| blob | 4d40d9d54a20c151671b11ecc9004a64e7521504 |
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 {
324 {
325 /* it really is the end of this request */
327 /* free extra copy of the data pages */
333 }
334 /* clear the bitmap if all writes complete successfully */
341 }
342 }
345 {
357 /*
358 * 'one mirror IO has finished' event handler:
359 */
364 /* an I/O failed, we can't clear the bitmap */
367 /*
368 * Set R1BIO_Uptodate in our master bio, so that we
369 * will return a good error code for to the higher
370 * levels even if IO on some other mirrored buffer
371 * fails.
372 *
373 * The 'master' represents the composite IO operation
374 * to user-side. So if something waits for IO, then it
375 * will wait for the 'master' bio.
376 */
385 /*
386 * In behind mode, we ACK the master bio once the I/O
387 * has safely reached all non-writemostly
388 * disks. Setting the Returned bit ensures that this
389 * gets done only once -- we don't ever want to return
390 * -EIO here, instead we'll wait
391 */
394 /* Maybe we can return now */
402 }
403 }
404 }
407 /*
408 * Let's see if all mirrored write operations have finished
409 * already.
410 */
415 }
418 /*
419 * This routine returns the disk from which the requested read should
420 * be done. There is a per-array 'next expected sequential IO' sector
421 * number - if this matches on the next IO then we use the last disk.
422 * There is also a per-disk 'last know head position' sector that is
423 * maintained from IRQ contexts, both the normal and the resync IO
424 * completion handlers update this position correctly. If there is no
425 * perfect sequential match then we pick the disk whose head is closest.
426 *
427 * If there are 2 mirrors in the same 2 devices, performance degrades
428 * because position is mirror, not device based.
429 *
430 * The rdev for the device selected will have nr_pending incremented.
431 */
433 {
440 sector_t best_dist;
445 /*
446 * Check if we can balance. We can balance on the whole
447 * device if no resync is going on, or below the resync window.
448 * We take the first readable disk when above the resync window.
449 */
450 retry:
463 }
466 sector_t dist;
467 sector_t first_bad;
483 /* Don't balance among write-mostly, just
484 * use the first as a last resort */
488 }
489 /* This is a reasonable device to use. It might
490 * even be best.
491 */
495 /* already have a better device */
498 /* cannot read here. If this is the 'primary'
499 * device, then we must not read beyond
500 * bad_sectors from another device..
501 */
513 }
516 }
523 /* Don't change to another disk for sequential reads */
526 /* If device is idle, use it */
530 }
534 }
535 }
543 /* cannot risk returning a device that failed
544 * before we inc'ed nr_pending
545 */
548 }
552 }
557 }
560 {
572 /* Note the '|| 1' - when read_balance prefers
573 * non-congested targets, it can be removed
574 */
577 else
579 }
580 }
583 }
587 {
592 }
595 {
596 /* Any writes that have been queued but are awaiting
597 * bitmap updates get flushed here.
598 */
605 /* flush any pending bitmap writes to
606 * disk before proceeding w/ I/O */
614 }
617 }
619 /* Barriers....
620 * Sometimes we need to suspend IO while we do something else,
621 * either some resync/recovery, or reconfigure the array.
622 * To do this we raise a 'barrier'.
623 * The 'barrier' is a counter that can be raised multiple times
624 * to count how many activities are happening which preclude
625 * normal IO.
626 * We can only raise the barrier if there is no pending IO.
627 * i.e. if nr_pending == 0.
628 * We choose only to raise the barrier if no-one is waiting for the
629 * barrier to go down. This means that as soon as an IO request
630 * is ready, no other operations which require a barrier will start
631 * until the IO request has had a chance.
632 *
633 * So: regular IO calls 'wait_barrier'. When that returns there
634 * is no backgroup IO happening, It must arrange to call
635 * allow_barrier when it has finished its IO.
636 * backgroup IO calls must call raise_barrier. Once that returns
637 * there is no normal IO happeing. It must arrange to call
638 * lower_barrier when the particular background IO completes.
639 */
640 #define RESYNC_DEPTH 32
643 {
646 /* Wait until no block IO is waiting */
650 /* block any new IO from starting */
653 /* Now wait for all pending IO to complete */
659 }
662 {
669 }
672 {
678 );
680 }
683 }
686 {
692 }
695 {
696 /* stop syncio and normal IO and wait for everything to
697 * go quite.
698 * We increment barrier and nr_waiting, and then
699 * wait until nr_pending match nr_queued+1
700 * This is called in the context of one normal IO request
701 * that has failed. Thus any sync request that might be pending
702 * will be blocked by nr_pending, and we need to wait for
703 * pending IO requests to complete or be queued for re-try.
704 * Thus the number queued (nr_queued) plus this request (1)
705 * must match the number of pending IOs (nr_pending) before
706 * we continue.
707 */
716 }
718 {
719 /* reverse the effect of the freeze */
725 }
728 /* duplicate the data pages for behind I/O
729 */
731 {
735 GFP_NOIO);
747 }
753 do_sync_io:
759 }
762 {
776 /*
777 * Register the new request and wait if the reconstruction
778 * thread has put up a bar for new requests.
779 * Continue immediately if no resync is active currently.
780 */
787 /* As the suspend_* range is controlled by
788 * userspace, we want an interruptible
789 * wait.
790 */
800 }
802 }
808 /*
809 * make_request() can abort the operation when READA is being
810 * used and no empty request is available.
811 *
812 */
821 /* We might need to issue multiple reads to different
822 * devices if there are bad blocks around, so we keep
823 * track of the number of reads in bio->bi_phys_segments.
824 * If this is 0, there is only one r1_bio and no locking
825 * will be needed when requests complete. If it is
826 * non-zero, then it is the number of not-completed requests.
827 */
832 /*
833 * read balancing logic:
834 */
838 read_again:
842 /* couldn't find anywhere to read from */
845 }
849 bitmap) {
850 /* Reading from a write-mostly device must
851 * take care not to over-take any writes
852 * that are 'behind'
853 */
856 }
861 max_sectors);
872 /* could not read all from this device, so we will
873 * need another r1_bio.
874 */
883 else
886 /* Cannot call generic_make_request directly
887 * as that will be queued in __make_request
888 * and subsequent mempool_alloc might block waiting
889 * for it. So hand bio over to raid1d.
890 */
904 }
906 /*
907 * WRITE:
908 */
909 /* first select target devices under spinlock and
910 * inc refcount on their rdev. Record them by setting
911 * bios[x] to bio
912 */
916 retry_write:
925 }
933 targets++;
934 }
937 }
941 /* Wait for this device to become unblocked */
952 }
955 /* array is degraded, we will not clear the bitmap
956 * on I/O completion (see raid1_end_write_request) */
958 }
960 /* do behind I/O ?
961 * Not if there are too many, or cannot allocate memory,
962 * or a reader on WriteMostly is waiting for behind writes
963 * to flush */
993 /* Yes, I really want the '__' version so that
994 * we clear any unused pointer in the io_vec, rather
995 * than leave them unchanged. This is important
996 * because when we come to free the pages, we won't
997 * know the original bi_idx, so we just free
998 * them all
999 */
1004 }
1010 }
1013 /* In case raid1d snuck in to freeze_array */
1020 }
1023 {
1034 }
1037 }
1041 {
1045 /*
1046 * If it is not operational, then we have already marked it as dead
1047 * else if it is the last working disks, ignore the error, let the
1048 * next level up know.
1049 * else mark the drive as failed
1050 */
1053 /*
1054 * Don't fail the drive, act as though we were just a
1055 * normal single drive.
1056 * However don't try a recovery from this drive as
1057 * it is very likely to fail.
1058 */
1061 }
1069 /*
1070 * if recovery is running, make sure it aborts.
1071 */
1081 }
1084 {
1091 }
1104 }
1106 }
1109 {
1115 }
1118 {
1124 /*
1125 * Find all failed disks within the RAID1 configuration
1126 * and mark them readable.
1127 * Called under mddev lock, so rcu protection not needed.
1128 */
1134 count++;
1136 }
1137 }
1144 }
1148 {
1167 /* as we don't honour merge_bvec_fn, we must
1168 * never risk violating it, so limit
1169 * ->max_segments to one lying with a single
1170 * page, as a one page request is never in
1171 * violation.
1172 */
1177 }
1182 /* As all devices are equivalent, we don't need a full recovery
1183 * if this was recently any drive of the array
1184 */
1189 }
1193 }
1196 {
1209 }
1210 /* Only remove non-faulty devices if recovery
1211 * is not possible.
1212 */
1218 }
1222 /* lost the race, try later */
1226 }
1228 }
1229 abort:
1233 }
1237 {
1246 /*
1247 * we have read a block, now it needs to be re-written,
1248 * or re-read if the read failed.
1249 * We don't do much here, just schedule handling by raid1d
1250 */
1256 }
1259 {
1271 }
1276 /* make sure these bits doesn't get cleared. */
1284 }
1292 }
1293 }
1296 {
1297 /* Try some synchronous reads of other devices to get
1298 * good data, much like with normal read errors. Only
1299 * read into the pages we already have so we don't
1300 * need to re-issue the read request.
1301 * We don't need to freeze the array, because being in an
1302 * active sync request, there is no normal IO, and
1303 * no overlapping syncs.
1304 * We don't need to check is_badblock() again as we
1305 * made sure that anything with a bad block in range
1306 * will have bi_end_io clear.
1307 */
1326 /* No rcu protection needed here devices
1327 * can only be removed when no resync is
1328 * active, and resync is currently active
1329 */
1336 }
1337 }
1338 d++;
1345 /* Cannot read from anywhere, array is toast */
1355 }
1358 /* write it back and re-read */
1362 d--;
1372 }
1373 }
1378 d--;
1386 else
1388 }
1391 idx ++;
1392 }
1396 }
1399 {
1400 /* We have read all readable devices. If we haven't
1401 * got the block, then there is no hope left.
1402 * If we have, then we want to do a comparison
1403 * and skip the write if everything is the same.
1404 * If any blocks failed to read, then we need to
1405 * attempt an over-write
1406 */
1418 }
1437 PAGE_SIZE))
1439 }
1446 /* No need to write to this device. */
1450 }
1451 /* fixup the bio for reuse */
1469 else
1474 PAGE_SIZE);
1475 }
1476 }
1478 }
1481 {
1490 /* ouch - failed to read all of that. */
1497 /*
1498 * schedule writes
1499 */
1515 }
1518 /* if we're here, all write(s) have completed, so clean up */
1521 }
1522 }
1524 /*
1525 * This is a kernel thread which:
1526 *
1527 * 1. Retries failed read operations on working mirrors.
1528 * 2. Updates the raid superblock when problems encounter.
1529 * 3. Performs writes following reads for array synchronising.
1530 */
1534 {
1547 /* Note: no rcu protection needed here
1548 * as this is synchronous in the raid1d thread
1549 * which is the thread that might remove
1550 * a device. If raid1d ever becomes multi-threaded....
1551 */
1552 sector_t first_bad;
1564 d++;
1567 }
1571 /* Cannot read from anywhere -- bye bye array */
1574 }
1575 /* write it back and re-read */
1580 d--;
1587 /* Well, this device is dead */
1589 }
1590 }
1596 d--;
1603 /* Well, this device is dead */
1608 "md/raid1:%s: read error corrected "
1614 }
1615 }
1616 }
1619 }
1620 }
1623 {
1645 }
1660 /* we got a read error. Maybe the drive is bad. Maybe just
1661 * the block and we can fix it.
1662 * We freeze all other IO, and try reading the block from
1663 * other devices. When we find one, we re-write
1664 * and check it that fixes the read error.
1665 * This is all done synchronously while the array is
1666 * frozen
1667 */
1680 read_more:
1694 }
1700 max_sectors);
1704 KERN_ERR
1705 "md/raid1:%s: redirecting sector %llu"
1716 /* Drat - have to split this up more */
1725 else
1732 GFP_NOIO);
1747 }
1749 /* just a partial read to be scheduled from separate
1750 * context
1751 */
1753 }
1757 }
1759 }
1763 {
1774 }
1776 /*
1777 * perform a "sync" on one "block"
1778 *
1779 * We need to make sure that no normal I/O request - particularly write
1780 * requests - conflict with active sync requests.
1781 *
1782 * This is achieved by tracking pending requests and a 'barrier' concept
1783 * that can be installed to exclude normal IO requests.
1784 */
1787 {
1796 sector_t sync_blocks;
1807 /* If we aborted, we need to abort the
1808 * sync on the 'current' bitmap chunk (there will
1809 * only be one in raid1 resync.
1810 * We can find the current addess in mddev->curr_resync
1811 */
1821 }
1829 }
1830 /* before building a request, check if we can skip these blocks..
1831 * This call the bitmap_start_sync doesn't actually record anything
1832 */
1835 /* We can skip this block, and probably several more */
1838 }
1839 /*
1840 * If there is non-resync activity waiting for a turn,
1841 * and resync is going fast enough,
1842 * then let it though before starting on this new sync request.
1843 */
1854 /*
1855 * If we get a correctably read error during resync or recovery,
1856 * we might want to read from a different device. So we
1857 * flag all drives that could conceivably be read from for READ,
1858 * and any others (which will be non-In_sync devices) for WRITE.
1859 * If a read fails, we try reading from something else for which READ
1860 * is OK.
1861 */
1872 /* take from bio_init */
1892 write_targets ++;
1894 /* may need to read from here */
1907 }
1908 }
1916 }
1919 read_targets++;
1920 }
1921 }
1927 }
1928 }
1935 /* These sectors are bad on all InSync devices, so we
1936 * need to mark them bad on all write targets
1937 */
1946 }
1952 /* Cannot record the badblocks, so need to
1953 * abort the resync.
1954 * If there are multiple read targets, could just
1955 * fail the really bad ones ???
1956 */
1963 }
1965 /* only resync enough to reach the next bad->good
1966 * transition */
1968 }
1971 /* extra read targets are also write targets */
1975 /* There is nowhere to write, so all non-sync
1976 * drives must be failed - so we are finished
1977 */
1982 }
2006 }
2013 /* stop here */
2016 i--;
2020 /* remove last page from this bio */
2024 }
2026 }
2027 }
2028 }
2033 bio_full:
2036 /* For a user-requested sync, we read all readable devices and do a
2037 * compare
2038 */
2046 }
2047 }
2054 }
2056 }
2059 {
2064 }
2067 {
2079 GFP_KERNEL);
2092 r1bio_pool_free,
2110 }
2131 /*
2132 * The first working device is used as a
2133 * starting point to read balancing.
2134 */
2136 }
2143 }
2151 }
2155 abort:
2163 }
2165 }
2168 {
2177 }
2182 }
2183 /*
2184 * copy the already verified devices into our private RAID1
2185 * bookkeeping area. [whatever we allocate in run(),
2186 * should be freed in stop()]
2187 */
2190 else
2201 /* as we don't honour merge_bvec_fn, we must never risk
2202 * violating it, so limit ->max_segments to 1 lying within
2203 * a single page, as a one page request is never in violation.
2204 */
2209 }
2210 }
2231 /*
2232 * Ok, everything is just fine now
2233 */
2243 }
2245 }
2248 {
2252 /* wait for behind writes to complete */
2256 /* need to kick something here to make sure I/O goes? */
2259 }
2273 }
2276 {
2277 /* no resync is happening, and there is enough space
2278 * on all devices, so we can resize.
2279 * We need to make sure resync covers any new space.
2280 * If the array is shrinking we should possibly wait until
2281 * any io in the removed space completes, but it hardly seems
2282 * worth it.
2283 */
2293 }
2297 }
2300 {
2301 /* We need to:
2302 * 1/ resize the r1bio_pool
2303 * 2/ resize conf->mirrors
2304 *
2305 * We allocate a new r1bio_pool if we can.
2306 * Then raise a device barrier and wait until all IO stops.
2307 * Then resize conf->mirrors and swap in the new r1bio pool.
2308 *
2309 * At the same time, we "pack" the devices so that all the missing
2310 * devices have the higher raid_disk numbers.
2311 */
2320 /* Cannot change chunk_size, layout, or level */
2328 }
2340 cnt++;
2343 }
2356 }
2362 }
2366 /* ok, everything is stopped */
2380 }
2383 }
2403 }
2406 {
2419 }
2420 }
2423 {
2424 /* raid1 can take over:
2425 * raid5 with 2 devices, any layout or chunk size
2426 */
2436 }
2438 }
2441 {
2459 };
2462 {
2464 }
2467 {
2469 }