| blob | 56db3230c674343b7951a6f14a892dd1465c4a65 |
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 #define DEBUG 0
45 #define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
47 /*
48 * Number of guaranteed r1bios in case of extreme VM load:
49 */
50 #define NR_RAID1_BIOS 256
57 {
61 /* allocate a r1bio with room for raid_disks entries in the bios array */
63 }
66 {
68 }
70 #define RESYNC_BLOCK_SIZE (64*1024)
71 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
72 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
73 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
74 #define RESYNC_WINDOW (2048*1024)
77 {
88 /*
89 * Allocate bios : 1 for reading, n-1 for writing
90 */
96 }
97 /*
98 * Allocate RESYNC_PAGES data pages and attach them to
99 * the first bio.
100 * If this is a user-requested check/repair, allocate
101 * RESYNC_PAGES for each bio.
102 */
105 else
116 }
117 }
118 /* If not user-requests, copy the page pointers to all bios */
124 }
130 out_free_pages:
135 out_free_bio:
140 }
143 {
154 }
159 }
162 {
170 }
171 }
174 {
179 }
182 {
190 }
195 }
198 {
210 }
212 /*
213 * raid_end_bio_io() is called when we have finished servicing a mirrored
214 * operation and are ready to return a success/failure code to the buffer
215 * cache layer.
216 */
218 {
236 /*
237 * Wake up any possible resync thread that waits for the device
238 * to go idle.
239 */
241 }
242 }
245 {
248 /* if nobody has done the final endio yet, do it now */
257 }
259 }
261 /*
262 * Update disk head position estimator based on IRQ completion info.
263 */
265 {
270 }
273 {
280 /*
281 * this branch is our 'one mirror IO has finished' event handler:
282 */
288 /* If all other devices have failed, we want to return
289 * the error upwards rather than fail the last device.
290 * Here we redefine "uptodate" to mean "Don't want to retry"
291 */
299 }
304 /*
305 * oops, read error:
306 */
309 KERN_ERR "md/raid1:%s: %s: "
313 b),
317 }
320 }
323 {
324 /* it really is the end of this request */
326 /* free extra copy of the data pages */
332 }
333 /* clear the bitmap if all writes complete successfully */
339 }
342 {
352 else
354 }
355 }
358 {
370 /*
371 * 'one mirror IO has finished' event handler:
372 */
378 /*
379 * Set R1BIO_Uptodate in our master bio, so that we
380 * will return a good error code for to the higher
381 * levels even if IO on some other mirrored buffer
382 * fails.
383 *
384 * The 'master' represents the composite IO operation
385 * to user-side. So if something waits for IO, then it
386 * will wait for the 'master' bio.
387 */
388 sector_t first_bad;
395 /* Maybe we can clear some bad blocks. */
401 }
402 }
410 /*
411 * In behind mode, we ACK the master bio once the I/O
412 * has safely reached all non-writemostly
413 * disks. Setting the Returned bit ensures that this
414 * gets done only once -- we don't ever want to return
415 * -EIO here, instead we'll wait
416 */
419 /* Maybe we can return now */
427 }
428 }
429 }
434 /*
435 * Let's see if all mirrored write operations have finished
436 * already.
437 */
442 }
445 /*
446 * This routine returns the disk from which the requested read should
447 * be done. There is a per-array 'next expected sequential IO' sector
448 * number - if this matches on the next IO then we use the last disk.
449 * There is also a per-disk 'last know head position' sector that is
450 * maintained from IRQ contexts, both the normal and the resync IO
451 * completion handlers update this position correctly. If there is no
452 * perfect sequential match then we pick the disk whose head is closest.
453 *
454 * If there are 2 mirrors in the same 2 devices, performance degrades
455 * because position is mirror, not device based.
456 *
457 * The rdev for the device selected will have nr_pending incremented.
458 */
460 {
467 sector_t best_dist;
472 /*
473 * Check if we can balance. We can balance on the whole
474 * device if no resync is going on, or below the resync window.
475 * We take the first readable disk when above the resync window.
476 */
477 retry:
490 }
493 sector_t dist;
494 sector_t first_bad;
510 /* Don't balance among write-mostly, just
511 * use the first as a last resort */
515 }
516 /* This is a reasonable device to use. It might
517 * even be best.
518 */
522 /* already have a better device */
525 /* cannot read here. If this is the 'primary'
526 * device, then we must not read beyond
527 * bad_sectors from another device..
528 */
540 }
543 }
550 /* Don't change to another disk for sequential reads */
553 /* If device is idle, use it */
557 }
561 }
562 }
570 /* cannot risk returning a device that failed
571 * before we inc'ed nr_pending
572 */
575 }
579 }
584 }
587 {
599 /* Note the '|| 1' - when read_balance prefers
600 * non-congested targets, it can be removed
601 */
604 else
606 }
607 }
610 }
614 {
619 }
622 {
623 /* Any writes that have been queued but are awaiting
624 * bitmap updates get flushed here.
625 */
632 /* flush any pending bitmap writes to
633 * disk before proceeding w/ I/O */
641 }
644 }
646 /* Barriers....
647 * Sometimes we need to suspend IO while we do something else,
648 * either some resync/recovery, or reconfigure the array.
649 * To do this we raise a 'barrier'.
650 * The 'barrier' is a counter that can be raised multiple times
651 * to count how many activities are happening which preclude
652 * normal IO.
653 * We can only raise the barrier if there is no pending IO.
654 * i.e. if nr_pending == 0.
655 * We choose only to raise the barrier if no-one is waiting for the
656 * barrier to go down. This means that as soon as an IO request
657 * is ready, no other operations which require a barrier will start
658 * until the IO request has had a chance.
659 *
660 * So: regular IO calls 'wait_barrier'. When that returns there
661 * is no backgroup IO happening, It must arrange to call
662 * allow_barrier when it has finished its IO.
663 * backgroup IO calls must call raise_barrier. Once that returns
664 * there is no normal IO happeing. It must arrange to call
665 * lower_barrier when the particular background IO completes.
666 */
667 #define RESYNC_DEPTH 32
670 {
673 /* Wait until no block IO is waiting */
677 /* block any new IO from starting */
680 /* Now wait for all pending IO to complete */
686 }
689 {
696 }
699 {
705 );
707 }
710 }
713 {
719 }
722 {
723 /* stop syncio and normal IO and wait for everything to
724 * go quite.
725 * We increment barrier and nr_waiting, and then
726 * wait until nr_pending match nr_queued+1
727 * This is called in the context of one normal IO request
728 * that has failed. Thus any sync request that might be pending
729 * will be blocked by nr_pending, and we need to wait for
730 * pending IO requests to complete or be queued for re-try.
731 * Thus the number queued (nr_queued) plus this request (1)
732 * must match the number of pending IOs (nr_pending) before
733 * we continue.
734 */
743 }
745 {
746 /* reverse the effect of the freeze */
752 }
755 /* duplicate the data pages for behind I/O
756 */
758 {
762 GFP_NOIO);
775 }
781 do_sync_io:
787 }
790 {
807 /*
808 * Register the new request and wait if the reconstruction
809 * thread has put up a bar for new requests.
810 * Continue immediately if no resync is active currently.
811 */
818 /* As the suspend_* range is controlled by
819 * userspace, we want an interruptible
820 * wait.
821 */
831 }
833 }
839 /*
840 * make_request() can abort the operation when READA is being
841 * used and no empty request is available.
842 *
843 */
852 /* We might need to issue multiple reads to different
853 * devices if there are bad blocks around, so we keep
854 * track of the number of reads in bio->bi_phys_segments.
855 * If this is 0, there is only one r1_bio and no locking
856 * will be needed when requests complete. If it is
857 * non-zero, then it is the number of not-completed requests.
858 */
863 /*
864 * read balancing logic:
865 */
868 read_again:
872 /* couldn't find anywhere to read from */
875 }
879 bitmap) {
880 /* Reading from a write-mostly device must
881 * take care not to over-take any writes
882 * that are 'behind'
883 */
886 }
891 max_sectors);
902 /* could not read all from this device, so we will
903 * need another r1_bio.
904 */
912 else
915 /* Cannot call generic_make_request directly
916 * as that will be queued in __make_request
917 * and subsequent mempool_alloc might block waiting
918 * for it. So hand bio over to raid1d.
919 */
933 }
935 /*
936 * WRITE:
937 */
938 /* first select target devices under rcu_lock and
939 * inc refcount on their rdev. Record them by setting
940 * bios[x] to bio
941 * If there are known/acknowledged bad blocks on any device on
942 * which we have seen a write error, we want to avoid writing those
943 * blocks.
944 * This potentially requires several writes to write around
945 * the bad blocks. Each set of writes gets it's own r1bio
946 * with a set of bios attached.
947 */
951 retry_write:
961 }
966 }
970 sector_t first_bad;
975 max_sectors,
978 /* mustn't write here until the bad block is
979 * acknowledged*/
983 }
985 /* Cannot write here at all */
988 /* mustn't write more than bad_sectors
989 * to other devices yet
990 */
993 /* We don't set R1BIO_Degraded as that
994 * only applies if the disk is
995 * missing, so it might be re-added,
996 * and we want to know to recover this
997 * chunk.
998 * In this case the device is here,
999 * and the fact that this chunk is not
1000 * in-sync is recorded in the bad
1001 * block log
1002 */
1004 }
1009 }
1010 }
1012 }
1016 /* Wait for this device to become unblocked */
1027 }
1030 /* We are splitting this write into multiple parts, so
1031 * we need to prepare for allocating another r1_bio.
1032 */
1037 else
1040 }
1056 /* do behind I/O ?
1057 * Not if there are too many, or cannot
1058 * allocate memory, or a reader on WriteMostly
1059 * is waiting for behind writes to flush */
1071 }
1076 /* Yes, I really want the '__' version so that
1077 * we clear any unused pointer in the io_vec, rather
1078 * than leave them unchanged. This is important
1079 * because when we come to free the pages, we won't
1080 * know the original bi_idx, so we just free
1081 * them all
1082 */
1087 }
1102 }
1105 /* In case raid1d snuck in to freeze_array */
1109 /* We need another r1_bio. It has already been counted
1110 * in bio->bi_phys_segments
1111 */
1119 }
1125 }
1128 {
1139 }
1142 }
1146 {
1150 /*
1151 * If it is not operational, then we have already marked it as dead
1152 * else if it is the last working disks, ignore the error, let the
1153 * next level up know.
1154 * else mark the drive as failed
1155 */
1158 /*
1159 * Don't fail the drive, act as though we were just a
1160 * normal single drive.
1161 * However don't try a recovery from this drive as
1162 * it is very likely to fail.
1163 */
1166 }
1174 /*
1175 * if recovery is running, make sure it aborts.
1176 */
1186 }
1189 {
1196 }
1209 }
1211 }
1214 {
1220 }
1223 {
1229 /*
1230 * Find all failed disks within the RAID1 configuration
1231 * and mark them readable.
1232 * Called under mddev lock, so rcu protection not needed.
1233 */
1239 count++;
1241 }
1242 }
1249 }
1253 {
1272 /* as we don't honour merge_bvec_fn, we must
1273 * never risk violating it, so limit
1274 * ->max_segments to one lying with a single
1275 * page, as a one page request is never in
1276 * violation.
1277 */
1282 }
1287 /* As all devices are equivalent, we don't need a full recovery
1288 * if this was recently any drive of the array
1289 */
1294 }
1298 }
1301 {
1314 }
1315 /* Only remove non-faulty devices if recovery
1316 * is not possible.
1317 */
1323 }
1327 /* lost the race, try later */
1331 }
1333 }
1334 abort:
1338 }
1342 {
1351 /*
1352 * we have read a block, now it needs to be re-written,
1353 * or re-read if the read failed.
1354 * We don't do much here, just schedule handling by raid1d
1355 */
1361 }
1364 {
1371 sector_t first_bad;
1378 }
1383 /* make sure these bits doesn't get cleared. */
1401 )
1414 }
1415 }
1416 }
1420 {
1422 /* success */
1426 /* need to record an error - either for the block or the device */
1430 }
1433 {
1434 /* Try some synchronous reads of other devices to get
1435 * good data, much like with normal read errors. Only
1436 * read into the pages we already have so we don't
1437 * need to re-issue the read request.
1438 * We don't need to freeze the array, because being in an
1439 * active sync request, there is no normal IO, and
1440 * no overlapping syncs.
1441 * We don't need to check is_badblock() again as we
1442 * made sure that anything with a bad block in range
1443 * will have bi_end_io clear.
1444 */
1463 /* No rcu protection needed here devices
1464 * can only be removed when no resync is
1465 * active, and resync is currently active
1466 */
1473 }
1474 }
1475 d++;
1483 /* Cannot read from anywhere, this block is lost.
1484 * Record a bad block on each device. If that doesn't
1485 * work just disable and interrupt the recovery.
1486 * Don't fail devices as that won't really help.
1487 */
1499 }
1506 }
1507 /* Try next page */
1510 idx++;
1512 }
1515 /* write it back and re-read */
1519 d--;
1528 }
1529 }
1534 d--;
1542 }
1545 idx ++;
1546 }
1550 }
1553 {
1554 /* We have read all readable devices. If we haven't
1555 * got the block, then there is no hope left.
1556 * If we have, then we want to do a comparison
1557 * and skip the write if everything is the same.
1558 * If any blocks failed to read, then we need to
1559 * attempt an over-write
1560 */
1572 }
1591 PAGE_SIZE))
1593 }
1600 /* No need to write to this device. */
1604 }
1605 /* fixup the bio for reuse */
1623 else
1628 PAGE_SIZE);
1629 }
1630 }
1632 }
1635 {
1644 /* ouch - failed to read all of that. */
1651 /*
1652 * schedule writes
1653 */
1669 }
1672 /* if we're here, all write(s) have completed, so clean up */
1675 }
1676 }
1678 /*
1679 * This is a kernel thread which:
1680 *
1681 * 1. Retries failed read operations on working mirrors.
1682 * 2. Updates the raid superblock when problems encounter.
1683 * 3. Performs writes following reads for array synchronising.
1684 */
1688 {
1701 /* Note: no rcu protection needed here
1702 * as this is synchronous in the raid1d thread
1703 * which is the thread that might remove
1704 * a device. If raid1d ever becomes multi-threaded....
1705 */
1706 sector_t first_bad;
1718 d++;
1721 }
1725 /* Cannot read from anywhere - mark it bad */
1730 }
1731 /* write it back and re-read */
1736 d--;
1742 }
1748 d--;
1756 "md/raid1:%s: read error corrected "
1762 }
1763 }
1764 }
1767 }
1768 }
1771 {
1773 }
1776 {
1787 }
1790 {
1797 /* bio has the data to be written to device 'i' where
1798 * we just recently had a write error.
1799 * We repeatedly clone the bio and trim down to one block,
1800 * then try the write. Where the write fails we record
1801 * a bad block.
1802 * It is conceivable that the bio doesn't exactly align with
1803 * blocks. We must handle this somehow.
1804 *
1805 * We currently own a reference on the rdev.
1806 */
1809 sector_t sector;
1828 idx++;
1833 }
1838 /* Write at 'sector' for 'sectors'*/
1852 /* failure! */
1861 }
1863 }
1866 {
1877 }
1882 }
1883 }
1886 }
1889 {
1899 /* This drive got a write error. We need to
1900 * narrow down and record precise write
1901 * errors.
1902 */
1906 /* an I/O failed, we can't clear the bitmap */
1908 }
1911 }
1915 }
1918 {
1927 /* we got a read error. Maybe the drive is bad. Maybe just
1928 * the block and we can fix it.
1929 * We freeze all other IO, and try reading the block from
1930 * other devices. When we find one, we re-write
1931 * and check it that fixes the read error.
1932 * This is all done synchronously while the array is
1933 * frozen
1934 */
1945 read_more:
1953 const unsigned long do_sync
1959 }
1966 "md/raid1:%s: redirecting sector %llu"
1977 /* Drat - have to split this up more */
1985 else
2004 }
2005 }
2008 {
2027 }
2039 else
2046 else
2047 /* just a partial read to be scheduled from separate
2048 * context
2049 */
2055 }
2057 }
2061 {
2072 }
2074 /*
2075 * perform a "sync" on one "block"
2076 *
2077 * We need to make sure that no normal I/O request - particularly write
2078 * requests - conflict with active sync requests.
2079 *
2080 * This is achieved by tracking pending requests and a 'barrier' concept
2081 * that can be installed to exclude normal IO requests.
2082 */
2085 {
2094 sector_t sync_blocks;
2105 /* If we aborted, we need to abort the
2106 * sync on the 'current' bitmap chunk (there will
2107 * only be one in raid1 resync.
2108 * We can find the current addess in mddev->curr_resync
2109 */
2119 }
2127 }
2128 /* before building a request, check if we can skip these blocks..
2129 * This call the bitmap_start_sync doesn't actually record anything
2130 */
2133 /* We can skip this block, and probably several more */
2136 }
2137 /*
2138 * If there is non-resync activity waiting for a turn,
2139 * and resync is going fast enough,
2140 * then let it though before starting on this new sync request.
2141 */
2152 /*
2153 * If we get a correctably read error during resync or recovery,
2154 * we might want to read from a different device. So we
2155 * flag all drives that could conceivably be read from for READ,
2156 * and any others (which will be non-In_sync devices) for WRITE.
2157 * If a read fails, we try reading from something else for which READ
2158 * is OK.
2159 */
2170 /* take from bio_init */
2190 write_targets ++;
2192 /* may need to read from here */
2205 }
2206 }
2214 }
2217 read_targets++;
2218 }
2219 }
2225 }
2226 }
2233 /* These sectors are bad on all InSync devices, so we
2234 * need to mark them bad on all write targets
2235 */
2244 }
2250 /* Cannot record the badblocks, so need to
2251 * abort the resync.
2252 * If there are multiple read targets, could just
2253 * fail the really bad ones ???
2254 */
2261 }
2263 /* only resync enough to reach the next bad->good
2264 * transition */
2266 }
2269 /* extra read targets are also write targets */
2273 /* There is nowhere to write, so all non-sync
2274 * drives must be failed - so we are finished
2275 */
2280 }
2304 }
2311 /* stop here */
2314 i--;
2318 /* remove last page from this bio */
2322 }
2324 }
2325 }
2326 }
2331 bio_full:
2334 /* For a user-requested sync, we read all readable devices and do a
2335 * compare
2336 */
2344 }
2345 }
2352 }
2354 }
2357 {
2362 }
2365 {
2377 GFP_KERNEL);
2390 r1bio_pool_free,
2408 }
2429 /*
2430 * The first working device is used as a
2431 * starting point to read balancing.
2432 */
2434 }
2441 }
2449 }
2453 abort:
2461 }
2463 }
2466 {
2475 }
2480 }
2481 /*
2482 * copy the already verified devices into our private RAID1
2483 * bookkeeping area. [whatever we allocate in run(),
2484 * should be freed in stop()]
2485 */
2488 else
2499 /* as we don't honour merge_bvec_fn, we must never risk
2500 * violating it, so limit ->max_segments to 1 lying within
2501 * a single page, as a one page request is never in violation.
2502 */
2507 }
2508 }
2529 /*
2530 * Ok, everything is just fine now
2531 */
2541 }
2543 }
2546 {
2550 /* wait for behind writes to complete */
2554 /* need to kick something here to make sure I/O goes? */
2557 }
2571 }
2574 {
2575 /* no resync is happening, and there is enough space
2576 * on all devices, so we can resize.
2577 * We need to make sure resync covers any new space.
2578 * If the array is shrinking we should possibly wait until
2579 * any io in the removed space completes, but it hardly seems
2580 * worth it.
2581 */
2591 }
2595 }
2598 {
2599 /* We need to:
2600 * 1/ resize the r1bio_pool
2601 * 2/ resize conf->mirrors
2602 *
2603 * We allocate a new r1bio_pool if we can.
2604 * Then raise a device barrier and wait until all IO stops.
2605 * Then resize conf->mirrors and swap in the new r1bio pool.
2606 *
2607 * At the same time, we "pack" the devices so that all the missing
2608 * devices have the higher raid_disk numbers.
2609 */
2618 /* Cannot change chunk_size, layout, or level */
2626 }
2638 cnt++;
2641 }
2654 }
2660 }
2664 /* ok, everything is stopped */
2678 }
2681 }
2701 }
2704 {
2717 }
2718 }
2721 {
2722 /* raid1 can take over:
2723 * raid5 with 2 devices, any layout or chunk size
2724 */
2734 }
2736 }
2739 {
2757 };
2760 {
2762 }
2765 {
2767 }