| blob | d9587dffe533e69c81b6adf5221ee1da745cfa4b |
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/seq_file.h>
38 #include <linux/ratelimit.h>
43 #define DEBUG 0
44 #define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
46 /*
47 * Number of guaranteed r1bios in case of extreme VM load:
48 */
49 #define NR_RAID1_BIOS 256
56 {
60 /* allocate a r1bio with room for raid_disks entries in the bios array */
62 }
65 {
67 }
69 #define RESYNC_BLOCK_SIZE (64*1024)
70 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
71 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
72 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
73 #define RESYNC_WINDOW (2048*1024)
76 {
87 /*
88 * Allocate bios : 1 for reading, n-1 for writing
89 */
95 }
96 /*
97 * Allocate RESYNC_PAGES data pages and attach them to
98 * the first bio.
99 * If this is a user-requested check/repair, allocate
100 * RESYNC_PAGES for each bio.
101 */
104 else
115 }
116 }
117 /* If not user-requests, copy the page pointers to all bios */
123 }
129 out_free_pages:
134 out_free_bio:
139 }
142 {
153 }
158 }
161 {
169 }
170 }
173 {
178 }
181 {
189 }
194 }
197 {
209 }
211 /*
212 * raid_end_bio_io() is called when we have finished servicing a mirrored
213 * operation and are ready to return a success/failure code to the buffer
214 * cache layer.
215 */
217 {
235 /*
236 * Wake up any possible resync thread that waits for the device
237 * to go idle.
238 */
240 }
241 }
244 {
247 /* if nobody has done the final endio yet, do it now */
256 }
258 }
260 /*
261 * Update disk head position estimator based on IRQ completion info.
262 */
264 {
269 }
272 {
279 /*
280 * this branch is our 'one mirror IO has finished' event handler:
281 */
287 /* If all other devices have failed, we want to return
288 * the error upwards rather than fail the last device.
289 * Here we redefine "uptodate" to mean "Don't want to retry"
290 */
298 }
303 /*
304 * oops, read error:
305 */
308 KERN_ERR "md/raid1:%s: %s: "
312 b),
316 }
319 }
322 {
323 /* it really is the end of this request */
325 /* free extra copy of the data pages */
331 }
332 /* clear the bitmap if all writes complete successfully */
338 }
341 {
351 else
353 }
354 }
357 {
369 /*
370 * 'one mirror IO has finished' event handler:
371 */
377 /*
378 * Set R1BIO_Uptodate in our master bio, so that we
379 * will return a good error code for to the higher
380 * levels even if IO on some other mirrored buffer
381 * fails.
382 *
383 * The 'master' represents the composite IO operation
384 * to user-side. So if something waits for IO, then it
385 * will wait for the 'master' bio.
386 */
387 sector_t first_bad;
394 /* Maybe we can clear some bad blocks. */
400 }
401 }
409 /*
410 * In behind mode, we ACK the master bio once the I/O
411 * has safely reached all non-writemostly
412 * disks. Setting the Returned bit ensures that this
413 * gets done only once -- we don't ever want to return
414 * -EIO here, instead we'll wait
415 */
418 /* Maybe we can return now */
426 }
427 }
428 }
433 /*
434 * Let's see if all mirrored write operations have finished
435 * already.
436 */
441 }
444 /*
445 * This routine returns the disk from which the requested read should
446 * be done. There is a per-array 'next expected sequential IO' sector
447 * number - if this matches on the next IO then we use the last disk.
448 * There is also a per-disk 'last know head position' sector that is
449 * maintained from IRQ contexts, both the normal and the resync IO
450 * completion handlers update this position correctly. If there is no
451 * perfect sequential match then we pick the disk whose head is closest.
452 *
453 * If there are 2 mirrors in the same 2 devices, performance degrades
454 * because position is mirror, not device based.
455 *
456 * The rdev for the device selected will have nr_pending incremented.
457 */
459 {
466 sector_t best_dist;
471 /*
472 * Check if we can balance. We can balance on the whole
473 * device if no resync is going on, or below the resync window.
474 * We take the first readable disk when above the resync window.
475 */
476 retry:
489 }
492 sector_t dist;
493 sector_t first_bad;
509 /* Don't balance among write-mostly, just
510 * use the first as a last resort */
514 }
515 /* This is a reasonable device to use. It might
516 * even be best.
517 */
521 /* already have a better device */
524 /* cannot read here. If this is the 'primary'
525 * device, then we must not read beyond
526 * bad_sectors from another device..
527 */
539 }
542 }
549 /* Don't change to another disk for sequential reads */
552 /* If device is idle, use it */
556 }
560 }
561 }
569 /* cannot risk returning a device that failed
570 * before we inc'ed nr_pending
571 */
574 }
578 }
583 }
586 {
598 /* Note the '|| 1' - when read_balance prefers
599 * non-congested targets, it can be removed
600 */
603 else
605 }
606 }
609 }
613 {
618 }
621 {
622 /* Any writes that have been queued but are awaiting
623 * bitmap updates get flushed here.
624 */
631 /* flush any pending bitmap writes to
632 * disk before proceeding w/ I/O */
640 }
643 }
645 /* Barriers....
646 * Sometimes we need to suspend IO while we do something else,
647 * either some resync/recovery, or reconfigure the array.
648 * To do this we raise a 'barrier'.
649 * The 'barrier' is a counter that can be raised multiple times
650 * to count how many activities are happening which preclude
651 * normal IO.
652 * We can only raise the barrier if there is no pending IO.
653 * i.e. if nr_pending == 0.
654 * We choose only to raise the barrier if no-one is waiting for the
655 * barrier to go down. This means that as soon as an IO request
656 * is ready, no other operations which require a barrier will start
657 * until the IO request has had a chance.
658 *
659 * So: regular IO calls 'wait_barrier'. When that returns there
660 * is no backgroup IO happening, It must arrange to call
661 * allow_barrier when it has finished its IO.
662 * backgroup IO calls must call raise_barrier. Once that returns
663 * there is no normal IO happeing. It must arrange to call
664 * lower_barrier when the particular background IO completes.
665 */
666 #define RESYNC_DEPTH 32
669 {
672 /* Wait until no block IO is waiting */
676 /* block any new IO from starting */
679 /* Now wait for all pending IO to complete */
685 }
688 {
695 }
698 {
704 );
706 }
709 }
712 {
718 }
721 {
722 /* stop syncio and normal IO and wait for everything to
723 * go quite.
724 * We increment barrier and nr_waiting, and then
725 * wait until nr_pending match nr_queued+1
726 * This is called in the context of one normal IO request
727 * that has failed. Thus any sync request that might be pending
728 * will be blocked by nr_pending, and we need to wait for
729 * pending IO requests to complete or be queued for re-try.
730 * Thus the number queued (nr_queued) plus this request (1)
731 * must match the number of pending IOs (nr_pending) before
732 * we continue.
733 */
742 }
744 {
745 /* reverse the effect of the freeze */
751 }
754 /* duplicate the data pages for behind I/O
755 */
757 {
761 GFP_NOIO);
774 }
780 do_sync_io:
786 }
789 {
806 /*
807 * Register the new request and wait if the reconstruction
808 * thread has put up a bar for new requests.
809 * Continue immediately if no resync is active currently.
810 */
817 /* As the suspend_* range is controlled by
818 * userspace, we want an interruptible
819 * wait.
820 */
830 }
832 }
838 /*
839 * make_request() can abort the operation when READA is being
840 * used and no empty request is available.
841 *
842 */
851 /* We might need to issue multiple reads to different
852 * devices if there are bad blocks around, so we keep
853 * track of the number of reads in bio->bi_phys_segments.
854 * If this is 0, there is only one r1_bio and no locking
855 * will be needed when requests complete. If it is
856 * non-zero, then it is the number of not-completed requests.
857 */
862 /*
863 * read balancing logic:
864 */
867 read_again:
871 /* couldn't find anywhere to read from */
874 }
878 bitmap) {
879 /* Reading from a write-mostly device must
880 * take care not to over-take any writes
881 * that are 'behind'
882 */
885 }
890 max_sectors);
901 /* could not read all from this device, so we will
902 * need another r1_bio.
903 */
911 else
914 /* Cannot call generic_make_request directly
915 * as that will be queued in __make_request
916 * and subsequent mempool_alloc might block waiting
917 * for it. So hand bio over to raid1d.
918 */
932 }
934 /*
935 * WRITE:
936 */
937 /* first select target devices under rcu_lock and
938 * inc refcount on their rdev. Record them by setting
939 * bios[x] to bio
940 * If there are known/acknowledged bad blocks on any device on
941 * which we have seen a write error, we want to avoid writing those
942 * blocks.
943 * This potentially requires several writes to write around
944 * the bad blocks. Each set of writes gets it's own r1bio
945 * with a set of bios attached.
946 */
950 retry_write:
960 }
965 }
969 sector_t first_bad;
974 max_sectors,
977 /* mustn't write here until the bad block is
978 * acknowledged*/
982 }
984 /* Cannot write here at all */
987 /* mustn't write more than bad_sectors
988 * to other devices yet
989 */
992 /* We don't set R1BIO_Degraded as that
993 * only applies if the disk is
994 * missing, so it might be re-added,
995 * and we want to know to recover this
996 * chunk.
997 * In this case the device is here,
998 * and the fact that this chunk is not
999 * in-sync is recorded in the bad
1000 * block log
1001 */
1003 }
1008 }
1009 }
1011 }
1015 /* Wait for this device to become unblocked */
1026 }
1029 /* We are splitting this write into multiple parts, so
1030 * we need to prepare for allocating another r1_bio.
1031 */
1036 else
1039 }
1055 /* do behind I/O ?
1056 * Not if there are too many, or cannot
1057 * allocate memory, or a reader on WriteMostly
1058 * is waiting for behind writes to flush */
1070 }
1075 /* Yes, I really want the '__' version so that
1076 * we clear any unused pointer in the io_vec, rather
1077 * than leave them unchanged. This is important
1078 * because when we come to free the pages, we won't
1079 * know the original bi_idx, so we just free
1080 * them all
1081 */
1086 }
1101 }
1102 /* Mustn't call r1_bio_write_done before this next test,
1103 * as it could result in the bio being freed.
1104 */
1107 /* We need another r1_bio. It has already been counted
1108 * in bio->bi_phys_segments
1109 */
1117 }
1121 /* In case raid1d snuck in to freeze_array */
1128 }
1131 {
1142 }
1145 }
1149 {
1153 /*
1154 * If it is not operational, then we have already marked it as dead
1155 * else if it is the last working disks, ignore the error, let the
1156 * next level up know.
1157 * else mark the drive as failed
1158 */
1161 /*
1162 * Don't fail the drive, act as though we were just a
1163 * normal single drive.
1164 * However don't try a recovery from this drive as
1165 * it is very likely to fail.
1166 */
1169 }
1177 /*
1178 * if recovery is running, make sure it aborts.
1179 */
1189 }
1192 {
1199 }
1212 }
1214 }
1217 {
1223 }
1226 {
1232 /*
1233 * Find all failed disks within the RAID1 configuration
1234 * and mark them readable.
1235 * Called under mddev lock, so rcu protection not needed.
1236 */
1242 count++;
1244 }
1245 }
1252 }
1256 {
1275 /* as we don't honour merge_bvec_fn, we must
1276 * never risk violating it, so limit
1277 * ->max_segments to one lying with a single
1278 * page, as a one page request is never in
1279 * violation.
1280 */
1285 }
1290 /* As all devices are equivalent, we don't need a full recovery
1291 * if this was recently any drive of the array
1292 */
1297 }
1301 }
1304 {
1317 }
1318 /* Only remove non-faulty devices if recovery
1319 * is not possible.
1320 */
1326 }
1330 /* lost the race, try later */
1334 }
1336 }
1337 abort:
1341 }
1345 {
1354 /*
1355 * we have read a block, now it needs to be re-written,
1356 * or re-read if the read failed.
1357 * We don't do much here, just schedule handling by raid1d
1358 */
1364 }
1367 {
1374 sector_t first_bad;
1381 }
1386 /* make sure these bits doesn't get cleared. */
1404 )
1417 }
1418 }
1419 }
1423 {
1425 /* success */
1429 /* need to record an error - either for the block or the device */
1433 }
1436 {
1437 /* Try some synchronous reads of other devices to get
1438 * good data, much like with normal read errors. Only
1439 * read into the pages we already have so we don't
1440 * need to re-issue the read request.
1441 * We don't need to freeze the array, because being in an
1442 * active sync request, there is no normal IO, and
1443 * no overlapping syncs.
1444 * We don't need to check is_badblock() again as we
1445 * made sure that anything with a bad block in range
1446 * will have bi_end_io clear.
1447 */
1466 /* No rcu protection needed here devices
1467 * can only be removed when no resync is
1468 * active, and resync is currently active
1469 */
1476 }
1477 }
1478 d++;
1486 /* Cannot read from anywhere, this block is lost.
1487 * Record a bad block on each device. If that doesn't
1488 * work just disable and interrupt the recovery.
1489 * Don't fail devices as that won't really help.
1490 */
1502 }
1509 }
1510 /* Try next page */
1513 idx++;
1515 }
1518 /* write it back and re-read */
1522 d--;
1531 }
1532 }
1537 d--;
1545 }
1548 idx ++;
1549 }
1553 }
1556 {
1557 /* We have read all readable devices. If we haven't
1558 * got the block, then there is no hope left.
1559 * If we have, then we want to do a comparison
1560 * and skip the write if everything is the same.
1561 * If any blocks failed to read, then we need to
1562 * attempt an over-write
1563 */
1575 }
1594 PAGE_SIZE))
1596 }
1603 /* No need to write to this device. */
1607 }
1608 /* fixup the bio for reuse */
1626 else
1631 PAGE_SIZE);
1632 }
1633 }
1635 }
1638 {
1647 /* ouch - failed to read all of that. */
1654 /*
1655 * schedule writes
1656 */
1672 }
1675 /* if we're here, all write(s) have completed, so clean up */
1678 }
1679 }
1681 /*
1682 * This is a kernel thread which:
1683 *
1684 * 1. Retries failed read operations on working mirrors.
1685 * 2. Updates the raid superblock when problems encounter.
1686 * 3. Performs writes following reads for array synchronising.
1687 */
1691 {
1704 /* Note: no rcu protection needed here
1705 * as this is synchronous in the raid1d thread
1706 * which is the thread that might remove
1707 * a device. If raid1d ever becomes multi-threaded....
1708 */
1709 sector_t first_bad;
1721 d++;
1724 }
1728 /* Cannot read from anywhere - mark it bad */
1733 }
1734 /* write it back and re-read */
1739 d--;
1745 }
1751 d--;
1759 "md/raid1:%s: read error corrected "
1765 }
1766 }
1767 }
1770 }
1771 }
1774 {
1776 }
1779 {
1790 }
1793 {
1800 /* bio has the data to be written to device 'i' where
1801 * we just recently had a write error.
1802 * We repeatedly clone the bio and trim down to one block,
1803 * then try the write. Where the write fails we record
1804 * a bad block.
1805 * It is conceivable that the bio doesn't exactly align with
1806 * blocks. We must handle this somehow.
1807 *
1808 * We currently own a reference on the rdev.
1809 */
1812 sector_t sector;
1831 idx++;
1836 }
1841 /* Write at 'sector' for 'sectors'*/
1855 /* failure! */
1864 }
1866 }
1869 {
1880 }
1885 }
1886 }
1889 }
1892 {
1902 /* This drive got a write error. We need to
1903 * narrow down and record precise write
1904 * errors.
1905 */
1909 /* an I/O failed, we can't clear the bitmap */
1911 }
1914 }
1918 }
1921 {
1930 /* we got a read error. Maybe the drive is bad. Maybe just
1931 * the block and we can fix it.
1932 * We freeze all other IO, and try reading the block from
1933 * other devices. When we find one, we re-write
1934 * and check it that fixes the read error.
1935 * This is all done synchronously while the array is
1936 * frozen
1937 */
1948 read_more:
1956 const unsigned long do_sync
1962 }
1969 "md/raid1:%s: redirecting sector %llu"
1980 /* Drat - have to split this up more */
1988 else
2007 }
2008 }
2011 {
2030 }
2042 else
2049 else
2050 /* just a partial read to be scheduled from separate
2051 * context
2052 */
2058 }
2060 }
2064 {
2075 }
2077 /*
2078 * perform a "sync" on one "block"
2079 *
2080 * We need to make sure that no normal I/O request - particularly write
2081 * requests - conflict with active sync requests.
2082 *
2083 * This is achieved by tracking pending requests and a 'barrier' concept
2084 * that can be installed to exclude normal IO requests.
2085 */
2088 {
2097 sector_t sync_blocks;
2108 /* If we aborted, we need to abort the
2109 * sync on the 'current' bitmap chunk (there will
2110 * only be one in raid1 resync.
2111 * We can find the current addess in mddev->curr_resync
2112 */
2122 }
2130 }
2131 /* before building a request, check if we can skip these blocks..
2132 * This call the bitmap_start_sync doesn't actually record anything
2133 */
2136 /* We can skip this block, and probably several more */
2139 }
2140 /*
2141 * If there is non-resync activity waiting for a turn,
2142 * and resync is going fast enough,
2143 * then let it though before starting on this new sync request.
2144 */
2155 /*
2156 * If we get a correctably read error during resync or recovery,
2157 * we might want to read from a different device. So we
2158 * flag all drives that could conceivably be read from for READ,
2159 * and any others (which will be non-In_sync devices) for WRITE.
2160 * If a read fails, we try reading from something else for which READ
2161 * is OK.
2162 */
2173 /* take from bio_init */
2193 write_targets ++;
2195 /* may need to read from here */
2208 }
2209 }
2217 }
2220 read_targets++;
2221 }
2222 }
2228 }
2229 }
2236 /* These sectors are bad on all InSync devices, so we
2237 * need to mark them bad on all write targets
2238 */
2247 }
2253 /* Cannot record the badblocks, so need to
2254 * abort the resync.
2255 * If there are multiple read targets, could just
2256 * fail the really bad ones ???
2257 */
2264 }
2266 /* only resync enough to reach the next bad->good
2267 * transition */
2269 }
2272 /* extra read targets are also write targets */
2276 /* There is nowhere to write, so all non-sync
2277 * drives must be failed - so we are finished
2278 */
2283 }
2307 }
2314 /* stop here */
2317 i--;
2321 /* remove last page from this bio */
2325 }
2327 }
2328 }
2329 }
2334 bio_full:
2337 /* For a user-requested sync, we read all readable devices and do a
2338 * compare
2339 */
2347 }
2348 }
2355 }
2357 }
2360 {
2365 }
2368 {
2380 GFP_KERNEL);
2393 r1bio_pool_free,
2411 }
2432 /*
2433 * The first working device is used as a
2434 * starting point to read balancing.
2435 */
2437 }
2444 }
2452 }
2456 abort:
2464 }
2466 }
2469 {
2478 }
2483 }
2484 /*
2485 * copy the already verified devices into our private RAID1
2486 * bookkeeping area. [whatever we allocate in run(),
2487 * should be freed in stop()]
2488 */
2491 else
2502 /* as we don't honour merge_bvec_fn, we must never risk
2503 * violating it, so limit ->max_segments to 1 lying within
2504 * a single page, as a one page request is never in violation.
2505 */
2510 }
2511 }
2532 /*
2533 * Ok, everything is just fine now
2534 */
2544 }
2546 }
2549 {
2553 /* wait for behind writes to complete */
2557 /* need to kick something here to make sure I/O goes? */
2560 }
2573 }
2576 {
2577 /* no resync is happening, and there is enough space
2578 * on all devices, so we can resize.
2579 * We need to make sure resync covers any new space.
2580 * If the array is shrinking we should possibly wait until
2581 * any io in the removed space completes, but it hardly seems
2582 * worth it.
2583 */
2593 }
2597 }
2600 {
2601 /* We need to:
2602 * 1/ resize the r1bio_pool
2603 * 2/ resize conf->mirrors
2604 *
2605 * We allocate a new r1bio_pool if we can.
2606 * Then raise a device barrier and wait until all IO stops.
2607 * Then resize conf->mirrors and swap in the new r1bio pool.
2608 *
2609 * At the same time, we "pack" the devices so that all the missing
2610 * devices have the higher raid_disk numbers.
2611 */
2620 /* Cannot change chunk_size, layout, or level */
2628 }
2640 cnt++;
2643 }
2656 }
2662 }
2666 /* ok, everything is stopped */
2680 }
2683 }
2703 }
2706 {
2719 }
2720 }
2723 {
2724 /* raid1 can take over:
2725 * raid5 with 2 devices, any layout or chunk size
2726 */
2736 }
2738 }
2741 {
2759 };
2762 {
2764 }
2767 {
2769 }