| blob | 7d9e071f2304aa57ef4a68fae878b4af88d046cb |
1 /*
2 * raid1.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5 *
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7 *
8 * RAID-1 management functions.
9 *
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11 *
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
27 * any later version.
28 *
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32 */
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
44 /*
45 * Number of guaranteed r1bios in case of extreme VM load:
46 */
47 #define NR_RAID1_BIOS 256
49 /* When there are this many requests queue to be written by
50 * the raid1 thread, we become 'congested' to provide back-pressure
51 * for writeback.
52 */
59 {
63 /* allocate a r1bio with room for raid_disks entries in the bios array */
65 }
68 {
70 }
72 #define RESYNC_BLOCK_SIZE (64*1024)
73 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
74 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
75 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
76 #define RESYNC_WINDOW (2048*1024)
79 {
90 /*
91 * Allocate bios : 1 for reading, n-1 for writing
92 */
98 }
99 /*
100 * Allocate RESYNC_PAGES data pages and attach them to
101 * the first bio.
102 * If this is a user-requested check/repair, allocate
103 * RESYNC_PAGES for each bio.
104 */
107 else
118 }
119 }
120 /* If not user-requests, copy the page pointers to all bios */
126 }
132 out_free_pages:
137 out_free_bio:
142 }
145 {
156 }
161 }
164 {
172 }
173 }
176 {
181 }
184 {
192 }
197 }
200 {
212 }
214 /*
215 * raid_end_bio_io() is called when we have finished servicing a mirrored
216 * operation and are ready to return a success/failure code to the buffer
217 * cache layer.
218 */
220 {
238 /*
239 * Wake up any possible resync thread that waits for the device
240 * to go idle.
241 */
243 }
244 }
247 {
250 /* if nobody has done the final endio yet, do it now */
259 }
261 }
263 /*
264 * Update disk head position estimator based on IRQ completion info.
265 */
267 {
272 }
274 /*
275 * Find the disk number which triggered given bio
276 */
278 {
290 }
293 {
300 /*
301 * this branch is our 'one mirror IO has finished' event handler:
302 */
308 /* If all other devices have failed, we want to return
309 * the error upwards rather than fail the last device.
310 * Here we redefine "uptodate" to mean "Don't want to retry"
311 */
319 }
324 /*
325 * oops, read error:
326 */
329 KERN_ERR "md/raid1:%s: %s: "
333 b),
337 }
340 }
343 {
344 /* it really is the end of this request */
346 /* free extra copy of the data pages */
352 }
353 /* clear the bitmap if all writes complete successfully */
359 }
362 {
372 else
374 }
375 }
378 {
387 /*
388 * 'one mirror IO has finished' event handler:
389 */
395 /*
396 * Set R1BIO_Uptodate in our master bio, so that we
397 * will return a good error code for to the higher
398 * levels even if IO on some other mirrored buffer
399 * fails.
400 *
401 * The 'master' represents the composite IO operation
402 * to user-side. So if something waits for IO, then it
403 * will wait for the 'master' bio.
404 */
405 sector_t first_bad;
412 /* Maybe we can clear some bad blocks. */
418 }
419 }
425 /*
426 * In behind mode, we ACK the master bio once the I/O
427 * has safely reached all non-writemostly
428 * disks. Setting the Returned bit ensures that this
429 * gets done only once -- we don't ever want to return
430 * -EIO here, instead we'll wait
431 */
434 /* Maybe we can return now */
443 }
444 }
445 }
450 /*
451 * Let's see if all mirrored write operations have finished
452 * already.
453 */
458 }
461 /*
462 * This routine returns the disk from which the requested read should
463 * be done. There is a per-array 'next expected sequential IO' sector
464 * number - if this matches on the next IO then we use the last disk.
465 * There is also a per-disk 'last know head position' sector that is
466 * maintained from IRQ contexts, both the normal and the resync IO
467 * completion handlers update this position correctly. If there is no
468 * perfect sequential match then we pick the disk whose head is closest.
469 *
470 * If there are 2 mirrors in the same 2 devices, performance degrades
471 * because position is mirror, not device based.
472 *
473 * The rdev for the device selected will have nr_pending incremented.
474 */
476 {
483 sector_t best_dist;
488 /*
489 * Check if we can balance. We can balance on the whole
490 * device if no resync is going on, or below the resync window.
491 * We take the first readable disk when above the resync window.
492 */
493 retry:
506 }
509 sector_t dist;
510 sector_t first_bad;
526 /* Don't balance among write-mostly, just
527 * use the first as a last resort */
532 /* Cannot use this */
538 }
540 }
541 /* This is a reasonable device to use. It might
542 * even be best.
543 */
547 /* already have a better device */
550 /* cannot read here. If this is the 'primary'
551 * device, then we must not read beyond
552 * bad_sectors from another device..
553 */
565 }
568 }
575 /* Don't change to another disk for sequential reads */
578 /* If device is idle, use it */
582 }
586 }
587 }
595 /* cannot risk returning a device that failed
596 * before we inc'ed nr_pending
597 */
600 }
604 }
609 }
612 {
628 /* Note the '|| 1' - when read_balance prefers
629 * non-congested targets, it can be removed
630 */
633 else
635 }
636 }
639 }
643 {
648 }
651 {
652 /* Any writes that have been queued but are awaiting
653 * bitmap updates get flushed here.
654 */
662 /* flush any pending bitmap writes to
663 * disk before proceeding w/ I/O */
672 }
675 }
677 /* Barriers....
678 * Sometimes we need to suspend IO while we do something else,
679 * either some resync/recovery, or reconfigure the array.
680 * To do this we raise a 'barrier'.
681 * The 'barrier' is a counter that can be raised multiple times
682 * to count how many activities are happening which preclude
683 * normal IO.
684 * We can only raise the barrier if there is no pending IO.
685 * i.e. if nr_pending == 0.
686 * We choose only to raise the barrier if no-one is waiting for the
687 * barrier to go down. This means that as soon as an IO request
688 * is ready, no other operations which require a barrier will start
689 * until the IO request has had a chance.
690 *
691 * So: regular IO calls 'wait_barrier'. When that returns there
692 * is no backgroup IO happening, It must arrange to call
693 * allow_barrier when it has finished its IO.
694 * backgroup IO calls must call raise_barrier. Once that returns
695 * there is no normal IO happeing. It must arrange to call
696 * lower_barrier when the particular background IO completes.
697 */
698 #define RESYNC_DEPTH 32
701 {
704 /* Wait until no block IO is waiting */
708 /* block any new IO from starting */
711 /* Now wait for all pending IO to complete */
717 }
720 {
727 }
730 {
736 );
738 }
741 }
744 {
750 }
753 {
754 /* stop syncio and normal IO and wait for everything to
755 * go quite.
756 * We increment barrier and nr_waiting, and then
757 * wait until nr_pending match nr_queued+1
758 * This is called in the context of one normal IO request
759 * that has failed. Thus any sync request that might be pending
760 * will be blocked by nr_pending, and we need to wait for
761 * pending IO requests to complete or be queued for re-try.
762 * Thus the number queued (nr_queued) plus this request (1)
763 * must match the number of pending IOs (nr_pending) before
764 * we continue.
765 */
774 }
776 {
777 /* reverse the effect of the freeze */
783 }
786 /* duplicate the data pages for behind I/O
787 */
789 {
793 GFP_NOIO);
806 }
812 do_sync_io:
818 }
821 {
838 /*
839 * Register the new request and wait if the reconstruction
840 * thread has put up a bar for new requests.
841 * Continue immediately if no resync is active currently.
842 */
849 /* As the suspend_* range is controlled by
850 * userspace, we want an interruptible
851 * wait.
852 */
862 }
864 }
870 /*
871 * make_request() can abort the operation when READA is being
872 * used and no empty request is available.
873 *
874 */
883 /* We might need to issue multiple reads to different
884 * devices if there are bad blocks around, so we keep
885 * track of the number of reads in bio->bi_phys_segments.
886 * If this is 0, there is only one r1_bio and no locking
887 * will be needed when requests complete. If it is
888 * non-zero, then it is the number of not-completed requests.
889 */
894 /*
895 * read balancing logic:
896 */
899 read_again:
903 /* couldn't find anywhere to read from */
906 }
910 bitmap) {
911 /* Reading from a write-mostly device must
912 * take care not to over-take any writes
913 * that are 'behind'
914 */
917 }
922 max_sectors);
933 /* could not read all from this device, so we will
934 * need another r1_bio.
935 */
943 else
946 /* Cannot call generic_make_request directly
947 * as that will be queued in __make_request
948 * and subsequent mempool_alloc might block waiting
949 * for it. So hand bio over to raid1d.
950 */
964 }
966 /*
967 * WRITE:
968 */
973 }
974 /* first select target devices under rcu_lock and
975 * inc refcount on their rdev. Record them by setting
976 * bios[x] to bio
977 * If there are known/acknowledged bad blocks on any device on
978 * which we have seen a write error, we want to avoid writing those
979 * blocks.
980 * This potentially requires several writes to write around
981 * the bad blocks. Each set of writes gets it's own r1bio
982 * with a set of bios attached.
983 */
987 retry_write:
997 }
1002 }
1006 sector_t first_bad;
1011 max_sectors,
1014 /* mustn't write here until the bad block is
1015 * acknowledged*/
1019 }
1021 /* Cannot write here at all */
1024 /* mustn't write more than bad_sectors
1025 * to other devices yet
1026 */
1029 /* We don't set R1BIO_Degraded as that
1030 * only applies if the disk is
1031 * missing, so it might be re-added,
1032 * and we want to know to recover this
1033 * chunk.
1034 * In this case the device is here,
1035 * and the fact that this chunk is not
1036 * in-sync is recorded in the bad
1037 * block log
1038 */
1040 }
1045 }
1046 }
1048 }
1052 /* Wait for this device to become unblocked */
1063 }
1066 /* We are splitting this write into multiple parts, so
1067 * we need to prepare for allocating another r1_bio.
1068 */
1073 else
1076 }
1092 /* do behind I/O ?
1093 * Not if there are too many, or cannot
1094 * allocate memory, or a reader on WriteMostly
1095 * is waiting for behind writes to flush */
1107 }
1112 /* Yes, I really want the '__' version so that
1113 * we clear any unused pointer in the io_vec, rather
1114 * than leave them unchanged. This is important
1115 * because when we come to free the pages, we won't
1116 * know the original bi_idx, so we just free
1117 * them all
1118 */
1123 }
1139 }
1140 /* Mustn't call r1_bio_write_done before this next test,
1141 * as it could result in the bio being freed.
1142 */
1145 /* We need another r1_bio. It has already been counted
1146 * in bio->bi_phys_segments
1147 */
1155 }
1159 /* In case raid1d snuck in to freeze_array */
1164 }
1167 {
1178 }
1181 }
1185 {
1189 /*
1190 * If it is not operational, then we have already marked it as dead
1191 * else if it is the last working disks, ignore the error, let the
1192 * next level up know.
1193 * else mark the drive as failed
1194 */
1197 /*
1198 * Don't fail the drive, act as though we were just a
1199 * normal single drive.
1200 * However don't try a recovery from this drive as
1201 * it is very likely to fail.
1202 */
1205 }
1213 /*
1214 * if recovery is running, make sure it aborts.
1215 */
1225 }
1228 {
1235 }
1248 }
1250 }
1253 {
1259 }
1262 {
1268 /*
1269 * Find all failed disks within the RAID1 configuration
1270 * and mark them readable.
1271 * Called under mddev lock, so rcu protection not needed.
1272 */
1278 count++;
1280 }
1281 }
1288 }
1292 {
1311 /* as we don't honour merge_bvec_fn, we must
1312 * never risk violating it, so limit
1313 * ->max_segments to one lying with a single
1314 * page, as a one page request is never in
1315 * violation.
1316 */
1321 }
1326 /* As all devices are equivalent, we don't need a full recovery
1327 * if this was recently any drive of the array
1328 */
1333 }
1337 }
1340 {
1353 }
1354 /* Only remove non-faulty devices if recovery
1355 * is not possible.
1356 */
1362 }
1366 /* lost the race, try later */
1370 }
1372 }
1373 abort:
1377 }
1381 {
1386 /*
1387 * we have read a block, now it needs to be re-written,
1388 * or re-read if the read failed.
1389 * We don't do much here, just schedule handling by raid1d
1390 */
1396 }
1399 {
1405 sector_t first_bad;
1414 /* make sure these bits doesn't get cleared. */
1432 )
1443 }
1444 }
1445 }
1449 {
1451 /* success */
1455 /* need to record an error - either for the block or the device */
1459 }
1462 {
1463 /* Try some synchronous reads of other devices to get
1464 * good data, much like with normal read errors. Only
1465 * read into the pages we already have so we don't
1466 * need to re-issue the read request.
1467 * We don't need to freeze the array, because being in an
1468 * active sync request, there is no normal IO, and
1469 * no overlapping syncs.
1470 * We don't need to check is_badblock() again as we
1471 * made sure that anything with a bad block in range
1472 * will have bi_end_io clear.
1473 */
1492 /* No rcu protection needed here devices
1493 * can only be removed when no resync is
1494 * active, and resync is currently active
1495 */
1502 }
1503 }
1504 d++;
1512 /* Cannot read from anywhere, this block is lost.
1513 * Record a bad block on each device. If that doesn't
1514 * work just disable and interrupt the recovery.
1515 * Don't fail devices as that won't really help.
1516 */
1528 }
1536 }
1537 /* Try next page */
1540 idx++;
1542 }
1545 /* write it back and re-read */
1549 d--;
1558 }
1559 }
1564 d--;
1572 }
1575 idx ++;
1576 }
1580 }
1583 {
1584 /* We have read all readable devices. If we haven't
1585 * got the block, then there is no hope left.
1586 * If we have, then we want to do a comparison
1587 * and skip the write if everything is the same.
1588 * If any blocks failed to read, then we need to
1589 * attempt an over-write
1590 */
1602 }
1621 PAGE_SIZE))
1623 }
1630 /* No need to write to this device. */
1634 }
1635 /* fixup the bio for reuse */
1653 else
1658 PAGE_SIZE);
1659 }
1660 }
1662 }
1665 {
1674 /* ouch - failed to read all of that. */
1681 /*
1682 * schedule writes
1683 */
1699 }
1702 /* if we're here, all write(s) have completed, so clean up */
1705 }
1706 }
1708 /*
1709 * This is a kernel thread which:
1710 *
1711 * 1. Retries failed read operations on working mirrors.
1712 * 2. Updates the raid superblock when problems encounter.
1713 * 3. Performs writes following reads for array synchronising.
1714 */
1718 {
1731 /* Note: no rcu protection needed here
1732 * as this is synchronous in the raid1d thread
1733 * which is the thread that might remove
1734 * a device. If raid1d ever becomes multi-threaded....
1735 */
1736 sector_t first_bad;
1748 d++;
1751 }
1755 /* Cannot read from anywhere - mark it bad */
1760 }
1761 /* write it back and re-read */
1766 d--;
1772 }
1778 d--;
1786 "md/raid1:%s: read error corrected "
1792 }
1793 }
1794 }
1797 }
1798 }
1801 {
1803 }
1806 {
1817 }
1820 {
1827 /* bio has the data to be written to device 'i' where
1828 * we just recently had a write error.
1829 * We repeatedly clone the bio and trim down to one block,
1830 * then try the write. Where the write fails we record
1831 * a bad block.
1832 * It is conceivable that the bio doesn't exactly align with
1833 * blocks. We must handle this somehow.
1834 *
1835 * We currently own a reference on the rdev.
1836 */
1839 sector_t sector;
1858 idx++;
1863 }
1868 /* Write at 'sector' for 'sectors'*/
1882 /* failure! */
1891 }
1893 }
1896 {
1907 }
1912 }
1913 }
1916 }
1919 {
1929 /* This drive got a write error. We need to
1930 * narrow down and record precise write
1931 * errors.
1932 */
1936 /* an I/O failed, we can't clear the bitmap */
1938 }
1941 }
1945 }
1948 {
1957 /* we got a read error. Maybe the drive is bad. Maybe just
1958 * the block and we can fix it.
1959 * We freeze all other IO, and try reading the block from
1960 * other devices. When we find one, we re-write
1961 * and check it that fixes the read error.
1962 * This is all done synchronously while the array is
1963 * frozen
1964 */
1975 read_more:
1983 const unsigned long do_sync
1989 }
1996 "md/raid1:%s: redirecting sector %llu"
2007 /* Drat - have to split this up more */
2015 else
2034 }
2035 }
2038 {
2057 }
2069 else
2076 else
2077 /* just a partial read to be scheduled from separate
2078 * context
2079 */
2085 }
2087 }
2091 {
2102 }
2104 /*
2105 * perform a "sync" on one "block"
2106 *
2107 * We need to make sure that no normal I/O request - particularly write
2108 * requests - conflict with active sync requests.
2109 *
2110 * This is achieved by tracking pending requests and a 'barrier' concept
2111 * that can be installed to exclude normal IO requests.
2112 */
2114 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2115 {
2124 sector_t sync_blocks;
2135 /* If we aborted, we need to abort the
2136 * sync on the 'current' bitmap chunk (there will
2137 * only be one in raid1 resync.
2138 * We can find the current addess in mddev->curr_resync
2139 */
2149 }
2157 }
2158 /* before building a request, check if we can skip these blocks..
2159 * This call the bitmap_start_sync doesn't actually record anything
2160 */
2163 /* We can skip this block, and probably several more */
2166 }
2167 /*
2168 * If there is non-resync activity waiting for a turn,
2169 * and resync is going fast enough,
2170 * then let it though before starting on this new sync request.
2171 */
2182 /*
2183 * If we get a correctably read error during resync or recovery,
2184 * we might want to read from a different device. So we
2185 * flag all drives that could conceivably be read from for READ,
2186 * and any others (which will be non-In_sync devices) for WRITE.
2187 * If a read fails, we try reading from something else for which READ
2188 * is OK.
2189 */
2200 /* take from bio_init */
2219 write_targets ++;
2221 /* may need to read from here */
2234 }
2235 }
2243 }
2246 read_targets++;
2247 }
2248 }
2254 }
2255 }
2262 /* These sectors are bad on all InSync devices, so we
2263 * need to mark them bad on all write targets
2264 */
2273 }
2279 /* Cannot record the badblocks, so need to
2280 * abort the resync.
2281 * If there are multiple read targets, could just
2282 * fail the really bad ones ???
2283 */
2290 }
2292 /* only resync enough to reach the next bad->good
2293 * transition */
2295 }
2298 /* extra read targets are also write targets */
2302 /* There is nowhere to write, so all non-sync
2303 * drives must be failed - so we are finished
2304 */
2309 }
2333 }
2340 /* stop here */
2343 i--;
2347 /* remove last page from this bio */
2351 }
2353 }
2354 }
2355 }
2360 bio_full:
2363 /* For a user-requested sync, we read all readable devices and do a
2364 * compare
2365 */
2373 }
2374 }
2381 }
2383 }
2386 {
2391 }
2394 {
2406 GFP_KERNEL);
2419 r1bio_pool_free,
2437 }
2460 /*
2461 * The first working device is used as a
2462 * starting point to read balancing.
2463 */
2465 }
2472 }
2480 }
2484 abort:
2492 }
2494 }
2497 {
2506 }
2511 }
2512 /*
2513 * copy the already verified devices into our private RAID1
2514 * bookkeeping area. [whatever we allocate in run(),
2515 * should be freed in stop()]
2516 */
2519 else
2530 /* as we don't honour merge_bvec_fn, we must never risk
2531 * violating it, so limit ->max_segments to 1 lying within
2532 * a single page, as a one page request is never in violation.
2533 */
2538 }
2539 }
2560 /*
2561 * Ok, everything is just fine now
2562 */
2572 }
2574 }
2577 {
2581 /* wait for behind writes to complete */
2585 /* need to kick something here to make sure I/O goes? */
2588 }
2601 }
2604 {
2605 /* no resync is happening, and there is enough space
2606 * on all devices, so we can resize.
2607 * We need to make sure resync covers any new space.
2608 * If the array is shrinking we should possibly wait until
2609 * any io in the removed space completes, but it hardly seems
2610 * worth it.
2611 */
2621 }
2625 }
2628 {
2629 /* We need to:
2630 * 1/ resize the r1bio_pool
2631 * 2/ resize conf->mirrors
2632 *
2633 * We allocate a new r1bio_pool if we can.
2634 * Then raise a device barrier and wait until all IO stops.
2635 * Then resize conf->mirrors and swap in the new r1bio pool.
2636 *
2637 * At the same time, we "pack" the devices so that all the missing
2638 * devices have the higher raid_disk numbers.
2639 */
2648 /* Cannot change chunk_size, layout, or level */
2656 }
2668 cnt++;
2671 }
2684 }
2690 }
2694 /* ok, everything is stopped */
2708 }
2711 }
2731 }
2734 {
2747 }
2748 }
2751 {
2752 /* raid1 can take over:
2753 * raid5 with 2 devices, any layout or chunk size
2754 */
2764 }
2766 }
2769 {
2787 };
2790 {
2792 }
2795 {
2797 }