| blob | ede2461e79c51e8d3bd1ac04fcac4c7a974bbbbe |
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 {
290 }
293 {
300 /*
301 * this branch is our 'one mirror IO has finished' event handler:
302 */
308 /* If all other devices have failed, we want to return
309 * the error upwards rather than fail the last device.
310 * Here we redefine "uptodate" to mean "Don't want to retry"
311 */
319 }
324 /*
325 * oops, read error:
326 */
329 KERN_ERR "md/raid1:%s: %s: "
333 b),
337 }
340 }
343 {
344 /* it really is the end of this request */
346 /* free extra copy of the data pages */
352 }
353 /* clear the bitmap if all writes complete successfully */
359 }
362 {
372 else
374 }
375 }
378 {
387 /*
388 * 'one mirror IO has finished' event handler:
389 */
395 /*
396 * Set R1BIO_Uptodate in our master bio, so that we
397 * will return a good error code for to the higher
398 * levels even if IO on some other mirrored buffer
399 * fails.
400 *
401 * The 'master' represents the composite IO operation
402 * to user-side. So if something waits for IO, then it
403 * will wait for the 'master' bio.
404 */
405 sector_t first_bad;
412 /* Maybe we can clear some bad blocks. */
418 }
419 }
425 /*
426 * In behind mode, we ACK the master bio once the I/O
427 * has safely reached all non-writemostly
428 * disks. Setting the Returned bit ensures that this
429 * gets done only once -- we don't ever want to return
430 * -EIO here, instead we'll wait
431 */
434 /* Maybe we can return now */
443 }
444 }
445 }
450 /*
451 * Let's see if all mirrored write operations have finished
452 * already.
453 */
458 }
461 /*
462 * This routine returns the disk from which the requested read should
463 * be done. There is a per-array 'next expected sequential IO' sector
464 * number - if this matches on the next IO then we use the last disk.
465 * There is also a per-disk 'last know head position' sector that is
466 * maintained from IRQ contexts, both the normal and the resync IO
467 * completion handlers update this position correctly. If there is no
468 * perfect sequential match then we pick the disk whose head is closest.
469 *
470 * If there are 2 mirrors in the same 2 devices, performance degrades
471 * because position is mirror, not device based.
472 *
473 * The rdev for the device selected will have nr_pending incremented.
474 */
476 {
483 sector_t best_dist;
488 /*
489 * Check if we can balance. We can balance on the whole
490 * device if no resync is going on, or below the resync window.
491 * We take the first readable disk when above the resync window.
492 */
493 retry:
506 }
509 sector_t dist;
510 sector_t first_bad;
526 /* Don't balance among write-mostly, just
527 * use the first as a last resort */
531 }
532 /* This is a reasonable device to use. It might
533 * even be best.
534 */
538 /* already have a better device */
541 /* cannot read here. If this is the 'primary'
542 * device, then we must not read beyond
543 * bad_sectors from another device..
544 */
556 }
559 }
566 /* Don't change to another disk for sequential reads */
569 /* If device is idle, use it */
573 }
577 }
578 }
586 /* cannot risk returning a device that failed
587 * before we inc'ed nr_pending
588 */
591 }
595 }
600 }
603 {
619 /* Note the '|| 1' - when read_balance prefers
620 * non-congested targets, it can be removed
621 */
624 else
626 }
627 }
630 }
634 {
639 }
642 {
643 /* Any writes that have been queued but are awaiting
644 * bitmap updates get flushed here.
645 */
653 /* flush any pending bitmap writes to
654 * disk before proceeding w/ I/O */
663 }
666 }
668 /* Barriers....
669 * Sometimes we need to suspend IO while we do something else,
670 * either some resync/recovery, or reconfigure the array.
671 * To do this we raise a 'barrier'.
672 * The 'barrier' is a counter that can be raised multiple times
673 * to count how many activities are happening which preclude
674 * normal IO.
675 * We can only raise the barrier if there is no pending IO.
676 * i.e. if nr_pending == 0.
677 * We choose only to raise the barrier if no-one is waiting for the
678 * barrier to go down. This means that as soon as an IO request
679 * is ready, no other operations which require a barrier will start
680 * until the IO request has had a chance.
681 *
682 * So: regular IO calls 'wait_barrier'. When that returns there
683 * is no backgroup IO happening, It must arrange to call
684 * allow_barrier when it has finished its IO.
685 * backgroup IO calls must call raise_barrier. Once that returns
686 * there is no normal IO happeing. It must arrange to call
687 * lower_barrier when the particular background IO completes.
688 */
689 #define RESYNC_DEPTH 32
692 {
695 /* Wait until no block IO is waiting */
699 /* block any new IO from starting */
702 /* Now wait for all pending IO to complete */
708 }
711 {
718 }
721 {
727 );
729 }
732 }
735 {
741 }
744 {
745 /* stop syncio and normal IO and wait for everything to
746 * go quite.
747 * We increment barrier and nr_waiting, and then
748 * wait until nr_pending match nr_queued+1
749 * This is called in the context of one normal IO request
750 * that has failed. Thus any sync request that might be pending
751 * will be blocked by nr_pending, and we need to wait for
752 * pending IO requests to complete or be queued for re-try.
753 * Thus the number queued (nr_queued) plus this request (1)
754 * must match the number of pending IOs (nr_pending) before
755 * we continue.
756 */
765 }
767 {
768 /* reverse the effect of the freeze */
774 }
777 /* duplicate the data pages for behind I/O
778 */
780 {
784 GFP_NOIO);
797 }
803 do_sync_io:
809 }
812 {
829 /*
830 * Register the new request and wait if the reconstruction
831 * thread has put up a bar for new requests.
832 * Continue immediately if no resync is active currently.
833 */
840 /* As the suspend_* range is controlled by
841 * userspace, we want an interruptible
842 * wait.
843 */
853 }
855 }
861 /*
862 * make_request() can abort the operation when READA is being
863 * used and no empty request is available.
864 *
865 */
874 /* We might need to issue multiple reads to different
875 * devices if there are bad blocks around, so we keep
876 * track of the number of reads in bio->bi_phys_segments.
877 * If this is 0, there is only one r1_bio and no locking
878 * will be needed when requests complete. If it is
879 * non-zero, then it is the number of not-completed requests.
880 */
885 /*
886 * read balancing logic:
887 */
890 read_again:
894 /* couldn't find anywhere to read from */
897 }
901 bitmap) {
902 /* Reading from a write-mostly device must
903 * take care not to over-take any writes
904 * that are 'behind'
905 */
908 }
913 max_sectors);
924 /* could not read all from this device, so we will
925 * need another r1_bio.
926 */
934 else
937 /* Cannot call generic_make_request directly
938 * as that will be queued in __make_request
939 * and subsequent mempool_alloc might block waiting
940 * for it. So hand bio over to raid1d.
941 */
955 }
957 /*
958 * WRITE:
959 */
964 }
965 /* first select target devices under rcu_lock and
966 * inc refcount on their rdev. Record them by setting
967 * bios[x] to bio
968 * If there are known/acknowledged bad blocks on any device on
969 * which we have seen a write error, we want to avoid writing those
970 * blocks.
971 * This potentially requires several writes to write around
972 * the bad blocks. Each set of writes gets it's own r1bio
973 * with a set of bios attached.
974 */
978 retry_write:
988 }
993 }
997 sector_t first_bad;
1002 max_sectors,
1005 /* mustn't write here until the bad block is
1006 * acknowledged*/
1010 }
1012 /* Cannot write here at all */
1015 /* mustn't write more than bad_sectors
1016 * to other devices yet
1017 */
1020 /* We don't set R1BIO_Degraded as that
1021 * only applies if the disk is
1022 * missing, so it might be re-added,
1023 * and we want to know to recover this
1024 * chunk.
1025 * In this case the device is here,
1026 * and the fact that this chunk is not
1027 * in-sync is recorded in the bad
1028 * block log
1029 */
1031 }
1036 }
1037 }
1039 }
1043 /* Wait for this device to become unblocked */
1054 }
1057 /* We are splitting this write into multiple parts, so
1058 * we need to prepare for allocating another r1_bio.
1059 */
1064 else
1067 }
1083 /* do behind I/O ?
1084 * Not if there are too many, or cannot
1085 * allocate memory, or a reader on WriteMostly
1086 * is waiting for behind writes to flush */
1098 }
1103 /* Yes, I really want the '__' version so that
1104 * we clear any unused pointer in the io_vec, rather
1105 * than leave them unchanged. This is important
1106 * because when we come to free the pages, we won't
1107 * know the original bi_idx, so we just free
1108 * them all
1109 */
1114 }
1130 }
1131 /* Mustn't call r1_bio_write_done before this next test,
1132 * as it could result in the bio being freed.
1133 */
1136 /* We need another r1_bio. It has already been counted
1137 * in bio->bi_phys_segments
1138 */
1146 }
1150 /* In case raid1d snuck in to freeze_array */
1155 }
1158 {
1169 }
1172 }
1176 {
1180 /*
1181 * If it is not operational, then we have already marked it as dead
1182 * else if it is the last working disks, ignore the error, let the
1183 * next level up know.
1184 * else mark the drive as failed
1185 */
1188 /*
1189 * Don't fail the drive, act as though we were just a
1190 * normal single drive.
1191 * However don't try a recovery from this drive as
1192 * it is very likely to fail.
1193 */
1196 }
1204 /*
1205 * if recovery is running, make sure it aborts.
1206 */
1216 }
1219 {
1226 }
1239 }
1241 }
1244 {
1250 }
1253 {
1259 /*
1260 * Find all failed disks within the RAID1 configuration
1261 * and mark them readable.
1262 * Called under mddev lock, so rcu protection not needed.
1263 */
1269 count++;
1271 }
1272 }
1279 }
1283 {
1302 /* as we don't honour merge_bvec_fn, we must
1303 * never risk violating it, so limit
1304 * ->max_segments to one lying with a single
1305 * page, as a one page request is never in
1306 * violation.
1307 */
1312 }
1317 /* As all devices are equivalent, we don't need a full recovery
1318 * if this was recently any drive of the array
1319 */
1324 }
1328 }
1331 {
1344 }
1345 /* Only remove non-faulty devices if recovery
1346 * is not possible.
1347 */
1353 }
1357 /* lost the race, try later */
1361 }
1363 }
1364 abort:
1368 }
1372 {
1377 /*
1378 * we have read a block, now it needs to be re-written,
1379 * or re-read if the read failed.
1380 * We don't do much here, just schedule handling by raid1d
1381 */
1387 }
1390 {
1396 sector_t first_bad;
1405 /* make sure these bits doesn't get cleared. */
1423 )
1434 }
1435 }
1436 }
1440 {
1442 /* success */
1446 /* need to record an error - either for the block or the device */
1450 }
1453 {
1454 /* Try some synchronous reads of other devices to get
1455 * good data, much like with normal read errors. Only
1456 * read into the pages we already have so we don't
1457 * need to re-issue the read request.
1458 * We don't need to freeze the array, because being in an
1459 * active sync request, there is no normal IO, and
1460 * no overlapping syncs.
1461 * We don't need to check is_badblock() again as we
1462 * made sure that anything with a bad block in range
1463 * will have bi_end_io clear.
1464 */
1483 /* No rcu protection needed here devices
1484 * can only be removed when no resync is
1485 * active, and resync is currently active
1486 */
1493 }
1494 }
1495 d++;
1503 /* Cannot read from anywhere, this block is lost.
1504 * Record a bad block on each device. If that doesn't
1505 * work just disable and interrupt the recovery.
1506 * Don't fail devices as that won't really help.
1507 */
1519 }
1527 }
1528 /* Try next page */
1531 idx++;
1533 }
1536 /* write it back and re-read */
1540 d--;
1549 }
1550 }
1555 d--;
1563 }
1566 idx ++;
1567 }
1571 }
1574 {
1575 /* We have read all readable devices. If we haven't
1576 * got the block, then there is no hope left.
1577 * If we have, then we want to do a comparison
1578 * and skip the write if everything is the same.
1579 * If any blocks failed to read, then we need to
1580 * attempt an over-write
1581 */
1593 }
1612 PAGE_SIZE))
1614 }
1621 /* No need to write to this device. */
1625 }
1626 /* fixup the bio for reuse */
1644 else
1649 PAGE_SIZE);
1650 }
1651 }
1653 }
1656 {
1665 /* ouch - failed to read all of that. */
1672 /*
1673 * schedule writes
1674 */
1690 }
1693 /* if we're here, all write(s) have completed, so clean up */
1696 }
1697 }
1699 /*
1700 * This is a kernel thread which:
1701 *
1702 * 1. Retries failed read operations on working mirrors.
1703 * 2. Updates the raid superblock when problems encounter.
1704 * 3. Performs writes following reads for array synchronising.
1705 */
1709 {
1722 /* Note: no rcu protection needed here
1723 * as this is synchronous in the raid1d thread
1724 * which is the thread that might remove
1725 * a device. If raid1d ever becomes multi-threaded....
1726 */
1727 sector_t first_bad;
1739 d++;
1742 }
1746 /* Cannot read from anywhere - mark it bad */
1751 }
1752 /* write it back and re-read */
1757 d--;
1763 }
1769 d--;
1777 "md/raid1:%s: read error corrected "
1783 }
1784 }
1785 }
1788 }
1789 }
1792 {
1794 }
1797 {
1808 }
1811 {
1818 /* bio has the data to be written to device 'i' where
1819 * we just recently had a write error.
1820 * We repeatedly clone the bio and trim down to one block,
1821 * then try the write. Where the write fails we record
1822 * a bad block.
1823 * It is conceivable that the bio doesn't exactly align with
1824 * blocks. We must handle this somehow.
1825 *
1826 * We currently own a reference on the rdev.
1827 */
1830 sector_t sector;
1849 idx++;
1854 }
1859 /* Write at 'sector' for 'sectors'*/
1873 /* failure! */
1882 }
1884 }
1887 {
1898 }
1903 }
1904 }
1907 }
1910 {
1920 /* This drive got a write error. We need to
1921 * narrow down and record precise write
1922 * errors.
1923 */
1927 /* an I/O failed, we can't clear the bitmap */
1929 }
1932 }
1936 }
1939 {
1948 /* we got a read error. Maybe the drive is bad. Maybe just
1949 * the block and we can fix it.
1950 * We freeze all other IO, and try reading the block from
1951 * other devices. When we find one, we re-write
1952 * and check it that fixes the read error.
1953 * This is all done synchronously while the array is
1954 * frozen
1955 */
1966 read_more:
1974 const unsigned long do_sync
1980 }
1987 "md/raid1:%s: redirecting sector %llu"
1998 /* Drat - have to split this up more */
2006 else
2025 }
2026 }
2029 {
2048 }
2060 else
2067 else
2068 /* just a partial read to be scheduled from separate
2069 * context
2070 */
2076 }
2078 }
2082 {
2093 }
2095 /*
2096 * perform a "sync" on one "block"
2097 *
2098 * We need to make sure that no normal I/O request - particularly write
2099 * requests - conflict with active sync requests.
2100 *
2101 * This is achieved by tracking pending requests and a 'barrier' concept
2102 * that can be installed to exclude normal IO requests.
2103 */
2105 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2106 {
2115 sector_t sync_blocks;
2126 /* If we aborted, we need to abort the
2127 * sync on the 'current' bitmap chunk (there will
2128 * only be one in raid1 resync.
2129 * We can find the current addess in mddev->curr_resync
2130 */
2140 }
2148 }
2149 /* before building a request, check if we can skip these blocks..
2150 * This call the bitmap_start_sync doesn't actually record anything
2151 */
2154 /* We can skip this block, and probably several more */
2157 }
2158 /*
2159 * If there is non-resync activity waiting for a turn,
2160 * and resync is going fast enough,
2161 * then let it though before starting on this new sync request.
2162 */
2173 /*
2174 * If we get a correctably read error during resync or recovery,
2175 * we might want to read from a different device. So we
2176 * flag all drives that could conceivably be read from for READ,
2177 * and any others (which will be non-In_sync devices) for WRITE.
2178 * If a read fails, we try reading from something else for which READ
2179 * is OK.
2180 */
2191 /* take from bio_init */
2210 write_targets ++;
2212 /* may need to read from here */
2225 }
2226 }
2234 }
2237 read_targets++;
2238 }
2239 }
2245 }
2246 }
2253 /* These sectors are bad on all InSync devices, so we
2254 * need to mark them bad on all write targets
2255 */
2264 }
2270 /* Cannot record the badblocks, so need to
2271 * abort the resync.
2272 * If there are multiple read targets, could just
2273 * fail the really bad ones ???
2274 */
2281 }
2283 /* only resync enough to reach the next bad->good
2284 * transition */
2286 }
2289 /* extra read targets are also write targets */
2293 /* There is nowhere to write, so all non-sync
2294 * drives must be failed - so we are finished
2295 */
2300 }
2324 }
2331 /* stop here */
2334 i--;
2338 /* remove last page from this bio */
2342 }
2344 }
2345 }
2346 }
2351 bio_full:
2354 /* For a user-requested sync, we read all readable devices and do a
2355 * compare
2356 */
2364 }
2365 }
2372 }
2374 }
2377 {
2382 }
2385 {
2397 GFP_KERNEL);
2410 r1bio_pool_free,
2428 }
2451 /*
2452 * The first working device is used as a
2453 * starting point to read balancing.
2454 */
2456 }
2463 }
2471 }
2475 abort:
2483 }
2485 }
2488 {
2497 }
2502 }
2503 /*
2504 * copy the already verified devices into our private RAID1
2505 * bookkeeping area. [whatever we allocate in run(),
2506 * should be freed in stop()]
2507 */
2510 else
2521 /* as we don't honour merge_bvec_fn, we must never risk
2522 * violating it, so limit ->max_segments to 1 lying within
2523 * a single page, as a one page request is never in violation.
2524 */
2529 }
2530 }
2551 /*
2552 * Ok, everything is just fine now
2553 */
2563 }
2565 }
2568 {
2572 /* wait for behind writes to complete */
2576 /* need to kick something here to make sure I/O goes? */
2579 }
2592 }
2595 {
2596 /* no resync is happening, and there is enough space
2597 * on all devices, so we can resize.
2598 * We need to make sure resync covers any new space.
2599 * If the array is shrinking we should possibly wait until
2600 * any io in the removed space completes, but it hardly seems
2601 * worth it.
2602 */
2612 }
2616 }
2619 {
2620 /* We need to:
2621 * 1/ resize the r1bio_pool
2622 * 2/ resize conf->mirrors
2623 *
2624 * We allocate a new r1bio_pool if we can.
2625 * Then raise a device barrier and wait until all IO stops.
2626 * Then resize conf->mirrors and swap in the new r1bio pool.
2627 *
2628 * At the same time, we "pack" the devices so that all the missing
2629 * devices have the higher raid_disk numbers.
2630 */
2639 /* Cannot change chunk_size, layout, or level */
2647 }
2659 cnt++;
2662 }
2675 }
2681 }
2685 /* ok, everything is stopped */
2699 }
2702 }
2722 }
2725 {
2738 }
2739 }
2742 {
2743 /* raid1 can take over:
2744 * raid5 with 2 devices, any layout or chunk size
2745 */
2755 }
2757 }
2760 {
2778 };
2781 {
2783 }
2786 {
2788 }