| blob | 606fc04fd76a25aa8d3530e612fcb477cc8e660c |
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 */
515 /* Cannot use this */
521 }
523 }
524 /* This is a reasonable device to use. It might
525 * even be best.
526 */
530 /* already have a better device */
533 /* cannot read here. If this is the 'primary'
534 * device, then we must not read beyond
535 * bad_sectors from another device..
536 */
548 }
551 }
558 /* Don't change to another disk for sequential reads */
561 /* If device is idle, use it */
565 }
569 }
570 }
578 /* cannot risk returning a device that failed
579 * before we inc'ed nr_pending
580 */
583 }
587 }
592 }
595 {
607 /* Note the '|| 1' - when read_balance prefers
608 * non-congested targets, it can be removed
609 */
612 else
614 }
615 }
618 }
622 {
627 }
630 {
631 /* Any writes that have been queued but are awaiting
632 * bitmap updates get flushed here.
633 */
640 /* flush any pending bitmap writes to
641 * disk before proceeding w/ I/O */
649 }
652 }
654 /* Barriers....
655 * Sometimes we need to suspend IO while we do something else,
656 * either some resync/recovery, or reconfigure the array.
657 * To do this we raise a 'barrier'.
658 * The 'barrier' is a counter that can be raised multiple times
659 * to count how many activities are happening which preclude
660 * normal IO.
661 * We can only raise the barrier if there is no pending IO.
662 * i.e. if nr_pending == 0.
663 * We choose only to raise the barrier if no-one is waiting for the
664 * barrier to go down. This means that as soon as an IO request
665 * is ready, no other operations which require a barrier will start
666 * until the IO request has had a chance.
667 *
668 * So: regular IO calls 'wait_barrier'. When that returns there
669 * is no backgroup IO happening, It must arrange to call
670 * allow_barrier when it has finished its IO.
671 * backgroup IO calls must call raise_barrier. Once that returns
672 * there is no normal IO happeing. It must arrange to call
673 * lower_barrier when the particular background IO completes.
674 */
675 #define RESYNC_DEPTH 32
678 {
681 /* Wait until no block IO is waiting */
685 /* block any new IO from starting */
688 /* Now wait for all pending IO to complete */
694 }
697 {
704 }
707 {
713 );
715 }
718 }
721 {
727 }
730 {
731 /* stop syncio and normal IO and wait for everything to
732 * go quite.
733 * We increment barrier and nr_waiting, and then
734 * wait until nr_pending match nr_queued+1
735 * This is called in the context of one normal IO request
736 * that has failed. Thus any sync request that might be pending
737 * will be blocked by nr_pending, and we need to wait for
738 * pending IO requests to complete or be queued for re-try.
739 * Thus the number queued (nr_queued) plus this request (1)
740 * must match the number of pending IOs (nr_pending) before
741 * we continue.
742 */
751 }
753 {
754 /* reverse the effect of the freeze */
760 }
763 /* duplicate the data pages for behind I/O
764 */
766 {
770 GFP_NOIO);
783 }
789 do_sync_io:
795 }
798 {
815 /*
816 * Register the new request and wait if the reconstruction
817 * thread has put up a bar for new requests.
818 * Continue immediately if no resync is active currently.
819 */
826 /* As the suspend_* range is controlled by
827 * userspace, we want an interruptible
828 * wait.
829 */
839 }
841 }
847 /*
848 * make_request() can abort the operation when READA is being
849 * used and no empty request is available.
850 *
851 */
860 /* We might need to issue multiple reads to different
861 * devices if there are bad blocks around, so we keep
862 * track of the number of reads in bio->bi_phys_segments.
863 * If this is 0, there is only one r1_bio and no locking
864 * will be needed when requests complete. If it is
865 * non-zero, then it is the number of not-completed requests.
866 */
871 /*
872 * read balancing logic:
873 */
876 read_again:
880 /* couldn't find anywhere to read from */
883 }
887 bitmap) {
888 /* Reading from a write-mostly device must
889 * take care not to over-take any writes
890 * that are 'behind'
891 */
894 }
899 max_sectors);
910 /* could not read all from this device, so we will
911 * need another r1_bio.
912 */
920 else
923 /* Cannot call generic_make_request directly
924 * as that will be queued in __make_request
925 * and subsequent mempool_alloc might block waiting
926 * for it. So hand bio over to raid1d.
927 */
941 }
943 /*
944 * WRITE:
945 */
946 /* first select target devices under rcu_lock and
947 * inc refcount on their rdev. Record them by setting
948 * bios[x] to bio
949 * If there are known/acknowledged bad blocks on any device on
950 * which we have seen a write error, we want to avoid writing those
951 * blocks.
952 * This potentially requires several writes to write around
953 * the bad blocks. Each set of writes gets it's own r1bio
954 * with a set of bios attached.
955 */
959 retry_write:
969 }
974 }
978 sector_t first_bad;
983 max_sectors,
986 /* mustn't write here until the bad block is
987 * acknowledged*/
991 }
993 /* Cannot write here at all */
996 /* mustn't write more than bad_sectors
997 * to other devices yet
998 */
1001 /* We don't set R1BIO_Degraded as that
1002 * only applies if the disk is
1003 * missing, so it might be re-added,
1004 * and we want to know to recover this
1005 * chunk.
1006 * In this case the device is here,
1007 * and the fact that this chunk is not
1008 * in-sync is recorded in the bad
1009 * block log
1010 */
1012 }
1017 }
1018 }
1020 }
1024 /* Wait for this device to become unblocked */
1035 }
1038 /* We are splitting this write into multiple parts, so
1039 * we need to prepare for allocating another r1_bio.
1040 */
1045 else
1048 }
1064 /* do behind I/O ?
1065 * Not if there are too many, or cannot
1066 * allocate memory, or a reader on WriteMostly
1067 * is waiting for behind writes to flush */
1079 }
1084 /* Yes, I really want the '__' version so that
1085 * we clear any unused pointer in the io_vec, rather
1086 * than leave them unchanged. This is important
1087 * because when we come to free the pages, we won't
1088 * know the original bi_idx, so we just free
1089 * them all
1090 */
1095 }
1110 }
1111 /* Mustn't call r1_bio_write_done before this next test,
1112 * as it could result in the bio being freed.
1113 */
1116 /* We need another r1_bio. It has already been counted
1117 * in bio->bi_phys_segments
1118 */
1126 }
1130 /* In case raid1d snuck in to freeze_array */
1137 }
1140 {
1151 }
1154 }
1158 {
1162 /*
1163 * If it is not operational, then we have already marked it as dead
1164 * else if it is the last working disks, ignore the error, let the
1165 * next level up know.
1166 * else mark the drive as failed
1167 */
1170 /*
1171 * Don't fail the drive, act as though we were just a
1172 * normal single drive.
1173 * However don't try a recovery from this drive as
1174 * it is very likely to fail.
1175 */
1178 }
1186 /*
1187 * if recovery is running, make sure it aborts.
1188 */
1198 }
1201 {
1208 }
1221 }
1223 }
1226 {
1232 }
1235 {
1241 /*
1242 * Find all failed disks within the RAID1 configuration
1243 * and mark them readable.
1244 * Called under mddev lock, so rcu protection not needed.
1245 */
1251 count++;
1253 }
1254 }
1261 }
1265 {
1284 /* as we don't honour merge_bvec_fn, we must
1285 * never risk violating it, so limit
1286 * ->max_segments to one lying with a single
1287 * page, as a one page request is never in
1288 * violation.
1289 */
1294 }
1299 /* As all devices are equivalent, we don't need a full recovery
1300 * if this was recently any drive of the array
1301 */
1306 }
1310 }
1313 {
1326 }
1327 /* Only remove non-faulty devices if recovery
1328 * is not possible.
1329 */
1335 }
1339 /* lost the race, try later */
1343 }
1345 }
1346 abort:
1350 }
1354 {
1363 /*
1364 * we have read a block, now it needs to be re-written,
1365 * or re-read if the read failed.
1366 * We don't do much here, just schedule handling by raid1d
1367 */
1373 }
1376 {
1383 sector_t first_bad;
1390 }
1395 /* make sure these bits doesn't get cleared. */
1413 )
1426 }
1427 }
1428 }
1432 {
1434 /* success */
1438 /* need to record an error - either for the block or the device */
1442 }
1445 {
1446 /* Try some synchronous reads of other devices to get
1447 * good data, much like with normal read errors. Only
1448 * read into the pages we already have so we don't
1449 * need to re-issue the read request.
1450 * We don't need to freeze the array, because being in an
1451 * active sync request, there is no normal IO, and
1452 * no overlapping syncs.
1453 * We don't need to check is_badblock() again as we
1454 * made sure that anything with a bad block in range
1455 * will have bi_end_io clear.
1456 */
1475 /* No rcu protection needed here devices
1476 * can only be removed when no resync is
1477 * active, and resync is currently active
1478 */
1485 }
1486 }
1487 d++;
1495 /* Cannot read from anywhere, this block is lost.
1496 * Record a bad block on each device. If that doesn't
1497 * work just disable and interrupt the recovery.
1498 * Don't fail devices as that won't really help.
1499 */
1511 }
1518 }
1519 /* Try next page */
1522 idx++;
1524 }
1527 /* write it back and re-read */
1531 d--;
1540 }
1541 }
1546 d--;
1554 }
1557 idx ++;
1558 }
1562 }
1565 {
1566 /* We have read all readable devices. If we haven't
1567 * got the block, then there is no hope left.
1568 * If we have, then we want to do a comparison
1569 * and skip the write if everything is the same.
1570 * If any blocks failed to read, then we need to
1571 * attempt an over-write
1572 */
1584 }
1603 PAGE_SIZE))
1605 }
1612 /* No need to write to this device. */
1616 }
1617 /* fixup the bio for reuse */
1635 else
1640 PAGE_SIZE);
1641 }
1642 }
1644 }
1647 {
1656 /* ouch - failed to read all of that. */
1663 /*
1664 * schedule writes
1665 */
1681 }
1684 /* if we're here, all write(s) have completed, so clean up */
1687 }
1688 }
1690 /*
1691 * This is a kernel thread which:
1692 *
1693 * 1. Retries failed read operations on working mirrors.
1694 * 2. Updates the raid superblock when problems encounter.
1695 * 3. Performs writes following reads for array synchronising.
1696 */
1700 {
1713 /* Note: no rcu protection needed here
1714 * as this is synchronous in the raid1d thread
1715 * which is the thread that might remove
1716 * a device. If raid1d ever becomes multi-threaded....
1717 */
1718 sector_t first_bad;
1730 d++;
1733 }
1737 /* Cannot read from anywhere - mark it bad */
1742 }
1743 /* write it back and re-read */
1748 d--;
1754 }
1760 d--;
1768 "md/raid1:%s: read error corrected "
1774 }
1775 }
1776 }
1779 }
1780 }
1783 {
1785 }
1788 {
1799 }
1802 {
1809 /* bio has the data to be written to device 'i' where
1810 * we just recently had a write error.
1811 * We repeatedly clone the bio and trim down to one block,
1812 * then try the write. Where the write fails we record
1813 * a bad block.
1814 * It is conceivable that the bio doesn't exactly align with
1815 * blocks. We must handle this somehow.
1816 *
1817 * We currently own a reference on the rdev.
1818 */
1821 sector_t sector;
1840 idx++;
1845 }
1850 /* Write at 'sector' for 'sectors'*/
1864 /* failure! */
1873 }
1875 }
1878 {
1889 }
1894 }
1895 }
1898 }
1901 {
1911 /* This drive got a write error. We need to
1912 * narrow down and record precise write
1913 * errors.
1914 */
1918 /* an I/O failed, we can't clear the bitmap */
1920 }
1923 }
1927 }
1930 {
1939 /* we got a read error. Maybe the drive is bad. Maybe just
1940 * the block and we can fix it.
1941 * We freeze all other IO, and try reading the block from
1942 * other devices. When we find one, we re-write
1943 * and check it that fixes the read error.
1944 * This is all done synchronously while the array is
1945 * frozen
1946 */
1957 read_more:
1965 const unsigned long do_sync
1971 }
1978 "md/raid1:%s: redirecting sector %llu"
1989 /* Drat - have to split this up more */
1997 else
2016 }
2017 }
2020 {
2039 }
2051 else
2058 else
2059 /* just a partial read to be scheduled from separate
2060 * context
2061 */
2067 }
2069 }
2073 {
2084 }
2086 /*
2087 * perform a "sync" on one "block"
2088 *
2089 * We need to make sure that no normal I/O request - particularly write
2090 * requests - conflict with active sync requests.
2091 *
2092 * This is achieved by tracking pending requests and a 'barrier' concept
2093 * that can be installed to exclude normal IO requests.
2094 */
2097 {
2106 sector_t sync_blocks;
2117 /* If we aborted, we need to abort the
2118 * sync on the 'current' bitmap chunk (there will
2119 * only be one in raid1 resync.
2120 * We can find the current addess in mddev->curr_resync
2121 */
2131 }
2139 }
2140 /* before building a request, check if we can skip these blocks..
2141 * This call the bitmap_start_sync doesn't actually record anything
2142 */
2145 /* We can skip this block, and probably several more */
2148 }
2149 /*
2150 * If there is non-resync activity waiting for a turn,
2151 * and resync is going fast enough,
2152 * then let it though before starting on this new sync request.
2153 */
2164 /*
2165 * If we get a correctably read error during resync or recovery,
2166 * we might want to read from a different device. So we
2167 * flag all drives that could conceivably be read from for READ,
2168 * and any others (which will be non-In_sync devices) for WRITE.
2169 * If a read fails, we try reading from something else for which READ
2170 * is OK.
2171 */
2182 /* take from bio_init */
2202 write_targets ++;
2204 /* may need to read from here */
2217 }
2218 }
2226 }
2229 read_targets++;
2230 }
2231 }
2237 }
2238 }
2245 /* These sectors are bad on all InSync devices, so we
2246 * need to mark them bad on all write targets
2247 */
2256 }
2262 /* Cannot record the badblocks, so need to
2263 * abort the resync.
2264 * If there are multiple read targets, could just
2265 * fail the really bad ones ???
2266 */
2273 }
2275 /* only resync enough to reach the next bad->good
2276 * transition */
2278 }
2281 /* extra read targets are also write targets */
2285 /* There is nowhere to write, so all non-sync
2286 * drives must be failed - so we are finished
2287 */
2292 }
2316 }
2323 /* stop here */
2326 i--;
2330 /* remove last page from this bio */
2334 }
2336 }
2337 }
2338 }
2343 bio_full:
2346 /* For a user-requested sync, we read all readable devices and do a
2347 * compare
2348 */
2356 }
2357 }
2364 }
2366 }
2369 {
2374 }
2377 {
2389 GFP_KERNEL);
2402 r1bio_pool_free,
2420 }
2441 /*
2442 * The first working device is used as a
2443 * starting point to read balancing.
2444 */
2446 }
2453 }
2461 }
2465 abort:
2473 }
2475 }
2478 {
2487 }
2492 }
2493 /*
2494 * copy the already verified devices into our private RAID1
2495 * bookkeeping area. [whatever we allocate in run(),
2496 * should be freed in stop()]
2497 */
2500 else
2511 /* as we don't honour merge_bvec_fn, we must never risk
2512 * violating it, so limit ->max_segments to 1 lying within
2513 * a single page, as a one page request is never in violation.
2514 */
2519 }
2520 }
2541 /*
2542 * Ok, everything is just fine now
2543 */
2553 }
2555 }
2558 {
2562 /* wait for behind writes to complete */
2566 /* need to kick something here to make sure I/O goes? */
2569 }
2582 }
2585 {
2586 /* no resync is happening, and there is enough space
2587 * on all devices, so we can resize.
2588 * We need to make sure resync covers any new space.
2589 * If the array is shrinking we should possibly wait until
2590 * any io in the removed space completes, but it hardly seems
2591 * worth it.
2592 */
2602 }
2606 }
2609 {
2610 /* We need to:
2611 * 1/ resize the r1bio_pool
2612 * 2/ resize conf->mirrors
2613 *
2614 * We allocate a new r1bio_pool if we can.
2615 * Then raise a device barrier and wait until all IO stops.
2616 * Then resize conf->mirrors and swap in the new r1bio pool.
2617 *
2618 * At the same time, we "pack" the devices so that all the missing
2619 * devices have the higher raid_disk numbers.
2620 */
2629 /* Cannot change chunk_size, layout, or level */
2637 }
2649 cnt++;
2652 }
2665 }
2671 }
2675 /* ok, everything is stopped */
2689 }
2692 }
2712 }
2715 {
2728 }
2729 }
2732 {
2733 /* raid1 can take over:
2734 * raid5 with 2 devices, any layout or chunk size
2735 */
2745 }
2747 }
2750 {
2768 };
2771 {
2773 }
2776 {
2778 }