| blob | a368db2431a596020a98a3e64f67685ccef9c6cc |
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;
532 /* Don't balance among write-mostly, just
533 * use the first as a last resort */
538 /* Cannot use this */
544 }
546 }
547 /* This is a reasonable device to use. It might
548 * even be best.
549 */
553 /* already have a better device */
556 /* cannot read here. If this is the 'primary'
557 * device, then we must not read beyond
558 * bad_sectors from another device..
559 */
571 }
574 }
581 /* Don't change to another disk for sequential reads */
584 /* If device is idle, use it */
588 }
592 }
593 }
601 /* cannot risk returning a device that failed
602 * before we inc'ed nr_pending
603 */
606 }
610 }
615 }
618 {
634 /* Note the '|| 1' - when read_balance prefers
635 * non-congested targets, it can be removed
636 */
639 else
641 }
642 }
645 }
649 {
654 }
657 {
658 /* Any writes that have been queued but are awaiting
659 * bitmap updates get flushed here.
660 */
668 /* flush any pending bitmap writes to
669 * disk before proceeding w/ I/O */
678 }
681 }
683 /* Barriers....
684 * Sometimes we need to suspend IO while we do something else,
685 * either some resync/recovery, or reconfigure the array.
686 * To do this we raise a 'barrier'.
687 * The 'barrier' is a counter that can be raised multiple times
688 * to count how many activities are happening which preclude
689 * normal IO.
690 * We can only raise the barrier if there is no pending IO.
691 * i.e. if nr_pending == 0.
692 * We choose only to raise the barrier if no-one is waiting for the
693 * barrier to go down. This means that as soon as an IO request
694 * is ready, no other operations which require a barrier will start
695 * until the IO request has had a chance.
696 *
697 * So: regular IO calls 'wait_barrier'. When that returns there
698 * is no backgroup IO happening, It must arrange to call
699 * allow_barrier when it has finished its IO.
700 * backgroup IO calls must call raise_barrier. Once that returns
701 * there is no normal IO happeing. It must arrange to call
702 * lower_barrier when the particular background IO completes.
703 */
704 #define RESYNC_DEPTH 32
707 {
710 /* Wait until no block IO is waiting */
714 /* block any new IO from starting */
717 /* Now wait for all pending IO to complete */
723 }
726 {
733 }
736 {
742 );
744 }
747 }
750 {
756 }
759 {
760 /* stop syncio and normal IO and wait for everything to
761 * go quite.
762 * We increment barrier and nr_waiting, and then
763 * wait until nr_pending match nr_queued+1
764 * This is called in the context of one normal IO request
765 * that has failed. Thus any sync request that might be pending
766 * will be blocked by nr_pending, and we need to wait for
767 * pending IO requests to complete or be queued for re-try.
768 * Thus the number queued (nr_queued) plus this request (1)
769 * must match the number of pending IOs (nr_pending) before
770 * we continue.
771 */
780 }
782 {
783 /* reverse the effect of the freeze */
789 }
792 /* duplicate the data pages for behind I/O
793 */
795 {
799 GFP_NOIO);
812 }
818 do_sync_io:
824 }
827 {
844 /*
845 * Register the new request and wait if the reconstruction
846 * thread has put up a bar for new requests.
847 * Continue immediately if no resync is active currently.
848 */
855 /* As the suspend_* range is controlled by
856 * userspace, we want an interruptible
857 * wait.
858 */
868 }
870 }
876 /*
877 * make_request() can abort the operation when READA is being
878 * used and no empty request is available.
879 *
880 */
889 /* We might need to issue multiple reads to different
890 * devices if there are bad blocks around, so we keep
891 * track of the number of reads in bio->bi_phys_segments.
892 * If this is 0, there is only one r1_bio and no locking
893 * will be needed when requests complete. If it is
894 * non-zero, then it is the number of not-completed requests.
895 */
900 /*
901 * read balancing logic:
902 */
905 read_again:
909 /* couldn't find anywhere to read from */
912 }
916 bitmap) {
917 /* Reading from a write-mostly device must
918 * take care not to over-take any writes
919 * that are 'behind'
920 */
923 }
928 max_sectors);
939 /* could not read all from this device, so we will
940 * need another r1_bio.
941 */
949 else
952 /* Cannot call generic_make_request directly
953 * as that will be queued in __make_request
954 * and subsequent mempool_alloc might block waiting
955 * for it. So hand bio over to raid1d.
956 */
970 }
972 /*
973 * WRITE:
974 */
979 }
980 /* first select target devices under rcu_lock and
981 * inc refcount on their rdev. Record them by setting
982 * bios[x] to bio
983 * If there are known/acknowledged bad blocks on any device on
984 * which we have seen a write error, we want to avoid writing those
985 * blocks.
986 * This potentially requires several writes to write around
987 * the bad blocks. Each set of writes gets it's own r1bio
988 * with a set of bios attached.
989 */
993 retry_write:
1003 }
1009 }
1013 sector_t first_bad;
1018 max_sectors,
1021 /* mustn't write here until the bad block is
1022 * acknowledged*/
1026 }
1028 /* Cannot write here at all */
1031 /* mustn't write more than bad_sectors
1032 * to other devices yet
1033 */
1036 /* We don't set R1BIO_Degraded as that
1037 * only applies if the disk is
1038 * missing, so it might be re-added,
1039 * and we want to know to recover this
1040 * chunk.
1041 * In this case the device is here,
1042 * and the fact that this chunk is not
1043 * in-sync is recorded in the bad
1044 * block log
1045 */
1047 }
1052 }
1053 }
1055 }
1059 /* Wait for this device to become unblocked */
1070 }
1073 /* We are splitting this write into multiple parts, so
1074 * we need to prepare for allocating another r1_bio.
1075 */
1080 else
1083 }
1099 /* do behind I/O ?
1100 * Not if there are too many, or cannot
1101 * allocate memory, or a reader on WriteMostly
1102 * is waiting for behind writes to flush */
1114 }
1119 /* Yes, I really want the '__' version so that
1120 * we clear any unused pointer in the io_vec, rather
1121 * than leave them unchanged. This is important
1122 * because when we come to free the pages, we won't
1123 * know the original bi_idx, so we just free
1124 * them all
1125 */
1130 }
1146 }
1147 /* Mustn't call r1_bio_write_done before this next test,
1148 * as it could result in the bio being freed.
1149 */
1152 /* We need another r1_bio. It has already been counted
1153 * in bio->bi_phys_segments
1154 */
1162 }
1166 /* In case raid1d snuck in to freeze_array */
1171 }
1174 {
1185 }
1188 }
1192 {
1196 /*
1197 * If it is not operational, then we have already marked it as dead
1198 * else if it is the last working disks, ignore the error, let the
1199 * next level up know.
1200 * else mark the drive as failed
1201 */
1204 /*
1205 * Don't fail the drive, act as though we were just a
1206 * normal single drive.
1207 * However don't try a recovery from this drive as
1208 * it is very likely to fail.
1209 */
1212 }
1220 /*
1221 * if recovery is running, make sure it aborts.
1222 */
1232 }
1235 {
1242 }
1255 }
1257 }
1260 {
1266 }
1269 {
1275 /*
1276 * Find all failed disks within the RAID1 configuration
1277 * and mark them readable.
1278 * Called under mddev lock, so rcu protection not needed.
1279 */
1287 /* replacement has just become active */
1290 count++;
1292 /* Replaced device not technically
1293 * faulty, but we need to be sure
1294 * it gets removed and never re-added
1295 */
1299 }
1300 }
1304 count++;
1306 }
1307 }
1314 }
1318 {
1338 /* as we don't honour merge_bvec_fn, we must
1339 * never risk violating it, so limit
1340 * ->max_segments to one lying with a single
1341 * page, as a one page request is never in
1342 * violation.
1343 */
1348 }
1353 /* As all devices are equivalent, we don't need a full recovery
1354 * if this was recently any drive of the array
1355 */
1360 }
1363 /* Add this device as a replacement */
1371 }
1372 }
1376 }
1379 {
1394 }
1395 /* Only remove non-faulty devices if recovery
1396 * is not possible.
1397 */
1403 }
1407 /* lost the race, try later */
1412 /* We just removed a device that is being replaced.
1413 * Move down the replacement. We drain all IO before
1414 * doing this to avoid confusion.
1415 */
1427 }
1428 abort:
1432 }
1436 {
1441 /*
1442 * we have read a block, now it needs to be re-written,
1443 * or re-read if the read failed.
1444 * We don't do much here, just schedule handling by raid1d
1445 */
1451 }
1454 {
1460 sector_t first_bad;
1469 /* make sure these bits doesn't get cleared. */
1491 )
1502 }
1503 }
1504 }
1508 {
1510 /* success */
1518 }
1519 /* need to record an error - either for the block or the device */
1523 }
1526 {
1527 /* Try some synchronous reads of other devices to get
1528 * good data, much like with normal read errors. Only
1529 * read into the pages we already have so we don't
1530 * need to re-issue the read request.
1531 * We don't need to freeze the array, because being in an
1532 * active sync request, there is no normal IO, and
1533 * no overlapping syncs.
1534 * We don't need to check is_badblock() again as we
1535 * made sure that anything with a bad block in range
1536 * will have bi_end_io clear.
1537 */
1556 /* No rcu protection needed here devices
1557 * can only be removed when no resync is
1558 * active, and resync is currently active
1559 */
1566 }
1567 }
1568 d++;
1576 /* Cannot read from anywhere, this block is lost.
1577 * Record a bad block on each device. If that doesn't
1578 * work just disable and interrupt the recovery.
1579 * Don't fail devices as that won't really help.
1580 */
1592 }
1600 }
1601 /* Try next page */
1604 idx++;
1606 }
1609 /* write it back and re-read */
1613 d--;
1622 }
1623 }
1628 d--;
1636 }
1639 idx ++;
1640 }
1644 }
1647 {
1648 /* We have read all readable devices. If we haven't
1649 * got the block, then there is no hope left.
1650 * If we have, then we want to do a comparison
1651 * and skip the write if everything is the same.
1652 * If any blocks failed to read, then we need to
1653 * attempt an over-write
1654 */
1666 }
1685 PAGE_SIZE))
1687 }
1694 /* No need to write to this device. */
1698 }
1699 /* fixup the bio for reuse */
1717 else
1722 PAGE_SIZE);
1723 }
1724 }
1726 }
1729 {
1738 /* ouch - failed to read all of that. */
1745 /*
1746 * schedule writes
1747 */
1763 }
1766 /* if we're here, all write(s) have completed, so clean up */
1769 }
1770 }
1772 /*
1773 * This is a kernel thread which:
1774 *
1775 * 1. Retries failed read operations on working mirrors.
1776 * 2. Updates the raid superblock when problems encounter.
1777 * 3. Performs writes following reads for array synchronising.
1778 */
1782 {
1795 /* Note: no rcu protection needed here
1796 * as this is synchronous in the raid1d thread
1797 * which is the thread that might remove
1798 * a device. If raid1d ever becomes multi-threaded....
1799 */
1800 sector_t first_bad;
1812 d++;
1815 }
1819 /* Cannot read from anywhere - mark it bad */
1824 }
1825 /* write it back and re-read */
1830 d--;
1836 }
1842 d--;
1850 "md/raid1:%s: read error corrected "
1856 }
1857 }
1858 }
1861 }
1862 }
1865 {
1867 }
1870 {
1881 }
1884 {
1891 /* bio has the data to be written to device 'i' where
1892 * we just recently had a write error.
1893 * We repeatedly clone the bio and trim down to one block,
1894 * then try the write. Where the write fails we record
1895 * a bad block.
1896 * It is conceivable that the bio doesn't exactly align with
1897 * blocks. We must handle this somehow.
1898 *
1899 * We currently own a reference on the rdev.
1900 */
1903 sector_t sector;
1922 idx++;
1927 }
1932 /* Write at 'sector' for 'sectors'*/
1946 /* failure! */
1955 }
1957 }
1960 {
1971 }
1976 }
1977 }
1980 }
1983 {
1993 /* This drive got a write error. We need to
1994 * narrow down and record precise write
1995 * errors.
1996 */
2000 /* an I/O failed, we can't clear the bitmap */
2002 }
2005 }
2009 }
2012 {
2021 /* we got a read error. Maybe the drive is bad. Maybe just
2022 * the block and we can fix it.
2023 * We freeze all other IO, and try reading the block from
2024 * other devices. When we find one, we re-write
2025 * and check it that fixes the read error.
2026 * This is all done synchronously while the array is
2027 * frozen
2028 */
2039 read_more:
2047 const unsigned long do_sync
2053 }
2060 "md/raid1:%s: redirecting sector %llu"
2071 /* Drat - have to split this up more */
2079 else
2098 }
2099 }
2102 {
2121 }
2133 else
2140 else
2141 /* just a partial read to be scheduled from separate
2142 * context
2143 */
2149 }
2151 }
2155 {
2166 }
2168 /*
2169 * perform a "sync" on one "block"
2170 *
2171 * We need to make sure that no normal I/O request - particularly write
2172 * requests - conflict with active sync requests.
2173 *
2174 * This is achieved by tracking pending requests and a 'barrier' concept
2175 * that can be installed to exclude normal IO requests.
2176 */
2178 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2179 {
2188 sector_t sync_blocks;
2199 /* If we aborted, we need to abort the
2200 * sync on the 'current' bitmap chunk (there will
2201 * only be one in raid1 resync.
2202 * We can find the current addess in mddev->curr_resync
2203 */
2213 }
2221 }
2222 /* before building a request, check if we can skip these blocks..
2223 * This call the bitmap_start_sync doesn't actually record anything
2224 */
2227 /* We can skip this block, and probably several more */
2230 }
2231 /*
2232 * If there is non-resync activity waiting for a turn,
2233 * and resync is going fast enough,
2234 * then let it though before starting on this new sync request.
2235 */
2246 /*
2247 * If we get a correctably read error during resync or recovery,
2248 * we might want to read from a different device. So we
2249 * flag all drives that could conceivably be read from for READ,
2250 * and any others (which will be non-In_sync devices) for WRITE.
2251 * If a read fails, we try reading from something else for which READ
2252 * is OK.
2253 */
2264 /* take from bio_init */
2284 write_targets ++;
2286 /* may need to read from here */
2299 }
2300 }
2308 }
2311 read_targets++;
2312 }
2313 }
2319 }
2320 }
2327 /* These sectors are bad on all InSync devices, so we
2328 * need to mark them bad on all write targets
2329 */
2338 }
2344 /* Cannot record the badblocks, so need to
2345 * abort the resync.
2346 * If there are multiple read targets, could just
2347 * fail the really bad ones ???
2348 */
2355 }
2357 /* only resync enough to reach the next bad->good
2358 * transition */
2360 }
2363 /* extra read targets are also write targets */
2367 /* There is nowhere to write, so all non-sync
2368 * drives must be failed - so we are finished
2369 */
2374 }
2398 }
2405 /* stop here */
2408 i--;
2412 /* remove last page from this bio */
2416 }
2418 }
2419 }
2420 }
2425 bio_full:
2428 /* For a user-requested sync, we read all readable devices and do a
2429 * compare
2430 */
2438 }
2439 }
2446 }
2448 }
2451 {
2456 }
2459 {
2472 GFP_KERNEL);
2485 r1bio_pool_free,
2501 else
2509 }
2529 /* This slot has a replacement. */
2531 /* No original, just make the replacement
2532 * a recovering spare
2533 */
2538 /* Original is not in_sync - bad */
2540 }
2548 /*
2549 * The first working device is used as a
2550 * starting point to read balancing.
2551 */
2553 }
2559 }
2567 }
2571 abort:
2579 }
2581 }
2584 {
2593 }
2598 }
2599 /*
2600 * copy the already verified devices into our private RAID1
2601 * bookkeeping area. [whatever we allocate in run(),
2602 * should be freed in stop()]
2603 */
2606 else
2617 /* as we don't honour merge_bvec_fn, we must never risk
2618 * violating it, so limit ->max_segments to 1 lying within
2619 * a single page, as a one page request is never in violation.
2620 */
2625 }
2626 }
2647 /*
2648 * Ok, everything is just fine now
2649 */
2659 }
2661 }
2664 {
2668 /* wait for behind writes to complete */
2672 /* need to kick something here to make sure I/O goes? */
2675 }
2688 }
2691 {
2692 /* no resync is happening, and there is enough space
2693 * on all devices, so we can resize.
2694 * We need to make sure resync covers any new space.
2695 * If the array is shrinking we should possibly wait until
2696 * any io in the removed space completes, but it hardly seems
2697 * worth it.
2698 */
2708 }
2712 }
2715 {
2716 /* We need to:
2717 * 1/ resize the r1bio_pool
2718 * 2/ resize conf->mirrors
2719 *
2720 * We allocate a new r1bio_pool if we can.
2721 * Then raise a device barrier and wait until all IO stops.
2722 * Then resize conf->mirrors and swap in the new r1bio pool.
2723 *
2724 * At the same time, we "pack" the devices so that all the missing
2725 * devices have the higher raid_disk numbers.
2726 */
2735 /* Cannot change chunk_size, layout, or level */
2743 }
2755 cnt++;
2758 }
2771 }
2773 GFP_KERNEL);
2778 }
2782 /* ok, everything is stopped */
2796 }
2799 }
2819 }
2822 {
2835 }
2836 }
2839 {
2840 /* raid1 can take over:
2841 * raid5 with 2 devices, any layout or chunk size
2842 */
2852 }
2854 }
2857 {
2875 };
2878 {
2880 }
2883 {
2885 }