Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * raid1.c : Multiple Devices driver for Linux
   4 *
   5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6 *
   7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8 *
   9 * RAID-1 management functions.
  10 *
  11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  12 *
  13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  15 *
  16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  17 * bitmapped intelligence in resync:
  18 *
  19 *      - bitmap marked during normal i/o
  20 *      - bitmap used to skip nondirty blocks during sync
  21 *
  22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  23 * - persistent bitmap code
 
 
 
 
 
 
 
 
 
  24 */
  25
  26#include <linux/slab.h>
  27#include <linux/delay.h>
  28#include <linux/blkdev.h>
  29#include <linux/module.h>
  30#include <linux/seq_file.h>
  31#include <linux/ratelimit.h>
  32
  33#include <trace/events/block.h>
  34
  35#include "md.h"
  36#include "raid1.h"
  37#include "md-bitmap.h"
  38
  39#define UNSUPPORTED_MDDEV_FLAGS		\
  40	((1L << MD_HAS_JOURNAL) |	\
  41	 (1L << MD_JOURNAL_CLEAN) |	\
  42	 (1L << MD_HAS_PPL) |		\
  43	 (1L << MD_HAS_MULTIPLE_PPLS))
  44
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  45static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  46static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  47
  48#define raid1_log(md, fmt, args...)				\
  49	do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  50
  51#include "raid1-10.c"
  52
  53static int check_and_add_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
  54{
  55	struct wb_info *wi, *temp_wi;
  56	unsigned long flags;
  57	int ret = 0;
  58	struct mddev *mddev = rdev->mddev;
  59
  60	wi = mempool_alloc(mddev->wb_info_pool, GFP_NOIO);
  61
  62	spin_lock_irqsave(&rdev->wb_list_lock, flags);
  63	list_for_each_entry(temp_wi, &rdev->wb_list, list) {
  64		/* collision happened */
  65		if (hi > temp_wi->lo && lo < temp_wi->hi) {
  66			ret = -EBUSY;
  67			break;
  68		}
  69	}
  70
  71	if (!ret) {
  72		wi->lo = lo;
  73		wi->hi = hi;
  74		list_add(&wi->list, &rdev->wb_list);
  75	} else
  76		mempool_free(wi, mddev->wb_info_pool);
  77	spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
  78
  79	return ret;
  80}
  81
  82static void remove_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
  83{
  84	struct wb_info *wi;
  85	unsigned long flags;
  86	int found = 0;
  87	struct mddev *mddev = rdev->mddev;
  88
  89	spin_lock_irqsave(&rdev->wb_list_lock, flags);
  90	list_for_each_entry(wi, &rdev->wb_list, list)
  91		if (hi == wi->hi && lo == wi->lo) {
  92			list_del(&wi->list);
  93			mempool_free(wi, mddev->wb_info_pool);
  94			found = 1;
  95			break;
  96		}
  97
  98	if (!found)
  99		WARN(1, "The write behind IO is not recorded\n");
 100	spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
 101	wake_up(&rdev->wb_io_wait);
 102}
 103
 104/*
 105 * for resync bio, r1bio pointer can be retrieved from the per-bio
 106 * 'struct resync_pages'.
 107 */
 108static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 109{
 110	return get_resync_pages(bio)->raid_bio;
 111}
 112
 113static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 114{
 115	struct pool_info *pi = data;
 116	int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 117
 118	/* allocate a r1bio with room for raid_disks entries in the bios array */
 119	return kzalloc(size, gfp_flags);
 120}
 121
 
 
 
 
 
 122#define RESYNC_DEPTH 32
 123#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 124#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 125#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 126#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 127#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 128
 129static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 130{
 131	struct pool_info *pi = data;
 132	struct r1bio *r1_bio;
 133	struct bio *bio;
 134	int need_pages;
 135	int j;
 136	struct resync_pages *rps;
 137
 138	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 139	if (!r1_bio)
 140		return NULL;
 141
 142	rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 143			    gfp_flags);
 144	if (!rps)
 145		goto out_free_r1bio;
 146
 147	/*
 148	 * Allocate bios : 1 for reading, n-1 for writing
 149	 */
 150	for (j = pi->raid_disks ; j-- ; ) {
 151		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 152		if (!bio)
 153			goto out_free_bio;
 154		r1_bio->bios[j] = bio;
 155	}
 156	/*
 157	 * Allocate RESYNC_PAGES data pages and attach them to
 158	 * the first bio.
 159	 * If this is a user-requested check/repair, allocate
 160	 * RESYNC_PAGES for each bio.
 161	 */
 162	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 163		need_pages = pi->raid_disks;
 164	else
 165		need_pages = 1;
 166	for (j = 0; j < pi->raid_disks; j++) {
 167		struct resync_pages *rp = &rps[j];
 168
 169		bio = r1_bio->bios[j];
 170
 171		if (j < need_pages) {
 172			if (resync_alloc_pages(rp, gfp_flags))
 173				goto out_free_pages;
 174		} else {
 175			memcpy(rp, &rps[0], sizeof(*rp));
 176			resync_get_all_pages(rp);
 177		}
 178
 179		rp->raid_bio = r1_bio;
 180		bio->bi_private = rp;
 181	}
 182
 183	r1_bio->master_bio = NULL;
 184
 185	return r1_bio;
 186
 187out_free_pages:
 188	while (--j >= 0)
 189		resync_free_pages(&rps[j]);
 190
 191out_free_bio:
 192	while (++j < pi->raid_disks)
 193		bio_put(r1_bio->bios[j]);
 194	kfree(rps);
 195
 196out_free_r1bio:
 197	rbio_pool_free(r1_bio, data);
 198	return NULL;
 199}
 200
 201static void r1buf_pool_free(void *__r1_bio, void *data)
 202{
 203	struct pool_info *pi = data;
 204	int i;
 205	struct r1bio *r1bio = __r1_bio;
 206	struct resync_pages *rp = NULL;
 207
 208	for (i = pi->raid_disks; i--; ) {
 209		rp = get_resync_pages(r1bio->bios[i]);
 210		resync_free_pages(rp);
 211		bio_put(r1bio->bios[i]);
 212	}
 213
 214	/* resync pages array stored in the 1st bio's .bi_private */
 215	kfree(rp);
 216
 217	rbio_pool_free(r1bio, data);
 218}
 219
 220static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 221{
 222	int i;
 223
 224	for (i = 0; i < conf->raid_disks * 2; i++) {
 225		struct bio **bio = r1_bio->bios + i;
 226		if (!BIO_SPECIAL(*bio))
 227			bio_put(*bio);
 228		*bio = NULL;
 229	}
 230}
 231
 232static void free_r1bio(struct r1bio *r1_bio)
 233{
 234	struct r1conf *conf = r1_bio->mddev->private;
 235
 236	put_all_bios(conf, r1_bio);
 237	mempool_free(r1_bio, &conf->r1bio_pool);
 238}
 239
 240static void put_buf(struct r1bio *r1_bio)
 241{
 242	struct r1conf *conf = r1_bio->mddev->private;
 243	sector_t sect = r1_bio->sector;
 244	int i;
 245
 246	for (i = 0; i < conf->raid_disks * 2; i++) {
 247		struct bio *bio = r1_bio->bios[i];
 248		if (bio->bi_end_io)
 249			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 250	}
 251
 252	mempool_free(r1_bio, &conf->r1buf_pool);
 253
 254	lower_barrier(conf, sect);
 255}
 256
 257static void reschedule_retry(struct r1bio *r1_bio)
 258{
 259	unsigned long flags;
 260	struct mddev *mddev = r1_bio->mddev;
 261	struct r1conf *conf = mddev->private;
 262	int idx;
 263
 264	idx = sector_to_idx(r1_bio->sector);
 265	spin_lock_irqsave(&conf->device_lock, flags);
 266	list_add(&r1_bio->retry_list, &conf->retry_list);
 267	atomic_inc(&conf->nr_queued[idx]);
 268	spin_unlock_irqrestore(&conf->device_lock, flags);
 269
 270	wake_up(&conf->wait_barrier);
 271	md_wakeup_thread(mddev->thread);
 272}
 273
 274/*
 275 * raid_end_bio_io() is called when we have finished servicing a mirrored
 276 * operation and are ready to return a success/failure code to the buffer
 277 * cache layer.
 278 */
 279static void call_bio_endio(struct r1bio *r1_bio)
 280{
 281	struct bio *bio = r1_bio->master_bio;
 282	struct r1conf *conf = r1_bio->mddev->private;
 283
 284	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 285		bio->bi_status = BLK_STS_IOERR;
 286
 287	bio_endio(bio);
 288	/*
 289	 * Wake up any possible resync thread that waits for the device
 290	 * to go idle.
 291	 */
 292	allow_barrier(conf, r1_bio->sector);
 293}
 294
 295static void raid_end_bio_io(struct r1bio *r1_bio)
 296{
 297	struct bio *bio = r1_bio->master_bio;
 298
 299	/* if nobody has done the final endio yet, do it now */
 300	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 301		pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 302			 (bio_data_dir(bio) == WRITE) ? "write" : "read",
 303			 (unsigned long long) bio->bi_iter.bi_sector,
 304			 (unsigned long long) bio_end_sector(bio) - 1);
 305
 306		call_bio_endio(r1_bio);
 307	}
 308	free_r1bio(r1_bio);
 309}
 310
 311/*
 312 * Update disk head position estimator based on IRQ completion info.
 313 */
 314static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 315{
 316	struct r1conf *conf = r1_bio->mddev->private;
 317
 318	conf->mirrors[disk].head_position =
 319		r1_bio->sector + (r1_bio->sectors);
 320}
 321
 322/*
 323 * Find the disk number which triggered given bio
 324 */
 325static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 326{
 327	int mirror;
 328	struct r1conf *conf = r1_bio->mddev->private;
 329	int raid_disks = conf->raid_disks;
 330
 331	for (mirror = 0; mirror < raid_disks * 2; mirror++)
 332		if (r1_bio->bios[mirror] == bio)
 333			break;
 334
 335	BUG_ON(mirror == raid_disks * 2);
 336	update_head_pos(mirror, r1_bio);
 337
 338	return mirror;
 339}
 340
 341static void raid1_end_read_request(struct bio *bio)
 342{
 343	int uptodate = !bio->bi_status;
 344	struct r1bio *r1_bio = bio->bi_private;
 345	struct r1conf *conf = r1_bio->mddev->private;
 346	struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 347
 348	/*
 349	 * this branch is our 'one mirror IO has finished' event handler:
 350	 */
 351	update_head_pos(r1_bio->read_disk, r1_bio);
 352
 353	if (uptodate)
 354		set_bit(R1BIO_Uptodate, &r1_bio->state);
 355	else if (test_bit(FailFast, &rdev->flags) &&
 356		 test_bit(R1BIO_FailFast, &r1_bio->state))
 357		/* This was a fail-fast read so we definitely
 358		 * want to retry */
 359		;
 360	else {
 361		/* If all other devices have failed, we want to return
 362		 * the error upwards rather than fail the last device.
 363		 * Here we redefine "uptodate" to mean "Don't want to retry"
 364		 */
 365		unsigned long flags;
 366		spin_lock_irqsave(&conf->device_lock, flags);
 367		if (r1_bio->mddev->degraded == conf->raid_disks ||
 368		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 369		     test_bit(In_sync, &rdev->flags)))
 370			uptodate = 1;
 371		spin_unlock_irqrestore(&conf->device_lock, flags);
 372	}
 373
 374	if (uptodate) {
 375		raid_end_bio_io(r1_bio);
 376		rdev_dec_pending(rdev, conf->mddev);
 377	} else {
 378		/*
 379		 * oops, read error:
 380		 */
 381		char b[BDEVNAME_SIZE];
 382		pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 383				   mdname(conf->mddev),
 384				   bdevname(rdev->bdev, b),
 385				   (unsigned long long)r1_bio->sector);
 386		set_bit(R1BIO_ReadError, &r1_bio->state);
 387		reschedule_retry(r1_bio);
 388		/* don't drop the reference on read_disk yet */
 389	}
 390}
 391
 392static void close_write(struct r1bio *r1_bio)
 393{
 394	/* it really is the end of this request */
 395	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 396		bio_free_pages(r1_bio->behind_master_bio);
 397		bio_put(r1_bio->behind_master_bio);
 398		r1_bio->behind_master_bio = NULL;
 399	}
 400	/* clear the bitmap if all writes complete successfully */
 401	md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 402			   r1_bio->sectors,
 403			   !test_bit(R1BIO_Degraded, &r1_bio->state),
 404			   test_bit(R1BIO_BehindIO, &r1_bio->state));
 405	md_write_end(r1_bio->mddev);
 406}
 407
 408static void r1_bio_write_done(struct r1bio *r1_bio)
 409{
 410	if (!atomic_dec_and_test(&r1_bio->remaining))
 411		return;
 412
 413	if (test_bit(R1BIO_WriteError, &r1_bio->state))
 414		reschedule_retry(r1_bio);
 415	else {
 416		close_write(r1_bio);
 417		if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 418			reschedule_retry(r1_bio);
 419		else
 420			raid_end_bio_io(r1_bio);
 421	}
 422}
 423
 424static void raid1_end_write_request(struct bio *bio)
 425{
 426	struct r1bio *r1_bio = bio->bi_private;
 427	int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 428	struct r1conf *conf = r1_bio->mddev->private;
 429	struct bio *to_put = NULL;
 430	int mirror = find_bio_disk(r1_bio, bio);
 431	struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 432	bool discard_error;
 433
 434	discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 435
 436	/*
 437	 * 'one mirror IO has finished' event handler:
 438	 */
 439	if (bio->bi_status && !discard_error) {
 440		set_bit(WriteErrorSeen,	&rdev->flags);
 441		if (!test_and_set_bit(WantReplacement, &rdev->flags))
 442			set_bit(MD_RECOVERY_NEEDED, &
 443				conf->mddev->recovery);
 444
 445		if (test_bit(FailFast, &rdev->flags) &&
 446		    (bio->bi_opf & MD_FAILFAST) &&
 447		    /* We never try FailFast to WriteMostly devices */
 448		    !test_bit(WriteMostly, &rdev->flags)) {
 449			md_error(r1_bio->mddev, rdev);
 450		}
 451
 452		/*
 453		 * When the device is faulty, it is not necessary to
 454		 * handle write error.
 455		 * For failfast, this is the only remaining device,
 456		 * We need to retry the write without FailFast.
 457		 */
 458		if (!test_bit(Faulty, &rdev->flags))
 
 
 
 459			set_bit(R1BIO_WriteError, &r1_bio->state);
 460		else {
 461			/* Finished with this branch */
 462			r1_bio->bios[mirror] = NULL;
 463			to_put = bio;
 464		}
 465	} else {
 466		/*
 467		 * Set R1BIO_Uptodate in our master bio, so that we
 468		 * will return a good error code for to the higher
 469		 * levels even if IO on some other mirrored buffer
 470		 * fails.
 471		 *
 472		 * The 'master' represents the composite IO operation
 473		 * to user-side. So if something waits for IO, then it
 474		 * will wait for the 'master' bio.
 475		 */
 476		sector_t first_bad;
 477		int bad_sectors;
 478
 479		r1_bio->bios[mirror] = NULL;
 480		to_put = bio;
 481		/*
 482		 * Do not set R1BIO_Uptodate if the current device is
 483		 * rebuilding or Faulty. This is because we cannot use
 484		 * such device for properly reading the data back (we could
 485		 * potentially use it, if the current write would have felt
 486		 * before rdev->recovery_offset, but for simplicity we don't
 487		 * check this here.
 488		 */
 489		if (test_bit(In_sync, &rdev->flags) &&
 490		    !test_bit(Faulty, &rdev->flags))
 491			set_bit(R1BIO_Uptodate, &r1_bio->state);
 492
 493		/* Maybe we can clear some bad blocks. */
 494		if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 495				&first_bad, &bad_sectors) && !discard_error) {
 496			r1_bio->bios[mirror] = IO_MADE_GOOD;
 497			set_bit(R1BIO_MadeGood, &r1_bio->state);
 498		}
 499	}
 500
 501	if (behind) {
 502		if (test_bit(WBCollisionCheck, &rdev->flags)) {
 503			sector_t lo = r1_bio->sector;
 504			sector_t hi = r1_bio->sector + r1_bio->sectors;
 505
 506			remove_wb(rdev, lo, hi);
 507		}
 508		if (test_bit(WriteMostly, &rdev->flags))
 509			atomic_dec(&r1_bio->behind_remaining);
 510
 511		/*
 512		 * In behind mode, we ACK the master bio once the I/O
 513		 * has safely reached all non-writemostly
 514		 * disks. Setting the Returned bit ensures that this
 515		 * gets done only once -- we don't ever want to return
 516		 * -EIO here, instead we'll wait
 517		 */
 518		if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 519		    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 520			/* Maybe we can return now */
 521			if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 522				struct bio *mbio = r1_bio->master_bio;
 523				pr_debug("raid1: behind end write sectors"
 524					 " %llu-%llu\n",
 525					 (unsigned long long) mbio->bi_iter.bi_sector,
 526					 (unsigned long long) bio_end_sector(mbio) - 1);
 527				call_bio_endio(r1_bio);
 528			}
 529		}
 530	}
 531	if (r1_bio->bios[mirror] == NULL)
 532		rdev_dec_pending(rdev, conf->mddev);
 533
 534	/*
 535	 * Let's see if all mirrored write operations have finished
 536	 * already.
 537	 */
 538	r1_bio_write_done(r1_bio);
 539
 540	if (to_put)
 541		bio_put(to_put);
 542}
 543
 544static sector_t align_to_barrier_unit_end(sector_t start_sector,
 545					  sector_t sectors)
 546{
 547	sector_t len;
 548
 549	WARN_ON(sectors == 0);
 550	/*
 551	 * len is the number of sectors from start_sector to end of the
 552	 * barrier unit which start_sector belongs to.
 553	 */
 554	len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 555	      start_sector;
 556
 557	if (len > sectors)
 558		len = sectors;
 559
 560	return len;
 561}
 562
 563/*
 564 * This routine returns the disk from which the requested read should
 565 * be done. There is a per-array 'next expected sequential IO' sector
 566 * number - if this matches on the next IO then we use the last disk.
 567 * There is also a per-disk 'last know head position' sector that is
 568 * maintained from IRQ contexts, both the normal and the resync IO
 569 * completion handlers update this position correctly. If there is no
 570 * perfect sequential match then we pick the disk whose head is closest.
 571 *
 572 * If there are 2 mirrors in the same 2 devices, performance degrades
 573 * because position is mirror, not device based.
 574 *
 575 * The rdev for the device selected will have nr_pending incremented.
 576 */
 577static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 578{
 579	const sector_t this_sector = r1_bio->sector;
 580	int sectors;
 581	int best_good_sectors;
 582	int best_disk, best_dist_disk, best_pending_disk;
 583	int has_nonrot_disk;
 584	int disk;
 585	sector_t best_dist;
 586	unsigned int min_pending;
 587	struct md_rdev *rdev;
 588	int choose_first;
 589	int choose_next_idle;
 590
 591	rcu_read_lock();
 592	/*
 593	 * Check if we can balance. We can balance on the whole
 594	 * device if no resync is going on, or below the resync window.
 595	 * We take the first readable disk when above the resync window.
 596	 */
 597 retry:
 598	sectors = r1_bio->sectors;
 599	best_disk = -1;
 600	best_dist_disk = -1;
 601	best_dist = MaxSector;
 602	best_pending_disk = -1;
 603	min_pending = UINT_MAX;
 604	best_good_sectors = 0;
 605	has_nonrot_disk = 0;
 606	choose_next_idle = 0;
 607	clear_bit(R1BIO_FailFast, &r1_bio->state);
 608
 609	if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 610	    (mddev_is_clustered(conf->mddev) &&
 611	    md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 612		    this_sector + sectors)))
 613		choose_first = 1;
 614	else
 615		choose_first = 0;
 616
 617	for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 618		sector_t dist;
 619		sector_t first_bad;
 620		int bad_sectors;
 621		unsigned int pending;
 622		bool nonrot;
 623
 624		rdev = rcu_dereference(conf->mirrors[disk].rdev);
 625		if (r1_bio->bios[disk] == IO_BLOCKED
 626		    || rdev == NULL
 627		    || test_bit(Faulty, &rdev->flags))
 628			continue;
 629		if (!test_bit(In_sync, &rdev->flags) &&
 630		    rdev->recovery_offset < this_sector + sectors)
 631			continue;
 632		if (test_bit(WriteMostly, &rdev->flags)) {
 633			/* Don't balance among write-mostly, just
 634			 * use the first as a last resort */
 635			if (best_dist_disk < 0) {
 636				if (is_badblock(rdev, this_sector, sectors,
 637						&first_bad, &bad_sectors)) {
 638					if (first_bad <= this_sector)
 639						/* Cannot use this */
 640						continue;
 641					best_good_sectors = first_bad - this_sector;
 642				} else
 643					best_good_sectors = sectors;
 644				best_dist_disk = disk;
 645				best_pending_disk = disk;
 646			}
 647			continue;
 648		}
 649		/* This is a reasonable device to use.  It might
 650		 * even be best.
 651		 */
 652		if (is_badblock(rdev, this_sector, sectors,
 653				&first_bad, &bad_sectors)) {
 654			if (best_dist < MaxSector)
 655				/* already have a better device */
 656				continue;
 657			if (first_bad <= this_sector) {
 658				/* cannot read here. If this is the 'primary'
 659				 * device, then we must not read beyond
 660				 * bad_sectors from another device..
 661				 */
 662				bad_sectors -= (this_sector - first_bad);
 663				if (choose_first && sectors > bad_sectors)
 664					sectors = bad_sectors;
 665				if (best_good_sectors > sectors)
 666					best_good_sectors = sectors;
 667
 668			} else {
 669				sector_t good_sectors = first_bad - this_sector;
 670				if (good_sectors > best_good_sectors) {
 671					best_good_sectors = good_sectors;
 672					best_disk = disk;
 673				}
 674				if (choose_first)
 675					break;
 676			}
 677			continue;
 678		} else {
 679			if ((sectors > best_good_sectors) && (best_disk >= 0))
 680				best_disk = -1;
 681			best_good_sectors = sectors;
 682		}
 683
 684		if (best_disk >= 0)
 685			/* At least two disks to choose from so failfast is OK */
 686			set_bit(R1BIO_FailFast, &r1_bio->state);
 687
 688		nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 689		has_nonrot_disk |= nonrot;
 690		pending = atomic_read(&rdev->nr_pending);
 691		dist = abs(this_sector - conf->mirrors[disk].head_position);
 692		if (choose_first) {
 693			best_disk = disk;
 694			break;
 695		}
 696		/* Don't change to another disk for sequential reads */
 697		if (conf->mirrors[disk].next_seq_sect == this_sector
 698		    || dist == 0) {
 699			int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 700			struct raid1_info *mirror = &conf->mirrors[disk];
 701
 702			best_disk = disk;
 703			/*
 704			 * If buffered sequential IO size exceeds optimal
 705			 * iosize, check if there is idle disk. If yes, choose
 706			 * the idle disk. read_balance could already choose an
 707			 * idle disk before noticing it's a sequential IO in
 708			 * this disk. This doesn't matter because this disk
 709			 * will idle, next time it will be utilized after the
 710			 * first disk has IO size exceeds optimal iosize. In
 711			 * this way, iosize of the first disk will be optimal
 712			 * iosize at least. iosize of the second disk might be
 713			 * small, but not a big deal since when the second disk
 714			 * starts IO, the first disk is likely still busy.
 715			 */
 716			if (nonrot && opt_iosize > 0 &&
 717			    mirror->seq_start != MaxSector &&
 718			    mirror->next_seq_sect > opt_iosize &&
 719			    mirror->next_seq_sect - opt_iosize >=
 720			    mirror->seq_start) {
 721				choose_next_idle = 1;
 722				continue;
 723			}
 724			break;
 725		}
 726
 727		if (choose_next_idle)
 728			continue;
 729
 730		if (min_pending > pending) {
 731			min_pending = pending;
 732			best_pending_disk = disk;
 733		}
 734
 735		if (dist < best_dist) {
 736			best_dist = dist;
 737			best_dist_disk = disk;
 738		}
 739	}
 740
 741	/*
 742	 * If all disks are rotational, choose the closest disk. If any disk is
 743	 * non-rotational, choose the disk with less pending request even the
 744	 * disk is rotational, which might/might not be optimal for raids with
 745	 * mixed ratation/non-rotational disks depending on workload.
 746	 */
 747	if (best_disk == -1) {
 748		if (has_nonrot_disk || min_pending == 0)
 749			best_disk = best_pending_disk;
 750		else
 751			best_disk = best_dist_disk;
 752	}
 753
 754	if (best_disk >= 0) {
 755		rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 756		if (!rdev)
 757			goto retry;
 758		atomic_inc(&rdev->nr_pending);
 759		sectors = best_good_sectors;
 760
 761		if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 762			conf->mirrors[best_disk].seq_start = this_sector;
 763
 764		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 765	}
 766	rcu_read_unlock();
 767	*max_sectors = sectors;
 768
 769	return best_disk;
 770}
 771
 772static int raid1_congested(struct mddev *mddev, int bits)
 773{
 774	struct r1conf *conf = mddev->private;
 775	int i, ret = 0;
 776
 777	if ((bits & (1 << WB_async_congested)) &&
 778	    conf->pending_count >= max_queued_requests)
 779		return 1;
 780
 781	rcu_read_lock();
 782	for (i = 0; i < conf->raid_disks * 2; i++) {
 783		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 784		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 785			struct request_queue *q = bdev_get_queue(rdev->bdev);
 786
 787			BUG_ON(!q);
 788
 789			/* Note the '|| 1' - when read_balance prefers
 790			 * non-congested targets, it can be removed
 791			 */
 792			if ((bits & (1 << WB_async_congested)) || 1)
 793				ret |= bdi_congested(q->backing_dev_info, bits);
 794			else
 795				ret &= bdi_congested(q->backing_dev_info, bits);
 796		}
 797	}
 798	rcu_read_unlock();
 799	return ret;
 800}
 801
 802static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 803{
 804	/* flush any pending bitmap writes to disk before proceeding w/ I/O */
 805	md_bitmap_unplug(conf->mddev->bitmap);
 806	wake_up(&conf->wait_barrier);
 807
 808	while (bio) { /* submit pending writes */
 809		struct bio *next = bio->bi_next;
 810		struct md_rdev *rdev = (void *)bio->bi_disk;
 811		bio->bi_next = NULL;
 812		bio_set_dev(bio, rdev->bdev);
 813		if (test_bit(Faulty, &rdev->flags)) {
 814			bio_io_error(bio);
 815		} else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 816				    !blk_queue_discard(bio->bi_disk->queue)))
 817			/* Just ignore it */
 818			bio_endio(bio);
 819		else
 820			generic_make_request(bio);
 821		bio = next;
 822	}
 823}
 824
 825static void flush_pending_writes(struct r1conf *conf)
 826{
 827	/* Any writes that have been queued but are awaiting
 828	 * bitmap updates get flushed here.
 829	 */
 830	spin_lock_irq(&conf->device_lock);
 831
 832	if (conf->pending_bio_list.head) {
 833		struct blk_plug plug;
 834		struct bio *bio;
 835
 836		bio = bio_list_get(&conf->pending_bio_list);
 837		conf->pending_count = 0;
 838		spin_unlock_irq(&conf->device_lock);
 839
 840		/*
 841		 * As this is called in a wait_event() loop (see freeze_array),
 842		 * current->state might be TASK_UNINTERRUPTIBLE which will
 843		 * cause a warning when we prepare to wait again.  As it is
 844		 * rare that this path is taken, it is perfectly safe to force
 845		 * us to go around the wait_event() loop again, so the warning
 846		 * is a false-positive.  Silence the warning by resetting
 847		 * thread state
 848		 */
 849		__set_current_state(TASK_RUNNING);
 850		blk_start_plug(&plug);
 851		flush_bio_list(conf, bio);
 852		blk_finish_plug(&plug);
 853	} else
 854		spin_unlock_irq(&conf->device_lock);
 855}
 856
 857/* Barriers....
 858 * Sometimes we need to suspend IO while we do something else,
 859 * either some resync/recovery, or reconfigure the array.
 860 * To do this we raise a 'barrier'.
 861 * The 'barrier' is a counter that can be raised multiple times
 862 * to count how many activities are happening which preclude
 863 * normal IO.
 864 * We can only raise the barrier if there is no pending IO.
 865 * i.e. if nr_pending == 0.
 866 * We choose only to raise the barrier if no-one is waiting for the
 867 * barrier to go down.  This means that as soon as an IO request
 868 * is ready, no other operations which require a barrier will start
 869 * until the IO request has had a chance.
 870 *
 871 * So: regular IO calls 'wait_barrier'.  When that returns there
 872 *    is no backgroup IO happening,  It must arrange to call
 873 *    allow_barrier when it has finished its IO.
 874 * backgroup IO calls must call raise_barrier.  Once that returns
 875 *    there is no normal IO happeing.  It must arrange to call
 876 *    lower_barrier when the particular background IO completes.
 877 *
 878 * If resync/recovery is interrupted, returns -EINTR;
 879 * Otherwise, returns 0.
 880 */
 881static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 882{
 883	int idx = sector_to_idx(sector_nr);
 884
 885	spin_lock_irq(&conf->resync_lock);
 886
 887	/* Wait until no block IO is waiting */
 888	wait_event_lock_irq(conf->wait_barrier,
 889			    !atomic_read(&conf->nr_waiting[idx]),
 890			    conf->resync_lock);
 891
 892	/* block any new IO from starting */
 893	atomic_inc(&conf->barrier[idx]);
 894	/*
 895	 * In raise_barrier() we firstly increase conf->barrier[idx] then
 896	 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 897	 * increase conf->nr_pending[idx] then check conf->barrier[idx].
 898	 * A memory barrier here to make sure conf->nr_pending[idx] won't
 899	 * be fetched before conf->barrier[idx] is increased. Otherwise
 900	 * there will be a race between raise_barrier() and _wait_barrier().
 901	 */
 902	smp_mb__after_atomic();
 903
 904	/* For these conditions we must wait:
 905	 * A: while the array is in frozen state
 906	 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 907	 *    existing in corresponding I/O barrier bucket.
 908	 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 909	 *    max resync count which allowed on current I/O barrier bucket.
 910	 */
 911	wait_event_lock_irq(conf->wait_barrier,
 912			    (!conf->array_frozen &&
 913			     !atomic_read(&conf->nr_pending[idx]) &&
 914			     atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 915				test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 916			    conf->resync_lock);
 917
 918	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 919		atomic_dec(&conf->barrier[idx]);
 920		spin_unlock_irq(&conf->resync_lock);
 921		wake_up(&conf->wait_barrier);
 922		return -EINTR;
 923	}
 924
 925	atomic_inc(&conf->nr_sync_pending);
 926	spin_unlock_irq(&conf->resync_lock);
 927
 928	return 0;
 929}
 930
 931static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 932{
 933	int idx = sector_to_idx(sector_nr);
 934
 935	BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 936
 937	atomic_dec(&conf->barrier[idx]);
 938	atomic_dec(&conf->nr_sync_pending);
 939	wake_up(&conf->wait_barrier);
 940}
 941
 942static void _wait_barrier(struct r1conf *conf, int idx)
 943{
 944	/*
 945	 * We need to increase conf->nr_pending[idx] very early here,
 946	 * then raise_barrier() can be blocked when it waits for
 947	 * conf->nr_pending[idx] to be 0. Then we can avoid holding
 948	 * conf->resync_lock when there is no barrier raised in same
 949	 * barrier unit bucket. Also if the array is frozen, I/O
 950	 * should be blocked until array is unfrozen.
 951	 */
 952	atomic_inc(&conf->nr_pending[idx]);
 953	/*
 954	 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 955	 * check conf->barrier[idx]. In raise_barrier() we firstly increase
 956	 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 957	 * barrier is necessary here to make sure conf->barrier[idx] won't be
 958	 * fetched before conf->nr_pending[idx] is increased. Otherwise there
 959	 * will be a race between _wait_barrier() and raise_barrier().
 960	 */
 961	smp_mb__after_atomic();
 962
 963	/*
 964	 * Don't worry about checking two atomic_t variables at same time
 965	 * here. If during we check conf->barrier[idx], the array is
 966	 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 967	 * 0, it is safe to return and make the I/O continue. Because the
 968	 * array is frozen, all I/O returned here will eventually complete
 969	 * or be queued, no race will happen. See code comment in
 970	 * frozen_array().
 971	 */
 972	if (!READ_ONCE(conf->array_frozen) &&
 973	    !atomic_read(&conf->barrier[idx]))
 974		return;
 975
 976	/*
 977	 * After holding conf->resync_lock, conf->nr_pending[idx]
 978	 * should be decreased before waiting for barrier to drop.
 979	 * Otherwise, we may encounter a race condition because
 980	 * raise_barrer() might be waiting for conf->nr_pending[idx]
 981	 * to be 0 at same time.
 982	 */
 983	spin_lock_irq(&conf->resync_lock);
 984	atomic_inc(&conf->nr_waiting[idx]);
 985	atomic_dec(&conf->nr_pending[idx]);
 986	/*
 987	 * In case freeze_array() is waiting for
 988	 * get_unqueued_pending() == extra
 989	 */
 990	wake_up(&conf->wait_barrier);
 991	/* Wait for the barrier in same barrier unit bucket to drop. */
 992	wait_event_lock_irq(conf->wait_barrier,
 993			    !conf->array_frozen &&
 994			     !atomic_read(&conf->barrier[idx]),
 995			    conf->resync_lock);
 996	atomic_inc(&conf->nr_pending[idx]);
 997	atomic_dec(&conf->nr_waiting[idx]);
 998	spin_unlock_irq(&conf->resync_lock);
 999}
1000
1001static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
1002{
1003	int idx = sector_to_idx(sector_nr);
1004
1005	/*
1006	 * Very similar to _wait_barrier(). The difference is, for read
1007	 * I/O we don't need wait for sync I/O, but if the whole array
1008	 * is frozen, the read I/O still has to wait until the array is
1009	 * unfrozen. Since there is no ordering requirement with
1010	 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1011	 */
1012	atomic_inc(&conf->nr_pending[idx]);
1013
1014	if (!READ_ONCE(conf->array_frozen))
1015		return;
1016
1017	spin_lock_irq(&conf->resync_lock);
1018	atomic_inc(&conf->nr_waiting[idx]);
1019	atomic_dec(&conf->nr_pending[idx]);
1020	/*
1021	 * In case freeze_array() is waiting for
1022	 * get_unqueued_pending() == extra
1023	 */
1024	wake_up(&conf->wait_barrier);
1025	/* Wait for array to be unfrozen */
1026	wait_event_lock_irq(conf->wait_barrier,
1027			    !conf->array_frozen,
1028			    conf->resync_lock);
1029	atomic_inc(&conf->nr_pending[idx]);
1030	atomic_dec(&conf->nr_waiting[idx]);
1031	spin_unlock_irq(&conf->resync_lock);
1032}
1033
1034static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1035{
1036	int idx = sector_to_idx(sector_nr);
1037
1038	_wait_barrier(conf, idx);
1039}
1040
1041static void _allow_barrier(struct r1conf *conf, int idx)
1042{
1043	atomic_dec(&conf->nr_pending[idx]);
1044	wake_up(&conf->wait_barrier);
1045}
1046
1047static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1048{
1049	int idx = sector_to_idx(sector_nr);
1050
1051	_allow_barrier(conf, idx);
1052}
1053
1054/* conf->resync_lock should be held */
1055static int get_unqueued_pending(struct r1conf *conf)
1056{
1057	int idx, ret;
1058
1059	ret = atomic_read(&conf->nr_sync_pending);
1060	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1061		ret += atomic_read(&conf->nr_pending[idx]) -
1062			atomic_read(&conf->nr_queued[idx]);
1063
1064	return ret;
1065}
1066
1067static void freeze_array(struct r1conf *conf, int extra)
1068{
1069	/* Stop sync I/O and normal I/O and wait for everything to
1070	 * go quiet.
1071	 * This is called in two situations:
1072	 * 1) management command handlers (reshape, remove disk, quiesce).
1073	 * 2) one normal I/O request failed.
1074
1075	 * After array_frozen is set to 1, new sync IO will be blocked at
1076	 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1077	 * or wait_read_barrier(). The flying I/Os will either complete or be
1078	 * queued. When everything goes quite, there are only queued I/Os left.
1079
1080	 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1081	 * barrier bucket index which this I/O request hits. When all sync and
1082	 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1083	 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1084	 * in handle_read_error(), we may call freeze_array() before trying to
1085	 * fix the read error. In this case, the error read I/O is not queued,
1086	 * so get_unqueued_pending() == 1.
1087	 *
1088	 * Therefore before this function returns, we need to wait until
1089	 * get_unqueued_pendings(conf) gets equal to extra. For
1090	 * normal I/O context, extra is 1, in rested situations extra is 0.
1091	 */
1092	spin_lock_irq(&conf->resync_lock);
1093	conf->array_frozen = 1;
1094	raid1_log(conf->mddev, "wait freeze");
1095	wait_event_lock_irq_cmd(
1096		conf->wait_barrier,
1097		get_unqueued_pending(conf) == extra,
1098		conf->resync_lock,
1099		flush_pending_writes(conf));
1100	spin_unlock_irq(&conf->resync_lock);
1101}
1102static void unfreeze_array(struct r1conf *conf)
1103{
1104	/* reverse the effect of the freeze */
1105	spin_lock_irq(&conf->resync_lock);
1106	conf->array_frozen = 0;
1107	spin_unlock_irq(&conf->resync_lock);
1108	wake_up(&conf->wait_barrier);
1109}
1110
1111static void alloc_behind_master_bio(struct r1bio *r1_bio,
1112					   struct bio *bio)
1113{
1114	int size = bio->bi_iter.bi_size;
1115	unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1116	int i = 0;
1117	struct bio *behind_bio = NULL;
1118
1119	behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1120	if (!behind_bio)
1121		return;
1122
1123	/* discard op, we don't support writezero/writesame yet */
1124	if (!bio_has_data(bio)) {
1125		behind_bio->bi_iter.bi_size = size;
1126		goto skip_copy;
1127	}
1128
1129	behind_bio->bi_write_hint = bio->bi_write_hint;
1130
1131	while (i < vcnt && size) {
1132		struct page *page;
1133		int len = min_t(int, PAGE_SIZE, size);
1134
1135		page = alloc_page(GFP_NOIO);
1136		if (unlikely(!page))
1137			goto free_pages;
1138
1139		bio_add_page(behind_bio, page, len, 0);
1140
1141		size -= len;
1142		i++;
1143	}
1144
1145	bio_copy_data(behind_bio, bio);
1146skip_copy:
1147	r1_bio->behind_master_bio = behind_bio;
1148	set_bit(R1BIO_BehindIO, &r1_bio->state);
1149
1150	return;
1151
1152free_pages:
1153	pr_debug("%dB behind alloc failed, doing sync I/O\n",
1154		 bio->bi_iter.bi_size);
1155	bio_free_pages(behind_bio);
1156	bio_put(behind_bio);
1157}
1158
1159struct raid1_plug_cb {
1160	struct blk_plug_cb	cb;
1161	struct bio_list		pending;
1162	int			pending_cnt;
1163};
1164
1165static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1166{
1167	struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1168						  cb);
1169	struct mddev *mddev = plug->cb.data;
1170	struct r1conf *conf = mddev->private;
1171	struct bio *bio;
1172
1173	if (from_schedule || current->bio_list) {
1174		spin_lock_irq(&conf->device_lock);
1175		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1176		conf->pending_count += plug->pending_cnt;
1177		spin_unlock_irq(&conf->device_lock);
1178		wake_up(&conf->wait_barrier);
1179		md_wakeup_thread(mddev->thread);
1180		kfree(plug);
1181		return;
1182	}
1183
1184	/* we aren't scheduling, so we can do the write-out directly. */
1185	bio = bio_list_get(&plug->pending);
1186	flush_bio_list(conf, bio);
1187	kfree(plug);
1188}
1189
1190static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1191{
1192	r1_bio->master_bio = bio;
1193	r1_bio->sectors = bio_sectors(bio);
1194	r1_bio->state = 0;
1195	r1_bio->mddev = mddev;
1196	r1_bio->sector = bio->bi_iter.bi_sector;
1197}
1198
1199static inline struct r1bio *
1200alloc_r1bio(struct mddev *mddev, struct bio *bio)
1201{
1202	struct r1conf *conf = mddev->private;
1203	struct r1bio *r1_bio;
1204
1205	r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1206	/* Ensure no bio records IO_BLOCKED */
1207	memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1208	init_r1bio(r1_bio, mddev, bio);
1209	return r1_bio;
1210}
1211
1212static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1213			       int max_read_sectors, struct r1bio *r1_bio)
1214{
1215	struct r1conf *conf = mddev->private;
1216	struct raid1_info *mirror;
1217	struct bio *read_bio;
1218	struct bitmap *bitmap = mddev->bitmap;
1219	const int op = bio_op(bio);
1220	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1221	int max_sectors;
1222	int rdisk;
1223	bool print_msg = !!r1_bio;
1224	char b[BDEVNAME_SIZE];
1225
1226	/*
1227	 * If r1_bio is set, we are blocking the raid1d thread
1228	 * so there is a tiny risk of deadlock.  So ask for
1229	 * emergency memory if needed.
1230	 */
1231	gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1232
1233	if (print_msg) {
1234		/* Need to get the block device name carefully */
1235		struct md_rdev *rdev;
1236		rcu_read_lock();
1237		rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1238		if (rdev)
1239			bdevname(rdev->bdev, b);
1240		else
1241			strcpy(b, "???");
1242		rcu_read_unlock();
1243	}
1244
1245	/*
1246	 * Still need barrier for READ in case that whole
1247	 * array is frozen.
1248	 */
1249	wait_read_barrier(conf, bio->bi_iter.bi_sector);
1250
1251	if (!r1_bio)
1252		r1_bio = alloc_r1bio(mddev, bio);
1253	else
1254		init_r1bio(r1_bio, mddev, bio);
1255	r1_bio->sectors = max_read_sectors;
1256
1257	/*
1258	 * make_request() can abort the operation when read-ahead is being
1259	 * used and no empty request is available.
1260	 */
1261	rdisk = read_balance(conf, r1_bio, &max_sectors);
1262
1263	if (rdisk < 0) {
1264		/* couldn't find anywhere to read from */
1265		if (print_msg) {
1266			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1267					    mdname(mddev),
1268					    b,
1269					    (unsigned long long)r1_bio->sector);
1270		}
1271		raid_end_bio_io(r1_bio);
1272		return;
1273	}
1274	mirror = conf->mirrors + rdisk;
1275
1276	if (print_msg)
1277		pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1278				    mdname(mddev),
1279				    (unsigned long long)r1_bio->sector,
1280				    bdevname(mirror->rdev->bdev, b));
1281
1282	if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1283	    bitmap) {
1284		/*
1285		 * Reading from a write-mostly device must take care not to
1286		 * over-take any writes that are 'behind'
1287		 */
1288		raid1_log(mddev, "wait behind writes");
1289		wait_event(bitmap->behind_wait,
1290			   atomic_read(&bitmap->behind_writes) == 0);
1291	}
1292
1293	if (max_sectors < bio_sectors(bio)) {
1294		struct bio *split = bio_split(bio, max_sectors,
1295					      gfp, &conf->bio_split);
1296		bio_chain(split, bio);
1297		generic_make_request(bio);
1298		bio = split;
1299		r1_bio->master_bio = bio;
1300		r1_bio->sectors = max_sectors;
1301	}
1302
1303	r1_bio->read_disk = rdisk;
1304
1305	read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1306
1307	r1_bio->bios[rdisk] = read_bio;
1308
1309	read_bio->bi_iter.bi_sector = r1_bio->sector +
1310		mirror->rdev->data_offset;
1311	bio_set_dev(read_bio, mirror->rdev->bdev);
1312	read_bio->bi_end_io = raid1_end_read_request;
1313	bio_set_op_attrs(read_bio, op, do_sync);
1314	if (test_bit(FailFast, &mirror->rdev->flags) &&
1315	    test_bit(R1BIO_FailFast, &r1_bio->state))
1316	        read_bio->bi_opf |= MD_FAILFAST;
1317	read_bio->bi_private = r1_bio;
1318
1319	if (mddev->gendisk)
1320	        trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1321				disk_devt(mddev->gendisk), r1_bio->sector);
1322
1323	generic_make_request(read_bio);
1324}
1325
1326static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1327				int max_write_sectors)
1328{
1329	struct r1conf *conf = mddev->private;
1330	struct r1bio *r1_bio;
1331	int i, disks;
1332	struct bitmap *bitmap = mddev->bitmap;
1333	unsigned long flags;
1334	struct md_rdev *blocked_rdev;
1335	struct blk_plug_cb *cb;
1336	struct raid1_plug_cb *plug = NULL;
1337	int first_clone;
1338	int max_sectors;
1339
1340	if (mddev_is_clustered(mddev) &&
1341	     md_cluster_ops->area_resyncing(mddev, WRITE,
1342		     bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1343
1344		DEFINE_WAIT(w);
1345		for (;;) {
1346			prepare_to_wait(&conf->wait_barrier,
1347					&w, TASK_IDLE);
1348			if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1349							bio->bi_iter.bi_sector,
1350							bio_end_sector(bio)))
1351				break;
1352			schedule();
1353		}
1354		finish_wait(&conf->wait_barrier, &w);
1355	}
1356
1357	/*
1358	 * Register the new request and wait if the reconstruction
1359	 * thread has put up a bar for new requests.
1360	 * Continue immediately if no resync is active currently.
1361	 */
1362	wait_barrier(conf, bio->bi_iter.bi_sector);
1363
1364	r1_bio = alloc_r1bio(mddev, bio);
1365	r1_bio->sectors = max_write_sectors;
1366
1367	if (conf->pending_count >= max_queued_requests) {
1368		md_wakeup_thread(mddev->thread);
1369		raid1_log(mddev, "wait queued");
1370		wait_event(conf->wait_barrier,
1371			   conf->pending_count < max_queued_requests);
1372	}
1373	/* first select target devices under rcu_lock and
1374	 * inc refcount on their rdev.  Record them by setting
1375	 * bios[x] to bio
1376	 * If there are known/acknowledged bad blocks on any device on
1377	 * which we have seen a write error, we want to avoid writing those
1378	 * blocks.
1379	 * This potentially requires several writes to write around
1380	 * the bad blocks.  Each set of writes gets it's own r1bio
1381	 * with a set of bios attached.
1382	 */
1383
1384	disks = conf->raid_disks * 2;
1385 retry_write:
1386	blocked_rdev = NULL;
1387	rcu_read_lock();
1388	max_sectors = r1_bio->sectors;
1389	for (i = 0;  i < disks; i++) {
1390		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1391		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1392			atomic_inc(&rdev->nr_pending);
1393			blocked_rdev = rdev;
1394			break;
1395		}
1396		r1_bio->bios[i] = NULL;
1397		if (!rdev || test_bit(Faulty, &rdev->flags)) {
1398			if (i < conf->raid_disks)
1399				set_bit(R1BIO_Degraded, &r1_bio->state);
1400			continue;
1401		}
1402
1403		atomic_inc(&rdev->nr_pending);
1404		if (test_bit(WriteErrorSeen, &rdev->flags)) {
1405			sector_t first_bad;
1406			int bad_sectors;
1407			int is_bad;
1408
1409			is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1410					     &first_bad, &bad_sectors);
1411			if (is_bad < 0) {
1412				/* mustn't write here until the bad block is
1413				 * acknowledged*/
1414				set_bit(BlockedBadBlocks, &rdev->flags);
1415				blocked_rdev = rdev;
1416				break;
1417			}
1418			if (is_bad && first_bad <= r1_bio->sector) {
1419				/* Cannot write here at all */
1420				bad_sectors -= (r1_bio->sector - first_bad);
1421				if (bad_sectors < max_sectors)
1422					/* mustn't write more than bad_sectors
1423					 * to other devices yet
1424					 */
1425					max_sectors = bad_sectors;
1426				rdev_dec_pending(rdev, mddev);
1427				/* We don't set R1BIO_Degraded as that
1428				 * only applies if the disk is
1429				 * missing, so it might be re-added,
1430				 * and we want to know to recover this
1431				 * chunk.
1432				 * In this case the device is here,
1433				 * and the fact that this chunk is not
1434				 * in-sync is recorded in the bad
1435				 * block log
1436				 */
1437				continue;
1438			}
1439			if (is_bad) {
1440				int good_sectors = first_bad - r1_bio->sector;
1441				if (good_sectors < max_sectors)
1442					max_sectors = good_sectors;
1443			}
1444		}
1445		r1_bio->bios[i] = bio;
1446	}
1447	rcu_read_unlock();
1448
1449	if (unlikely(blocked_rdev)) {
1450		/* Wait for this device to become unblocked */
1451		int j;
1452
1453		for (j = 0; j < i; j++)
1454			if (r1_bio->bios[j])
1455				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1456		r1_bio->state = 0;
1457		allow_barrier(conf, bio->bi_iter.bi_sector);
1458		raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1459		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1460		wait_barrier(conf, bio->bi_iter.bi_sector);
1461		goto retry_write;
1462	}
1463
1464	if (max_sectors < bio_sectors(bio)) {
1465		struct bio *split = bio_split(bio, max_sectors,
1466					      GFP_NOIO, &conf->bio_split);
1467		bio_chain(split, bio);
1468		generic_make_request(bio);
1469		bio = split;
1470		r1_bio->master_bio = bio;
1471		r1_bio->sectors = max_sectors;
1472	}
1473
1474	atomic_set(&r1_bio->remaining, 1);
1475	atomic_set(&r1_bio->behind_remaining, 0);
1476
1477	first_clone = 1;
1478
1479	for (i = 0; i < disks; i++) {
1480		struct bio *mbio = NULL;
1481		if (!r1_bio->bios[i])
1482			continue;
1483
 
1484		if (first_clone) {
1485			/* do behind I/O ?
1486			 * Not if there are too many, or cannot
1487			 * allocate memory, or a reader on WriteMostly
1488			 * is waiting for behind writes to flush */
1489			if (bitmap &&
1490			    (atomic_read(&bitmap->behind_writes)
1491			     < mddev->bitmap_info.max_write_behind) &&
1492			    !waitqueue_active(&bitmap->behind_wait)) {
1493				alloc_behind_master_bio(r1_bio, bio);
1494			}
1495
1496			md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1497					     test_bit(R1BIO_BehindIO, &r1_bio->state));
 
 
1498			first_clone = 0;
1499		}
1500
1501		if (r1_bio->behind_master_bio)
1502			mbio = bio_clone_fast(r1_bio->behind_master_bio,
1503					      GFP_NOIO, &mddev->bio_set);
1504		else
1505			mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1506
1507		if (r1_bio->behind_master_bio) {
1508			struct md_rdev *rdev = conf->mirrors[i].rdev;
1509
1510			if (test_bit(WBCollisionCheck, &rdev->flags)) {
1511				sector_t lo = r1_bio->sector;
1512				sector_t hi = r1_bio->sector + r1_bio->sectors;
1513
1514				wait_event(rdev->wb_io_wait,
1515					   check_and_add_wb(rdev, lo, hi) == 0);
1516			}
1517			if (test_bit(WriteMostly, &rdev->flags))
1518				atomic_inc(&r1_bio->behind_remaining);
1519		}
1520
1521		r1_bio->bios[i] = mbio;
1522
1523		mbio->bi_iter.bi_sector	= (r1_bio->sector +
1524				   conf->mirrors[i].rdev->data_offset);
1525		bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1526		mbio->bi_end_io	= raid1_end_write_request;
1527		mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1528		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1529		    !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1530		    conf->raid_disks - mddev->degraded > 1)
1531			mbio->bi_opf |= MD_FAILFAST;
1532		mbio->bi_private = r1_bio;
1533
1534		atomic_inc(&r1_bio->remaining);
1535
1536		if (mddev->gendisk)
1537			trace_block_bio_remap(mbio->bi_disk->queue,
1538					      mbio, disk_devt(mddev->gendisk),
1539					      r1_bio->sector);
1540		/* flush_pending_writes() needs access to the rdev so...*/
1541		mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1542
1543		cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1544		if (cb)
1545			plug = container_of(cb, struct raid1_plug_cb, cb);
1546		else
1547			plug = NULL;
1548		if (plug) {
1549			bio_list_add(&plug->pending, mbio);
1550			plug->pending_cnt++;
1551		} else {
1552			spin_lock_irqsave(&conf->device_lock, flags);
1553			bio_list_add(&conf->pending_bio_list, mbio);
1554			conf->pending_count++;
1555			spin_unlock_irqrestore(&conf->device_lock, flags);
1556			md_wakeup_thread(mddev->thread);
1557		}
1558	}
1559
1560	r1_bio_write_done(r1_bio);
1561
1562	/* In case raid1d snuck in to freeze_array */
1563	wake_up(&conf->wait_barrier);
1564}
1565
1566static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1567{
1568	sector_t sectors;
1569
1570	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1571		md_flush_request(mddev, bio);
1572		return true;
1573	}
1574
1575	/*
1576	 * There is a limit to the maximum size, but
1577	 * the read/write handler might find a lower limit
1578	 * due to bad blocks.  To avoid multiple splits,
1579	 * we pass the maximum number of sectors down
1580	 * and let the lower level perform the split.
1581	 */
1582	sectors = align_to_barrier_unit_end(
1583		bio->bi_iter.bi_sector, bio_sectors(bio));
1584
1585	if (bio_data_dir(bio) == READ)
1586		raid1_read_request(mddev, bio, sectors, NULL);
1587	else {
1588		if (!md_write_start(mddev,bio))
1589			return false;
1590		raid1_write_request(mddev, bio, sectors);
1591	}
1592	return true;
1593}
1594
1595static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1596{
1597	struct r1conf *conf = mddev->private;
1598	int i;
1599
1600	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1601		   conf->raid_disks - mddev->degraded);
1602	rcu_read_lock();
1603	for (i = 0; i < conf->raid_disks; i++) {
1604		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1605		seq_printf(seq, "%s",
1606			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1607	}
1608	rcu_read_unlock();
1609	seq_printf(seq, "]");
1610}
1611
1612static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1613{
1614	char b[BDEVNAME_SIZE];
1615	struct r1conf *conf = mddev->private;
1616	unsigned long flags;
1617
1618	/*
1619	 * If it is not operational, then we have already marked it as dead
1620	 * else if it is the last working disks with "fail_last_dev == false",
1621	 * ignore the error, let the next level up know.
1622	 * else mark the drive as failed
1623	 */
1624	spin_lock_irqsave(&conf->device_lock, flags);
1625	if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1626	    && (conf->raid_disks - mddev->degraded) == 1) {
1627		/*
1628		 * Don't fail the drive, act as though we were just a
1629		 * normal single drive.
1630		 * However don't try a recovery from this drive as
1631		 * it is very likely to fail.
1632		 */
1633		conf->recovery_disabled = mddev->recovery_disabled;
1634		spin_unlock_irqrestore(&conf->device_lock, flags);
1635		return;
1636	}
1637	set_bit(Blocked, &rdev->flags);
1638	if (test_and_clear_bit(In_sync, &rdev->flags))
1639		mddev->degraded++;
1640	set_bit(Faulty, &rdev->flags);
 
 
1641	spin_unlock_irqrestore(&conf->device_lock, flags);
1642	/*
1643	 * if recovery is running, make sure it aborts.
1644	 */
1645	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1646	set_mask_bits(&mddev->sb_flags, 0,
1647		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1648	pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1649		"md/raid1:%s: Operation continuing on %d devices.\n",
1650		mdname(mddev), bdevname(rdev->bdev, b),
1651		mdname(mddev), conf->raid_disks - mddev->degraded);
1652}
1653
1654static void print_conf(struct r1conf *conf)
1655{
1656	int i;
1657
1658	pr_debug("RAID1 conf printout:\n");
1659	if (!conf) {
1660		pr_debug("(!conf)\n");
1661		return;
1662	}
1663	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1664		 conf->raid_disks);
1665
1666	rcu_read_lock();
1667	for (i = 0; i < conf->raid_disks; i++) {
1668		char b[BDEVNAME_SIZE];
1669		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1670		if (rdev)
1671			pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1672				 i, !test_bit(In_sync, &rdev->flags),
1673				 !test_bit(Faulty, &rdev->flags),
1674				 bdevname(rdev->bdev,b));
1675	}
1676	rcu_read_unlock();
1677}
1678
1679static void close_sync(struct r1conf *conf)
1680{
1681	int idx;
1682
1683	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1684		_wait_barrier(conf, idx);
1685		_allow_barrier(conf, idx);
1686	}
1687
1688	mempool_exit(&conf->r1buf_pool);
 
1689}
1690
1691static int raid1_spare_active(struct mddev *mddev)
1692{
1693	int i;
1694	struct r1conf *conf = mddev->private;
1695	int count = 0;
1696	unsigned long flags;
1697
1698	/*
1699	 * Find all failed disks within the RAID1 configuration
1700	 * and mark them readable.
1701	 * Called under mddev lock, so rcu protection not needed.
1702	 * device_lock used to avoid races with raid1_end_read_request
1703	 * which expects 'In_sync' flags and ->degraded to be consistent.
1704	 */
1705	spin_lock_irqsave(&conf->device_lock, flags);
1706	for (i = 0; i < conf->raid_disks; i++) {
1707		struct md_rdev *rdev = conf->mirrors[i].rdev;
1708		struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1709		if (repl
1710		    && !test_bit(Candidate, &repl->flags)
1711		    && repl->recovery_offset == MaxSector
1712		    && !test_bit(Faulty, &repl->flags)
1713		    && !test_and_set_bit(In_sync, &repl->flags)) {
1714			/* replacement has just become active */
1715			if (!rdev ||
1716			    !test_and_clear_bit(In_sync, &rdev->flags))
1717				count++;
1718			if (rdev) {
1719				/* Replaced device not technically
1720				 * faulty, but we need to be sure
1721				 * it gets removed and never re-added
1722				 */
1723				set_bit(Faulty, &rdev->flags);
1724				sysfs_notify_dirent_safe(
1725					rdev->sysfs_state);
1726			}
1727		}
1728		if (rdev
1729		    && rdev->recovery_offset == MaxSector
1730		    && !test_bit(Faulty, &rdev->flags)
1731		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1732			count++;
1733			sysfs_notify_dirent_safe(rdev->sysfs_state);
1734		}
1735	}
1736	mddev->degraded -= count;
1737	spin_unlock_irqrestore(&conf->device_lock, flags);
1738
1739	print_conf(conf);
1740	return count;
1741}
1742
1743static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1744{
1745	struct r1conf *conf = mddev->private;
1746	int err = -EEXIST;
1747	int mirror = 0;
1748	struct raid1_info *p;
1749	int first = 0;
1750	int last = conf->raid_disks - 1;
1751
1752	if (mddev->recovery_disabled == conf->recovery_disabled)
1753		return -EBUSY;
1754
1755	if (md_integrity_add_rdev(rdev, mddev))
1756		return -ENXIO;
1757
1758	if (rdev->raid_disk >= 0)
1759		first = last = rdev->raid_disk;
1760
1761	/*
1762	 * find the disk ... but prefer rdev->saved_raid_disk
1763	 * if possible.
1764	 */
1765	if (rdev->saved_raid_disk >= 0 &&
1766	    rdev->saved_raid_disk >= first &&
1767	    rdev->saved_raid_disk < conf->raid_disks &&
1768	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1769		first = last = rdev->saved_raid_disk;
1770
1771	for (mirror = first; mirror <= last; mirror++) {
1772		p = conf->mirrors + mirror;
1773		if (!p->rdev) {
 
1774			if (mddev->gendisk)
1775				disk_stack_limits(mddev->gendisk, rdev->bdev,
1776						  rdev->data_offset << 9);
1777
1778			p->head_position = 0;
1779			rdev->raid_disk = mirror;
1780			err = 0;
1781			/* As all devices are equivalent, we don't need a full recovery
1782			 * if this was recently any drive of the array
1783			 */
1784			if (rdev->saved_raid_disk < 0)
1785				conf->fullsync = 1;
1786			rcu_assign_pointer(p->rdev, rdev);
1787			break;
1788		}
1789		if (test_bit(WantReplacement, &p->rdev->flags) &&
1790		    p[conf->raid_disks].rdev == NULL) {
1791			/* Add this device as a replacement */
1792			clear_bit(In_sync, &rdev->flags);
1793			set_bit(Replacement, &rdev->flags);
1794			rdev->raid_disk = mirror;
1795			err = 0;
1796			conf->fullsync = 1;
1797			rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1798			break;
1799		}
1800	}
1801	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1802		blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1803	print_conf(conf);
1804	return err;
1805}
1806
1807static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1808{
1809	struct r1conf *conf = mddev->private;
1810	int err = 0;
1811	int number = rdev->raid_disk;
1812	struct raid1_info *p = conf->mirrors + number;
1813
1814	if (rdev != p->rdev)
1815		p = conf->mirrors + conf->raid_disks + number;
1816
1817	print_conf(conf);
1818	if (rdev == p->rdev) {
1819		if (test_bit(In_sync, &rdev->flags) ||
1820		    atomic_read(&rdev->nr_pending)) {
1821			err = -EBUSY;
1822			goto abort;
1823		}
1824		/* Only remove non-faulty devices if recovery
1825		 * is not possible.
1826		 */
1827		if (!test_bit(Faulty, &rdev->flags) &&
1828		    mddev->recovery_disabled != conf->recovery_disabled &&
1829		    mddev->degraded < conf->raid_disks) {
1830			err = -EBUSY;
1831			goto abort;
1832		}
1833		p->rdev = NULL;
1834		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1835			synchronize_rcu();
1836			if (atomic_read(&rdev->nr_pending)) {
1837				/* lost the race, try later */
1838				err = -EBUSY;
1839				p->rdev = rdev;
1840				goto abort;
1841			}
1842		}
1843		if (conf->mirrors[conf->raid_disks + number].rdev) {
1844			/* We just removed a device that is being replaced.
1845			 * Move down the replacement.  We drain all IO before
1846			 * doing this to avoid confusion.
1847			 */
1848			struct md_rdev *repl =
1849				conf->mirrors[conf->raid_disks + number].rdev;
1850			freeze_array(conf, 0);
1851			if (atomic_read(&repl->nr_pending)) {
1852				/* It means that some queued IO of retry_list
1853				 * hold repl. Thus, we cannot set replacement
1854				 * as NULL, avoiding rdev NULL pointer
1855				 * dereference in sync_request_write and
1856				 * handle_write_finished.
1857				 */
1858				err = -EBUSY;
1859				unfreeze_array(conf);
1860				goto abort;
1861			}
1862			clear_bit(Replacement, &repl->flags);
1863			p->rdev = repl;
1864			conf->mirrors[conf->raid_disks + number].rdev = NULL;
1865			unfreeze_array(conf);
1866		}
1867
1868		clear_bit(WantReplacement, &rdev->flags);
1869		err = md_integrity_register(mddev);
1870	}
1871abort:
1872
1873	print_conf(conf);
1874	return err;
1875}
1876
1877static void end_sync_read(struct bio *bio)
1878{
1879	struct r1bio *r1_bio = get_resync_r1bio(bio);
1880
1881	update_head_pos(r1_bio->read_disk, r1_bio);
1882
1883	/*
1884	 * we have read a block, now it needs to be re-written,
1885	 * or re-read if the read failed.
1886	 * We don't do much here, just schedule handling by raid1d
1887	 */
1888	if (!bio->bi_status)
1889		set_bit(R1BIO_Uptodate, &r1_bio->state);
1890
1891	if (atomic_dec_and_test(&r1_bio->remaining))
1892		reschedule_retry(r1_bio);
1893}
1894
1895static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1896{
1897	sector_t sync_blocks = 0;
1898	sector_t s = r1_bio->sector;
1899	long sectors_to_go = r1_bio->sectors;
1900
1901	/* make sure these bits don't get cleared. */
1902	do {
1903		md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1904		s += sync_blocks;
1905		sectors_to_go -= sync_blocks;
1906	} while (sectors_to_go > 0);
1907}
1908
1909static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1910{
1911	if (atomic_dec_and_test(&r1_bio->remaining)) {
1912		struct mddev *mddev = r1_bio->mddev;
1913		int s = r1_bio->sectors;
1914
1915		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1916		    test_bit(R1BIO_WriteError, &r1_bio->state))
1917			reschedule_retry(r1_bio);
1918		else {
1919			put_buf(r1_bio);
1920			md_done_sync(mddev, s, uptodate);
1921		}
1922	}
1923}
1924
1925static void end_sync_write(struct bio *bio)
1926{
1927	int uptodate = !bio->bi_status;
1928	struct r1bio *r1_bio = get_resync_r1bio(bio);
1929	struct mddev *mddev = r1_bio->mddev;
1930	struct r1conf *conf = mddev->private;
1931	sector_t first_bad;
1932	int bad_sectors;
1933	struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1934
1935	if (!uptodate) {
1936		abort_sync_write(mddev, r1_bio);
 
 
 
 
 
 
 
 
 
1937		set_bit(WriteErrorSeen, &rdev->flags);
1938		if (!test_and_set_bit(WantReplacement, &rdev->flags))
1939			set_bit(MD_RECOVERY_NEEDED, &
1940				mddev->recovery);
1941		set_bit(R1BIO_WriteError, &r1_bio->state);
1942	} else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1943			       &first_bad, &bad_sectors) &&
1944		   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1945				r1_bio->sector,
1946				r1_bio->sectors,
1947				&first_bad, &bad_sectors)
1948		)
1949		set_bit(R1BIO_MadeGood, &r1_bio->state);
1950
1951	put_sync_write_buf(r1_bio, uptodate);
 
 
 
 
 
 
 
 
 
1952}
1953
1954static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1955			    int sectors, struct page *page, int rw)
1956{
1957	if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1958		/* success */
1959		return 1;
1960	if (rw == WRITE) {
1961		set_bit(WriteErrorSeen, &rdev->flags);
1962		if (!test_and_set_bit(WantReplacement,
1963				      &rdev->flags))
1964			set_bit(MD_RECOVERY_NEEDED, &
1965				rdev->mddev->recovery);
1966	}
1967	/* need to record an error - either for the block or the device */
1968	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1969		md_error(rdev->mddev, rdev);
1970	return 0;
1971}
1972
1973static int fix_sync_read_error(struct r1bio *r1_bio)
1974{
1975	/* Try some synchronous reads of other devices to get
1976	 * good data, much like with normal read errors.  Only
1977	 * read into the pages we already have so we don't
1978	 * need to re-issue the read request.
1979	 * We don't need to freeze the array, because being in an
1980	 * active sync request, there is no normal IO, and
1981	 * no overlapping syncs.
1982	 * We don't need to check is_badblock() again as we
1983	 * made sure that anything with a bad block in range
1984	 * will have bi_end_io clear.
1985	 */
1986	struct mddev *mddev = r1_bio->mddev;
1987	struct r1conf *conf = mddev->private;
1988	struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1989	struct page **pages = get_resync_pages(bio)->pages;
1990	sector_t sect = r1_bio->sector;
1991	int sectors = r1_bio->sectors;
1992	int idx = 0;
1993	struct md_rdev *rdev;
1994
1995	rdev = conf->mirrors[r1_bio->read_disk].rdev;
1996	if (test_bit(FailFast, &rdev->flags)) {
1997		/* Don't try recovering from here - just fail it
1998		 * ... unless it is the last working device of course */
1999		md_error(mddev, rdev);
2000		if (test_bit(Faulty, &rdev->flags))
2001			/* Don't try to read from here, but make sure
2002			 * put_buf does it's thing
2003			 */
2004			bio->bi_end_io = end_sync_write;
2005	}
2006
2007	while(sectors) {
2008		int s = sectors;
2009		int d = r1_bio->read_disk;
2010		int success = 0;
2011		int start;
2012
2013		if (s > (PAGE_SIZE>>9))
2014			s = PAGE_SIZE >> 9;
2015		do {
2016			if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2017				/* No rcu protection needed here devices
2018				 * can only be removed when no resync is
2019				 * active, and resync is currently active
2020				 */
2021				rdev = conf->mirrors[d].rdev;
2022				if (sync_page_io(rdev, sect, s<<9,
2023						 pages[idx],
2024						 REQ_OP_READ, 0, false)) {
2025					success = 1;
2026					break;
2027				}
2028			}
2029			d++;
2030			if (d == conf->raid_disks * 2)
2031				d = 0;
2032		} while (!success && d != r1_bio->read_disk);
2033
2034		if (!success) {
2035			char b[BDEVNAME_SIZE];
2036			int abort = 0;
2037			/* Cannot read from anywhere, this block is lost.
2038			 * Record a bad block on each device.  If that doesn't
2039			 * work just disable and interrupt the recovery.
2040			 * Don't fail devices as that won't really help.
2041			 */
2042			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2043					    mdname(mddev), bio_devname(bio, b),
2044					    (unsigned long long)r1_bio->sector);
2045			for (d = 0; d < conf->raid_disks * 2; d++) {
2046				rdev = conf->mirrors[d].rdev;
2047				if (!rdev || test_bit(Faulty, &rdev->flags))
2048					continue;
2049				if (!rdev_set_badblocks(rdev, sect, s, 0))
2050					abort = 1;
2051			}
2052			if (abort) {
2053				conf->recovery_disabled =
2054					mddev->recovery_disabled;
2055				set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2056				md_done_sync(mddev, r1_bio->sectors, 0);
2057				put_buf(r1_bio);
2058				return 0;
2059			}
2060			/* Try next page */
2061			sectors -= s;
2062			sect += s;
2063			idx++;
2064			continue;
2065		}
2066
2067		start = d;
2068		/* write it back and re-read */
2069		while (d != r1_bio->read_disk) {
2070			if (d == 0)
2071				d = conf->raid_disks * 2;
2072			d--;
2073			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2074				continue;
2075			rdev = conf->mirrors[d].rdev;
2076			if (r1_sync_page_io(rdev, sect, s,
2077					    pages[idx],
2078					    WRITE) == 0) {
2079				r1_bio->bios[d]->bi_end_io = NULL;
2080				rdev_dec_pending(rdev, mddev);
2081			}
2082		}
2083		d = start;
2084		while (d != r1_bio->read_disk) {
2085			if (d == 0)
2086				d = conf->raid_disks * 2;
2087			d--;
2088			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2089				continue;
2090			rdev = conf->mirrors[d].rdev;
2091			if (r1_sync_page_io(rdev, sect, s,
2092					    pages[idx],
2093					    READ) != 0)
2094				atomic_add(s, &rdev->corrected_errors);
2095		}
2096		sectors -= s;
2097		sect += s;
2098		idx ++;
2099	}
2100	set_bit(R1BIO_Uptodate, &r1_bio->state);
2101	bio->bi_status = 0;
2102	return 1;
2103}
2104
2105static void process_checks(struct r1bio *r1_bio)
2106{
2107	/* We have read all readable devices.  If we haven't
2108	 * got the block, then there is no hope left.
2109	 * If we have, then we want to do a comparison
2110	 * and skip the write if everything is the same.
2111	 * If any blocks failed to read, then we need to
2112	 * attempt an over-write
2113	 */
2114	struct mddev *mddev = r1_bio->mddev;
2115	struct r1conf *conf = mddev->private;
2116	int primary;
2117	int i;
2118	int vcnt;
2119
2120	/* Fix variable parts of all bios */
2121	vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2122	for (i = 0; i < conf->raid_disks * 2; i++) {
2123		blk_status_t status;
2124		struct bio *b = r1_bio->bios[i];
2125		struct resync_pages *rp = get_resync_pages(b);
2126		if (b->bi_end_io != end_sync_read)
2127			continue;
2128		/* fixup the bio for reuse, but preserve errno */
2129		status = b->bi_status;
2130		bio_reset(b);
2131		b->bi_status = status;
2132		b->bi_iter.bi_sector = r1_bio->sector +
2133			conf->mirrors[i].rdev->data_offset;
2134		bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2135		b->bi_end_io = end_sync_read;
2136		rp->raid_bio = r1_bio;
2137		b->bi_private = rp;
2138
2139		/* initialize bvec table again */
2140		md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2141	}
2142	for (primary = 0; primary < conf->raid_disks * 2; primary++)
2143		if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2144		    !r1_bio->bios[primary]->bi_status) {
2145			r1_bio->bios[primary]->bi_end_io = NULL;
2146			rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2147			break;
2148		}
2149	r1_bio->read_disk = primary;
2150	for (i = 0; i < conf->raid_disks * 2; i++) {
2151		int j = 0;
2152		struct bio *pbio = r1_bio->bios[primary];
2153		struct bio *sbio = r1_bio->bios[i];
2154		blk_status_t status = sbio->bi_status;
2155		struct page **ppages = get_resync_pages(pbio)->pages;
2156		struct page **spages = get_resync_pages(sbio)->pages;
2157		struct bio_vec *bi;
2158		int page_len[RESYNC_PAGES] = { 0 };
2159		struct bvec_iter_all iter_all;
2160
2161		if (sbio->bi_end_io != end_sync_read)
2162			continue;
2163		/* Now we can 'fixup' the error value */
2164		sbio->bi_status = 0;
2165
2166		bio_for_each_segment_all(bi, sbio, iter_all)
2167			page_len[j++] = bi->bv_len;
2168
2169		if (!status) {
2170			for (j = vcnt; j-- ; ) {
2171				if (memcmp(page_address(ppages[j]),
2172					   page_address(spages[j]),
2173					   page_len[j]))
2174					break;
2175			}
2176		} else
2177			j = 0;
2178		if (j >= 0)
2179			atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2180		if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2181			      && !status)) {
2182			/* No need to write to this device. */
2183			sbio->bi_end_io = NULL;
2184			rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2185			continue;
2186		}
2187
2188		bio_copy_data(sbio, pbio);
2189	}
2190}
2191
2192static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2193{
2194	struct r1conf *conf = mddev->private;
2195	int i;
2196	int disks = conf->raid_disks * 2;
2197	struct bio *wbio;
2198
2199	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2200		/* ouch - failed to read all of that. */
2201		if (!fix_sync_read_error(r1_bio))
2202			return;
2203
2204	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2205		process_checks(r1_bio);
2206
2207	/*
2208	 * schedule writes
2209	 */
2210	atomic_set(&r1_bio->remaining, 1);
2211	for (i = 0; i < disks ; i++) {
2212		wbio = r1_bio->bios[i];
2213		if (wbio->bi_end_io == NULL ||
2214		    (wbio->bi_end_io == end_sync_read &&
2215		     (i == r1_bio->read_disk ||
2216		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2217			continue;
2218		if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2219			abort_sync_write(mddev, r1_bio);
2220			continue;
2221		}
2222
2223		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2224		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2225			wbio->bi_opf |= MD_FAILFAST;
2226
2227		wbio->bi_end_io = end_sync_write;
2228		atomic_inc(&r1_bio->remaining);
2229		md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2230
2231		generic_make_request(wbio);
2232	}
2233
2234	put_sync_write_buf(r1_bio, 1);
 
 
 
 
 
 
 
 
 
 
2235}
2236
2237/*
2238 * This is a kernel thread which:
2239 *
2240 *	1.	Retries failed read operations on working mirrors.
2241 *	2.	Updates the raid superblock when problems encounter.
2242 *	3.	Performs writes following reads for array synchronising.
2243 */
2244
2245static void fix_read_error(struct r1conf *conf, int read_disk,
2246			   sector_t sect, int sectors)
2247{
2248	struct mddev *mddev = conf->mddev;
2249	while(sectors) {
2250		int s = sectors;
2251		int d = read_disk;
2252		int success = 0;
2253		int start;
2254		struct md_rdev *rdev;
2255
2256		if (s > (PAGE_SIZE>>9))
2257			s = PAGE_SIZE >> 9;
2258
2259		do {
2260			sector_t first_bad;
2261			int bad_sectors;
2262
2263			rcu_read_lock();
2264			rdev = rcu_dereference(conf->mirrors[d].rdev);
2265			if (rdev &&
2266			    (test_bit(In_sync, &rdev->flags) ||
2267			     (!test_bit(Faulty, &rdev->flags) &&
2268			      rdev->recovery_offset >= sect + s)) &&
2269			    is_badblock(rdev, sect, s,
2270					&first_bad, &bad_sectors) == 0) {
2271				atomic_inc(&rdev->nr_pending);
2272				rcu_read_unlock();
2273				if (sync_page_io(rdev, sect, s<<9,
2274					 conf->tmppage, REQ_OP_READ, 0, false))
2275					success = 1;
2276				rdev_dec_pending(rdev, mddev);
2277				if (success)
2278					break;
2279			} else
2280				rcu_read_unlock();
2281			d++;
2282			if (d == conf->raid_disks * 2)
2283				d = 0;
2284		} while (!success && d != read_disk);
2285
2286		if (!success) {
2287			/* Cannot read from anywhere - mark it bad */
2288			struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2289			if (!rdev_set_badblocks(rdev, sect, s, 0))
2290				md_error(mddev, rdev);
2291			break;
2292		}
2293		/* write it back and re-read */
2294		start = d;
2295		while (d != read_disk) {
2296			if (d==0)
2297				d = conf->raid_disks * 2;
2298			d--;
2299			rcu_read_lock();
2300			rdev = rcu_dereference(conf->mirrors[d].rdev);
2301			if (rdev &&
2302			    !test_bit(Faulty, &rdev->flags)) {
2303				atomic_inc(&rdev->nr_pending);
2304				rcu_read_unlock();
2305				r1_sync_page_io(rdev, sect, s,
2306						conf->tmppage, WRITE);
2307				rdev_dec_pending(rdev, mddev);
2308			} else
2309				rcu_read_unlock();
2310		}
2311		d = start;
2312		while (d != read_disk) {
2313			char b[BDEVNAME_SIZE];
2314			if (d==0)
2315				d = conf->raid_disks * 2;
2316			d--;
2317			rcu_read_lock();
2318			rdev = rcu_dereference(conf->mirrors[d].rdev);
2319			if (rdev &&
2320			    !test_bit(Faulty, &rdev->flags)) {
2321				atomic_inc(&rdev->nr_pending);
2322				rcu_read_unlock();
2323				if (r1_sync_page_io(rdev, sect, s,
2324						    conf->tmppage, READ)) {
2325					atomic_add(s, &rdev->corrected_errors);
2326					pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2327						mdname(mddev), s,
2328						(unsigned long long)(sect +
2329								     rdev->data_offset),
2330						bdevname(rdev->bdev, b));
2331				}
2332				rdev_dec_pending(rdev, mddev);
2333			} else
2334				rcu_read_unlock();
2335		}
2336		sectors -= s;
2337		sect += s;
2338	}
2339}
2340
2341static int narrow_write_error(struct r1bio *r1_bio, int i)
2342{
2343	struct mddev *mddev = r1_bio->mddev;
2344	struct r1conf *conf = mddev->private;
2345	struct md_rdev *rdev = conf->mirrors[i].rdev;
2346
2347	/* bio has the data to be written to device 'i' where
2348	 * we just recently had a write error.
2349	 * We repeatedly clone the bio and trim down to one block,
2350	 * then try the write.  Where the write fails we record
2351	 * a bad block.
2352	 * It is conceivable that the bio doesn't exactly align with
2353	 * blocks.  We must handle this somehow.
2354	 *
2355	 * We currently own a reference on the rdev.
2356	 */
2357
2358	int block_sectors;
2359	sector_t sector;
2360	int sectors;
2361	int sect_to_write = r1_bio->sectors;
2362	int ok = 1;
2363
2364	if (rdev->badblocks.shift < 0)
2365		return 0;
2366
2367	block_sectors = roundup(1 << rdev->badblocks.shift,
2368				bdev_logical_block_size(rdev->bdev) >> 9);
2369	sector = r1_bio->sector;
2370	sectors = ((sector + block_sectors)
2371		   & ~(sector_t)(block_sectors - 1))
2372		- sector;
2373
2374	while (sect_to_write) {
2375		struct bio *wbio;
2376		if (sectors > sect_to_write)
2377			sectors = sect_to_write;
2378		/* Write at 'sector' for 'sectors'*/
2379
2380		if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2381			wbio = bio_clone_fast(r1_bio->behind_master_bio,
2382					      GFP_NOIO,
2383					      &mddev->bio_set);
2384		} else {
2385			wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2386					      &mddev->bio_set);
2387		}
2388
2389		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2390		wbio->bi_iter.bi_sector = r1_bio->sector;
2391		wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2392
2393		bio_trim(wbio, sector - r1_bio->sector, sectors);
2394		wbio->bi_iter.bi_sector += rdev->data_offset;
2395		bio_set_dev(wbio, rdev->bdev);
2396
2397		if (submit_bio_wait(wbio) < 0)
2398			/* failure! */
2399			ok = rdev_set_badblocks(rdev, sector,
2400						sectors, 0)
2401				&& ok;
2402
2403		bio_put(wbio);
2404		sect_to_write -= sectors;
2405		sector += sectors;
2406		sectors = block_sectors;
2407	}
2408	return ok;
2409}
2410
2411static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2412{
2413	int m;
2414	int s = r1_bio->sectors;
2415	for (m = 0; m < conf->raid_disks * 2 ; m++) {
2416		struct md_rdev *rdev = conf->mirrors[m].rdev;
2417		struct bio *bio = r1_bio->bios[m];
2418		if (bio->bi_end_io == NULL)
2419			continue;
2420		if (!bio->bi_status &&
2421		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2422			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2423		}
2424		if (bio->bi_status &&
2425		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2426			if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2427				md_error(conf->mddev, rdev);
2428		}
2429	}
2430	put_buf(r1_bio);
2431	md_done_sync(conf->mddev, s, 1);
2432}
2433
2434static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2435{
2436	int m, idx;
2437	bool fail = false;
2438
2439	for (m = 0; m < conf->raid_disks * 2 ; m++)
2440		if (r1_bio->bios[m] == IO_MADE_GOOD) {
2441			struct md_rdev *rdev = conf->mirrors[m].rdev;
2442			rdev_clear_badblocks(rdev,
2443					     r1_bio->sector,
2444					     r1_bio->sectors, 0);
2445			rdev_dec_pending(rdev, conf->mddev);
2446		} else if (r1_bio->bios[m] != NULL) {
2447			/* This drive got a write error.  We need to
2448			 * narrow down and record precise write
2449			 * errors.
2450			 */
2451			fail = true;
2452			if (!narrow_write_error(r1_bio, m)) {
2453				md_error(conf->mddev,
2454					 conf->mirrors[m].rdev);
2455				/* an I/O failed, we can't clear the bitmap */
2456				set_bit(R1BIO_Degraded, &r1_bio->state);
2457			}
2458			rdev_dec_pending(conf->mirrors[m].rdev,
2459					 conf->mddev);
2460		}
2461	if (fail) {
2462		spin_lock_irq(&conf->device_lock);
2463		list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2464		idx = sector_to_idx(r1_bio->sector);
2465		atomic_inc(&conf->nr_queued[idx]);
2466		spin_unlock_irq(&conf->device_lock);
2467		/*
2468		 * In case freeze_array() is waiting for condition
2469		 * get_unqueued_pending() == extra to be true.
2470		 */
2471		wake_up(&conf->wait_barrier);
2472		md_wakeup_thread(conf->mddev->thread);
2473	} else {
2474		if (test_bit(R1BIO_WriteError, &r1_bio->state))
2475			close_write(r1_bio);
2476		raid_end_bio_io(r1_bio);
2477	}
2478}
2479
2480static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2481{
2482	struct mddev *mddev = conf->mddev;
2483	struct bio *bio;
2484	struct md_rdev *rdev;
 
2485
2486	clear_bit(R1BIO_ReadError, &r1_bio->state);
2487	/* we got a read error. Maybe the drive is bad.  Maybe just
2488	 * the block and we can fix it.
2489	 * We freeze all other IO, and try reading the block from
2490	 * other devices.  When we find one, we re-write
2491	 * and check it that fixes the read error.
2492	 * This is all done synchronously while the array is
2493	 * frozen
2494	 */
2495
2496	bio = r1_bio->bios[r1_bio->read_disk];
 
2497	bio_put(bio);
2498	r1_bio->bios[r1_bio->read_disk] = NULL;
2499
2500	rdev = conf->mirrors[r1_bio->read_disk].rdev;
2501	if (mddev->ro == 0
2502	    && !test_bit(FailFast, &rdev->flags)) {
2503		freeze_array(conf, 1);
2504		fix_read_error(conf, r1_bio->read_disk,
2505			       r1_bio->sector, r1_bio->sectors);
2506		unfreeze_array(conf);
2507	} else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2508		md_error(mddev, rdev);
2509	} else {
2510		r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2511	}
2512
2513	rdev_dec_pending(rdev, conf->mddev);
2514	allow_barrier(conf, r1_bio->sector);
2515	bio = r1_bio->master_bio;
2516
2517	/* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2518	r1_bio->state = 0;
2519	raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2520}
2521
2522static void raid1d(struct md_thread *thread)
2523{
2524	struct mddev *mddev = thread->mddev;
2525	struct r1bio *r1_bio;
2526	unsigned long flags;
2527	struct r1conf *conf = mddev->private;
2528	struct list_head *head = &conf->retry_list;
2529	struct blk_plug plug;
2530	int idx;
2531
2532	md_check_recovery(mddev);
2533
2534	if (!list_empty_careful(&conf->bio_end_io_list) &&
2535	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2536		LIST_HEAD(tmp);
2537		spin_lock_irqsave(&conf->device_lock, flags);
2538		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2539			list_splice_init(&conf->bio_end_io_list, &tmp);
2540		spin_unlock_irqrestore(&conf->device_lock, flags);
2541		while (!list_empty(&tmp)) {
2542			r1_bio = list_first_entry(&tmp, struct r1bio,
2543						  retry_list);
2544			list_del(&r1_bio->retry_list);
2545			idx = sector_to_idx(r1_bio->sector);
2546			atomic_dec(&conf->nr_queued[idx]);
2547			if (mddev->degraded)
2548				set_bit(R1BIO_Degraded, &r1_bio->state);
2549			if (test_bit(R1BIO_WriteError, &r1_bio->state))
2550				close_write(r1_bio);
2551			raid_end_bio_io(r1_bio);
2552		}
2553	}
2554
2555	blk_start_plug(&plug);
2556	for (;;) {
2557
2558		flush_pending_writes(conf);
2559
2560		spin_lock_irqsave(&conf->device_lock, flags);
2561		if (list_empty(head)) {
2562			spin_unlock_irqrestore(&conf->device_lock, flags);
2563			break;
2564		}
2565		r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2566		list_del(head->prev);
2567		idx = sector_to_idx(r1_bio->sector);
2568		atomic_dec(&conf->nr_queued[idx]);
2569		spin_unlock_irqrestore(&conf->device_lock, flags);
2570
2571		mddev = r1_bio->mddev;
2572		conf = mddev->private;
2573		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2574			if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2575			    test_bit(R1BIO_WriteError, &r1_bio->state))
2576				handle_sync_write_finished(conf, r1_bio);
2577			else
2578				sync_request_write(mddev, r1_bio);
2579		} else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2580			   test_bit(R1BIO_WriteError, &r1_bio->state))
2581			handle_write_finished(conf, r1_bio);
2582		else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2583			handle_read_error(conf, r1_bio);
2584		else
2585			WARN_ON_ONCE(1);
2586
2587		cond_resched();
2588		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2589			md_check_recovery(mddev);
2590	}
2591	blk_finish_plug(&plug);
2592}
2593
2594static int init_resync(struct r1conf *conf)
2595{
2596	int buffs;
2597
2598	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2599	BUG_ON(mempool_initialized(&conf->r1buf_pool));
2600
2601	return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2602			    r1buf_pool_free, conf->poolinfo);
 
 
2603}
2604
2605static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2606{
2607	struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2608	struct resync_pages *rps;
2609	struct bio *bio;
2610	int i;
2611
2612	for (i = conf->poolinfo->raid_disks; i--; ) {
2613		bio = r1bio->bios[i];
2614		rps = bio->bi_private;
2615		bio_reset(bio);
2616		bio->bi_private = rps;
2617	}
2618	r1bio->master_bio = NULL;
2619	return r1bio;
2620}
2621
2622/*
2623 * perform a "sync" on one "block"
2624 *
2625 * We need to make sure that no normal I/O request - particularly write
2626 * requests - conflict with active sync requests.
2627 *
2628 * This is achieved by tracking pending requests and a 'barrier' concept
2629 * that can be installed to exclude normal IO requests.
2630 */
2631
2632static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2633				   int *skipped)
2634{
2635	struct r1conf *conf = mddev->private;
2636	struct r1bio *r1_bio;
2637	struct bio *bio;
2638	sector_t max_sector, nr_sectors;
2639	int disk = -1;
2640	int i;
2641	int wonly = -1;
2642	int write_targets = 0, read_targets = 0;
2643	sector_t sync_blocks;
2644	int still_degraded = 0;
2645	int good_sectors = RESYNC_SECTORS;
2646	int min_bad = 0; /* number of sectors that are bad in all devices */
2647	int idx = sector_to_idx(sector_nr);
2648	int page_idx = 0;
2649
2650	if (!mempool_initialized(&conf->r1buf_pool))
2651		if (init_resync(conf))
2652			return 0;
2653
2654	max_sector = mddev->dev_sectors;
2655	if (sector_nr >= max_sector) {
2656		/* If we aborted, we need to abort the
2657		 * sync on the 'current' bitmap chunk (there will
2658		 * only be one in raid1 resync.
2659		 * We can find the current addess in mddev->curr_resync
2660		 */
2661		if (mddev->curr_resync < max_sector) /* aborted */
2662			md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2663					   &sync_blocks, 1);
2664		else /* completed sync */
2665			conf->fullsync = 0;
2666
2667		md_bitmap_close_sync(mddev->bitmap);
2668		close_sync(conf);
2669
2670		if (mddev_is_clustered(mddev)) {
2671			conf->cluster_sync_low = 0;
2672			conf->cluster_sync_high = 0;
2673		}
2674		return 0;
2675	}
2676
2677	if (mddev->bitmap == NULL &&
2678	    mddev->recovery_cp == MaxSector &&
2679	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2680	    conf->fullsync == 0) {
2681		*skipped = 1;
2682		return max_sector - sector_nr;
2683	}
2684	/* before building a request, check if we can skip these blocks..
2685	 * This call the bitmap_start_sync doesn't actually record anything
2686	 */
2687	if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2688	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2689		/* We can skip this block, and probably several more */
2690		*skipped = 1;
2691		return sync_blocks;
2692	}
2693
2694	/*
2695	 * If there is non-resync activity waiting for a turn, then let it
2696	 * though before starting on this new sync request.
2697	 */
2698	if (atomic_read(&conf->nr_waiting[idx]))
2699		schedule_timeout_uninterruptible(1);
2700
2701	/* we are incrementing sector_nr below. To be safe, we check against
2702	 * sector_nr + two times RESYNC_SECTORS
2703	 */
2704
2705	md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2706		mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2707
2708
2709	if (raise_barrier(conf, sector_nr))
2710		return 0;
2711
2712	r1_bio = raid1_alloc_init_r1buf(conf);
2713
2714	rcu_read_lock();
2715	/*
2716	 * If we get a correctably read error during resync or recovery,
2717	 * we might want to read from a different device.  So we
2718	 * flag all drives that could conceivably be read from for READ,
2719	 * and any others (which will be non-In_sync devices) for WRITE.
2720	 * If a read fails, we try reading from something else for which READ
2721	 * is OK.
2722	 */
2723
2724	r1_bio->mddev = mddev;
2725	r1_bio->sector = sector_nr;
2726	r1_bio->state = 0;
2727	set_bit(R1BIO_IsSync, &r1_bio->state);
2728	/* make sure good_sectors won't go across barrier unit boundary */
2729	good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2730
2731	for (i = 0; i < conf->raid_disks * 2; i++) {
2732		struct md_rdev *rdev;
2733		bio = r1_bio->bios[i];
2734
2735		rdev = rcu_dereference(conf->mirrors[i].rdev);
2736		if (rdev == NULL ||
2737		    test_bit(Faulty, &rdev->flags)) {
2738			if (i < conf->raid_disks)
2739				still_degraded = 1;
2740		} else if (!test_bit(In_sync, &rdev->flags)) {
2741			bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2742			bio->bi_end_io = end_sync_write;
2743			write_targets ++;
2744		} else {
2745			/* may need to read from here */
2746			sector_t first_bad = MaxSector;
2747			int bad_sectors;
2748
2749			if (is_badblock(rdev, sector_nr, good_sectors,
2750					&first_bad, &bad_sectors)) {
2751				if (first_bad > sector_nr)
2752					good_sectors = first_bad - sector_nr;
2753				else {
2754					bad_sectors -= (sector_nr - first_bad);
2755					if (min_bad == 0 ||
2756					    min_bad > bad_sectors)
2757						min_bad = bad_sectors;
2758				}
2759			}
2760			if (sector_nr < first_bad) {
2761				if (test_bit(WriteMostly, &rdev->flags)) {
2762					if (wonly < 0)
2763						wonly = i;
2764				} else {
2765					if (disk < 0)
2766						disk = i;
2767				}
2768				bio_set_op_attrs(bio, REQ_OP_READ, 0);
2769				bio->bi_end_io = end_sync_read;
2770				read_targets++;
2771			} else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2772				test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2773				!test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2774				/*
2775				 * The device is suitable for reading (InSync),
2776				 * but has bad block(s) here. Let's try to correct them,
2777				 * if we are doing resync or repair. Otherwise, leave
2778				 * this device alone for this sync request.
2779				 */
2780				bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2781				bio->bi_end_io = end_sync_write;
2782				write_targets++;
2783			}
2784		}
2785		if (bio->bi_end_io) {
2786			atomic_inc(&rdev->nr_pending);
2787			bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2788			bio_set_dev(bio, rdev->bdev);
2789			if (test_bit(FailFast, &rdev->flags))
2790				bio->bi_opf |= MD_FAILFAST;
2791		}
2792	}
2793	rcu_read_unlock();
2794	if (disk < 0)
2795		disk = wonly;
2796	r1_bio->read_disk = disk;
2797
2798	if (read_targets == 0 && min_bad > 0) {
2799		/* These sectors are bad on all InSync devices, so we
2800		 * need to mark them bad on all write targets
2801		 */
2802		int ok = 1;
2803		for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2804			if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2805				struct md_rdev *rdev = conf->mirrors[i].rdev;
2806				ok = rdev_set_badblocks(rdev, sector_nr,
2807							min_bad, 0
2808					) && ok;
2809			}
2810		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2811		*skipped = 1;
2812		put_buf(r1_bio);
2813
2814		if (!ok) {
2815			/* Cannot record the badblocks, so need to
2816			 * abort the resync.
2817			 * If there are multiple read targets, could just
2818			 * fail the really bad ones ???
2819			 */
2820			conf->recovery_disabled = mddev->recovery_disabled;
2821			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2822			return 0;
2823		} else
2824			return min_bad;
2825
2826	}
2827	if (min_bad > 0 && min_bad < good_sectors) {
2828		/* only resync enough to reach the next bad->good
2829		 * transition */
2830		good_sectors = min_bad;
2831	}
2832
2833	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2834		/* extra read targets are also write targets */
2835		write_targets += read_targets-1;
2836
2837	if (write_targets == 0 || read_targets == 0) {
2838		/* There is nowhere to write, so all non-sync
2839		 * drives must be failed - so we are finished
2840		 */
2841		sector_t rv;
2842		if (min_bad > 0)
2843			max_sector = sector_nr + min_bad;
2844		rv = max_sector - sector_nr;
2845		*skipped = 1;
2846		put_buf(r1_bio);
2847		return rv;
2848	}
2849
2850	if (max_sector > mddev->resync_max)
2851		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2852	if (max_sector > sector_nr + good_sectors)
2853		max_sector = sector_nr + good_sectors;
2854	nr_sectors = 0;
2855	sync_blocks = 0;
2856	do {
2857		struct page *page;
2858		int len = PAGE_SIZE;
2859		if (sector_nr + (len>>9) > max_sector)
2860			len = (max_sector - sector_nr) << 9;
2861		if (len == 0)
2862			break;
2863		if (sync_blocks == 0) {
2864			if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2865						  &sync_blocks, still_degraded) &&
2866			    !conf->fullsync &&
2867			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2868				break;
2869			if ((len >> 9) > sync_blocks)
2870				len = sync_blocks<<9;
2871		}
2872
2873		for (i = 0 ; i < conf->raid_disks * 2; i++) {
2874			struct resync_pages *rp;
2875
2876			bio = r1_bio->bios[i];
2877			rp = get_resync_pages(bio);
2878			if (bio->bi_end_io) {
2879				page = resync_fetch_page(rp, page_idx);
2880
2881				/*
2882				 * won't fail because the vec table is big
2883				 * enough to hold all these pages
2884				 */
2885				bio_add_page(bio, page, len, 0);
2886			}
2887		}
2888		nr_sectors += len>>9;
2889		sector_nr += len>>9;
2890		sync_blocks -= (len>>9);
2891	} while (++page_idx < RESYNC_PAGES);
2892
2893	r1_bio->sectors = nr_sectors;
2894
2895	if (mddev_is_clustered(mddev) &&
2896			conf->cluster_sync_high < sector_nr + nr_sectors) {
2897		conf->cluster_sync_low = mddev->curr_resync_completed;
2898		conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2899		/* Send resync message */
2900		md_cluster_ops->resync_info_update(mddev,
2901				conf->cluster_sync_low,
2902				conf->cluster_sync_high);
2903	}
2904
2905	/* For a user-requested sync, we read all readable devices and do a
2906	 * compare
2907	 */
2908	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2909		atomic_set(&r1_bio->remaining, read_targets);
2910		for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2911			bio = r1_bio->bios[i];
2912			if (bio->bi_end_io == end_sync_read) {
2913				read_targets--;
2914				md_sync_acct_bio(bio, nr_sectors);
2915				if (read_targets == 1)
2916					bio->bi_opf &= ~MD_FAILFAST;
2917				generic_make_request(bio);
2918			}
2919		}
2920	} else {
2921		atomic_set(&r1_bio->remaining, 1);
2922		bio = r1_bio->bios[r1_bio->read_disk];
2923		md_sync_acct_bio(bio, nr_sectors);
2924		if (read_targets == 1)
2925			bio->bi_opf &= ~MD_FAILFAST;
2926		generic_make_request(bio);
 
2927	}
2928	return nr_sectors;
2929}
2930
2931static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2932{
2933	if (sectors)
2934		return sectors;
2935
2936	return mddev->dev_sectors;
2937}
2938
2939static struct r1conf *setup_conf(struct mddev *mddev)
2940{
2941	struct r1conf *conf;
2942	int i;
2943	struct raid1_info *disk;
2944	struct md_rdev *rdev;
2945	int err = -ENOMEM;
2946
2947	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2948	if (!conf)
2949		goto abort;
2950
2951	conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2952				   sizeof(atomic_t), GFP_KERNEL);
2953	if (!conf->nr_pending)
2954		goto abort;
2955
2956	conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2957				   sizeof(atomic_t), GFP_KERNEL);
2958	if (!conf->nr_waiting)
2959		goto abort;
2960
2961	conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2962				  sizeof(atomic_t), GFP_KERNEL);
2963	if (!conf->nr_queued)
2964		goto abort;
2965
2966	conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2967				sizeof(atomic_t), GFP_KERNEL);
2968	if (!conf->barrier)
2969		goto abort;
2970
2971	conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2972					    mddev->raid_disks, 2),
2973				GFP_KERNEL);
2974	if (!conf->mirrors)
2975		goto abort;
2976
2977	conf->tmppage = alloc_page(GFP_KERNEL);
2978	if (!conf->tmppage)
2979		goto abort;
2980
2981	conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2982	if (!conf->poolinfo)
2983		goto abort;
2984	conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2985	err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2986			   rbio_pool_free, conf->poolinfo);
2987	if (err)
 
2988		goto abort;
2989
2990	err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2991	if (err)
2992		goto abort;
2993
2994	conf->poolinfo->mddev = mddev;
2995
2996	err = -EINVAL;
2997	spin_lock_init(&conf->device_lock);
2998	rdev_for_each(rdev, mddev) {
2999		int disk_idx = rdev->raid_disk;
3000		if (disk_idx >= mddev->raid_disks
3001		    || disk_idx < 0)
3002			continue;
3003		if (test_bit(Replacement, &rdev->flags))
3004			disk = conf->mirrors + mddev->raid_disks + disk_idx;
3005		else
3006			disk = conf->mirrors + disk_idx;
3007
3008		if (disk->rdev)
3009			goto abort;
3010		disk->rdev = rdev;
3011		disk->head_position = 0;
3012		disk->seq_start = MaxSector;
3013	}
3014	conf->raid_disks = mddev->raid_disks;
3015	conf->mddev = mddev;
3016	INIT_LIST_HEAD(&conf->retry_list);
3017	INIT_LIST_HEAD(&conf->bio_end_io_list);
3018
3019	spin_lock_init(&conf->resync_lock);
3020	init_waitqueue_head(&conf->wait_barrier);
3021
3022	bio_list_init(&conf->pending_bio_list);
3023	conf->pending_count = 0;
3024	conf->recovery_disabled = mddev->recovery_disabled - 1;
3025
3026	err = -EIO;
3027	for (i = 0; i < conf->raid_disks * 2; i++) {
3028
3029		disk = conf->mirrors + i;
3030
3031		if (i < conf->raid_disks &&
3032		    disk[conf->raid_disks].rdev) {
3033			/* This slot has a replacement. */
3034			if (!disk->rdev) {
3035				/* No original, just make the replacement
3036				 * a recovering spare
3037				 */
3038				disk->rdev =
3039					disk[conf->raid_disks].rdev;
3040				disk[conf->raid_disks].rdev = NULL;
3041			} else if (!test_bit(In_sync, &disk->rdev->flags))
3042				/* Original is not in_sync - bad */
3043				goto abort;
3044		}
3045
3046		if (!disk->rdev ||
3047		    !test_bit(In_sync, &disk->rdev->flags)) {
3048			disk->head_position = 0;
3049			if (disk->rdev &&
3050			    (disk->rdev->saved_raid_disk < 0))
3051				conf->fullsync = 1;
3052		}
3053	}
3054
3055	err = -ENOMEM;
3056	conf->thread = md_register_thread(raid1d, mddev, "raid1");
3057	if (!conf->thread)
3058		goto abort;
3059
3060	return conf;
3061
3062 abort:
3063	if (conf) {
3064		mempool_exit(&conf->r1bio_pool);
3065		kfree(conf->mirrors);
3066		safe_put_page(conf->tmppage);
3067		kfree(conf->poolinfo);
3068		kfree(conf->nr_pending);
3069		kfree(conf->nr_waiting);
3070		kfree(conf->nr_queued);
3071		kfree(conf->barrier);
3072		bioset_exit(&conf->bio_split);
 
3073		kfree(conf);
3074	}
3075	return ERR_PTR(err);
3076}
3077
3078static void raid1_free(struct mddev *mddev, void *priv);
3079static int raid1_run(struct mddev *mddev)
3080{
3081	struct r1conf *conf;
3082	int i;
3083	struct md_rdev *rdev;
3084	int ret;
3085	bool discard_supported = false;
3086
3087	if (mddev->level != 1) {
3088		pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3089			mdname(mddev), mddev->level);
3090		return -EIO;
3091	}
3092	if (mddev->reshape_position != MaxSector) {
3093		pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3094			mdname(mddev));
3095		return -EIO;
3096	}
3097	if (mddev_init_writes_pending(mddev) < 0)
3098		return -ENOMEM;
3099	/*
3100	 * copy the already verified devices into our private RAID1
3101	 * bookkeeping area. [whatever we allocate in run(),
3102	 * should be freed in raid1_free()]
3103	 */
3104	if (mddev->private == NULL)
3105		conf = setup_conf(mddev);
3106	else
3107		conf = mddev->private;
3108
3109	if (IS_ERR(conf))
3110		return PTR_ERR(conf);
3111
3112	if (mddev->queue) {
3113		blk_queue_max_write_same_sectors(mddev->queue, 0);
3114		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3115	}
3116
3117	rdev_for_each(rdev, mddev) {
3118		if (!mddev->gendisk)
3119			continue;
3120		disk_stack_limits(mddev->gendisk, rdev->bdev,
3121				  rdev->data_offset << 9);
3122		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3123			discard_supported = true;
3124	}
3125
3126	mddev->degraded = 0;
3127	for (i = 0; i < conf->raid_disks; i++)
3128		if (conf->mirrors[i].rdev == NULL ||
3129		    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3130		    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3131			mddev->degraded++;
3132	/*
3133	 * RAID1 needs at least one disk in active
3134	 */
3135	if (conf->raid_disks - mddev->degraded < 1) {
3136		ret = -EINVAL;
3137		goto abort;
3138	}
3139
3140	if (conf->raid_disks - mddev->degraded == 1)
3141		mddev->recovery_cp = MaxSector;
3142
3143	if (mddev->recovery_cp != MaxSector)
3144		pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3145			mdname(mddev));
3146	pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3147		mdname(mddev), mddev->raid_disks - mddev->degraded,
3148		mddev->raid_disks);
3149
3150	/*
3151	 * Ok, everything is just fine now
3152	 */
3153	mddev->thread = conf->thread;
3154	conf->thread = NULL;
3155	mddev->private = conf;
3156	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3157
3158	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3159
3160	if (mddev->queue) {
3161		if (discard_supported)
3162			blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3163						mddev->queue);
3164		else
3165			blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3166						  mddev->queue);
3167	}
3168
3169	ret = md_integrity_register(mddev);
3170	if (ret) {
3171		md_unregister_thread(&mddev->thread);
3172		goto abort;
3173	}
3174	return 0;
3175
3176abort:
3177	raid1_free(mddev, conf);
3178	return ret;
3179}
3180
3181static void raid1_free(struct mddev *mddev, void *priv)
3182{
3183	struct r1conf *conf = priv;
3184
3185	mempool_exit(&conf->r1bio_pool);
3186	kfree(conf->mirrors);
3187	safe_put_page(conf->tmppage);
3188	kfree(conf->poolinfo);
3189	kfree(conf->nr_pending);
3190	kfree(conf->nr_waiting);
3191	kfree(conf->nr_queued);
3192	kfree(conf->barrier);
3193	bioset_exit(&conf->bio_split);
 
3194	kfree(conf);
3195}
3196
3197static int raid1_resize(struct mddev *mddev, sector_t sectors)
3198{
3199	/* no resync is happening, and there is enough space
3200	 * on all devices, so we can resize.
3201	 * We need to make sure resync covers any new space.
3202	 * If the array is shrinking we should possibly wait until
3203	 * any io in the removed space completes, but it hardly seems
3204	 * worth it.
3205	 */
3206	sector_t newsize = raid1_size(mddev, sectors, 0);
3207	if (mddev->external_size &&
3208	    mddev->array_sectors > newsize)
3209		return -EINVAL;
3210	if (mddev->bitmap) {
3211		int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3212		if (ret)
3213			return ret;
3214	}
3215	md_set_array_sectors(mddev, newsize);
3216	if (sectors > mddev->dev_sectors &&
3217	    mddev->recovery_cp > mddev->dev_sectors) {
3218		mddev->recovery_cp = mddev->dev_sectors;
3219		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3220	}
3221	mddev->dev_sectors = sectors;
3222	mddev->resync_max_sectors = sectors;
3223	return 0;
3224}
3225
3226static int raid1_reshape(struct mddev *mddev)
3227{
3228	/* We need to:
3229	 * 1/ resize the r1bio_pool
3230	 * 2/ resize conf->mirrors
3231	 *
3232	 * We allocate a new r1bio_pool if we can.
3233	 * Then raise a device barrier and wait until all IO stops.
3234	 * Then resize conf->mirrors and swap in the new r1bio pool.
3235	 *
3236	 * At the same time, we "pack" the devices so that all the missing
3237	 * devices have the higher raid_disk numbers.
3238	 */
3239	mempool_t newpool, oldpool;
3240	struct pool_info *newpoolinfo;
3241	struct raid1_info *newmirrors;
3242	struct r1conf *conf = mddev->private;
3243	int cnt, raid_disks;
3244	unsigned long flags;
3245	int d, d2;
3246	int ret;
3247
3248	memset(&newpool, 0, sizeof(newpool));
3249	memset(&oldpool, 0, sizeof(oldpool));
3250
3251	/* Cannot change chunk_size, layout, or level */
3252	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3253	    mddev->layout != mddev->new_layout ||
3254	    mddev->level != mddev->new_level) {
3255		mddev->new_chunk_sectors = mddev->chunk_sectors;
3256		mddev->new_layout = mddev->layout;
3257		mddev->new_level = mddev->level;
3258		return -EINVAL;
3259	}
3260
3261	if (!mddev_is_clustered(mddev))
3262		md_allow_write(mddev);
3263
3264	raid_disks = mddev->raid_disks + mddev->delta_disks;
3265
3266	if (raid_disks < conf->raid_disks) {
3267		cnt=0;
3268		for (d= 0; d < conf->raid_disks; d++)
3269			if (conf->mirrors[d].rdev)
3270				cnt++;
3271		if (cnt > raid_disks)
3272			return -EBUSY;
3273	}
3274
3275	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3276	if (!newpoolinfo)
3277		return -ENOMEM;
3278	newpoolinfo->mddev = mddev;
3279	newpoolinfo->raid_disks = raid_disks * 2;
3280
3281	ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3282			   rbio_pool_free, newpoolinfo);
3283	if (ret) {
3284		kfree(newpoolinfo);
3285		return ret;
3286	}
3287	newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3288					 raid_disks, 2),
3289			     GFP_KERNEL);
3290	if (!newmirrors) {
3291		kfree(newpoolinfo);
3292		mempool_exit(&newpool);
3293		return -ENOMEM;
3294	}
3295
3296	freeze_array(conf, 0);
3297
3298	/* ok, everything is stopped */
3299	oldpool = conf->r1bio_pool;
3300	conf->r1bio_pool = newpool;
3301
3302	for (d = d2 = 0; d < conf->raid_disks; d++) {
3303		struct md_rdev *rdev = conf->mirrors[d].rdev;
3304		if (rdev && rdev->raid_disk != d2) {
3305			sysfs_unlink_rdev(mddev, rdev);
3306			rdev->raid_disk = d2;
3307			sysfs_unlink_rdev(mddev, rdev);
3308			if (sysfs_link_rdev(mddev, rdev))
3309				pr_warn("md/raid1:%s: cannot register rd%d\n",
3310					mdname(mddev), rdev->raid_disk);
3311		}
3312		if (rdev)
3313			newmirrors[d2++].rdev = rdev;
3314	}
3315	kfree(conf->mirrors);
3316	conf->mirrors = newmirrors;
3317	kfree(conf->poolinfo);
3318	conf->poolinfo = newpoolinfo;
3319
3320	spin_lock_irqsave(&conf->device_lock, flags);
3321	mddev->degraded += (raid_disks - conf->raid_disks);
3322	spin_unlock_irqrestore(&conf->device_lock, flags);
3323	conf->raid_disks = mddev->raid_disks = raid_disks;
3324	mddev->delta_disks = 0;
3325
3326	unfreeze_array(conf);
3327
3328	set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3329	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3330	md_wakeup_thread(mddev->thread);
3331
3332	mempool_exit(&oldpool);
3333	return 0;
3334}
3335
3336static void raid1_quiesce(struct mddev *mddev, int quiesce)
3337{
3338	struct r1conf *conf = mddev->private;
3339
3340	if (quiesce)
3341		freeze_array(conf, 0);
3342	else
3343		unfreeze_array(conf);
3344}
3345
3346static void *raid1_takeover(struct mddev *mddev)
3347{
3348	/* raid1 can take over:
3349	 *  raid5 with 2 devices, any layout or chunk size
3350	 */
3351	if (mddev->level == 5 && mddev->raid_disks == 2) {
3352		struct r1conf *conf;
3353		mddev->new_level = 1;
3354		mddev->new_layout = 0;
3355		mddev->new_chunk_sectors = 0;
3356		conf = setup_conf(mddev);
3357		if (!IS_ERR(conf)) {
3358			/* Array must appear to be quiesced */
3359			conf->array_frozen = 1;
3360			mddev_clear_unsupported_flags(mddev,
3361				UNSUPPORTED_MDDEV_FLAGS);
3362		}
3363		return conf;
3364	}
3365	return ERR_PTR(-EINVAL);
3366}
3367
3368static struct md_personality raid1_personality =
3369{
3370	.name		= "raid1",
3371	.level		= 1,
3372	.owner		= THIS_MODULE,
3373	.make_request	= raid1_make_request,
3374	.run		= raid1_run,
3375	.free		= raid1_free,
3376	.status		= raid1_status,
3377	.error_handler	= raid1_error,
3378	.hot_add_disk	= raid1_add_disk,
3379	.hot_remove_disk= raid1_remove_disk,
3380	.spare_active	= raid1_spare_active,
3381	.sync_request	= raid1_sync_request,
3382	.resize		= raid1_resize,
3383	.size		= raid1_size,
3384	.check_reshape	= raid1_reshape,
3385	.quiesce	= raid1_quiesce,
3386	.takeover	= raid1_takeover,
3387	.congested	= raid1_congested,
3388};
3389
3390static int __init raid_init(void)
3391{
3392	return register_md_personality(&raid1_personality);
3393}
3394
3395static void raid_exit(void)
3396{
3397	unregister_md_personality(&raid1_personality);
3398}
3399
3400module_init(raid_init);
3401module_exit(raid_exit);
3402MODULE_LICENSE("GPL");
3403MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3404MODULE_ALIAS("md-personality-3"); /* RAID1 */
3405MODULE_ALIAS("md-raid1");
3406MODULE_ALIAS("md-level-1");
3407
3408module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
v4.17
 
   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 */
  33
  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>
  40
  41#include <trace/events/block.h>
  42
  43#include "md.h"
  44#include "raid1.h"
  45#include "md-bitmap.h"
  46
  47#define UNSUPPORTED_MDDEV_FLAGS		\
  48	((1L << MD_HAS_JOURNAL) |	\
  49	 (1L << MD_JOURNAL_CLEAN) |	\
  50	 (1L << MD_HAS_PPL) |		\
  51	 (1L << MD_HAS_MULTIPLE_PPLS))
  52
  53/*
  54 * Number of guaranteed r1bios in case of extreme VM load:
  55 */
  56#define	NR_RAID1_BIOS 256
  57
  58/* when we get a read error on a read-only array, we redirect to another
  59 * device without failing the first device, or trying to over-write to
  60 * correct the read error.  To keep track of bad blocks on a per-bio
  61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  62 */
  63#define IO_BLOCKED ((struct bio *)1)
  64/* When we successfully write to a known bad-block, we need to remove the
  65 * bad-block marking which must be done from process context.  So we record
  66 * the success by setting devs[n].bio to IO_MADE_GOOD
  67 */
  68#define IO_MADE_GOOD ((struct bio *)2)
  69
  70#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  71
  72/* When there are this many requests queue to be written by
  73 * the raid1 thread, we become 'congested' to provide back-pressure
  74 * for writeback.
  75 */
  76static int max_queued_requests = 1024;
  77
  78static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  79static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  80
  81#define raid1_log(md, fmt, args...)				\
  82	do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  83
  84#include "raid1-10.c"
  85
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  86/*
  87 * for resync bio, r1bio pointer can be retrieved from the per-bio
  88 * 'struct resync_pages'.
  89 */
  90static inline struct r1bio *get_resync_r1bio(struct bio *bio)
  91{
  92	return get_resync_pages(bio)->raid_bio;
  93}
  94
  95static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  96{
  97	struct pool_info *pi = data;
  98	int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  99
 100	/* allocate a r1bio with room for raid_disks entries in the bios array */
 101	return kzalloc(size, gfp_flags);
 102}
 103
 104static void r1bio_pool_free(void *r1_bio, void *data)
 105{
 106	kfree(r1_bio);
 107}
 108
 109#define RESYNC_DEPTH 32
 110#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 111#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 112#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 113#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 114#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 115
 116static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 117{
 118	struct pool_info *pi = data;
 119	struct r1bio *r1_bio;
 120	struct bio *bio;
 121	int need_pages;
 122	int j;
 123	struct resync_pages *rps;
 124
 125	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 126	if (!r1_bio)
 127		return NULL;
 128
 129	rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
 130		      gfp_flags);
 131	if (!rps)
 132		goto out_free_r1bio;
 133
 134	/*
 135	 * Allocate bios : 1 for reading, n-1 for writing
 136	 */
 137	for (j = pi->raid_disks ; j-- ; ) {
 138		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 139		if (!bio)
 140			goto out_free_bio;
 141		r1_bio->bios[j] = bio;
 142	}
 143	/*
 144	 * Allocate RESYNC_PAGES data pages and attach them to
 145	 * the first bio.
 146	 * If this is a user-requested check/repair, allocate
 147	 * RESYNC_PAGES for each bio.
 148	 */
 149	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 150		need_pages = pi->raid_disks;
 151	else
 152		need_pages = 1;
 153	for (j = 0; j < pi->raid_disks; j++) {
 154		struct resync_pages *rp = &rps[j];
 155
 156		bio = r1_bio->bios[j];
 157
 158		if (j < need_pages) {
 159			if (resync_alloc_pages(rp, gfp_flags))
 160				goto out_free_pages;
 161		} else {
 162			memcpy(rp, &rps[0], sizeof(*rp));
 163			resync_get_all_pages(rp);
 164		}
 165
 166		rp->raid_bio = r1_bio;
 167		bio->bi_private = rp;
 168	}
 169
 170	r1_bio->master_bio = NULL;
 171
 172	return r1_bio;
 173
 174out_free_pages:
 175	while (--j >= 0)
 176		resync_free_pages(&rps[j]);
 177
 178out_free_bio:
 179	while (++j < pi->raid_disks)
 180		bio_put(r1_bio->bios[j]);
 181	kfree(rps);
 182
 183out_free_r1bio:
 184	r1bio_pool_free(r1_bio, data);
 185	return NULL;
 186}
 187
 188static void r1buf_pool_free(void *__r1_bio, void *data)
 189{
 190	struct pool_info *pi = data;
 191	int i;
 192	struct r1bio *r1bio = __r1_bio;
 193	struct resync_pages *rp = NULL;
 194
 195	for (i = pi->raid_disks; i--; ) {
 196		rp = get_resync_pages(r1bio->bios[i]);
 197		resync_free_pages(rp);
 198		bio_put(r1bio->bios[i]);
 199	}
 200
 201	/* resync pages array stored in the 1st bio's .bi_private */
 202	kfree(rp);
 203
 204	r1bio_pool_free(r1bio, data);
 205}
 206
 207static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 208{
 209	int i;
 210
 211	for (i = 0; i < conf->raid_disks * 2; i++) {
 212		struct bio **bio = r1_bio->bios + i;
 213		if (!BIO_SPECIAL(*bio))
 214			bio_put(*bio);
 215		*bio = NULL;
 216	}
 217}
 218
 219static void free_r1bio(struct r1bio *r1_bio)
 220{
 221	struct r1conf *conf = r1_bio->mddev->private;
 222
 223	put_all_bios(conf, r1_bio);
 224	mempool_free(r1_bio, conf->r1bio_pool);
 225}
 226
 227static void put_buf(struct r1bio *r1_bio)
 228{
 229	struct r1conf *conf = r1_bio->mddev->private;
 230	sector_t sect = r1_bio->sector;
 231	int i;
 232
 233	for (i = 0; i < conf->raid_disks * 2; i++) {
 234		struct bio *bio = r1_bio->bios[i];
 235		if (bio->bi_end_io)
 236			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 237	}
 238
 239	mempool_free(r1_bio, conf->r1buf_pool);
 240
 241	lower_barrier(conf, sect);
 242}
 243
 244static void reschedule_retry(struct r1bio *r1_bio)
 245{
 246	unsigned long flags;
 247	struct mddev *mddev = r1_bio->mddev;
 248	struct r1conf *conf = mddev->private;
 249	int idx;
 250
 251	idx = sector_to_idx(r1_bio->sector);
 252	spin_lock_irqsave(&conf->device_lock, flags);
 253	list_add(&r1_bio->retry_list, &conf->retry_list);
 254	atomic_inc(&conf->nr_queued[idx]);
 255	spin_unlock_irqrestore(&conf->device_lock, flags);
 256
 257	wake_up(&conf->wait_barrier);
 258	md_wakeup_thread(mddev->thread);
 259}
 260
 261/*
 262 * raid_end_bio_io() is called when we have finished servicing a mirrored
 263 * operation and are ready to return a success/failure code to the buffer
 264 * cache layer.
 265 */
 266static void call_bio_endio(struct r1bio *r1_bio)
 267{
 268	struct bio *bio = r1_bio->master_bio;
 269	struct r1conf *conf = r1_bio->mddev->private;
 270
 271	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 272		bio->bi_status = BLK_STS_IOERR;
 273
 274	bio_endio(bio);
 275	/*
 276	 * Wake up any possible resync thread that waits for the device
 277	 * to go idle.
 278	 */
 279	allow_barrier(conf, r1_bio->sector);
 280}
 281
 282static void raid_end_bio_io(struct r1bio *r1_bio)
 283{
 284	struct bio *bio = r1_bio->master_bio;
 285
 286	/* if nobody has done the final endio yet, do it now */
 287	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 288		pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 289			 (bio_data_dir(bio) == WRITE) ? "write" : "read",
 290			 (unsigned long long) bio->bi_iter.bi_sector,
 291			 (unsigned long long) bio_end_sector(bio) - 1);
 292
 293		call_bio_endio(r1_bio);
 294	}
 295	free_r1bio(r1_bio);
 296}
 297
 298/*
 299 * Update disk head position estimator based on IRQ completion info.
 300 */
 301static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 302{
 303	struct r1conf *conf = r1_bio->mddev->private;
 304
 305	conf->mirrors[disk].head_position =
 306		r1_bio->sector + (r1_bio->sectors);
 307}
 308
 309/*
 310 * Find the disk number which triggered given bio
 311 */
 312static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 313{
 314	int mirror;
 315	struct r1conf *conf = r1_bio->mddev->private;
 316	int raid_disks = conf->raid_disks;
 317
 318	for (mirror = 0; mirror < raid_disks * 2; mirror++)
 319		if (r1_bio->bios[mirror] == bio)
 320			break;
 321
 322	BUG_ON(mirror == raid_disks * 2);
 323	update_head_pos(mirror, r1_bio);
 324
 325	return mirror;
 326}
 327
 328static void raid1_end_read_request(struct bio *bio)
 329{
 330	int uptodate = !bio->bi_status;
 331	struct r1bio *r1_bio = bio->bi_private;
 332	struct r1conf *conf = r1_bio->mddev->private;
 333	struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 334
 335	/*
 336	 * this branch is our 'one mirror IO has finished' event handler:
 337	 */
 338	update_head_pos(r1_bio->read_disk, r1_bio);
 339
 340	if (uptodate)
 341		set_bit(R1BIO_Uptodate, &r1_bio->state);
 342	else if (test_bit(FailFast, &rdev->flags) &&
 343		 test_bit(R1BIO_FailFast, &r1_bio->state))
 344		/* This was a fail-fast read so we definitely
 345		 * want to retry */
 346		;
 347	else {
 348		/* If all other devices have failed, we want to return
 349		 * the error upwards rather than fail the last device.
 350		 * Here we redefine "uptodate" to mean "Don't want to retry"
 351		 */
 352		unsigned long flags;
 353		spin_lock_irqsave(&conf->device_lock, flags);
 354		if (r1_bio->mddev->degraded == conf->raid_disks ||
 355		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 356		     test_bit(In_sync, &rdev->flags)))
 357			uptodate = 1;
 358		spin_unlock_irqrestore(&conf->device_lock, flags);
 359	}
 360
 361	if (uptodate) {
 362		raid_end_bio_io(r1_bio);
 363		rdev_dec_pending(rdev, conf->mddev);
 364	} else {
 365		/*
 366		 * oops, read error:
 367		 */
 368		char b[BDEVNAME_SIZE];
 369		pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 370				   mdname(conf->mddev),
 371				   bdevname(rdev->bdev, b),
 372				   (unsigned long long)r1_bio->sector);
 373		set_bit(R1BIO_ReadError, &r1_bio->state);
 374		reschedule_retry(r1_bio);
 375		/* don't drop the reference on read_disk yet */
 376	}
 377}
 378
 379static void close_write(struct r1bio *r1_bio)
 380{
 381	/* it really is the end of this request */
 382	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 383		bio_free_pages(r1_bio->behind_master_bio);
 384		bio_put(r1_bio->behind_master_bio);
 385		r1_bio->behind_master_bio = NULL;
 386	}
 387	/* clear the bitmap if all writes complete successfully */
 388	bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 389			r1_bio->sectors,
 390			!test_bit(R1BIO_Degraded, &r1_bio->state),
 391			test_bit(R1BIO_BehindIO, &r1_bio->state));
 392	md_write_end(r1_bio->mddev);
 393}
 394
 395static void r1_bio_write_done(struct r1bio *r1_bio)
 396{
 397	if (!atomic_dec_and_test(&r1_bio->remaining))
 398		return;
 399
 400	if (test_bit(R1BIO_WriteError, &r1_bio->state))
 401		reschedule_retry(r1_bio);
 402	else {
 403		close_write(r1_bio);
 404		if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 405			reschedule_retry(r1_bio);
 406		else
 407			raid_end_bio_io(r1_bio);
 408	}
 409}
 410
 411static void raid1_end_write_request(struct bio *bio)
 412{
 413	struct r1bio *r1_bio = bio->bi_private;
 414	int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 415	struct r1conf *conf = r1_bio->mddev->private;
 416	struct bio *to_put = NULL;
 417	int mirror = find_bio_disk(r1_bio, bio);
 418	struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 419	bool discard_error;
 420
 421	discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 422
 423	/*
 424	 * 'one mirror IO has finished' event handler:
 425	 */
 426	if (bio->bi_status && !discard_error) {
 427		set_bit(WriteErrorSeen,	&rdev->flags);
 428		if (!test_and_set_bit(WantReplacement, &rdev->flags))
 429			set_bit(MD_RECOVERY_NEEDED, &
 430				conf->mddev->recovery);
 431
 432		if (test_bit(FailFast, &rdev->flags) &&
 433		    (bio->bi_opf & MD_FAILFAST) &&
 434		    /* We never try FailFast to WriteMostly devices */
 435		    !test_bit(WriteMostly, &rdev->flags)) {
 436			md_error(r1_bio->mddev, rdev);
 437			if (!test_bit(Faulty, &rdev->flags))
 438				/* This is the only remaining device,
 439				 * We need to retry the write without
 440				 * FailFast
 441				 */
 442				set_bit(R1BIO_WriteError, &r1_bio->state);
 443			else {
 444				/* Finished with this branch */
 445				r1_bio->bios[mirror] = NULL;
 446				to_put = bio;
 447			}
 448		} else
 449			set_bit(R1BIO_WriteError, &r1_bio->state);
 
 
 
 
 
 450	} else {
 451		/*
 452		 * Set R1BIO_Uptodate in our master bio, so that we
 453		 * will return a good error code for to the higher
 454		 * levels even if IO on some other mirrored buffer
 455		 * fails.
 456		 *
 457		 * The 'master' represents the composite IO operation
 458		 * to user-side. So if something waits for IO, then it
 459		 * will wait for the 'master' bio.
 460		 */
 461		sector_t first_bad;
 462		int bad_sectors;
 463
 464		r1_bio->bios[mirror] = NULL;
 465		to_put = bio;
 466		/*
 467		 * Do not set R1BIO_Uptodate if the current device is
 468		 * rebuilding or Faulty. This is because we cannot use
 469		 * such device for properly reading the data back (we could
 470		 * potentially use it, if the current write would have felt
 471		 * before rdev->recovery_offset, but for simplicity we don't
 472		 * check this here.
 473		 */
 474		if (test_bit(In_sync, &rdev->flags) &&
 475		    !test_bit(Faulty, &rdev->flags))
 476			set_bit(R1BIO_Uptodate, &r1_bio->state);
 477
 478		/* Maybe we can clear some bad blocks. */
 479		if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 480				&first_bad, &bad_sectors) && !discard_error) {
 481			r1_bio->bios[mirror] = IO_MADE_GOOD;
 482			set_bit(R1BIO_MadeGood, &r1_bio->state);
 483		}
 484	}
 485
 486	if (behind) {
 
 
 
 
 
 
 487		if (test_bit(WriteMostly, &rdev->flags))
 488			atomic_dec(&r1_bio->behind_remaining);
 489
 490		/*
 491		 * In behind mode, we ACK the master bio once the I/O
 492		 * has safely reached all non-writemostly
 493		 * disks. Setting the Returned bit ensures that this
 494		 * gets done only once -- we don't ever want to return
 495		 * -EIO here, instead we'll wait
 496		 */
 497		if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 498		    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 499			/* Maybe we can return now */
 500			if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 501				struct bio *mbio = r1_bio->master_bio;
 502				pr_debug("raid1: behind end write sectors"
 503					 " %llu-%llu\n",
 504					 (unsigned long long) mbio->bi_iter.bi_sector,
 505					 (unsigned long long) bio_end_sector(mbio) - 1);
 506				call_bio_endio(r1_bio);
 507			}
 508		}
 509	}
 510	if (r1_bio->bios[mirror] == NULL)
 511		rdev_dec_pending(rdev, conf->mddev);
 512
 513	/*
 514	 * Let's see if all mirrored write operations have finished
 515	 * already.
 516	 */
 517	r1_bio_write_done(r1_bio);
 518
 519	if (to_put)
 520		bio_put(to_put);
 521}
 522
 523static sector_t align_to_barrier_unit_end(sector_t start_sector,
 524					  sector_t sectors)
 525{
 526	sector_t len;
 527
 528	WARN_ON(sectors == 0);
 529	/*
 530	 * len is the number of sectors from start_sector to end of the
 531	 * barrier unit which start_sector belongs to.
 532	 */
 533	len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 534	      start_sector;
 535
 536	if (len > sectors)
 537		len = sectors;
 538
 539	return len;
 540}
 541
 542/*
 543 * This routine returns the disk from which the requested read should
 544 * be done. There is a per-array 'next expected sequential IO' sector
 545 * number - if this matches on the next IO then we use the last disk.
 546 * There is also a per-disk 'last know head position' sector that is
 547 * maintained from IRQ contexts, both the normal and the resync IO
 548 * completion handlers update this position correctly. If there is no
 549 * perfect sequential match then we pick the disk whose head is closest.
 550 *
 551 * If there are 2 mirrors in the same 2 devices, performance degrades
 552 * because position is mirror, not device based.
 553 *
 554 * The rdev for the device selected will have nr_pending incremented.
 555 */
 556static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 557{
 558	const sector_t this_sector = r1_bio->sector;
 559	int sectors;
 560	int best_good_sectors;
 561	int best_disk, best_dist_disk, best_pending_disk;
 562	int has_nonrot_disk;
 563	int disk;
 564	sector_t best_dist;
 565	unsigned int min_pending;
 566	struct md_rdev *rdev;
 567	int choose_first;
 568	int choose_next_idle;
 569
 570	rcu_read_lock();
 571	/*
 572	 * Check if we can balance. We can balance on the whole
 573	 * device if no resync is going on, or below the resync window.
 574	 * We take the first readable disk when above the resync window.
 575	 */
 576 retry:
 577	sectors = r1_bio->sectors;
 578	best_disk = -1;
 579	best_dist_disk = -1;
 580	best_dist = MaxSector;
 581	best_pending_disk = -1;
 582	min_pending = UINT_MAX;
 583	best_good_sectors = 0;
 584	has_nonrot_disk = 0;
 585	choose_next_idle = 0;
 586	clear_bit(R1BIO_FailFast, &r1_bio->state);
 587
 588	if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 589	    (mddev_is_clustered(conf->mddev) &&
 590	    md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 591		    this_sector + sectors)))
 592		choose_first = 1;
 593	else
 594		choose_first = 0;
 595
 596	for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 597		sector_t dist;
 598		sector_t first_bad;
 599		int bad_sectors;
 600		unsigned int pending;
 601		bool nonrot;
 602
 603		rdev = rcu_dereference(conf->mirrors[disk].rdev);
 604		if (r1_bio->bios[disk] == IO_BLOCKED
 605		    || rdev == NULL
 606		    || test_bit(Faulty, &rdev->flags))
 607			continue;
 608		if (!test_bit(In_sync, &rdev->flags) &&
 609		    rdev->recovery_offset < this_sector + sectors)
 610			continue;
 611		if (test_bit(WriteMostly, &rdev->flags)) {
 612			/* Don't balance among write-mostly, just
 613			 * use the first as a last resort */
 614			if (best_dist_disk < 0) {
 615				if (is_badblock(rdev, this_sector, sectors,
 616						&first_bad, &bad_sectors)) {
 617					if (first_bad <= this_sector)
 618						/* Cannot use this */
 619						continue;
 620					best_good_sectors = first_bad - this_sector;
 621				} else
 622					best_good_sectors = sectors;
 623				best_dist_disk = disk;
 624				best_pending_disk = disk;
 625			}
 626			continue;
 627		}
 628		/* This is a reasonable device to use.  It might
 629		 * even be best.
 630		 */
 631		if (is_badblock(rdev, this_sector, sectors,
 632				&first_bad, &bad_sectors)) {
 633			if (best_dist < MaxSector)
 634				/* already have a better device */
 635				continue;
 636			if (first_bad <= this_sector) {
 637				/* cannot read here. If this is the 'primary'
 638				 * device, then we must not read beyond
 639				 * bad_sectors from another device..
 640				 */
 641				bad_sectors -= (this_sector - first_bad);
 642				if (choose_first && sectors > bad_sectors)
 643					sectors = bad_sectors;
 644				if (best_good_sectors > sectors)
 645					best_good_sectors = sectors;
 646
 647			} else {
 648				sector_t good_sectors = first_bad - this_sector;
 649				if (good_sectors > best_good_sectors) {
 650					best_good_sectors = good_sectors;
 651					best_disk = disk;
 652				}
 653				if (choose_first)
 654					break;
 655			}
 656			continue;
 657		} else {
 658			if ((sectors > best_good_sectors) && (best_disk >= 0))
 659				best_disk = -1;
 660			best_good_sectors = sectors;
 661		}
 662
 663		if (best_disk >= 0)
 664			/* At least two disks to choose from so failfast is OK */
 665			set_bit(R1BIO_FailFast, &r1_bio->state);
 666
 667		nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 668		has_nonrot_disk |= nonrot;
 669		pending = atomic_read(&rdev->nr_pending);
 670		dist = abs(this_sector - conf->mirrors[disk].head_position);
 671		if (choose_first) {
 672			best_disk = disk;
 673			break;
 674		}
 675		/* Don't change to another disk for sequential reads */
 676		if (conf->mirrors[disk].next_seq_sect == this_sector
 677		    || dist == 0) {
 678			int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 679			struct raid1_info *mirror = &conf->mirrors[disk];
 680
 681			best_disk = disk;
 682			/*
 683			 * If buffered sequential IO size exceeds optimal
 684			 * iosize, check if there is idle disk. If yes, choose
 685			 * the idle disk. read_balance could already choose an
 686			 * idle disk before noticing it's a sequential IO in
 687			 * this disk. This doesn't matter because this disk
 688			 * will idle, next time it will be utilized after the
 689			 * first disk has IO size exceeds optimal iosize. In
 690			 * this way, iosize of the first disk will be optimal
 691			 * iosize at least. iosize of the second disk might be
 692			 * small, but not a big deal since when the second disk
 693			 * starts IO, the first disk is likely still busy.
 694			 */
 695			if (nonrot && opt_iosize > 0 &&
 696			    mirror->seq_start != MaxSector &&
 697			    mirror->next_seq_sect > opt_iosize &&
 698			    mirror->next_seq_sect - opt_iosize >=
 699			    mirror->seq_start) {
 700				choose_next_idle = 1;
 701				continue;
 702			}
 703			break;
 704		}
 705
 706		if (choose_next_idle)
 707			continue;
 708
 709		if (min_pending > pending) {
 710			min_pending = pending;
 711			best_pending_disk = disk;
 712		}
 713
 714		if (dist < best_dist) {
 715			best_dist = dist;
 716			best_dist_disk = disk;
 717		}
 718	}
 719
 720	/*
 721	 * If all disks are rotational, choose the closest disk. If any disk is
 722	 * non-rotational, choose the disk with less pending request even the
 723	 * disk is rotational, which might/might not be optimal for raids with
 724	 * mixed ratation/non-rotational disks depending on workload.
 725	 */
 726	if (best_disk == -1) {
 727		if (has_nonrot_disk || min_pending == 0)
 728			best_disk = best_pending_disk;
 729		else
 730			best_disk = best_dist_disk;
 731	}
 732
 733	if (best_disk >= 0) {
 734		rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 735		if (!rdev)
 736			goto retry;
 737		atomic_inc(&rdev->nr_pending);
 738		sectors = best_good_sectors;
 739
 740		if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 741			conf->mirrors[best_disk].seq_start = this_sector;
 742
 743		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 744	}
 745	rcu_read_unlock();
 746	*max_sectors = sectors;
 747
 748	return best_disk;
 749}
 750
 751static int raid1_congested(struct mddev *mddev, int bits)
 752{
 753	struct r1conf *conf = mddev->private;
 754	int i, ret = 0;
 755
 756	if ((bits & (1 << WB_async_congested)) &&
 757	    conf->pending_count >= max_queued_requests)
 758		return 1;
 759
 760	rcu_read_lock();
 761	for (i = 0; i < conf->raid_disks * 2; i++) {
 762		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 763		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 764			struct request_queue *q = bdev_get_queue(rdev->bdev);
 765
 766			BUG_ON(!q);
 767
 768			/* Note the '|| 1' - when read_balance prefers
 769			 * non-congested targets, it can be removed
 770			 */
 771			if ((bits & (1 << WB_async_congested)) || 1)
 772				ret |= bdi_congested(q->backing_dev_info, bits);
 773			else
 774				ret &= bdi_congested(q->backing_dev_info, bits);
 775		}
 776	}
 777	rcu_read_unlock();
 778	return ret;
 779}
 780
 781static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 782{
 783	/* flush any pending bitmap writes to disk before proceeding w/ I/O */
 784	bitmap_unplug(conf->mddev->bitmap);
 785	wake_up(&conf->wait_barrier);
 786
 787	while (bio) { /* submit pending writes */
 788		struct bio *next = bio->bi_next;
 789		struct md_rdev *rdev = (void *)bio->bi_disk;
 790		bio->bi_next = NULL;
 791		bio_set_dev(bio, rdev->bdev);
 792		if (test_bit(Faulty, &rdev->flags)) {
 793			bio_io_error(bio);
 794		} else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 795				    !blk_queue_discard(bio->bi_disk->queue)))
 796			/* Just ignore it */
 797			bio_endio(bio);
 798		else
 799			generic_make_request(bio);
 800		bio = next;
 801	}
 802}
 803
 804static void flush_pending_writes(struct r1conf *conf)
 805{
 806	/* Any writes that have been queued but are awaiting
 807	 * bitmap updates get flushed here.
 808	 */
 809	spin_lock_irq(&conf->device_lock);
 810
 811	if (conf->pending_bio_list.head) {
 812		struct blk_plug plug;
 813		struct bio *bio;
 814
 815		bio = bio_list_get(&conf->pending_bio_list);
 816		conf->pending_count = 0;
 817		spin_unlock_irq(&conf->device_lock);
 818
 819		/*
 820		 * As this is called in a wait_event() loop (see freeze_array),
 821		 * current->state might be TASK_UNINTERRUPTIBLE which will
 822		 * cause a warning when we prepare to wait again.  As it is
 823		 * rare that this path is taken, it is perfectly safe to force
 824		 * us to go around the wait_event() loop again, so the warning
 825		 * is a false-positive.  Silence the warning by resetting
 826		 * thread state
 827		 */
 828		__set_current_state(TASK_RUNNING);
 829		blk_start_plug(&plug);
 830		flush_bio_list(conf, bio);
 831		blk_finish_plug(&plug);
 832	} else
 833		spin_unlock_irq(&conf->device_lock);
 834}
 835
 836/* Barriers....
 837 * Sometimes we need to suspend IO while we do something else,
 838 * either some resync/recovery, or reconfigure the array.
 839 * To do this we raise a 'barrier'.
 840 * The 'barrier' is a counter that can be raised multiple times
 841 * to count how many activities are happening which preclude
 842 * normal IO.
 843 * We can only raise the barrier if there is no pending IO.
 844 * i.e. if nr_pending == 0.
 845 * We choose only to raise the barrier if no-one is waiting for the
 846 * barrier to go down.  This means that as soon as an IO request
 847 * is ready, no other operations which require a barrier will start
 848 * until the IO request has had a chance.
 849 *
 850 * So: regular IO calls 'wait_barrier'.  When that returns there
 851 *    is no backgroup IO happening,  It must arrange to call
 852 *    allow_barrier when it has finished its IO.
 853 * backgroup IO calls must call raise_barrier.  Once that returns
 854 *    there is no normal IO happeing.  It must arrange to call
 855 *    lower_barrier when the particular background IO completes.
 
 
 
 856 */
 857static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
 858{
 859	int idx = sector_to_idx(sector_nr);
 860
 861	spin_lock_irq(&conf->resync_lock);
 862
 863	/* Wait until no block IO is waiting */
 864	wait_event_lock_irq(conf->wait_barrier,
 865			    !atomic_read(&conf->nr_waiting[idx]),
 866			    conf->resync_lock);
 867
 868	/* block any new IO from starting */
 869	atomic_inc(&conf->barrier[idx]);
 870	/*
 871	 * In raise_barrier() we firstly increase conf->barrier[idx] then
 872	 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 873	 * increase conf->nr_pending[idx] then check conf->barrier[idx].
 874	 * A memory barrier here to make sure conf->nr_pending[idx] won't
 875	 * be fetched before conf->barrier[idx] is increased. Otherwise
 876	 * there will be a race between raise_barrier() and _wait_barrier().
 877	 */
 878	smp_mb__after_atomic();
 879
 880	/* For these conditions we must wait:
 881	 * A: while the array is in frozen state
 882	 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 883	 *    existing in corresponding I/O barrier bucket.
 884	 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 885	 *    max resync count which allowed on current I/O barrier bucket.
 886	 */
 887	wait_event_lock_irq(conf->wait_barrier,
 888			    (!conf->array_frozen &&
 889			     !atomic_read(&conf->nr_pending[idx]) &&
 890			     atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 891				test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 892			    conf->resync_lock);
 893
 894	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 895		atomic_dec(&conf->barrier[idx]);
 896		spin_unlock_irq(&conf->resync_lock);
 897		wake_up(&conf->wait_barrier);
 898		return -EINTR;
 899	}
 900
 901	atomic_inc(&conf->nr_sync_pending);
 902	spin_unlock_irq(&conf->resync_lock);
 903
 904	return 0;
 905}
 906
 907static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 908{
 909	int idx = sector_to_idx(sector_nr);
 910
 911	BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 912
 913	atomic_dec(&conf->barrier[idx]);
 914	atomic_dec(&conf->nr_sync_pending);
 915	wake_up(&conf->wait_barrier);
 916}
 917
 918static void _wait_barrier(struct r1conf *conf, int idx)
 919{
 920	/*
 921	 * We need to increase conf->nr_pending[idx] very early here,
 922	 * then raise_barrier() can be blocked when it waits for
 923	 * conf->nr_pending[idx] to be 0. Then we can avoid holding
 924	 * conf->resync_lock when there is no barrier raised in same
 925	 * barrier unit bucket. Also if the array is frozen, I/O
 926	 * should be blocked until array is unfrozen.
 927	 */
 928	atomic_inc(&conf->nr_pending[idx]);
 929	/*
 930	 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 931	 * check conf->barrier[idx]. In raise_barrier() we firstly increase
 932	 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 933	 * barrier is necessary here to make sure conf->barrier[idx] won't be
 934	 * fetched before conf->nr_pending[idx] is increased. Otherwise there
 935	 * will be a race between _wait_barrier() and raise_barrier().
 936	 */
 937	smp_mb__after_atomic();
 938
 939	/*
 940	 * Don't worry about checking two atomic_t variables at same time
 941	 * here. If during we check conf->barrier[idx], the array is
 942	 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 943	 * 0, it is safe to return and make the I/O continue. Because the
 944	 * array is frozen, all I/O returned here will eventually complete
 945	 * or be queued, no race will happen. See code comment in
 946	 * frozen_array().
 947	 */
 948	if (!READ_ONCE(conf->array_frozen) &&
 949	    !atomic_read(&conf->barrier[idx]))
 950		return;
 951
 952	/*
 953	 * After holding conf->resync_lock, conf->nr_pending[idx]
 954	 * should be decreased before waiting for barrier to drop.
 955	 * Otherwise, we may encounter a race condition because
 956	 * raise_barrer() might be waiting for conf->nr_pending[idx]
 957	 * to be 0 at same time.
 958	 */
 959	spin_lock_irq(&conf->resync_lock);
 960	atomic_inc(&conf->nr_waiting[idx]);
 961	atomic_dec(&conf->nr_pending[idx]);
 962	/*
 963	 * In case freeze_array() is waiting for
 964	 * get_unqueued_pending() == extra
 965	 */
 966	wake_up(&conf->wait_barrier);
 967	/* Wait for the barrier in same barrier unit bucket to drop. */
 968	wait_event_lock_irq(conf->wait_barrier,
 969			    !conf->array_frozen &&
 970			     !atomic_read(&conf->barrier[idx]),
 971			    conf->resync_lock);
 972	atomic_inc(&conf->nr_pending[idx]);
 973	atomic_dec(&conf->nr_waiting[idx]);
 974	spin_unlock_irq(&conf->resync_lock);
 975}
 976
 977static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 978{
 979	int idx = sector_to_idx(sector_nr);
 980
 981	/*
 982	 * Very similar to _wait_barrier(). The difference is, for read
 983	 * I/O we don't need wait for sync I/O, but if the whole array
 984	 * is frozen, the read I/O still has to wait until the array is
 985	 * unfrozen. Since there is no ordering requirement with
 986	 * conf->barrier[idx] here, memory barrier is unnecessary as well.
 987	 */
 988	atomic_inc(&conf->nr_pending[idx]);
 989
 990	if (!READ_ONCE(conf->array_frozen))
 991		return;
 992
 993	spin_lock_irq(&conf->resync_lock);
 994	atomic_inc(&conf->nr_waiting[idx]);
 995	atomic_dec(&conf->nr_pending[idx]);
 996	/*
 997	 * In case freeze_array() is waiting for
 998	 * get_unqueued_pending() == extra
 999	 */
1000	wake_up(&conf->wait_barrier);
1001	/* Wait for array to be unfrozen */
1002	wait_event_lock_irq(conf->wait_barrier,
1003			    !conf->array_frozen,
1004			    conf->resync_lock);
1005	atomic_inc(&conf->nr_pending[idx]);
1006	atomic_dec(&conf->nr_waiting[idx]);
1007	spin_unlock_irq(&conf->resync_lock);
1008}
1009
1010static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1011{
1012	int idx = sector_to_idx(sector_nr);
1013
1014	_wait_barrier(conf, idx);
1015}
1016
1017static void _allow_barrier(struct r1conf *conf, int idx)
1018{
1019	atomic_dec(&conf->nr_pending[idx]);
1020	wake_up(&conf->wait_barrier);
1021}
1022
1023static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1024{
1025	int idx = sector_to_idx(sector_nr);
1026
1027	_allow_barrier(conf, idx);
1028}
1029
1030/* conf->resync_lock should be held */
1031static int get_unqueued_pending(struct r1conf *conf)
1032{
1033	int idx, ret;
1034
1035	ret = atomic_read(&conf->nr_sync_pending);
1036	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1037		ret += atomic_read(&conf->nr_pending[idx]) -
1038			atomic_read(&conf->nr_queued[idx]);
1039
1040	return ret;
1041}
1042
1043static void freeze_array(struct r1conf *conf, int extra)
1044{
1045	/* Stop sync I/O and normal I/O and wait for everything to
1046	 * go quiet.
1047	 * This is called in two situations:
1048	 * 1) management command handlers (reshape, remove disk, quiesce).
1049	 * 2) one normal I/O request failed.
1050
1051	 * After array_frozen is set to 1, new sync IO will be blocked at
1052	 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1053	 * or wait_read_barrier(). The flying I/Os will either complete or be
1054	 * queued. When everything goes quite, there are only queued I/Os left.
1055
1056	 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1057	 * barrier bucket index which this I/O request hits. When all sync and
1058	 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1059	 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1060	 * in handle_read_error(), we may call freeze_array() before trying to
1061	 * fix the read error. In this case, the error read I/O is not queued,
1062	 * so get_unqueued_pending() == 1.
1063	 *
1064	 * Therefore before this function returns, we need to wait until
1065	 * get_unqueued_pendings(conf) gets equal to extra. For
1066	 * normal I/O context, extra is 1, in rested situations extra is 0.
1067	 */
1068	spin_lock_irq(&conf->resync_lock);
1069	conf->array_frozen = 1;
1070	raid1_log(conf->mddev, "wait freeze");
1071	wait_event_lock_irq_cmd(
1072		conf->wait_barrier,
1073		get_unqueued_pending(conf) == extra,
1074		conf->resync_lock,
1075		flush_pending_writes(conf));
1076	spin_unlock_irq(&conf->resync_lock);
1077}
1078static void unfreeze_array(struct r1conf *conf)
1079{
1080	/* reverse the effect of the freeze */
1081	spin_lock_irq(&conf->resync_lock);
1082	conf->array_frozen = 0;
1083	spin_unlock_irq(&conf->resync_lock);
1084	wake_up(&conf->wait_barrier);
1085}
1086
1087static void alloc_behind_master_bio(struct r1bio *r1_bio,
1088					   struct bio *bio)
1089{
1090	int size = bio->bi_iter.bi_size;
1091	unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1092	int i = 0;
1093	struct bio *behind_bio = NULL;
1094
1095	behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1096	if (!behind_bio)
1097		return;
1098
1099	/* discard op, we don't support writezero/writesame yet */
1100	if (!bio_has_data(bio)) {
1101		behind_bio->bi_iter.bi_size = size;
1102		goto skip_copy;
1103	}
1104
1105	behind_bio->bi_write_hint = bio->bi_write_hint;
1106
1107	while (i < vcnt && size) {
1108		struct page *page;
1109		int len = min_t(int, PAGE_SIZE, size);
1110
1111		page = alloc_page(GFP_NOIO);
1112		if (unlikely(!page))
1113			goto free_pages;
1114
1115		bio_add_page(behind_bio, page, len, 0);
1116
1117		size -= len;
1118		i++;
1119	}
1120
1121	bio_copy_data(behind_bio, bio);
1122skip_copy:
1123	r1_bio->behind_master_bio = behind_bio;
1124	set_bit(R1BIO_BehindIO, &r1_bio->state);
1125
1126	return;
1127
1128free_pages:
1129	pr_debug("%dB behind alloc failed, doing sync I/O\n",
1130		 bio->bi_iter.bi_size);
1131	bio_free_pages(behind_bio);
1132	bio_put(behind_bio);
1133}
1134
1135struct raid1_plug_cb {
1136	struct blk_plug_cb	cb;
1137	struct bio_list		pending;
1138	int			pending_cnt;
1139};
1140
1141static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1142{
1143	struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1144						  cb);
1145	struct mddev *mddev = plug->cb.data;
1146	struct r1conf *conf = mddev->private;
1147	struct bio *bio;
1148
1149	if (from_schedule || current->bio_list) {
1150		spin_lock_irq(&conf->device_lock);
1151		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1152		conf->pending_count += plug->pending_cnt;
1153		spin_unlock_irq(&conf->device_lock);
1154		wake_up(&conf->wait_barrier);
1155		md_wakeup_thread(mddev->thread);
1156		kfree(plug);
1157		return;
1158	}
1159
1160	/* we aren't scheduling, so we can do the write-out directly. */
1161	bio = bio_list_get(&plug->pending);
1162	flush_bio_list(conf, bio);
1163	kfree(plug);
1164}
1165
1166static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1167{
1168	r1_bio->master_bio = bio;
1169	r1_bio->sectors = bio_sectors(bio);
1170	r1_bio->state = 0;
1171	r1_bio->mddev = mddev;
1172	r1_bio->sector = bio->bi_iter.bi_sector;
1173}
1174
1175static inline struct r1bio *
1176alloc_r1bio(struct mddev *mddev, struct bio *bio)
1177{
1178	struct r1conf *conf = mddev->private;
1179	struct r1bio *r1_bio;
1180
1181	r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1182	/* Ensure no bio records IO_BLOCKED */
1183	memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1184	init_r1bio(r1_bio, mddev, bio);
1185	return r1_bio;
1186}
1187
1188static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1189			       int max_read_sectors, struct r1bio *r1_bio)
1190{
1191	struct r1conf *conf = mddev->private;
1192	struct raid1_info *mirror;
1193	struct bio *read_bio;
1194	struct bitmap *bitmap = mddev->bitmap;
1195	const int op = bio_op(bio);
1196	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1197	int max_sectors;
1198	int rdisk;
1199	bool print_msg = !!r1_bio;
1200	char b[BDEVNAME_SIZE];
1201
1202	/*
1203	 * If r1_bio is set, we are blocking the raid1d thread
1204	 * so there is a tiny risk of deadlock.  So ask for
1205	 * emergency memory if needed.
1206	 */
1207	gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1208
1209	if (print_msg) {
1210		/* Need to get the block device name carefully */
1211		struct md_rdev *rdev;
1212		rcu_read_lock();
1213		rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1214		if (rdev)
1215			bdevname(rdev->bdev, b);
1216		else
1217			strcpy(b, "???");
1218		rcu_read_unlock();
1219	}
1220
1221	/*
1222	 * Still need barrier for READ in case that whole
1223	 * array is frozen.
1224	 */
1225	wait_read_barrier(conf, bio->bi_iter.bi_sector);
1226
1227	if (!r1_bio)
1228		r1_bio = alloc_r1bio(mddev, bio);
1229	else
1230		init_r1bio(r1_bio, mddev, bio);
1231	r1_bio->sectors = max_read_sectors;
1232
1233	/*
1234	 * make_request() can abort the operation when read-ahead is being
1235	 * used and no empty request is available.
1236	 */
1237	rdisk = read_balance(conf, r1_bio, &max_sectors);
1238
1239	if (rdisk < 0) {
1240		/* couldn't find anywhere to read from */
1241		if (print_msg) {
1242			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1243					    mdname(mddev),
1244					    b,
1245					    (unsigned long long)r1_bio->sector);
1246		}
1247		raid_end_bio_io(r1_bio);
1248		return;
1249	}
1250	mirror = conf->mirrors + rdisk;
1251
1252	if (print_msg)
1253		pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1254				    mdname(mddev),
1255				    (unsigned long long)r1_bio->sector,
1256				    bdevname(mirror->rdev->bdev, b));
1257
1258	if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1259	    bitmap) {
1260		/*
1261		 * Reading from a write-mostly device must take care not to
1262		 * over-take any writes that are 'behind'
1263		 */
1264		raid1_log(mddev, "wait behind writes");
1265		wait_event(bitmap->behind_wait,
1266			   atomic_read(&bitmap->behind_writes) == 0);
1267	}
1268
1269	if (max_sectors < bio_sectors(bio)) {
1270		struct bio *split = bio_split(bio, max_sectors,
1271					      gfp, conf->bio_split);
1272		bio_chain(split, bio);
1273		generic_make_request(bio);
1274		bio = split;
1275		r1_bio->master_bio = bio;
1276		r1_bio->sectors = max_sectors;
1277	}
1278
1279	r1_bio->read_disk = rdisk;
1280
1281	read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1282
1283	r1_bio->bios[rdisk] = read_bio;
1284
1285	read_bio->bi_iter.bi_sector = r1_bio->sector +
1286		mirror->rdev->data_offset;
1287	bio_set_dev(read_bio, mirror->rdev->bdev);
1288	read_bio->bi_end_io = raid1_end_read_request;
1289	bio_set_op_attrs(read_bio, op, do_sync);
1290	if (test_bit(FailFast, &mirror->rdev->flags) &&
1291	    test_bit(R1BIO_FailFast, &r1_bio->state))
1292	        read_bio->bi_opf |= MD_FAILFAST;
1293	read_bio->bi_private = r1_bio;
1294
1295	if (mddev->gendisk)
1296	        trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1297				disk_devt(mddev->gendisk), r1_bio->sector);
1298
1299	generic_make_request(read_bio);
1300}
1301
1302static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1303				int max_write_sectors)
1304{
1305	struct r1conf *conf = mddev->private;
1306	struct r1bio *r1_bio;
1307	int i, disks;
1308	struct bitmap *bitmap = mddev->bitmap;
1309	unsigned long flags;
1310	struct md_rdev *blocked_rdev;
1311	struct blk_plug_cb *cb;
1312	struct raid1_plug_cb *plug = NULL;
1313	int first_clone;
1314	int max_sectors;
1315
1316	if (mddev_is_clustered(mddev) &&
1317	     md_cluster_ops->area_resyncing(mddev, WRITE,
1318		     bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1319
1320		DEFINE_WAIT(w);
1321		for (;;) {
1322			prepare_to_wait(&conf->wait_barrier,
1323					&w, TASK_IDLE);
1324			if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325							bio->bi_iter.bi_sector,
1326							bio_end_sector(bio)))
1327				break;
1328			schedule();
1329		}
1330		finish_wait(&conf->wait_barrier, &w);
1331	}
1332
1333	/*
1334	 * Register the new request and wait if the reconstruction
1335	 * thread has put up a bar for new requests.
1336	 * Continue immediately if no resync is active currently.
1337	 */
1338	wait_barrier(conf, bio->bi_iter.bi_sector);
1339
1340	r1_bio = alloc_r1bio(mddev, bio);
1341	r1_bio->sectors = max_write_sectors;
1342
1343	if (conf->pending_count >= max_queued_requests) {
1344		md_wakeup_thread(mddev->thread);
1345		raid1_log(mddev, "wait queued");
1346		wait_event(conf->wait_barrier,
1347			   conf->pending_count < max_queued_requests);
1348	}
1349	/* first select target devices under rcu_lock and
1350	 * inc refcount on their rdev.  Record them by setting
1351	 * bios[x] to bio
1352	 * If there are known/acknowledged bad blocks on any device on
1353	 * which we have seen a write error, we want to avoid writing those
1354	 * blocks.
1355	 * This potentially requires several writes to write around
1356	 * the bad blocks.  Each set of writes gets it's own r1bio
1357	 * with a set of bios attached.
1358	 */
1359
1360	disks = conf->raid_disks * 2;
1361 retry_write:
1362	blocked_rdev = NULL;
1363	rcu_read_lock();
1364	max_sectors = r1_bio->sectors;
1365	for (i = 0;  i < disks; i++) {
1366		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1367		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1368			atomic_inc(&rdev->nr_pending);
1369			blocked_rdev = rdev;
1370			break;
1371		}
1372		r1_bio->bios[i] = NULL;
1373		if (!rdev || test_bit(Faulty, &rdev->flags)) {
1374			if (i < conf->raid_disks)
1375				set_bit(R1BIO_Degraded, &r1_bio->state);
1376			continue;
1377		}
1378
1379		atomic_inc(&rdev->nr_pending);
1380		if (test_bit(WriteErrorSeen, &rdev->flags)) {
1381			sector_t first_bad;
1382			int bad_sectors;
1383			int is_bad;
1384
1385			is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1386					     &first_bad, &bad_sectors);
1387			if (is_bad < 0) {
1388				/* mustn't write here until the bad block is
1389				 * acknowledged*/
1390				set_bit(BlockedBadBlocks, &rdev->flags);
1391				blocked_rdev = rdev;
1392				break;
1393			}
1394			if (is_bad && first_bad <= r1_bio->sector) {
1395				/* Cannot write here at all */
1396				bad_sectors -= (r1_bio->sector - first_bad);
1397				if (bad_sectors < max_sectors)
1398					/* mustn't write more than bad_sectors
1399					 * to other devices yet
1400					 */
1401					max_sectors = bad_sectors;
1402				rdev_dec_pending(rdev, mddev);
1403				/* We don't set R1BIO_Degraded as that
1404				 * only applies if the disk is
1405				 * missing, so it might be re-added,
1406				 * and we want to know to recover this
1407				 * chunk.
1408				 * In this case the device is here,
1409				 * and the fact that this chunk is not
1410				 * in-sync is recorded in the bad
1411				 * block log
1412				 */
1413				continue;
1414			}
1415			if (is_bad) {
1416				int good_sectors = first_bad - r1_bio->sector;
1417				if (good_sectors < max_sectors)
1418					max_sectors = good_sectors;
1419			}
1420		}
1421		r1_bio->bios[i] = bio;
1422	}
1423	rcu_read_unlock();
1424
1425	if (unlikely(blocked_rdev)) {
1426		/* Wait for this device to become unblocked */
1427		int j;
1428
1429		for (j = 0; j < i; j++)
1430			if (r1_bio->bios[j])
1431				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1432		r1_bio->state = 0;
1433		allow_barrier(conf, bio->bi_iter.bi_sector);
1434		raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1435		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1436		wait_barrier(conf, bio->bi_iter.bi_sector);
1437		goto retry_write;
1438	}
1439
1440	if (max_sectors < bio_sectors(bio)) {
1441		struct bio *split = bio_split(bio, max_sectors,
1442					      GFP_NOIO, conf->bio_split);
1443		bio_chain(split, bio);
1444		generic_make_request(bio);
1445		bio = split;
1446		r1_bio->master_bio = bio;
1447		r1_bio->sectors = max_sectors;
1448	}
1449
1450	atomic_set(&r1_bio->remaining, 1);
1451	atomic_set(&r1_bio->behind_remaining, 0);
1452
1453	first_clone = 1;
1454
1455	for (i = 0; i < disks; i++) {
1456		struct bio *mbio = NULL;
1457		if (!r1_bio->bios[i])
1458			continue;
1459
1460
1461		if (first_clone) {
1462			/* do behind I/O ?
1463			 * Not if there are too many, or cannot
1464			 * allocate memory, or a reader on WriteMostly
1465			 * is waiting for behind writes to flush */
1466			if (bitmap &&
1467			    (atomic_read(&bitmap->behind_writes)
1468			     < mddev->bitmap_info.max_write_behind) &&
1469			    !waitqueue_active(&bitmap->behind_wait)) {
1470				alloc_behind_master_bio(r1_bio, bio);
1471			}
1472
1473			bitmap_startwrite(bitmap, r1_bio->sector,
1474					  r1_bio->sectors,
1475					  test_bit(R1BIO_BehindIO,
1476						   &r1_bio->state));
1477			first_clone = 0;
1478		}
1479
1480		if (r1_bio->behind_master_bio)
1481			mbio = bio_clone_fast(r1_bio->behind_master_bio,
1482					      GFP_NOIO, mddev->bio_set);
1483		else
1484			mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1485
1486		if (r1_bio->behind_master_bio) {
1487			if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 
 
 
 
 
 
 
 
 
1488				atomic_inc(&r1_bio->behind_remaining);
1489		}
1490
1491		r1_bio->bios[i] = mbio;
1492
1493		mbio->bi_iter.bi_sector	= (r1_bio->sector +
1494				   conf->mirrors[i].rdev->data_offset);
1495		bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1496		mbio->bi_end_io	= raid1_end_write_request;
1497		mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1498		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1499		    !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1500		    conf->raid_disks - mddev->degraded > 1)
1501			mbio->bi_opf |= MD_FAILFAST;
1502		mbio->bi_private = r1_bio;
1503
1504		atomic_inc(&r1_bio->remaining);
1505
1506		if (mddev->gendisk)
1507			trace_block_bio_remap(mbio->bi_disk->queue,
1508					      mbio, disk_devt(mddev->gendisk),
1509					      r1_bio->sector);
1510		/* flush_pending_writes() needs access to the rdev so...*/
1511		mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1512
1513		cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1514		if (cb)
1515			plug = container_of(cb, struct raid1_plug_cb, cb);
1516		else
1517			plug = NULL;
1518		if (plug) {
1519			bio_list_add(&plug->pending, mbio);
1520			plug->pending_cnt++;
1521		} else {
1522			spin_lock_irqsave(&conf->device_lock, flags);
1523			bio_list_add(&conf->pending_bio_list, mbio);
1524			conf->pending_count++;
1525			spin_unlock_irqrestore(&conf->device_lock, flags);
1526			md_wakeup_thread(mddev->thread);
1527		}
1528	}
1529
1530	r1_bio_write_done(r1_bio);
1531
1532	/* In case raid1d snuck in to freeze_array */
1533	wake_up(&conf->wait_barrier);
1534}
1535
1536static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1537{
1538	sector_t sectors;
1539
1540	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1541		md_flush_request(mddev, bio);
1542		return true;
1543	}
1544
1545	/*
1546	 * There is a limit to the maximum size, but
1547	 * the read/write handler might find a lower limit
1548	 * due to bad blocks.  To avoid multiple splits,
1549	 * we pass the maximum number of sectors down
1550	 * and let the lower level perform the split.
1551	 */
1552	sectors = align_to_barrier_unit_end(
1553		bio->bi_iter.bi_sector, bio_sectors(bio));
1554
1555	if (bio_data_dir(bio) == READ)
1556		raid1_read_request(mddev, bio, sectors, NULL);
1557	else {
1558		if (!md_write_start(mddev,bio))
1559			return false;
1560		raid1_write_request(mddev, bio, sectors);
1561	}
1562	return true;
1563}
1564
1565static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1566{
1567	struct r1conf *conf = mddev->private;
1568	int i;
1569
1570	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1571		   conf->raid_disks - mddev->degraded);
1572	rcu_read_lock();
1573	for (i = 0; i < conf->raid_disks; i++) {
1574		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1575		seq_printf(seq, "%s",
1576			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1577	}
1578	rcu_read_unlock();
1579	seq_printf(seq, "]");
1580}
1581
1582static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1583{
1584	char b[BDEVNAME_SIZE];
1585	struct r1conf *conf = mddev->private;
1586	unsigned long flags;
1587
1588	/*
1589	 * If it is not operational, then we have already marked it as dead
1590	 * else if it is the last working disks, ignore the error, let the
1591	 * next level up know.
1592	 * else mark the drive as failed
1593	 */
1594	spin_lock_irqsave(&conf->device_lock, flags);
1595	if (test_bit(In_sync, &rdev->flags)
1596	    && (conf->raid_disks - mddev->degraded) == 1) {
1597		/*
1598		 * Don't fail the drive, act as though we were just a
1599		 * normal single drive.
1600		 * However don't try a recovery from this drive as
1601		 * it is very likely to fail.
1602		 */
1603		conf->recovery_disabled = mddev->recovery_disabled;
1604		spin_unlock_irqrestore(&conf->device_lock, flags);
1605		return;
1606	}
1607	set_bit(Blocked, &rdev->flags);
1608	if (test_and_clear_bit(In_sync, &rdev->flags)) {
1609		mddev->degraded++;
1610		set_bit(Faulty, &rdev->flags);
1611	} else
1612		set_bit(Faulty, &rdev->flags);
1613	spin_unlock_irqrestore(&conf->device_lock, flags);
1614	/*
1615	 * if recovery is running, make sure it aborts.
1616	 */
1617	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1618	set_mask_bits(&mddev->sb_flags, 0,
1619		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1620	pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1621		"md/raid1:%s: Operation continuing on %d devices.\n",
1622		mdname(mddev), bdevname(rdev->bdev, b),
1623		mdname(mddev), conf->raid_disks - mddev->degraded);
1624}
1625
1626static void print_conf(struct r1conf *conf)
1627{
1628	int i;
1629
1630	pr_debug("RAID1 conf printout:\n");
1631	if (!conf) {
1632		pr_debug("(!conf)\n");
1633		return;
1634	}
1635	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1636		 conf->raid_disks);
1637
1638	rcu_read_lock();
1639	for (i = 0; i < conf->raid_disks; i++) {
1640		char b[BDEVNAME_SIZE];
1641		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1642		if (rdev)
1643			pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1644				 i, !test_bit(In_sync, &rdev->flags),
1645				 !test_bit(Faulty, &rdev->flags),
1646				 bdevname(rdev->bdev,b));
1647	}
1648	rcu_read_unlock();
1649}
1650
1651static void close_sync(struct r1conf *conf)
1652{
1653	int idx;
1654
1655	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1656		_wait_barrier(conf, idx);
1657		_allow_barrier(conf, idx);
1658	}
1659
1660	mempool_destroy(conf->r1buf_pool);
1661	conf->r1buf_pool = NULL;
1662}
1663
1664static int raid1_spare_active(struct mddev *mddev)
1665{
1666	int i;
1667	struct r1conf *conf = mddev->private;
1668	int count = 0;
1669	unsigned long flags;
1670
1671	/*
1672	 * Find all failed disks within the RAID1 configuration
1673	 * and mark them readable.
1674	 * Called under mddev lock, so rcu protection not needed.
1675	 * device_lock used to avoid races with raid1_end_read_request
1676	 * which expects 'In_sync' flags and ->degraded to be consistent.
1677	 */
1678	spin_lock_irqsave(&conf->device_lock, flags);
1679	for (i = 0; i < conf->raid_disks; i++) {
1680		struct md_rdev *rdev = conf->mirrors[i].rdev;
1681		struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1682		if (repl
1683		    && !test_bit(Candidate, &repl->flags)
1684		    && repl->recovery_offset == MaxSector
1685		    && !test_bit(Faulty, &repl->flags)
1686		    && !test_and_set_bit(In_sync, &repl->flags)) {
1687			/* replacement has just become active */
1688			if (!rdev ||
1689			    !test_and_clear_bit(In_sync, &rdev->flags))
1690				count++;
1691			if (rdev) {
1692				/* Replaced device not technically
1693				 * faulty, but we need to be sure
1694				 * it gets removed and never re-added
1695				 */
1696				set_bit(Faulty, &rdev->flags);
1697				sysfs_notify_dirent_safe(
1698					rdev->sysfs_state);
1699			}
1700		}
1701		if (rdev
1702		    && rdev->recovery_offset == MaxSector
1703		    && !test_bit(Faulty, &rdev->flags)
1704		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1705			count++;
1706			sysfs_notify_dirent_safe(rdev->sysfs_state);
1707		}
1708	}
1709	mddev->degraded -= count;
1710	spin_unlock_irqrestore(&conf->device_lock, flags);
1711
1712	print_conf(conf);
1713	return count;
1714}
1715
1716static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1717{
1718	struct r1conf *conf = mddev->private;
1719	int err = -EEXIST;
1720	int mirror = 0;
1721	struct raid1_info *p;
1722	int first = 0;
1723	int last = conf->raid_disks - 1;
1724
1725	if (mddev->recovery_disabled == conf->recovery_disabled)
1726		return -EBUSY;
1727
1728	if (md_integrity_add_rdev(rdev, mddev))
1729		return -ENXIO;
1730
1731	if (rdev->raid_disk >= 0)
1732		first = last = rdev->raid_disk;
1733
1734	/*
1735	 * find the disk ... but prefer rdev->saved_raid_disk
1736	 * if possible.
1737	 */
1738	if (rdev->saved_raid_disk >= 0 &&
1739	    rdev->saved_raid_disk >= first &&
 
1740	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1741		first = last = rdev->saved_raid_disk;
1742
1743	for (mirror = first; mirror <= last; mirror++) {
1744		p = conf->mirrors+mirror;
1745		if (!p->rdev) {
1746
1747			if (mddev->gendisk)
1748				disk_stack_limits(mddev->gendisk, rdev->bdev,
1749						  rdev->data_offset << 9);
1750
1751			p->head_position = 0;
1752			rdev->raid_disk = mirror;
1753			err = 0;
1754			/* As all devices are equivalent, we don't need a full recovery
1755			 * if this was recently any drive of the array
1756			 */
1757			if (rdev->saved_raid_disk < 0)
1758				conf->fullsync = 1;
1759			rcu_assign_pointer(p->rdev, rdev);
1760			break;
1761		}
1762		if (test_bit(WantReplacement, &p->rdev->flags) &&
1763		    p[conf->raid_disks].rdev == NULL) {
1764			/* Add this device as a replacement */
1765			clear_bit(In_sync, &rdev->flags);
1766			set_bit(Replacement, &rdev->flags);
1767			rdev->raid_disk = mirror;
1768			err = 0;
1769			conf->fullsync = 1;
1770			rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1771			break;
1772		}
1773	}
1774	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1775		blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1776	print_conf(conf);
1777	return err;
1778}
1779
1780static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1781{
1782	struct r1conf *conf = mddev->private;
1783	int err = 0;
1784	int number = rdev->raid_disk;
1785	struct raid1_info *p = conf->mirrors + number;
1786
1787	if (rdev != p->rdev)
1788		p = conf->mirrors + conf->raid_disks + number;
1789
1790	print_conf(conf);
1791	if (rdev == p->rdev) {
1792		if (test_bit(In_sync, &rdev->flags) ||
1793		    atomic_read(&rdev->nr_pending)) {
1794			err = -EBUSY;
1795			goto abort;
1796		}
1797		/* Only remove non-faulty devices if recovery
1798		 * is not possible.
1799		 */
1800		if (!test_bit(Faulty, &rdev->flags) &&
1801		    mddev->recovery_disabled != conf->recovery_disabled &&
1802		    mddev->degraded < conf->raid_disks) {
1803			err = -EBUSY;
1804			goto abort;
1805		}
1806		p->rdev = NULL;
1807		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1808			synchronize_rcu();
1809			if (atomic_read(&rdev->nr_pending)) {
1810				/* lost the race, try later */
1811				err = -EBUSY;
1812				p->rdev = rdev;
1813				goto abort;
1814			}
1815		}
1816		if (conf->mirrors[conf->raid_disks + number].rdev) {
1817			/* We just removed a device that is being replaced.
1818			 * Move down the replacement.  We drain all IO before
1819			 * doing this to avoid confusion.
1820			 */
1821			struct md_rdev *repl =
1822				conf->mirrors[conf->raid_disks + number].rdev;
1823			freeze_array(conf, 0);
1824			if (atomic_read(&repl->nr_pending)) {
1825				/* It means that some queued IO of retry_list
1826				 * hold repl. Thus, we cannot set replacement
1827				 * as NULL, avoiding rdev NULL pointer
1828				 * dereference in sync_request_write and
1829				 * handle_write_finished.
1830				 */
1831				err = -EBUSY;
1832				unfreeze_array(conf);
1833				goto abort;
1834			}
1835			clear_bit(Replacement, &repl->flags);
1836			p->rdev = repl;
1837			conf->mirrors[conf->raid_disks + number].rdev = NULL;
1838			unfreeze_array(conf);
1839		}
1840
1841		clear_bit(WantReplacement, &rdev->flags);
1842		err = md_integrity_register(mddev);
1843	}
1844abort:
1845
1846	print_conf(conf);
1847	return err;
1848}
1849
1850static void end_sync_read(struct bio *bio)
1851{
1852	struct r1bio *r1_bio = get_resync_r1bio(bio);
1853
1854	update_head_pos(r1_bio->read_disk, r1_bio);
1855
1856	/*
1857	 * we have read a block, now it needs to be re-written,
1858	 * or re-read if the read failed.
1859	 * We don't do much here, just schedule handling by raid1d
1860	 */
1861	if (!bio->bi_status)
1862		set_bit(R1BIO_Uptodate, &r1_bio->state);
1863
1864	if (atomic_dec_and_test(&r1_bio->remaining))
1865		reschedule_retry(r1_bio);
1866}
1867
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1868static void end_sync_write(struct bio *bio)
1869{
1870	int uptodate = !bio->bi_status;
1871	struct r1bio *r1_bio = get_resync_r1bio(bio);
1872	struct mddev *mddev = r1_bio->mddev;
1873	struct r1conf *conf = mddev->private;
1874	sector_t first_bad;
1875	int bad_sectors;
1876	struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1877
1878	if (!uptodate) {
1879		sector_t sync_blocks = 0;
1880		sector_t s = r1_bio->sector;
1881		long sectors_to_go = r1_bio->sectors;
1882		/* make sure these bits doesn't get cleared. */
1883		do {
1884			bitmap_end_sync(mddev->bitmap, s,
1885					&sync_blocks, 1);
1886			s += sync_blocks;
1887			sectors_to_go -= sync_blocks;
1888		} while (sectors_to_go > 0);
1889		set_bit(WriteErrorSeen, &rdev->flags);
1890		if (!test_and_set_bit(WantReplacement, &rdev->flags))
1891			set_bit(MD_RECOVERY_NEEDED, &
1892				mddev->recovery);
1893		set_bit(R1BIO_WriteError, &r1_bio->state);
1894	} else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1895			       &first_bad, &bad_sectors) &&
1896		   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1897				r1_bio->sector,
1898				r1_bio->sectors,
1899				&first_bad, &bad_sectors)
1900		)
1901		set_bit(R1BIO_MadeGood, &r1_bio->state);
1902
1903	if (atomic_dec_and_test(&r1_bio->remaining)) {
1904		int s = r1_bio->sectors;
1905		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1906		    test_bit(R1BIO_WriteError, &r1_bio->state))
1907			reschedule_retry(r1_bio);
1908		else {
1909			put_buf(r1_bio);
1910			md_done_sync(mddev, s, uptodate);
1911		}
1912	}
1913}
1914
1915static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1916			    int sectors, struct page *page, int rw)
1917{
1918	if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1919		/* success */
1920		return 1;
1921	if (rw == WRITE) {
1922		set_bit(WriteErrorSeen, &rdev->flags);
1923		if (!test_and_set_bit(WantReplacement,
1924				      &rdev->flags))
1925			set_bit(MD_RECOVERY_NEEDED, &
1926				rdev->mddev->recovery);
1927	}
1928	/* need to record an error - either for the block or the device */
1929	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1930		md_error(rdev->mddev, rdev);
1931	return 0;
1932}
1933
1934static int fix_sync_read_error(struct r1bio *r1_bio)
1935{
1936	/* Try some synchronous reads of other devices to get
1937	 * good data, much like with normal read errors.  Only
1938	 * read into the pages we already have so we don't
1939	 * need to re-issue the read request.
1940	 * We don't need to freeze the array, because being in an
1941	 * active sync request, there is no normal IO, and
1942	 * no overlapping syncs.
1943	 * We don't need to check is_badblock() again as we
1944	 * made sure that anything with a bad block in range
1945	 * will have bi_end_io clear.
1946	 */
1947	struct mddev *mddev = r1_bio->mddev;
1948	struct r1conf *conf = mddev->private;
1949	struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1950	struct page **pages = get_resync_pages(bio)->pages;
1951	sector_t sect = r1_bio->sector;
1952	int sectors = r1_bio->sectors;
1953	int idx = 0;
1954	struct md_rdev *rdev;
1955
1956	rdev = conf->mirrors[r1_bio->read_disk].rdev;
1957	if (test_bit(FailFast, &rdev->flags)) {
1958		/* Don't try recovering from here - just fail it
1959		 * ... unless it is the last working device of course */
1960		md_error(mddev, rdev);
1961		if (test_bit(Faulty, &rdev->flags))
1962			/* Don't try to read from here, but make sure
1963			 * put_buf does it's thing
1964			 */
1965			bio->bi_end_io = end_sync_write;
1966	}
1967
1968	while(sectors) {
1969		int s = sectors;
1970		int d = r1_bio->read_disk;
1971		int success = 0;
1972		int start;
1973
1974		if (s > (PAGE_SIZE>>9))
1975			s = PAGE_SIZE >> 9;
1976		do {
1977			if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1978				/* No rcu protection needed here devices
1979				 * can only be removed when no resync is
1980				 * active, and resync is currently active
1981				 */
1982				rdev = conf->mirrors[d].rdev;
1983				if (sync_page_io(rdev, sect, s<<9,
1984						 pages[idx],
1985						 REQ_OP_READ, 0, false)) {
1986					success = 1;
1987					break;
1988				}
1989			}
1990			d++;
1991			if (d == conf->raid_disks * 2)
1992				d = 0;
1993		} while (!success && d != r1_bio->read_disk);
1994
1995		if (!success) {
1996			char b[BDEVNAME_SIZE];
1997			int abort = 0;
1998			/* Cannot read from anywhere, this block is lost.
1999			 * Record a bad block on each device.  If that doesn't
2000			 * work just disable and interrupt the recovery.
2001			 * Don't fail devices as that won't really help.
2002			 */
2003			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2004					    mdname(mddev), bio_devname(bio, b),
2005					    (unsigned long long)r1_bio->sector);
2006			for (d = 0; d < conf->raid_disks * 2; d++) {
2007				rdev = conf->mirrors[d].rdev;
2008				if (!rdev || test_bit(Faulty, &rdev->flags))
2009					continue;
2010				if (!rdev_set_badblocks(rdev, sect, s, 0))
2011					abort = 1;
2012			}
2013			if (abort) {
2014				conf->recovery_disabled =
2015					mddev->recovery_disabled;
2016				set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2017				md_done_sync(mddev, r1_bio->sectors, 0);
2018				put_buf(r1_bio);
2019				return 0;
2020			}
2021			/* Try next page */
2022			sectors -= s;
2023			sect += s;
2024			idx++;
2025			continue;
2026		}
2027
2028		start = d;
2029		/* write it back and re-read */
2030		while (d != r1_bio->read_disk) {
2031			if (d == 0)
2032				d = conf->raid_disks * 2;
2033			d--;
2034			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2035				continue;
2036			rdev = conf->mirrors[d].rdev;
2037			if (r1_sync_page_io(rdev, sect, s,
2038					    pages[idx],
2039					    WRITE) == 0) {
2040				r1_bio->bios[d]->bi_end_io = NULL;
2041				rdev_dec_pending(rdev, mddev);
2042			}
2043		}
2044		d = start;
2045		while (d != r1_bio->read_disk) {
2046			if (d == 0)
2047				d = conf->raid_disks * 2;
2048			d--;
2049			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2050				continue;
2051			rdev = conf->mirrors[d].rdev;
2052			if (r1_sync_page_io(rdev, sect, s,
2053					    pages[idx],
2054					    READ) != 0)
2055				atomic_add(s, &rdev->corrected_errors);
2056		}
2057		sectors -= s;
2058		sect += s;
2059		idx ++;
2060	}
2061	set_bit(R1BIO_Uptodate, &r1_bio->state);
2062	bio->bi_status = 0;
2063	return 1;
2064}
2065
2066static void process_checks(struct r1bio *r1_bio)
2067{
2068	/* We have read all readable devices.  If we haven't
2069	 * got the block, then there is no hope left.
2070	 * If we have, then we want to do a comparison
2071	 * and skip the write if everything is the same.
2072	 * If any blocks failed to read, then we need to
2073	 * attempt an over-write
2074	 */
2075	struct mddev *mddev = r1_bio->mddev;
2076	struct r1conf *conf = mddev->private;
2077	int primary;
2078	int i;
2079	int vcnt;
2080
2081	/* Fix variable parts of all bios */
2082	vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2083	for (i = 0; i < conf->raid_disks * 2; i++) {
2084		blk_status_t status;
2085		struct bio *b = r1_bio->bios[i];
2086		struct resync_pages *rp = get_resync_pages(b);
2087		if (b->bi_end_io != end_sync_read)
2088			continue;
2089		/* fixup the bio for reuse, but preserve errno */
2090		status = b->bi_status;
2091		bio_reset(b);
2092		b->bi_status = status;
2093		b->bi_iter.bi_sector = r1_bio->sector +
2094			conf->mirrors[i].rdev->data_offset;
2095		bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2096		b->bi_end_io = end_sync_read;
2097		rp->raid_bio = r1_bio;
2098		b->bi_private = rp;
2099
2100		/* initialize bvec table again */
2101		md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2102	}
2103	for (primary = 0; primary < conf->raid_disks * 2; primary++)
2104		if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2105		    !r1_bio->bios[primary]->bi_status) {
2106			r1_bio->bios[primary]->bi_end_io = NULL;
2107			rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2108			break;
2109		}
2110	r1_bio->read_disk = primary;
2111	for (i = 0; i < conf->raid_disks * 2; i++) {
2112		int j;
2113		struct bio *pbio = r1_bio->bios[primary];
2114		struct bio *sbio = r1_bio->bios[i];
2115		blk_status_t status = sbio->bi_status;
2116		struct page **ppages = get_resync_pages(pbio)->pages;
2117		struct page **spages = get_resync_pages(sbio)->pages;
2118		struct bio_vec *bi;
2119		int page_len[RESYNC_PAGES] = { 0 };
 
2120
2121		if (sbio->bi_end_io != end_sync_read)
2122			continue;
2123		/* Now we can 'fixup' the error value */
2124		sbio->bi_status = 0;
2125
2126		bio_for_each_segment_all(bi, sbio, j)
2127			page_len[j] = bi->bv_len;
2128
2129		if (!status) {
2130			for (j = vcnt; j-- ; ) {
2131				if (memcmp(page_address(ppages[j]),
2132					   page_address(spages[j]),
2133					   page_len[j]))
2134					break;
2135			}
2136		} else
2137			j = 0;
2138		if (j >= 0)
2139			atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2140		if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2141			      && !status)) {
2142			/* No need to write to this device. */
2143			sbio->bi_end_io = NULL;
2144			rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2145			continue;
2146		}
2147
2148		bio_copy_data(sbio, pbio);
2149	}
2150}
2151
2152static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2153{
2154	struct r1conf *conf = mddev->private;
2155	int i;
2156	int disks = conf->raid_disks * 2;
2157	struct bio *wbio;
2158
2159	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2160		/* ouch - failed to read all of that. */
2161		if (!fix_sync_read_error(r1_bio))
2162			return;
2163
2164	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2165		process_checks(r1_bio);
2166
2167	/*
2168	 * schedule writes
2169	 */
2170	atomic_set(&r1_bio->remaining, 1);
2171	for (i = 0; i < disks ; i++) {
2172		wbio = r1_bio->bios[i];
2173		if (wbio->bi_end_io == NULL ||
2174		    (wbio->bi_end_io == end_sync_read &&
2175		     (i == r1_bio->read_disk ||
2176		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2177			continue;
2178		if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
 
2179			continue;
 
2180
2181		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2182		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2183			wbio->bi_opf |= MD_FAILFAST;
2184
2185		wbio->bi_end_io = end_sync_write;
2186		atomic_inc(&r1_bio->remaining);
2187		md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2188
2189		generic_make_request(wbio);
2190	}
2191
2192	if (atomic_dec_and_test(&r1_bio->remaining)) {
2193		/* if we're here, all write(s) have completed, so clean up */
2194		int s = r1_bio->sectors;
2195		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2196		    test_bit(R1BIO_WriteError, &r1_bio->state))
2197			reschedule_retry(r1_bio);
2198		else {
2199			put_buf(r1_bio);
2200			md_done_sync(mddev, s, 1);
2201		}
2202	}
2203}
2204
2205/*
2206 * This is a kernel thread which:
2207 *
2208 *	1.	Retries failed read operations on working mirrors.
2209 *	2.	Updates the raid superblock when problems encounter.
2210 *	3.	Performs writes following reads for array synchronising.
2211 */
2212
2213static void fix_read_error(struct r1conf *conf, int read_disk,
2214			   sector_t sect, int sectors)
2215{
2216	struct mddev *mddev = conf->mddev;
2217	while(sectors) {
2218		int s = sectors;
2219		int d = read_disk;
2220		int success = 0;
2221		int start;
2222		struct md_rdev *rdev;
2223
2224		if (s > (PAGE_SIZE>>9))
2225			s = PAGE_SIZE >> 9;
2226
2227		do {
2228			sector_t first_bad;
2229			int bad_sectors;
2230
2231			rcu_read_lock();
2232			rdev = rcu_dereference(conf->mirrors[d].rdev);
2233			if (rdev &&
2234			    (test_bit(In_sync, &rdev->flags) ||
2235			     (!test_bit(Faulty, &rdev->flags) &&
2236			      rdev->recovery_offset >= sect + s)) &&
2237			    is_badblock(rdev, sect, s,
2238					&first_bad, &bad_sectors) == 0) {
2239				atomic_inc(&rdev->nr_pending);
2240				rcu_read_unlock();
2241				if (sync_page_io(rdev, sect, s<<9,
2242					 conf->tmppage, REQ_OP_READ, 0, false))
2243					success = 1;
2244				rdev_dec_pending(rdev, mddev);
2245				if (success)
2246					break;
2247			} else
2248				rcu_read_unlock();
2249			d++;
2250			if (d == conf->raid_disks * 2)
2251				d = 0;
2252		} while (!success && d != read_disk);
2253
2254		if (!success) {
2255			/* Cannot read from anywhere - mark it bad */
2256			struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2257			if (!rdev_set_badblocks(rdev, sect, s, 0))
2258				md_error(mddev, rdev);
2259			break;
2260		}
2261		/* write it back and re-read */
2262		start = d;
2263		while (d != read_disk) {
2264			if (d==0)
2265				d = conf->raid_disks * 2;
2266			d--;
2267			rcu_read_lock();
2268			rdev = rcu_dereference(conf->mirrors[d].rdev);
2269			if (rdev &&
2270			    !test_bit(Faulty, &rdev->flags)) {
2271				atomic_inc(&rdev->nr_pending);
2272				rcu_read_unlock();
2273				r1_sync_page_io(rdev, sect, s,
2274						conf->tmppage, WRITE);
2275				rdev_dec_pending(rdev, mddev);
2276			} else
2277				rcu_read_unlock();
2278		}
2279		d = start;
2280		while (d != read_disk) {
2281			char b[BDEVNAME_SIZE];
2282			if (d==0)
2283				d = conf->raid_disks * 2;
2284			d--;
2285			rcu_read_lock();
2286			rdev = rcu_dereference(conf->mirrors[d].rdev);
2287			if (rdev &&
2288			    !test_bit(Faulty, &rdev->flags)) {
2289				atomic_inc(&rdev->nr_pending);
2290				rcu_read_unlock();
2291				if (r1_sync_page_io(rdev, sect, s,
2292						    conf->tmppage, READ)) {
2293					atomic_add(s, &rdev->corrected_errors);
2294					pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2295						mdname(mddev), s,
2296						(unsigned long long)(sect +
2297								     rdev->data_offset),
2298						bdevname(rdev->bdev, b));
2299				}
2300				rdev_dec_pending(rdev, mddev);
2301			} else
2302				rcu_read_unlock();
2303		}
2304		sectors -= s;
2305		sect += s;
2306	}
2307}
2308
2309static int narrow_write_error(struct r1bio *r1_bio, int i)
2310{
2311	struct mddev *mddev = r1_bio->mddev;
2312	struct r1conf *conf = mddev->private;
2313	struct md_rdev *rdev = conf->mirrors[i].rdev;
2314
2315	/* bio has the data to be written to device 'i' where
2316	 * we just recently had a write error.
2317	 * We repeatedly clone the bio and trim down to one block,
2318	 * then try the write.  Where the write fails we record
2319	 * a bad block.
2320	 * It is conceivable that the bio doesn't exactly align with
2321	 * blocks.  We must handle this somehow.
2322	 *
2323	 * We currently own a reference on the rdev.
2324	 */
2325
2326	int block_sectors;
2327	sector_t sector;
2328	int sectors;
2329	int sect_to_write = r1_bio->sectors;
2330	int ok = 1;
2331
2332	if (rdev->badblocks.shift < 0)
2333		return 0;
2334
2335	block_sectors = roundup(1 << rdev->badblocks.shift,
2336				bdev_logical_block_size(rdev->bdev) >> 9);
2337	sector = r1_bio->sector;
2338	sectors = ((sector + block_sectors)
2339		   & ~(sector_t)(block_sectors - 1))
2340		- sector;
2341
2342	while (sect_to_write) {
2343		struct bio *wbio;
2344		if (sectors > sect_to_write)
2345			sectors = sect_to_write;
2346		/* Write at 'sector' for 'sectors'*/
2347
2348		if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2349			wbio = bio_clone_fast(r1_bio->behind_master_bio,
2350					      GFP_NOIO,
2351					      mddev->bio_set);
2352		} else {
2353			wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2354					      mddev->bio_set);
2355		}
2356
2357		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2358		wbio->bi_iter.bi_sector = r1_bio->sector;
2359		wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2360
2361		bio_trim(wbio, sector - r1_bio->sector, sectors);
2362		wbio->bi_iter.bi_sector += rdev->data_offset;
2363		bio_set_dev(wbio, rdev->bdev);
2364
2365		if (submit_bio_wait(wbio) < 0)
2366			/* failure! */
2367			ok = rdev_set_badblocks(rdev, sector,
2368						sectors, 0)
2369				&& ok;
2370
2371		bio_put(wbio);
2372		sect_to_write -= sectors;
2373		sector += sectors;
2374		sectors = block_sectors;
2375	}
2376	return ok;
2377}
2378
2379static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2380{
2381	int m;
2382	int s = r1_bio->sectors;
2383	for (m = 0; m < conf->raid_disks * 2 ; m++) {
2384		struct md_rdev *rdev = conf->mirrors[m].rdev;
2385		struct bio *bio = r1_bio->bios[m];
2386		if (bio->bi_end_io == NULL)
2387			continue;
2388		if (!bio->bi_status &&
2389		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2390			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2391		}
2392		if (bio->bi_status &&
2393		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2394			if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2395				md_error(conf->mddev, rdev);
2396		}
2397	}
2398	put_buf(r1_bio);
2399	md_done_sync(conf->mddev, s, 1);
2400}
2401
2402static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2403{
2404	int m, idx;
2405	bool fail = false;
2406
2407	for (m = 0; m < conf->raid_disks * 2 ; m++)
2408		if (r1_bio->bios[m] == IO_MADE_GOOD) {
2409			struct md_rdev *rdev = conf->mirrors[m].rdev;
2410			rdev_clear_badblocks(rdev,
2411					     r1_bio->sector,
2412					     r1_bio->sectors, 0);
2413			rdev_dec_pending(rdev, conf->mddev);
2414		} else if (r1_bio->bios[m] != NULL) {
2415			/* This drive got a write error.  We need to
2416			 * narrow down and record precise write
2417			 * errors.
2418			 */
2419			fail = true;
2420			if (!narrow_write_error(r1_bio, m)) {
2421				md_error(conf->mddev,
2422					 conf->mirrors[m].rdev);
2423				/* an I/O failed, we can't clear the bitmap */
2424				set_bit(R1BIO_Degraded, &r1_bio->state);
2425			}
2426			rdev_dec_pending(conf->mirrors[m].rdev,
2427					 conf->mddev);
2428		}
2429	if (fail) {
2430		spin_lock_irq(&conf->device_lock);
2431		list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2432		idx = sector_to_idx(r1_bio->sector);
2433		atomic_inc(&conf->nr_queued[idx]);
2434		spin_unlock_irq(&conf->device_lock);
2435		/*
2436		 * In case freeze_array() is waiting for condition
2437		 * get_unqueued_pending() == extra to be true.
2438		 */
2439		wake_up(&conf->wait_barrier);
2440		md_wakeup_thread(conf->mddev->thread);
2441	} else {
2442		if (test_bit(R1BIO_WriteError, &r1_bio->state))
2443			close_write(r1_bio);
2444		raid_end_bio_io(r1_bio);
2445	}
2446}
2447
2448static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2449{
2450	struct mddev *mddev = conf->mddev;
2451	struct bio *bio;
2452	struct md_rdev *rdev;
2453	sector_t bio_sector;
2454
2455	clear_bit(R1BIO_ReadError, &r1_bio->state);
2456	/* we got a read error. Maybe the drive is bad.  Maybe just
2457	 * the block and we can fix it.
2458	 * We freeze all other IO, and try reading the block from
2459	 * other devices.  When we find one, we re-write
2460	 * and check it that fixes the read error.
2461	 * This is all done synchronously while the array is
2462	 * frozen
2463	 */
2464
2465	bio = r1_bio->bios[r1_bio->read_disk];
2466	bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2467	bio_put(bio);
2468	r1_bio->bios[r1_bio->read_disk] = NULL;
2469
2470	rdev = conf->mirrors[r1_bio->read_disk].rdev;
2471	if (mddev->ro == 0
2472	    && !test_bit(FailFast, &rdev->flags)) {
2473		freeze_array(conf, 1);
2474		fix_read_error(conf, r1_bio->read_disk,
2475			       r1_bio->sector, r1_bio->sectors);
2476		unfreeze_array(conf);
 
 
2477	} else {
2478		r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2479	}
2480
2481	rdev_dec_pending(rdev, conf->mddev);
2482	allow_barrier(conf, r1_bio->sector);
2483	bio = r1_bio->master_bio;
2484
2485	/* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2486	r1_bio->state = 0;
2487	raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2488}
2489
2490static void raid1d(struct md_thread *thread)
2491{
2492	struct mddev *mddev = thread->mddev;
2493	struct r1bio *r1_bio;
2494	unsigned long flags;
2495	struct r1conf *conf = mddev->private;
2496	struct list_head *head = &conf->retry_list;
2497	struct blk_plug plug;
2498	int idx;
2499
2500	md_check_recovery(mddev);
2501
2502	if (!list_empty_careful(&conf->bio_end_io_list) &&
2503	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2504		LIST_HEAD(tmp);
2505		spin_lock_irqsave(&conf->device_lock, flags);
2506		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2507			list_splice_init(&conf->bio_end_io_list, &tmp);
2508		spin_unlock_irqrestore(&conf->device_lock, flags);
2509		while (!list_empty(&tmp)) {
2510			r1_bio = list_first_entry(&tmp, struct r1bio,
2511						  retry_list);
2512			list_del(&r1_bio->retry_list);
2513			idx = sector_to_idx(r1_bio->sector);
2514			atomic_dec(&conf->nr_queued[idx]);
2515			if (mddev->degraded)
2516				set_bit(R1BIO_Degraded, &r1_bio->state);
2517			if (test_bit(R1BIO_WriteError, &r1_bio->state))
2518				close_write(r1_bio);
2519			raid_end_bio_io(r1_bio);
2520		}
2521	}
2522
2523	blk_start_plug(&plug);
2524	for (;;) {
2525
2526		flush_pending_writes(conf);
2527
2528		spin_lock_irqsave(&conf->device_lock, flags);
2529		if (list_empty(head)) {
2530			spin_unlock_irqrestore(&conf->device_lock, flags);
2531			break;
2532		}
2533		r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2534		list_del(head->prev);
2535		idx = sector_to_idx(r1_bio->sector);
2536		atomic_dec(&conf->nr_queued[idx]);
2537		spin_unlock_irqrestore(&conf->device_lock, flags);
2538
2539		mddev = r1_bio->mddev;
2540		conf = mddev->private;
2541		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2542			if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2543			    test_bit(R1BIO_WriteError, &r1_bio->state))
2544				handle_sync_write_finished(conf, r1_bio);
2545			else
2546				sync_request_write(mddev, r1_bio);
2547		} else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2548			   test_bit(R1BIO_WriteError, &r1_bio->state))
2549			handle_write_finished(conf, r1_bio);
2550		else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2551			handle_read_error(conf, r1_bio);
2552		else
2553			WARN_ON_ONCE(1);
2554
2555		cond_resched();
2556		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2557			md_check_recovery(mddev);
2558	}
2559	blk_finish_plug(&plug);
2560}
2561
2562static int init_resync(struct r1conf *conf)
2563{
2564	int buffs;
2565
2566	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2567	BUG_ON(conf->r1buf_pool);
2568	conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2569					  conf->poolinfo);
2570	if (!conf->r1buf_pool)
2571		return -ENOMEM;
2572	return 0;
2573}
2574
2575static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2576{
2577	struct r1bio *r1bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2578	struct resync_pages *rps;
2579	struct bio *bio;
2580	int i;
2581
2582	for (i = conf->poolinfo->raid_disks; i--; ) {
2583		bio = r1bio->bios[i];
2584		rps = bio->bi_private;
2585		bio_reset(bio);
2586		bio->bi_private = rps;
2587	}
2588	r1bio->master_bio = NULL;
2589	return r1bio;
2590}
2591
2592/*
2593 * perform a "sync" on one "block"
2594 *
2595 * We need to make sure that no normal I/O request - particularly write
2596 * requests - conflict with active sync requests.
2597 *
2598 * This is achieved by tracking pending requests and a 'barrier' concept
2599 * that can be installed to exclude normal IO requests.
2600 */
2601
2602static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2603				   int *skipped)
2604{
2605	struct r1conf *conf = mddev->private;
2606	struct r1bio *r1_bio;
2607	struct bio *bio;
2608	sector_t max_sector, nr_sectors;
2609	int disk = -1;
2610	int i;
2611	int wonly = -1;
2612	int write_targets = 0, read_targets = 0;
2613	sector_t sync_blocks;
2614	int still_degraded = 0;
2615	int good_sectors = RESYNC_SECTORS;
2616	int min_bad = 0; /* number of sectors that are bad in all devices */
2617	int idx = sector_to_idx(sector_nr);
2618	int page_idx = 0;
2619
2620	if (!conf->r1buf_pool)
2621		if (init_resync(conf))
2622			return 0;
2623
2624	max_sector = mddev->dev_sectors;
2625	if (sector_nr >= max_sector) {
2626		/* If we aborted, we need to abort the
2627		 * sync on the 'current' bitmap chunk (there will
2628		 * only be one in raid1 resync.
2629		 * We can find the current addess in mddev->curr_resync
2630		 */
2631		if (mddev->curr_resync < max_sector) /* aborted */
2632			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2633						&sync_blocks, 1);
2634		else /* completed sync */
2635			conf->fullsync = 0;
2636
2637		bitmap_close_sync(mddev->bitmap);
2638		close_sync(conf);
2639
2640		if (mddev_is_clustered(mddev)) {
2641			conf->cluster_sync_low = 0;
2642			conf->cluster_sync_high = 0;
2643		}
2644		return 0;
2645	}
2646
2647	if (mddev->bitmap == NULL &&
2648	    mddev->recovery_cp == MaxSector &&
2649	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2650	    conf->fullsync == 0) {
2651		*skipped = 1;
2652		return max_sector - sector_nr;
2653	}
2654	/* before building a request, check if we can skip these blocks..
2655	 * This call the bitmap_start_sync doesn't actually record anything
2656	 */
2657	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2658	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2659		/* We can skip this block, and probably several more */
2660		*skipped = 1;
2661		return sync_blocks;
2662	}
2663
2664	/*
2665	 * If there is non-resync activity waiting for a turn, then let it
2666	 * though before starting on this new sync request.
2667	 */
2668	if (atomic_read(&conf->nr_waiting[idx]))
2669		schedule_timeout_uninterruptible(1);
2670
2671	/* we are incrementing sector_nr below. To be safe, we check against
2672	 * sector_nr + two times RESYNC_SECTORS
2673	 */
2674
2675	bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2676		mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2677
2678
2679	if (raise_barrier(conf, sector_nr))
2680		return 0;
2681
2682	r1_bio = raid1_alloc_init_r1buf(conf);
2683
2684	rcu_read_lock();
2685	/*
2686	 * If we get a correctably read error during resync or recovery,
2687	 * we might want to read from a different device.  So we
2688	 * flag all drives that could conceivably be read from for READ,
2689	 * and any others (which will be non-In_sync devices) for WRITE.
2690	 * If a read fails, we try reading from something else for which READ
2691	 * is OK.
2692	 */
2693
2694	r1_bio->mddev = mddev;
2695	r1_bio->sector = sector_nr;
2696	r1_bio->state = 0;
2697	set_bit(R1BIO_IsSync, &r1_bio->state);
2698	/* make sure good_sectors won't go across barrier unit boundary */
2699	good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2700
2701	for (i = 0; i < conf->raid_disks * 2; i++) {
2702		struct md_rdev *rdev;
2703		bio = r1_bio->bios[i];
2704
2705		rdev = rcu_dereference(conf->mirrors[i].rdev);
2706		if (rdev == NULL ||
2707		    test_bit(Faulty, &rdev->flags)) {
2708			if (i < conf->raid_disks)
2709				still_degraded = 1;
2710		} else if (!test_bit(In_sync, &rdev->flags)) {
2711			bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2712			bio->bi_end_io = end_sync_write;
2713			write_targets ++;
2714		} else {
2715			/* may need to read from here */
2716			sector_t first_bad = MaxSector;
2717			int bad_sectors;
2718
2719			if (is_badblock(rdev, sector_nr, good_sectors,
2720					&first_bad, &bad_sectors)) {
2721				if (first_bad > sector_nr)
2722					good_sectors = first_bad - sector_nr;
2723				else {
2724					bad_sectors -= (sector_nr - first_bad);
2725					if (min_bad == 0 ||
2726					    min_bad > bad_sectors)
2727						min_bad = bad_sectors;
2728				}
2729			}
2730			if (sector_nr < first_bad) {
2731				if (test_bit(WriteMostly, &rdev->flags)) {
2732					if (wonly < 0)
2733						wonly = i;
2734				} else {
2735					if (disk < 0)
2736						disk = i;
2737				}
2738				bio_set_op_attrs(bio, REQ_OP_READ, 0);
2739				bio->bi_end_io = end_sync_read;
2740				read_targets++;
2741			} else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2742				test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2743				!test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2744				/*
2745				 * The device is suitable for reading (InSync),
2746				 * but has bad block(s) here. Let's try to correct them,
2747				 * if we are doing resync or repair. Otherwise, leave
2748				 * this device alone for this sync request.
2749				 */
2750				bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2751				bio->bi_end_io = end_sync_write;
2752				write_targets++;
2753			}
2754		}
2755		if (bio->bi_end_io) {
2756			atomic_inc(&rdev->nr_pending);
2757			bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2758			bio_set_dev(bio, rdev->bdev);
2759			if (test_bit(FailFast, &rdev->flags))
2760				bio->bi_opf |= MD_FAILFAST;
2761		}
2762	}
2763	rcu_read_unlock();
2764	if (disk < 0)
2765		disk = wonly;
2766	r1_bio->read_disk = disk;
2767
2768	if (read_targets == 0 && min_bad > 0) {
2769		/* These sectors are bad on all InSync devices, so we
2770		 * need to mark them bad on all write targets
2771		 */
2772		int ok = 1;
2773		for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2774			if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2775				struct md_rdev *rdev = conf->mirrors[i].rdev;
2776				ok = rdev_set_badblocks(rdev, sector_nr,
2777							min_bad, 0
2778					) && ok;
2779			}
2780		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2781		*skipped = 1;
2782		put_buf(r1_bio);
2783
2784		if (!ok) {
2785			/* Cannot record the badblocks, so need to
2786			 * abort the resync.
2787			 * If there are multiple read targets, could just
2788			 * fail the really bad ones ???
2789			 */
2790			conf->recovery_disabled = mddev->recovery_disabled;
2791			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2792			return 0;
2793		} else
2794			return min_bad;
2795
2796	}
2797	if (min_bad > 0 && min_bad < good_sectors) {
2798		/* only resync enough to reach the next bad->good
2799		 * transition */
2800		good_sectors = min_bad;
2801	}
2802
2803	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2804		/* extra read targets are also write targets */
2805		write_targets += read_targets-1;
2806
2807	if (write_targets == 0 || read_targets == 0) {
2808		/* There is nowhere to write, so all non-sync
2809		 * drives must be failed - so we are finished
2810		 */
2811		sector_t rv;
2812		if (min_bad > 0)
2813			max_sector = sector_nr + min_bad;
2814		rv = max_sector - sector_nr;
2815		*skipped = 1;
2816		put_buf(r1_bio);
2817		return rv;
2818	}
2819
2820	if (max_sector > mddev->resync_max)
2821		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2822	if (max_sector > sector_nr + good_sectors)
2823		max_sector = sector_nr + good_sectors;
2824	nr_sectors = 0;
2825	sync_blocks = 0;
2826	do {
2827		struct page *page;
2828		int len = PAGE_SIZE;
2829		if (sector_nr + (len>>9) > max_sector)
2830			len = (max_sector - sector_nr) << 9;
2831		if (len == 0)
2832			break;
2833		if (sync_blocks == 0) {
2834			if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2835					       &sync_blocks, still_degraded) &&
2836			    !conf->fullsync &&
2837			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2838				break;
2839			if ((len >> 9) > sync_blocks)
2840				len = sync_blocks<<9;
2841		}
2842
2843		for (i = 0 ; i < conf->raid_disks * 2; i++) {
2844			struct resync_pages *rp;
2845
2846			bio = r1_bio->bios[i];
2847			rp = get_resync_pages(bio);
2848			if (bio->bi_end_io) {
2849				page = resync_fetch_page(rp, page_idx);
2850
2851				/*
2852				 * won't fail because the vec table is big
2853				 * enough to hold all these pages
2854				 */
2855				bio_add_page(bio, page, len, 0);
2856			}
2857		}
2858		nr_sectors += len>>9;
2859		sector_nr += len>>9;
2860		sync_blocks -= (len>>9);
2861	} while (++page_idx < RESYNC_PAGES);
2862
2863	r1_bio->sectors = nr_sectors;
2864
2865	if (mddev_is_clustered(mddev) &&
2866			conf->cluster_sync_high < sector_nr + nr_sectors) {
2867		conf->cluster_sync_low = mddev->curr_resync_completed;
2868		conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2869		/* Send resync message */
2870		md_cluster_ops->resync_info_update(mddev,
2871				conf->cluster_sync_low,
2872				conf->cluster_sync_high);
2873	}
2874
2875	/* For a user-requested sync, we read all readable devices and do a
2876	 * compare
2877	 */
2878	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2879		atomic_set(&r1_bio->remaining, read_targets);
2880		for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2881			bio = r1_bio->bios[i];
2882			if (bio->bi_end_io == end_sync_read) {
2883				read_targets--;
2884				md_sync_acct_bio(bio, nr_sectors);
2885				if (read_targets == 1)
2886					bio->bi_opf &= ~MD_FAILFAST;
2887				generic_make_request(bio);
2888			}
2889		}
2890	} else {
2891		atomic_set(&r1_bio->remaining, 1);
2892		bio = r1_bio->bios[r1_bio->read_disk];
2893		md_sync_acct_bio(bio, nr_sectors);
2894		if (read_targets == 1)
2895			bio->bi_opf &= ~MD_FAILFAST;
2896		generic_make_request(bio);
2897
2898	}
2899	return nr_sectors;
2900}
2901
2902static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2903{
2904	if (sectors)
2905		return sectors;
2906
2907	return mddev->dev_sectors;
2908}
2909
2910static struct r1conf *setup_conf(struct mddev *mddev)
2911{
2912	struct r1conf *conf;
2913	int i;
2914	struct raid1_info *disk;
2915	struct md_rdev *rdev;
2916	int err = -ENOMEM;
2917
2918	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2919	if (!conf)
2920		goto abort;
2921
2922	conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2923				   sizeof(atomic_t), GFP_KERNEL);
2924	if (!conf->nr_pending)
2925		goto abort;
2926
2927	conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2928				   sizeof(atomic_t), GFP_KERNEL);
2929	if (!conf->nr_waiting)
2930		goto abort;
2931
2932	conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2933				  sizeof(atomic_t), GFP_KERNEL);
2934	if (!conf->nr_queued)
2935		goto abort;
2936
2937	conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2938				sizeof(atomic_t), GFP_KERNEL);
2939	if (!conf->barrier)
2940		goto abort;
2941
2942	conf->mirrors = kzalloc(sizeof(struct raid1_info)
2943				* mddev->raid_disks * 2,
2944				 GFP_KERNEL);
2945	if (!conf->mirrors)
2946		goto abort;
2947
2948	conf->tmppage = alloc_page(GFP_KERNEL);
2949	if (!conf->tmppage)
2950		goto abort;
2951
2952	conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2953	if (!conf->poolinfo)
2954		goto abort;
2955	conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2956	conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2957					  r1bio_pool_free,
2958					  conf->poolinfo);
2959	if (!conf->r1bio_pool)
2960		goto abort;
2961
2962	conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2963	if (!conf->bio_split)
2964		goto abort;
2965
2966	conf->poolinfo->mddev = mddev;
2967
2968	err = -EINVAL;
2969	spin_lock_init(&conf->device_lock);
2970	rdev_for_each(rdev, mddev) {
2971		int disk_idx = rdev->raid_disk;
2972		if (disk_idx >= mddev->raid_disks
2973		    || disk_idx < 0)
2974			continue;
2975		if (test_bit(Replacement, &rdev->flags))
2976			disk = conf->mirrors + mddev->raid_disks + disk_idx;
2977		else
2978			disk = conf->mirrors + disk_idx;
2979
2980		if (disk->rdev)
2981			goto abort;
2982		disk->rdev = rdev;
2983		disk->head_position = 0;
2984		disk->seq_start = MaxSector;
2985	}
2986	conf->raid_disks = mddev->raid_disks;
2987	conf->mddev = mddev;
2988	INIT_LIST_HEAD(&conf->retry_list);
2989	INIT_LIST_HEAD(&conf->bio_end_io_list);
2990
2991	spin_lock_init(&conf->resync_lock);
2992	init_waitqueue_head(&conf->wait_barrier);
2993
2994	bio_list_init(&conf->pending_bio_list);
2995	conf->pending_count = 0;
2996	conf->recovery_disabled = mddev->recovery_disabled - 1;
2997
2998	err = -EIO;
2999	for (i = 0; i < conf->raid_disks * 2; i++) {
3000
3001		disk = conf->mirrors + i;
3002
3003		if (i < conf->raid_disks &&
3004		    disk[conf->raid_disks].rdev) {
3005			/* This slot has a replacement. */
3006			if (!disk->rdev) {
3007				/* No original, just make the replacement
3008				 * a recovering spare
3009				 */
3010				disk->rdev =
3011					disk[conf->raid_disks].rdev;
3012				disk[conf->raid_disks].rdev = NULL;
3013			} else if (!test_bit(In_sync, &disk->rdev->flags))
3014				/* Original is not in_sync - bad */
3015				goto abort;
3016		}
3017
3018		if (!disk->rdev ||
3019		    !test_bit(In_sync, &disk->rdev->flags)) {
3020			disk->head_position = 0;
3021			if (disk->rdev &&
3022			    (disk->rdev->saved_raid_disk < 0))
3023				conf->fullsync = 1;
3024		}
3025	}
3026
3027	err = -ENOMEM;
3028	conf->thread = md_register_thread(raid1d, mddev, "raid1");
3029	if (!conf->thread)
3030		goto abort;
3031
3032	return conf;
3033
3034 abort:
3035	if (conf) {
3036		mempool_destroy(conf->r1bio_pool);
3037		kfree(conf->mirrors);
3038		safe_put_page(conf->tmppage);
3039		kfree(conf->poolinfo);
3040		kfree(conf->nr_pending);
3041		kfree(conf->nr_waiting);
3042		kfree(conf->nr_queued);
3043		kfree(conf->barrier);
3044		if (conf->bio_split)
3045			bioset_free(conf->bio_split);
3046		kfree(conf);
3047	}
3048	return ERR_PTR(err);
3049}
3050
3051static void raid1_free(struct mddev *mddev, void *priv);
3052static int raid1_run(struct mddev *mddev)
3053{
3054	struct r1conf *conf;
3055	int i;
3056	struct md_rdev *rdev;
3057	int ret;
3058	bool discard_supported = false;
3059
3060	if (mddev->level != 1) {
3061		pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3062			mdname(mddev), mddev->level);
3063		return -EIO;
3064	}
3065	if (mddev->reshape_position != MaxSector) {
3066		pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3067			mdname(mddev));
3068		return -EIO;
3069	}
3070	if (mddev_init_writes_pending(mddev) < 0)
3071		return -ENOMEM;
3072	/*
3073	 * copy the already verified devices into our private RAID1
3074	 * bookkeeping area. [whatever we allocate in run(),
3075	 * should be freed in raid1_free()]
3076	 */
3077	if (mddev->private == NULL)
3078		conf = setup_conf(mddev);
3079	else
3080		conf = mddev->private;
3081
3082	if (IS_ERR(conf))
3083		return PTR_ERR(conf);
3084
3085	if (mddev->queue) {
3086		blk_queue_max_write_same_sectors(mddev->queue, 0);
3087		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3088	}
3089
3090	rdev_for_each(rdev, mddev) {
3091		if (!mddev->gendisk)
3092			continue;
3093		disk_stack_limits(mddev->gendisk, rdev->bdev,
3094				  rdev->data_offset << 9);
3095		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3096			discard_supported = true;
3097	}
3098
3099	mddev->degraded = 0;
3100	for (i=0; i < conf->raid_disks; i++)
3101		if (conf->mirrors[i].rdev == NULL ||
3102		    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3103		    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3104			mddev->degraded++;
 
 
 
 
 
 
 
3105
3106	if (conf->raid_disks - mddev->degraded == 1)
3107		mddev->recovery_cp = MaxSector;
3108
3109	if (mddev->recovery_cp != MaxSector)
3110		pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3111			mdname(mddev));
3112	pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3113		mdname(mddev), mddev->raid_disks - mddev->degraded,
3114		mddev->raid_disks);
3115
3116	/*
3117	 * Ok, everything is just fine now
3118	 */
3119	mddev->thread = conf->thread;
3120	conf->thread = NULL;
3121	mddev->private = conf;
3122	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3123
3124	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3125
3126	if (mddev->queue) {
3127		if (discard_supported)
3128			blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3129						mddev->queue);
3130		else
3131			blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3132						  mddev->queue);
3133	}
3134
3135	ret =  md_integrity_register(mddev);
3136	if (ret) {
3137		md_unregister_thread(&mddev->thread);
3138		raid1_free(mddev, conf);
3139	}
 
 
 
 
3140	return ret;
3141}
3142
3143static void raid1_free(struct mddev *mddev, void *priv)
3144{
3145	struct r1conf *conf = priv;
3146
3147	mempool_destroy(conf->r1bio_pool);
3148	kfree(conf->mirrors);
3149	safe_put_page(conf->tmppage);
3150	kfree(conf->poolinfo);
3151	kfree(conf->nr_pending);
3152	kfree(conf->nr_waiting);
3153	kfree(conf->nr_queued);
3154	kfree(conf->barrier);
3155	if (conf->bio_split)
3156		bioset_free(conf->bio_split);
3157	kfree(conf);
3158}
3159
3160static int raid1_resize(struct mddev *mddev, sector_t sectors)
3161{
3162	/* no resync is happening, and there is enough space
3163	 * on all devices, so we can resize.
3164	 * We need to make sure resync covers any new space.
3165	 * If the array is shrinking we should possibly wait until
3166	 * any io in the removed space completes, but it hardly seems
3167	 * worth it.
3168	 */
3169	sector_t newsize = raid1_size(mddev, sectors, 0);
3170	if (mddev->external_size &&
3171	    mddev->array_sectors > newsize)
3172		return -EINVAL;
3173	if (mddev->bitmap) {
3174		int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3175		if (ret)
3176			return ret;
3177	}
3178	md_set_array_sectors(mddev, newsize);
3179	if (sectors > mddev->dev_sectors &&
3180	    mddev->recovery_cp > mddev->dev_sectors) {
3181		mddev->recovery_cp = mddev->dev_sectors;
3182		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3183	}
3184	mddev->dev_sectors = sectors;
3185	mddev->resync_max_sectors = sectors;
3186	return 0;
3187}
3188
3189static int raid1_reshape(struct mddev *mddev)
3190{
3191	/* We need to:
3192	 * 1/ resize the r1bio_pool
3193	 * 2/ resize conf->mirrors
3194	 *
3195	 * We allocate a new r1bio_pool if we can.
3196	 * Then raise a device barrier and wait until all IO stops.
3197	 * Then resize conf->mirrors and swap in the new r1bio pool.
3198	 *
3199	 * At the same time, we "pack" the devices so that all the missing
3200	 * devices have the higher raid_disk numbers.
3201	 */
3202	mempool_t *newpool, *oldpool;
3203	struct pool_info *newpoolinfo;
3204	struct raid1_info *newmirrors;
3205	struct r1conf *conf = mddev->private;
3206	int cnt, raid_disks;
3207	unsigned long flags;
3208	int d, d2;
 
 
 
 
3209
3210	/* Cannot change chunk_size, layout, or level */
3211	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3212	    mddev->layout != mddev->new_layout ||
3213	    mddev->level != mddev->new_level) {
3214		mddev->new_chunk_sectors = mddev->chunk_sectors;
3215		mddev->new_layout = mddev->layout;
3216		mddev->new_level = mddev->level;
3217		return -EINVAL;
3218	}
3219
3220	if (!mddev_is_clustered(mddev))
3221		md_allow_write(mddev);
3222
3223	raid_disks = mddev->raid_disks + mddev->delta_disks;
3224
3225	if (raid_disks < conf->raid_disks) {
3226		cnt=0;
3227		for (d= 0; d < conf->raid_disks; d++)
3228			if (conf->mirrors[d].rdev)
3229				cnt++;
3230		if (cnt > raid_disks)
3231			return -EBUSY;
3232	}
3233
3234	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3235	if (!newpoolinfo)
3236		return -ENOMEM;
3237	newpoolinfo->mddev = mddev;
3238	newpoolinfo->raid_disks = raid_disks * 2;
3239
3240	newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3241				 r1bio_pool_free, newpoolinfo);
3242	if (!newpool) {
3243		kfree(newpoolinfo);
3244		return -ENOMEM;
3245	}
3246	newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
 
3247			     GFP_KERNEL);
3248	if (!newmirrors) {
3249		kfree(newpoolinfo);
3250		mempool_destroy(newpool);
3251		return -ENOMEM;
3252	}
3253
3254	freeze_array(conf, 0);
3255
3256	/* ok, everything is stopped */
3257	oldpool = conf->r1bio_pool;
3258	conf->r1bio_pool = newpool;
3259
3260	for (d = d2 = 0; d < conf->raid_disks; d++) {
3261		struct md_rdev *rdev = conf->mirrors[d].rdev;
3262		if (rdev && rdev->raid_disk != d2) {
3263			sysfs_unlink_rdev(mddev, rdev);
3264			rdev->raid_disk = d2;
3265			sysfs_unlink_rdev(mddev, rdev);
3266			if (sysfs_link_rdev(mddev, rdev))
3267				pr_warn("md/raid1:%s: cannot register rd%d\n",
3268					mdname(mddev), rdev->raid_disk);
3269		}
3270		if (rdev)
3271			newmirrors[d2++].rdev = rdev;
3272	}
3273	kfree(conf->mirrors);
3274	conf->mirrors = newmirrors;
3275	kfree(conf->poolinfo);
3276	conf->poolinfo = newpoolinfo;
3277
3278	spin_lock_irqsave(&conf->device_lock, flags);
3279	mddev->degraded += (raid_disks - conf->raid_disks);
3280	spin_unlock_irqrestore(&conf->device_lock, flags);
3281	conf->raid_disks = mddev->raid_disks = raid_disks;
3282	mddev->delta_disks = 0;
3283
3284	unfreeze_array(conf);
3285
3286	set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3287	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3288	md_wakeup_thread(mddev->thread);
3289
3290	mempool_destroy(oldpool);
3291	return 0;
3292}
3293
3294static void raid1_quiesce(struct mddev *mddev, int quiesce)
3295{
3296	struct r1conf *conf = mddev->private;
3297
3298	if (quiesce)
3299		freeze_array(conf, 0);
3300	else
3301		unfreeze_array(conf);
3302}
3303
3304static void *raid1_takeover(struct mddev *mddev)
3305{
3306	/* raid1 can take over:
3307	 *  raid5 with 2 devices, any layout or chunk size
3308	 */
3309	if (mddev->level == 5 && mddev->raid_disks == 2) {
3310		struct r1conf *conf;
3311		mddev->new_level = 1;
3312		mddev->new_layout = 0;
3313		mddev->new_chunk_sectors = 0;
3314		conf = setup_conf(mddev);
3315		if (!IS_ERR(conf)) {
3316			/* Array must appear to be quiesced */
3317			conf->array_frozen = 1;
3318			mddev_clear_unsupported_flags(mddev,
3319				UNSUPPORTED_MDDEV_FLAGS);
3320		}
3321		return conf;
3322	}
3323	return ERR_PTR(-EINVAL);
3324}
3325
3326static struct md_personality raid1_personality =
3327{
3328	.name		= "raid1",
3329	.level		= 1,
3330	.owner		= THIS_MODULE,
3331	.make_request	= raid1_make_request,
3332	.run		= raid1_run,
3333	.free		= raid1_free,
3334	.status		= raid1_status,
3335	.error_handler	= raid1_error,
3336	.hot_add_disk	= raid1_add_disk,
3337	.hot_remove_disk= raid1_remove_disk,
3338	.spare_active	= raid1_spare_active,
3339	.sync_request	= raid1_sync_request,
3340	.resize		= raid1_resize,
3341	.size		= raid1_size,
3342	.check_reshape	= raid1_reshape,
3343	.quiesce	= raid1_quiesce,
3344	.takeover	= raid1_takeover,
3345	.congested	= raid1_congested,
3346};
3347
3348static int __init raid_init(void)
3349{
3350	return register_md_personality(&raid1_personality);
3351}
3352
3353static void raid_exit(void)
3354{
3355	unregister_md_personality(&raid1_personality);
3356}
3357
3358module_init(raid_init);
3359module_exit(raid_exit);
3360MODULE_LICENSE("GPL");
3361MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3362MODULE_ALIAS("md-personality-3"); /* RAID1 */
3363MODULE_ALIAS("md-raid1");
3364MODULE_ALIAS("md-level-1");
3365
3366module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);