| blob | edc735a98bdcc3319e661120d82af76b84c2806a |
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 {
740 /* Wait for the barrier to drop.
741 * However if there are already pending
742 * requests (preventing the barrier from
743 * rising completely), and the
744 * pre-process bio queue isn't empty,
745 * then don't wait, as we need to empty
746 * that queue to get the nr_pending
747 * count down.
748 */
755 );
757 }
760 }
763 {
769 }
772 {
773 /* stop syncio and normal IO and wait for everything to
774 * go quite.
775 * We increment barrier and nr_waiting, and then
776 * wait until nr_pending match nr_queued+1
777 * This is called in the context of one normal IO request
778 * that has failed. Thus any sync request that might be pending
779 * will be blocked by nr_pending, and we need to wait for
780 * pending IO requests to complete or be queued for re-try.
781 * Thus the number queued (nr_queued) plus this request (1)
782 * must match the number of pending IOs (nr_pending) before
783 * we continue.
784 */
793 }
795 {
796 /* reverse the effect of the freeze */
802 }
805 /* duplicate the data pages for behind I/O
806 */
808 {
812 GFP_NOIO);
825 }
831 do_sync_io:
837 }
840 {
857 /*
858 * Register the new request and wait if the reconstruction
859 * thread has put up a bar for new requests.
860 * Continue immediately if no resync is active currently.
861 */
868 /* As the suspend_* range is controlled by
869 * userspace, we want an interruptible
870 * wait.
871 */
881 }
883 }
889 /*
890 * make_request() can abort the operation when READA is being
891 * used and no empty request is available.
892 *
893 */
902 /* We might need to issue multiple reads to different
903 * devices if there are bad blocks around, so we keep
904 * track of the number of reads in bio->bi_phys_segments.
905 * If this is 0, there is only one r1_bio and no locking
906 * will be needed when requests complete. If it is
907 * non-zero, then it is the number of not-completed requests.
908 */
913 /*
914 * read balancing logic:
915 */
918 read_again:
922 /* couldn't find anywhere to read from */
925 }
929 bitmap) {
930 /* Reading from a write-mostly device must
931 * take care not to over-take any writes
932 * that are 'behind'
933 */
936 }
941 max_sectors);
952 /* could not read all from this device, so we will
953 * need another r1_bio.
954 */
962 else
965 /* Cannot call generic_make_request directly
966 * as that will be queued in __make_request
967 * and subsequent mempool_alloc might block waiting
968 * for it. So hand bio over to raid1d.
969 */
983 }
985 /*
986 * WRITE:
987 */
992 }
993 /* first select target devices under rcu_lock and
994 * inc refcount on their rdev. Record them by setting
995 * bios[x] to bio
996 * If there are known/acknowledged bad blocks on any device on
997 * which we have seen a write error, we want to avoid writing those
998 * blocks.
999 * This potentially requires several writes to write around
1000 * the bad blocks. Each set of writes gets it's own r1bio
1001 * with a set of bios attached.
1002 */
1006 retry_write:
1016 }
1022 }
1026 sector_t first_bad;
1031 max_sectors,
1034 /* mustn't write here until the bad block is
1035 * acknowledged*/
1039 }
1041 /* Cannot write here at all */
1044 /* mustn't write more than bad_sectors
1045 * to other devices yet
1046 */
1049 /* We don't set R1BIO_Degraded as that
1050 * only applies if the disk is
1051 * missing, so it might be re-added,
1052 * and we want to know to recover this
1053 * chunk.
1054 * In this case the device is here,
1055 * and the fact that this chunk is not
1056 * in-sync is recorded in the bad
1057 * block log
1058 */
1060 }
1065 }
1066 }
1068 }
1072 /* Wait for this device to become unblocked */
1083 }
1086 /* We are splitting this write into multiple parts, so
1087 * we need to prepare for allocating another r1_bio.
1088 */
1093 else
1096 }
1112 /* do behind I/O ?
1113 * Not if there are too many, or cannot
1114 * allocate memory, or a reader on WriteMostly
1115 * is waiting for behind writes to flush */
1127 }
1132 /* Yes, I really want the '__' version so that
1133 * we clear any unused pointer in the io_vec, rather
1134 * than leave them unchanged. This is important
1135 * because when we come to free the pages, we won't
1136 * know the original bi_idx, so we just free
1137 * them all
1138 */
1143 }
1159 }
1160 /* Mustn't call r1_bio_write_done before this next test,
1161 * as it could result in the bio being freed.
1162 */
1165 /* We need another r1_bio. It has already been counted
1166 * in bio->bi_phys_segments
1167 */
1175 }
1179 /* In case raid1d snuck in to freeze_array */
1184 }
1187 {
1198 }
1201 }
1205 {
1209 /*
1210 * If it is not operational, then we have already marked it as dead
1211 * else if it is the last working disks, ignore the error, let the
1212 * next level up know.
1213 * else mark the drive as failed
1214 */
1217 /*
1218 * Don't fail the drive, act as though we were just a
1219 * normal single drive.
1220 * However don't try a recovery from this drive as
1221 * it is very likely to fail.
1222 */
1225 }
1233 /*
1234 * if recovery is running, make sure it aborts.
1235 */
1245 }
1248 {
1255 }
1268 }
1270 }
1273 {
1279 }
1282 {
1288 /*
1289 * Find all failed disks within the RAID1 configuration
1290 * and mark them readable.
1291 * Called under mddev lock, so rcu protection not needed.
1292 */
1300 /* replacement has just become active */
1303 count++;
1305 /* Replaced device not technically
1306 * faulty, but we need to be sure
1307 * it gets removed and never re-added
1308 */
1312 }
1313 }
1317 count++;
1319 }
1320 }
1327 }
1331 {
1351 /* as we don't honour merge_bvec_fn, we must
1352 * never risk violating it, so limit
1353 * ->max_segments to one lying with a single
1354 * page, as a one page request is never in
1355 * violation.
1356 */
1361 }
1366 /* As all devices are equivalent, we don't need a full recovery
1367 * if this was recently any drive of the array
1368 */
1373 }
1376 /* Add this device as a replacement */
1384 }
1385 }
1389 }
1392 {
1407 }
1408 /* Only remove non-faulty devices if recovery
1409 * is not possible.
1410 */
1416 }
1420 /* lost the race, try later */
1425 /* We just removed a device that is being replaced.
1426 * Move down the replacement. We drain all IO before
1427 * doing this to avoid confusion.
1428 */
1440 }
1441 abort:
1445 }
1449 {
1454 /*
1455 * we have read a block, now it needs to be re-written,
1456 * or re-read if the read failed.
1457 * We don't do much here, just schedule handling by raid1d
1458 */
1464 }
1467 {
1473 sector_t first_bad;
1482 /* make sure these bits doesn't get cleared. */
1504 )
1515 }
1516 }
1517 }
1521 {
1523 /* success */
1531 }
1532 /* need to record an error - either for the block or the device */
1536 }
1539 {
1540 /* Try some synchronous reads of other devices to get
1541 * good data, much like with normal read errors. Only
1542 * read into the pages we already have so we don't
1543 * need to re-issue the read request.
1544 * We don't need to freeze the array, because being in an
1545 * active sync request, there is no normal IO, and
1546 * no overlapping syncs.
1547 * We don't need to check is_badblock() again as we
1548 * made sure that anything with a bad block in range
1549 * will have bi_end_io clear.
1550 */
1569 /* No rcu protection needed here devices
1570 * can only be removed when no resync is
1571 * active, and resync is currently active
1572 */
1579 }
1580 }
1581 d++;
1589 /* Cannot read from anywhere, this block is lost.
1590 * Record a bad block on each device. If that doesn't
1591 * work just disable and interrupt the recovery.
1592 * Don't fail devices as that won't really help.
1593 */
1605 }
1613 }
1614 /* Try next page */
1617 idx++;
1619 }
1622 /* write it back and re-read */
1626 d--;
1635 }
1636 }
1641 d--;
1649 }
1652 idx ++;
1653 }
1657 }
1660 {
1661 /* We have read all readable devices. If we haven't
1662 * got the block, then there is no hope left.
1663 * If we have, then we want to do a comparison
1664 * and skip the write if everything is the same.
1665 * If any blocks failed to read, then we need to
1666 * attempt an over-write
1667 */
1680 }
1699 PAGE_SIZE))
1701 }
1708 /* No need to write to this device. */
1712 }
1713 /* fixup the bio for reuse */
1731 else
1736 PAGE_SIZE);
1737 }
1738 }
1740 }
1743 {
1752 /* ouch - failed to read all of that. */
1759 /*
1760 * schedule writes
1761 */
1777 }
1780 /* if we're here, all write(s) have completed, so clean up */
1783 }
1784 }
1786 /*
1787 * This is a kernel thread which:
1788 *
1789 * 1. Retries failed read operations on working mirrors.
1790 * 2. Updates the raid superblock when problems encounter.
1791 * 3. Performs writes following reads for array synchronising.
1792 */
1796 {
1809 /* Note: no rcu protection needed here
1810 * as this is synchronous in the raid1d thread
1811 * which is the thread that might remove
1812 * a device. If raid1d ever becomes multi-threaded....
1813 */
1814 sector_t first_bad;
1826 d++;
1829 }
1833 /* Cannot read from anywhere - mark it bad */
1838 }
1839 /* write it back and re-read */
1844 d--;
1850 }
1856 d--;
1864 "md/raid1:%s: read error corrected "
1870 }
1871 }
1872 }
1875 }
1876 }
1879 {
1881 }
1884 {
1895 }
1898 {
1905 /* bio has the data to be written to device 'i' where
1906 * we just recently had a write error.
1907 * We repeatedly clone the bio and trim down to one block,
1908 * then try the write. Where the write fails we record
1909 * a bad block.
1910 * It is conceivable that the bio doesn't exactly align with
1911 * blocks. We must handle this somehow.
1912 *
1913 * We currently own a reference on the rdev.
1914 */
1917 sector_t sector;
1936 idx++;
1941 }
1946 /* Write at 'sector' for 'sectors'*/
1960 /* failure! */
1969 }
1971 }
1974 {
1985 }
1990 }
1991 }
1994 }
1997 {
2007 /* This drive got a write error. We need to
2008 * narrow down and record precise write
2009 * errors.
2010 */
2014 /* an I/O failed, we can't clear the bitmap */
2016 }
2019 }
2023 }
2026 {
2035 /* we got a read error. Maybe the drive is bad. Maybe just
2036 * the block and we can fix it.
2037 * We freeze all other IO, and try reading the block from
2038 * other devices. When we find one, we re-write
2039 * and check it that fixes the read error.
2040 * This is all done synchronously while the array is
2041 * frozen
2042 */
2053 read_more:
2061 const unsigned long do_sync
2067 }
2074 "md/raid1:%s: redirecting sector %llu"
2085 /* Drat - have to split this up more */
2093 else
2112 }
2113 }
2116 {
2135 }
2147 else
2154 else
2155 /* just a partial read to be scheduled from separate
2156 * context
2157 */
2163 }
2165 }
2169 {
2180 }
2182 /*
2183 * perform a "sync" on one "block"
2184 *
2185 * We need to make sure that no normal I/O request - particularly write
2186 * requests - conflict with active sync requests.
2187 *
2188 * This is achieved by tracking pending requests and a 'barrier' concept
2189 * that can be installed to exclude normal IO requests.
2190 */
2192 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2193 {
2202 sector_t sync_blocks;
2213 /* If we aborted, we need to abort the
2214 * sync on the 'current' bitmap chunk (there will
2215 * only be one in raid1 resync.
2216 * We can find the current addess in mddev->curr_resync
2217 */
2227 }
2235 }
2236 /* before building a request, check if we can skip these blocks..
2237 * This call the bitmap_start_sync doesn't actually record anything
2238 */
2241 /* We can skip this block, and probably several more */
2244 }
2245 /*
2246 * If there is non-resync activity waiting for a turn,
2247 * and resync is going fast enough,
2248 * then let it though before starting on this new sync request.
2249 */
2260 /*
2261 * If we get a correctably read error during resync or recovery,
2262 * we might want to read from a different device. So we
2263 * flag all drives that could conceivably be read from for READ,
2264 * and any others (which will be non-In_sync devices) for WRITE.
2265 * If a read fails, we try reading from something else for which READ
2266 * is OK.
2267 */
2278 /* take from bio_init */
2298 write_targets ++;
2300 /* may need to read from here */
2313 }
2314 }
2322 }
2325 read_targets++;
2326 }
2327 }
2333 }
2334 }
2341 /* These sectors are bad on all InSync devices, so we
2342 * need to mark them bad on all write targets
2343 */
2352 }
2358 /* Cannot record the badblocks, so need to
2359 * abort the resync.
2360 * If there are multiple read targets, could just
2361 * fail the really bad ones ???
2362 */
2369 }
2371 /* only resync enough to reach the next bad->good
2372 * transition */
2374 }
2377 /* extra read targets are also write targets */
2381 /* There is nowhere to write, so all non-sync
2382 * drives must be failed - so we are finished
2383 */
2388 }
2412 }
2419 /* stop here */
2422 i--;
2426 /* remove last page from this bio */
2430 }
2432 }
2433 }
2434 }
2439 bio_full:
2442 /* For a user-requested sync, we read all readable devices and do a
2443 * compare
2444 */
2452 }
2453 }
2460 }
2462 }
2465 {
2470 }
2473 {
2486 GFP_KERNEL);
2499 r1bio_pool_free,
2515 else
2523 }
2543 /* This slot has a replacement. */
2545 /* No original, just make the replacement
2546 * a recovering spare
2547 */
2552 /* Original is not in_sync - bad */
2554 }
2562 /*
2563 * The first working device is used as a
2564 * starting point to read balancing.
2565 */
2567 }
2573 }
2581 }
2585 abort:
2593 }
2595 }
2598 {
2607 }
2612 }
2613 /*
2614 * copy the already verified devices into our private RAID1
2615 * bookkeeping area. [whatever we allocate in run(),
2616 * should be freed in stop()]
2617 */
2620 else
2631 /* as we don't honour merge_bvec_fn, we must never risk
2632 * violating it, so limit ->max_segments to 1 lying within
2633 * a single page, as a one page request is never in violation.
2634 */
2639 }
2640 }
2661 /*
2662 * Ok, everything is just fine now
2663 */
2673 }
2675 }
2678 {
2682 /* wait for behind writes to complete */
2686 /* need to kick something here to make sure I/O goes? */
2689 }
2702 }
2705 {
2706 /* no resync is happening, and there is enough space
2707 * on all devices, so we can resize.
2708 * We need to make sure resync covers any new space.
2709 * If the array is shrinking we should possibly wait until
2710 * any io in the removed space completes, but it hardly seems
2711 * worth it.
2712 */
2722 }
2726 }
2729 {
2730 /* We need to:
2731 * 1/ resize the r1bio_pool
2732 * 2/ resize conf->mirrors
2733 *
2734 * We allocate a new r1bio_pool if we can.
2735 * Then raise a device barrier and wait until all IO stops.
2736 * Then resize conf->mirrors and swap in the new r1bio pool.
2737 *
2738 * At the same time, we "pack" the devices so that all the missing
2739 * devices have the higher raid_disk numbers.
2740 */
2749 /* Cannot change chunk_size, layout, or level */
2757 }
2769 cnt++;
2772 }
2785 }
2787 GFP_KERNEL);
2792 }
2796 /* ok, everything is stopped */
2810 }
2813 }
2833 }
2836 {
2849 }
2850 }
2853 {
2854 /* raid1 can take over:
2855 * raid5 with 2 devices, any layout or chunk size
2856 */
2866 }
2868 }
2871 {
2889 };
2892 {
2894 }
2897 {
2899 }