this by allocating
1727 * a number of r10_bio structures, one for each out-of-sync device.
1728 * As we setup these structures, we collect all bio's together into a list
1729 * which we then process collectively to add pages, and then process again
1730 * to pass to generic_make_request.
1731 *
1732 * The r10_bio structures are linked using a borrowed master_bio pointer.
1733 * This link is counted in ->remaining. When the r10_bio that points to NULL
1734 * has its remaining count decremented to 0, the whole complex operation
1735 * is complete.
1736 *
1737 */
1741 {
1748 sector_t sync_blocks;
1757 skipped:
1762 /* If we aborted, we need to abort the
1763 * sync on the 'current' bitmap chucks (there can
1764 * be several when recovering multiple devices).
1765 * as we may have started syncing it but not finished.
1766 * We can find the current address in
1767 * mddev->curr_resync, but for recovery,
1768 * we need to convert that to several
1769 * virtual addresses.
1770 */
1776 sector_t sect =
1780 }
1788 }
1790 /* if there has been nothing to do on any drive,
1791 * then there is nothing to do at all..
1792 */
1795 }
1800 /* make sure whole request will fit in a chunk - if chunks
1801 * are meaningful
1802 */
1806 /*
1807 * If there is non-resync activity waiting for us then
1808 * put in a delay to throttle resync.
1809 */
1813 /* Again, very different code for resync and recovery.
1814 * Both must result in an r10bio with a list of bios that
1815 * have bi_end_io, bi_sector, bi_bdev set,
1816 * and bi_private set to the r10bio.
1817 * For recovery, we may actually create several r10bios
1818 * with 2 bios in each, that correspond to the bios in the main one.
1819 * In this case, the subordinate r10bios link back through a
1820 * borrowed master_bio pointer, and the counter in the master
1821 * includes a ref from each subordinate.
1822 */
1823 /* First, we decide what to do and set ->bi_end_io
1824 * To end_sync_read if we want to read, and
1825 * end_sync_write if we will want to write.
1826 */
1830 /* recovery... the complicated one */
1837 sector_t sect;
1845 /* want to reconstruct this device */
1848 /* Unless we are doing a full sync, we only need
1849 * to recover the block if it is set in the bitmap
1850 */
1857 /* yep, skip the sync_blocks here, but don't assume
1858 * that there will never be anything to do here
1859 */
1862 }
1877 /* Need to check if the array will still be
1878 * degraded
1879 */
1885 }
1895 /* This is where we read from */
1907 /* and we write to 'i' */
1927 }
1929 /* Cannot recover, so abort the recovery */
1940 }
1941 }
1948 }
1950 }
1952 /* resync. Schedule a read for every block at this virt offset */
1961 /* We can skip this block */
1964 }
1998 count++;
1999 }
2006 mddev);
2007 }
2011 }
2012 }
2023 }
2041 /* stop here */
2046 /* remove last page from this bio */
2050 }
2052 }
2056 bio_full:
2070 }
2071 }
2074 /* pretend they weren't skipped, it makes
2075 * no important difference in this case
2076 */
2080 giveup:
2081 /* There is nowhere to write, so all non-sync
2082 * drives must be failed, so try the next chunk...
2083 */
2088 chunks_skipped ++;
2091 }
2093 static sector_t
2095 {
2096 sector_t size;
2110 }
2114 {
2126 }
2137 }
2145 GFP_KERNEL);
2171 /* 'size' is now the number of chunks in the array */
2172 /* calculate "used chunks per device" in 'stride' */
2175 /* We need to round up when dividing by raid_disks to
2176 * get the stride size.
2177 */
2185 else
2203 out:
2212 }
2214 }
2217 {
2222 sector_t size;
2224 /*
2225 * copy the already verified devices into our private RAID10
2226 * bookkeeping area. [whatever we allocate in run(),
2227 * should be freed in stop()]
2228 */
2235 }
2247 else
2261 /* as we don't honour merge_bvec_fn, we must never risk
2262 * violating it, so limit max_segments to 1 lying
2263 * within a single page.
2264 */
2269 }
2272 }
2273 /* need to check that every block has at least one working mirror */
2278 }
2291 }
2292 }
2302 /*
2303 * Ok, everything is just fine now
2304 */
2313 /* Calculate max read-ahead size.
2314 * We need to readahead at least twice a whole stripe....
2315 * maybe...
2316 */
2317 {
2323 }
2333 out_free_conf:
2341 out:
2343 }
2346 {
2360 }
2363 {
2373 }
2374 }
2377 {
2385 }
2387 /* Set new parameters */
2389 /* new layout: far_copies = 1, near_copies = 2 */
2394 /* make sure it will be not marked as dirty */
2403 }
2406 }
2409 {
2412 /* raid10 can take over:
2413 * raid0 - providing it has only two drives
2414 */
2416 /* for raid0 takeover only one zone is supported */
2423 }
2425 }
2427 }
2430 {
2446 };
2449 {
2451 }
2454 {
2456 }