| blob | 8ea0acad606bad395f8680d22a5a0de96b583986 |
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 */
275 /*
276 * oops, read error:
277 */
284 }
287 }
290 {
301 /*
302 * this branch is our 'one mirror IO has finished' event handler:
303 */
306 /* an I/O failed, we can't clear the bitmap */
309 /*
310 * Set R10BIO_Uptodate in our master bio, so that
311 * we will return a good error code for to the higher
312 * levels even if IO on some other mirrored buffer fails.
313 *
314 * The 'master' represents the composite IO operation to
315 * user-side. So if something waits for IO, then it will
316 * wait for the 'master' bio.
317 */
322 /*
323 *
324 * Let's see if all mirrored write operations have finished
325 * already.
326 */
328 /* clear the bitmap if all writes complete successfully */
335 }
338 }
341 /*
342 * RAID10 layout manager
343 * As well 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 laid 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.
361 *
362 * raid10_find_virt does the reverse mapping, from a device and a
363 * sector offset to a virtual address
364 */
367 {
369 sector_t sector;
370 sector_t chunk;
371 sector_t stripe;
376 /* now calculate first sector/dev */
388 /* and calculate all the others */
394 slot++;
403 slot++;
404 }
405 dev++;
409 }
410 }
412 }
415 {
431 else
433 }
435 }
439 }
441 /**
442 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
443 * @q: request queue
444 * @bvm: properties of new bio
445 * @biovec: the request that could be merged to it.
446 *
447 * Return amount of bytes we can accept at this offset
448 * If near_copies == raid_disk, there are no striping issues,
449 * but in that case, the function isn't called at all.
450 */
454 {
465 else
467 }
469 /*
470 * This routine returns the disk from which the requested read should
471 * be done. There is a per-array 'next expected sequential IO' sector
472 * number - if this matches on the next IO then we use the last disk.
473 * There is also a per-disk 'last know head position' sector that is
474 * maintained from IRQ contexts, both the normal and the resync IO
475 * completion handlers update this position correctly. If there is no
476 * perfect sequential match then we pick the disk whose head is closest.
477 *
478 * If there are 2 mirrors in the same 2 devices, performance degrades
479 * because position is mirror, not device based.
480 *
481 * The rdev for the device selected will have nr_pending incremented.
482 */
484 /*
485 * FIXME: possibly should rethink readbalancing and do it differently
486 * depending on near_copies / far_copies geometry.
487 */
489 {
500 retry:
504 /*
505 * Check if we can balance. We can balance on the whole
506 * device if no resync is going on (recovery is ok), or below
507 * the resync window. We take the first readable disk when
508 * above the resync window.
509 */
527 /* This optimisation is debatable, and completely destroys
528 * sequential read speed for 'far copies' arrays. So only
529 * keep it for 'near' arrays, and review those later.
530 */
534 /* for far > 1 always use the lowest address */
537 else
543 }
544 }
555 /* Cannot risk returning a device that failed
556 * before we inc'ed nr_pending
557 */
560 }
567 }
570 {
584 }
585 }
588 }
591 {
592 /* Any writes that have been queued but are awaiting
593 * bitmap updates get flushed here.
594 */
601 /* flush any pending bitmap writes to disk
602 * before proceeding w/ I/O */
610 }
613 }
615 /* Barriers....
616 * Sometimes we need to suspend IO while we do something else,
617 * either some resync/recovery, or reconfigure the array.
618 * To do this we raise a 'barrier'.
619 * The 'barrier' is a counter that can be raised multiple times
620 * to count how many activities are happening which preclude
621 * normal IO.
622 * We can only raise the barrier if there is no pending IO.
623 * i.e. if nr_pending == 0.
624 * We choose only to raise the barrier if no-one is waiting for the
625 * barrier to go down. This means that as soon as an IO request
626 * is ready, no other operations which require a barrier will start
627 * until the IO request has had a chance.
628 *
629 * So: regular IO calls 'wait_barrier'. When that returns there
630 * is no backgroup IO happening, It must arrange to call
631 * allow_barrier when it has finished its IO.
632 * backgroup IO calls must call raise_barrier. Once that returns
633 * there is no normal IO happeing. It must arrange to call
634 * lower_barrier when the particular background IO completes.
635 */
638 {
642 /* Wait until no block IO is waiting (unless 'force') */
646 /* block any new IO from starting */
649 /* Now wait for all pending IO to complete */
655 }
658 {
664 }
667 {
673 );
675 }
678 }
681 {
687 }
690 {
691 /* stop syncio and normal IO and wait for everything to
692 * go quiet.
693 * We increment barrier and nr_waiting, and then
694 * wait until nr_pending match nr_queued+1
695 * This is called in the context of one normal IO request
696 * that has failed. Thus any sync request that might be pending
697 * will be blocked by nr_pending, and we need to wait for
698 * pending IO requests to complete or be queued for re-try.
699 * Thus the number queued (nr_queued) plus this request (1)
700 * must match the number of pending IOs (nr_pending) before
701 * we continue.
702 */
712 }
715 {
716 /* reverse the effect of the freeze */
722 }
725 {
742 }
744 /* If this request crosses a chunk boundary, we need to
745 * split it. This will only happen for 1 PAGE (or less) requests.
746 */
751 /* Sanity check -- queue functions should prevent this happening */
755 /* This is a one page bio that upper layers
756 * refuse to split for us, so we need to split it.
757 */
761 /* Each of these 'make_request' calls will call 'wait_barrier'.
762 * If the first succeeds but the second blocks due to the resync
763 * thread raising the barrier, we will deadlock because the
764 * IO to the underlying device will be queued in generic_make_request
765 * and will never complete, so will never reduce nr_pending.
766 * So increment nr_waiting here so no new raise_barriers will
767 * succeed, and so the second wait_barrier cannot block.
768 */
785 bad_map:
792 }
796 /*
797 * Register the new request and wait if the reconstruction
798 * thread has put up a bar for new requests.
799 * Continue immediately if no resync is active currently.
800 */
813 /*
814 * read balancing logic:
815 */
821 }
837 }
839 /*
840 * WRITE:
841 */
842 /* first select target devices under rcu_lock and
843 * inc refcount on their rdev. Record them by setting
844 * bios[x] to bio
845 */
849 retry_write:
859 }
866 }
867 }
871 /* Have to wait for this device to get unblocked, then retry */
879 }
884 }
909 }
912 /* This matches the end of raid10_end_write_request() */
919 }
921 /* In case raid10d snuck in to freeze_array */
927 }
930 {
941 else
943 }
951 }
954 {
958 /*
959 * If it is not operational, then we have already marked it as dead
960 * else if it is the last working disks, ignore the error, let the
961 * next level up know.
962 * else mark the drive as failed
963 */
966 /*
967 * Don't fail the drive, just return an IO error.
968 * The test should really be more sophisticated than
969 * "working_disks == 1", but it isn't critical, and
970 * can wait until we do more sophisticated "is the drive
971 * really dead" tests...
972 */
979 /*
980 * if recovery is running, make sure it aborts.
981 */
983 }
991 }
994 {
1002 }
1014 }
1015 }
1018 {
1024 }
1026 /* check if there are enough drives for
1027 * every block to appear on atleast one
1028 */
1030 {
1038 cnt++;
1040 }
1045 }
1048 {
1055 /*
1056 * Find all non-in_sync disks within the RAID10 configuration
1057 * and mark them in_sync
1058 */
1064 count++;
1066 }
1067 }
1074 }
1078 {
1087 /* only hot-add to in-sync arrays, as recovery is
1088 * very different from resync
1089 */
1101 else
1108 /* as we don't honour merge_bvec_fn, we must
1109 * never risk violating it, so limit
1110 * ->max_segments to one lying with a single
1111 * page, as a one page request is never in
1112 * violation.
1113 */
1118 }
1127 }
1132 }
1135 {
1148 }
1149 /* Only remove faulty devices in recovery
1150 * is not possible.
1151 */
1156 }
1160 /* lost the race, try later */
1164 }
1166 }
1167 abort:
1171 }
1175 {
1195 }
1197 /* for reconstruct, we always reschedule after a read.
1198 * for resync, only after all reads
1199 */
1203 /* we have read all the blocks,
1204 * do the comparison in process context in raid10d
1205 */
1207 }
1208 }
1211 {
1231 /* the primary of several recovery bios */
1240 }
1241 }
1242 }
1244 /*
1245 * Note: sync and recover and handled very differently for raid10
1246 * This code is for resync.
1247 * For resync, we read through virtual addresses and read all blocks.
1248 * If there is any error, we schedule a write. The lowest numbered
1249 * drive is authoritative.
1250 * However requests come for physical address, so we need to map.
1251 * For every physical address there are raid_disks/copies virtual addresses,
1252 * which is always are least one, but is not necessarly an integer.
1253 * This means that a physical address can span multiple chunks, so we may
1254 * have to submit multiple io requests for a single sync request.
1255 */
1256 /*
1257 * We check if all blocks are in-sync and only write to blocks that
1258 * aren't in sync
1259 */
1261 {
1268 /* find the first device with a block */
1279 /* now find blocks with errors */
1291 /* We know that the bi_io_vec layout is the same for
1292 * both 'first' and 'i', so we just compare them.
1293 * All vec entries are PAGE_SIZE;
1294 */
1298 PAGE_SIZE))
1303 }
1305 /* Don't fix anything. */
1307 /* Ok, we need to write this bio
1308 * First we need to fixup bv_offset, bv_len and
1309 * bi_vecs, as the read request might have corrupted these
1310 */
1328 PAGE_SIZE);
1329 }
1340 }
1342 done:
1346 }
1347 }
1349 /*
1350 * Now for the recovery code.
1351 * Recovery happens across physical sectors.
1352 * We recover all non-is_sync drives by finding the virtual address of
1353 * each, and then choose a working drive that also has that virt address.
1354 * There is a separate r10_bio for each non-in_sync drive.
1355 * Only the first two slots are in use. The first for reading,
1356 * The second for writing.
1357 *
1358 */
1361 {
1367 /* move the pages across to the second bio
1368 * and submit the write request
1369 */
1376 }
1383 else
1385 }
1388 /*
1389 * Used by fix_read_error() to decay the per rdev read_errors.
1390 * We halve the read error count for every hour that has elapsed
1391 * since the last recorded read error.
1392 *
1393 */
1395 {
1404 /* first time we've seen a read error */
1407 }
1414 /*
1415 * if hours_since_last is > the number of bits in read_errors
1416 * just set read errors to 0. We do this to avoid
1417 * overflowing the shift of read_errors by hours_since_last.
1418 */
1421 else
1423 }
1425 /*
1426 * This is a kernel thread which:
1427 *
1428 * 1. Retries failed read operations on working mirrors.
1429 * 2. Updates the raid superblock when problems encounter.
1430 * 3. Performs writes following reads for array synchronising.
1431 */
1434 {
1451 /* drive has already been failed, just ignore any
1452 more fix_read_error() attempts */
1454 }
1462 "md/raid10:%s: %s: Raid device exceeded "
1463 "read_error threshold "
1468 "md/raid10:%s: %s: Failing raid "
1472 }
1473 }
1495 sect,
1502 }
1503 sl++;
1510 /* Cannot read from anywhere -- bye bye array */
1514 }
1517 /* write it back and re-read */
1524 sl--;
1534 sect,
1537 /* Well, this device is dead */
1539 "md/raid10:%s: read correction "
1540 "write failed"
1551 }
1554 }
1555 }
1561 sl--;
1571 sect,
1574 /* Well, this device is dead */
1576 "md/raid10:%s: unable to read back "
1577 "corrected sectors"
1590 "md/raid10:%s: read error corrected"
1596 }
1600 }
1601 }
1606 }
1607 }
1610 {
1631 }
1645 /* we got a read error. Maybe the drive is bad. Maybe just
1646 * the block and we can fix it.
1647 * We freeze all other IO, and try reading the block from
1648 * other devices. When we find one, we re-write
1649 * and check it that fixes the read error.
1650 * This is all done synchronously while the array is
1651 * frozen.
1652 */
1657 }
1691 }
1692 }
1694 }
1696 }
1700 {
1710 }
1712 /*
1713 * perform a "sync" on one "block"
1714 *
1715 * We need to make sure that no normal I/O request - particularly write
1716 * requests - conflict with active sync requests.
1717 *
1718 * This is achieved by tracking pending requests and a 'barrier' concept
1719 * that can be installed to exclude normal IO requests.
1720 *
1721 * Resync and recovery are handled very differently.
1722 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1723 *
1724 * For resync, we iterate over virtual addresses, read all copies,
1725 * and update if there are differences. If only one copy is live,
1726 * skip it.
1727 * For recovery, we iterate over physical addresses, read a good
1728 * value for each non-in_sync drive, and over-write.
1729 *
1730 * So, for recovery we may have several outstanding complex requests for a
1731 * given address, one for each out-of-sync device. We model this by allocating
1732 * a number of r10_bio structures, one for each out-of-sync device.
1733 * As we setup these structures, we collect all bio's together into a list
1734 * which we then process collectively to add pages, and then process again
1735 * to pass to generic_make_request.
1736 *
1737 * The r10_bio structures are linked using a borrowed master_bio pointer.
1738 * This link is counted in ->remaining. When the r10_bio that points to NULL
1739 * has its remaining count decremented to 0, the whole complex operation
1740 * is complete.
1741 *
1742 */
1745 {
1753 sector_t sync_blocks;
1762 skipped:
1767 /* If we aborted, we need to abort the
1768 * sync on the 'current' bitmap chucks (there can
1769 * be several when recovering multiple devices).
1770 * as we may have started syncing it but not finished.
1771 * We can find the current address in
1772 * mddev->curr_resync, but for recovery,
1773 * we need to convert that to several
1774 * virtual addresses.
1775 */
1781 sector_t sect =
1785 }
1793 }
1795 /* if there has been nothing to do on any drive,
1796 * then there is nothing to do at all..
1797 */
1800 }
1805 /* make sure whole request will fit in a chunk - if chunks
1806 * are meaningful
1807 */
1811 /*
1812 * If there is non-resync activity waiting for us then
1813 * put in a delay to throttle resync.
1814 */
1818 /* Again, very different code for resync and recovery.
1819 * Both must result in an r10bio with a list of bios that
1820 * have bi_end_io, bi_sector, bi_bdev set,
1821 * and bi_private set to the r10bio.
1822 * For recovery, we may actually create several r10bios
1823 * with 2 bios in each, that correspond to the bios in the main one.
1824 * In this case, the subordinate r10bios link back through a
1825 * borrowed master_bio pointer, and the counter in the master
1826 * includes a ref from each subordinate.
1827 */
1828 /* First, we decide what to do and set ->bi_end_io
1829 * To end_sync_read if we want to read, and
1830 * end_sync_write if we will want to write.
1831 */
1835 /* recovery... the complicated one */
1843 /* want to reconstruct this device */
1847 /* Unless we are doing a full sync, we only need
1848 * to recover the block if it is set in the bitmap
1849 */
1856 /* yep, skip the sync_blocks here, but don't assume
1857 * that there will never be anything to do here
1858 */
1861 }
1876 /* Need to check if the array will still be
1877 * degraded
1878 */
1884 }
1893 /* This is where we read from */
1905 /* and we write to 'i' */
1925 }
1926 }
1928 /* Cannot recover, so abort the recovery */
1939 }
1940 }
1947 }
1949 }
1951 /* resync. Schedule a read for every block at this virt offset */
1959 /* We can skip this block */
1962 }
1996 count++;
1997 }
2004 }
2008 }
2009 }
2020 }
2036 /* stop here */
2040 /* remove last page from this bio */
2044 }
2046 }
2048 }
2052 bio_full:
2066 }
2067 }
2070 /* pretend they weren't skipped, it makes
2071 * no important difference in this case
2072 */
2076 giveup:
2077 /* There is nowhere to write, so all non-sync
2078 * drives must be failed, so try the next chunk...
2079 */
2084 chunks_skipped ++;
2087 }
2089 static sector_t
2091 {
2092 sector_t size;
2106 }
2110 {
2122 }
2133 }
2141 GFP_KERNEL);
2167 /* 'size' is now the number of chunks in the array */
2168 /* calculate "used chunks per device" in 'stride' */
2171 /* We need to round up when dividing by raid_disks to
2172 * get the stride size.
2173 */
2181 else
2199 out:
2208 }
2210 }
2213 {
2218 sector_t size;
2220 /*
2221 * copy the already verified devices into our private RAID10
2222 * bookkeeping area. [whatever we allocate in run(),
2223 * should be freed in stop()]
2224 */
2231 }
2243 else
2257 /* as we don't honour merge_bvec_fn, we must never risk
2258 * violating it, so limit max_segments to 1 lying
2259 * within a single page.
2260 */
2265 }
2268 }
2269 /* need to check that every block has at least one working mirror */
2274 }
2287 }
2288 }
2298 /*
2299 * Ok, everything is just fine now
2300 */
2309 /* Calculate max read-ahead size.
2310 * We need to readahead at least twice a whole stripe....
2311 * maybe...
2312 */
2313 {
2319 }
2329 out_free_conf:
2337 out:
2339 }
2342 {
2357 }
2360 {
2370 }
2371 }
2374 {
2382 }
2384 /* Set new parameters */
2386 /* new layout: far_copies = 1, near_copies = 2 */
2391 /* make sure it will be not marked as dirty */
2400 }
2403 }
2406 {
2409 /* raid10 can take over:
2410 * raid0 - providing it has only two drives
2411 */
2413 /* for raid0 takeover only one zone is supported */
2420 }
2422 }
2424 }
2427 {
2443 };
2446 {
2448 }
2451 {
2453 }