Linux Audio

Check our new training course

Loading...
v6.8
   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#include <linux/interval_tree_generic.h>
  33
  34#include <trace/events/block.h>
  35
  36#include "md.h"
  37#include "raid1.h"
  38#include "md-bitmap.h"
  39
  40#define UNSUPPORTED_MDDEV_FLAGS		\
  41	((1L << MD_HAS_JOURNAL) |	\
  42	 (1L << MD_JOURNAL_CLEAN) |	\
  43	 (1L << MD_HAS_PPL) |		\
  44	 (1L << MD_HAS_MULTIPLE_PPLS))
  45
  46static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  47static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  48
  49#define raid1_log(md, fmt, args...)				\
  50	do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  51
  52#define RAID_1_10_NAME "raid1"
  53#include "raid1-10.c"
  54
  55#define START(node) ((node)->start)
  56#define LAST(node) ((node)->last)
  57INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  58		     START, LAST, static inline, raid1_rb);
  59
  60static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  61				struct serial_info *si, int idx)
  62{
 
  63	unsigned long flags;
  64	int ret = 0;
  65	sector_t lo = r1_bio->sector;
  66	sector_t hi = lo + r1_bio->sectors;
  67	struct serial_in_rdev *serial = &rdev->serial[idx];
  68
  69	spin_lock_irqsave(&serial->serial_lock, flags);
  70	/* collision happened */
  71	if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  72		ret = -EBUSY;
  73	else {
  74		si->start = lo;
  75		si->last = hi;
  76		raid1_rb_insert(si, &serial->serial_rb);
  77	}
  78	spin_unlock_irqrestore(&serial->serial_lock, flags);
  79
  80	return ret;
  81}
  82
  83static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  84{
  85	struct mddev *mddev = rdev->mddev;
  86	struct serial_info *si;
  87	int idx = sector_to_idx(r1_bio->sector);
  88	struct serial_in_rdev *serial = &rdev->serial[idx];
 
 
  89
  90	if (WARN_ON(!mddev->serial_info_pool))
  91		return;
  92	si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  93	wait_event(serial->serial_io_wait,
  94		   check_and_add_serial(rdev, r1_bio, si, idx) == 0);
 
 
 
 
  95}
  96
  97static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  98{
  99	struct serial_info *si;
 100	unsigned long flags;
 101	int found = 0;
 102	struct mddev *mddev = rdev->mddev;
 103	int idx = sector_to_idx(lo);
 104	struct serial_in_rdev *serial = &rdev->serial[idx];
 105
 106	spin_lock_irqsave(&serial->serial_lock, flags);
 107	for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 108	     si; si = raid1_rb_iter_next(si, lo, hi)) {
 109		if (si->start == lo && si->last == hi) {
 110			raid1_rb_remove(si, &serial->serial_rb);
 111			mempool_free(si, mddev->serial_info_pool);
 112			found = 1;
 113			break;
 114		}
 115	}
 116	if (!found)
 117		WARN(1, "The write IO is not recorded for serialization\n");
 118	spin_unlock_irqrestore(&serial->serial_lock, flags);
 119	wake_up(&serial->serial_io_wait);
 120}
 121
 122/*
 123 * for resync bio, r1bio pointer can be retrieved from the per-bio
 124 * 'struct resync_pages'.
 125 */
 126static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 127{
 128	return get_resync_pages(bio)->raid_bio;
 129}
 130
 131static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 132{
 133	struct pool_info *pi = data;
 134	int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 135
 136	/* allocate a r1bio with room for raid_disks entries in the bios array */
 137	return kzalloc(size, gfp_flags);
 138}
 139
 140#define RESYNC_DEPTH 32
 141#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 142#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 143#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 144#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 145#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 146
 147static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 148{
 149	struct pool_info *pi = data;
 150	struct r1bio *r1_bio;
 151	struct bio *bio;
 152	int need_pages;
 153	int j;
 154	struct resync_pages *rps;
 155
 156	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 157	if (!r1_bio)
 158		return NULL;
 159
 160	rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 161			    gfp_flags);
 162	if (!rps)
 163		goto out_free_r1bio;
 164
 165	/*
 166	 * Allocate bios : 1 for reading, n-1 for writing
 167	 */
 168	for (j = pi->raid_disks ; j-- ; ) {
 169		bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
 170		if (!bio)
 171			goto out_free_bio;
 172		bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
 173		r1_bio->bios[j] = bio;
 174	}
 175	/*
 176	 * Allocate RESYNC_PAGES data pages and attach them to
 177	 * the first bio.
 178	 * If this is a user-requested check/repair, allocate
 179	 * RESYNC_PAGES for each bio.
 180	 */
 181	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 182		need_pages = pi->raid_disks;
 183	else
 184		need_pages = 1;
 185	for (j = 0; j < pi->raid_disks; j++) {
 186		struct resync_pages *rp = &rps[j];
 187
 188		bio = r1_bio->bios[j];
 189
 190		if (j < need_pages) {
 191			if (resync_alloc_pages(rp, gfp_flags))
 192				goto out_free_pages;
 193		} else {
 194			memcpy(rp, &rps[0], sizeof(*rp));
 195			resync_get_all_pages(rp);
 196		}
 197
 198		rp->raid_bio = r1_bio;
 199		bio->bi_private = rp;
 200	}
 201
 202	r1_bio->master_bio = NULL;
 203
 204	return r1_bio;
 205
 206out_free_pages:
 207	while (--j >= 0)
 208		resync_free_pages(&rps[j]);
 209
 210out_free_bio:
 211	while (++j < pi->raid_disks) {
 212		bio_uninit(r1_bio->bios[j]);
 213		kfree(r1_bio->bios[j]);
 214	}
 215	kfree(rps);
 216
 217out_free_r1bio:
 218	rbio_pool_free(r1_bio, data);
 219	return NULL;
 220}
 221
 222static void r1buf_pool_free(void *__r1_bio, void *data)
 223{
 224	struct pool_info *pi = data;
 225	int i;
 226	struct r1bio *r1bio = __r1_bio;
 227	struct resync_pages *rp = NULL;
 228
 229	for (i = pi->raid_disks; i--; ) {
 230		rp = get_resync_pages(r1bio->bios[i]);
 231		resync_free_pages(rp);
 232		bio_uninit(r1bio->bios[i]);
 233		kfree(r1bio->bios[i]);
 234	}
 235
 236	/* resync pages array stored in the 1st bio's .bi_private */
 237	kfree(rp);
 238
 239	rbio_pool_free(r1bio, data);
 240}
 241
 242static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 243{
 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_SPECIAL(*bio))
 249			bio_put(*bio);
 250		*bio = NULL;
 251	}
 252}
 253
 254static void free_r1bio(struct r1bio *r1_bio)
 255{
 256	struct r1conf *conf = r1_bio->mddev->private;
 257
 258	put_all_bios(conf, r1_bio);
 259	mempool_free(r1_bio, &conf->r1bio_pool);
 260}
 261
 262static void put_buf(struct r1bio *r1_bio)
 263{
 264	struct r1conf *conf = r1_bio->mddev->private;
 265	sector_t sect = r1_bio->sector;
 266	int i;
 267
 268	for (i = 0; i < conf->raid_disks * 2; i++) {
 269		struct bio *bio = r1_bio->bios[i];
 270		if (bio->bi_end_io)
 271			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 272	}
 273
 274	mempool_free(r1_bio, &conf->r1buf_pool);
 275
 276	lower_barrier(conf, sect);
 277}
 278
 279static void reschedule_retry(struct r1bio *r1_bio)
 280{
 281	unsigned long flags;
 282	struct mddev *mddev = r1_bio->mddev;
 283	struct r1conf *conf = mddev->private;
 284	int idx;
 285
 286	idx = sector_to_idx(r1_bio->sector);
 287	spin_lock_irqsave(&conf->device_lock, flags);
 288	list_add(&r1_bio->retry_list, &conf->retry_list);
 289	atomic_inc(&conf->nr_queued[idx]);
 290	spin_unlock_irqrestore(&conf->device_lock, flags);
 291
 292	wake_up(&conf->wait_barrier);
 293	md_wakeup_thread(mddev->thread);
 294}
 295
 296/*
 297 * raid_end_bio_io() is called when we have finished servicing a mirrored
 298 * operation and are ready to return a success/failure code to the buffer
 299 * cache layer.
 300 */
 301static void call_bio_endio(struct r1bio *r1_bio)
 302{
 303	struct bio *bio = r1_bio->master_bio;
 
 304
 305	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 306		bio->bi_status = BLK_STS_IOERR;
 307
 308	bio_endio(bio);
 
 
 
 
 
 309}
 310
 311static void raid_end_bio_io(struct r1bio *r1_bio)
 312{
 313	struct bio *bio = r1_bio->master_bio;
 314	struct r1conf *conf = r1_bio->mddev->private;
 315	sector_t sector = r1_bio->sector;
 316
 317	/* if nobody has done the final endio yet, do it now */
 318	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 319		pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 320			 (bio_data_dir(bio) == WRITE) ? "write" : "read",
 321			 (unsigned long long) bio->bi_iter.bi_sector,
 322			 (unsigned long long) bio_end_sector(bio) - 1);
 323
 324		call_bio_endio(r1_bio);
 325	}
 326
 327	free_r1bio(r1_bio);
 328	/*
 329	 * Wake up any possible resync thread that waits for the device
 330	 * to go idle.  All I/Os, even write-behind writes, are done.
 331	 */
 332	allow_barrier(conf, sector);
 333}
 334
 335/*
 336 * Update disk head position estimator based on IRQ completion info.
 337 */
 338static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 339{
 340	struct r1conf *conf = r1_bio->mddev->private;
 341
 342	conf->mirrors[disk].head_position =
 343		r1_bio->sector + (r1_bio->sectors);
 344}
 345
 346/*
 347 * Find the disk number which triggered given bio
 348 */
 349static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 350{
 351	int mirror;
 352	struct r1conf *conf = r1_bio->mddev->private;
 353	int raid_disks = conf->raid_disks;
 354
 355	for (mirror = 0; mirror < raid_disks * 2; mirror++)
 356		if (r1_bio->bios[mirror] == bio)
 357			break;
 358
 359	BUG_ON(mirror == raid_disks * 2);
 360	update_head_pos(mirror, r1_bio);
 361
 362	return mirror;
 363}
 364
 365static void raid1_end_read_request(struct bio *bio)
 366{
 367	int uptodate = !bio->bi_status;
 368	struct r1bio *r1_bio = bio->bi_private;
 369	struct r1conf *conf = r1_bio->mddev->private;
 370	struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 371
 372	/*
 373	 * this branch is our 'one mirror IO has finished' event handler:
 374	 */
 375	update_head_pos(r1_bio->read_disk, r1_bio);
 376
 377	if (uptodate)
 378		set_bit(R1BIO_Uptodate, &r1_bio->state);
 379	else if (test_bit(FailFast, &rdev->flags) &&
 380		 test_bit(R1BIO_FailFast, &r1_bio->state))
 381		/* This was a fail-fast read so we definitely
 382		 * want to retry */
 383		;
 384	else {
 385		/* If all other devices have failed, we want to return
 386		 * the error upwards rather than fail the last device.
 387		 * Here we redefine "uptodate" to mean "Don't want to retry"
 388		 */
 389		unsigned long flags;
 390		spin_lock_irqsave(&conf->device_lock, flags);
 391		if (r1_bio->mddev->degraded == conf->raid_disks ||
 392		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 393		     test_bit(In_sync, &rdev->flags)))
 394			uptodate = 1;
 395		spin_unlock_irqrestore(&conf->device_lock, flags);
 396	}
 397
 398	if (uptodate) {
 399		raid_end_bio_io(r1_bio);
 400		rdev_dec_pending(rdev, conf->mddev);
 401	} else {
 402		/*
 403		 * oops, read error:
 404		 */
 405		pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
 
 406				   mdname(conf->mddev),
 407				   rdev->bdev,
 408				   (unsigned long long)r1_bio->sector);
 409		set_bit(R1BIO_ReadError, &r1_bio->state);
 410		reschedule_retry(r1_bio);
 411		/* don't drop the reference on read_disk yet */
 412	}
 413}
 414
 415static void close_write(struct r1bio *r1_bio)
 416{
 417	/* it really is the end of this request */
 418	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 419		bio_free_pages(r1_bio->behind_master_bio);
 420		bio_put(r1_bio->behind_master_bio);
 421		r1_bio->behind_master_bio = NULL;
 422	}
 423	/* clear the bitmap if all writes complete successfully */
 424	md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 425			   r1_bio->sectors,
 426			   !test_bit(R1BIO_Degraded, &r1_bio->state),
 427			   test_bit(R1BIO_BehindIO, &r1_bio->state));
 428	md_write_end(r1_bio->mddev);
 429}
 430
 431static void r1_bio_write_done(struct r1bio *r1_bio)
 432{
 433	if (!atomic_dec_and_test(&r1_bio->remaining))
 434		return;
 435
 436	if (test_bit(R1BIO_WriteError, &r1_bio->state))
 437		reschedule_retry(r1_bio);
 438	else {
 439		close_write(r1_bio);
 440		if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 441			reschedule_retry(r1_bio);
 442		else
 443			raid_end_bio_io(r1_bio);
 444	}
 445}
 446
 447static void raid1_end_write_request(struct bio *bio)
 448{
 449	struct r1bio *r1_bio = bio->bi_private;
 450	int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 451	struct r1conf *conf = r1_bio->mddev->private;
 452	struct bio *to_put = NULL;
 453	int mirror = find_bio_disk(r1_bio, bio);
 454	struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 455	bool discard_error;
 456	sector_t lo = r1_bio->sector;
 457	sector_t hi = r1_bio->sector + r1_bio->sectors;
 458
 459	discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 460
 461	/*
 462	 * 'one mirror IO has finished' event handler:
 463	 */
 464	if (bio->bi_status && !discard_error) {
 465		set_bit(WriteErrorSeen,	&rdev->flags);
 466		if (!test_and_set_bit(WantReplacement, &rdev->flags))
 467			set_bit(MD_RECOVERY_NEEDED, &
 468				conf->mddev->recovery);
 469
 470		if (test_bit(FailFast, &rdev->flags) &&
 471		    (bio->bi_opf & MD_FAILFAST) &&
 472		    /* We never try FailFast to WriteMostly devices */
 473		    !test_bit(WriteMostly, &rdev->flags)) {
 474			md_error(r1_bio->mddev, rdev);
 475		}
 476
 477		/*
 478		 * When the device is faulty, it is not necessary to
 479		 * handle write error.
 
 
 480		 */
 481		if (!test_bit(Faulty, &rdev->flags))
 482			set_bit(R1BIO_WriteError, &r1_bio->state);
 483		else {
 484			/* Fail the request */
 485			set_bit(R1BIO_Degraded, &r1_bio->state);
 486			/* Finished with this branch */
 487			r1_bio->bios[mirror] = NULL;
 488			to_put = bio;
 489		}
 490	} else {
 491		/*
 492		 * Set R1BIO_Uptodate in our master bio, so that we
 493		 * will return a good error code for to the higher
 494		 * levels even if IO on some other mirrored buffer
 495		 * fails.
 496		 *
 497		 * The 'master' represents the composite IO operation
 498		 * to user-side. So if something waits for IO, then it
 499		 * will wait for the 'master' bio.
 500		 */
 501		sector_t first_bad;
 502		int bad_sectors;
 503
 504		r1_bio->bios[mirror] = NULL;
 505		to_put = bio;
 506		/*
 507		 * Do not set R1BIO_Uptodate if the current device is
 508		 * rebuilding or Faulty. This is because we cannot use
 509		 * such device for properly reading the data back (we could
 510		 * potentially use it, if the current write would have felt
 511		 * before rdev->recovery_offset, but for simplicity we don't
 512		 * check this here.
 513		 */
 514		if (test_bit(In_sync, &rdev->flags) &&
 515		    !test_bit(Faulty, &rdev->flags))
 516			set_bit(R1BIO_Uptodate, &r1_bio->state);
 517
 518		/* Maybe we can clear some bad blocks. */
 519		if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 520				&first_bad, &bad_sectors) && !discard_error) {
 521			r1_bio->bios[mirror] = IO_MADE_GOOD;
 522			set_bit(R1BIO_MadeGood, &r1_bio->state);
 523		}
 524	}
 525
 526	if (behind) {
 527		if (test_bit(CollisionCheck, &rdev->flags))
 528			remove_serial(rdev, lo, hi);
 
 
 
 
 529		if (test_bit(WriteMostly, &rdev->flags))
 530			atomic_dec(&r1_bio->behind_remaining);
 531
 532		/*
 533		 * In behind mode, we ACK the master bio once the I/O
 534		 * has safely reached all non-writemostly
 535		 * disks. Setting the Returned bit ensures that this
 536		 * gets done only once -- we don't ever want to return
 537		 * -EIO here, instead we'll wait
 538		 */
 539		if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 540		    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 541			/* Maybe we can return now */
 542			if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 543				struct bio *mbio = r1_bio->master_bio;
 544				pr_debug("raid1: behind end write sectors"
 545					 " %llu-%llu\n",
 546					 (unsigned long long) mbio->bi_iter.bi_sector,
 547					 (unsigned long long) bio_end_sector(mbio) - 1);
 548				call_bio_endio(r1_bio);
 549			}
 550		}
 551	} else if (rdev->mddev->serialize_policy)
 552		remove_serial(rdev, lo, hi);
 553	if (r1_bio->bios[mirror] == NULL)
 554		rdev_dec_pending(rdev, conf->mddev);
 555
 556	/*
 557	 * Let's see if all mirrored write operations have finished
 558	 * already.
 559	 */
 560	r1_bio_write_done(r1_bio);
 561
 562	if (to_put)
 563		bio_put(to_put);
 564}
 565
 566static sector_t align_to_barrier_unit_end(sector_t start_sector,
 567					  sector_t sectors)
 568{
 569	sector_t len;
 570
 571	WARN_ON(sectors == 0);
 572	/*
 573	 * len is the number of sectors from start_sector to end of the
 574	 * barrier unit which start_sector belongs to.
 575	 */
 576	len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 577	      start_sector;
 578
 579	if (len > sectors)
 580		len = sectors;
 581
 582	return len;
 583}
 584
 585/*
 586 * This routine returns the disk from which the requested read should
 587 * be done. There is a per-array 'next expected sequential IO' sector
 588 * number - if this matches on the next IO then we use the last disk.
 589 * There is also a per-disk 'last know head position' sector that is
 590 * maintained from IRQ contexts, both the normal and the resync IO
 591 * completion handlers update this position correctly. If there is no
 592 * perfect sequential match then we pick the disk whose head is closest.
 593 *
 594 * If there are 2 mirrors in the same 2 devices, performance degrades
 595 * because position is mirror, not device based.
 596 *
 597 * The rdev for the device selected will have nr_pending incremented.
 598 */
 599static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 600{
 601	const sector_t this_sector = r1_bio->sector;
 602	int sectors;
 603	int best_good_sectors;
 604	int best_disk, best_dist_disk, best_pending_disk;
 605	int has_nonrot_disk;
 606	int disk;
 607	sector_t best_dist;
 608	unsigned int min_pending;
 609	struct md_rdev *rdev;
 610	int choose_first;
 611	int choose_next_idle;
 612
 
 613	/*
 614	 * Check if we can balance. We can balance on the whole
 615	 * device if no resync is going on, or below the resync window.
 616	 * We take the first readable disk when above the resync window.
 617	 */
 618 retry:
 619	sectors = r1_bio->sectors;
 620	best_disk = -1;
 621	best_dist_disk = -1;
 622	best_dist = MaxSector;
 623	best_pending_disk = -1;
 624	min_pending = UINT_MAX;
 625	best_good_sectors = 0;
 626	has_nonrot_disk = 0;
 627	choose_next_idle = 0;
 628	clear_bit(R1BIO_FailFast, &r1_bio->state);
 629
 630	if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 631	    (mddev_is_clustered(conf->mddev) &&
 632	    md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 633		    this_sector + sectors)))
 634		choose_first = 1;
 635	else
 636		choose_first = 0;
 637
 638	for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 639		sector_t dist;
 640		sector_t first_bad;
 641		int bad_sectors;
 642		unsigned int pending;
 643		bool nonrot;
 644
 645		rdev = conf->mirrors[disk].rdev;
 646		if (r1_bio->bios[disk] == IO_BLOCKED
 647		    || rdev == NULL
 648		    || test_bit(Faulty, &rdev->flags))
 649			continue;
 650		if (!test_bit(In_sync, &rdev->flags) &&
 651		    rdev->recovery_offset < this_sector + sectors)
 652			continue;
 653		if (test_bit(WriteMostly, &rdev->flags)) {
 654			/* Don't balance among write-mostly, just
 655			 * use the first as a last resort */
 656			if (best_dist_disk < 0) {
 657				if (is_badblock(rdev, this_sector, sectors,
 658						&first_bad, &bad_sectors)) {
 659					if (first_bad <= this_sector)
 660						/* Cannot use this */
 661						continue;
 662					best_good_sectors = first_bad - this_sector;
 663				} else
 664					best_good_sectors = sectors;
 665				best_dist_disk = disk;
 666				best_pending_disk = disk;
 667			}
 668			continue;
 669		}
 670		/* This is a reasonable device to use.  It might
 671		 * even be best.
 672		 */
 673		if (is_badblock(rdev, this_sector, sectors,
 674				&first_bad, &bad_sectors)) {
 675			if (best_dist < MaxSector)
 676				/* already have a better device */
 677				continue;
 678			if (first_bad <= this_sector) {
 679				/* cannot read here. If this is the 'primary'
 680				 * device, then we must not read beyond
 681				 * bad_sectors from another device..
 682				 */
 683				bad_sectors -= (this_sector - first_bad);
 684				if (choose_first && sectors > bad_sectors)
 685					sectors = bad_sectors;
 686				if (best_good_sectors > sectors)
 687					best_good_sectors = sectors;
 688
 689			} else {
 690				sector_t good_sectors = first_bad - this_sector;
 691				if (good_sectors > best_good_sectors) {
 692					best_good_sectors = good_sectors;
 693					best_disk = disk;
 694				}
 695				if (choose_first)
 696					break;
 697			}
 698			continue;
 699		} else {
 700			if ((sectors > best_good_sectors) && (best_disk >= 0))
 701				best_disk = -1;
 702			best_good_sectors = sectors;
 703		}
 704
 705		if (best_disk >= 0)
 706			/* At least two disks to choose from so failfast is OK */
 707			set_bit(R1BIO_FailFast, &r1_bio->state);
 708
 709		nonrot = bdev_nonrot(rdev->bdev);
 710		has_nonrot_disk |= nonrot;
 711		pending = atomic_read(&rdev->nr_pending);
 712		dist = abs(this_sector - conf->mirrors[disk].head_position);
 713		if (choose_first) {
 714			best_disk = disk;
 715			break;
 716		}
 717		/* Don't change to another disk for sequential reads */
 718		if (conf->mirrors[disk].next_seq_sect == this_sector
 719		    || dist == 0) {
 720			int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 721			struct raid1_info *mirror = &conf->mirrors[disk];
 722
 723			best_disk = disk;
 724			/*
 725			 * If buffered sequential IO size exceeds optimal
 726			 * iosize, check if there is idle disk. If yes, choose
 727			 * the idle disk. read_balance could already choose an
 728			 * idle disk before noticing it's a sequential IO in
 729			 * this disk. This doesn't matter because this disk
 730			 * will idle, next time it will be utilized after the
 731			 * first disk has IO size exceeds optimal iosize. In
 732			 * this way, iosize of the first disk will be optimal
 733			 * iosize at least. iosize of the second disk might be
 734			 * small, but not a big deal since when the second disk
 735			 * starts IO, the first disk is likely still busy.
 736			 */
 737			if (nonrot && opt_iosize > 0 &&
 738			    mirror->seq_start != MaxSector &&
 739			    mirror->next_seq_sect > opt_iosize &&
 740			    mirror->next_seq_sect - opt_iosize >=
 741			    mirror->seq_start) {
 742				choose_next_idle = 1;
 743				continue;
 744			}
 745			break;
 746		}
 747
 748		if (choose_next_idle)
 749			continue;
 750
 751		if (min_pending > pending) {
 752			min_pending = pending;
 753			best_pending_disk = disk;
 754		}
 755
 756		if (dist < best_dist) {
 757			best_dist = dist;
 758			best_dist_disk = disk;
 759		}
 760	}
 761
 762	/*
 763	 * If all disks are rotational, choose the closest disk. If any disk is
 764	 * non-rotational, choose the disk with less pending request even the
 765	 * disk is rotational, which might/might not be optimal for raids with
 766	 * mixed ratation/non-rotational disks depending on workload.
 767	 */
 768	if (best_disk == -1) {
 769		if (has_nonrot_disk || min_pending == 0)
 770			best_disk = best_pending_disk;
 771		else
 772			best_disk = best_dist_disk;
 773	}
 774
 775	if (best_disk >= 0) {
 776		rdev = conf->mirrors[best_disk].rdev;
 777		if (!rdev)
 778			goto retry;
 779		atomic_inc(&rdev->nr_pending);
 780		sectors = best_good_sectors;
 781
 782		if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 783			conf->mirrors[best_disk].seq_start = this_sector;
 784
 785		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 786	}
 
 787	*max_sectors = sectors;
 788
 789	return best_disk;
 790}
 791
 792static void wake_up_barrier(struct r1conf *conf)
 793{
 794	if (wq_has_sleeper(&conf->wait_barrier))
 795		wake_up(&conf->wait_barrier);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 796}
 797
 798static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 799{
 800	/* flush any pending bitmap writes to disk before proceeding w/ I/O */
 801	raid1_prepare_flush_writes(conf->mddev->bitmap);
 802	wake_up_barrier(conf);
 803
 804	while (bio) { /* submit pending writes */
 805		struct bio *next = bio->bi_next;
 806
 807		raid1_submit_write(bio);
 
 
 
 
 
 
 
 
 
 808		bio = next;
 809		cond_resched();
 810	}
 811}
 812
 813static void flush_pending_writes(struct r1conf *conf)
 814{
 815	/* Any writes that have been queued but are awaiting
 816	 * bitmap updates get flushed here.
 817	 */
 818	spin_lock_irq(&conf->device_lock);
 819
 820	if (conf->pending_bio_list.head) {
 821		struct blk_plug plug;
 822		struct bio *bio;
 823
 824		bio = bio_list_get(&conf->pending_bio_list);
 
 825		spin_unlock_irq(&conf->device_lock);
 826
 827		/*
 828		 * As this is called in a wait_event() loop (see freeze_array),
 829		 * current->state might be TASK_UNINTERRUPTIBLE which will
 830		 * cause a warning when we prepare to wait again.  As it is
 831		 * rare that this path is taken, it is perfectly safe to force
 832		 * us to go around the wait_event() loop again, so the warning
 833		 * is a false-positive.  Silence the warning by resetting
 834		 * thread state
 835		 */
 836		__set_current_state(TASK_RUNNING);
 837		blk_start_plug(&plug);
 838		flush_bio_list(conf, bio);
 839		blk_finish_plug(&plug);
 840	} else
 841		spin_unlock_irq(&conf->device_lock);
 842}
 843
 844/* Barriers....
 845 * Sometimes we need to suspend IO while we do something else,
 846 * either some resync/recovery, or reconfigure the array.
 847 * To do this we raise a 'barrier'.
 848 * The 'barrier' is a counter that can be raised multiple times
 849 * to count how many activities are happening which preclude
 850 * normal IO.
 851 * We can only raise the barrier if there is no pending IO.
 852 * i.e. if nr_pending == 0.
 853 * We choose only to raise the barrier if no-one is waiting for the
 854 * barrier to go down.  This means that as soon as an IO request
 855 * is ready, no other operations which require a barrier will start
 856 * until the IO request has had a chance.
 857 *
 858 * So: regular IO calls 'wait_barrier'.  When that returns there
 859 *    is no backgroup IO happening,  It must arrange to call
 860 *    allow_barrier when it has finished its IO.
 861 * backgroup IO calls must call raise_barrier.  Once that returns
 862 *    there is no normal IO happeing.  It must arrange to call
 863 *    lower_barrier when the particular background IO completes.
 864 *
 865 * If resync/recovery is interrupted, returns -EINTR;
 866 * Otherwise, returns 0.
 867 */
 868static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 869{
 870	int idx = sector_to_idx(sector_nr);
 871
 872	spin_lock_irq(&conf->resync_lock);
 873
 874	/* Wait until no block IO is waiting */
 875	wait_event_lock_irq(conf->wait_barrier,
 876			    !atomic_read(&conf->nr_waiting[idx]),
 877			    conf->resync_lock);
 878
 879	/* block any new IO from starting */
 880	atomic_inc(&conf->barrier[idx]);
 881	/*
 882	 * In raise_barrier() we firstly increase conf->barrier[idx] then
 883	 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 884	 * increase conf->nr_pending[idx] then check conf->barrier[idx].
 885	 * A memory barrier here to make sure conf->nr_pending[idx] won't
 886	 * be fetched before conf->barrier[idx] is increased. Otherwise
 887	 * there will be a race between raise_barrier() and _wait_barrier().
 888	 */
 889	smp_mb__after_atomic();
 890
 891	/* For these conditions we must wait:
 892	 * A: while the array is in frozen state
 893	 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 894	 *    existing in corresponding I/O barrier bucket.
 895	 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 896	 *    max resync count which allowed on current I/O barrier bucket.
 897	 */
 898	wait_event_lock_irq(conf->wait_barrier,
 899			    (!conf->array_frozen &&
 900			     !atomic_read(&conf->nr_pending[idx]) &&
 901			     atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 902				test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 903			    conf->resync_lock);
 904
 905	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 906		atomic_dec(&conf->barrier[idx]);
 907		spin_unlock_irq(&conf->resync_lock);
 908		wake_up(&conf->wait_barrier);
 909		return -EINTR;
 910	}
 911
 912	atomic_inc(&conf->nr_sync_pending);
 913	spin_unlock_irq(&conf->resync_lock);
 914
 915	return 0;
 916}
 917
 918static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 919{
 920	int idx = sector_to_idx(sector_nr);
 921
 922	BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 923
 924	atomic_dec(&conf->barrier[idx]);
 925	atomic_dec(&conf->nr_sync_pending);
 926	wake_up(&conf->wait_barrier);
 927}
 928
 929static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
 930{
 931	bool ret = true;
 932
 933	/*
 934	 * We need to increase conf->nr_pending[idx] very early here,
 935	 * then raise_barrier() can be blocked when it waits for
 936	 * conf->nr_pending[idx] to be 0. Then we can avoid holding
 937	 * conf->resync_lock when there is no barrier raised in same
 938	 * barrier unit bucket. Also if the array is frozen, I/O
 939	 * should be blocked until array is unfrozen.
 940	 */
 941	atomic_inc(&conf->nr_pending[idx]);
 942	/*
 943	 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 944	 * check conf->barrier[idx]. In raise_barrier() we firstly increase
 945	 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 946	 * barrier is necessary here to make sure conf->barrier[idx] won't be
 947	 * fetched before conf->nr_pending[idx] is increased. Otherwise there
 948	 * will be a race between _wait_barrier() and raise_barrier().
 949	 */
 950	smp_mb__after_atomic();
 951
 952	/*
 953	 * Don't worry about checking two atomic_t variables at same time
 954	 * here. If during we check conf->barrier[idx], the array is
 955	 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 956	 * 0, it is safe to return and make the I/O continue. Because the
 957	 * array is frozen, all I/O returned here will eventually complete
 958	 * or be queued, no race will happen. See code comment in
 959	 * frozen_array().
 960	 */
 961	if (!READ_ONCE(conf->array_frozen) &&
 962	    !atomic_read(&conf->barrier[idx]))
 963		return ret;
 964
 965	/*
 966	 * After holding conf->resync_lock, conf->nr_pending[idx]
 967	 * should be decreased before waiting for barrier to drop.
 968	 * Otherwise, we may encounter a race condition because
 969	 * raise_barrer() might be waiting for conf->nr_pending[idx]
 970	 * to be 0 at same time.
 971	 */
 972	spin_lock_irq(&conf->resync_lock);
 973	atomic_inc(&conf->nr_waiting[idx]);
 974	atomic_dec(&conf->nr_pending[idx]);
 975	/*
 976	 * In case freeze_array() is waiting for
 977	 * get_unqueued_pending() == extra
 978	 */
 979	wake_up_barrier(conf);
 980	/* Wait for the barrier in same barrier unit bucket to drop. */
 981
 982	/* Return false when nowait flag is set */
 983	if (nowait) {
 984		ret = false;
 985	} else {
 986		wait_event_lock_irq(conf->wait_barrier,
 987				!conf->array_frozen &&
 988				!atomic_read(&conf->barrier[idx]),
 989				conf->resync_lock);
 990		atomic_inc(&conf->nr_pending[idx]);
 991	}
 992
 993	atomic_dec(&conf->nr_waiting[idx]);
 994	spin_unlock_irq(&conf->resync_lock);
 995	return ret;
 996}
 997
 998static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
 999{
1000	int idx = sector_to_idx(sector_nr);
1001	bool ret = true;
1002
1003	/*
1004	 * Very similar to _wait_barrier(). The difference is, for read
1005	 * I/O we don't need wait for sync I/O, but if the whole array
1006	 * is frozen, the read I/O still has to wait until the array is
1007	 * unfrozen. Since there is no ordering requirement with
1008	 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1009	 */
1010	atomic_inc(&conf->nr_pending[idx]);
1011
1012	if (!READ_ONCE(conf->array_frozen))
1013		return ret;
1014
1015	spin_lock_irq(&conf->resync_lock);
1016	atomic_inc(&conf->nr_waiting[idx]);
1017	atomic_dec(&conf->nr_pending[idx]);
1018	/*
1019	 * In case freeze_array() is waiting for
1020	 * get_unqueued_pending() == extra
1021	 */
1022	wake_up_barrier(conf);
1023	/* Wait for array to be unfrozen */
1024
1025	/* Return false when nowait flag is set */
1026	if (nowait) {
1027		/* Return false when nowait flag is set */
1028		ret = false;
1029	} else {
1030		wait_event_lock_irq(conf->wait_barrier,
1031				!conf->array_frozen,
1032				conf->resync_lock);
1033		atomic_inc(&conf->nr_pending[idx]);
1034	}
1035
1036	atomic_dec(&conf->nr_waiting[idx]);
1037	spin_unlock_irq(&conf->resync_lock);
1038	return ret;
1039}
1040
1041static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1042{
1043	int idx = sector_to_idx(sector_nr);
1044
1045	return _wait_barrier(conf, idx, nowait);
1046}
1047
1048static void _allow_barrier(struct r1conf *conf, int idx)
1049{
1050	atomic_dec(&conf->nr_pending[idx]);
1051	wake_up_barrier(conf);
1052}
1053
1054static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1055{
1056	int idx = sector_to_idx(sector_nr);
1057
1058	_allow_barrier(conf, idx);
1059}
1060
1061/* conf->resync_lock should be held */
1062static int get_unqueued_pending(struct r1conf *conf)
1063{
1064	int idx, ret;
1065
1066	ret = atomic_read(&conf->nr_sync_pending);
1067	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1068		ret += atomic_read(&conf->nr_pending[idx]) -
1069			atomic_read(&conf->nr_queued[idx]);
1070
1071	return ret;
1072}
1073
1074static void freeze_array(struct r1conf *conf, int extra)
1075{
1076	/* Stop sync I/O and normal I/O and wait for everything to
1077	 * go quiet.
1078	 * This is called in two situations:
1079	 * 1) management command handlers (reshape, remove disk, quiesce).
1080	 * 2) one normal I/O request failed.
1081
1082	 * After array_frozen is set to 1, new sync IO will be blocked at
1083	 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1084	 * or wait_read_barrier(). The flying I/Os will either complete or be
1085	 * queued. When everything goes quite, there are only queued I/Os left.
1086
1087	 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1088	 * barrier bucket index which this I/O request hits. When all sync and
1089	 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1090	 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1091	 * in handle_read_error(), we may call freeze_array() before trying to
1092	 * fix the read error. In this case, the error read I/O is not queued,
1093	 * so get_unqueued_pending() == 1.
1094	 *
1095	 * Therefore before this function returns, we need to wait until
1096	 * get_unqueued_pendings(conf) gets equal to extra. For
1097	 * normal I/O context, extra is 1, in rested situations extra is 0.
1098	 */
1099	spin_lock_irq(&conf->resync_lock);
1100	conf->array_frozen = 1;
1101	raid1_log(conf->mddev, "wait freeze");
1102	wait_event_lock_irq_cmd(
1103		conf->wait_barrier,
1104		get_unqueued_pending(conf) == extra,
1105		conf->resync_lock,
1106		flush_pending_writes(conf));
1107	spin_unlock_irq(&conf->resync_lock);
1108}
1109static void unfreeze_array(struct r1conf *conf)
1110{
1111	/* reverse the effect of the freeze */
1112	spin_lock_irq(&conf->resync_lock);
1113	conf->array_frozen = 0;
1114	spin_unlock_irq(&conf->resync_lock);
1115	wake_up(&conf->wait_barrier);
1116}
1117
1118static void alloc_behind_master_bio(struct r1bio *r1_bio,
1119					   struct bio *bio)
1120{
1121	int size = bio->bi_iter.bi_size;
1122	unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1123	int i = 0;
1124	struct bio *behind_bio = NULL;
1125
1126	behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1127				      &r1_bio->mddev->bio_set);
 
1128
1129	/* discard op, we don't support writezero/writesame yet */
1130	if (!bio_has_data(bio)) {
1131		behind_bio->bi_iter.bi_size = size;
1132		goto skip_copy;
1133	}
1134
 
 
1135	while (i < vcnt && size) {
1136		struct page *page;
1137		int len = min_t(int, PAGE_SIZE, size);
1138
1139		page = alloc_page(GFP_NOIO);
1140		if (unlikely(!page))
1141			goto free_pages;
1142
1143		if (!bio_add_page(behind_bio, page, len, 0)) {
1144			put_page(page);
1145			goto free_pages;
1146		}
1147
1148		size -= len;
1149		i++;
1150	}
1151
1152	bio_copy_data(behind_bio, bio);
1153skip_copy:
1154	r1_bio->behind_master_bio = behind_bio;
1155	set_bit(R1BIO_BehindIO, &r1_bio->state);
1156
1157	return;
1158
1159free_pages:
1160	pr_debug("%dB behind alloc failed, doing sync I/O\n",
1161		 bio->bi_iter.bi_size);
1162	bio_free_pages(behind_bio);
1163	bio_put(behind_bio);
1164}
1165
 
 
 
 
 
 
1166static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1167{
1168	struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1169						  cb);
1170	struct mddev *mddev = plug->cb.data;
1171	struct r1conf *conf = mddev->private;
1172	struct bio *bio;
1173
1174	if (from_schedule) {
1175		spin_lock_irq(&conf->device_lock);
1176		bio_list_merge(&conf->pending_bio_list, &plug->pending);
 
1177		spin_unlock_irq(&conf->device_lock);
1178		wake_up_barrier(conf);
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 enum req_op op = bio_op(bio);
1220	const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1221	int max_sectors;
1222	int rdisk;
1223	bool r1bio_existed = !!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 (r1bio_existed) {
1234		/* Need to get the block device name carefully */
1235		struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
1236
 
1237		if (rdev)
1238			snprintf(b, sizeof(b), "%pg", rdev->bdev);
1239		else
1240			strcpy(b, "???");
 
1241	}
1242
1243	/*
1244	 * Still need barrier for READ in case that whole
1245	 * array is frozen.
1246	 */
1247	if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1248				bio->bi_opf & REQ_NOWAIT)) {
1249		bio_wouldblock_error(bio);
1250		return;
1251	}
1252
1253	if (!r1_bio)
1254		r1_bio = alloc_r1bio(mddev, bio);
1255	else
1256		init_r1bio(r1_bio, mddev, bio);
1257	r1_bio->sectors = max_read_sectors;
1258
1259	/*
1260	 * make_request() can abort the operation when read-ahead is being
1261	 * used and no empty request is available.
1262	 */
1263	rdisk = read_balance(conf, r1_bio, &max_sectors);
1264
1265	if (rdisk < 0) {
1266		/* couldn't find anywhere to read from */
1267		if (r1bio_existed) {
1268			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1269					    mdname(mddev),
1270					    b,
1271					    (unsigned long long)r1_bio->sector);
1272		}
1273		raid_end_bio_io(r1_bio);
1274		return;
1275	}
1276	mirror = conf->mirrors + rdisk;
1277
1278	if (r1bio_existed)
1279		pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1280				    mdname(mddev),
1281				    (unsigned long long)r1_bio->sector,
1282				    mirror->rdev->bdev);
1283
1284	if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1285	    bitmap) {
1286		/*
1287		 * Reading from a write-mostly device must take care not to
1288		 * over-take any writes that are 'behind'
1289		 */
1290		raid1_log(mddev, "wait behind writes");
1291		wait_event(bitmap->behind_wait,
1292			   atomic_read(&bitmap->behind_writes) == 0);
1293	}
1294
1295	if (max_sectors < bio_sectors(bio)) {
1296		struct bio *split = bio_split(bio, max_sectors,
1297					      gfp, &conf->bio_split);
1298		bio_chain(split, bio);
1299		submit_bio_noacct(bio);
1300		bio = split;
1301		r1_bio->master_bio = bio;
1302		r1_bio->sectors = max_sectors;
1303	}
1304
1305	r1_bio->read_disk = rdisk;
1306	if (!r1bio_existed) {
1307		md_account_bio(mddev, &bio);
1308		r1_bio->master_bio = bio;
1309	}
1310	read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1311				   &mddev->bio_set);
1312
1313	r1_bio->bios[rdisk] = read_bio;
1314
1315	read_bio->bi_iter.bi_sector = r1_bio->sector +
1316		mirror->rdev->data_offset;
 
1317	read_bio->bi_end_io = raid1_end_read_request;
1318	read_bio->bi_opf = op | do_sync;
1319	if (test_bit(FailFast, &mirror->rdev->flags) &&
1320	    test_bit(R1BIO_FailFast, &r1_bio->state))
1321	        read_bio->bi_opf |= MD_FAILFAST;
1322	read_bio->bi_private = r1_bio;
1323
1324	if (mddev->gendisk)
1325	        trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1326				      r1_bio->sector);
1327
1328	submit_bio_noacct(read_bio);
1329}
1330
1331static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1332				int max_write_sectors)
1333{
1334	struct r1conf *conf = mddev->private;
1335	struct r1bio *r1_bio;
1336	int i, disks;
1337	struct bitmap *bitmap = mddev->bitmap;
1338	unsigned long flags;
1339	struct md_rdev *blocked_rdev;
 
 
1340	int first_clone;
1341	int max_sectors;
1342	bool write_behind = false;
1343	bool is_discard = (bio_op(bio) == REQ_OP_DISCARD);
1344
1345	if (mddev_is_clustered(mddev) &&
1346	     md_cluster_ops->area_resyncing(mddev, WRITE,
1347		     bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1348
1349		DEFINE_WAIT(w);
1350		if (bio->bi_opf & REQ_NOWAIT) {
1351			bio_wouldblock_error(bio);
1352			return;
1353		}
1354		for (;;) {
1355			prepare_to_wait(&conf->wait_barrier,
1356					&w, TASK_IDLE);
1357			if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1358							bio->bi_iter.bi_sector,
1359							bio_end_sector(bio)))
1360				break;
1361			schedule();
1362		}
1363		finish_wait(&conf->wait_barrier, &w);
1364	}
1365
1366	/*
1367	 * Register the new request and wait if the reconstruction
1368	 * thread has put up a bar for new requests.
1369	 * Continue immediately if no resync is active currently.
1370	 */
1371	if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1372				bio->bi_opf & REQ_NOWAIT)) {
1373		bio_wouldblock_error(bio);
1374		return;
1375	}
1376
1377 retry_write:
1378	r1_bio = alloc_r1bio(mddev, bio);
1379	r1_bio->sectors = max_write_sectors;
1380
 
 
 
 
 
 
1381	/* first select target devices under rcu_lock and
1382	 * inc refcount on their rdev.  Record them by setting
1383	 * bios[x] to bio
1384	 * If there are known/acknowledged bad blocks on any device on
1385	 * which we have seen a write error, we want to avoid writing those
1386	 * blocks.
1387	 * This potentially requires several writes to write around
1388	 * the bad blocks.  Each set of writes gets it's own r1bio
1389	 * with a set of bios attached.
1390	 */
1391
1392	disks = conf->raid_disks * 2;
 
1393	blocked_rdev = NULL;
 
1394	max_sectors = r1_bio->sectors;
1395	for (i = 0;  i < disks; i++) {
1396		struct md_rdev *rdev = conf->mirrors[i].rdev;
1397
1398		/*
1399		 * The write-behind io is only attempted on drives marked as
1400		 * write-mostly, which means we could allocate write behind
1401		 * bio later.
1402		 */
1403		if (!is_discard && rdev && test_bit(WriteMostly, &rdev->flags))
1404			write_behind = true;
1405
1406		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1407			atomic_inc(&rdev->nr_pending);
1408			blocked_rdev = rdev;
1409			break;
1410		}
1411		r1_bio->bios[i] = NULL;
1412		if (!rdev || test_bit(Faulty, &rdev->flags)) {
1413			if (i < conf->raid_disks)
1414				set_bit(R1BIO_Degraded, &r1_bio->state);
1415			continue;
1416		}
1417
1418		atomic_inc(&rdev->nr_pending);
1419		if (test_bit(WriteErrorSeen, &rdev->flags)) {
1420			sector_t first_bad;
1421			int bad_sectors;
1422			int is_bad;
1423
1424			is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1425					     &first_bad, &bad_sectors);
1426			if (is_bad < 0) {
1427				/* mustn't write here until the bad block is
1428				 * acknowledged*/
1429				set_bit(BlockedBadBlocks, &rdev->flags);
1430				blocked_rdev = rdev;
1431				break;
1432			}
1433			if (is_bad && first_bad <= r1_bio->sector) {
1434				/* Cannot write here at all */
1435				bad_sectors -= (r1_bio->sector - first_bad);
1436				if (bad_sectors < max_sectors)
1437					/* mustn't write more than bad_sectors
1438					 * to other devices yet
1439					 */
1440					max_sectors = bad_sectors;
1441				rdev_dec_pending(rdev, mddev);
1442				/* We don't set R1BIO_Degraded as that
1443				 * only applies if the disk is
1444				 * missing, so it might be re-added,
1445				 * and we want to know to recover this
1446				 * chunk.
1447				 * In this case the device is here,
1448				 * and the fact that this chunk is not
1449				 * in-sync is recorded in the bad
1450				 * block log
1451				 */
1452				continue;
1453			}
1454			if (is_bad) {
1455				int good_sectors = first_bad - r1_bio->sector;
1456				if (good_sectors < max_sectors)
1457					max_sectors = good_sectors;
1458			}
1459		}
1460		r1_bio->bios[i] = bio;
1461	}
 
1462
1463	if (unlikely(blocked_rdev)) {
1464		/* Wait for this device to become unblocked */
1465		int j;
1466
1467		for (j = 0; j < i; j++)
1468			if (r1_bio->bios[j])
1469				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1470		free_r1bio(r1_bio);
1471		allow_barrier(conf, bio->bi_iter.bi_sector);
1472
1473		if (bio->bi_opf & REQ_NOWAIT) {
1474			bio_wouldblock_error(bio);
1475			return;
1476		}
1477		raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1478		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1479		wait_barrier(conf, bio->bi_iter.bi_sector, false);
1480		goto retry_write;
1481	}
1482
1483	/*
1484	 * When using a bitmap, we may call alloc_behind_master_bio below.
1485	 * alloc_behind_master_bio allocates a copy of the data payload a page
1486	 * at a time and thus needs a new bio that can fit the whole payload
1487	 * this bio in page sized chunks.
1488	 */
1489	if (write_behind && bitmap)
1490		max_sectors = min_t(int, max_sectors,
1491				    BIO_MAX_VECS * (PAGE_SIZE >> 9));
1492	if (max_sectors < bio_sectors(bio)) {
1493		struct bio *split = bio_split(bio, max_sectors,
1494					      GFP_NOIO, &conf->bio_split);
1495		bio_chain(split, bio);
1496		submit_bio_noacct(bio);
1497		bio = split;
1498		r1_bio->master_bio = bio;
1499		r1_bio->sectors = max_sectors;
1500	}
1501
1502	md_account_bio(mddev, &bio);
1503	r1_bio->master_bio = bio;
1504	atomic_set(&r1_bio->remaining, 1);
1505	atomic_set(&r1_bio->behind_remaining, 0);
1506
1507	first_clone = 1;
1508
1509	for (i = 0; i < disks; i++) {
1510		struct bio *mbio = NULL;
1511		struct md_rdev *rdev = conf->mirrors[i].rdev;
1512		if (!r1_bio->bios[i])
1513			continue;
1514
1515		if (first_clone) {
1516			/* do behind I/O ?
1517			 * Not if there are too many, or cannot
1518			 * allocate memory, or a reader on WriteMostly
1519			 * is waiting for behind writes to flush */
1520			if (bitmap && write_behind &&
1521			    (atomic_read(&bitmap->behind_writes)
1522			     < mddev->bitmap_info.max_write_behind) &&
1523			    !waitqueue_active(&bitmap->behind_wait)) {
1524				alloc_behind_master_bio(r1_bio, bio);
1525			}
1526
1527			md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1528					     test_bit(R1BIO_BehindIO, &r1_bio->state));
1529			first_clone = 0;
1530		}
1531
 
 
 
 
 
 
1532		if (r1_bio->behind_master_bio) {
1533			mbio = bio_alloc_clone(rdev->bdev,
1534					       r1_bio->behind_master_bio,
1535					       GFP_NOIO, &mddev->bio_set);
1536			if (test_bit(CollisionCheck, &rdev->flags))
1537				wait_for_serialization(rdev, r1_bio);
 
 
 
 
1538			if (test_bit(WriteMostly, &rdev->flags))
1539				atomic_inc(&r1_bio->behind_remaining);
1540		} else {
1541			mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1542					       &mddev->bio_set);
1543
1544			if (mddev->serialize_policy)
1545				wait_for_serialization(rdev, r1_bio);
1546		}
1547
1548		r1_bio->bios[i] = mbio;
1549
1550		mbio->bi_iter.bi_sector	= (r1_bio->sector + rdev->data_offset);
 
 
1551		mbio->bi_end_io	= raid1_end_write_request;
1552		mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1553		if (test_bit(FailFast, &rdev->flags) &&
1554		    !test_bit(WriteMostly, &rdev->flags) &&
1555		    conf->raid_disks - mddev->degraded > 1)
1556			mbio->bi_opf |= MD_FAILFAST;
1557		mbio->bi_private = r1_bio;
1558
1559		atomic_inc(&r1_bio->remaining);
1560
1561		if (mddev->gendisk)
1562			trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
 
1563					      r1_bio->sector);
1564		/* flush_pending_writes() needs access to the rdev so...*/
1565		mbio->bi_bdev = (void *)rdev;
1566		if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
 
 
 
 
 
 
 
 
 
1567			spin_lock_irqsave(&conf->device_lock, flags);
1568			bio_list_add(&conf->pending_bio_list, mbio);
 
1569			spin_unlock_irqrestore(&conf->device_lock, flags);
1570			md_wakeup_thread(mddev->thread);
1571		}
1572	}
1573
1574	r1_bio_write_done(r1_bio);
1575
1576	/* In case raid1d snuck in to freeze_array */
1577	wake_up_barrier(conf);
1578}
1579
1580static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1581{
1582	sector_t sectors;
1583
1584	if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1585	    && md_flush_request(mddev, bio))
1586		return true;
 
1587
1588	/*
1589	 * There is a limit to the maximum size, but
1590	 * the read/write handler might find a lower limit
1591	 * due to bad blocks.  To avoid multiple splits,
1592	 * we pass the maximum number of sectors down
1593	 * and let the lower level perform the split.
1594	 */
1595	sectors = align_to_barrier_unit_end(
1596		bio->bi_iter.bi_sector, bio_sectors(bio));
1597
1598	if (bio_data_dir(bio) == READ)
1599		raid1_read_request(mddev, bio, sectors, NULL);
1600	else {
1601		if (!md_write_start(mddev,bio))
1602			return false;
1603		raid1_write_request(mddev, bio, sectors);
1604	}
1605	return true;
1606}
1607
1608static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1609{
1610	struct r1conf *conf = mddev->private;
1611	int i;
1612
1613	lockdep_assert_held(&mddev->lock);
1614
1615	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1616		   conf->raid_disks - mddev->degraded);
 
1617	for (i = 0; i < conf->raid_disks; i++) {
1618		struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1619
1620		seq_printf(seq, "%s",
1621			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1622	}
 
1623	seq_printf(seq, "]");
1624}
1625
1626/**
1627 * raid1_error() - RAID1 error handler.
1628 * @mddev: affected md device.
1629 * @rdev: member device to fail.
1630 *
1631 * The routine acknowledges &rdev failure and determines new @mddev state.
1632 * If it failed, then:
1633 *	- &MD_BROKEN flag is set in &mddev->flags.
1634 *	- recovery is disabled.
1635 * Otherwise, it must be degraded:
1636 *	- recovery is interrupted.
1637 *	- &mddev->degraded is bumped.
1638 *
1639 * @rdev is marked as &Faulty excluding case when array is failed and
1640 * &mddev->fail_last_dev is off.
1641 */
1642static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1643{
 
1644	struct r1conf *conf = mddev->private;
1645	unsigned long flags;
1646
 
 
 
 
 
 
1647	spin_lock_irqsave(&conf->device_lock, flags);
1648
1649	if (test_bit(In_sync, &rdev->flags) &&
1650	    (conf->raid_disks - mddev->degraded) == 1) {
1651		set_bit(MD_BROKEN, &mddev->flags);
1652
1653		if (!mddev->fail_last_dev) {
1654			conf->recovery_disabled = mddev->recovery_disabled;
1655			spin_unlock_irqrestore(&conf->device_lock, flags);
1656			return;
1657		}
 
1658	}
1659	set_bit(Blocked, &rdev->flags);
1660	if (test_and_clear_bit(In_sync, &rdev->flags))
1661		mddev->degraded++;
1662	set_bit(Faulty, &rdev->flags);
1663	spin_unlock_irqrestore(&conf->device_lock, flags);
1664	/*
1665	 * if recovery is running, make sure it aborts.
1666	 */
1667	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1668	set_mask_bits(&mddev->sb_flags, 0,
1669		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1670	pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1671		"md/raid1:%s: Operation continuing on %d devices.\n",
1672		mdname(mddev), rdev->bdev,
1673		mdname(mddev), conf->raid_disks - mddev->degraded);
1674}
1675
1676static void print_conf(struct r1conf *conf)
1677{
1678	int i;
1679
1680	pr_debug("RAID1 conf printout:\n");
1681	if (!conf) {
1682		pr_debug("(!conf)\n");
1683		return;
1684	}
1685	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1686		 conf->raid_disks);
1687
1688	lockdep_assert_held(&conf->mddev->reconfig_mutex);
1689	for (i = 0; i < conf->raid_disks; i++) {
1690		struct md_rdev *rdev = conf->mirrors[i].rdev;
 
1691		if (rdev)
1692			pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1693				 i, !test_bit(In_sync, &rdev->flags),
1694				 !test_bit(Faulty, &rdev->flags),
1695				 rdev->bdev);
1696	}
 
1697}
1698
1699static void close_sync(struct r1conf *conf)
1700{
1701	int idx;
1702
1703	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1704		_wait_barrier(conf, idx, false);
1705		_allow_barrier(conf, idx);
1706	}
1707
1708	mempool_exit(&conf->r1buf_pool);
1709}
1710
1711static int raid1_spare_active(struct mddev *mddev)
1712{
1713	int i;
1714	struct r1conf *conf = mddev->private;
1715	int count = 0;
1716	unsigned long flags;
1717
1718	/*
1719	 * Find all failed disks within the RAID1 configuration
1720	 * and mark them readable.
1721	 * Called under mddev lock, so rcu protection not needed.
1722	 * device_lock used to avoid races with raid1_end_read_request
1723	 * which expects 'In_sync' flags and ->degraded to be consistent.
1724	 */
1725	spin_lock_irqsave(&conf->device_lock, flags);
1726	for (i = 0; i < conf->raid_disks; i++) {
1727		struct md_rdev *rdev = conf->mirrors[i].rdev;
1728		struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1729		if (repl
1730		    && !test_bit(Candidate, &repl->flags)
1731		    && repl->recovery_offset == MaxSector
1732		    && !test_bit(Faulty, &repl->flags)
1733		    && !test_and_set_bit(In_sync, &repl->flags)) {
1734			/* replacement has just become active */
1735			if (!rdev ||
1736			    !test_and_clear_bit(In_sync, &rdev->flags))
1737				count++;
1738			if (rdev) {
1739				/* Replaced device not technically
1740				 * faulty, but we need to be sure
1741				 * it gets removed and never re-added
1742				 */
1743				set_bit(Faulty, &rdev->flags);
1744				sysfs_notify_dirent_safe(
1745					rdev->sysfs_state);
1746			}
1747		}
1748		if (rdev
1749		    && rdev->recovery_offset == MaxSector
1750		    && !test_bit(Faulty, &rdev->flags)
1751		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1752			count++;
1753			sysfs_notify_dirent_safe(rdev->sysfs_state);
1754		}
1755	}
1756	mddev->degraded -= count;
1757	spin_unlock_irqrestore(&conf->device_lock, flags);
1758
1759	print_conf(conf);
1760	return count;
1761}
1762
1763static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1764{
1765	struct r1conf *conf = mddev->private;
1766	int err = -EEXIST;
1767	int mirror = 0, repl_slot = -1;
1768	struct raid1_info *p;
1769	int first = 0;
1770	int last = conf->raid_disks - 1;
1771
1772	if (mddev->recovery_disabled == conf->recovery_disabled)
1773		return -EBUSY;
1774
1775	if (md_integrity_add_rdev(rdev, mddev))
1776		return -ENXIO;
1777
1778	if (rdev->raid_disk >= 0)
1779		first = last = rdev->raid_disk;
1780
1781	/*
1782	 * find the disk ... but prefer rdev->saved_raid_disk
1783	 * if possible.
1784	 */
1785	if (rdev->saved_raid_disk >= 0 &&
1786	    rdev->saved_raid_disk >= first &&
1787	    rdev->saved_raid_disk < conf->raid_disks &&
1788	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1789		first = last = rdev->saved_raid_disk;
1790
1791	for (mirror = first; mirror <= last; mirror++) {
1792		p = conf->mirrors + mirror;
1793		if (!p->rdev) {
1794			if (mddev->gendisk)
1795				disk_stack_limits(mddev->gendisk, rdev->bdev,
1796						  rdev->data_offset << 9);
1797
1798			p->head_position = 0;
1799			rdev->raid_disk = mirror;
1800			err = 0;
1801			/* As all devices are equivalent, we don't need a full recovery
1802			 * if this was recently any drive of the array
1803			 */
1804			if (rdev->saved_raid_disk < 0)
1805				conf->fullsync = 1;
1806			WRITE_ONCE(p->rdev, rdev);
1807			break;
1808		}
1809		if (test_bit(WantReplacement, &p->rdev->flags) &&
1810		    p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1811			repl_slot = mirror;
1812	}
1813
1814	if (err && repl_slot >= 0) {
1815		/* Add this device as a replacement */
1816		p = conf->mirrors + repl_slot;
1817		clear_bit(In_sync, &rdev->flags);
1818		set_bit(Replacement, &rdev->flags);
1819		rdev->raid_disk = repl_slot;
1820		err = 0;
1821		conf->fullsync = 1;
1822		WRITE_ONCE(p[conf->raid_disks].rdev, rdev);
1823	}
1824
 
1825	print_conf(conf);
1826	return err;
1827}
1828
1829static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1830{
1831	struct r1conf *conf = mddev->private;
1832	int err = 0;
1833	int number = rdev->raid_disk;
1834	struct raid1_info *p = conf->mirrors + number;
1835
1836	if (unlikely(number >= conf->raid_disks))
1837		goto abort;
1838
1839	if (rdev != p->rdev)
1840		p = conf->mirrors + conf->raid_disks + number;
1841
1842	print_conf(conf);
1843	if (rdev == p->rdev) {
1844		if (test_bit(In_sync, &rdev->flags) ||
1845		    atomic_read(&rdev->nr_pending)) {
1846			err = -EBUSY;
1847			goto abort;
1848		}
1849		/* Only remove non-faulty devices if recovery
1850		 * is not possible.
1851		 */
1852		if (!test_bit(Faulty, &rdev->flags) &&
1853		    mddev->recovery_disabled != conf->recovery_disabled &&
1854		    mddev->degraded < conf->raid_disks) {
1855			err = -EBUSY;
1856			goto abort;
1857		}
1858		WRITE_ONCE(p->rdev, NULL);
 
 
 
 
 
 
 
 
 
1859		if (conf->mirrors[conf->raid_disks + number].rdev) {
1860			/* We just removed a device that is being replaced.
1861			 * Move down the replacement.  We drain all IO before
1862			 * doing this to avoid confusion.
1863			 */
1864			struct md_rdev *repl =
1865				conf->mirrors[conf->raid_disks + number].rdev;
1866			freeze_array(conf, 0);
1867			if (atomic_read(&repl->nr_pending)) {
1868				/* It means that some queued IO of retry_list
1869				 * hold repl. Thus, we cannot set replacement
1870				 * as NULL, avoiding rdev NULL pointer
1871				 * dereference in sync_request_write and
1872				 * handle_write_finished.
1873				 */
1874				err = -EBUSY;
1875				unfreeze_array(conf);
1876				goto abort;
1877			}
1878			clear_bit(Replacement, &repl->flags);
1879			WRITE_ONCE(p->rdev, repl);
1880			conf->mirrors[conf->raid_disks + number].rdev = NULL;
1881			unfreeze_array(conf);
1882		}
1883
1884		clear_bit(WantReplacement, &rdev->flags);
1885		err = md_integrity_register(mddev);
1886	}
1887abort:
1888
1889	print_conf(conf);
1890	return err;
1891}
1892
1893static void end_sync_read(struct bio *bio)
1894{
1895	struct r1bio *r1_bio = get_resync_r1bio(bio);
1896
1897	update_head_pos(r1_bio->read_disk, r1_bio);
1898
1899	/*
1900	 * we have read a block, now it needs to be re-written,
1901	 * or re-read if the read failed.
1902	 * We don't do much here, just schedule handling by raid1d
1903	 */
1904	if (!bio->bi_status)
1905		set_bit(R1BIO_Uptodate, &r1_bio->state);
1906
1907	if (atomic_dec_and_test(&r1_bio->remaining))
1908		reschedule_retry(r1_bio);
1909}
1910
1911static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1912{
1913	sector_t sync_blocks = 0;
1914	sector_t s = r1_bio->sector;
1915	long sectors_to_go = r1_bio->sectors;
1916
1917	/* make sure these bits don't get cleared. */
1918	do {
1919		md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1920		s += sync_blocks;
1921		sectors_to_go -= sync_blocks;
1922	} while (sectors_to_go > 0);
1923}
1924
1925static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1926{
1927	if (atomic_dec_and_test(&r1_bio->remaining)) {
1928		struct mddev *mddev = r1_bio->mddev;
1929		int s = r1_bio->sectors;
1930
1931		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1932		    test_bit(R1BIO_WriteError, &r1_bio->state))
1933			reschedule_retry(r1_bio);
1934		else {
1935			put_buf(r1_bio);
1936			md_done_sync(mddev, s, uptodate);
1937		}
1938	}
1939}
1940
1941static void end_sync_write(struct bio *bio)
1942{
1943	int uptodate = !bio->bi_status;
1944	struct r1bio *r1_bio = get_resync_r1bio(bio);
1945	struct mddev *mddev = r1_bio->mddev;
1946	struct r1conf *conf = mddev->private;
1947	sector_t first_bad;
1948	int bad_sectors;
1949	struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1950
1951	if (!uptodate) {
1952		abort_sync_write(mddev, r1_bio);
1953		set_bit(WriteErrorSeen, &rdev->flags);
1954		if (!test_and_set_bit(WantReplacement, &rdev->flags))
1955			set_bit(MD_RECOVERY_NEEDED, &
1956				mddev->recovery);
1957		set_bit(R1BIO_WriteError, &r1_bio->state);
1958	} else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1959			       &first_bad, &bad_sectors) &&
1960		   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1961				r1_bio->sector,
1962				r1_bio->sectors,
1963				&first_bad, &bad_sectors)
1964		)
1965		set_bit(R1BIO_MadeGood, &r1_bio->state);
1966
1967	put_sync_write_buf(r1_bio, uptodate);
1968}
1969
1970static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1971			   int sectors, struct page *page, blk_opf_t rw)
1972{
1973	if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1974		/* success */
1975		return 1;
1976	if (rw == REQ_OP_WRITE) {
1977		set_bit(WriteErrorSeen, &rdev->flags);
1978		if (!test_and_set_bit(WantReplacement,
1979				      &rdev->flags))
1980			set_bit(MD_RECOVERY_NEEDED, &
1981				rdev->mddev->recovery);
1982	}
1983	/* need to record an error - either for the block or the device */
1984	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1985		md_error(rdev->mddev, rdev);
1986	return 0;
1987}
1988
1989static int fix_sync_read_error(struct r1bio *r1_bio)
1990{
1991	/* Try some synchronous reads of other devices to get
1992	 * good data, much like with normal read errors.  Only
1993	 * read into the pages we already have so we don't
1994	 * need to re-issue the read request.
1995	 * We don't need to freeze the array, because being in an
1996	 * active sync request, there is no normal IO, and
1997	 * no overlapping syncs.
1998	 * We don't need to check is_badblock() again as we
1999	 * made sure that anything with a bad block in range
2000	 * will have bi_end_io clear.
2001	 */
2002	struct mddev *mddev = r1_bio->mddev;
2003	struct r1conf *conf = mddev->private;
2004	struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2005	struct page **pages = get_resync_pages(bio)->pages;
2006	sector_t sect = r1_bio->sector;
2007	int sectors = r1_bio->sectors;
2008	int idx = 0;
2009	struct md_rdev *rdev;
2010
2011	rdev = conf->mirrors[r1_bio->read_disk].rdev;
2012	if (test_bit(FailFast, &rdev->flags)) {
2013		/* Don't try recovering from here - just fail it
2014		 * ... unless it is the last working device of course */
2015		md_error(mddev, rdev);
2016		if (test_bit(Faulty, &rdev->flags))
2017			/* Don't try to read from here, but make sure
2018			 * put_buf does it's thing
2019			 */
2020			bio->bi_end_io = end_sync_write;
2021	}
2022
2023	while(sectors) {
2024		int s = sectors;
2025		int d = r1_bio->read_disk;
2026		int success = 0;
2027		int start;
2028
2029		if (s > (PAGE_SIZE>>9))
2030			s = PAGE_SIZE >> 9;
2031		do {
2032			if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2033				/* No rcu protection needed here devices
2034				 * can only be removed when no resync is
2035				 * active, and resync is currently active
2036				 */
2037				rdev = conf->mirrors[d].rdev;
2038				if (sync_page_io(rdev, sect, s<<9,
2039						 pages[idx],
2040						 REQ_OP_READ, false)) {
2041					success = 1;
2042					break;
2043				}
2044			}
2045			d++;
2046			if (d == conf->raid_disks * 2)
2047				d = 0;
2048		} while (!success && d != r1_bio->read_disk);
2049
2050		if (!success) {
 
2051			int abort = 0;
2052			/* Cannot read from anywhere, this block is lost.
2053			 * Record a bad block on each device.  If that doesn't
2054			 * work just disable and interrupt the recovery.
2055			 * Don't fail devices as that won't really help.
2056			 */
2057			pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2058					    mdname(mddev), bio->bi_bdev,
2059					    (unsigned long long)r1_bio->sector);
2060			for (d = 0; d < conf->raid_disks * 2; d++) {
2061				rdev = conf->mirrors[d].rdev;
2062				if (!rdev || test_bit(Faulty, &rdev->flags))
2063					continue;
2064				if (!rdev_set_badblocks(rdev, sect, s, 0))
2065					abort = 1;
2066			}
2067			if (abort) {
2068				conf->recovery_disabled =
2069					mddev->recovery_disabled;
2070				set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2071				md_done_sync(mddev, r1_bio->sectors, 0);
2072				put_buf(r1_bio);
2073				return 0;
2074			}
2075			/* Try next page */
2076			sectors -= s;
2077			sect += s;
2078			idx++;
2079			continue;
2080		}
2081
2082		start = d;
2083		/* write it back and re-read */
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					    REQ_OP_WRITE) == 0) {
2094				r1_bio->bios[d]->bi_end_io = NULL;
2095				rdev_dec_pending(rdev, mddev);
2096			}
2097		}
2098		d = start;
2099		while (d != r1_bio->read_disk) {
2100			if (d == 0)
2101				d = conf->raid_disks * 2;
2102			d--;
2103			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2104				continue;
2105			rdev = conf->mirrors[d].rdev;
2106			if (r1_sync_page_io(rdev, sect, s,
2107					    pages[idx],
2108					    REQ_OP_READ) != 0)
2109				atomic_add(s, &rdev->corrected_errors);
2110		}
2111		sectors -= s;
2112		sect += s;
2113		idx ++;
2114	}
2115	set_bit(R1BIO_Uptodate, &r1_bio->state);
2116	bio->bi_status = 0;
2117	return 1;
2118}
2119
2120static void process_checks(struct r1bio *r1_bio)
2121{
2122	/* We have read all readable devices.  If we haven't
2123	 * got the block, then there is no hope left.
2124	 * If we have, then we want to do a comparison
2125	 * and skip the write if everything is the same.
2126	 * If any blocks failed to read, then we need to
2127	 * attempt an over-write
2128	 */
2129	struct mddev *mddev = r1_bio->mddev;
2130	struct r1conf *conf = mddev->private;
2131	int primary;
2132	int i;
2133	int vcnt;
2134
2135	/* Fix variable parts of all bios */
2136	vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2137	for (i = 0; i < conf->raid_disks * 2; i++) {
2138		blk_status_t status;
2139		struct bio *b = r1_bio->bios[i];
2140		struct resync_pages *rp = get_resync_pages(b);
2141		if (b->bi_end_io != end_sync_read)
2142			continue;
2143		/* fixup the bio for reuse, but preserve errno */
2144		status = b->bi_status;
2145		bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2146		b->bi_status = status;
2147		b->bi_iter.bi_sector = r1_bio->sector +
2148			conf->mirrors[i].rdev->data_offset;
 
2149		b->bi_end_io = end_sync_read;
2150		rp->raid_bio = r1_bio;
2151		b->bi_private = rp;
2152
2153		/* initialize bvec table again */
2154		md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2155	}
2156	for (primary = 0; primary < conf->raid_disks * 2; primary++)
2157		if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2158		    !r1_bio->bios[primary]->bi_status) {
2159			r1_bio->bios[primary]->bi_end_io = NULL;
2160			rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2161			break;
2162		}
2163	r1_bio->read_disk = primary;
2164	for (i = 0; i < conf->raid_disks * 2; i++) {
2165		int j = 0;
2166		struct bio *pbio = r1_bio->bios[primary];
2167		struct bio *sbio = r1_bio->bios[i];
2168		blk_status_t status = sbio->bi_status;
2169		struct page **ppages = get_resync_pages(pbio)->pages;
2170		struct page **spages = get_resync_pages(sbio)->pages;
2171		struct bio_vec *bi;
2172		int page_len[RESYNC_PAGES] = { 0 };
2173		struct bvec_iter_all iter_all;
2174
2175		if (sbio->bi_end_io != end_sync_read)
2176			continue;
2177		/* Now we can 'fixup' the error value */
2178		sbio->bi_status = 0;
2179
2180		bio_for_each_segment_all(bi, sbio, iter_all)
2181			page_len[j++] = bi->bv_len;
2182
2183		if (!status) {
2184			for (j = vcnt; j-- ; ) {
2185				if (memcmp(page_address(ppages[j]),
2186					   page_address(spages[j]),
2187					   page_len[j]))
2188					break;
2189			}
2190		} else
2191			j = 0;
2192		if (j >= 0)
2193			atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2194		if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2195			      && !status)) {
2196			/* No need to write to this device. */
2197			sbio->bi_end_io = NULL;
2198			rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2199			continue;
2200		}
2201
2202		bio_copy_data(sbio, pbio);
2203	}
2204}
2205
2206static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2207{
2208	struct r1conf *conf = mddev->private;
2209	int i;
2210	int disks = conf->raid_disks * 2;
2211	struct bio *wbio;
2212
2213	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2214		/* ouch - failed to read all of that. */
2215		if (!fix_sync_read_error(r1_bio))
2216			return;
2217
2218	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2219		process_checks(r1_bio);
2220
2221	/*
2222	 * schedule writes
2223	 */
2224	atomic_set(&r1_bio->remaining, 1);
2225	for (i = 0; i < disks ; i++) {
2226		wbio = r1_bio->bios[i];
2227		if (wbio->bi_end_io == NULL ||
2228		    (wbio->bi_end_io == end_sync_read &&
2229		     (i == r1_bio->read_disk ||
2230		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2231			continue;
2232		if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2233			abort_sync_write(mddev, r1_bio);
2234			continue;
2235		}
2236
2237		wbio->bi_opf = REQ_OP_WRITE;
2238		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2239			wbio->bi_opf |= MD_FAILFAST;
2240
2241		wbio->bi_end_io = end_sync_write;
2242		atomic_inc(&r1_bio->remaining);
2243		md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2244
2245		submit_bio_noacct(wbio);
2246	}
2247
2248	put_sync_write_buf(r1_bio, 1);
2249}
2250
2251/*
2252 * This is a kernel thread which:
2253 *
2254 *	1.	Retries failed read operations on working mirrors.
2255 *	2.	Updates the raid superblock when problems encounter.
2256 *	3.	Performs writes following reads for array synchronising.
2257 */
2258
2259static void fix_read_error(struct r1conf *conf, struct r1bio *r1_bio)
 
2260{
2261	sector_t sect = r1_bio->sector;
2262	int sectors = r1_bio->sectors;
2263	int read_disk = r1_bio->read_disk;
2264	struct mddev *mddev = conf->mddev;
2265	struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2266
2267	if (exceed_read_errors(mddev, rdev)) {
2268		r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2269		return;
2270	}
2271
2272	while(sectors) {
2273		int s = sectors;
2274		int d = read_disk;
2275		int success = 0;
2276		int start;
 
2277
2278		if (s > (PAGE_SIZE>>9))
2279			s = PAGE_SIZE >> 9;
2280
2281		do {
2282			sector_t first_bad;
2283			int bad_sectors;
2284
2285			rdev = conf->mirrors[d].rdev;
 
2286			if (rdev &&
2287			    (test_bit(In_sync, &rdev->flags) ||
2288			     (!test_bit(Faulty, &rdev->flags) &&
2289			      rdev->recovery_offset >= sect + s)) &&
2290			    is_badblock(rdev, sect, s,
2291					&first_bad, &bad_sectors) == 0) {
2292				atomic_inc(&rdev->nr_pending);
 
2293				if (sync_page_io(rdev, sect, s<<9,
2294					 conf->tmppage, REQ_OP_READ, false))
2295					success = 1;
2296				rdev_dec_pending(rdev, mddev);
2297				if (success)
2298					break;
2299			}
2300
2301			d++;
2302			if (d == conf->raid_disks * 2)
2303				d = 0;
2304		} while (d != read_disk);
2305
2306		if (!success) {
2307			/* Cannot read from anywhere - mark it bad */
2308			struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2309			if (!rdev_set_badblocks(rdev, sect, s, 0))
2310				md_error(mddev, rdev);
2311			break;
2312		}
2313		/* write it back and re-read */
2314		start = d;
2315		while (d != read_disk) {
2316			if (d==0)
2317				d = conf->raid_disks * 2;
2318			d--;
2319			rdev = conf->mirrors[d].rdev;
 
2320			if (rdev &&
2321			    !test_bit(Faulty, &rdev->flags)) {
2322				atomic_inc(&rdev->nr_pending);
 
2323				r1_sync_page_io(rdev, sect, s,
2324						conf->tmppage, REQ_OP_WRITE);
2325				rdev_dec_pending(rdev, mddev);
2326			}
 
2327		}
2328		d = start;
2329		while (d != read_disk) {
 
2330			if (d==0)
2331				d = conf->raid_disks * 2;
2332			d--;
2333			rdev = conf->mirrors[d].rdev;
 
2334			if (rdev &&
2335			    !test_bit(Faulty, &rdev->flags)) {
2336				atomic_inc(&rdev->nr_pending);
 
2337				if (r1_sync_page_io(rdev, sect, s,
2338						conf->tmppage, REQ_OP_READ)) {
2339					atomic_add(s, &rdev->corrected_errors);
2340					pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2341						mdname(mddev), s,
2342						(unsigned long long)(sect +
2343								     rdev->data_offset),
2344						rdev->bdev);
2345				}
2346				rdev_dec_pending(rdev, mddev);
2347			}
 
2348		}
2349		sectors -= s;
2350		sect += s;
2351	}
2352}
2353
2354static int narrow_write_error(struct r1bio *r1_bio, int i)
2355{
2356	struct mddev *mddev = r1_bio->mddev;
2357	struct r1conf *conf = mddev->private;
2358	struct md_rdev *rdev = conf->mirrors[i].rdev;
2359
2360	/* bio has the data to be written to device 'i' where
2361	 * we just recently had a write error.
2362	 * We repeatedly clone the bio and trim down to one block,
2363	 * then try the write.  Where the write fails we record
2364	 * a bad block.
2365	 * It is conceivable that the bio doesn't exactly align with
2366	 * blocks.  We must handle this somehow.
2367	 *
2368	 * We currently own a reference on the rdev.
2369	 */
2370
2371	int block_sectors;
2372	sector_t sector;
2373	int sectors;
2374	int sect_to_write = r1_bio->sectors;
2375	int ok = 1;
2376
2377	if (rdev->badblocks.shift < 0)
2378		return 0;
2379
2380	block_sectors = roundup(1 << rdev->badblocks.shift,
2381				bdev_logical_block_size(rdev->bdev) >> 9);
2382	sector = r1_bio->sector;
2383	sectors = ((sector + block_sectors)
2384		   & ~(sector_t)(block_sectors - 1))
2385		- sector;
2386
2387	while (sect_to_write) {
2388		struct bio *wbio;
2389		if (sectors > sect_to_write)
2390			sectors = sect_to_write;
2391		/* Write at 'sector' for 'sectors'*/
2392
2393		if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2394			wbio = bio_alloc_clone(rdev->bdev,
2395					       r1_bio->behind_master_bio,
2396					       GFP_NOIO, &mddev->bio_set);
2397		} else {
2398			wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2399					       GFP_NOIO, &mddev->bio_set);
2400		}
2401
2402		wbio->bi_opf = REQ_OP_WRITE;
2403		wbio->bi_iter.bi_sector = r1_bio->sector;
2404		wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2405
2406		bio_trim(wbio, sector - r1_bio->sector, sectors);
2407		wbio->bi_iter.bi_sector += rdev->data_offset;
 
2408
2409		if (submit_bio_wait(wbio) < 0)
2410			/* failure! */
2411			ok = rdev_set_badblocks(rdev, sector,
2412						sectors, 0)
2413				&& ok;
2414
2415		bio_put(wbio);
2416		sect_to_write -= sectors;
2417		sector += sectors;
2418		sectors = block_sectors;
2419	}
2420	return ok;
2421}
2422
2423static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2424{
2425	int m;
2426	int s = r1_bio->sectors;
2427	for (m = 0; m < conf->raid_disks * 2 ; m++) {
2428		struct md_rdev *rdev = conf->mirrors[m].rdev;
2429		struct bio *bio = r1_bio->bios[m];
2430		if (bio->bi_end_io == NULL)
2431			continue;
2432		if (!bio->bi_status &&
2433		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2434			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2435		}
2436		if (bio->bi_status &&
2437		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2438			if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2439				md_error(conf->mddev, rdev);
2440		}
2441	}
2442	put_buf(r1_bio);
2443	md_done_sync(conf->mddev, s, 1);
2444}
2445
2446static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2447{
2448	int m, idx;
2449	bool fail = false;
2450
2451	for (m = 0; m < conf->raid_disks * 2 ; m++)
2452		if (r1_bio->bios[m] == IO_MADE_GOOD) {
2453			struct md_rdev *rdev = conf->mirrors[m].rdev;
2454			rdev_clear_badblocks(rdev,
2455					     r1_bio->sector,
2456					     r1_bio->sectors, 0);
2457			rdev_dec_pending(rdev, conf->mddev);
2458		} else if (r1_bio->bios[m] != NULL) {
2459			/* This drive got a write error.  We need to
2460			 * narrow down and record precise write
2461			 * errors.
2462			 */
2463			fail = true;
2464			if (!narrow_write_error(r1_bio, m)) {
2465				md_error(conf->mddev,
2466					 conf->mirrors[m].rdev);
2467				/* an I/O failed, we can't clear the bitmap */
2468				set_bit(R1BIO_Degraded, &r1_bio->state);
2469			}
2470			rdev_dec_pending(conf->mirrors[m].rdev,
2471					 conf->mddev);
2472		}
2473	if (fail) {
2474		spin_lock_irq(&conf->device_lock);
2475		list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2476		idx = sector_to_idx(r1_bio->sector);
2477		atomic_inc(&conf->nr_queued[idx]);
2478		spin_unlock_irq(&conf->device_lock);
2479		/*
2480		 * In case freeze_array() is waiting for condition
2481		 * get_unqueued_pending() == extra to be true.
2482		 */
2483		wake_up(&conf->wait_barrier);
2484		md_wakeup_thread(conf->mddev->thread);
2485	} else {
2486		if (test_bit(R1BIO_WriteError, &r1_bio->state))
2487			close_write(r1_bio);
2488		raid_end_bio_io(r1_bio);
2489	}
2490}
2491
2492static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2493{
2494	struct mddev *mddev = conf->mddev;
2495	struct bio *bio;
2496	struct md_rdev *rdev;
2497	sector_t sector;
2498
2499	clear_bit(R1BIO_ReadError, &r1_bio->state);
2500	/* we got a read error. Maybe the drive is bad.  Maybe just
2501	 * the block and we can fix it.
2502	 * We freeze all other IO, and try reading the block from
2503	 * other devices.  When we find one, we re-write
2504	 * and check it that fixes the read error.
2505	 * This is all done synchronously while the array is
2506	 * frozen
2507	 */
2508
2509	bio = r1_bio->bios[r1_bio->read_disk];
2510	bio_put(bio);
2511	r1_bio->bios[r1_bio->read_disk] = NULL;
2512
2513	rdev = conf->mirrors[r1_bio->read_disk].rdev;
2514	if (mddev->ro == 0
2515	    && !test_bit(FailFast, &rdev->flags)) {
2516		freeze_array(conf, 1);
2517		fix_read_error(conf, r1_bio);
 
2518		unfreeze_array(conf);
2519	} else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2520		md_error(mddev, rdev);
2521	} else {
2522		r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2523	}
2524
2525	rdev_dec_pending(rdev, conf->mddev);
2526	sector = r1_bio->sector;
2527	bio = r1_bio->master_bio;
2528
2529	/* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2530	r1_bio->state = 0;
2531	raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2532	allow_barrier(conf, sector);
2533}
2534
2535static void raid1d(struct md_thread *thread)
2536{
2537	struct mddev *mddev = thread->mddev;
2538	struct r1bio *r1_bio;
2539	unsigned long flags;
2540	struct r1conf *conf = mddev->private;
2541	struct list_head *head = &conf->retry_list;
2542	struct blk_plug plug;
2543	int idx;
2544
2545	md_check_recovery(mddev);
2546
2547	if (!list_empty_careful(&conf->bio_end_io_list) &&
2548	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2549		LIST_HEAD(tmp);
2550		spin_lock_irqsave(&conf->device_lock, flags);
2551		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2552			list_splice_init(&conf->bio_end_io_list, &tmp);
2553		spin_unlock_irqrestore(&conf->device_lock, flags);
2554		while (!list_empty(&tmp)) {
2555			r1_bio = list_first_entry(&tmp, struct r1bio,
2556						  retry_list);
2557			list_del(&r1_bio->retry_list);
2558			idx = sector_to_idx(r1_bio->sector);
2559			atomic_dec(&conf->nr_queued[idx]);
2560			if (mddev->degraded)
2561				set_bit(R1BIO_Degraded, &r1_bio->state);
2562			if (test_bit(R1BIO_WriteError, &r1_bio->state))
2563				close_write(r1_bio);
2564			raid_end_bio_io(r1_bio);
2565		}
2566	}
2567
2568	blk_start_plug(&plug);
2569	for (;;) {
2570
2571		flush_pending_writes(conf);
2572
2573		spin_lock_irqsave(&conf->device_lock, flags);
2574		if (list_empty(head)) {
2575			spin_unlock_irqrestore(&conf->device_lock, flags);
2576			break;
2577		}
2578		r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2579		list_del(head->prev);
2580		idx = sector_to_idx(r1_bio->sector);
2581		atomic_dec(&conf->nr_queued[idx]);
2582		spin_unlock_irqrestore(&conf->device_lock, flags);
2583
2584		mddev = r1_bio->mddev;
2585		conf = mddev->private;
2586		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2587			if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2588			    test_bit(R1BIO_WriteError, &r1_bio->state))
2589				handle_sync_write_finished(conf, r1_bio);
2590			else
2591				sync_request_write(mddev, r1_bio);
2592		} else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2593			   test_bit(R1BIO_WriteError, &r1_bio->state))
2594			handle_write_finished(conf, r1_bio);
2595		else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2596			handle_read_error(conf, r1_bio);
2597		else
2598			WARN_ON_ONCE(1);
2599
2600		cond_resched();
2601		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2602			md_check_recovery(mddev);
2603	}
2604	blk_finish_plug(&plug);
2605}
2606
2607static int init_resync(struct r1conf *conf)
2608{
2609	int buffs;
2610
2611	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2612	BUG_ON(mempool_initialized(&conf->r1buf_pool));
2613
2614	return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2615			    r1buf_pool_free, conf->poolinfo);
2616}
2617
2618static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2619{
2620	struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2621	struct resync_pages *rps;
2622	struct bio *bio;
2623	int i;
2624
2625	for (i = conf->poolinfo->raid_disks; i--; ) {
2626		bio = r1bio->bios[i];
2627		rps = bio->bi_private;
2628		bio_reset(bio, NULL, 0);
2629		bio->bi_private = rps;
2630	}
2631	r1bio->master_bio = NULL;
2632	return r1bio;
2633}
2634
2635/*
2636 * perform a "sync" on one "block"
2637 *
2638 * We need to make sure that no normal I/O request - particularly write
2639 * requests - conflict with active sync requests.
2640 *
2641 * This is achieved by tracking pending requests and a 'barrier' concept
2642 * that can be installed to exclude normal IO requests.
2643 */
2644
2645static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2646				   int *skipped)
2647{
2648	struct r1conf *conf = mddev->private;
2649	struct r1bio *r1_bio;
2650	struct bio *bio;
2651	sector_t max_sector, nr_sectors;
2652	int disk = -1;
2653	int i;
2654	int wonly = -1;
2655	int write_targets = 0, read_targets = 0;
2656	sector_t sync_blocks;
2657	int still_degraded = 0;
2658	int good_sectors = RESYNC_SECTORS;
2659	int min_bad = 0; /* number of sectors that are bad in all devices */
2660	int idx = sector_to_idx(sector_nr);
2661	int page_idx = 0;
2662
2663	if (!mempool_initialized(&conf->r1buf_pool))
2664		if (init_resync(conf))
2665			return 0;
2666
2667	max_sector = mddev->dev_sectors;
2668	if (sector_nr >= max_sector) {
2669		/* If we aborted, we need to abort the
2670		 * sync on the 'current' bitmap chunk (there will
2671		 * only be one in raid1 resync.
2672		 * We can find the current addess in mddev->curr_resync
2673		 */
2674		if (mddev->curr_resync < max_sector) /* aborted */
2675			md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2676					   &sync_blocks, 1);
2677		else /* completed sync */
2678			conf->fullsync = 0;
2679
2680		md_bitmap_close_sync(mddev->bitmap);
2681		close_sync(conf);
2682
2683		if (mddev_is_clustered(mddev)) {
2684			conf->cluster_sync_low = 0;
2685			conf->cluster_sync_high = 0;
2686		}
2687		return 0;
2688	}
2689
2690	if (mddev->bitmap == NULL &&
2691	    mddev->recovery_cp == MaxSector &&
2692	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2693	    conf->fullsync == 0) {
2694		*skipped = 1;
2695		return max_sector - sector_nr;
2696	}
2697	/* before building a request, check if we can skip these blocks..
2698	 * This call the bitmap_start_sync doesn't actually record anything
2699	 */
2700	if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2701	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2702		/* We can skip this block, and probably several more */
2703		*skipped = 1;
2704		return sync_blocks;
2705	}
2706
2707	/*
2708	 * If there is non-resync activity waiting for a turn, then let it
2709	 * though before starting on this new sync request.
2710	 */
2711	if (atomic_read(&conf->nr_waiting[idx]))
2712		schedule_timeout_uninterruptible(1);
2713
2714	/* we are incrementing sector_nr below. To be safe, we check against
2715	 * sector_nr + two times RESYNC_SECTORS
2716	 */
2717
2718	md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2719		mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2720
2721
2722	if (raise_barrier(conf, sector_nr))
2723		return 0;
2724
2725	r1_bio = raid1_alloc_init_r1buf(conf);
2726
 
2727	/*
2728	 * If we get a correctably read error during resync or recovery,
2729	 * we might want to read from a different device.  So we
2730	 * flag all drives that could conceivably be read from for READ,
2731	 * and any others (which will be non-In_sync devices) for WRITE.
2732	 * If a read fails, we try reading from something else for which READ
2733	 * is OK.
2734	 */
2735
2736	r1_bio->mddev = mddev;
2737	r1_bio->sector = sector_nr;
2738	r1_bio->state = 0;
2739	set_bit(R1BIO_IsSync, &r1_bio->state);
2740	/* make sure good_sectors won't go across barrier unit boundary */
2741	good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2742
2743	for (i = 0; i < conf->raid_disks * 2; i++) {
2744		struct md_rdev *rdev;
2745		bio = r1_bio->bios[i];
2746
2747		rdev = conf->mirrors[i].rdev;
2748		if (rdev == NULL ||
2749		    test_bit(Faulty, &rdev->flags)) {
2750			if (i < conf->raid_disks)
2751				still_degraded = 1;
2752		} else if (!test_bit(In_sync, &rdev->flags)) {
2753			bio->bi_opf = REQ_OP_WRITE;
2754			bio->bi_end_io = end_sync_write;
2755			write_targets ++;
2756		} else {
2757			/* may need to read from here */
2758			sector_t first_bad = MaxSector;
2759			int bad_sectors;
2760
2761			if (is_badblock(rdev, sector_nr, good_sectors,
2762					&first_bad, &bad_sectors)) {
2763				if (first_bad > sector_nr)
2764					good_sectors = first_bad - sector_nr;
2765				else {
2766					bad_sectors -= (sector_nr - first_bad);
2767					if (min_bad == 0 ||
2768					    min_bad > bad_sectors)
2769						min_bad = bad_sectors;
2770				}
2771			}
2772			if (sector_nr < first_bad) {
2773				if (test_bit(WriteMostly, &rdev->flags)) {
2774					if (wonly < 0)
2775						wonly = i;
2776				} else {
2777					if (disk < 0)
2778						disk = i;
2779				}
2780				bio->bi_opf = REQ_OP_READ;
2781				bio->bi_end_io = end_sync_read;
2782				read_targets++;
2783			} else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2784				test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2785				!test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2786				/*
2787				 * The device is suitable for reading (InSync),
2788				 * but has bad block(s) here. Let's try to correct them,
2789				 * if we are doing resync or repair. Otherwise, leave
2790				 * this device alone for this sync request.
2791				 */
2792				bio->bi_opf = REQ_OP_WRITE;
2793				bio->bi_end_io = end_sync_write;
2794				write_targets++;
2795			}
2796		}
2797		if (rdev && bio->bi_end_io) {
2798			atomic_inc(&rdev->nr_pending);
2799			bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2800			bio_set_dev(bio, rdev->bdev);
2801			if (test_bit(FailFast, &rdev->flags))
2802				bio->bi_opf |= MD_FAILFAST;
2803		}
2804	}
 
2805	if (disk < 0)
2806		disk = wonly;
2807	r1_bio->read_disk = disk;
2808
2809	if (read_targets == 0 && min_bad > 0) {
2810		/* These sectors are bad on all InSync devices, so we
2811		 * need to mark them bad on all write targets
2812		 */
2813		int ok = 1;
2814		for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2815			if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2816				struct md_rdev *rdev = conf->mirrors[i].rdev;
2817				ok = rdev_set_badblocks(rdev, sector_nr,
2818							min_bad, 0
2819					) && ok;
2820			}
2821		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2822		*skipped = 1;
2823		put_buf(r1_bio);
2824
2825		if (!ok) {
2826			/* Cannot record the badblocks, so need to
2827			 * abort the resync.
2828			 * If there are multiple read targets, could just
2829			 * fail the really bad ones ???
2830			 */
2831			conf->recovery_disabled = mddev->recovery_disabled;
2832			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2833			return 0;
2834		} else
2835			return min_bad;
2836
2837	}
2838	if (min_bad > 0 && min_bad < good_sectors) {
2839		/* only resync enough to reach the next bad->good
2840		 * transition */
2841		good_sectors = min_bad;
2842	}
2843
2844	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2845		/* extra read targets are also write targets */
2846		write_targets += read_targets-1;
2847
2848	if (write_targets == 0 || read_targets == 0) {
2849		/* There is nowhere to write, so all non-sync
2850		 * drives must be failed - so we are finished
2851		 */
2852		sector_t rv;
2853		if (min_bad > 0)
2854			max_sector = sector_nr + min_bad;
2855		rv = max_sector - sector_nr;
2856		*skipped = 1;
2857		put_buf(r1_bio);
2858		return rv;
2859	}
2860
2861	if (max_sector > mddev->resync_max)
2862		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2863	if (max_sector > sector_nr + good_sectors)
2864		max_sector = sector_nr + good_sectors;
2865	nr_sectors = 0;
2866	sync_blocks = 0;
2867	do {
2868		struct page *page;
2869		int len = PAGE_SIZE;
2870		if (sector_nr + (len>>9) > max_sector)
2871			len = (max_sector - sector_nr) << 9;
2872		if (len == 0)
2873			break;
2874		if (sync_blocks == 0) {
2875			if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2876						  &sync_blocks, still_degraded) &&
2877			    !conf->fullsync &&
2878			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2879				break;
2880			if ((len >> 9) > sync_blocks)
2881				len = sync_blocks<<9;
2882		}
2883
2884		for (i = 0 ; i < conf->raid_disks * 2; i++) {
2885			struct resync_pages *rp;
2886
2887			bio = r1_bio->bios[i];
2888			rp = get_resync_pages(bio);
2889			if (bio->bi_end_io) {
2890				page = resync_fetch_page(rp, page_idx);
2891
2892				/*
2893				 * won't fail because the vec table is big
2894				 * enough to hold all these pages
2895				 */
2896				__bio_add_page(bio, page, len, 0);
2897			}
2898		}
2899		nr_sectors += len>>9;
2900		sector_nr += len>>9;
2901		sync_blocks -= (len>>9);
2902	} while (++page_idx < RESYNC_PAGES);
2903
2904	r1_bio->sectors = nr_sectors;
2905
2906	if (mddev_is_clustered(mddev) &&
2907			conf->cluster_sync_high < sector_nr + nr_sectors) {
2908		conf->cluster_sync_low = mddev->curr_resync_completed;
2909		conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2910		/* Send resync message */
2911		md_cluster_ops->resync_info_update(mddev,
2912				conf->cluster_sync_low,
2913				conf->cluster_sync_high);
2914	}
2915
2916	/* For a user-requested sync, we read all readable devices and do a
2917	 * compare
2918	 */
2919	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2920		atomic_set(&r1_bio->remaining, read_targets);
2921		for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2922			bio = r1_bio->bios[i];
2923			if (bio->bi_end_io == end_sync_read) {
2924				read_targets--;
2925				md_sync_acct_bio(bio, nr_sectors);
2926				if (read_targets == 1)
2927					bio->bi_opf &= ~MD_FAILFAST;
2928				submit_bio_noacct(bio);
2929			}
2930		}
2931	} else {
2932		atomic_set(&r1_bio->remaining, 1);
2933		bio = r1_bio->bios[r1_bio->read_disk];
2934		md_sync_acct_bio(bio, nr_sectors);
2935		if (read_targets == 1)
2936			bio->bi_opf &= ~MD_FAILFAST;
2937		submit_bio_noacct(bio);
2938	}
2939	return nr_sectors;
2940}
2941
2942static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2943{
2944	if (sectors)
2945		return sectors;
2946
2947	return mddev->dev_sectors;
2948}
2949
2950static struct r1conf *setup_conf(struct mddev *mddev)
2951{
2952	struct r1conf *conf;
2953	int i;
2954	struct raid1_info *disk;
2955	struct md_rdev *rdev;
2956	int err = -ENOMEM;
2957
2958	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2959	if (!conf)
2960		goto abort;
2961
2962	conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2963				   sizeof(atomic_t), GFP_KERNEL);
2964	if (!conf->nr_pending)
2965		goto abort;
2966
2967	conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2968				   sizeof(atomic_t), GFP_KERNEL);
2969	if (!conf->nr_waiting)
2970		goto abort;
2971
2972	conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2973				  sizeof(atomic_t), GFP_KERNEL);
2974	if (!conf->nr_queued)
2975		goto abort;
2976
2977	conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2978				sizeof(atomic_t), GFP_KERNEL);
2979	if (!conf->barrier)
2980		goto abort;
2981
2982	conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2983					    mddev->raid_disks, 2),
2984				GFP_KERNEL);
2985	if (!conf->mirrors)
2986		goto abort;
2987
2988	conf->tmppage = alloc_page(GFP_KERNEL);
2989	if (!conf->tmppage)
2990		goto abort;
2991
2992	conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2993	if (!conf->poolinfo)
2994		goto abort;
2995	conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2996	err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2997			   rbio_pool_free, conf->poolinfo);
2998	if (err)
2999		goto abort;
3000
3001	err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3002	if (err)
3003		goto abort;
3004
3005	conf->poolinfo->mddev = mddev;
3006
3007	err = -EINVAL;
3008	spin_lock_init(&conf->device_lock);
3009	rdev_for_each(rdev, mddev) {
3010		int disk_idx = rdev->raid_disk;
3011		if (disk_idx >= mddev->raid_disks
3012		    || disk_idx < 0)
3013			continue;
3014		if (test_bit(Replacement, &rdev->flags))
3015			disk = conf->mirrors + mddev->raid_disks + disk_idx;
3016		else
3017			disk = conf->mirrors + disk_idx;
3018
3019		if (disk->rdev)
3020			goto abort;
3021		disk->rdev = rdev;
3022		disk->head_position = 0;
3023		disk->seq_start = MaxSector;
3024	}
3025	conf->raid_disks = mddev->raid_disks;
3026	conf->mddev = mddev;
3027	INIT_LIST_HEAD(&conf->retry_list);
3028	INIT_LIST_HEAD(&conf->bio_end_io_list);
3029
3030	spin_lock_init(&conf->resync_lock);
3031	init_waitqueue_head(&conf->wait_barrier);
3032
3033	bio_list_init(&conf->pending_bio_list);
 
3034	conf->recovery_disabled = mddev->recovery_disabled - 1;
3035
3036	err = -EIO;
3037	for (i = 0; i < conf->raid_disks * 2; i++) {
3038
3039		disk = conf->mirrors + i;
3040
3041		if (i < conf->raid_disks &&
3042		    disk[conf->raid_disks].rdev) {
3043			/* This slot has a replacement. */
3044			if (!disk->rdev) {
3045				/* No original, just make the replacement
3046				 * a recovering spare
3047				 */
3048				disk->rdev =
3049					disk[conf->raid_disks].rdev;
3050				disk[conf->raid_disks].rdev = NULL;
3051			} else if (!test_bit(In_sync, &disk->rdev->flags))
3052				/* Original is not in_sync - bad */
3053				goto abort;
3054		}
3055
3056		if (!disk->rdev ||
3057		    !test_bit(In_sync, &disk->rdev->flags)) {
3058			disk->head_position = 0;
3059			if (disk->rdev &&
3060			    (disk->rdev->saved_raid_disk < 0))
3061				conf->fullsync = 1;
3062		}
3063	}
3064
3065	err = -ENOMEM;
3066	rcu_assign_pointer(conf->thread,
3067			   md_register_thread(raid1d, mddev, "raid1"));
3068	if (!conf->thread)
3069		goto abort;
3070
3071	return conf;
3072
3073 abort:
3074	if (conf) {
3075		mempool_exit(&conf->r1bio_pool);
3076		kfree(conf->mirrors);
3077		safe_put_page(conf->tmppage);
3078		kfree(conf->poolinfo);
3079		kfree(conf->nr_pending);
3080		kfree(conf->nr_waiting);
3081		kfree(conf->nr_queued);
3082		kfree(conf->barrier);
3083		bioset_exit(&conf->bio_split);
3084		kfree(conf);
3085	}
3086	return ERR_PTR(err);
3087}
3088
3089static void raid1_free(struct mddev *mddev, void *priv);
3090static int raid1_run(struct mddev *mddev)
3091{
3092	struct r1conf *conf;
3093	int i;
3094	struct md_rdev *rdev;
3095	int ret;
 
3096
3097	if (mddev->level != 1) {
3098		pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3099			mdname(mddev), mddev->level);
3100		return -EIO;
3101	}
3102	if (mddev->reshape_position != MaxSector) {
3103		pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3104			mdname(mddev));
3105		return -EIO;
3106	}
3107
 
3108	/*
3109	 * copy the already verified devices into our private RAID1
3110	 * bookkeeping area. [whatever we allocate in run(),
3111	 * should be freed in raid1_free()]
3112	 */
3113	if (mddev->private == NULL)
3114		conf = setup_conf(mddev);
3115	else
3116		conf = mddev->private;
3117
3118	if (IS_ERR(conf))
3119		return PTR_ERR(conf);
3120
3121	if (mddev->queue)
 
3122		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
 
3123
3124	rdev_for_each(rdev, mddev) {
3125		if (!mddev->gendisk)
3126			continue;
3127		disk_stack_limits(mddev->gendisk, rdev->bdev,
3128				  rdev->data_offset << 9);
 
 
3129	}
3130
3131	mddev->degraded = 0;
3132	for (i = 0; i < conf->raid_disks; i++)
3133		if (conf->mirrors[i].rdev == NULL ||
3134		    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3135		    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3136			mddev->degraded++;
3137	/*
3138	 * RAID1 needs at least one disk in active
3139	 */
3140	if (conf->raid_disks - mddev->degraded < 1) {
3141		md_unregister_thread(mddev, &conf->thread);
3142		ret = -EINVAL;
3143		goto abort;
3144	}
3145
3146	if (conf->raid_disks - mddev->degraded == 1)
3147		mddev->recovery_cp = MaxSector;
3148
3149	if (mddev->recovery_cp != MaxSector)
3150		pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3151			mdname(mddev));
3152	pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3153		mdname(mddev), mddev->raid_disks - mddev->degraded,
3154		mddev->raid_disks);
3155
3156	/*
3157	 * Ok, everything is just fine now
3158	 */
3159	rcu_assign_pointer(mddev->thread, conf->thread);
3160	rcu_assign_pointer(conf->thread, NULL);
3161	mddev->private = conf;
3162	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3163
3164	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3165
 
 
 
 
 
 
 
 
 
3166	ret = md_integrity_register(mddev);
3167	if (ret) {
3168		md_unregister_thread(mddev, &mddev->thread);
3169		goto abort;
3170	}
3171	return 0;
3172
3173abort:
3174	raid1_free(mddev, conf);
3175	return ret;
3176}
3177
3178static void raid1_free(struct mddev *mddev, void *priv)
3179{
3180	struct r1conf *conf = priv;
3181
3182	mempool_exit(&conf->r1bio_pool);
3183	kfree(conf->mirrors);
3184	safe_put_page(conf->tmppage);
3185	kfree(conf->poolinfo);
3186	kfree(conf->nr_pending);
3187	kfree(conf->nr_waiting);
3188	kfree(conf->nr_queued);
3189	kfree(conf->barrier);
3190	bioset_exit(&conf->bio_split);
3191	kfree(conf);
3192}
3193
3194static int raid1_resize(struct mddev *mddev, sector_t sectors)
3195{
3196	/* no resync is happening, and there is enough space
3197	 * on all devices, so we can resize.
3198	 * We need to make sure resync covers any new space.
3199	 * If the array is shrinking we should possibly wait until
3200	 * any io in the removed space completes, but it hardly seems
3201	 * worth it.
3202	 */
3203	sector_t newsize = raid1_size(mddev, sectors, 0);
3204	if (mddev->external_size &&
3205	    mddev->array_sectors > newsize)
3206		return -EINVAL;
3207	if (mddev->bitmap) {
3208		int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3209		if (ret)
3210			return ret;
3211	}
3212	md_set_array_sectors(mddev, newsize);
3213	if (sectors > mddev->dev_sectors &&
3214	    mddev->recovery_cp > mddev->dev_sectors) {
3215		mddev->recovery_cp = mddev->dev_sectors;
3216		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3217	}
3218	mddev->dev_sectors = sectors;
3219	mddev->resync_max_sectors = sectors;
3220	return 0;
3221}
3222
3223static int raid1_reshape(struct mddev *mddev)
3224{
3225	/* We need to:
3226	 * 1/ resize the r1bio_pool
3227	 * 2/ resize conf->mirrors
3228	 *
3229	 * We allocate a new r1bio_pool if we can.
3230	 * Then raise a device barrier and wait until all IO stops.
3231	 * Then resize conf->mirrors and swap in the new r1bio pool.
3232	 *
3233	 * At the same time, we "pack" the devices so that all the missing
3234	 * devices have the higher raid_disk numbers.
3235	 */
3236	mempool_t newpool, oldpool;
3237	struct pool_info *newpoolinfo;
3238	struct raid1_info *newmirrors;
3239	struct r1conf *conf = mddev->private;
3240	int cnt, raid_disks;
3241	unsigned long flags;
3242	int d, d2;
3243	int ret;
3244
3245	memset(&newpool, 0, sizeof(newpool));
3246	memset(&oldpool, 0, sizeof(oldpool));
3247
3248	/* Cannot change chunk_size, layout, or level */
3249	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3250	    mddev->layout != mddev->new_layout ||
3251	    mddev->level != mddev->new_level) {
3252		mddev->new_chunk_sectors = mddev->chunk_sectors;
3253		mddev->new_layout = mddev->layout;
3254		mddev->new_level = mddev->level;
3255		return -EINVAL;
3256	}
3257
3258	if (!mddev_is_clustered(mddev))
3259		md_allow_write(mddev);
3260
3261	raid_disks = mddev->raid_disks + mddev->delta_disks;
3262
3263	if (raid_disks < conf->raid_disks) {
3264		cnt=0;
3265		for (d= 0; d < conf->raid_disks; d++)
3266			if (conf->mirrors[d].rdev)
3267				cnt++;
3268		if (cnt > raid_disks)
3269			return -EBUSY;
3270	}
3271
3272	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3273	if (!newpoolinfo)
3274		return -ENOMEM;
3275	newpoolinfo->mddev = mddev;
3276	newpoolinfo->raid_disks = raid_disks * 2;
3277
3278	ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3279			   rbio_pool_free, newpoolinfo);
3280	if (ret) {
3281		kfree(newpoolinfo);
3282		return ret;
3283	}
3284	newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3285					 raid_disks, 2),
3286			     GFP_KERNEL);
3287	if (!newmirrors) {
3288		kfree(newpoolinfo);
3289		mempool_exit(&newpool);
3290		return -ENOMEM;
3291	}
3292
3293	freeze_array(conf, 0);
3294
3295	/* ok, everything is stopped */
3296	oldpool = conf->r1bio_pool;
3297	conf->r1bio_pool = newpool;
3298
3299	for (d = d2 = 0; d < conf->raid_disks; d++) {
3300		struct md_rdev *rdev = conf->mirrors[d].rdev;
3301		if (rdev && rdev->raid_disk != d2) {
3302			sysfs_unlink_rdev(mddev, rdev);
3303			rdev->raid_disk = d2;
3304			sysfs_unlink_rdev(mddev, rdev);
3305			if (sysfs_link_rdev(mddev, rdev))
3306				pr_warn("md/raid1:%s: cannot register rd%d\n",
3307					mdname(mddev), rdev->raid_disk);
3308		}
3309		if (rdev)
3310			newmirrors[d2++].rdev = rdev;
3311	}
3312	kfree(conf->mirrors);
3313	conf->mirrors = newmirrors;
3314	kfree(conf->poolinfo);
3315	conf->poolinfo = newpoolinfo;
3316
3317	spin_lock_irqsave(&conf->device_lock, flags);
3318	mddev->degraded += (raid_disks - conf->raid_disks);
3319	spin_unlock_irqrestore(&conf->device_lock, flags);
3320	conf->raid_disks = mddev->raid_disks = raid_disks;
3321	mddev->delta_disks = 0;
3322
3323	unfreeze_array(conf);
3324
3325	set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3326	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3327	md_wakeup_thread(mddev->thread);
3328
3329	mempool_exit(&oldpool);
3330	return 0;
3331}
3332
3333static void raid1_quiesce(struct mddev *mddev, int quiesce)
3334{
3335	struct r1conf *conf = mddev->private;
3336
3337	if (quiesce)
3338		freeze_array(conf, 0);
3339	else
3340		unfreeze_array(conf);
3341}
3342
3343static void *raid1_takeover(struct mddev *mddev)
3344{
3345	/* raid1 can take over:
3346	 *  raid5 with 2 devices, any layout or chunk size
3347	 */
3348	if (mddev->level == 5 && mddev->raid_disks == 2) {
3349		struct r1conf *conf;
3350		mddev->new_level = 1;
3351		mddev->new_layout = 0;
3352		mddev->new_chunk_sectors = 0;
3353		conf = setup_conf(mddev);
3354		if (!IS_ERR(conf)) {
3355			/* Array must appear to be quiesced */
3356			conf->array_frozen = 1;
3357			mddev_clear_unsupported_flags(mddev,
3358				UNSUPPORTED_MDDEV_FLAGS);
3359		}
3360		return conf;
3361	}
3362	return ERR_PTR(-EINVAL);
3363}
3364
3365static struct md_personality raid1_personality =
3366{
3367	.name		= "raid1",
3368	.level		= 1,
3369	.owner		= THIS_MODULE,
3370	.make_request	= raid1_make_request,
3371	.run		= raid1_run,
3372	.free		= raid1_free,
3373	.status		= raid1_status,
3374	.error_handler	= raid1_error,
3375	.hot_add_disk	= raid1_add_disk,
3376	.hot_remove_disk= raid1_remove_disk,
3377	.spare_active	= raid1_spare_active,
3378	.sync_request	= raid1_sync_request,
3379	.resize		= raid1_resize,
3380	.size		= raid1_size,
3381	.check_reshape	= raid1_reshape,
3382	.quiesce	= raid1_quiesce,
3383	.takeover	= raid1_takeover,
 
3384};
3385
3386static int __init raid_init(void)
3387{
3388	return register_md_personality(&raid1_personality);
3389}
3390
3391static void raid_exit(void)
3392{
3393	unregister_md_personality(&raid1_personality);
3394}
3395
3396module_init(raid_init);
3397module_exit(raid_exit);
3398MODULE_LICENSE("GPL");
3399MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3400MODULE_ALIAS("md-personality-3"); /* RAID1 */
3401MODULE_ALIAS("md-raid1");
3402MODULE_ALIAS("md-level-1");
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);