| blob | 240ff3125040ab51c1a010966e6e231fafd13062 |
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 {
291 }
294 {
301 /*
302 * this branch is our 'one mirror IO has finished' event handler:
303 */
309 /* If all other devices have failed, we want to return
310 * the error upwards rather than fail the last device.
311 * Here we redefine "uptodate" to mean "Don't want to retry"
312 */
320 }
325 /*
326 * oops, read error:
327 */
330 KERN_ERR "md/raid1:%s: %s: "
334 b),
338 }
341 }
344 {
345 /* it really is the end of this request */
347 /* free extra copy of the data pages */
353 }
354 /* clear the bitmap if all writes complete successfully */
360 }
363 {
373 else
375 }
376 }
379 {
388 /*
389 * 'one mirror IO has finished' event handler:
390 */
401 /*
402 * Set R1BIO_Uptodate in our master bio, so that we
403 * will return a good error code for to the higher
404 * levels even if IO on some other mirrored buffer
405 * fails.
406 *
407 * The 'master' represents the composite IO operation
408 * to user-side. So if something waits for IO, then it
409 * will wait for the 'master' bio.
410 */
411 sector_t first_bad;
418 /* Maybe we can clear some bad blocks. */
424 }
425 }
431 /*
432 * In behind mode, we ACK the master bio once the I/O
433 * has safely reached all non-writemostly
434 * disks. Setting the Returned bit ensures that this
435 * gets done only once -- we don't ever want to return
436 * -EIO here, instead we'll wait
437 */
440 /* Maybe we can return now */
449 }
450 }
451 }
456 /*
457 * Let's see if all mirrored write operations have finished
458 * already.
459 */
464 }
467 /*
468 * This routine returns the disk from which the requested read should
469 * be done. There is a per-array 'next expected sequential IO' sector
470 * number - if this matches on the next IO then we use the last disk.
471 * There is also a per-disk 'last know head position' sector that is
472 * maintained from IRQ contexts, both the normal and the resync IO
473 * completion handlers update this position correctly. If there is no
474 * perfect sequential match then we pick the disk whose head is closest.
475 *
476 * If there are 2 mirrors in the same 2 devices, performance degrades
477 * because position is mirror, not device based.
478 *
479 * The rdev for the device selected will have nr_pending incremented.
480 */
482 {
489 sector_t best_dist;
494 /*
495 * Check if we can balance. We can balance on the whole
496 * device if no resync is going on, or below the resync window.
497 * We take the first readable disk when above the resync window.
498 */
499 retry:
512 }
515 sector_t dist;
516 sector_t first_bad;
533 /* Don't balance among write-mostly, just
534 * use the first as a last resort */
539 /* Cannot use this */
545 }
547 }
548 /* This is a reasonable device to use. It might
549 * even be best.
550 */
554 /* already have a better device */
557 /* cannot read here. If this is the 'primary'
558 * device, then we must not read beyond
559 * bad_sectors from another device..
560 */
572 }
575 }
582 /* Don't change to another disk for sequential reads */
585 /* If device is idle, use it */
589 }
593 }
594 }
602 /* cannot risk returning a device that failed
603 * before we inc'ed nr_pending
604 */
607 }
611 }
616 }
621 {
642 }
643 }
644 }
646 }
649 }
652 {
668 /* Note the '|| 1' - when read_balance prefers
669 * non-congested targets, it can be removed
670 */
673 else
675 }
676 }
679 }
683 {
688 }
691 {
692 /* Any writes that have been queued but are awaiting
693 * bitmap updates get flushed here.
694 */
702 /* flush any pending bitmap writes to
703 * disk before proceeding w/ I/O */
712 }
715 }
717 /* Barriers....
718 * Sometimes we need to suspend IO while we do something else,
719 * either some resync/recovery, or reconfigure the array.
720 * To do this we raise a 'barrier'.
721 * The 'barrier' is a counter that can be raised multiple times
722 * to count how many activities are happening which preclude
723 * normal IO.
724 * We can only raise the barrier if there is no pending IO.
725 * i.e. if nr_pending == 0.
726 * We choose only to raise the barrier if no-one is waiting for the
727 * barrier to go down. This means that as soon as an IO request
728 * is ready, no other operations which require a barrier will start
729 * until the IO request has had a chance.
730 *
731 * So: regular IO calls 'wait_barrier'. When that returns there
732 * is no backgroup IO happening, It must arrange to call
733 * allow_barrier when it has finished its IO.
734 * backgroup IO calls must call raise_barrier. Once that returns
735 * there is no normal IO happeing. It must arrange to call
736 * lower_barrier when the particular background IO completes.
737 */
738 #define RESYNC_DEPTH 32
741 {
744 /* Wait until no block IO is waiting */
748 /* block any new IO from starting */
751 /* Now wait for all pending IO to complete */
757 }
760 {
767 }
770 {
774 /* Wait for the barrier to drop.
775 * However if there are already pending
776 * requests (preventing the barrier from
777 * rising completely), and the
778 * pre-process bio queue isn't empty,
779 * then don't wait, as we need to empty
780 * that queue to get the nr_pending
781 * count down.
782 */
789 );
791 }
794 }
797 {
803 }
806 {
807 /* stop syncio and normal IO and wait for everything to
808 * go quite.
809 * We increment barrier and nr_waiting, and then
810 * wait until nr_pending match nr_queued+1
811 * This is called in the context of one normal IO request
812 * that has failed. Thus any sync request that might be pending
813 * will be blocked by nr_pending, and we need to wait for
814 * pending IO requests to complete or be queued for re-try.
815 * Thus the number queued (nr_queued) plus this request (1)
816 * must match the number of pending IOs (nr_pending) before
817 * we continue.
818 */
827 }
829 {
830 /* reverse the effect of the freeze */
836 }
839 /* duplicate the data pages for behind I/O
840 */
842 {
846 GFP_NOIO);
859 }
865 do_sync_io:
871 }
874 {
890 /*
891 * Register the new request and wait if the reconstruction
892 * thread has put up a bar for new requests.
893 * Continue immediately if no resync is active currently.
894 */
901 /* As the suspend_* range is controlled by
902 * userspace, we want an interruptible
903 * wait.
904 */
914 }
916 }
922 /*
923 * make_request() can abort the operation when READA is being
924 * used and no empty request is available.
925 *
926 */
935 /* We might need to issue multiple reads to different
936 * devices if there are bad blocks around, so we keep
937 * track of the number of reads in bio->bi_phys_segments.
938 * If this is 0, there is only one r1_bio and no locking
939 * will be needed when requests complete. If it is
940 * non-zero, then it is the number of not-completed requests.
941 */
946 /*
947 * read balancing logic:
948 */
951 read_again:
955 /* couldn't find anywhere to read from */
958 }
962 bitmap) {
963 /* Reading from a write-mostly device must
964 * take care not to over-take any writes
965 * that are 'behind'
966 */
969 }
974 max_sectors);
985 /* could not read all from this device, so we will
986 * need another r1_bio.
987 */
995 else
998 /* Cannot call generic_make_request directly
999 * as that will be queued in __make_request
1000 * and subsequent mempool_alloc might block waiting
1001 * for it. So hand bio over to raid1d.
1002 */
1016 }
1018 /*
1019 * WRITE:
1020 */
1025 }
1026 /* first select target devices under rcu_lock and
1027 * inc refcount on their rdev. Record them by setting
1028 * bios[x] to bio
1029 * If there are known/acknowledged bad blocks on any device on
1030 * which we have seen a write error, we want to avoid writing those
1031 * blocks.
1032 * This potentially requires several writes to write around
1033 * the bad blocks. Each set of writes gets it's own r1bio
1034 * with a set of bios attached.
1035 */
1038 retry_write:
1048 }
1055 }
1059 sector_t first_bad;
1064 max_sectors,
1067 /* mustn't write here until the bad block is
1068 * acknowledged*/
1072 }
1074 /* Cannot write here at all */
1077 /* mustn't write more than bad_sectors
1078 * to other devices yet
1079 */
1082 /* We don't set R1BIO_Degraded as that
1083 * only applies if the disk is
1084 * missing, so it might be re-added,
1085 * and we want to know to recover this
1086 * chunk.
1087 * In this case the device is here,
1088 * and the fact that this chunk is not
1089 * in-sync is recorded in the bad
1090 * block log
1091 */
1093 }
1098 }
1099 }
1101 }
1105 /* Wait for this device to become unblocked */
1116 }
1119 /* We are splitting this write into multiple parts, so
1120 * we need to prepare for allocating another r1_bio.
1121 */
1126 else
1129 }
1145 /* do behind I/O ?
1146 * Not if there are too many, or cannot
1147 * allocate memory, or a reader on WriteMostly
1148 * is waiting for behind writes to flush */
1160 }
1165 /* Yes, I really want the '__' version so that
1166 * we clear any unused pointer in the io_vec, rather
1167 * than leave them unchanged. This is important
1168 * because when we come to free the pages, we won't
1169 * know the original bi_idx, so we just free
1170 * them all
1171 */
1176 }
1194 }
1195 /* Mustn't call r1_bio_write_done before this next test,
1196 * as it could result in the bio being freed.
1197 */
1200 /* We need another r1_bio. It has already been counted
1201 * in bio->bi_phys_segments
1202 */
1210 }
1214 /* In case raid1d snuck in to freeze_array */
1216 }
1219 {
1230 }
1233 }
1237 {
1241 /*
1242 * If it is not operational, then we have already marked it as dead
1243 * else if it is the last working disks, ignore the error, let the
1244 * next level up know.
1245 * else mark the drive as failed
1246 */
1249 /*
1250 * Don't fail the drive, act as though we were just a
1251 * normal single drive.
1252 * However don't try a recovery from this drive as
1253 * it is very likely to fail.
1254 */
1257 }
1265 /*
1266 * if recovery is running, make sure it aborts.
1267 */
1277 }
1280 {
1287 }
1300 }
1302 }
1305 {
1311 }
1314 {
1320 /*
1321 * Find all failed disks within the RAID1 configuration
1322 * and mark them readable.
1323 * Called under mddev lock, so rcu protection not needed.
1324 */
1332 /* replacement has just become active */
1335 count++;
1337 /* Replaced device not technically
1338 * faulty, but we need to be sure
1339 * it gets removed and never re-added
1340 */
1344 }
1345 }
1349 count++;
1351 }
1352 }
1359 }
1363 {
1381 }
1393 /* As all devices are equivalent, we don't need a full recovery
1394 * if this was recently any drive of the array
1395 */
1400 }
1403 /* Add this device as a replacement */
1411 }
1412 }
1414 /* Some requests might not have seen this new
1415 * merge_bvec_fn. We must wait for them to complete
1416 * before merging the device fully.
1417 * First we make sure any code which has tested
1418 * our function has submitted the request, then
1419 * we wait for all outstanding requests to complete.
1420 */
1425 }
1429 }
1432 {
1447 }
1448 /* Only remove non-faulty devices if recovery
1449 * is not possible.
1450 */
1456 }
1460 /* lost the race, try later */
1465 /* We just removed a device that is being replaced.
1466 * Move down the replacement. We drain all IO before
1467 * doing this to avoid confusion.
1468 */
1480 }
1481 abort:
1485 }
1489 {
1494 /*
1495 * we have read a block, now it needs to be re-written,
1496 * or re-read if the read failed.
1497 * We don't do much here, just schedule handling by raid1d
1498 */
1504 }
1507 {
1513 sector_t first_bad;
1522 /* make sure these bits doesn't get cleared. */
1544 )
1555 }
1556 }
1557 }
1561 {
1563 /* success */
1571 }
1572 /* need to record an error - either for the block or the device */
1576 }
1579 {
1580 /* Try some synchronous reads of other devices to get
1581 * good data, much like with normal read errors. Only
1582 * read into the pages we already have so we don't
1583 * need to re-issue the read request.
1584 * We don't need to freeze the array, because being in an
1585 * active sync request, there is no normal IO, and
1586 * no overlapping syncs.
1587 * We don't need to check is_badblock() again as we
1588 * made sure that anything with a bad block in range
1589 * will have bi_end_io clear.
1590 */
1609 /* No rcu protection needed here devices
1610 * can only be removed when no resync is
1611 * active, and resync is currently active
1612 */
1619 }
1620 }
1621 d++;
1629 /* Cannot read from anywhere, this block is lost.
1630 * Record a bad block on each device. If that doesn't
1631 * work just disable and interrupt the recovery.
1632 * Don't fail devices as that won't really help.
1633 */
1645 }
1653 }
1654 /* Try next page */
1657 idx++;
1659 }
1662 /* write it back and re-read */
1666 d--;
1675 }
1676 }
1681 d--;
1689 }
1692 idx ++;
1693 }
1697 }
1700 {
1701 /* We have read all readable devices. If we haven't
1702 * got the block, then there is no hope left.
1703 * If we have, then we want to do a comparison
1704 * and skip the write if everything is the same.
1705 * If any blocks failed to read, then we need to
1706 * attempt an over-write
1707 */
1720 }
1741 }
1748 /* No need to write to this device. */
1752 }
1753 /* fixup the bio for reuse */
1771 else
1776 PAGE_SIZE);
1777 }
1778 }
1780 }
1783 {
1792 /* ouch - failed to read all of that. */
1799 /*
1800 * schedule writes
1801 */
1817 }
1820 /* if we're here, all write(s) have completed, so clean up */
1823 }
1824 }
1826 /*
1827 * This is a kernel thread which:
1828 *
1829 * 1. Retries failed read operations on working mirrors.
1830 * 2. Updates the raid superblock when problems encounter.
1831 * 3. Performs writes following reads for array synchronising.
1832 */
1836 {
1849 /* Note: no rcu protection needed here
1850 * as this is synchronous in the raid1d thread
1851 * which is the thread that might remove
1852 * a device. If raid1d ever becomes multi-threaded....
1853 */
1854 sector_t first_bad;
1868 d++;
1871 }
1875 /* Cannot read from anywhere - mark it bad */
1880 }
1881 /* write it back and re-read */
1886 d--;
1892 }
1898 d--;
1906 "md/raid1:%s: read error corrected "
1912 }
1913 }
1914 }
1917 }
1918 }
1921 {
1923 }
1926 {
1937 }
1940 {
1947 /* bio has the data to be written to device 'i' where
1948 * we just recently had a write error.
1949 * We repeatedly clone the bio and trim down to one block,
1950 * then try the write. Where the write fails we record
1951 * a bad block.
1952 * It is conceivable that the bio doesn't exactly align with
1953 * blocks. We must handle this somehow.
1954 *
1955 * We currently own a reference on the rdev.
1956 */
1959 sector_t sector;
1978 idx++;
1983 }
1988 /* Write at 'sector' for 'sectors'*/
2002 /* failure! */
2011 }
2013 }
2016 {
2027 }
2032 }
2033 }
2036 }
2039 {
2049 /* This drive got a write error. We need to
2050 * narrow down and record precise write
2051 * errors.
2052 */
2056 /* an I/O failed, we can't clear the bitmap */
2058 }
2061 }
2065 }
2068 {
2077 /* we got a read error. Maybe the drive is bad. Maybe just
2078 * the block and we can fix it.
2079 * We freeze all other IO, and try reading the block from
2080 * other devices. When we find one, we re-write
2081 * and check it that fixes the read error.
2082 * This is all done synchronously while the array is
2083 * frozen
2084 */
2095 read_more:
2103 const unsigned long do_sync
2109 }
2116 "md/raid1:%s: redirecting sector %llu"
2127 /* Drat - have to split this up more */
2135 else
2154 }
2155 }
2158 {
2177 }
2189 else
2196 else
2197 /* just a partial read to be scheduled from separate
2198 * context
2199 */
2205 }
2207 }
2211 {
2222 }
2224 /*
2225 * perform a "sync" on one "block"
2226 *
2227 * We need to make sure that no normal I/O request - particularly write
2228 * requests - conflict with active sync requests.
2229 *
2230 * This is achieved by tracking pending requests and a 'barrier' concept
2231 * that can be installed to exclude normal IO requests.
2232 */
2234 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2235 {
2244 sector_t sync_blocks;
2255 /* If we aborted, we need to abort the
2256 * sync on the 'current' bitmap chunk (there will
2257 * only be one in raid1 resync.
2258 * We can find the current addess in mddev->curr_resync
2259 */
2269 }
2277 }
2278 /* before building a request, check if we can skip these blocks..
2279 * This call the bitmap_start_sync doesn't actually record anything
2280 */
2283 /* We can skip this block, and probably several more */
2286 }
2287 /*
2288 * If there is non-resync activity waiting for a turn,
2289 * and resync is going fast enough,
2290 * then let it though before starting on this new sync request.
2291 */
2302 /*
2303 * If we get a correctably read error during resync or recovery,
2304 * we might want to read from a different device. So we
2305 * flag all drives that could conceivably be read from for READ,
2306 * and any others (which will be non-In_sync devices) for WRITE.
2307 * If a read fails, we try reading from something else for which READ
2308 * is OK.
2309 */
2320 /* take from bio_init */
2340 write_targets ++;
2342 /* may need to read from here */
2355 }
2356 }
2364 }
2367 read_targets++;
2368 }
2369 }
2375 }
2376 }
2383 /* These sectors are bad on all InSync devices, so we
2384 * need to mark them bad on all write targets
2385 */
2393 }
2399 /* Cannot record the badblocks, so need to
2400 * abort the resync.
2401 * If there are multiple read targets, could just
2402 * fail the really bad ones ???
2403 */
2410 }
2412 /* only resync enough to reach the next bad->good
2413 * transition */
2415 }
2418 /* extra read targets are also write targets */
2422 /* There is nowhere to write, so all non-sync
2423 * drives must be failed - so we are finished
2424 */
2429 }
2453 }
2460 /* stop here */
2463 i--;
2467 /* remove last page from this bio */
2471 }
2473 }
2474 }
2475 }
2480 bio_full:
2483 /* For a user-requested sync, we read all readable devices and do a
2484 * compare
2485 */
2491 read_targets--;
2494 }
2495 }
2502 }
2504 }
2507 {
2512 }
2515 {
2528 GFP_KERNEL);
2541 r1bio_pool_free,
2558 else
2569 }
2589 /* This slot has a replacement. */
2591 /* No original, just make the replacement
2592 * a recovering spare
2593 */
2598 /* Original is not in_sync - bad */
2600 }
2609 /*
2610 * The first working device is used as a
2611 * starting point to read balancing.
2612 */
2614 }
2620 }
2628 }
2632 abort:
2640 }
2642 }
2646 {
2656 }
2661 }
2662 /*
2663 * copy the already verified devices into our private RAID1
2664 * bookkeeping area. [whatever we allocate in run(),
2665 * should be freed in stop()]
2666 */
2669 else
2680 }
2701 /*
2702 * Ok, everything is just fine now
2703 */
2714 }
2720 }
2723 {
2727 /* wait for behind writes to complete */
2731 /* need to kick something here to make sure I/O goes? */
2734 }
2747 }
2750 {
2751 /* no resync is happening, and there is enough space
2752 * on all devices, so we can resize.
2753 * We need to make sure resync covers any new space.
2754 * If the array is shrinking we should possibly wait until
2755 * any io in the removed space completes, but it hardly seems
2756 * worth it.
2757 */
2766 }
2774 }
2778 }
2781 {
2782 /* We need to:
2783 * 1/ resize the r1bio_pool
2784 * 2/ resize conf->mirrors
2785 *
2786 * We allocate a new r1bio_pool if we can.
2787 * Then raise a device barrier and wait until all IO stops.
2788 * Then resize conf->mirrors and swap in the new r1bio pool.
2789 *
2790 * At the same time, we "pack" the devices so that all the missing
2791 * devices have the higher raid_disk numbers.
2792 */
2801 /* Cannot change chunk_size, layout, or level */
2809 }
2821 cnt++;
2824 }
2837 }
2839 GFP_KERNEL);
2844 }
2848 /* ok, everything is stopped */
2862 }
2865 }
2885 }
2888 {
2901 }
2902 }
2905 {
2906 /* raid1 can take over:
2907 * raid5 with 2 devices, any layout or chunk size
2908 */
2918 }
2920 }
2923 {
2941 };
2944 {
2946 }
2949 {
2951 }