Linux Audio

Check our new training course

Loading...
v4.6
 
   1/*
   2 * raid1.c : Multiple Devices driver for Linux
   3 *
   4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5 *
   6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7 *
   8 * RAID-1 management functions.
   9 *
  10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11 *
  12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14 *
  15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16 * bitmapped intelligence in resync:
  17 *
  18 *      - bitmap marked during normal i/o
  19 *      - bitmap used to skip nondirty blocks during sync
  20 *
  21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22 * - persistent bitmap code
  23 *
  24 * This program is free software; you can redistribute it and/or modify
  25 * it under the terms of the GNU General Public License as published by
  26 * the Free Software Foundation; either version 2, or (at your option)
  27 * any later version.
  28 *
  29 * You should have received a copy of the GNU General Public License
  30 * (for example /usr/src/linux/COPYING); if not, write to the Free
  31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32 */
  33
  34#include <linux/slab.h>
  35#include <linux/delay.h>
  36#include <linux/blkdev.h>
  37#include <linux/module.h>
  38#include <linux/seq_file.h>
  39#include <linux/ratelimit.h>
 
 
 
 
  40#include "md.h"
  41#include "raid1.h"
  42#include "bitmap.h"
  43
  44/*
  45 * Number of guaranteed r1bios in case of extreme VM load:
  46 */
  47#define	NR_RAID1_BIOS 256
 
  48
  49/* when we get a read error on a read-only array, we redirect to another
  50 * device without failing the first device, or trying to over-write to
  51 * correct the read error.  To keep track of bad blocks on a per-bio
  52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  53 */
  54#define IO_BLOCKED ((struct bio *)1)
  55/* When we successfully write to a known bad-block, we need to remove the
  56 * bad-block marking which must be done from process context.  So we record
  57 * the success by setting devs[n].bio to IO_MADE_GOOD
  58 */
  59#define IO_MADE_GOOD ((struct bio *)2)
  60
  61#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
 
  62
  63/* When there are this many requests queue to be written by
  64 * the raid1 thread, we become 'congested' to provide back-pressure
  65 * for writeback.
  66 */
  67static int max_queued_requests = 1024;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  68
  69static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
  70			  sector_t bi_sector);
  71static void lower_barrier(struct r1conf *conf);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  72
  73static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  74{
  75	struct pool_info *pi = data;
  76	int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  77
  78	/* allocate a r1bio with room for raid_disks entries in the bios array */
  79	return kzalloc(size, gfp_flags);
  80}
  81
  82static void r1bio_pool_free(void *r1_bio, void *data)
  83{
  84	kfree(r1_bio);
  85}
  86
  87#define RESYNC_BLOCK_SIZE (64*1024)
  88#define RESYNC_DEPTH 32
  89#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  90#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  91#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
  92#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
  93#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
  94#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
  95#define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
  96
  97static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  98{
  99	struct pool_info *pi = data;
 100	struct r1bio *r1_bio;
 101	struct bio *bio;
 102	int need_pages;
 103	int i, j;
 
 104
 105	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 106	if (!r1_bio)
 107		return NULL;
 108
 
 
 
 
 
 109	/*
 110	 * Allocate bios : 1 for reading, n-1 for writing
 111	 */
 112	for (j = pi->raid_disks ; j-- ; ) {
 113		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 114		if (!bio)
 115			goto out_free_bio;
 116		r1_bio->bios[j] = bio;
 117	}
 118	/*
 119	 * Allocate RESYNC_PAGES data pages and attach them to
 120	 * the first bio.
 121	 * If this is a user-requested check/repair, allocate
 122	 * RESYNC_PAGES for each bio.
 123	 */
 124	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 125		need_pages = pi->raid_disks;
 126	else
 127		need_pages = 1;
 128	for (j = 0; j < need_pages; j++) {
 
 
 129		bio = r1_bio->bios[j];
 130		bio->bi_vcnt = RESYNC_PAGES;
 131
 132		if (bio_alloc_pages(bio, gfp_flags))
 133			goto out_free_pages;
 134	}
 135	/* If not user-requests, copy the page pointers to all bios */
 136	if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 137		for (i=0; i<RESYNC_PAGES ; i++)
 138			for (j=1; j<pi->raid_disks; j++)
 139				r1_bio->bios[j]->bi_io_vec[i].bv_page =
 140					r1_bio->bios[0]->bi_io_vec[i].bv_page;
 
 141	}
 142
 143	r1_bio->master_bio = NULL;
 144
 145	return r1_bio;
 146
 147out_free_pages:
 148	while (--j >= 0) {
 149		struct bio_vec *bv;
 150
 151		bio_for_each_segment_all(bv, r1_bio->bios[j], i)
 152			__free_page(bv->bv_page);
 153	}
 154
 155out_free_bio:
 156	while (++j < pi->raid_disks)
 157		bio_put(r1_bio->bios[j]);
 158	r1bio_pool_free(r1_bio, data);
 
 
 
 159	return NULL;
 160}
 161
 162static void r1buf_pool_free(void *__r1_bio, void *data)
 163{
 164	struct pool_info *pi = data;
 165	int i,j;
 166	struct r1bio *r1bio = __r1_bio;
 
 167
 168	for (i = 0; i < RESYNC_PAGES; i++)
 169		for (j = pi->raid_disks; j-- ;) {
 170			if (j == 0 ||
 171			    r1bio->bios[j]->bi_io_vec[i].bv_page !=
 172			    r1bio->bios[0]->bi_io_vec[i].bv_page)
 173				safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 174		}
 175	for (i=0 ; i < pi->raid_disks; i++)
 176		bio_put(r1bio->bios[i]);
 
 177
 178	r1bio_pool_free(r1bio, data);
 
 
 
 179}
 180
 181static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 182{
 183	int i;
 184
 185	for (i = 0; i < conf->raid_disks * 2; i++) {
 186		struct bio **bio = r1_bio->bios + i;
 187		if (!BIO_SPECIAL(*bio))
 188			bio_put(*bio);
 189		*bio = NULL;
 190	}
 191}
 192
 193static void free_r1bio(struct r1bio *r1_bio)
 194{
 195	struct r1conf *conf = r1_bio->mddev->private;
 196
 197	put_all_bios(conf, r1_bio);
 198	mempool_free(r1_bio, conf->r1bio_pool);
 199}
 200
 201static void put_buf(struct r1bio *r1_bio)
 202{
 203	struct r1conf *conf = r1_bio->mddev->private;
 
 204	int i;
 205
 206	for (i = 0; i < conf->raid_disks * 2; i++) {
 207		struct bio *bio = r1_bio->bios[i];
 208		if (bio->bi_end_io)
 209			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 210	}
 211
 212	mempool_free(r1_bio, conf->r1buf_pool);
 213
 214	lower_barrier(conf);
 215}
 216
 217static void reschedule_retry(struct r1bio *r1_bio)
 218{
 219	unsigned long flags;
 220	struct mddev *mddev = r1_bio->mddev;
 221	struct r1conf *conf = mddev->private;
 
 222
 
 223	spin_lock_irqsave(&conf->device_lock, flags);
 224	list_add(&r1_bio->retry_list, &conf->retry_list);
 225	conf->nr_queued ++;
 226	spin_unlock_irqrestore(&conf->device_lock, flags);
 227
 228	wake_up(&conf->wait_barrier);
 229	md_wakeup_thread(mddev->thread);
 230}
 231
 232/*
 233 * raid_end_bio_io() is called when we have finished servicing a mirrored
 234 * operation and are ready to return a success/failure code to the buffer
 235 * cache layer.
 236 */
 237static void call_bio_endio(struct r1bio *r1_bio)
 238{
 239	struct bio *bio = r1_bio->master_bio;
 240	int done;
 241	struct r1conf *conf = r1_bio->mddev->private;
 242	sector_t start_next_window = r1_bio->start_next_window;
 243	sector_t bi_sector = bio->bi_iter.bi_sector;
 244
 245	if (bio->bi_phys_segments) {
 246		unsigned long flags;
 247		spin_lock_irqsave(&conf->device_lock, flags);
 248		bio->bi_phys_segments--;
 249		done = (bio->bi_phys_segments == 0);
 250		spin_unlock_irqrestore(&conf->device_lock, flags);
 251		/*
 252		 * make_request() might be waiting for
 253		 * bi_phys_segments to decrease
 254		 */
 255		wake_up(&conf->wait_barrier);
 256	} else
 257		done = 1;
 258
 259	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 260		bio->bi_error = -EIO;
 261
 262	if (done) {
 263		bio_endio(bio);
 264		/*
 265		 * Wake up any possible resync thread that waits for the device
 266		 * to go idle.
 267		 */
 268		allow_barrier(conf, start_next_window, bi_sector);
 269	}
 270}
 271
 272static void raid_end_bio_io(struct r1bio *r1_bio)
 273{
 274	struct bio *bio = r1_bio->master_bio;
 
 275
 276	/* if nobody has done the final endio yet, do it now */
 277	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 278		pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 279			 (bio_data_dir(bio) == WRITE) ? "write" : "read",
 280			 (unsigned long long) bio->bi_iter.bi_sector,
 281			 (unsigned long long) bio_end_sector(bio) - 1);
 282
 283		call_bio_endio(r1_bio);
 284	}
 
 
 
 
 
 
 285	free_r1bio(r1_bio);
 286}
 287
 288/*
 289 * Update disk head position estimator based on IRQ completion info.
 290 */
 291static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 292{
 293	struct r1conf *conf = r1_bio->mddev->private;
 294
 295	conf->mirrors[disk].head_position =
 296		r1_bio->sector + (r1_bio->sectors);
 297}
 298
 299/*
 300 * Find the disk number which triggered given bio
 301 */
 302static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 303{
 304	int mirror;
 305	struct r1conf *conf = r1_bio->mddev->private;
 306	int raid_disks = conf->raid_disks;
 307
 308	for (mirror = 0; mirror < raid_disks * 2; mirror++)
 309		if (r1_bio->bios[mirror] == bio)
 310			break;
 311
 312	BUG_ON(mirror == raid_disks * 2);
 313	update_head_pos(mirror, r1_bio);
 314
 315	return mirror;
 316}
 317
 318static void raid1_end_read_request(struct bio *bio)
 319{
 320	int uptodate = !bio->bi_error;
 321	struct r1bio *r1_bio = bio->bi_private;
 322	int mirror;
 323	struct r1conf *conf = r1_bio->mddev->private;
 
 324
 325	mirror = r1_bio->read_disk;
 326	/*
 327	 * this branch is our 'one mirror IO has finished' event handler:
 328	 */
 329	update_head_pos(mirror, r1_bio);
 330
 331	if (uptodate)
 332		set_bit(R1BIO_Uptodate, &r1_bio->state);
 
 
 
 
 
 333	else {
 334		/* If all other devices have failed, we want to return
 335		 * the error upwards rather than fail the last device.
 336		 * Here we redefine "uptodate" to mean "Don't want to retry"
 337		 */
 338		unsigned long flags;
 339		spin_lock_irqsave(&conf->device_lock, flags);
 340		if (r1_bio->mddev->degraded == conf->raid_disks ||
 341		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 342		     test_bit(In_sync, &conf->mirrors[mirror].rdev->flags)))
 343			uptodate = 1;
 344		spin_unlock_irqrestore(&conf->device_lock, flags);
 345	}
 346
 347	if (uptodate) {
 348		raid_end_bio_io(r1_bio);
 349		rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 350	} else {
 351		/*
 352		 * oops, read error:
 353		 */
 354		char b[BDEVNAME_SIZE];
 355		printk_ratelimited(
 356			KERN_ERR "md/raid1:%s: %s: "
 357			"rescheduling sector %llu\n",
 358			mdname(conf->mddev),
 359			bdevname(conf->mirrors[mirror].rdev->bdev,
 360				 b),
 361			(unsigned long long)r1_bio->sector);
 362		set_bit(R1BIO_ReadError, &r1_bio->state);
 363		reschedule_retry(r1_bio);
 364		/* don't drop the reference on read_disk yet */
 365	}
 366}
 367
 368static void close_write(struct r1bio *r1_bio)
 369{
 370	/* it really is the end of this request */
 371	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 372		/* free extra copy of the data pages */
 373		int i = r1_bio->behind_page_count;
 374		while (i--)
 375			safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 376		kfree(r1_bio->behind_bvecs);
 377		r1_bio->behind_bvecs = NULL;
 378	}
 379	/* clear the bitmap if all writes complete successfully */
 380	bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 381			r1_bio->sectors,
 382			!test_bit(R1BIO_Degraded, &r1_bio->state),
 383			test_bit(R1BIO_BehindIO, &r1_bio->state));
 384	md_write_end(r1_bio->mddev);
 385}
 386
 387static void r1_bio_write_done(struct r1bio *r1_bio)
 388{
 389	if (!atomic_dec_and_test(&r1_bio->remaining))
 390		return;
 391
 392	if (test_bit(R1BIO_WriteError, &r1_bio->state))
 393		reschedule_retry(r1_bio);
 394	else {
 395		close_write(r1_bio);
 396		if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 397			reschedule_retry(r1_bio);
 398		else
 399			raid_end_bio_io(r1_bio);
 400	}
 401}
 402
 403static void raid1_end_write_request(struct bio *bio)
 404{
 405	struct r1bio *r1_bio = bio->bi_private;
 406	int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 407	struct r1conf *conf = r1_bio->mddev->private;
 408	struct bio *to_put = NULL;
 
 
 
 
 
 409
 410	mirror = find_bio_disk(r1_bio, bio);
 411
 412	/*
 413	 * 'one mirror IO has finished' event handler:
 414	 */
 415	if (bio->bi_error) {
 416		set_bit(WriteErrorSeen,
 417			&conf->mirrors[mirror].rdev->flags);
 418		if (!test_and_set_bit(WantReplacement,
 419				      &conf->mirrors[mirror].rdev->flags))
 420			set_bit(MD_RECOVERY_NEEDED, &
 421				conf->mddev->recovery);
 422
 423		set_bit(R1BIO_WriteError, &r1_bio->state);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 424	} else {
 425		/*
 426		 * Set R1BIO_Uptodate in our master bio, so that we
 427		 * will return a good error code for to the higher
 428		 * levels even if IO on some other mirrored buffer
 429		 * fails.
 430		 *
 431		 * The 'master' represents the composite IO operation
 432		 * to user-side. So if something waits for IO, then it
 433		 * will wait for the 'master' bio.
 434		 */
 435		sector_t first_bad;
 436		int bad_sectors;
 437
 438		r1_bio->bios[mirror] = NULL;
 439		to_put = bio;
 440		/*
 441		 * Do not set R1BIO_Uptodate if the current device is
 442		 * rebuilding or Faulty. This is because we cannot use
 443		 * such device for properly reading the data back (we could
 444		 * potentially use it, if the current write would have felt
 445		 * before rdev->recovery_offset, but for simplicity we don't
 446		 * check this here.
 447		 */
 448		if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
 449		    !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
 450			set_bit(R1BIO_Uptodate, &r1_bio->state);
 451
 452		/* Maybe we can clear some bad blocks. */
 453		if (is_badblock(conf->mirrors[mirror].rdev,
 454				r1_bio->sector, r1_bio->sectors,
 455				&first_bad, &bad_sectors)) {
 456			r1_bio->bios[mirror] = IO_MADE_GOOD;
 457			set_bit(R1BIO_MadeGood, &r1_bio->state);
 458		}
 459	}
 460
 461	if (behind) {
 462		if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 
 
 463			atomic_dec(&r1_bio->behind_remaining);
 464
 465		/*
 466		 * In behind mode, we ACK the master bio once the I/O
 467		 * has safely reached all non-writemostly
 468		 * disks. Setting the Returned bit ensures that this
 469		 * gets done only once -- we don't ever want to return
 470		 * -EIO here, instead we'll wait
 471		 */
 472		if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 473		    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 474			/* Maybe we can return now */
 475			if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 476				struct bio *mbio = r1_bio->master_bio;
 477				pr_debug("raid1: behind end write sectors"
 478					 " %llu-%llu\n",
 479					 (unsigned long long) mbio->bi_iter.bi_sector,
 480					 (unsigned long long) bio_end_sector(mbio) - 1);
 481				call_bio_endio(r1_bio);
 482			}
 483		}
 484	}
 
 485	if (r1_bio->bios[mirror] == NULL)
 486		rdev_dec_pending(conf->mirrors[mirror].rdev,
 487				 conf->mddev);
 488
 489	/*
 490	 * Let's see if all mirrored write operations have finished
 491	 * already.
 492	 */
 493	r1_bio_write_done(r1_bio);
 494
 495	if (to_put)
 496		bio_put(to_put);
 497}
 498
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 499/*
 500 * This routine returns the disk from which the requested read should
 501 * be done. There is a per-array 'next expected sequential IO' sector
 502 * number - if this matches on the next IO then we use the last disk.
 503 * There is also a per-disk 'last know head position' sector that is
 504 * maintained from IRQ contexts, both the normal and the resync IO
 505 * completion handlers update this position correctly. If there is no
 506 * perfect sequential match then we pick the disk whose head is closest.
 507 *
 508 * If there are 2 mirrors in the same 2 devices, performance degrades
 509 * because position is mirror, not device based.
 510 *
 511 * The rdev for the device selected will have nr_pending incremented.
 512 */
 513static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 514{
 515	const sector_t this_sector = r1_bio->sector;
 516	int sectors;
 517	int best_good_sectors;
 518	int best_disk, best_dist_disk, best_pending_disk;
 519	int has_nonrot_disk;
 520	int disk;
 521	sector_t best_dist;
 522	unsigned int min_pending;
 523	struct md_rdev *rdev;
 524	int choose_first;
 525	int choose_next_idle;
 526
 527	rcu_read_lock();
 528	/*
 529	 * Check if we can balance. We can balance on the whole
 530	 * device if no resync is going on, or below the resync window.
 531	 * We take the first readable disk when above the resync window.
 532	 */
 533 retry:
 534	sectors = r1_bio->sectors;
 535	best_disk = -1;
 536	best_dist_disk = -1;
 537	best_dist = MaxSector;
 538	best_pending_disk = -1;
 539	min_pending = UINT_MAX;
 540	best_good_sectors = 0;
 541	has_nonrot_disk = 0;
 542	choose_next_idle = 0;
 
 543
 544	if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 545	    (mddev_is_clustered(conf->mddev) &&
 546	    md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 547		    this_sector + sectors)))
 548		choose_first = 1;
 549	else
 550		choose_first = 0;
 551
 552	for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 553		sector_t dist;
 554		sector_t first_bad;
 555		int bad_sectors;
 556		unsigned int pending;
 557		bool nonrot;
 558
 559		rdev = rcu_dereference(conf->mirrors[disk].rdev);
 560		if (r1_bio->bios[disk] == IO_BLOCKED
 561		    || rdev == NULL
 562		    || test_bit(Faulty, &rdev->flags))
 563			continue;
 564		if (!test_bit(In_sync, &rdev->flags) &&
 565		    rdev->recovery_offset < this_sector + sectors)
 566			continue;
 567		if (test_bit(WriteMostly, &rdev->flags)) {
 568			/* Don't balance among write-mostly, just
 569			 * use the first as a last resort */
 570			if (best_dist_disk < 0) {
 571				if (is_badblock(rdev, this_sector, sectors,
 572						&first_bad, &bad_sectors)) {
 573					if (first_bad <= this_sector)
 574						/* Cannot use this */
 575						continue;
 576					best_good_sectors = first_bad - this_sector;
 577				} else
 578					best_good_sectors = sectors;
 579				best_dist_disk = disk;
 580				best_pending_disk = disk;
 581			}
 582			continue;
 583		}
 584		/* This is a reasonable device to use.  It might
 585		 * even be best.
 586		 */
 587		if (is_badblock(rdev, this_sector, sectors,
 588				&first_bad, &bad_sectors)) {
 589			if (best_dist < MaxSector)
 590				/* already have a better device */
 591				continue;
 592			if (first_bad <= this_sector) {
 593				/* cannot read here. If this is the 'primary'
 594				 * device, then we must not read beyond
 595				 * bad_sectors from another device..
 596				 */
 597				bad_sectors -= (this_sector - first_bad);
 598				if (choose_first && sectors > bad_sectors)
 599					sectors = bad_sectors;
 600				if (best_good_sectors > sectors)
 601					best_good_sectors = sectors;
 602
 603			} else {
 604				sector_t good_sectors = first_bad - this_sector;
 605				if (good_sectors > best_good_sectors) {
 606					best_good_sectors = good_sectors;
 607					best_disk = disk;
 608				}
 609				if (choose_first)
 610					break;
 611			}
 612			continue;
 613		} else
 
 
 614			best_good_sectors = sectors;
 
 
 
 
 
 615
 616		nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 617		has_nonrot_disk |= nonrot;
 618		pending = atomic_read(&rdev->nr_pending);
 619		dist = abs(this_sector - conf->mirrors[disk].head_position);
 620		if (choose_first) {
 621			best_disk = disk;
 622			break;
 623		}
 624		/* Don't change to another disk for sequential reads */
 625		if (conf->mirrors[disk].next_seq_sect == this_sector
 626		    || dist == 0) {
 627			int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 628			struct raid1_info *mirror = &conf->mirrors[disk];
 629
 630			best_disk = disk;
 631			/*
 632			 * If buffered sequential IO size exceeds optimal
 633			 * iosize, check if there is idle disk. If yes, choose
 634			 * the idle disk. read_balance could already choose an
 635			 * idle disk before noticing it's a sequential IO in
 636			 * this disk. This doesn't matter because this disk
 637			 * will idle, next time it will be utilized after the
 638			 * first disk has IO size exceeds optimal iosize. In
 639			 * this way, iosize of the first disk will be optimal
 640			 * iosize at least. iosize of the second disk might be
 641			 * small, but not a big deal since when the second disk
 642			 * starts IO, the first disk is likely still busy.
 643			 */
 644			if (nonrot && opt_iosize > 0 &&
 645			    mirror->seq_start != MaxSector &&
 646			    mirror->next_seq_sect > opt_iosize &&
 647			    mirror->next_seq_sect - opt_iosize >=
 648			    mirror->seq_start) {
 649				choose_next_idle = 1;
 650				continue;
 651			}
 652			break;
 653		}
 654		/* If device is idle, use it */
 655		if (pending == 0) {
 656			best_disk = disk;
 657			break;
 658		}
 659
 660		if (choose_next_idle)
 661			continue;
 662
 663		if (min_pending > pending) {
 664			min_pending = pending;
 665			best_pending_disk = disk;
 666		}
 667
 668		if (dist < best_dist) {
 669			best_dist = dist;
 670			best_dist_disk = disk;
 671		}
 672	}
 673
 674	/*
 675	 * If all disks are rotational, choose the closest disk. If any disk is
 676	 * non-rotational, choose the disk with less pending request even the
 677	 * disk is rotational, which might/might not be optimal for raids with
 678	 * mixed ratation/non-rotational disks depending on workload.
 679	 */
 680	if (best_disk == -1) {
 681		if (has_nonrot_disk)
 682			best_disk = best_pending_disk;
 683		else
 684			best_disk = best_dist_disk;
 685	}
 686
 687	if (best_disk >= 0) {
 688		rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 689		if (!rdev)
 690			goto retry;
 691		atomic_inc(&rdev->nr_pending);
 692		if (test_bit(Faulty, &rdev->flags)) {
 693			/* cannot risk returning a device that failed
 694			 * before we inc'ed nr_pending
 695			 */
 696			rdev_dec_pending(rdev, conf->mddev);
 697			goto retry;
 698		}
 699		sectors = best_good_sectors;
 700
 701		if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 702			conf->mirrors[best_disk].seq_start = this_sector;
 703
 704		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 705	}
 706	rcu_read_unlock();
 707	*max_sectors = sectors;
 708
 709	return best_disk;
 710}
 711
 712static int raid1_congested(struct mddev *mddev, int bits)
 713{
 714	struct r1conf *conf = mddev->private;
 715	int i, ret = 0;
 716
 717	if ((bits & (1 << WB_async_congested)) &&
 718	    conf->pending_count >= max_queued_requests)
 719		return 1;
 720
 721	rcu_read_lock();
 722	for (i = 0; i < conf->raid_disks * 2; i++) {
 723		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 724		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 725			struct request_queue *q = bdev_get_queue(rdev->bdev);
 726
 727			BUG_ON(!q);
 728
 729			/* Note the '|| 1' - when read_balance prefers
 730			 * non-congested targets, it can be removed
 731			 */
 732			if ((bits & (1 << WB_async_congested)) || 1)
 733				ret |= bdi_congested(&q->backing_dev_info, bits);
 734			else
 735				ret &= bdi_congested(&q->backing_dev_info, bits);
 736		}
 
 
 
 
 
 
 
 737	}
 738	rcu_read_unlock();
 739	return ret;
 740}
 741
 742static void flush_pending_writes(struct r1conf *conf)
 743{
 744	/* Any writes that have been queued but are awaiting
 745	 * bitmap updates get flushed here.
 746	 */
 747	spin_lock_irq(&conf->device_lock);
 748
 749	if (conf->pending_bio_list.head) {
 
 750		struct bio *bio;
 
 751		bio = bio_list_get(&conf->pending_bio_list);
 752		conf->pending_count = 0;
 753		spin_unlock_irq(&conf->device_lock);
 754		/* flush any pending bitmap writes to
 755		 * disk before proceeding w/ I/O */
 756		bitmap_unplug(conf->mddev->bitmap);
 757		wake_up(&conf->wait_barrier);
 758
 759		while (bio) { /* submit pending writes */
 760			struct bio *next = bio->bi_next;
 761			bio->bi_next = NULL;
 762			if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 763			    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 764				/* Just ignore it */
 765				bio_endio(bio);
 766			else
 767				generic_make_request(bio);
 768			bio = next;
 769		}
 
 
 770	} else
 771		spin_unlock_irq(&conf->device_lock);
 772}
 773
 774/* Barriers....
 775 * Sometimes we need to suspend IO while we do something else,
 776 * either some resync/recovery, or reconfigure the array.
 777 * To do this we raise a 'barrier'.
 778 * The 'barrier' is a counter that can be raised multiple times
 779 * to count how many activities are happening which preclude
 780 * normal IO.
 781 * We can only raise the barrier if there is no pending IO.
 782 * i.e. if nr_pending == 0.
 783 * We choose only to raise the barrier if no-one is waiting for the
 784 * barrier to go down.  This means that as soon as an IO request
 785 * is ready, no other operations which require a barrier will start
 786 * until the IO request has had a chance.
 787 *
 788 * So: regular IO calls 'wait_barrier'.  When that returns there
 789 *    is no backgroup IO happening,  It must arrange to call
 790 *    allow_barrier when it has finished its IO.
 791 * backgroup IO calls must call raise_barrier.  Once that returns
 792 *    there is no normal IO happeing.  It must arrange to call
 793 *    lower_barrier when the particular background IO completes.
 
 
 
 794 */
 795static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
 796{
 
 
 797	spin_lock_irq(&conf->resync_lock);
 798
 799	/* Wait until no block IO is waiting */
 800	wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 
 801			    conf->resync_lock);
 802
 803	/* block any new IO from starting */
 804	conf->barrier++;
 805	conf->next_resync = sector_nr;
 
 
 
 
 
 
 
 
 806
 807	/* For these conditions we must wait:
 808	 * A: while the array is in frozen state
 809	 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
 810	 *    the max count which allowed.
 811	 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
 812	 *    next resync will reach to the window which normal bios are
 813	 *    handling.
 814	 * D: while there are any active requests in the current window.
 815	 */
 816	wait_event_lock_irq(conf->wait_barrier,
 817			    !conf->array_frozen &&
 818			    conf->barrier < RESYNC_DEPTH &&
 819			    conf->current_window_requests == 0 &&
 820			    (conf->start_next_window >=
 821			     conf->next_resync + RESYNC_SECTORS),
 822			    conf->resync_lock);
 823
 824	conf->nr_pending++;
 
 
 
 
 
 
 
 825	spin_unlock_irq(&conf->resync_lock);
 
 
 826}
 827
 828static void lower_barrier(struct r1conf *conf)
 829{
 830	unsigned long flags;
 831	BUG_ON(conf->barrier <= 0);
 832	spin_lock_irqsave(&conf->resync_lock, flags);
 833	conf->barrier--;
 834	conf->nr_pending--;
 835	spin_unlock_irqrestore(&conf->resync_lock, flags);
 836	wake_up(&conf->wait_barrier);
 837}
 838
 839static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
 840{
 841	bool wait = false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 842
 843	if (conf->array_frozen || !bio)
 844		wait = true;
 845	else if (conf->barrier && bio_data_dir(bio) == WRITE) {
 846		if ((conf->mddev->curr_resync_completed
 847		     >= bio_end_sector(bio)) ||
 848		    (conf->next_resync + NEXT_NORMALIO_DISTANCE
 849		     <= bio->bi_iter.bi_sector))
 850			wait = false;
 851		else
 852			wait = true;
 853	}
 
 854
 855	return wait;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 856}
 857
 858static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
 859{
 860	sector_t sector = 0;
 861
 862	spin_lock_irq(&conf->resync_lock);
 863	if (need_to_wait_for_sync(conf, bio)) {
 864		conf->nr_waiting++;
 865		/* Wait for the barrier to drop.
 866		 * However if there are already pending
 867		 * requests (preventing the barrier from
 868		 * rising completely), and the
 869		 * per-process bio queue isn't empty,
 870		 * then don't wait, as we need to empty
 871		 * that queue to allow conf->start_next_window
 872		 * to increase.
 873		 */
 874		wait_event_lock_irq(conf->wait_barrier,
 875				    !conf->array_frozen &&
 876				    (!conf->barrier ||
 877				     ((conf->start_next_window <
 878				       conf->next_resync + RESYNC_SECTORS) &&
 879				      current->bio_list &&
 880				      !bio_list_empty(current->bio_list))),
 881				    conf->resync_lock);
 882		conf->nr_waiting--;
 883	}
 884
 885	if (bio && bio_data_dir(bio) == WRITE) {
 886		if (bio->bi_iter.bi_sector >= conf->next_resync) {
 887			if (conf->start_next_window == MaxSector)
 888				conf->start_next_window =
 889					conf->next_resync +
 890					NEXT_NORMALIO_DISTANCE;
 891
 892			if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
 893			    <= bio->bi_iter.bi_sector)
 894				conf->next_window_requests++;
 895			else
 896				conf->current_window_requests++;
 897			sector = conf->start_next_window;
 898		}
 899	}
 900
 901	conf->nr_pending++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 902	spin_unlock_irq(&conf->resync_lock);
 903	return sector;
 904}
 905
 906static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
 907			  sector_t bi_sector)
 908{
 909	unsigned long flags;
 910
 911	spin_lock_irqsave(&conf->resync_lock, flags);
 912	conf->nr_pending--;
 913	if (start_next_window) {
 914		if (start_next_window == conf->start_next_window) {
 915			if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
 916			    <= bi_sector)
 917				conf->next_window_requests--;
 918			else
 919				conf->current_window_requests--;
 920		} else
 921			conf->current_window_requests--;
 922
 923		if (!conf->current_window_requests) {
 924			if (conf->next_window_requests) {
 925				conf->current_window_requests =
 926					conf->next_window_requests;
 927				conf->next_window_requests = 0;
 928				conf->start_next_window +=
 929					NEXT_NORMALIO_DISTANCE;
 930			} else
 931				conf->start_next_window = MaxSector;
 932		}
 933	}
 934	spin_unlock_irqrestore(&conf->resync_lock, flags);
 935	wake_up(&conf->wait_barrier);
 936}
 937
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 938static void freeze_array(struct r1conf *conf, int extra)
 939{
 940	/* stop syncio and normal IO and wait for everything to
 941	 * go quite.
 942	 * We wait until nr_pending match nr_queued+extra
 943	 * This is called in the context of one normal IO request
 944	 * that has failed. Thus any sync request that might be pending
 945	 * will be blocked by nr_pending, and we need to wait for
 946	 * pending IO requests to complete or be queued for re-try.
 947	 * Thus the number queued (nr_queued) plus this request (extra)
 948	 * must match the number of pending IOs (nr_pending) before
 949	 * we continue.
 
 
 
 
 
 
 
 
 
 
 
 
 950	 */
 951	spin_lock_irq(&conf->resync_lock);
 952	conf->array_frozen = 1;
 953	wait_event_lock_irq_cmd(conf->wait_barrier,
 954				conf->nr_pending == conf->nr_queued+extra,
 955				conf->resync_lock,
 956				flush_pending_writes(conf));
 
 
 957	spin_unlock_irq(&conf->resync_lock);
 958}
 959static void unfreeze_array(struct r1conf *conf)
 960{
 961	/* reverse the effect of the freeze */
 962	spin_lock_irq(&conf->resync_lock);
 963	conf->array_frozen = 0;
 964	wake_up(&conf->wait_barrier);
 965	spin_unlock_irq(&conf->resync_lock);
 
 966}
 967
 968/* duplicate the data pages for behind I/O
 969 */
 970static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 971{
 972	int i;
 973	struct bio_vec *bvec;
 974	struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 975					GFP_NOIO);
 976	if (unlikely(!bvecs))
 
 
 977		return;
 978
 979	bio_for_each_segment_all(bvec, bio, i) {
 980		bvecs[i] = *bvec;
 981		bvecs[i].bv_page = alloc_page(GFP_NOIO);
 982		if (unlikely(!bvecs[i].bv_page))
 983			goto do_sync_io;
 984		memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 985		       kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 986		kunmap(bvecs[i].bv_page);
 987		kunmap(bvec->bv_page);
 988	}
 989	r1_bio->behind_bvecs = bvecs;
 990	r1_bio->behind_page_count = bio->bi_vcnt;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 991	set_bit(R1BIO_BehindIO, &r1_bio->state);
 
 992	return;
 993
 994do_sync_io:
 995	for (i = 0; i < bio->bi_vcnt; i++)
 996		if (bvecs[i].bv_page)
 997			put_page(bvecs[i].bv_page);
 998	kfree(bvecs);
 999	pr_debug("%dB behind alloc failed, doing sync I/O\n",
1000		 bio->bi_iter.bi_size);
 
 
1001}
1002
1003struct raid1_plug_cb {
1004	struct blk_plug_cb	cb;
1005	struct bio_list		pending;
1006	int			pending_cnt;
1007};
1008
1009static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1010{
1011	struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1012						  cb);
1013	struct mddev *mddev = plug->cb.data;
1014	struct r1conf *conf = mddev->private;
1015	struct bio *bio;
1016
1017	if (from_schedule || current->bio_list) {
1018		spin_lock_irq(&conf->device_lock);
1019		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1020		conf->pending_count += plug->pending_cnt;
1021		spin_unlock_irq(&conf->device_lock);
1022		wake_up(&conf->wait_barrier);
1023		md_wakeup_thread(mddev->thread);
1024		kfree(plug);
1025		return;
1026	}
1027
1028	/* we aren't scheduling, so we can do the write-out directly. */
1029	bio = bio_list_get(&plug->pending);
1030	bitmap_unplug(mddev->bitmap);
1031	wake_up(&conf->wait_barrier);
1032
1033	while (bio) { /* submit pending writes */
1034		struct bio *next = bio->bi_next;
1035		bio->bi_next = NULL;
1036		if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1037		    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1038			/* Just ignore it */
1039			bio_endio(bio);
1040		else
1041			generic_make_request(bio);
1042		bio = next;
1043	}
1044	kfree(plug);
1045}
1046
1047static void raid1_make_request(struct mddev *mddev, struct bio * bio)
 
 
 
 
 
 
 
 
 
 
1048{
1049	struct r1conf *conf = mddev->private;
1050	struct raid1_info *mirror;
1051	struct r1bio *r1_bio;
 
 
 
 
 
 
 
 
 
 
 
 
 
1052	struct bio *read_bio;
1053	int i, disks;
1054	struct bitmap *bitmap;
1055	unsigned long flags;
1056	const int rw = bio_data_dir(bio);
1057	const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1058	const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1059	const unsigned long do_discard = (bio->bi_rw
1060					  & (REQ_DISCARD | REQ_SECURE));
1061	const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1062	struct md_rdev *blocked_rdev;
1063	struct blk_plug_cb *cb;
1064	struct raid1_plug_cb *plug = NULL;
1065	int first_clone;
1066	int sectors_handled;
1067	int max_sectors;
1068	sector_t start_next_window;
 
 
1069
1070	/*
1071	 * Register the new request and wait if the reconstruction
1072	 * thread has put up a bar for new requests.
1073	 * Continue immediately if no resync is active currently.
1074	 */
 
1075
1076	md_write_start(mddev, bio); /* wait on superblock update early */
1077
1078	if (bio_data_dir(bio) == WRITE &&
1079	    ((bio_end_sector(bio) > mddev->suspend_lo &&
1080	    bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1081	    (mddev_is_clustered(mddev) &&
1082	     md_cluster_ops->area_resyncing(mddev, WRITE,
1083		     bio->bi_iter.bi_sector, bio_end_sector(bio))))) {
1084		/* As the suspend_* range is controlled by
1085		 * userspace, we want an interruptible
1086		 * wait.
1087		 */
1088		DEFINE_WAIT(w);
1089		for (;;) {
1090			flush_signals(current);
1091			prepare_to_wait(&conf->wait_barrier,
1092					&w, TASK_INTERRUPTIBLE);
1093			if (bio_end_sector(bio) <= mddev->suspend_lo ||
1094			    bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1095			    (mddev_is_clustered(mddev) &&
1096			     !md_cluster_ops->area_resyncing(mddev, WRITE,
1097				     bio->bi_iter.bi_sector, bio_end_sector(bio))))
1098				break;
1099			schedule();
1100		}
1101		finish_wait(&conf->wait_barrier, &w);
1102	}
1103
1104	start_next_window = wait_barrier(conf, bio);
 
 
 
 
1105
1106	bitmap = mddev->bitmap;
 
 
 
 
1107
1108	/*
1109	 * make_request() can abort the operation when READA is being
1110	 * used and no empty request is available.
1111	 *
1112	 */
1113	r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1114
1115	r1_bio->master_bio = bio;
1116	r1_bio->sectors = bio_sectors(bio);
1117	r1_bio->state = 0;
1118	r1_bio->mddev = mddev;
1119	r1_bio->sector = bio->bi_iter.bi_sector;
 
 
 
 
 
 
 
1120
1121	/* We might need to issue multiple reads to different
1122	 * devices if there are bad blocks around, so we keep
1123	 * track of the number of reads in bio->bi_phys_segments.
1124	 * If this is 0, there is only one r1_bio and no locking
1125	 * will be needed when requests complete.  If it is
1126	 * non-zero, then it is the number of not-completed requests.
1127	 */
1128	bio->bi_phys_segments = 0;
1129	bio_clear_flag(bio, BIO_SEG_VALID);
1130
1131	if (rw == READ) {
 
1132		/*
1133		 * read balancing logic:
 
1134		 */
1135		int rdisk;
 
 
 
1136
1137read_again:
1138		rdisk = read_balance(conf, r1_bio, &max_sectors);
 
 
 
 
 
 
 
1139
1140		if (rdisk < 0) {
1141			/* couldn't find anywhere to read from */
1142			raid_end_bio_io(r1_bio);
1143			return;
1144		}
1145		mirror = conf->mirrors + rdisk;
1146
1147		if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1148		    bitmap) {
1149			/* Reading from a write-mostly device must
1150			 * take care not to over-take any writes
1151			 * that are 'behind'
1152			 */
1153			wait_event(bitmap->behind_wait,
1154				   atomic_read(&bitmap->behind_writes) == 0);
1155		}
1156		r1_bio->read_disk = rdisk;
1157		r1_bio->start_next_window = 0;
1158
1159		read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1160		bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1161			 max_sectors);
1162
1163		r1_bio->bios[rdisk] = read_bio;
1164
1165		read_bio->bi_iter.bi_sector = r1_bio->sector +
1166			mirror->rdev->data_offset;
1167		read_bio->bi_bdev = mirror->rdev->bdev;
1168		read_bio->bi_end_io = raid1_end_read_request;
1169		read_bio->bi_rw = READ | do_sync;
1170		read_bio->bi_private = r1_bio;
1171
1172		if (max_sectors < r1_bio->sectors) {
1173			/* could not read all from this device, so we will
1174			 * need another r1_bio.
1175			 */
1176
1177			sectors_handled = (r1_bio->sector + max_sectors
1178					   - bio->bi_iter.bi_sector);
1179			r1_bio->sectors = max_sectors;
1180			spin_lock_irq(&conf->device_lock);
1181			if (bio->bi_phys_segments == 0)
1182				bio->bi_phys_segments = 2;
1183			else
1184				bio->bi_phys_segments++;
1185			spin_unlock_irq(&conf->device_lock);
1186			/* Cannot call generic_make_request directly
1187			 * as that will be queued in __make_request
1188			 * and subsequent mempool_alloc might block waiting
1189			 * for it.  So hand bio over to raid1d.
1190			 */
1191			reschedule_retry(r1_bio);
1192
1193			r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 
 
1194
1195			r1_bio->master_bio = bio;
1196			r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1197			r1_bio->state = 0;
1198			r1_bio->mddev = mddev;
1199			r1_bio->sector = bio->bi_iter.bi_sector +
1200				sectors_handled;
1201			goto read_again;
1202		} else
1203			generic_make_request(read_bio);
1204		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1205	}
1206
1207	/*
1208	 * WRITE:
 
 
1209	 */
 
 
 
 
 
1210	if (conf->pending_count >= max_queued_requests) {
1211		md_wakeup_thread(mddev->thread);
 
1212		wait_event(conf->wait_barrier,
1213			   conf->pending_count < max_queued_requests);
1214	}
1215	/* first select target devices under rcu_lock and
1216	 * inc refcount on their rdev.  Record them by setting
1217	 * bios[x] to bio
1218	 * If there are known/acknowledged bad blocks on any device on
1219	 * which we have seen a write error, we want to avoid writing those
1220	 * blocks.
1221	 * This potentially requires several writes to write around
1222	 * the bad blocks.  Each set of writes gets it's own r1bio
1223	 * with a set of bios attached.
1224	 */
1225
1226	disks = conf->raid_disks * 2;
1227 retry_write:
1228	r1_bio->start_next_window = start_next_window;
1229	blocked_rdev = NULL;
1230	rcu_read_lock();
1231	max_sectors = r1_bio->sectors;
1232	for (i = 0;  i < disks; i++) {
1233		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1234		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1235			atomic_inc(&rdev->nr_pending);
1236			blocked_rdev = rdev;
1237			break;
1238		}
1239		r1_bio->bios[i] = NULL;
1240		if (!rdev || test_bit(Faulty, &rdev->flags)) {
1241			if (i < conf->raid_disks)
1242				set_bit(R1BIO_Degraded, &r1_bio->state);
1243			continue;
1244		}
1245
1246		atomic_inc(&rdev->nr_pending);
1247		if (test_bit(WriteErrorSeen, &rdev->flags)) {
1248			sector_t first_bad;
1249			int bad_sectors;
1250			int is_bad;
1251
1252			is_bad = is_badblock(rdev, r1_bio->sector,
1253					     max_sectors,
1254					     &first_bad, &bad_sectors);
1255			if (is_bad < 0) {
1256				/* mustn't write here until the bad block is
1257				 * acknowledged*/
1258				set_bit(BlockedBadBlocks, &rdev->flags);
1259				blocked_rdev = rdev;
1260				break;
1261			}
1262			if (is_bad && first_bad <= r1_bio->sector) {
1263				/* Cannot write here at all */
1264				bad_sectors -= (r1_bio->sector - first_bad);
1265				if (bad_sectors < max_sectors)
1266					/* mustn't write more than bad_sectors
1267					 * to other devices yet
1268					 */
1269					max_sectors = bad_sectors;
1270				rdev_dec_pending(rdev, mddev);
1271				/* We don't set R1BIO_Degraded as that
1272				 * only applies if the disk is
1273				 * missing, so it might be re-added,
1274				 * and we want to know to recover this
1275				 * chunk.
1276				 * In this case the device is here,
1277				 * and the fact that this chunk is not
1278				 * in-sync is recorded in the bad
1279				 * block log
1280				 */
1281				continue;
1282			}
1283			if (is_bad) {
1284				int good_sectors = first_bad - r1_bio->sector;
1285				if (good_sectors < max_sectors)
1286					max_sectors = good_sectors;
1287			}
1288		}
1289		r1_bio->bios[i] = bio;
1290	}
1291	rcu_read_unlock();
1292
1293	if (unlikely(blocked_rdev)) {
1294		/* Wait for this device to become unblocked */
1295		int j;
1296		sector_t old = start_next_window;
1297
1298		for (j = 0; j < i; j++)
1299			if (r1_bio->bios[j])
1300				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1301		r1_bio->state = 0;
1302		allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
 
1303		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1304		start_next_window = wait_barrier(conf, bio);
1305		/*
1306		 * We must make sure the multi r1bios of bio have
1307		 * the same value of bi_phys_segments
1308		 */
1309		if (bio->bi_phys_segments && old &&
1310		    old != start_next_window)
1311			/* Wait for the former r1bio(s) to complete */
1312			wait_event(conf->wait_barrier,
1313				   bio->bi_phys_segments == 1);
1314		goto retry_write;
1315	}
1316
1317	if (max_sectors < r1_bio->sectors) {
1318		/* We are splitting this write into multiple parts, so
1319		 * we need to prepare for allocating another r1_bio.
1320		 */
 
 
 
1321		r1_bio->sectors = max_sectors;
1322		spin_lock_irq(&conf->device_lock);
1323		if (bio->bi_phys_segments == 0)
1324			bio->bi_phys_segments = 2;
1325		else
1326			bio->bi_phys_segments++;
1327		spin_unlock_irq(&conf->device_lock);
1328	}
1329	sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1330
1331	atomic_set(&r1_bio->remaining, 1);
1332	atomic_set(&r1_bio->behind_remaining, 0);
1333
1334	first_clone = 1;
 
1335	for (i = 0; i < disks; i++) {
1336		struct bio *mbio;
 
1337		if (!r1_bio->bios[i])
1338			continue;
1339
1340		mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1341		bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1342
1343		if (first_clone) {
1344			/* do behind I/O ?
1345			 * Not if there are too many, or cannot
1346			 * allocate memory, or a reader on WriteMostly
1347			 * is waiting for behind writes to flush */
1348			if (bitmap &&
1349			    (atomic_read(&bitmap->behind_writes)
1350			     < mddev->bitmap_info.max_write_behind) &&
1351			    !waitqueue_active(&bitmap->behind_wait))
1352				alloc_behind_pages(mbio, r1_bio);
 
1353
1354			bitmap_startwrite(bitmap, r1_bio->sector,
1355					  r1_bio->sectors,
1356					  test_bit(R1BIO_BehindIO,
1357						   &r1_bio->state));
1358			first_clone = 0;
1359		}
1360		if (r1_bio->behind_bvecs) {
1361			struct bio_vec *bvec;
1362			int j;
1363
1364			/*
1365			 * We trimmed the bio, so _all is legit
1366			 */
1367			bio_for_each_segment_all(bvec, mbio, j)
1368				bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1369			if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 
 
 
 
1370				atomic_inc(&r1_bio->behind_remaining);
1371		}
 
1372
1373		r1_bio->bios[i] = mbio;
1374
1375		mbio->bi_iter.bi_sector	= (r1_bio->sector +
1376				   conf->mirrors[i].rdev->data_offset);
1377		mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1378		mbio->bi_end_io	= raid1_end_write_request;
1379		mbio->bi_rw =
1380			WRITE | do_flush_fua | do_sync | do_discard | do_same;
 
 
 
1381		mbio->bi_private = r1_bio;
1382
1383		atomic_inc(&r1_bio->remaining);
1384
 
 
 
 
 
 
 
1385		cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1386		if (cb)
1387			plug = container_of(cb, struct raid1_plug_cb, cb);
1388		else
1389			plug = NULL;
1390		spin_lock_irqsave(&conf->device_lock, flags);
1391		if (plug) {
1392			bio_list_add(&plug->pending, mbio);
1393			plug->pending_cnt++;
1394		} else {
 
1395			bio_list_add(&conf->pending_bio_list, mbio);
1396			conf->pending_count++;
1397		}
1398		spin_unlock_irqrestore(&conf->device_lock, flags);
1399		if (!plug)
1400			md_wakeup_thread(mddev->thread);
1401	}
1402	/* Mustn't call r1_bio_write_done before this next test,
1403	 * as it could result in the bio being freed.
1404	 */
1405	if (sectors_handled < bio_sectors(bio)) {
1406		r1_bio_write_done(r1_bio);
1407		/* We need another r1_bio.  It has already been counted
1408		 * in bio->bi_phys_segments
1409		 */
1410		r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1411		r1_bio->master_bio = bio;
1412		r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1413		r1_bio->state = 0;
1414		r1_bio->mddev = mddev;
1415		r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1416		goto retry_write;
1417	}
1418
1419	r1_bio_write_done(r1_bio);
1420
1421	/* In case raid1d snuck in to freeze_array */
1422	wake_up(&conf->wait_barrier);
1423}
1424
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1425static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1426{
1427	struct r1conf *conf = mddev->private;
1428	int i;
1429
1430	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1431		   conf->raid_disks - mddev->degraded);
1432	rcu_read_lock();
1433	for (i = 0; i < conf->raid_disks; i++) {
1434		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1435		seq_printf(seq, "%s",
1436			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1437	}
1438	rcu_read_unlock();
1439	seq_printf(seq, "]");
1440}
1441
1442static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1443{
1444	char b[BDEVNAME_SIZE];
1445	struct r1conf *conf = mddev->private;
1446	unsigned long flags;
1447
1448	/*
1449	 * If it is not operational, then we have already marked it as dead
1450	 * else if it is the last working disks, ignore the error, let the
1451	 * next level up know.
1452	 * else mark the drive as failed
1453	 */
1454	if (test_bit(In_sync, &rdev->flags)
 
1455	    && (conf->raid_disks - mddev->degraded) == 1) {
1456		/*
1457		 * Don't fail the drive, act as though we were just a
1458		 * normal single drive.
1459		 * However don't try a recovery from this drive as
1460		 * it is very likely to fail.
1461		 */
1462		conf->recovery_disabled = mddev->recovery_disabled;
 
1463		return;
1464	}
1465	set_bit(Blocked, &rdev->flags);
1466	spin_lock_irqsave(&conf->device_lock, flags);
1467	if (test_and_clear_bit(In_sync, &rdev->flags)) {
1468		mddev->degraded++;
1469		set_bit(Faulty, &rdev->flags);
1470	} else
1471		set_bit(Faulty, &rdev->flags);
1472	spin_unlock_irqrestore(&conf->device_lock, flags);
1473	/*
1474	 * if recovery is running, make sure it aborts.
1475	 */
1476	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1477	set_bit(MD_CHANGE_DEVS, &mddev->flags);
1478	set_bit(MD_CHANGE_PENDING, &mddev->flags);
1479	printk(KERN_ALERT
1480	       "md/raid1:%s: Disk failure on %s, disabling device.\n"
1481	       "md/raid1:%s: Operation continuing on %d devices.\n",
1482	       mdname(mddev), bdevname(rdev->bdev, b),
1483	       mdname(mddev), conf->raid_disks - mddev->degraded);
1484}
1485
1486static void print_conf(struct r1conf *conf)
1487{
1488	int i;
1489
1490	printk(KERN_DEBUG "RAID1 conf printout:\n");
1491	if (!conf) {
1492		printk(KERN_DEBUG "(!conf)\n");
1493		return;
1494	}
1495	printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1496		conf->raid_disks);
1497
1498	rcu_read_lock();
1499	for (i = 0; i < conf->raid_disks; i++) {
1500		char b[BDEVNAME_SIZE];
1501		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1502		if (rdev)
1503			printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1504			       i, !test_bit(In_sync, &rdev->flags),
1505			       !test_bit(Faulty, &rdev->flags),
1506			       bdevname(rdev->bdev,b));
1507	}
1508	rcu_read_unlock();
1509}
1510
1511static void close_sync(struct r1conf *conf)
1512{
1513	wait_barrier(conf, NULL);
1514	allow_barrier(conf, 0, 0);
1515
1516	mempool_destroy(conf->r1buf_pool);
1517	conf->r1buf_pool = NULL;
 
 
1518
1519	spin_lock_irq(&conf->resync_lock);
1520	conf->next_resync = MaxSector - 2 * NEXT_NORMALIO_DISTANCE;
1521	conf->start_next_window = MaxSector;
1522	conf->current_window_requests +=
1523		conf->next_window_requests;
1524	conf->next_window_requests = 0;
1525	spin_unlock_irq(&conf->resync_lock);
1526}
1527
1528static int raid1_spare_active(struct mddev *mddev)
1529{
1530	int i;
1531	struct r1conf *conf = mddev->private;
1532	int count = 0;
1533	unsigned long flags;
1534
1535	/*
1536	 * Find all failed disks within the RAID1 configuration
1537	 * and mark them readable.
1538	 * Called under mddev lock, so rcu protection not needed.
1539	 * device_lock used to avoid races with raid1_end_read_request
1540	 * which expects 'In_sync' flags and ->degraded to be consistent.
1541	 */
1542	spin_lock_irqsave(&conf->device_lock, flags);
1543	for (i = 0; i < conf->raid_disks; i++) {
1544		struct md_rdev *rdev = conf->mirrors[i].rdev;
1545		struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1546		if (repl
1547		    && !test_bit(Candidate, &repl->flags)
1548		    && repl->recovery_offset == MaxSector
1549		    && !test_bit(Faulty, &repl->flags)
1550		    && !test_and_set_bit(In_sync, &repl->flags)) {
1551			/* replacement has just become active */
1552			if (!rdev ||
1553			    !test_and_clear_bit(In_sync, &rdev->flags))
1554				count++;
1555			if (rdev) {
1556				/* Replaced device not technically
1557				 * faulty, but we need to be sure
1558				 * it gets removed and never re-added
1559				 */
1560				set_bit(Faulty, &rdev->flags);
1561				sysfs_notify_dirent_safe(
1562					rdev->sysfs_state);
1563			}
1564		}
1565		if (rdev
1566		    && rdev->recovery_offset == MaxSector
1567		    && !test_bit(Faulty, &rdev->flags)
1568		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1569			count++;
1570			sysfs_notify_dirent_safe(rdev->sysfs_state);
1571		}
1572	}
1573	mddev->degraded -= count;
1574	spin_unlock_irqrestore(&conf->device_lock, flags);
1575
1576	print_conf(conf);
1577	return count;
1578}
1579
1580static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1581{
1582	struct r1conf *conf = mddev->private;
1583	int err = -EEXIST;
1584	int mirror = 0;
1585	struct raid1_info *p;
1586	int first = 0;
1587	int last = conf->raid_disks - 1;
1588
1589	if (mddev->recovery_disabled == conf->recovery_disabled)
1590		return -EBUSY;
1591
1592	if (md_integrity_add_rdev(rdev, mddev))
1593		return -ENXIO;
1594
1595	if (rdev->raid_disk >= 0)
1596		first = last = rdev->raid_disk;
1597
1598	/*
1599	 * find the disk ... but prefer rdev->saved_raid_disk
1600	 * if possible.
1601	 */
1602	if (rdev->saved_raid_disk >= 0 &&
1603	    rdev->saved_raid_disk >= first &&
 
1604	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1605		first = last = rdev->saved_raid_disk;
1606
1607	for (mirror = first; mirror <= last; mirror++) {
1608		p = conf->mirrors+mirror;
1609		if (!p->rdev) {
1610
1611			if (mddev->gendisk)
1612				disk_stack_limits(mddev->gendisk, rdev->bdev,
1613						  rdev->data_offset << 9);
1614
1615			p->head_position = 0;
1616			rdev->raid_disk = mirror;
1617			err = 0;
1618			/* As all devices are equivalent, we don't need a full recovery
1619			 * if this was recently any drive of the array
1620			 */
1621			if (rdev->saved_raid_disk < 0)
1622				conf->fullsync = 1;
1623			rcu_assign_pointer(p->rdev, rdev);
1624			break;
1625		}
1626		if (test_bit(WantReplacement, &p->rdev->flags) &&
1627		    p[conf->raid_disks].rdev == NULL) {
1628			/* Add this device as a replacement */
1629			clear_bit(In_sync, &rdev->flags);
1630			set_bit(Replacement, &rdev->flags);
1631			rdev->raid_disk = mirror;
1632			err = 0;
1633			conf->fullsync = 1;
1634			rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1635			break;
1636		}
1637	}
1638	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1639		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1640	print_conf(conf);
1641	return err;
1642}
1643
1644static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1645{
1646	struct r1conf *conf = mddev->private;
1647	int err = 0;
1648	int number = rdev->raid_disk;
1649	struct raid1_info *p = conf->mirrors + number;
1650
1651	if (rdev != p->rdev)
1652		p = conf->mirrors + conf->raid_disks + number;
1653
1654	print_conf(conf);
1655	if (rdev == p->rdev) {
1656		if (test_bit(In_sync, &rdev->flags) ||
1657		    atomic_read(&rdev->nr_pending)) {
1658			err = -EBUSY;
1659			goto abort;
1660		}
1661		/* Only remove non-faulty devices if recovery
1662		 * is not possible.
1663		 */
1664		if (!test_bit(Faulty, &rdev->flags) &&
1665		    mddev->recovery_disabled != conf->recovery_disabled &&
1666		    mddev->degraded < conf->raid_disks) {
1667			err = -EBUSY;
1668			goto abort;
1669		}
1670		p->rdev = NULL;
1671		synchronize_rcu();
1672		if (atomic_read(&rdev->nr_pending)) {
1673			/* lost the race, try later */
1674			err = -EBUSY;
1675			p->rdev = rdev;
1676			goto abort;
1677		} else if (conf->mirrors[conf->raid_disks + number].rdev) {
 
 
 
1678			/* We just removed a device that is being replaced.
1679			 * Move down the replacement.  We drain all IO before
1680			 * doing this to avoid confusion.
1681			 */
1682			struct md_rdev *repl =
1683				conf->mirrors[conf->raid_disks + number].rdev;
1684			freeze_array(conf, 0);
 
 
 
 
 
 
 
 
 
 
 
1685			clear_bit(Replacement, &repl->flags);
1686			p->rdev = repl;
1687			conf->mirrors[conf->raid_disks + number].rdev = NULL;
1688			unfreeze_array(conf);
1689			clear_bit(WantReplacement, &rdev->flags);
1690		} else
1691			clear_bit(WantReplacement, &rdev->flags);
1692		err = md_integrity_register(mddev);
1693	}
1694abort:
1695
1696	print_conf(conf);
1697	return err;
1698}
1699
1700static void end_sync_read(struct bio *bio)
1701{
1702	struct r1bio *r1_bio = bio->bi_private;
1703
1704	update_head_pos(r1_bio->read_disk, r1_bio);
1705
1706	/*
1707	 * we have read a block, now it needs to be re-written,
1708	 * or re-read if the read failed.
1709	 * We don't do much here, just schedule handling by raid1d
1710	 */
1711	if (!bio->bi_error)
1712		set_bit(R1BIO_Uptodate, &r1_bio->state);
1713
1714	if (atomic_dec_and_test(&r1_bio->remaining))
1715		reschedule_retry(r1_bio);
1716}
1717
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1718static void end_sync_write(struct bio *bio)
1719{
1720	int uptodate = !bio->bi_error;
1721	struct r1bio *r1_bio = bio->bi_private;
1722	struct mddev *mddev = r1_bio->mddev;
1723	struct r1conf *conf = mddev->private;
1724	int mirror=0;
1725	sector_t first_bad;
1726	int bad_sectors;
1727
1728	mirror = find_bio_disk(r1_bio, bio);
1729
1730	if (!uptodate) {
1731		sector_t sync_blocks = 0;
1732		sector_t s = r1_bio->sector;
1733		long sectors_to_go = r1_bio->sectors;
1734		/* make sure these bits doesn't get cleared. */
1735		do {
1736			bitmap_end_sync(mddev->bitmap, s,
1737					&sync_blocks, 1);
1738			s += sync_blocks;
1739			sectors_to_go -= sync_blocks;
1740		} while (sectors_to_go > 0);
1741		set_bit(WriteErrorSeen,
1742			&conf->mirrors[mirror].rdev->flags);
1743		if (!test_and_set_bit(WantReplacement,
1744				      &conf->mirrors[mirror].rdev->flags))
1745			set_bit(MD_RECOVERY_NEEDED, &
1746				mddev->recovery);
1747		set_bit(R1BIO_WriteError, &r1_bio->state);
1748	} else if (is_badblock(conf->mirrors[mirror].rdev,
1749			       r1_bio->sector,
1750			       r1_bio->sectors,
1751			       &first_bad, &bad_sectors) &&
1752		   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1753				r1_bio->sector,
1754				r1_bio->sectors,
1755				&first_bad, &bad_sectors)
1756		)
1757		set_bit(R1BIO_MadeGood, &r1_bio->state);
1758
1759	if (atomic_dec_and_test(&r1_bio->remaining)) {
1760		int s = r1_bio->sectors;
1761		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1762		    test_bit(R1BIO_WriteError, &r1_bio->state))
1763			reschedule_retry(r1_bio);
1764		else {
1765			put_buf(r1_bio);
1766			md_done_sync(mddev, s, uptodate);
1767		}
1768	}
1769}
1770
1771static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1772			    int sectors, struct page *page, int rw)
1773{
1774	if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1775		/* success */
1776		return 1;
1777	if (rw == WRITE) {
1778		set_bit(WriteErrorSeen, &rdev->flags);
1779		if (!test_and_set_bit(WantReplacement,
1780				      &rdev->flags))
1781			set_bit(MD_RECOVERY_NEEDED, &
1782				rdev->mddev->recovery);
1783	}
1784	/* need to record an error - either for the block or the device */
1785	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1786		md_error(rdev->mddev, rdev);
1787	return 0;
1788}
1789
1790static int fix_sync_read_error(struct r1bio *r1_bio)
1791{
1792	/* Try some synchronous reads of other devices to get
1793	 * good data, much like with normal read errors.  Only
1794	 * read into the pages we already have so we don't
1795	 * need to re-issue the read request.
1796	 * We don't need to freeze the array, because being in an
1797	 * active sync request, there is no normal IO, and
1798	 * no overlapping syncs.
1799	 * We don't need to check is_badblock() again as we
1800	 * made sure that anything with a bad block in range
1801	 * will have bi_end_io clear.
1802	 */
1803	struct mddev *mddev = r1_bio->mddev;
1804	struct r1conf *conf = mddev->private;
1805	struct bio *bio = r1_bio->bios[r1_bio->read_disk];
 
1806	sector_t sect = r1_bio->sector;
1807	int sectors = r1_bio->sectors;
1808	int idx = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
1809
1810	while(sectors) {
1811		int s = sectors;
1812		int d = r1_bio->read_disk;
1813		int success = 0;
1814		struct md_rdev *rdev;
1815		int start;
1816
1817		if (s > (PAGE_SIZE>>9))
1818			s = PAGE_SIZE >> 9;
1819		do {
1820			if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1821				/* No rcu protection needed here devices
1822				 * can only be removed when no resync is
1823				 * active, and resync is currently active
1824				 */
1825				rdev = conf->mirrors[d].rdev;
1826				if (sync_page_io(rdev, sect, s<<9,
1827						 bio->bi_io_vec[idx].bv_page,
1828						 READ, false)) {
1829					success = 1;
1830					break;
1831				}
1832			}
1833			d++;
1834			if (d == conf->raid_disks * 2)
1835				d = 0;
1836		} while (!success && d != r1_bio->read_disk);
1837
1838		if (!success) {
1839			char b[BDEVNAME_SIZE];
1840			int abort = 0;
1841			/* Cannot read from anywhere, this block is lost.
1842			 * Record a bad block on each device.  If that doesn't
1843			 * work just disable and interrupt the recovery.
1844			 * Don't fail devices as that won't really help.
1845			 */
1846			printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1847			       " for block %llu\n",
1848			       mdname(mddev),
1849			       bdevname(bio->bi_bdev, b),
1850			       (unsigned long long)r1_bio->sector);
1851			for (d = 0; d < conf->raid_disks * 2; d++) {
1852				rdev = conf->mirrors[d].rdev;
1853				if (!rdev || test_bit(Faulty, &rdev->flags))
1854					continue;
1855				if (!rdev_set_badblocks(rdev, sect, s, 0))
1856					abort = 1;
1857			}
1858			if (abort) {
1859				conf->recovery_disabled =
1860					mddev->recovery_disabled;
1861				set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1862				md_done_sync(mddev, r1_bio->sectors, 0);
1863				put_buf(r1_bio);
1864				return 0;
1865			}
1866			/* Try next page */
1867			sectors -= s;
1868			sect += s;
1869			idx++;
1870			continue;
1871		}
1872
1873		start = d;
1874		/* write it back and re-read */
1875		while (d != r1_bio->read_disk) {
1876			if (d == 0)
1877				d = conf->raid_disks * 2;
1878			d--;
1879			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1880				continue;
1881			rdev = conf->mirrors[d].rdev;
1882			if (r1_sync_page_io(rdev, sect, s,
1883					    bio->bi_io_vec[idx].bv_page,
1884					    WRITE) == 0) {
1885				r1_bio->bios[d]->bi_end_io = NULL;
1886				rdev_dec_pending(rdev, mddev);
1887			}
1888		}
1889		d = start;
1890		while (d != r1_bio->read_disk) {
1891			if (d == 0)
1892				d = conf->raid_disks * 2;
1893			d--;
1894			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1895				continue;
1896			rdev = conf->mirrors[d].rdev;
1897			if (r1_sync_page_io(rdev, sect, s,
1898					    bio->bi_io_vec[idx].bv_page,
1899					    READ) != 0)
1900				atomic_add(s, &rdev->corrected_errors);
1901		}
1902		sectors -= s;
1903		sect += s;
1904		idx ++;
1905	}
1906	set_bit(R1BIO_Uptodate, &r1_bio->state);
1907	bio->bi_error = 0;
1908	return 1;
1909}
1910
1911static void process_checks(struct r1bio *r1_bio)
1912{
1913	/* We have read all readable devices.  If we haven't
1914	 * got the block, then there is no hope left.
1915	 * If we have, then we want to do a comparison
1916	 * and skip the write if everything is the same.
1917	 * If any blocks failed to read, then we need to
1918	 * attempt an over-write
1919	 */
1920	struct mddev *mddev = r1_bio->mddev;
1921	struct r1conf *conf = mddev->private;
1922	int primary;
1923	int i;
1924	int vcnt;
1925
1926	/* Fix variable parts of all bios */
1927	vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1928	for (i = 0; i < conf->raid_disks * 2; i++) {
1929		int j;
1930		int size;
1931		int error;
1932		struct bio *b = r1_bio->bios[i];
 
1933		if (b->bi_end_io != end_sync_read)
1934			continue;
1935		/* fixup the bio for reuse, but preserve errno */
1936		error = b->bi_error;
1937		bio_reset(b);
1938		b->bi_error = error;
1939		b->bi_vcnt = vcnt;
1940		b->bi_iter.bi_size = r1_bio->sectors << 9;
1941		b->bi_iter.bi_sector = r1_bio->sector +
1942			conf->mirrors[i].rdev->data_offset;
1943		b->bi_bdev = conf->mirrors[i].rdev->bdev;
1944		b->bi_end_io = end_sync_read;
1945		b->bi_private = r1_bio;
 
1946
1947		size = b->bi_iter.bi_size;
1948		for (j = 0; j < vcnt ; j++) {
1949			struct bio_vec *bi;
1950			bi = &b->bi_io_vec[j];
1951			bi->bv_offset = 0;
1952			if (size > PAGE_SIZE)
1953				bi->bv_len = PAGE_SIZE;
1954			else
1955				bi->bv_len = size;
1956			size -= PAGE_SIZE;
1957		}
1958	}
1959	for (primary = 0; primary < conf->raid_disks * 2; primary++)
1960		if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1961		    !r1_bio->bios[primary]->bi_error) {
1962			r1_bio->bios[primary]->bi_end_io = NULL;
1963			rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1964			break;
1965		}
1966	r1_bio->read_disk = primary;
1967	for (i = 0; i < conf->raid_disks * 2; i++) {
1968		int j;
1969		struct bio *pbio = r1_bio->bios[primary];
1970		struct bio *sbio = r1_bio->bios[i];
1971		int error = sbio->bi_error;
 
 
 
 
 
1972
1973		if (sbio->bi_end_io != end_sync_read)
1974			continue;
1975		/* Now we can 'fixup' the error value */
1976		sbio->bi_error = 0;
 
 
 
1977
1978		if (!error) {
1979			for (j = vcnt; j-- ; ) {
1980				struct page *p, *s;
1981				p = pbio->bi_io_vec[j].bv_page;
1982				s = sbio->bi_io_vec[j].bv_page;
1983				if (memcmp(page_address(p),
1984					   page_address(s),
1985					   sbio->bi_io_vec[j].bv_len))
1986					break;
1987			}
1988		} else
1989			j = 0;
1990		if (j >= 0)
1991			atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1992		if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1993			      && !error)) {
1994			/* No need to write to this device. */
1995			sbio->bi_end_io = NULL;
1996			rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1997			continue;
1998		}
1999
2000		bio_copy_data(sbio, pbio);
2001	}
2002}
2003
2004static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2005{
2006	struct r1conf *conf = mddev->private;
2007	int i;
2008	int disks = conf->raid_disks * 2;
2009	struct bio *bio, *wbio;
2010
2011	bio = r1_bio->bios[r1_bio->read_disk];
2012
2013	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2014		/* ouch - failed to read all of that. */
2015		if (!fix_sync_read_error(r1_bio))
2016			return;
2017
2018	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2019		process_checks(r1_bio);
2020
2021	/*
2022	 * schedule writes
2023	 */
2024	atomic_set(&r1_bio->remaining, 1);
2025	for (i = 0; i < disks ; i++) {
2026		wbio = r1_bio->bios[i];
2027		if (wbio->bi_end_io == NULL ||
2028		    (wbio->bi_end_io == end_sync_read &&
2029		     (i == r1_bio->read_disk ||
2030		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2031			continue;
 
 
 
 
 
 
 
 
2032
2033		wbio->bi_rw = WRITE;
2034		wbio->bi_end_io = end_sync_write;
2035		atomic_inc(&r1_bio->remaining);
2036		md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2037
2038		generic_make_request(wbio);
2039	}
2040
2041	if (atomic_dec_and_test(&r1_bio->remaining)) {
2042		/* if we're here, all write(s) have completed, so clean up */
2043		int s = r1_bio->sectors;
2044		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2045		    test_bit(R1BIO_WriteError, &r1_bio->state))
2046			reschedule_retry(r1_bio);
2047		else {
2048			put_buf(r1_bio);
2049			md_done_sync(mddev, s, 1);
2050		}
2051	}
2052}
2053
2054/*
2055 * This is a kernel thread which:
2056 *
2057 *	1.	Retries failed read operations on working mirrors.
2058 *	2.	Updates the raid superblock when problems encounter.
2059 *	3.	Performs writes following reads for array synchronising.
2060 */
2061
2062static void fix_read_error(struct r1conf *conf, int read_disk,
2063			   sector_t sect, int sectors)
2064{
2065	struct mddev *mddev = conf->mddev;
2066	while(sectors) {
2067		int s = sectors;
2068		int d = read_disk;
2069		int success = 0;
2070		int start;
2071		struct md_rdev *rdev;
2072
2073		if (s > (PAGE_SIZE>>9))
2074			s = PAGE_SIZE >> 9;
2075
2076		do {
2077			/* Note: no rcu protection needed here
2078			 * as this is synchronous in the raid1d thread
2079			 * which is the thread that might remove
2080			 * a device.  If raid1d ever becomes multi-threaded....
2081			 */
2082			sector_t first_bad;
2083			int bad_sectors;
2084
2085			rdev = conf->mirrors[d].rdev;
 
2086			if (rdev &&
2087			    (test_bit(In_sync, &rdev->flags) ||
2088			     (!test_bit(Faulty, &rdev->flags) &&
2089			      rdev->recovery_offset >= sect + s)) &&
2090			    is_badblock(rdev, sect, s,
2091					&first_bad, &bad_sectors) == 0 &&
2092			    sync_page_io(rdev, sect, s<<9,
2093					 conf->tmppage, READ, false))
2094				success = 1;
2095			else {
2096				d++;
2097				if (d == conf->raid_disks * 2)
2098					d = 0;
2099			}
 
 
 
 
 
2100		} while (!success && d != read_disk);
2101
2102		if (!success) {
2103			/* Cannot read from anywhere - mark it bad */
2104			struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2105			if (!rdev_set_badblocks(rdev, sect, s, 0))
2106				md_error(mddev, rdev);
2107			break;
2108		}
2109		/* write it back and re-read */
2110		start = d;
2111		while (d != read_disk) {
2112			if (d==0)
2113				d = conf->raid_disks * 2;
2114			d--;
2115			rdev = conf->mirrors[d].rdev;
 
2116			if (rdev &&
2117			    !test_bit(Faulty, &rdev->flags))
 
 
2118				r1_sync_page_io(rdev, sect, s,
2119						conf->tmppage, WRITE);
 
 
 
2120		}
2121		d = start;
2122		while (d != read_disk) {
2123			char b[BDEVNAME_SIZE];
2124			if (d==0)
2125				d = conf->raid_disks * 2;
2126			d--;
2127			rdev = conf->mirrors[d].rdev;
 
2128			if (rdev &&
2129			    !test_bit(Faulty, &rdev->flags)) {
 
 
2130				if (r1_sync_page_io(rdev, sect, s,
2131						    conf->tmppage, READ)) {
2132					atomic_add(s, &rdev->corrected_errors);
2133					printk(KERN_INFO
2134					       "md/raid1:%s: read error corrected "
2135					       "(%d sectors at %llu on %s)\n",
2136					       mdname(mddev), s,
2137					       (unsigned long long)(sect +
2138					           rdev->data_offset),
2139					       bdevname(rdev->bdev, b));
2140				}
2141			}
 
 
2142		}
2143		sectors -= s;
2144		sect += s;
2145	}
2146}
2147
2148static int narrow_write_error(struct r1bio *r1_bio, int i)
2149{
2150	struct mddev *mddev = r1_bio->mddev;
2151	struct r1conf *conf = mddev->private;
2152	struct md_rdev *rdev = conf->mirrors[i].rdev;
2153
2154	/* bio has the data to be written to device 'i' where
2155	 * we just recently had a write error.
2156	 * We repeatedly clone the bio and trim down to one block,
2157	 * then try the write.  Where the write fails we record
2158	 * a bad block.
2159	 * It is conceivable that the bio doesn't exactly align with
2160	 * blocks.  We must handle this somehow.
2161	 *
2162	 * We currently own a reference on the rdev.
2163	 */
2164
2165	int block_sectors;
2166	sector_t sector;
2167	int sectors;
2168	int sect_to_write = r1_bio->sectors;
2169	int ok = 1;
2170
2171	if (rdev->badblocks.shift < 0)
2172		return 0;
2173
2174	block_sectors = roundup(1 << rdev->badblocks.shift,
2175				bdev_logical_block_size(rdev->bdev) >> 9);
2176	sector = r1_bio->sector;
2177	sectors = ((sector + block_sectors)
2178		   & ~(sector_t)(block_sectors - 1))
2179		- sector;
2180
2181	while (sect_to_write) {
2182		struct bio *wbio;
2183		if (sectors > sect_to_write)
2184			sectors = sect_to_write;
2185		/* Write at 'sector' for 'sectors'*/
2186
2187		if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2188			unsigned vcnt = r1_bio->behind_page_count;
2189			struct bio_vec *vec = r1_bio->behind_bvecs;
2190
2191			while (!vec->bv_page) {
2192				vec++;
2193				vcnt--;
2194			}
2195
2196			wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2197			memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2198
2199			wbio->bi_vcnt = vcnt;
2200		} else {
2201			wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
 
2202		}
2203
2204		wbio->bi_rw = WRITE;
2205		wbio->bi_iter.bi_sector = r1_bio->sector;
2206		wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2207
2208		bio_trim(wbio, sector - r1_bio->sector, sectors);
2209		wbio->bi_iter.bi_sector += rdev->data_offset;
2210		wbio->bi_bdev = rdev->bdev;
2211		if (submit_bio_wait(WRITE, wbio) < 0)
 
2212			/* failure! */
2213			ok = rdev_set_badblocks(rdev, sector,
2214						sectors, 0)
2215				&& ok;
2216
2217		bio_put(wbio);
2218		sect_to_write -= sectors;
2219		sector += sectors;
2220		sectors = block_sectors;
2221	}
2222	return ok;
2223}
2224
2225static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2226{
2227	int m;
2228	int s = r1_bio->sectors;
2229	for (m = 0; m < conf->raid_disks * 2 ; m++) {
2230		struct md_rdev *rdev = conf->mirrors[m].rdev;
2231		struct bio *bio = r1_bio->bios[m];
2232		if (bio->bi_end_io == NULL)
2233			continue;
2234		if (!bio->bi_error &&
2235		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2236			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2237		}
2238		if (bio->bi_error &&
2239		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2240			if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2241				md_error(conf->mddev, rdev);
2242		}
2243	}
2244	put_buf(r1_bio);
2245	md_done_sync(conf->mddev, s, 1);
2246}
2247
2248static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2249{
2250	int m;
2251	bool fail = false;
 
2252	for (m = 0; m < conf->raid_disks * 2 ; m++)
2253		if (r1_bio->bios[m] == IO_MADE_GOOD) {
2254			struct md_rdev *rdev = conf->mirrors[m].rdev;
2255			rdev_clear_badblocks(rdev,
2256					     r1_bio->sector,
2257					     r1_bio->sectors, 0);
2258			rdev_dec_pending(rdev, conf->mddev);
2259		} else if (r1_bio->bios[m] != NULL) {
2260			/* This drive got a write error.  We need to
2261			 * narrow down and record precise write
2262			 * errors.
2263			 */
2264			fail = true;
2265			if (!narrow_write_error(r1_bio, m)) {
2266				md_error(conf->mddev,
2267					 conf->mirrors[m].rdev);
2268				/* an I/O failed, we can't clear the bitmap */
2269				set_bit(R1BIO_Degraded, &r1_bio->state);
2270			}
2271			rdev_dec_pending(conf->mirrors[m].rdev,
2272					 conf->mddev);
2273		}
2274	if (fail) {
2275		spin_lock_irq(&conf->device_lock);
2276		list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2277		conf->nr_queued++;
 
2278		spin_unlock_irq(&conf->device_lock);
 
 
 
 
 
2279		md_wakeup_thread(conf->mddev->thread);
2280	} else {
2281		if (test_bit(R1BIO_WriteError, &r1_bio->state))
2282			close_write(r1_bio);
2283		raid_end_bio_io(r1_bio);
2284	}
2285}
2286
2287static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2288{
2289	int disk;
2290	int max_sectors;
2291	struct mddev *mddev = conf->mddev;
2292	struct bio *bio;
2293	char b[BDEVNAME_SIZE];
2294	struct md_rdev *rdev;
2295
2296	clear_bit(R1BIO_ReadError, &r1_bio->state);
2297	/* we got a read error. Maybe the drive is bad.  Maybe just
2298	 * the block and we can fix it.
2299	 * We freeze all other IO, and try reading the block from
2300	 * other devices.  When we find one, we re-write
2301	 * and check it that fixes the read error.
2302	 * This is all done synchronously while the array is
2303	 * frozen
2304	 */
2305	if (mddev->ro == 0) {
 
 
 
 
 
 
 
2306		freeze_array(conf, 1);
2307		fix_read_error(conf, r1_bio->read_disk,
2308			       r1_bio->sector, r1_bio->sectors);
2309		unfreeze_array(conf);
2310	} else
2311		md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2312	rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2313
2314	bio = r1_bio->bios[r1_bio->read_disk];
2315	bdevname(bio->bi_bdev, b);
2316read_more:
2317	disk = read_balance(conf, r1_bio, &max_sectors);
2318	if (disk == -1) {
2319		printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2320		       " read error for block %llu\n",
2321		       mdname(mddev), b, (unsigned long long)r1_bio->sector);
2322		raid_end_bio_io(r1_bio);
2323	} else {
2324		const unsigned long do_sync
2325			= r1_bio->master_bio->bi_rw & REQ_SYNC;
2326		if (bio) {
2327			r1_bio->bios[r1_bio->read_disk] =
2328				mddev->ro ? IO_BLOCKED : NULL;
2329			bio_put(bio);
2330		}
2331		r1_bio->read_disk = disk;
2332		bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2333		bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2334			 max_sectors);
2335		r1_bio->bios[r1_bio->read_disk] = bio;
2336		rdev = conf->mirrors[disk].rdev;
2337		printk_ratelimited(KERN_ERR
2338				   "md/raid1:%s: redirecting sector %llu"
2339				   " to other mirror: %s\n",
2340				   mdname(mddev),
2341				   (unsigned long long)r1_bio->sector,
2342				   bdevname(rdev->bdev, b));
2343		bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2344		bio->bi_bdev = rdev->bdev;
2345		bio->bi_end_io = raid1_end_read_request;
2346		bio->bi_rw = READ | do_sync;
2347		bio->bi_private = r1_bio;
2348		if (max_sectors < r1_bio->sectors) {
2349			/* Drat - have to split this up more */
2350			struct bio *mbio = r1_bio->master_bio;
2351			int sectors_handled = (r1_bio->sector + max_sectors
2352					       - mbio->bi_iter.bi_sector);
2353			r1_bio->sectors = max_sectors;
2354			spin_lock_irq(&conf->device_lock);
2355			if (mbio->bi_phys_segments == 0)
2356				mbio->bi_phys_segments = 2;
2357			else
2358				mbio->bi_phys_segments++;
2359			spin_unlock_irq(&conf->device_lock);
2360			generic_make_request(bio);
2361			bio = NULL;
2362
2363			r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2364
2365			r1_bio->master_bio = mbio;
2366			r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2367			r1_bio->state = 0;
2368			set_bit(R1BIO_ReadError, &r1_bio->state);
2369			r1_bio->mddev = mddev;
2370			r1_bio->sector = mbio->bi_iter.bi_sector +
2371				sectors_handled;
2372
2373			goto read_more;
2374		} else
2375			generic_make_request(bio);
2376	}
 
 
 
 
 
 
 
 
2377}
2378
2379static void raid1d(struct md_thread *thread)
2380{
2381	struct mddev *mddev = thread->mddev;
2382	struct r1bio *r1_bio;
2383	unsigned long flags;
2384	struct r1conf *conf = mddev->private;
2385	struct list_head *head = &conf->retry_list;
2386	struct blk_plug plug;
 
2387
2388	md_check_recovery(mddev);
2389
2390	if (!list_empty_careful(&conf->bio_end_io_list) &&
2391	    !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2392		LIST_HEAD(tmp);
2393		spin_lock_irqsave(&conf->device_lock, flags);
2394		if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2395			while (!list_empty(&conf->bio_end_io_list)) {
2396				list_move(conf->bio_end_io_list.prev, &tmp);
2397				conf->nr_queued--;
2398			}
2399		}
2400		spin_unlock_irqrestore(&conf->device_lock, flags);
2401		while (!list_empty(&tmp)) {
2402			r1_bio = list_first_entry(&tmp, struct r1bio,
2403						  retry_list);
2404			list_del(&r1_bio->retry_list);
 
 
2405			if (mddev->degraded)
2406				set_bit(R1BIO_Degraded, &r1_bio->state);
2407			if (test_bit(R1BIO_WriteError, &r1_bio->state))
2408				close_write(r1_bio);
2409			raid_end_bio_io(r1_bio);
2410		}
2411	}
2412
2413	blk_start_plug(&plug);
2414	for (;;) {
2415
2416		flush_pending_writes(conf);
2417
2418		spin_lock_irqsave(&conf->device_lock, flags);
2419		if (list_empty(head)) {
2420			spin_unlock_irqrestore(&conf->device_lock, flags);
2421			break;
2422		}
2423		r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2424		list_del(head->prev);
2425		conf->nr_queued--;
 
2426		spin_unlock_irqrestore(&conf->device_lock, flags);
2427
2428		mddev = r1_bio->mddev;
2429		conf = mddev->private;
2430		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2431			if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2432			    test_bit(R1BIO_WriteError, &r1_bio->state))
2433				handle_sync_write_finished(conf, r1_bio);
2434			else
2435				sync_request_write(mddev, r1_bio);
2436		} else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2437			   test_bit(R1BIO_WriteError, &r1_bio->state))
2438			handle_write_finished(conf, r1_bio);
2439		else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2440			handle_read_error(conf, r1_bio);
2441		else
2442			/* just a partial read to be scheduled from separate
2443			 * context
2444			 */
2445			generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2446
2447		cond_resched();
2448		if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2449			md_check_recovery(mddev);
2450	}
2451	blk_finish_plug(&plug);
2452}
2453
2454static int init_resync(struct r1conf *conf)
2455{
2456	int buffs;
2457
2458	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2459	BUG_ON(conf->r1buf_pool);
2460	conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2461					  conf->poolinfo);
2462	if (!conf->r1buf_pool)
2463		return -ENOMEM;
2464	conf->next_resync = 0;
2465	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2466}
2467
2468/*
2469 * perform a "sync" on one "block"
2470 *
2471 * We need to make sure that no normal I/O request - particularly write
2472 * requests - conflict with active sync requests.
2473 *
2474 * This is achieved by tracking pending requests and a 'barrier' concept
2475 * that can be installed to exclude normal IO requests.
2476 */
2477
2478static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2479				   int *skipped)
2480{
2481	struct r1conf *conf = mddev->private;
2482	struct r1bio *r1_bio;
2483	struct bio *bio;
2484	sector_t max_sector, nr_sectors;
2485	int disk = -1;
2486	int i;
2487	int wonly = -1;
2488	int write_targets = 0, read_targets = 0;
2489	sector_t sync_blocks;
2490	int still_degraded = 0;
2491	int good_sectors = RESYNC_SECTORS;
2492	int min_bad = 0; /* number of sectors that are bad in all devices */
 
 
2493
2494	if (!conf->r1buf_pool)
2495		if (init_resync(conf))
2496			return 0;
2497
2498	max_sector = mddev->dev_sectors;
2499	if (sector_nr >= max_sector) {
2500		/* If we aborted, we need to abort the
2501		 * sync on the 'current' bitmap chunk (there will
2502		 * only be one in raid1 resync.
2503		 * We can find the current addess in mddev->curr_resync
2504		 */
2505		if (mddev->curr_resync < max_sector) /* aborted */
2506			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2507						&sync_blocks, 1);
2508		else /* completed sync */
2509			conf->fullsync = 0;
2510
2511		bitmap_close_sync(mddev->bitmap);
2512		close_sync(conf);
2513
2514		if (mddev_is_clustered(mddev)) {
2515			conf->cluster_sync_low = 0;
2516			conf->cluster_sync_high = 0;
2517		}
2518		return 0;
2519	}
2520
2521	if (mddev->bitmap == NULL &&
2522	    mddev->recovery_cp == MaxSector &&
2523	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2524	    conf->fullsync == 0) {
2525		*skipped = 1;
2526		return max_sector - sector_nr;
2527	}
2528	/* before building a request, check if we can skip these blocks..
2529	 * This call the bitmap_start_sync doesn't actually record anything
2530	 */
2531	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2532	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2533		/* We can skip this block, and probably several more */
2534		*skipped = 1;
2535		return sync_blocks;
2536	}
2537
 
 
 
 
 
 
 
2538	/* we are incrementing sector_nr below. To be safe, we check against
2539	 * sector_nr + two times RESYNC_SECTORS
2540	 */
2541
2542	bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2543		mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2544	r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2545
2546	raise_barrier(conf, sector_nr);
 
 
 
 
2547
2548	rcu_read_lock();
2549	/*
2550	 * If we get a correctably read error during resync or recovery,
2551	 * we might want to read from a different device.  So we
2552	 * flag all drives that could conceivably be read from for READ,
2553	 * and any others (which will be non-In_sync devices) for WRITE.
2554	 * If a read fails, we try reading from something else for which READ
2555	 * is OK.
2556	 */
2557
2558	r1_bio->mddev = mddev;
2559	r1_bio->sector = sector_nr;
2560	r1_bio->state = 0;
2561	set_bit(R1BIO_IsSync, &r1_bio->state);
 
 
2562
2563	for (i = 0; i < conf->raid_disks * 2; i++) {
2564		struct md_rdev *rdev;
2565		bio = r1_bio->bios[i];
2566		bio_reset(bio);
2567
2568		rdev = rcu_dereference(conf->mirrors[i].rdev);
2569		if (rdev == NULL ||
2570		    test_bit(Faulty, &rdev->flags)) {
2571			if (i < conf->raid_disks)
2572				still_degraded = 1;
2573		} else if (!test_bit(In_sync, &rdev->flags)) {
2574			bio->bi_rw = WRITE;
2575			bio->bi_end_io = end_sync_write;
2576			write_targets ++;
2577		} else {
2578			/* may need to read from here */
2579			sector_t first_bad = MaxSector;
2580			int bad_sectors;
2581
2582			if (is_badblock(rdev, sector_nr, good_sectors,
2583					&first_bad, &bad_sectors)) {
2584				if (first_bad > sector_nr)
2585					good_sectors = first_bad - sector_nr;
2586				else {
2587					bad_sectors -= (sector_nr - first_bad);
2588					if (min_bad == 0 ||
2589					    min_bad > bad_sectors)
2590						min_bad = bad_sectors;
2591				}
2592			}
2593			if (sector_nr < first_bad) {
2594				if (test_bit(WriteMostly, &rdev->flags)) {
2595					if (wonly < 0)
2596						wonly = i;
2597				} else {
2598					if (disk < 0)
2599						disk = i;
2600				}
2601				bio->bi_rw = READ;
2602				bio->bi_end_io = end_sync_read;
2603				read_targets++;
2604			} else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2605				test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2606				!test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2607				/*
2608				 * The device is suitable for reading (InSync),
2609				 * but has bad block(s) here. Let's try to correct them,
2610				 * if we are doing resync or repair. Otherwise, leave
2611				 * this device alone for this sync request.
2612				 */
2613				bio->bi_rw = WRITE;
2614				bio->bi_end_io = end_sync_write;
2615				write_targets++;
2616			}
2617		}
2618		if (bio->bi_end_io) {
2619			atomic_inc(&rdev->nr_pending);
2620			bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2621			bio->bi_bdev = rdev->bdev;
2622			bio->bi_private = r1_bio;
 
2623		}
2624	}
2625	rcu_read_unlock();
2626	if (disk < 0)
2627		disk = wonly;
2628	r1_bio->read_disk = disk;
2629
2630	if (read_targets == 0 && min_bad > 0) {
2631		/* These sectors are bad on all InSync devices, so we
2632		 * need to mark them bad on all write targets
2633		 */
2634		int ok = 1;
2635		for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2636			if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2637				struct md_rdev *rdev = conf->mirrors[i].rdev;
2638				ok = rdev_set_badblocks(rdev, sector_nr,
2639							min_bad, 0
2640					) && ok;
2641			}
2642		set_bit(MD_CHANGE_DEVS, &mddev->flags);
2643		*skipped = 1;
2644		put_buf(r1_bio);
2645
2646		if (!ok) {
2647			/* Cannot record the badblocks, so need to
2648			 * abort the resync.
2649			 * If there are multiple read targets, could just
2650			 * fail the really bad ones ???
2651			 */
2652			conf->recovery_disabled = mddev->recovery_disabled;
2653			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2654			return 0;
2655		} else
2656			return min_bad;
2657
2658	}
2659	if (min_bad > 0 && min_bad < good_sectors) {
2660		/* only resync enough to reach the next bad->good
2661		 * transition */
2662		good_sectors = min_bad;
2663	}
2664
2665	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2666		/* extra read targets are also write targets */
2667		write_targets += read_targets-1;
2668
2669	if (write_targets == 0 || read_targets == 0) {
2670		/* There is nowhere to write, so all non-sync
2671		 * drives must be failed - so we are finished
2672		 */
2673		sector_t rv;
2674		if (min_bad > 0)
2675			max_sector = sector_nr + min_bad;
2676		rv = max_sector - sector_nr;
2677		*skipped = 1;
2678		put_buf(r1_bio);
2679		return rv;
2680	}
2681
2682	if (max_sector > mddev->resync_max)
2683		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2684	if (max_sector > sector_nr + good_sectors)
2685		max_sector = sector_nr + good_sectors;
2686	nr_sectors = 0;
2687	sync_blocks = 0;
2688	do {
2689		struct page *page;
2690		int len = PAGE_SIZE;
2691		if (sector_nr + (len>>9) > max_sector)
2692			len = (max_sector - sector_nr) << 9;
2693		if (len == 0)
2694			break;
2695		if (sync_blocks == 0) {
2696			if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2697					       &sync_blocks, still_degraded) &&
2698			    !conf->fullsync &&
2699			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2700				break;
2701			if ((len >> 9) > sync_blocks)
2702				len = sync_blocks<<9;
2703		}
2704
2705		for (i = 0 ; i < conf->raid_disks * 2; i++) {
 
 
2706			bio = r1_bio->bios[i];
 
2707			if (bio->bi_end_io) {
2708				page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2709				if (bio_add_page(bio, page, len, 0) == 0) {
2710					/* stop here */
2711					bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2712					while (i > 0) {
2713						i--;
2714						bio = r1_bio->bios[i];
2715						if (bio->bi_end_io==NULL)
2716							continue;
2717						/* remove last page from this bio */
2718						bio->bi_vcnt--;
2719						bio->bi_iter.bi_size -= len;
2720						bio_clear_flag(bio, BIO_SEG_VALID);
2721					}
2722					goto bio_full;
2723				}
2724			}
2725		}
2726		nr_sectors += len>>9;
2727		sector_nr += len>>9;
2728		sync_blocks -= (len>>9);
2729	} while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2730 bio_full:
2731	r1_bio->sectors = nr_sectors;
2732
2733	if (mddev_is_clustered(mddev) &&
2734			conf->cluster_sync_high < sector_nr + nr_sectors) {
2735		conf->cluster_sync_low = mddev->curr_resync_completed;
2736		conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2737		/* Send resync message */
2738		md_cluster_ops->resync_info_update(mddev,
2739				conf->cluster_sync_low,
2740				conf->cluster_sync_high);
2741	}
2742
2743	/* For a user-requested sync, we read all readable devices and do a
2744	 * compare
2745	 */
2746	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2747		atomic_set(&r1_bio->remaining, read_targets);
2748		for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2749			bio = r1_bio->bios[i];
2750			if (bio->bi_end_io == end_sync_read) {
2751				read_targets--;
2752				md_sync_acct(bio->bi_bdev, nr_sectors);
2753				generic_make_request(bio);
 
 
2754			}
2755		}
2756	} else {
2757		atomic_set(&r1_bio->remaining, 1);
2758		bio = r1_bio->bios[r1_bio->read_disk];
2759		md_sync_acct(bio->bi_bdev, nr_sectors);
2760		generic_make_request(bio);
2761
 
2762	}
2763	return nr_sectors;
2764}
2765
2766static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2767{
2768	if (sectors)
2769		return sectors;
2770
2771	return mddev->dev_sectors;
2772}
2773
2774static struct r1conf *setup_conf(struct mddev *mddev)
2775{
2776	struct r1conf *conf;
2777	int i;
2778	struct raid1_info *disk;
2779	struct md_rdev *rdev;
2780	int err = -ENOMEM;
2781
2782	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2783	if (!conf)
2784		goto abort;
2785
2786	conf->mirrors = kzalloc(sizeof(struct raid1_info)
2787				* mddev->raid_disks * 2,
2788				 GFP_KERNEL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2789	if (!conf->mirrors)
2790		goto abort;
2791
2792	conf->tmppage = alloc_page(GFP_KERNEL);
2793	if (!conf->tmppage)
2794		goto abort;
2795
2796	conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2797	if (!conf->poolinfo)
2798		goto abort;
2799	conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2800	conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2801					  r1bio_pool_free,
2802					  conf->poolinfo);
2803	if (!conf->r1bio_pool)
 
 
 
2804		goto abort;
2805
2806	conf->poolinfo->mddev = mddev;
2807
2808	err = -EINVAL;
2809	spin_lock_init(&conf->device_lock);
2810	rdev_for_each(rdev, mddev) {
2811		struct request_queue *q;
2812		int disk_idx = rdev->raid_disk;
2813		if (disk_idx >= mddev->raid_disks
2814		    || disk_idx < 0)
2815			continue;
2816		if (test_bit(Replacement, &rdev->flags))
2817			disk = conf->mirrors + mddev->raid_disks + disk_idx;
2818		else
2819			disk = conf->mirrors + disk_idx;
2820
2821		if (disk->rdev)
2822			goto abort;
2823		disk->rdev = rdev;
2824		q = bdev_get_queue(rdev->bdev);
2825
2826		disk->head_position = 0;
2827		disk->seq_start = MaxSector;
2828	}
2829	conf->raid_disks = mddev->raid_disks;
2830	conf->mddev = mddev;
2831	INIT_LIST_HEAD(&conf->retry_list);
2832	INIT_LIST_HEAD(&conf->bio_end_io_list);
2833
2834	spin_lock_init(&conf->resync_lock);
2835	init_waitqueue_head(&conf->wait_barrier);
2836
2837	bio_list_init(&conf->pending_bio_list);
2838	conf->pending_count = 0;
2839	conf->recovery_disabled = mddev->recovery_disabled - 1;
2840
2841	conf->start_next_window = MaxSector;
2842	conf->current_window_requests = conf->next_window_requests = 0;
2843
2844	err = -EIO;
2845	for (i = 0; i < conf->raid_disks * 2; i++) {
2846
2847		disk = conf->mirrors + i;
2848
2849		if (i < conf->raid_disks &&
2850		    disk[conf->raid_disks].rdev) {
2851			/* This slot has a replacement. */
2852			if (!disk->rdev) {
2853				/* No original, just make the replacement
2854				 * a recovering spare
2855				 */
2856				disk->rdev =
2857					disk[conf->raid_disks].rdev;
2858				disk[conf->raid_disks].rdev = NULL;
2859			} else if (!test_bit(In_sync, &disk->rdev->flags))
2860				/* Original is not in_sync - bad */
2861				goto abort;
2862		}
2863
2864		if (!disk->rdev ||
2865		    !test_bit(In_sync, &disk->rdev->flags)) {
2866			disk->head_position = 0;
2867			if (disk->rdev &&
2868			    (disk->rdev->saved_raid_disk < 0))
2869				conf->fullsync = 1;
2870		}
2871	}
2872
2873	err = -ENOMEM;
2874	conf->thread = md_register_thread(raid1d, mddev, "raid1");
2875	if (!conf->thread) {
2876		printk(KERN_ERR
2877		       "md/raid1:%s: couldn't allocate thread\n",
2878		       mdname(mddev));
2879		goto abort;
2880	}
2881
2882	return conf;
2883
2884 abort:
2885	if (conf) {
2886		mempool_destroy(conf->r1bio_pool);
2887		kfree(conf->mirrors);
2888		safe_put_page(conf->tmppage);
2889		kfree(conf->poolinfo);
 
 
 
 
 
2890		kfree(conf);
2891	}
2892	return ERR_PTR(err);
2893}
2894
2895static void raid1_free(struct mddev *mddev, void *priv);
2896static int raid1_run(struct mddev *mddev)
2897{
2898	struct r1conf *conf;
2899	int i;
2900	struct md_rdev *rdev;
2901	int ret;
2902	bool discard_supported = false;
2903
2904	if (mddev->level != 1) {
2905		printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2906		       mdname(mddev), mddev->level);
2907		return -EIO;
2908	}
2909	if (mddev->reshape_position != MaxSector) {
2910		printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2911		       mdname(mddev));
2912		return -EIO;
2913	}
 
 
2914	/*
2915	 * copy the already verified devices into our private RAID1
2916	 * bookkeeping area. [whatever we allocate in run(),
2917	 * should be freed in raid1_free()]
2918	 */
2919	if (mddev->private == NULL)
2920		conf = setup_conf(mddev);
2921	else
2922		conf = mddev->private;
2923
2924	if (IS_ERR(conf))
2925		return PTR_ERR(conf);
2926
2927	if (mddev->queue)
2928		blk_queue_max_write_same_sectors(mddev->queue, 0);
 
 
2929
2930	rdev_for_each(rdev, mddev) {
2931		if (!mddev->gendisk)
2932			continue;
2933		disk_stack_limits(mddev->gendisk, rdev->bdev,
2934				  rdev->data_offset << 9);
2935		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2936			discard_supported = true;
2937	}
2938
2939	mddev->degraded = 0;
2940	for (i=0; i < conf->raid_disks; i++)
2941		if (conf->mirrors[i].rdev == NULL ||
2942		    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2943		    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2944			mddev->degraded++;
 
 
 
 
 
 
 
2945
2946	if (conf->raid_disks - mddev->degraded == 1)
2947		mddev->recovery_cp = MaxSector;
2948
2949	if (mddev->recovery_cp != MaxSector)
2950		printk(KERN_NOTICE "md/raid1:%s: not clean"
2951		       " -- starting background reconstruction\n",
2952		       mdname(mddev));
2953	printk(KERN_INFO
2954		"md/raid1:%s: active with %d out of %d mirrors\n",
2955		mdname(mddev), mddev->raid_disks - mddev->degraded,
2956		mddev->raid_disks);
2957
2958	/*
2959	 * Ok, everything is just fine now
2960	 */
2961	mddev->thread = conf->thread;
2962	conf->thread = NULL;
2963	mddev->private = conf;
 
2964
2965	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2966
2967	if (mddev->queue) {
2968		if (discard_supported)
2969			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2970						mddev->queue);
2971		else
2972			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2973						  mddev->queue);
2974	}
2975
2976	ret =  md_integrity_register(mddev);
2977	if (ret) {
2978		md_unregister_thread(&mddev->thread);
2979		raid1_free(mddev, conf);
2980	}
 
 
 
 
2981	return ret;
2982}
2983
2984static void raid1_free(struct mddev *mddev, void *priv)
2985{
2986	struct r1conf *conf = priv;
2987
2988	mempool_destroy(conf->r1bio_pool);
2989	kfree(conf->mirrors);
2990	safe_put_page(conf->tmppage);
2991	kfree(conf->poolinfo);
 
 
 
 
 
2992	kfree(conf);
2993}
2994
2995static int raid1_resize(struct mddev *mddev, sector_t sectors)
2996{
2997	/* no resync is happening, and there is enough space
2998	 * on all devices, so we can resize.
2999	 * We need to make sure resync covers any new space.
3000	 * If the array is shrinking we should possibly wait until
3001	 * any io in the removed space completes, but it hardly seems
3002	 * worth it.
3003	 */
3004	sector_t newsize = raid1_size(mddev, sectors, 0);
3005	if (mddev->external_size &&
3006	    mddev->array_sectors > newsize)
3007		return -EINVAL;
3008	if (mddev->bitmap) {
3009		int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3010		if (ret)
3011			return ret;
3012	}
3013	md_set_array_sectors(mddev, newsize);
3014	set_capacity(mddev->gendisk, mddev->array_sectors);
3015	revalidate_disk(mddev->gendisk);
3016	if (sectors > mddev->dev_sectors &&
3017	    mddev->recovery_cp > mddev->dev_sectors) {
3018		mddev->recovery_cp = mddev->dev_sectors;
3019		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3020	}
3021	mddev->dev_sectors = sectors;
3022	mddev->resync_max_sectors = sectors;
3023	return 0;
3024}
3025
3026static int raid1_reshape(struct mddev *mddev)
3027{
3028	/* We need to:
3029	 * 1/ resize the r1bio_pool
3030	 * 2/ resize conf->mirrors
3031	 *
3032	 * We allocate a new r1bio_pool if we can.
3033	 * Then raise a device barrier and wait until all IO stops.
3034	 * Then resize conf->mirrors and swap in the new r1bio pool.
3035	 *
3036	 * At the same time, we "pack" the devices so that all the missing
3037	 * devices have the higher raid_disk numbers.
3038	 */
3039	mempool_t *newpool, *oldpool;
3040	struct pool_info *newpoolinfo;
3041	struct raid1_info *newmirrors;
3042	struct r1conf *conf = mddev->private;
3043	int cnt, raid_disks;
3044	unsigned long flags;
3045	int d, d2, err;
 
 
 
 
3046
3047	/* Cannot change chunk_size, layout, or level */
3048	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3049	    mddev->layout != mddev->new_layout ||
3050	    mddev->level != mddev->new_level) {
3051		mddev->new_chunk_sectors = mddev->chunk_sectors;
3052		mddev->new_layout = mddev->layout;
3053		mddev->new_level = mddev->level;
3054		return -EINVAL;
3055	}
3056
3057	if (!mddev_is_clustered(mddev)) {
3058		err = md_allow_write(mddev);
3059		if (err)
3060			return err;
3061	}
3062
3063	raid_disks = mddev->raid_disks + mddev->delta_disks;
3064
3065	if (raid_disks < conf->raid_disks) {
3066		cnt=0;
3067		for (d= 0; d < conf->raid_disks; d++)
3068			if (conf->mirrors[d].rdev)
3069				cnt++;
3070		if (cnt > raid_disks)
3071			return -EBUSY;
3072	}
3073
3074	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3075	if (!newpoolinfo)
3076		return -ENOMEM;
3077	newpoolinfo->mddev = mddev;
3078	newpoolinfo->raid_disks = raid_disks * 2;
3079
3080	newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3081				 r1bio_pool_free, newpoolinfo);
3082	if (!newpool) {
3083		kfree(newpoolinfo);
3084		return -ENOMEM;
3085	}
3086	newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
 
3087			     GFP_KERNEL);
3088	if (!newmirrors) {
3089		kfree(newpoolinfo);
3090		mempool_destroy(newpool);
3091		return -ENOMEM;
3092	}
3093
3094	freeze_array(conf, 0);
3095
3096	/* ok, everything is stopped */
3097	oldpool = conf->r1bio_pool;
3098	conf->r1bio_pool = newpool;
3099
3100	for (d = d2 = 0; d < conf->raid_disks; d++) {
3101		struct md_rdev *rdev = conf->mirrors[d].rdev;
3102		if (rdev && rdev->raid_disk != d2) {
3103			sysfs_unlink_rdev(mddev, rdev);
3104			rdev->raid_disk = d2;
3105			sysfs_unlink_rdev(mddev, rdev);
3106			if (sysfs_link_rdev(mddev, rdev))
3107				printk(KERN_WARNING
3108				       "md/raid1:%s: cannot register rd%d\n",
3109				       mdname(mddev), rdev->raid_disk);
3110		}
3111		if (rdev)
3112			newmirrors[d2++].rdev = rdev;
3113	}
3114	kfree(conf->mirrors);
3115	conf->mirrors = newmirrors;
3116	kfree(conf->poolinfo);
3117	conf->poolinfo = newpoolinfo;
3118
3119	spin_lock_irqsave(&conf->device_lock, flags);
3120	mddev->degraded += (raid_disks - conf->raid_disks);
3121	spin_unlock_irqrestore(&conf->device_lock, flags);
3122	conf->raid_disks = mddev->raid_disks = raid_disks;
3123	mddev->delta_disks = 0;
3124
3125	unfreeze_array(conf);
3126
3127	set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3128	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3129	md_wakeup_thread(mddev->thread);
3130
3131	mempool_destroy(oldpool);
3132	return 0;
3133}
3134
3135static void raid1_quiesce(struct mddev *mddev, int state)
3136{
3137	struct r1conf *conf = mddev->private;
3138
3139	switch(state) {
3140	case 2: /* wake for suspend */
3141		wake_up(&conf->wait_barrier);
3142		break;
3143	case 1:
3144		freeze_array(conf, 0);
3145		break;
3146	case 0:
3147		unfreeze_array(conf);
3148		break;
3149	}
3150}
3151
3152static void *raid1_takeover(struct mddev *mddev)
3153{
3154	/* raid1 can take over:
3155	 *  raid5 with 2 devices, any layout or chunk size
3156	 */
3157	if (mddev->level == 5 && mddev->raid_disks == 2) {
3158		struct r1conf *conf;
3159		mddev->new_level = 1;
3160		mddev->new_layout = 0;
3161		mddev->new_chunk_sectors = 0;
3162		conf = setup_conf(mddev);
3163		if (!IS_ERR(conf))
3164			/* Array must appear to be quiesced */
3165			conf->array_frozen = 1;
 
 
 
3166		return conf;
3167	}
3168	return ERR_PTR(-EINVAL);
3169}
3170
3171static struct md_personality raid1_personality =
3172{
3173	.name		= "raid1",
3174	.level		= 1,
3175	.owner		= THIS_MODULE,
3176	.make_request	= raid1_make_request,
3177	.run		= raid1_run,
3178	.free		= raid1_free,
3179	.status		= raid1_status,
3180	.error_handler	= raid1_error,
3181	.hot_add_disk	= raid1_add_disk,
3182	.hot_remove_disk= raid1_remove_disk,
3183	.spare_active	= raid1_spare_active,
3184	.sync_request	= raid1_sync_request,
3185	.resize		= raid1_resize,
3186	.size		= raid1_size,
3187	.check_reshape	= raid1_reshape,
3188	.quiesce	= raid1_quiesce,
3189	.takeover	= raid1_takeover,
3190	.congested	= raid1_congested,
3191};
3192
3193static int __init raid_init(void)
3194{
3195	return register_md_personality(&raid1_personality);
3196}
3197
3198static void raid_exit(void)
3199{
3200	unregister_md_personality(&raid1_personality);
3201}
3202
3203module_init(raid_init);
3204module_exit(raid_exit);
3205MODULE_LICENSE("GPL");
3206MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3207MODULE_ALIAS("md-personality-3"); /* RAID1 */
3208MODULE_ALIAS("md-raid1");
3209MODULE_ALIAS("md-level-1");
3210
3211module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
v5.9
   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#include "raid1-10.c"
  53
  54#define START(node) ((node)->start)
  55#define LAST(node) ((node)->last)
  56INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  57		     START, LAST, static inline, raid1_rb);
  58
  59static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  60				struct serial_info *si, int idx)
  61{
  62	unsigned long flags;
  63	int ret = 0;
  64	sector_t lo = r1_bio->sector;
  65	sector_t hi = lo + r1_bio->sectors;
  66	struct serial_in_rdev *serial = &rdev->serial[idx];
  67
  68	spin_lock_irqsave(&serial->serial_lock, flags);
  69	/* collision happened */
  70	if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  71		ret = -EBUSY;
  72	else {
  73		si->start = lo;
  74		si->last = hi;
  75		raid1_rb_insert(si, &serial->serial_rb);
  76	}
  77	spin_unlock_irqrestore(&serial->serial_lock, flags);
  78
  79	return ret;
  80}
  81
  82static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  83{
  84	struct mddev *mddev = rdev->mddev;
  85	struct serial_info *si;
  86	int idx = sector_to_idx(r1_bio->sector);
  87	struct serial_in_rdev *serial = &rdev->serial[idx];
  88
  89	if (WARN_ON(!mddev->serial_info_pool))
  90		return;
  91	si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  92	wait_event(serial->serial_io_wait,
  93		   check_and_add_serial(rdev, r1_bio, si, idx) == 0);
  94}
  95
  96static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  97{
  98	struct serial_info *si;
  99	unsigned long flags;
 100	int found = 0;
 101	struct mddev *mddev = rdev->mddev;
 102	int idx = sector_to_idx(lo);
 103	struct serial_in_rdev *serial = &rdev->serial[idx];
 104
 105	spin_lock_irqsave(&serial->serial_lock, flags);
 106	for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 107	     si; si = raid1_rb_iter_next(si, lo, hi)) {
 108		if (si->start == lo && si->last == hi) {
 109			raid1_rb_remove(si, &serial->serial_rb);
 110			mempool_free(si, mddev->serial_info_pool);
 111			found = 1;
 112			break;
 113		}
 114	}
 115	if (!found)
 116		WARN(1, "The write IO is not recorded for serialization\n");
 117	spin_unlock_irqrestore(&serial->serial_lock, flags);
 118	wake_up(&serial->serial_io_wait);
 119}
 120
 121/*
 122 * for resync bio, r1bio pointer can be retrieved from the per-bio
 123 * 'struct resync_pages'.
 124 */
 125static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 126{
 127	return get_resync_pages(bio)->raid_bio;
 128}
 129
 130static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 131{
 132	struct pool_info *pi = data;
 133	int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 134
 135	/* allocate a r1bio with room for raid_disks entries in the bios array */
 136	return kzalloc(size, gfp_flags);
 137}
 138
 
 
 
 
 
 
 139#define RESYNC_DEPTH 32
 140#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 
 141#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 142#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 143#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 144#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 
 145
 146static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 147{
 148	struct pool_info *pi = data;
 149	struct r1bio *r1_bio;
 150	struct bio *bio;
 151	int need_pages;
 152	int j;
 153	struct resync_pages *rps;
 154
 155	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 156	if (!r1_bio)
 157		return NULL;
 158
 159	rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 160			    gfp_flags);
 161	if (!rps)
 162		goto out_free_r1bio;
 163
 164	/*
 165	 * Allocate bios : 1 for reading, n-1 for writing
 166	 */
 167	for (j = pi->raid_disks ; j-- ; ) {
 168		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 169		if (!bio)
 170			goto out_free_bio;
 171		r1_bio->bios[j] = bio;
 172	}
 173	/*
 174	 * Allocate RESYNC_PAGES data pages and attach them to
 175	 * the first bio.
 176	 * If this is a user-requested check/repair, allocate
 177	 * RESYNC_PAGES for each bio.
 178	 */
 179	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 180		need_pages = pi->raid_disks;
 181	else
 182		need_pages = 1;
 183	for (j = 0; j < pi->raid_disks; j++) {
 184		struct resync_pages *rp = &rps[j];
 185
 186		bio = r1_bio->bios[j];
 
 187
 188		if (j < need_pages) {
 189			if (resync_alloc_pages(rp, gfp_flags))
 190				goto out_free_pages;
 191		} else {
 192			memcpy(rp, &rps[0], sizeof(*rp));
 193			resync_get_all_pages(rp);
 194		}
 195
 196		rp->raid_bio = r1_bio;
 197		bio->bi_private = rp;
 198	}
 199
 200	r1_bio->master_bio = NULL;
 201
 202	return r1_bio;
 203
 204out_free_pages:
 205	while (--j >= 0)
 206		resync_free_pages(&rps[j]);
 
 
 
 
 207
 208out_free_bio:
 209	while (++j < pi->raid_disks)
 210		bio_put(r1_bio->bios[j]);
 211	kfree(rps);
 212
 213out_free_r1bio:
 214	rbio_pool_free(r1_bio, data);
 215	return NULL;
 216}
 217
 218static void r1buf_pool_free(void *__r1_bio, void *data)
 219{
 220	struct pool_info *pi = data;
 221	int i;
 222	struct r1bio *r1bio = __r1_bio;
 223	struct resync_pages *rp = NULL;
 224
 225	for (i = pi->raid_disks; i--; ) {
 226		rp = get_resync_pages(r1bio->bios[i]);
 227		resync_free_pages(rp);
 
 
 
 
 
 228		bio_put(r1bio->bios[i]);
 229	}
 230
 231	/* resync pages array stored in the 1st bio's .bi_private */
 232	kfree(rp);
 233
 234	rbio_pool_free(r1bio, data);
 235}
 236
 237static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 238{
 239	int i;
 240
 241	for (i = 0; i < conf->raid_disks * 2; i++) {
 242		struct bio **bio = r1_bio->bios + i;
 243		if (!BIO_SPECIAL(*bio))
 244			bio_put(*bio);
 245		*bio = NULL;
 246	}
 247}
 248
 249static void free_r1bio(struct r1bio *r1_bio)
 250{
 251	struct r1conf *conf = r1_bio->mddev->private;
 252
 253	put_all_bios(conf, r1_bio);
 254	mempool_free(r1_bio, &conf->r1bio_pool);
 255}
 256
 257static void put_buf(struct r1bio *r1_bio)
 258{
 259	struct r1conf *conf = r1_bio->mddev->private;
 260	sector_t sect = r1_bio->sector;
 261	int i;
 262
 263	for (i = 0; i < conf->raid_disks * 2; i++) {
 264		struct bio *bio = r1_bio->bios[i];
 265		if (bio->bi_end_io)
 266			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 267	}
 268
 269	mempool_free(r1_bio, &conf->r1buf_pool);
 270
 271	lower_barrier(conf, sect);
 272}
 273
 274static void reschedule_retry(struct r1bio *r1_bio)
 275{
 276	unsigned long flags;
 277	struct mddev *mddev = r1_bio->mddev;
 278	struct r1conf *conf = mddev->private;
 279	int idx;
 280
 281	idx = sector_to_idx(r1_bio->sector);
 282	spin_lock_irqsave(&conf->device_lock, flags);
 283	list_add(&r1_bio->retry_list, &conf->retry_list);
 284	atomic_inc(&conf->nr_queued[idx]);
 285	spin_unlock_irqrestore(&conf->device_lock, flags);
 286
 287	wake_up(&conf->wait_barrier);
 288	md_wakeup_thread(mddev->thread);
 289}
 290
 291/*
 292 * raid_end_bio_io() is called when we have finished servicing a mirrored
 293 * operation and are ready to return a success/failure code to the buffer
 294 * cache layer.
 295 */
 296static void call_bio_endio(struct r1bio *r1_bio)
 297{
 298	struct bio *bio = r1_bio->master_bio;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 299
 300	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 301		bio->bi_status = BLK_STS_IOERR;
 302
 303	bio_endio(bio);
 
 
 
 
 
 
 
 304}
 305
 306static void raid_end_bio_io(struct r1bio *r1_bio)
 307{
 308	struct bio *bio = r1_bio->master_bio;
 309	struct r1conf *conf = r1_bio->mddev->private;
 310
 311	/* if nobody has done the final endio yet, do it now */
 312	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 313		pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 314			 (bio_data_dir(bio) == WRITE) ? "write" : "read",
 315			 (unsigned long long) bio->bi_iter.bi_sector,
 316			 (unsigned long long) bio_end_sector(bio) - 1);
 317
 318		call_bio_endio(r1_bio);
 319	}
 320	/*
 321	 * Wake up any possible resync thread that waits for the device
 322	 * to go idle.  All I/Os, even write-behind writes, are done.
 323	 */
 324	allow_barrier(conf, r1_bio->sector);
 325
 326	free_r1bio(r1_bio);
 327}
 328
 329/*
 330 * Update disk head position estimator based on IRQ completion info.
 331 */
 332static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 333{
 334	struct r1conf *conf = r1_bio->mddev->private;
 335
 336	conf->mirrors[disk].head_position =
 337		r1_bio->sector + (r1_bio->sectors);
 338}
 339
 340/*
 341 * Find the disk number which triggered given bio
 342 */
 343static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 344{
 345	int mirror;
 346	struct r1conf *conf = r1_bio->mddev->private;
 347	int raid_disks = conf->raid_disks;
 348
 349	for (mirror = 0; mirror < raid_disks * 2; mirror++)
 350		if (r1_bio->bios[mirror] == bio)
 351			break;
 352
 353	BUG_ON(mirror == raid_disks * 2);
 354	update_head_pos(mirror, r1_bio);
 355
 356	return mirror;
 357}
 358
 359static void raid1_end_read_request(struct bio *bio)
 360{
 361	int uptodate = !bio->bi_status;
 362	struct r1bio *r1_bio = bio->bi_private;
 
 363	struct r1conf *conf = r1_bio->mddev->private;
 364	struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 365
 
 366	/*
 367	 * this branch is our 'one mirror IO has finished' event handler:
 368	 */
 369	update_head_pos(r1_bio->read_disk, r1_bio);
 370
 371	if (uptodate)
 372		set_bit(R1BIO_Uptodate, &r1_bio->state);
 373	else if (test_bit(FailFast, &rdev->flags) &&
 374		 test_bit(R1BIO_FailFast, &r1_bio->state))
 375		/* This was a fail-fast read so we definitely
 376		 * want to retry */
 377		;
 378	else {
 379		/* If all other devices have failed, we want to return
 380		 * the error upwards rather than fail the last device.
 381		 * Here we redefine "uptodate" to mean "Don't want to retry"
 382		 */
 383		unsigned long flags;
 384		spin_lock_irqsave(&conf->device_lock, flags);
 385		if (r1_bio->mddev->degraded == conf->raid_disks ||
 386		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 387		     test_bit(In_sync, &rdev->flags)))
 388			uptodate = 1;
 389		spin_unlock_irqrestore(&conf->device_lock, flags);
 390	}
 391
 392	if (uptodate) {
 393		raid_end_bio_io(r1_bio);
 394		rdev_dec_pending(rdev, conf->mddev);
 395	} else {
 396		/*
 397		 * oops, read error:
 398		 */
 399		char b[BDEVNAME_SIZE];
 400		pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 401				   mdname(conf->mddev),
 402				   bdevname(rdev->bdev, b),
 403				   (unsigned long long)r1_bio->sector);
 
 
 
 404		set_bit(R1BIO_ReadError, &r1_bio->state);
 405		reschedule_retry(r1_bio);
 406		/* don't drop the reference on read_disk yet */
 407	}
 408}
 409
 410static void close_write(struct r1bio *r1_bio)
 411{
 412	/* it really is the end of this request */
 413	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 414		bio_free_pages(r1_bio->behind_master_bio);
 415		bio_put(r1_bio->behind_master_bio);
 416		r1_bio->behind_master_bio = NULL;
 
 
 
 417	}
 418	/* clear the bitmap if all writes complete successfully */
 419	md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 420			   r1_bio->sectors,
 421			   !test_bit(R1BIO_Degraded, &r1_bio->state),
 422			   test_bit(R1BIO_BehindIO, &r1_bio->state));
 423	md_write_end(r1_bio->mddev);
 424}
 425
 426static void r1_bio_write_done(struct r1bio *r1_bio)
 427{
 428	if (!atomic_dec_and_test(&r1_bio->remaining))
 429		return;
 430
 431	if (test_bit(R1BIO_WriteError, &r1_bio->state))
 432		reschedule_retry(r1_bio);
 433	else {
 434		close_write(r1_bio);
 435		if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 436			reschedule_retry(r1_bio);
 437		else
 438			raid_end_bio_io(r1_bio);
 439	}
 440}
 441
 442static void raid1_end_write_request(struct bio *bio)
 443{
 444	struct r1bio *r1_bio = bio->bi_private;
 445	int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 446	struct r1conf *conf = r1_bio->mddev->private;
 447	struct bio *to_put = NULL;
 448	int mirror = find_bio_disk(r1_bio, bio);
 449	struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 450	bool discard_error;
 451	sector_t lo = r1_bio->sector;
 452	sector_t hi = r1_bio->sector + r1_bio->sectors;
 453
 454	discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 455
 456	/*
 457	 * 'one mirror IO has finished' event handler:
 458	 */
 459	if (bio->bi_status && !discard_error) {
 460		set_bit(WriteErrorSeen,	&rdev->flags);
 461		if (!test_and_set_bit(WantReplacement, &rdev->flags))
 
 
 462			set_bit(MD_RECOVERY_NEEDED, &
 463				conf->mddev->recovery);
 464
 465		if (test_bit(FailFast, &rdev->flags) &&
 466		    (bio->bi_opf & MD_FAILFAST) &&
 467		    /* We never try FailFast to WriteMostly devices */
 468		    !test_bit(WriteMostly, &rdev->flags)) {
 469			md_error(r1_bio->mddev, rdev);
 470		}
 471
 472		/*
 473		 * When the device is faulty, it is not necessary to
 474		 * handle write error.
 475		 * For failfast, this is the only remaining device,
 476		 * We need to retry the write without FailFast.
 477		 */
 478		if (!test_bit(Faulty, &rdev->flags))
 479			set_bit(R1BIO_WriteError, &r1_bio->state);
 480		else {
 481			/* Finished with this branch */
 482			r1_bio->bios[mirror] = NULL;
 483			to_put = bio;
 484		}
 485	} else {
 486		/*
 487		 * Set R1BIO_Uptodate in our master bio, so that we
 488		 * will return a good error code for to the higher
 489		 * levels even if IO on some other mirrored buffer
 490		 * fails.
 491		 *
 492		 * The 'master' represents the composite IO operation
 493		 * to user-side. So if something waits for IO, then it
 494		 * will wait for the 'master' bio.
 495		 */
 496		sector_t first_bad;
 497		int bad_sectors;
 498
 499		r1_bio->bios[mirror] = NULL;
 500		to_put = bio;
 501		/*
 502		 * Do not set R1BIO_Uptodate if the current device is
 503		 * rebuilding or Faulty. This is because we cannot use
 504		 * such device for properly reading the data back (we could
 505		 * potentially use it, if the current write would have felt
 506		 * before rdev->recovery_offset, but for simplicity we don't
 507		 * check this here.
 508		 */
 509		if (test_bit(In_sync, &rdev->flags) &&
 510		    !test_bit(Faulty, &rdev->flags))
 511			set_bit(R1BIO_Uptodate, &r1_bio->state);
 512
 513		/* Maybe we can clear some bad blocks. */
 514		if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 515				&first_bad, &bad_sectors) && !discard_error) {
 
 516			r1_bio->bios[mirror] = IO_MADE_GOOD;
 517			set_bit(R1BIO_MadeGood, &r1_bio->state);
 518		}
 519	}
 520
 521	if (behind) {
 522		if (test_bit(CollisionCheck, &rdev->flags))
 523			remove_serial(rdev, lo, hi);
 524		if (test_bit(WriteMostly, &rdev->flags))
 525			atomic_dec(&r1_bio->behind_remaining);
 526
 527		/*
 528		 * In behind mode, we ACK the master bio once the I/O
 529		 * has safely reached all non-writemostly
 530		 * disks. Setting the Returned bit ensures that this
 531		 * gets done only once -- we don't ever want to return
 532		 * -EIO here, instead we'll wait
 533		 */
 534		if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 535		    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 536			/* Maybe we can return now */
 537			if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 538				struct bio *mbio = r1_bio->master_bio;
 539				pr_debug("raid1: behind end write sectors"
 540					 " %llu-%llu\n",
 541					 (unsigned long long) mbio->bi_iter.bi_sector,
 542					 (unsigned long long) bio_end_sector(mbio) - 1);
 543				call_bio_endio(r1_bio);
 544			}
 545		}
 546	} else if (rdev->mddev->serialize_policy)
 547		remove_serial(rdev, lo, hi);
 548	if (r1_bio->bios[mirror] == NULL)
 549		rdev_dec_pending(rdev, conf->mddev);
 
 550
 551	/*
 552	 * Let's see if all mirrored write operations have finished
 553	 * already.
 554	 */
 555	r1_bio_write_done(r1_bio);
 556
 557	if (to_put)
 558		bio_put(to_put);
 559}
 560
 561static sector_t align_to_barrier_unit_end(sector_t start_sector,
 562					  sector_t sectors)
 563{
 564	sector_t len;
 565
 566	WARN_ON(sectors == 0);
 567	/*
 568	 * len is the number of sectors from start_sector to end of the
 569	 * barrier unit which start_sector belongs to.
 570	 */
 571	len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 572	      start_sector;
 573
 574	if (len > sectors)
 575		len = sectors;
 576
 577	return len;
 578}
 579
 580/*
 581 * This routine returns the disk from which the requested read should
 582 * be done. There is a per-array 'next expected sequential IO' sector
 583 * number - if this matches on the next IO then we use the last disk.
 584 * There is also a per-disk 'last know head position' sector that is
 585 * maintained from IRQ contexts, both the normal and the resync IO
 586 * completion handlers update this position correctly. If there is no
 587 * perfect sequential match then we pick the disk whose head is closest.
 588 *
 589 * If there are 2 mirrors in the same 2 devices, performance degrades
 590 * because position is mirror, not device based.
 591 *
 592 * The rdev for the device selected will have nr_pending incremented.
 593 */
 594static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 595{
 596	const sector_t this_sector = r1_bio->sector;
 597	int sectors;
 598	int best_good_sectors;
 599	int best_disk, best_dist_disk, best_pending_disk;
 600	int has_nonrot_disk;
 601	int disk;
 602	sector_t best_dist;
 603	unsigned int min_pending;
 604	struct md_rdev *rdev;
 605	int choose_first;
 606	int choose_next_idle;
 607
 608	rcu_read_lock();
 609	/*
 610	 * Check if we can balance. We can balance on the whole
 611	 * device if no resync is going on, or below the resync window.
 612	 * We take the first readable disk when above the resync window.
 613	 */
 614 retry:
 615	sectors = r1_bio->sectors;
 616	best_disk = -1;
 617	best_dist_disk = -1;
 618	best_dist = MaxSector;
 619	best_pending_disk = -1;
 620	min_pending = UINT_MAX;
 621	best_good_sectors = 0;
 622	has_nonrot_disk = 0;
 623	choose_next_idle = 0;
 624	clear_bit(R1BIO_FailFast, &r1_bio->state);
 625
 626	if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 627	    (mddev_is_clustered(conf->mddev) &&
 628	    md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 629		    this_sector + sectors)))
 630		choose_first = 1;
 631	else
 632		choose_first = 0;
 633
 634	for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 635		sector_t dist;
 636		sector_t first_bad;
 637		int bad_sectors;
 638		unsigned int pending;
 639		bool nonrot;
 640
 641		rdev = rcu_dereference(conf->mirrors[disk].rdev);
 642		if (r1_bio->bios[disk] == IO_BLOCKED
 643		    || rdev == NULL
 644		    || test_bit(Faulty, &rdev->flags))
 645			continue;
 646		if (!test_bit(In_sync, &rdev->flags) &&
 647		    rdev->recovery_offset < this_sector + sectors)
 648			continue;
 649		if (test_bit(WriteMostly, &rdev->flags)) {
 650			/* Don't balance among write-mostly, just
 651			 * use the first as a last resort */
 652			if (best_dist_disk < 0) {
 653				if (is_badblock(rdev, this_sector, sectors,
 654						&first_bad, &bad_sectors)) {
 655					if (first_bad <= this_sector)
 656						/* Cannot use this */
 657						continue;
 658					best_good_sectors = first_bad - this_sector;
 659				} else
 660					best_good_sectors = sectors;
 661				best_dist_disk = disk;
 662				best_pending_disk = disk;
 663			}
 664			continue;
 665		}
 666		/* This is a reasonable device to use.  It might
 667		 * even be best.
 668		 */
 669		if (is_badblock(rdev, this_sector, sectors,
 670				&first_bad, &bad_sectors)) {
 671			if (best_dist < MaxSector)
 672				/* already have a better device */
 673				continue;
 674			if (first_bad <= this_sector) {
 675				/* cannot read here. If this is the 'primary'
 676				 * device, then we must not read beyond
 677				 * bad_sectors from another device..
 678				 */
 679				bad_sectors -= (this_sector - first_bad);
 680				if (choose_first && sectors > bad_sectors)
 681					sectors = bad_sectors;
 682				if (best_good_sectors > sectors)
 683					best_good_sectors = sectors;
 684
 685			} else {
 686				sector_t good_sectors = first_bad - this_sector;
 687				if (good_sectors > best_good_sectors) {
 688					best_good_sectors = good_sectors;
 689					best_disk = disk;
 690				}
 691				if (choose_first)
 692					break;
 693			}
 694			continue;
 695		} else {
 696			if ((sectors > best_good_sectors) && (best_disk >= 0))
 697				best_disk = -1;
 698			best_good_sectors = sectors;
 699		}
 700
 701		if (best_disk >= 0)
 702			/* At least two disks to choose from so failfast is OK */
 703			set_bit(R1BIO_FailFast, &r1_bio->state);
 704
 705		nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 706		has_nonrot_disk |= nonrot;
 707		pending = atomic_read(&rdev->nr_pending);
 708		dist = abs(this_sector - conf->mirrors[disk].head_position);
 709		if (choose_first) {
 710			best_disk = disk;
 711			break;
 712		}
 713		/* Don't change to another disk for sequential reads */
 714		if (conf->mirrors[disk].next_seq_sect == this_sector
 715		    || dist == 0) {
 716			int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 717			struct raid1_info *mirror = &conf->mirrors[disk];
 718
 719			best_disk = disk;
 720			/*
 721			 * If buffered sequential IO size exceeds optimal
 722			 * iosize, check if there is idle disk. If yes, choose
 723			 * the idle disk. read_balance could already choose an
 724			 * idle disk before noticing it's a sequential IO in
 725			 * this disk. This doesn't matter because this disk
 726			 * will idle, next time it will be utilized after the
 727			 * first disk has IO size exceeds optimal iosize. In
 728			 * this way, iosize of the first disk will be optimal
 729			 * iosize at least. iosize of the second disk might be
 730			 * small, but not a big deal since when the second disk
 731			 * starts IO, the first disk is likely still busy.
 732			 */
 733			if (nonrot && opt_iosize > 0 &&
 734			    mirror->seq_start != MaxSector &&
 735			    mirror->next_seq_sect > opt_iosize &&
 736			    mirror->next_seq_sect - opt_iosize >=
 737			    mirror->seq_start) {
 738				choose_next_idle = 1;
 739				continue;
 740			}
 741			break;
 742		}
 
 
 
 
 
 743
 744		if (choose_next_idle)
 745			continue;
 746
 747		if (min_pending > pending) {
 748			min_pending = pending;
 749			best_pending_disk = disk;
 750		}
 751
 752		if (dist < best_dist) {
 753			best_dist = dist;
 754			best_dist_disk = disk;
 755		}
 756	}
 757
 758	/*
 759	 * If all disks are rotational, choose the closest disk. If any disk is
 760	 * non-rotational, choose the disk with less pending request even the
 761	 * disk is rotational, which might/might not be optimal for raids with
 762	 * mixed ratation/non-rotational disks depending on workload.
 763	 */
 764	if (best_disk == -1) {
 765		if (has_nonrot_disk || min_pending == 0)
 766			best_disk = best_pending_disk;
 767		else
 768			best_disk = best_dist_disk;
 769	}
 770
 771	if (best_disk >= 0) {
 772		rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 773		if (!rdev)
 774			goto retry;
 775		atomic_inc(&rdev->nr_pending);
 
 
 
 
 
 
 
 776		sectors = best_good_sectors;
 777
 778		if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 779			conf->mirrors[best_disk].seq_start = this_sector;
 780
 781		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 782	}
 783	rcu_read_unlock();
 784	*max_sectors = sectors;
 785
 786	return best_disk;
 787}
 788
 789static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 790{
 791	/* flush any pending bitmap writes to disk before proceeding w/ I/O */
 792	md_bitmap_unplug(conf->mddev->bitmap);
 793	wake_up(&conf->wait_barrier);
 
 
 
 
 
 
 
 
 
 
 
 794
 795	while (bio) { /* submit pending writes */
 796		struct bio *next = bio->bi_next;
 797		struct md_rdev *rdev = (void *)bio->bi_disk;
 798		bio->bi_next = NULL;
 799		bio_set_dev(bio, rdev->bdev);
 800		if (test_bit(Faulty, &rdev->flags)) {
 801			bio_io_error(bio);
 802		} else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 803				    !blk_queue_discard(bio->bi_disk->queue)))
 804			/* Just ignore it */
 805			bio_endio(bio);
 806		else
 807			submit_bio_noacct(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		conf->pending_count = 0;
 826		spin_unlock_irq(&conf->device_lock);
 
 
 
 
 827
 828		/*
 829		 * As this is called in a wait_event() loop (see freeze_array),
 830		 * current->state might be TASK_UNINTERRUPTIBLE which will
 831		 * cause a warning when we prepare to wait again.  As it is
 832		 * rare that this path is taken, it is perfectly safe to force
 833		 * us to go around the wait_event() loop again, so the warning
 834		 * is a false-positive.  Silence the warning by resetting
 835		 * thread state
 836		 */
 837		__set_current_state(TASK_RUNNING);
 838		blk_start_plug(&plug);
 839		flush_bio_list(conf, bio);
 840		blk_finish_plug(&plug);
 841	} else
 842		spin_unlock_irq(&conf->device_lock);
 843}
 844
 845/* Barriers....
 846 * Sometimes we need to suspend IO while we do something else,
 847 * either some resync/recovery, or reconfigure the array.
 848 * To do this we raise a 'barrier'.
 849 * The 'barrier' is a counter that can be raised multiple times
 850 * to count how many activities are happening which preclude
 851 * normal IO.
 852 * We can only raise the barrier if there is no pending IO.
 853 * i.e. if nr_pending == 0.
 854 * We choose only to raise the barrier if no-one is waiting for the
 855 * barrier to go down.  This means that as soon as an IO request
 856 * is ready, no other operations which require a barrier will start
 857 * until the IO request has had a chance.
 858 *
 859 * So: regular IO calls 'wait_barrier'.  When that returns there
 860 *    is no backgroup IO happening,  It must arrange to call
 861 *    allow_barrier when it has finished its IO.
 862 * backgroup IO calls must call raise_barrier.  Once that returns
 863 *    there is no normal IO happeing.  It must arrange to call
 864 *    lower_barrier when the particular background IO completes.
 865 *
 866 * If resync/recovery is interrupted, returns -EINTR;
 867 * Otherwise, returns 0.
 868 */
 869static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 870{
 871	int idx = sector_to_idx(sector_nr);
 872
 873	spin_lock_irq(&conf->resync_lock);
 874
 875	/* Wait until no block IO is waiting */
 876	wait_event_lock_irq(conf->wait_barrier,
 877			    !atomic_read(&conf->nr_waiting[idx]),
 878			    conf->resync_lock);
 879
 880	/* block any new IO from starting */
 881	atomic_inc(&conf->barrier[idx]);
 882	/*
 883	 * In raise_barrier() we firstly increase conf->barrier[idx] then
 884	 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 885	 * increase conf->nr_pending[idx] then check conf->barrier[idx].
 886	 * A memory barrier here to make sure conf->nr_pending[idx] won't
 887	 * be fetched before conf->barrier[idx] is increased. Otherwise
 888	 * there will be a race between raise_barrier() and _wait_barrier().
 889	 */
 890	smp_mb__after_atomic();
 891
 892	/* For these conditions we must wait:
 893	 * A: while the array is in frozen state
 894	 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 895	 *    existing in corresponding I/O barrier bucket.
 896	 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 897	 *    max resync count which allowed on current I/O barrier bucket.
 
 
 898	 */
 899	wait_event_lock_irq(conf->wait_barrier,
 900			    (!conf->array_frozen &&
 901			     !atomic_read(&conf->nr_pending[idx]) &&
 902			     atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 903				test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 
 904			    conf->resync_lock);
 905
 906	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 907		atomic_dec(&conf->barrier[idx]);
 908		spin_unlock_irq(&conf->resync_lock);
 909		wake_up(&conf->wait_barrier);
 910		return -EINTR;
 911	}
 912
 913	atomic_inc(&conf->nr_sync_pending);
 914	spin_unlock_irq(&conf->resync_lock);
 915
 916	return 0;
 917}
 918
 919static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 920{
 921	int idx = sector_to_idx(sector_nr);
 922
 923	BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 924
 925	atomic_dec(&conf->barrier[idx]);
 926	atomic_dec(&conf->nr_sync_pending);
 927	wake_up(&conf->wait_barrier);
 928}
 929
 930static void _wait_barrier(struct r1conf *conf, int idx)
 931{
 932	/*
 933	 * We need to increase conf->nr_pending[idx] very early here,
 934	 * then raise_barrier() can be blocked when it waits for
 935	 * conf->nr_pending[idx] to be 0. Then we can avoid holding
 936	 * conf->resync_lock when there is no barrier raised in same
 937	 * barrier unit bucket. Also if the array is frozen, I/O
 938	 * should be blocked until array is unfrozen.
 939	 */
 940	atomic_inc(&conf->nr_pending[idx]);
 941	/*
 942	 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 943	 * check conf->barrier[idx]. In raise_barrier() we firstly increase
 944	 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 945	 * barrier is necessary here to make sure conf->barrier[idx] won't be
 946	 * fetched before conf->nr_pending[idx] is increased. Otherwise there
 947	 * will be a race between _wait_barrier() and raise_barrier().
 948	 */
 949	smp_mb__after_atomic();
 950
 951	/*
 952	 * Don't worry about checking two atomic_t variables at same time
 953	 * here. If during we check conf->barrier[idx], the array is
 954	 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 955	 * 0, it is safe to return and make the I/O continue. Because the
 956	 * array is frozen, all I/O returned here will eventually complete
 957	 * or be queued, no race will happen. See code comment in
 958	 * frozen_array().
 959	 */
 960	if (!READ_ONCE(conf->array_frozen) &&
 961	    !atomic_read(&conf->barrier[idx]))
 962		return;
 963
 964	/*
 965	 * After holding conf->resync_lock, conf->nr_pending[idx]
 966	 * should be decreased before waiting for barrier to drop.
 967	 * Otherwise, we may encounter a race condition because
 968	 * raise_barrer() might be waiting for conf->nr_pending[idx]
 969	 * to be 0 at same time.
 970	 */
 971	spin_lock_irq(&conf->resync_lock);
 972	atomic_inc(&conf->nr_waiting[idx]);
 973	atomic_dec(&conf->nr_pending[idx]);
 974	/*
 975	 * In case freeze_array() is waiting for
 976	 * get_unqueued_pending() == extra
 977	 */
 978	wake_up(&conf->wait_barrier);
 979	/* Wait for the barrier in same barrier unit bucket to drop. */
 980	wait_event_lock_irq(conf->wait_barrier,
 981			    !conf->array_frozen &&
 982			     !atomic_read(&conf->barrier[idx]),
 983			    conf->resync_lock);
 984	atomic_inc(&conf->nr_pending[idx]);
 985	atomic_dec(&conf->nr_waiting[idx]);
 986	spin_unlock_irq(&conf->resync_lock);
 987}
 988
 989static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 990{
 991	int idx = sector_to_idx(sector_nr);
 992
 993	/*
 994	 * Very similar to _wait_barrier(). The difference is, for read
 995	 * I/O we don't need wait for sync I/O, but if the whole array
 996	 * is frozen, the read I/O still has to wait until the array is
 997	 * unfrozen. Since there is no ordering requirement with
 998	 * conf->barrier[idx] here, memory barrier is unnecessary as well.
 999	 */
1000	atomic_inc(&conf->nr_pending[idx]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1001
1002	if (!READ_ONCE(conf->array_frozen))
1003		return;
1004
1005	spin_lock_irq(&conf->resync_lock);
1006	atomic_inc(&conf->nr_waiting[idx]);
1007	atomic_dec(&conf->nr_pending[idx]);
1008	/*
1009	 * In case freeze_array() is waiting for
1010	 * get_unqueued_pending() == extra
1011	 */
1012	wake_up(&conf->wait_barrier);
1013	/* Wait for array to be unfrozen */
1014	wait_event_lock_irq(conf->wait_barrier,
1015			    !conf->array_frozen,
1016			    conf->resync_lock);
1017	atomic_inc(&conf->nr_pending[idx]);
1018	atomic_dec(&conf->nr_waiting[idx]);
1019	spin_unlock_irq(&conf->resync_lock);
 
1020}
1021
1022static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
 
1023{
1024	int idx = sector_to_idx(sector_nr);
1025
1026	_wait_barrier(conf, idx);
1027}
 
 
 
 
 
 
 
 
 
1028
1029static void _allow_barrier(struct r1conf *conf, int idx)
1030{
1031	atomic_dec(&conf->nr_pending[idx]);
 
 
 
 
 
 
 
 
 
1032	wake_up(&conf->wait_barrier);
1033}
1034
1035static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1036{
1037	int idx = sector_to_idx(sector_nr);
1038
1039	_allow_barrier(conf, idx);
1040}
1041
1042/* conf->resync_lock should be held */
1043static int get_unqueued_pending(struct r1conf *conf)
1044{
1045	int idx, ret;
1046
1047	ret = atomic_read(&conf->nr_sync_pending);
1048	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1049		ret += atomic_read(&conf->nr_pending[idx]) -
1050			atomic_read(&conf->nr_queued[idx]);
1051
1052	return ret;
1053}
1054
1055static void freeze_array(struct r1conf *conf, int extra)
1056{
1057	/* Stop sync I/O and normal I/O and wait for everything to
1058	 * go quiet.
1059	 * This is called in two situations:
1060	 * 1) management command handlers (reshape, remove disk, quiesce).
1061	 * 2) one normal I/O request failed.
1062
1063	 * After array_frozen is set to 1, new sync IO will be blocked at
1064	 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1065	 * or wait_read_barrier(). The flying I/Os will either complete or be
1066	 * queued. When everything goes quite, there are only queued I/Os left.
1067
1068	 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1069	 * barrier bucket index which this I/O request hits. When all sync and
1070	 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1071	 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1072	 * in handle_read_error(), we may call freeze_array() before trying to
1073	 * fix the read error. In this case, the error read I/O is not queued,
1074	 * so get_unqueued_pending() == 1.
1075	 *
1076	 * Therefore before this function returns, we need to wait until
1077	 * get_unqueued_pendings(conf) gets equal to extra. For
1078	 * normal I/O context, extra is 1, in rested situations extra is 0.
1079	 */
1080	spin_lock_irq(&conf->resync_lock);
1081	conf->array_frozen = 1;
1082	raid1_log(conf->mddev, "wait freeze");
1083	wait_event_lock_irq_cmd(
1084		conf->wait_barrier,
1085		get_unqueued_pending(conf) == extra,
1086		conf->resync_lock,
1087		flush_pending_writes(conf));
1088	spin_unlock_irq(&conf->resync_lock);
1089}
1090static void unfreeze_array(struct r1conf *conf)
1091{
1092	/* reverse the effect of the freeze */
1093	spin_lock_irq(&conf->resync_lock);
1094	conf->array_frozen = 0;
 
1095	spin_unlock_irq(&conf->resync_lock);
1096	wake_up(&conf->wait_barrier);
1097}
1098
1099static void alloc_behind_master_bio(struct r1bio *r1_bio,
1100					   struct bio *bio)
 
1101{
1102	int size = bio->bi_iter.bi_size;
1103	unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1104	int i = 0;
1105	struct bio *behind_bio = NULL;
1106
1107	behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1108	if (!behind_bio)
1109		return;
1110
1111	/* discard op, we don't support writezero/writesame yet */
1112	if (!bio_has_data(bio)) {
1113		behind_bio->bi_iter.bi_size = size;
1114		goto skip_copy;
 
 
 
 
 
1115	}
1116
1117	behind_bio->bi_write_hint = bio->bi_write_hint;
1118
1119	while (i < vcnt && size) {
1120		struct page *page;
1121		int len = min_t(int, PAGE_SIZE, size);
1122
1123		page = alloc_page(GFP_NOIO);
1124		if (unlikely(!page))
1125			goto free_pages;
1126
1127		bio_add_page(behind_bio, page, len, 0);
1128
1129		size -= len;
1130		i++;
1131	}
1132
1133	bio_copy_data(behind_bio, bio);
1134skip_copy:
1135	r1_bio->behind_master_bio = behind_bio;
1136	set_bit(R1BIO_BehindIO, &r1_bio->state);
1137
1138	return;
1139
1140free_pages:
 
 
 
 
1141	pr_debug("%dB behind alloc failed, doing sync I/O\n",
1142		 bio->bi_iter.bi_size);
1143	bio_free_pages(behind_bio);
1144	bio_put(behind_bio);
1145}
1146
1147struct raid1_plug_cb {
1148	struct blk_plug_cb	cb;
1149	struct bio_list		pending;
1150	int			pending_cnt;
1151};
1152
1153static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1154{
1155	struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1156						  cb);
1157	struct mddev *mddev = plug->cb.data;
1158	struct r1conf *conf = mddev->private;
1159	struct bio *bio;
1160
1161	if (from_schedule || current->bio_list) {
1162		spin_lock_irq(&conf->device_lock);
1163		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1164		conf->pending_count += plug->pending_cnt;
1165		spin_unlock_irq(&conf->device_lock);
1166		wake_up(&conf->wait_barrier);
1167		md_wakeup_thread(mddev->thread);
1168		kfree(plug);
1169		return;
1170	}
1171
1172	/* we aren't scheduling, so we can do the write-out directly. */
1173	bio = bio_list_get(&plug->pending);
1174	flush_bio_list(conf, bio);
 
 
 
 
 
 
 
 
 
 
 
 
 
1175	kfree(plug);
1176}
1177
1178static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1179{
1180	r1_bio->master_bio = bio;
1181	r1_bio->sectors = bio_sectors(bio);
1182	r1_bio->state = 0;
1183	r1_bio->mddev = mddev;
1184	r1_bio->sector = bio->bi_iter.bi_sector;
1185}
1186
1187static inline struct r1bio *
1188alloc_r1bio(struct mddev *mddev, struct bio *bio)
1189{
1190	struct r1conf *conf = mddev->private;
 
1191	struct r1bio *r1_bio;
1192
1193	r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1194	/* Ensure no bio records IO_BLOCKED */
1195	memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1196	init_r1bio(r1_bio, mddev, bio);
1197	return r1_bio;
1198}
1199
1200static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1201			       int max_read_sectors, struct r1bio *r1_bio)
1202{
1203	struct r1conf *conf = mddev->private;
1204	struct raid1_info *mirror;
1205	struct bio *read_bio;
1206	struct bitmap *bitmap = mddev->bitmap;
1207	const int op = bio_op(bio);
1208	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
 
 
 
 
 
 
 
 
 
 
 
1209	int max_sectors;
1210	int rdisk;
1211	bool print_msg = !!r1_bio;
1212	char b[BDEVNAME_SIZE];
1213
1214	/*
1215	 * If r1_bio is set, we are blocking the raid1d thread
1216	 * so there is a tiny risk of deadlock.  So ask for
1217	 * emergency memory if needed.
1218	 */
1219	gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1220
1221	if (print_msg) {
1222		/* Need to get the block device name carefully */
1223		struct md_rdev *rdev;
1224		rcu_read_lock();
1225		rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1226		if (rdev)
1227			bdevname(rdev->bdev, b);
1228		else
1229			strcpy(b, "???");
1230		rcu_read_unlock();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1231	}
1232
1233	/*
1234	 * Still need barrier for READ in case that whole
1235	 * array is frozen.
1236	 */
1237	wait_read_barrier(conf, bio->bi_iter.bi_sector);
1238
1239	if (!r1_bio)
1240		r1_bio = alloc_r1bio(mddev, bio);
1241	else
1242		init_r1bio(r1_bio, mddev, bio);
1243	r1_bio->sectors = max_read_sectors;
1244
1245	/*
1246	 * make_request() can abort the operation when read-ahead is being
1247	 * used and no empty request is available.
 
1248	 */
1249	rdisk = read_balance(conf, r1_bio, &max_sectors);
1250
1251	if (rdisk < 0) {
1252		/* couldn't find anywhere to read from */
1253		if (print_msg) {
1254			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1255					    mdname(mddev),
1256					    b,
1257					    (unsigned long long)r1_bio->sector);
1258		}
1259		raid_end_bio_io(r1_bio);
1260		return;
1261	}
1262	mirror = conf->mirrors + rdisk;
1263
1264	if (print_msg)
1265		pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1266				    mdname(mddev),
1267				    (unsigned long long)r1_bio->sector,
1268				    bdevname(mirror->rdev->bdev, b));
 
 
 
 
1269
1270	if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1271	    bitmap) {
1272		/*
1273		 * Reading from a write-mostly device must take care not to
1274		 * over-take any writes that are 'behind'
1275		 */
1276		raid1_log(mddev, "wait behind writes");
1277		wait_event(bitmap->behind_wait,
1278			   atomic_read(&bitmap->behind_writes) == 0);
1279	}
1280
1281	if (max_sectors < bio_sectors(bio)) {
1282		struct bio *split = bio_split(bio, max_sectors,
1283					      gfp, &conf->bio_split);
1284		bio_chain(split, bio);
1285		submit_bio_noacct(bio);
1286		bio = split;
1287		r1_bio->master_bio = bio;
1288		r1_bio->sectors = max_sectors;
1289	}
1290
1291	r1_bio->read_disk = rdisk;
 
 
 
 
 
1292
1293	read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
 
 
 
 
 
 
 
 
 
 
1294
1295	r1_bio->bios[rdisk] = read_bio;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1296
1297	read_bio->bi_iter.bi_sector = r1_bio->sector +
1298		mirror->rdev->data_offset;
1299	bio_set_dev(read_bio, mirror->rdev->bdev);
1300	read_bio->bi_end_io = raid1_end_read_request;
1301	bio_set_op_attrs(read_bio, op, do_sync);
1302	if (test_bit(FailFast, &mirror->rdev->flags) &&
1303	    test_bit(R1BIO_FailFast, &r1_bio->state))
1304	        read_bio->bi_opf |= MD_FAILFAST;
1305	read_bio->bi_private = r1_bio;
 
 
 
 
 
 
1306
1307	if (mddev->gendisk)
1308	        trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1309				disk_devt(mddev->gendisk), r1_bio->sector);
1310
1311	submit_bio_noacct(read_bio);
1312}
1313
1314static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1315				int max_write_sectors)
1316{
1317	struct r1conf *conf = mddev->private;
1318	struct r1bio *r1_bio;
1319	int i, disks;
1320	struct bitmap *bitmap = mddev->bitmap;
1321	unsigned long flags;
1322	struct md_rdev *blocked_rdev;
1323	struct blk_plug_cb *cb;
1324	struct raid1_plug_cb *plug = NULL;
1325	int first_clone;
1326	int max_sectors;
1327
1328	if (mddev_is_clustered(mddev) &&
1329	     md_cluster_ops->area_resyncing(mddev, WRITE,
1330		     bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1331
1332		DEFINE_WAIT(w);
1333		for (;;) {
1334			prepare_to_wait(&conf->wait_barrier,
1335					&w, TASK_IDLE);
1336			if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1337							bio->bi_iter.bi_sector,
1338							bio_end_sector(bio)))
1339				break;
1340			schedule();
1341		}
1342		finish_wait(&conf->wait_barrier, &w);
1343	}
1344
1345	/*
1346	 * Register the new request and wait if the reconstruction
1347	 * thread has put up a bar for new requests.
1348	 * Continue immediately if no resync is active currently.
1349	 */
1350	wait_barrier(conf, bio->bi_iter.bi_sector);
1351
1352	r1_bio = alloc_r1bio(mddev, bio);
1353	r1_bio->sectors = max_write_sectors;
1354
1355	if (conf->pending_count >= max_queued_requests) {
1356		md_wakeup_thread(mddev->thread);
1357		raid1_log(mddev, "wait queued");
1358		wait_event(conf->wait_barrier,
1359			   conf->pending_count < max_queued_requests);
1360	}
1361	/* first select target devices under rcu_lock and
1362	 * inc refcount on their rdev.  Record them by setting
1363	 * bios[x] to bio
1364	 * If there are known/acknowledged bad blocks on any device on
1365	 * which we have seen a write error, we want to avoid writing those
1366	 * blocks.
1367	 * This potentially requires several writes to write around
1368	 * the bad blocks.  Each set of writes gets it's own r1bio
1369	 * with a set of bios attached.
1370	 */
1371
1372	disks = conf->raid_disks * 2;
1373 retry_write:
 
1374	blocked_rdev = NULL;
1375	rcu_read_lock();
1376	max_sectors = r1_bio->sectors;
1377	for (i = 0;  i < disks; i++) {
1378		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1379		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1380			atomic_inc(&rdev->nr_pending);
1381			blocked_rdev = rdev;
1382			break;
1383		}
1384		r1_bio->bios[i] = NULL;
1385		if (!rdev || test_bit(Faulty, &rdev->flags)) {
1386			if (i < conf->raid_disks)
1387				set_bit(R1BIO_Degraded, &r1_bio->state);
1388			continue;
1389		}
1390
1391		atomic_inc(&rdev->nr_pending);
1392		if (test_bit(WriteErrorSeen, &rdev->flags)) {
1393			sector_t first_bad;
1394			int bad_sectors;
1395			int is_bad;
1396
1397			is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
 
1398					     &first_bad, &bad_sectors);
1399			if (is_bad < 0) {
1400				/* mustn't write here until the bad block is
1401				 * acknowledged*/
1402				set_bit(BlockedBadBlocks, &rdev->flags);
1403				blocked_rdev = rdev;
1404				break;
1405			}
1406			if (is_bad && first_bad <= r1_bio->sector) {
1407				/* Cannot write here at all */
1408				bad_sectors -= (r1_bio->sector - first_bad);
1409				if (bad_sectors < max_sectors)
1410					/* mustn't write more than bad_sectors
1411					 * to other devices yet
1412					 */
1413					max_sectors = bad_sectors;
1414				rdev_dec_pending(rdev, mddev);
1415				/* We don't set R1BIO_Degraded as that
1416				 * only applies if the disk is
1417				 * missing, so it might be re-added,
1418				 * and we want to know to recover this
1419				 * chunk.
1420				 * In this case the device is here,
1421				 * and the fact that this chunk is not
1422				 * in-sync is recorded in the bad
1423				 * block log
1424				 */
1425				continue;
1426			}
1427			if (is_bad) {
1428				int good_sectors = first_bad - r1_bio->sector;
1429				if (good_sectors < max_sectors)
1430					max_sectors = good_sectors;
1431			}
1432		}
1433		r1_bio->bios[i] = bio;
1434	}
1435	rcu_read_unlock();
1436
1437	if (unlikely(blocked_rdev)) {
1438		/* Wait for this device to become unblocked */
1439		int j;
 
1440
1441		for (j = 0; j < i; j++)
1442			if (r1_bio->bios[j])
1443				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1444		r1_bio->state = 0;
1445		allow_barrier(conf, bio->bi_iter.bi_sector);
1446		raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1447		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1448		wait_barrier(conf, bio->bi_iter.bi_sector);
 
 
 
 
 
 
 
 
 
1449		goto retry_write;
1450	}
1451
1452	if (max_sectors < bio_sectors(bio)) {
1453		struct bio *split = bio_split(bio, max_sectors,
1454					      GFP_NOIO, &conf->bio_split);
1455		bio_chain(split, bio);
1456		submit_bio_noacct(bio);
1457		bio = split;
1458		r1_bio->master_bio = bio;
1459		r1_bio->sectors = max_sectors;
 
 
 
 
 
 
1460	}
 
1461
1462	atomic_set(&r1_bio->remaining, 1);
1463	atomic_set(&r1_bio->behind_remaining, 0);
1464
1465	first_clone = 1;
1466
1467	for (i = 0; i < disks; i++) {
1468		struct bio *mbio = NULL;
1469		struct md_rdev *rdev = conf->mirrors[i].rdev;
1470		if (!r1_bio->bios[i])
1471			continue;
1472
 
 
 
1473		if (first_clone) {
1474			/* do behind I/O ?
1475			 * Not if there are too many, or cannot
1476			 * allocate memory, or a reader on WriteMostly
1477			 * is waiting for behind writes to flush */
1478			if (bitmap &&
1479			    (atomic_read(&bitmap->behind_writes)
1480			     < mddev->bitmap_info.max_write_behind) &&
1481			    !waitqueue_active(&bitmap->behind_wait)) {
1482				alloc_behind_master_bio(r1_bio, bio);
1483			}
1484
1485			md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1486					     test_bit(R1BIO_BehindIO, &r1_bio->state));
 
 
1487			first_clone = 0;
1488		}
 
 
 
1489
1490		if (r1_bio->behind_master_bio)
1491			mbio = bio_clone_fast(r1_bio->behind_master_bio,
1492					      GFP_NOIO, &mddev->bio_set);
1493		else
1494			mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1495
1496		if (r1_bio->behind_master_bio) {
1497			if (test_bit(CollisionCheck, &rdev->flags))
1498				wait_for_serialization(rdev, r1_bio);
1499			if (test_bit(WriteMostly, &rdev->flags))
1500				atomic_inc(&r1_bio->behind_remaining);
1501		} else if (mddev->serialize_policy)
1502			wait_for_serialization(rdev, r1_bio);
1503
1504		r1_bio->bios[i] = mbio;
1505
1506		mbio->bi_iter.bi_sector	= (r1_bio->sector +
1507				   conf->mirrors[i].rdev->data_offset);
1508		bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1509		mbio->bi_end_io	= raid1_end_write_request;
1510		mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1511		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1512		    !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1513		    conf->raid_disks - mddev->degraded > 1)
1514			mbio->bi_opf |= MD_FAILFAST;
1515		mbio->bi_private = r1_bio;
1516
1517		atomic_inc(&r1_bio->remaining);
1518
1519		if (mddev->gendisk)
1520			trace_block_bio_remap(mbio->bi_disk->queue,
1521					      mbio, disk_devt(mddev->gendisk),
1522					      r1_bio->sector);
1523		/* flush_pending_writes() needs access to the rdev so...*/
1524		mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1525
1526		cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1527		if (cb)
1528			plug = container_of(cb, struct raid1_plug_cb, cb);
1529		else
1530			plug = NULL;
 
1531		if (plug) {
1532			bio_list_add(&plug->pending, mbio);
1533			plug->pending_cnt++;
1534		} else {
1535			spin_lock_irqsave(&conf->device_lock, flags);
1536			bio_list_add(&conf->pending_bio_list, mbio);
1537			conf->pending_count++;
1538			spin_unlock_irqrestore(&conf->device_lock, flags);
 
 
1539			md_wakeup_thread(mddev->thread);
1540		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1541	}
1542
1543	r1_bio_write_done(r1_bio);
1544
1545	/* In case raid1d snuck in to freeze_array */
1546	wake_up(&conf->wait_barrier);
1547}
1548
1549static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1550{
1551	sector_t sectors;
1552
1553	if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1554	    && md_flush_request(mddev, bio))
1555		return true;
1556
1557	/*
1558	 * There is a limit to the maximum size, but
1559	 * the read/write handler might find a lower limit
1560	 * due to bad blocks.  To avoid multiple splits,
1561	 * we pass the maximum number of sectors down
1562	 * and let the lower level perform the split.
1563	 */
1564	sectors = align_to_barrier_unit_end(
1565		bio->bi_iter.bi_sector, bio_sectors(bio));
1566
1567	if (bio_data_dir(bio) == READ)
1568		raid1_read_request(mddev, bio, sectors, NULL);
1569	else {
1570		if (!md_write_start(mddev,bio))
1571			return false;
1572		raid1_write_request(mddev, bio, sectors);
1573	}
1574	return true;
1575}
1576
1577static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1578{
1579	struct r1conf *conf = mddev->private;
1580	int i;
1581
1582	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1583		   conf->raid_disks - mddev->degraded);
1584	rcu_read_lock();
1585	for (i = 0; i < conf->raid_disks; i++) {
1586		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1587		seq_printf(seq, "%s",
1588			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1589	}
1590	rcu_read_unlock();
1591	seq_printf(seq, "]");
1592}
1593
1594static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1595{
1596	char b[BDEVNAME_SIZE];
1597	struct r1conf *conf = mddev->private;
1598	unsigned long flags;
1599
1600	/*
1601	 * If it is not operational, then we have already marked it as dead
1602	 * else if it is the last working disks with "fail_last_dev == false",
1603	 * ignore the error, let the next level up know.
1604	 * else mark the drive as failed
1605	 */
1606	spin_lock_irqsave(&conf->device_lock, flags);
1607	if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1608	    && (conf->raid_disks - mddev->degraded) == 1) {
1609		/*
1610		 * Don't fail the drive, act as though we were just a
1611		 * normal single drive.
1612		 * However don't try a recovery from this drive as
1613		 * it is very likely to fail.
1614		 */
1615		conf->recovery_disabled = mddev->recovery_disabled;
1616		spin_unlock_irqrestore(&conf->device_lock, flags);
1617		return;
1618	}
1619	set_bit(Blocked, &rdev->flags);
1620	if (test_and_clear_bit(In_sync, &rdev->flags))
 
1621		mddev->degraded++;
1622	set_bit(Faulty, &rdev->flags);
 
 
1623	spin_unlock_irqrestore(&conf->device_lock, flags);
1624	/*
1625	 * if recovery is running, make sure it aborts.
1626	 */
1627	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1628	set_mask_bits(&mddev->sb_flags, 0,
1629		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1630	pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1631		"md/raid1:%s: Operation continuing on %d devices.\n",
1632		mdname(mddev), bdevname(rdev->bdev, b),
1633		mdname(mddev), conf->raid_disks - mddev->degraded);
 
1634}
1635
1636static void print_conf(struct r1conf *conf)
1637{
1638	int i;
1639
1640	pr_debug("RAID1 conf printout:\n");
1641	if (!conf) {
1642		pr_debug("(!conf)\n");
1643		return;
1644	}
1645	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1646		 conf->raid_disks);
1647
1648	rcu_read_lock();
1649	for (i = 0; i < conf->raid_disks; i++) {
1650		char b[BDEVNAME_SIZE];
1651		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1652		if (rdev)
1653			pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1654				 i, !test_bit(In_sync, &rdev->flags),
1655				 !test_bit(Faulty, &rdev->flags),
1656				 bdevname(rdev->bdev,b));
1657	}
1658	rcu_read_unlock();
1659}
1660
1661static void close_sync(struct r1conf *conf)
1662{
1663	int idx;
 
1664
1665	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1666		_wait_barrier(conf, idx);
1667		_allow_barrier(conf, idx);
1668	}
1669
1670	mempool_exit(&conf->r1buf_pool);
 
 
 
 
 
 
1671}
1672
1673static int raid1_spare_active(struct mddev *mddev)
1674{
1675	int i;
1676	struct r1conf *conf = mddev->private;
1677	int count = 0;
1678	unsigned long flags;
1679
1680	/*
1681	 * Find all failed disks within the RAID1 configuration
1682	 * and mark them readable.
1683	 * Called under mddev lock, so rcu protection not needed.
1684	 * device_lock used to avoid races with raid1_end_read_request
1685	 * which expects 'In_sync' flags and ->degraded to be consistent.
1686	 */
1687	spin_lock_irqsave(&conf->device_lock, flags);
1688	for (i = 0; i < conf->raid_disks; i++) {
1689		struct md_rdev *rdev = conf->mirrors[i].rdev;
1690		struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1691		if (repl
1692		    && !test_bit(Candidate, &repl->flags)
1693		    && repl->recovery_offset == MaxSector
1694		    && !test_bit(Faulty, &repl->flags)
1695		    && !test_and_set_bit(In_sync, &repl->flags)) {
1696			/* replacement has just become active */
1697			if (!rdev ||
1698			    !test_and_clear_bit(In_sync, &rdev->flags))
1699				count++;
1700			if (rdev) {
1701				/* Replaced device not technically
1702				 * faulty, but we need to be sure
1703				 * it gets removed and never re-added
1704				 */
1705				set_bit(Faulty, &rdev->flags);
1706				sysfs_notify_dirent_safe(
1707					rdev->sysfs_state);
1708			}
1709		}
1710		if (rdev
1711		    && rdev->recovery_offset == MaxSector
1712		    && !test_bit(Faulty, &rdev->flags)
1713		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1714			count++;
1715			sysfs_notify_dirent_safe(rdev->sysfs_state);
1716		}
1717	}
1718	mddev->degraded -= count;
1719	spin_unlock_irqrestore(&conf->device_lock, flags);
1720
1721	print_conf(conf);
1722	return count;
1723}
1724
1725static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1726{
1727	struct r1conf *conf = mddev->private;
1728	int err = -EEXIST;
1729	int mirror = 0;
1730	struct raid1_info *p;
1731	int first = 0;
1732	int last = conf->raid_disks - 1;
1733
1734	if (mddev->recovery_disabled == conf->recovery_disabled)
1735		return -EBUSY;
1736
1737	if (md_integrity_add_rdev(rdev, mddev))
1738		return -ENXIO;
1739
1740	if (rdev->raid_disk >= 0)
1741		first = last = rdev->raid_disk;
1742
1743	/*
1744	 * find the disk ... but prefer rdev->saved_raid_disk
1745	 * if possible.
1746	 */
1747	if (rdev->saved_raid_disk >= 0 &&
1748	    rdev->saved_raid_disk >= first &&
1749	    rdev->saved_raid_disk < conf->raid_disks &&
1750	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1751		first = last = rdev->saved_raid_disk;
1752
1753	for (mirror = first; mirror <= last; mirror++) {
1754		p = conf->mirrors + mirror;
1755		if (!p->rdev) {
 
1756			if (mddev->gendisk)
1757				disk_stack_limits(mddev->gendisk, rdev->bdev,
1758						  rdev->data_offset << 9);
1759
1760			p->head_position = 0;
1761			rdev->raid_disk = mirror;
1762			err = 0;
1763			/* As all devices are equivalent, we don't need a full recovery
1764			 * if this was recently any drive of the array
1765			 */
1766			if (rdev->saved_raid_disk < 0)
1767				conf->fullsync = 1;
1768			rcu_assign_pointer(p->rdev, rdev);
1769			break;
1770		}
1771		if (test_bit(WantReplacement, &p->rdev->flags) &&
1772		    p[conf->raid_disks].rdev == NULL) {
1773			/* Add this device as a replacement */
1774			clear_bit(In_sync, &rdev->flags);
1775			set_bit(Replacement, &rdev->flags);
1776			rdev->raid_disk = mirror;
1777			err = 0;
1778			conf->fullsync = 1;
1779			rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1780			break;
1781		}
1782	}
1783	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1784		blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1785	print_conf(conf);
1786	return err;
1787}
1788
1789static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1790{
1791	struct r1conf *conf = mddev->private;
1792	int err = 0;
1793	int number = rdev->raid_disk;
1794	struct raid1_info *p = conf->mirrors + number;
1795
1796	if (rdev != p->rdev)
1797		p = conf->mirrors + conf->raid_disks + number;
1798
1799	print_conf(conf);
1800	if (rdev == p->rdev) {
1801		if (test_bit(In_sync, &rdev->flags) ||
1802		    atomic_read(&rdev->nr_pending)) {
1803			err = -EBUSY;
1804			goto abort;
1805		}
1806		/* Only remove non-faulty devices if recovery
1807		 * is not possible.
1808		 */
1809		if (!test_bit(Faulty, &rdev->flags) &&
1810		    mddev->recovery_disabled != conf->recovery_disabled &&
1811		    mddev->degraded < conf->raid_disks) {
1812			err = -EBUSY;
1813			goto abort;
1814		}
1815		p->rdev = NULL;
1816		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1817			synchronize_rcu();
1818			if (atomic_read(&rdev->nr_pending)) {
1819				/* lost the race, try later */
1820				err = -EBUSY;
1821				p->rdev = rdev;
1822				goto abort;
1823			}
1824		}
1825		if (conf->mirrors[conf->raid_disks + number].rdev) {
1826			/* We just removed a device that is being replaced.
1827			 * Move down the replacement.  We drain all IO before
1828			 * doing this to avoid confusion.
1829			 */
1830			struct md_rdev *repl =
1831				conf->mirrors[conf->raid_disks + number].rdev;
1832			freeze_array(conf, 0);
1833			if (atomic_read(&repl->nr_pending)) {
1834				/* It means that some queued IO of retry_list
1835				 * hold repl. Thus, we cannot set replacement
1836				 * as NULL, avoiding rdev NULL pointer
1837				 * dereference in sync_request_write and
1838				 * handle_write_finished.
1839				 */
1840				err = -EBUSY;
1841				unfreeze_array(conf);
1842				goto abort;
1843			}
1844			clear_bit(Replacement, &repl->flags);
1845			p->rdev = repl;
1846			conf->mirrors[conf->raid_disks + number].rdev = NULL;
1847			unfreeze_array(conf);
1848		}
1849
1850		clear_bit(WantReplacement, &rdev->flags);
1851		err = md_integrity_register(mddev);
1852	}
1853abort:
1854
1855	print_conf(conf);
1856	return err;
1857}
1858
1859static void end_sync_read(struct bio *bio)
1860{
1861	struct r1bio *r1_bio = get_resync_r1bio(bio);
1862
1863	update_head_pos(r1_bio->read_disk, r1_bio);
1864
1865	/*
1866	 * we have read a block, now it needs to be re-written,
1867	 * or re-read if the read failed.
1868	 * We don't do much here, just schedule handling by raid1d
1869	 */
1870	if (!bio->bi_status)
1871		set_bit(R1BIO_Uptodate, &r1_bio->state);
1872
1873	if (atomic_dec_and_test(&r1_bio->remaining))
1874		reschedule_retry(r1_bio);
1875}
1876
1877static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1878{
1879	sector_t sync_blocks = 0;
1880	sector_t s = r1_bio->sector;
1881	long sectors_to_go = r1_bio->sectors;
1882
1883	/* make sure these bits don't get cleared. */
1884	do {
1885		md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1886		s += sync_blocks;
1887		sectors_to_go -= sync_blocks;
1888	} while (sectors_to_go > 0);
1889}
1890
1891static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1892{
1893	if (atomic_dec_and_test(&r1_bio->remaining)) {
1894		struct mddev *mddev = r1_bio->mddev;
1895		int s = r1_bio->sectors;
1896
1897		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1898		    test_bit(R1BIO_WriteError, &r1_bio->state))
1899			reschedule_retry(r1_bio);
1900		else {
1901			put_buf(r1_bio);
1902			md_done_sync(mddev, s, uptodate);
1903		}
1904	}
1905}
1906
1907static void end_sync_write(struct bio *bio)
1908{
1909	int uptodate = !bio->bi_status;
1910	struct r1bio *r1_bio = get_resync_r1bio(bio);
1911	struct mddev *mddev = r1_bio->mddev;
1912	struct r1conf *conf = mddev->private;
 
1913	sector_t first_bad;
1914	int bad_sectors;
1915	struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
 
1916
1917	if (!uptodate) {
1918		abort_sync_write(mddev, r1_bio);
1919		set_bit(WriteErrorSeen, &rdev->flags);
1920		if (!test_and_set_bit(WantReplacement, &rdev->flags))
 
 
 
 
 
 
 
 
 
 
 
1921			set_bit(MD_RECOVERY_NEEDED, &
1922				mddev->recovery);
1923		set_bit(R1BIO_WriteError, &r1_bio->state);
1924	} else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 
 
1925			       &first_bad, &bad_sectors) &&
1926		   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1927				r1_bio->sector,
1928				r1_bio->sectors,
1929				&first_bad, &bad_sectors)
1930		)
1931		set_bit(R1BIO_MadeGood, &r1_bio->state);
1932
1933	put_sync_write_buf(r1_bio, uptodate);
 
 
 
 
 
 
 
 
 
1934}
1935
1936static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1937			    int sectors, struct page *page, int rw)
1938{
1939	if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1940		/* success */
1941		return 1;
1942	if (rw == WRITE) {
1943		set_bit(WriteErrorSeen, &rdev->flags);
1944		if (!test_and_set_bit(WantReplacement,
1945				      &rdev->flags))
1946			set_bit(MD_RECOVERY_NEEDED, &
1947				rdev->mddev->recovery);
1948	}
1949	/* need to record an error - either for the block or the device */
1950	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1951		md_error(rdev->mddev, rdev);
1952	return 0;
1953}
1954
1955static int fix_sync_read_error(struct r1bio *r1_bio)
1956{
1957	/* Try some synchronous reads of other devices to get
1958	 * good data, much like with normal read errors.  Only
1959	 * read into the pages we already have so we don't
1960	 * need to re-issue the read request.
1961	 * We don't need to freeze the array, because being in an
1962	 * active sync request, there is no normal IO, and
1963	 * no overlapping syncs.
1964	 * We don't need to check is_badblock() again as we
1965	 * made sure that anything with a bad block in range
1966	 * will have bi_end_io clear.
1967	 */
1968	struct mddev *mddev = r1_bio->mddev;
1969	struct r1conf *conf = mddev->private;
1970	struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1971	struct page **pages = get_resync_pages(bio)->pages;
1972	sector_t sect = r1_bio->sector;
1973	int sectors = r1_bio->sectors;
1974	int idx = 0;
1975	struct md_rdev *rdev;
1976
1977	rdev = conf->mirrors[r1_bio->read_disk].rdev;
1978	if (test_bit(FailFast, &rdev->flags)) {
1979		/* Don't try recovering from here - just fail it
1980		 * ... unless it is the last working device of course */
1981		md_error(mddev, rdev);
1982		if (test_bit(Faulty, &rdev->flags))
1983			/* Don't try to read from here, but make sure
1984			 * put_buf does it's thing
1985			 */
1986			bio->bi_end_io = end_sync_write;
1987	}
1988
1989	while(sectors) {
1990		int s = sectors;
1991		int d = r1_bio->read_disk;
1992		int success = 0;
 
1993		int start;
1994
1995		if (s > (PAGE_SIZE>>9))
1996			s = PAGE_SIZE >> 9;
1997		do {
1998			if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1999				/* No rcu protection needed here devices
2000				 * can only be removed when no resync is
2001				 * active, and resync is currently active
2002				 */
2003				rdev = conf->mirrors[d].rdev;
2004				if (sync_page_io(rdev, sect, s<<9,
2005						 pages[idx],
2006						 REQ_OP_READ, 0, false)) {
2007					success = 1;
2008					break;
2009				}
2010			}
2011			d++;
2012			if (d == conf->raid_disks * 2)
2013				d = 0;
2014		} while (!success && d != r1_bio->read_disk);
2015
2016		if (!success) {
2017			char b[BDEVNAME_SIZE];
2018			int abort = 0;
2019			/* Cannot read from anywhere, this block is lost.
2020			 * Record a bad block on each device.  If that doesn't
2021			 * work just disable and interrupt the recovery.
2022			 * Don't fail devices as that won't really help.
2023			 */
2024			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2025					    mdname(mddev), bio_devname(bio, b),
2026					    (unsigned long long)r1_bio->sector);
 
 
2027			for (d = 0; d < conf->raid_disks * 2; d++) {
2028				rdev = conf->mirrors[d].rdev;
2029				if (!rdev || test_bit(Faulty, &rdev->flags))
2030					continue;
2031				if (!rdev_set_badblocks(rdev, sect, s, 0))
2032					abort = 1;
2033			}
2034			if (abort) {
2035				conf->recovery_disabled =
2036					mddev->recovery_disabled;
2037				set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2038				md_done_sync(mddev, r1_bio->sectors, 0);
2039				put_buf(r1_bio);
2040				return 0;
2041			}
2042			/* Try next page */
2043			sectors -= s;
2044			sect += s;
2045			idx++;
2046			continue;
2047		}
2048
2049		start = d;
2050		/* write it back and re-read */
2051		while (d != r1_bio->read_disk) {
2052			if (d == 0)
2053				d = conf->raid_disks * 2;
2054			d--;
2055			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2056				continue;
2057			rdev = conf->mirrors[d].rdev;
2058			if (r1_sync_page_io(rdev, sect, s,
2059					    pages[idx],
2060					    WRITE) == 0) {
2061				r1_bio->bios[d]->bi_end_io = NULL;
2062				rdev_dec_pending(rdev, mddev);
2063			}
2064		}
2065		d = start;
2066		while (d != r1_bio->read_disk) {
2067			if (d == 0)
2068				d = conf->raid_disks * 2;
2069			d--;
2070			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2071				continue;
2072			rdev = conf->mirrors[d].rdev;
2073			if (r1_sync_page_io(rdev, sect, s,
2074					    pages[idx],
2075					    READ) != 0)
2076				atomic_add(s, &rdev->corrected_errors);
2077		}
2078		sectors -= s;
2079		sect += s;
2080		idx ++;
2081	}
2082	set_bit(R1BIO_Uptodate, &r1_bio->state);
2083	bio->bi_status = 0;
2084	return 1;
2085}
2086
2087static void process_checks(struct r1bio *r1_bio)
2088{
2089	/* We have read all readable devices.  If we haven't
2090	 * got the block, then there is no hope left.
2091	 * If we have, then we want to do a comparison
2092	 * and skip the write if everything is the same.
2093	 * If any blocks failed to read, then we need to
2094	 * attempt an over-write
2095	 */
2096	struct mddev *mddev = r1_bio->mddev;
2097	struct r1conf *conf = mddev->private;
2098	int primary;
2099	int i;
2100	int vcnt;
2101
2102	/* Fix variable parts of all bios */
2103	vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2104	for (i = 0; i < conf->raid_disks * 2; i++) {
2105		blk_status_t status;
 
 
2106		struct bio *b = r1_bio->bios[i];
2107		struct resync_pages *rp = get_resync_pages(b);
2108		if (b->bi_end_io != end_sync_read)
2109			continue;
2110		/* fixup the bio for reuse, but preserve errno */
2111		status = b->bi_status;
2112		bio_reset(b);
2113		b->bi_status = status;
 
 
2114		b->bi_iter.bi_sector = r1_bio->sector +
2115			conf->mirrors[i].rdev->data_offset;
2116		bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2117		b->bi_end_io = end_sync_read;
2118		rp->raid_bio = r1_bio;
2119		b->bi_private = rp;
2120
2121		/* initialize bvec table again */
2122		md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
 
 
 
 
 
 
 
 
 
2123	}
2124	for (primary = 0; primary < conf->raid_disks * 2; primary++)
2125		if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2126		    !r1_bio->bios[primary]->bi_status) {
2127			r1_bio->bios[primary]->bi_end_io = NULL;
2128			rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2129			break;
2130		}
2131	r1_bio->read_disk = primary;
2132	for (i = 0; i < conf->raid_disks * 2; i++) {
2133		int j = 0;
2134		struct bio *pbio = r1_bio->bios[primary];
2135		struct bio *sbio = r1_bio->bios[i];
2136		blk_status_t status = sbio->bi_status;
2137		struct page **ppages = get_resync_pages(pbio)->pages;
2138		struct page **spages = get_resync_pages(sbio)->pages;
2139		struct bio_vec *bi;
2140		int page_len[RESYNC_PAGES] = { 0 };
2141		struct bvec_iter_all iter_all;
2142
2143		if (sbio->bi_end_io != end_sync_read)
2144			continue;
2145		/* Now we can 'fixup' the error value */
2146		sbio->bi_status = 0;
2147
2148		bio_for_each_segment_all(bi, sbio, iter_all)
2149			page_len[j++] = bi->bv_len;
2150
2151		if (!status) {
2152			for (j = vcnt; j-- ; ) {
2153				if (memcmp(page_address(ppages[j]),
2154					   page_address(spages[j]),
2155					   page_len[j]))
 
 
 
2156					break;
2157			}
2158		} else
2159			j = 0;
2160		if (j >= 0)
2161			atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2162		if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2163			      && !status)) {
2164			/* No need to write to this device. */
2165			sbio->bi_end_io = NULL;
2166			rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2167			continue;
2168		}
2169
2170		bio_copy_data(sbio, pbio);
2171	}
2172}
2173
2174static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2175{
2176	struct r1conf *conf = mddev->private;
2177	int i;
2178	int disks = conf->raid_disks * 2;
2179	struct bio *wbio;
 
 
2180
2181	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2182		/* ouch - failed to read all of that. */
2183		if (!fix_sync_read_error(r1_bio))
2184			return;
2185
2186	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2187		process_checks(r1_bio);
2188
2189	/*
2190	 * schedule writes
2191	 */
2192	atomic_set(&r1_bio->remaining, 1);
2193	for (i = 0; i < disks ; i++) {
2194		wbio = r1_bio->bios[i];
2195		if (wbio->bi_end_io == NULL ||
2196		    (wbio->bi_end_io == end_sync_read &&
2197		     (i == r1_bio->read_disk ||
2198		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2199			continue;
2200		if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2201			abort_sync_write(mddev, r1_bio);
2202			continue;
2203		}
2204
2205		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2206		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2207			wbio->bi_opf |= MD_FAILFAST;
2208
 
2209		wbio->bi_end_io = end_sync_write;
2210		atomic_inc(&r1_bio->remaining);
2211		md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2212
2213		submit_bio_noacct(wbio);
2214	}
2215
2216	put_sync_write_buf(r1_bio, 1);
 
 
 
 
 
 
 
 
 
 
2217}
2218
2219/*
2220 * This is a kernel thread which:
2221 *
2222 *	1.	Retries failed read operations on working mirrors.
2223 *	2.	Updates the raid superblock when problems encounter.
2224 *	3.	Performs writes following reads for array synchronising.
2225 */
2226
2227static void fix_read_error(struct r1conf *conf, int read_disk,
2228			   sector_t sect, int sectors)
2229{
2230	struct mddev *mddev = conf->mddev;
2231	while(sectors) {
2232		int s = sectors;
2233		int d = read_disk;
2234		int success = 0;
2235		int start;
2236		struct md_rdev *rdev;
2237
2238		if (s > (PAGE_SIZE>>9))
2239			s = PAGE_SIZE >> 9;
2240
2241		do {
 
 
 
 
 
2242			sector_t first_bad;
2243			int bad_sectors;
2244
2245			rcu_read_lock();
2246			rdev = rcu_dereference(conf->mirrors[d].rdev);
2247			if (rdev &&
2248			    (test_bit(In_sync, &rdev->flags) ||
2249			     (!test_bit(Faulty, &rdev->flags) &&
2250			      rdev->recovery_offset >= sect + s)) &&
2251			    is_badblock(rdev, sect, s,
2252					&first_bad, &bad_sectors) == 0) {
2253				atomic_inc(&rdev->nr_pending);
2254				rcu_read_unlock();
2255				if (sync_page_io(rdev, sect, s<<9,
2256					 conf->tmppage, REQ_OP_READ, 0, false))
2257					success = 1;
2258				rdev_dec_pending(rdev, mddev);
2259				if (success)
2260					break;
2261			} else
2262				rcu_read_unlock();
2263			d++;
2264			if (d == conf->raid_disks * 2)
2265				d = 0;
2266		} while (!success && d != read_disk);
2267
2268		if (!success) {
2269			/* Cannot read from anywhere - mark it bad */
2270			struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2271			if (!rdev_set_badblocks(rdev, sect, s, 0))
2272				md_error(mddev, rdev);
2273			break;
2274		}
2275		/* write it back and re-read */
2276		start = d;
2277		while (d != read_disk) {
2278			if (d==0)
2279				d = conf->raid_disks * 2;
2280			d--;
2281			rcu_read_lock();
2282			rdev = rcu_dereference(conf->mirrors[d].rdev);
2283			if (rdev &&
2284			    !test_bit(Faulty, &rdev->flags)) {
2285				atomic_inc(&rdev->nr_pending);
2286				rcu_read_unlock();
2287				r1_sync_page_io(rdev, sect, s,
2288						conf->tmppage, WRITE);
2289				rdev_dec_pending(rdev, mddev);
2290			} else
2291				rcu_read_unlock();
2292		}
2293		d = start;
2294		while (d != read_disk) {
2295			char b[BDEVNAME_SIZE];
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				if (r1_sync_page_io(rdev, sect, s,
2306						    conf->tmppage, READ)) {
2307					atomic_add(s, &rdev->corrected_errors);
2308					pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2309						mdname(mddev), s,
2310						(unsigned long long)(sect +
2311								     rdev->data_offset),
2312						bdevname(rdev->bdev, b));
 
 
2313				}
2314				rdev_dec_pending(rdev, mddev);
2315			} else
2316				rcu_read_unlock();
2317		}
2318		sectors -= s;
2319		sect += s;
2320	}
2321}
2322
2323static int narrow_write_error(struct r1bio *r1_bio, int i)
2324{
2325	struct mddev *mddev = r1_bio->mddev;
2326	struct r1conf *conf = mddev->private;
2327	struct md_rdev *rdev = conf->mirrors[i].rdev;
2328
2329	/* bio has the data to be written to device 'i' where
2330	 * we just recently had a write error.
2331	 * We repeatedly clone the bio and trim down to one block,
2332	 * then try the write.  Where the write fails we record
2333	 * a bad block.
2334	 * It is conceivable that the bio doesn't exactly align with
2335	 * blocks.  We must handle this somehow.
2336	 *
2337	 * We currently own a reference on the rdev.
2338	 */
2339
2340	int block_sectors;
2341	sector_t sector;
2342	int sectors;
2343	int sect_to_write = r1_bio->sectors;
2344	int ok = 1;
2345
2346	if (rdev->badblocks.shift < 0)
2347		return 0;
2348
2349	block_sectors = roundup(1 << rdev->badblocks.shift,
2350				bdev_logical_block_size(rdev->bdev) >> 9);
2351	sector = r1_bio->sector;
2352	sectors = ((sector + block_sectors)
2353		   & ~(sector_t)(block_sectors - 1))
2354		- sector;
2355
2356	while (sect_to_write) {
2357		struct bio *wbio;
2358		if (sectors > sect_to_write)
2359			sectors = sect_to_write;
2360		/* Write at 'sector' for 'sectors'*/
2361
2362		if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2363			wbio = bio_clone_fast(r1_bio->behind_master_bio,
2364					      GFP_NOIO,
2365					      &mddev->bio_set);
 
 
 
 
 
 
 
 
 
2366		} else {
2367			wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2368					      &mddev->bio_set);
2369		}
2370
2371		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2372		wbio->bi_iter.bi_sector = r1_bio->sector;
2373		wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2374
2375		bio_trim(wbio, sector - r1_bio->sector, sectors);
2376		wbio->bi_iter.bi_sector += rdev->data_offset;
2377		bio_set_dev(wbio, rdev->bdev);
2378
2379		if (submit_bio_wait(wbio) < 0)
2380			/* failure! */
2381			ok = rdev_set_badblocks(rdev, sector,
2382						sectors, 0)
2383				&& ok;
2384
2385		bio_put(wbio);
2386		sect_to_write -= sectors;
2387		sector += sectors;
2388		sectors = block_sectors;
2389	}
2390	return ok;
2391}
2392
2393static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2394{
2395	int m;
2396	int s = r1_bio->sectors;
2397	for (m = 0; m < conf->raid_disks * 2 ; m++) {
2398		struct md_rdev *rdev = conf->mirrors[m].rdev;
2399		struct bio *bio = r1_bio->bios[m];
2400		if (bio->bi_end_io == NULL)
2401			continue;
2402		if (!bio->bi_status &&
2403		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2404			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2405		}
2406		if (bio->bi_status &&
2407		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2408			if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2409				md_error(conf->mddev, rdev);
2410		}
2411	}
2412	put_buf(r1_bio);
2413	md_done_sync(conf->mddev, s, 1);
2414}
2415
2416static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2417{
2418	int m, idx;
2419	bool fail = false;
2420
2421	for (m = 0; m < conf->raid_disks * 2 ; m++)
2422		if (r1_bio->bios[m] == IO_MADE_GOOD) {
2423			struct md_rdev *rdev = conf->mirrors[m].rdev;
2424			rdev_clear_badblocks(rdev,
2425					     r1_bio->sector,
2426					     r1_bio->sectors, 0);
2427			rdev_dec_pending(rdev, conf->mddev);
2428		} else if (r1_bio->bios[m] != NULL) {
2429			/* This drive got a write error.  We need to
2430			 * narrow down and record precise write
2431			 * errors.
2432			 */
2433			fail = true;
2434			if (!narrow_write_error(r1_bio, m)) {
2435				md_error(conf->mddev,
2436					 conf->mirrors[m].rdev);
2437				/* an I/O failed, we can't clear the bitmap */
2438				set_bit(R1BIO_Degraded, &r1_bio->state);
2439			}
2440			rdev_dec_pending(conf->mirrors[m].rdev,
2441					 conf->mddev);
2442		}
2443	if (fail) {
2444		spin_lock_irq(&conf->device_lock);
2445		list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2446		idx = sector_to_idx(r1_bio->sector);
2447		atomic_inc(&conf->nr_queued[idx]);
2448		spin_unlock_irq(&conf->device_lock);
2449		/*
2450		 * In case freeze_array() is waiting for condition
2451		 * get_unqueued_pending() == extra to be true.
2452		 */
2453		wake_up(&conf->wait_barrier);
2454		md_wakeup_thread(conf->mddev->thread);
2455	} else {
2456		if (test_bit(R1BIO_WriteError, &r1_bio->state))
2457			close_write(r1_bio);
2458		raid_end_bio_io(r1_bio);
2459	}
2460}
2461
2462static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2463{
 
 
2464	struct mddev *mddev = conf->mddev;
2465	struct bio *bio;
 
2466	struct md_rdev *rdev;
2467
2468	clear_bit(R1BIO_ReadError, &r1_bio->state);
2469	/* we got a read error. Maybe the drive is bad.  Maybe just
2470	 * the block and we can fix it.
2471	 * We freeze all other IO, and try reading the block from
2472	 * other devices.  When we find one, we re-write
2473	 * and check it that fixes the read error.
2474	 * This is all done synchronously while the array is
2475	 * frozen
2476	 */
2477
2478	bio = r1_bio->bios[r1_bio->read_disk];
2479	bio_put(bio);
2480	r1_bio->bios[r1_bio->read_disk] = NULL;
2481
2482	rdev = conf->mirrors[r1_bio->read_disk].rdev;
2483	if (mddev->ro == 0
2484	    && !test_bit(FailFast, &rdev->flags)) {
2485		freeze_array(conf, 1);
2486		fix_read_error(conf, r1_bio->read_disk,
2487			       r1_bio->sector, r1_bio->sectors);
2488		unfreeze_array(conf);
2489	} else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2490		md_error(mddev, rdev);
 
 
 
 
 
 
 
 
 
 
 
2491	} else {
2492		r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2493	}
2494
2495	rdev_dec_pending(rdev, conf->mddev);
2496	allow_barrier(conf, r1_bio->sector);
2497	bio = r1_bio->master_bio;
2498
2499	/* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2500	r1_bio->state = 0;
2501	raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2502}
2503
2504static void raid1d(struct md_thread *thread)
2505{
2506	struct mddev *mddev = thread->mddev;
2507	struct r1bio *r1_bio;
2508	unsigned long flags;
2509	struct r1conf *conf = mddev->private;
2510	struct list_head *head = &conf->retry_list;
2511	struct blk_plug plug;
2512	int idx;
2513
2514	md_check_recovery(mddev);
2515
2516	if (!list_empty_careful(&conf->bio_end_io_list) &&
2517	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2518		LIST_HEAD(tmp);
2519		spin_lock_irqsave(&conf->device_lock, flags);
2520		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2521			list_splice_init(&conf->bio_end_io_list, &tmp);
 
 
 
 
2522		spin_unlock_irqrestore(&conf->device_lock, flags);
2523		while (!list_empty(&tmp)) {
2524			r1_bio = list_first_entry(&tmp, struct r1bio,
2525						  retry_list);
2526			list_del(&r1_bio->retry_list);
2527			idx = sector_to_idx(r1_bio->sector);
2528			atomic_dec(&conf->nr_queued[idx]);
2529			if (mddev->degraded)
2530				set_bit(R1BIO_Degraded, &r1_bio->state);
2531			if (test_bit(R1BIO_WriteError, &r1_bio->state))
2532				close_write(r1_bio);
2533			raid_end_bio_io(r1_bio);
2534		}
2535	}
2536
2537	blk_start_plug(&plug);
2538	for (;;) {
2539
2540		flush_pending_writes(conf);
2541
2542		spin_lock_irqsave(&conf->device_lock, flags);
2543		if (list_empty(head)) {
2544			spin_unlock_irqrestore(&conf->device_lock, flags);
2545			break;
2546		}
2547		r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2548		list_del(head->prev);
2549		idx = sector_to_idx(r1_bio->sector);
2550		atomic_dec(&conf->nr_queued[idx]);
2551		spin_unlock_irqrestore(&conf->device_lock, flags);
2552
2553		mddev = r1_bio->mddev;
2554		conf = mddev->private;
2555		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2556			if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2557			    test_bit(R1BIO_WriteError, &r1_bio->state))
2558				handle_sync_write_finished(conf, r1_bio);
2559			else
2560				sync_request_write(mddev, r1_bio);
2561		} else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2562			   test_bit(R1BIO_WriteError, &r1_bio->state))
2563			handle_write_finished(conf, r1_bio);
2564		else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2565			handle_read_error(conf, r1_bio);
2566		else
2567			WARN_ON_ONCE(1);
 
 
 
2568
2569		cond_resched();
2570		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2571			md_check_recovery(mddev);
2572	}
2573	blk_finish_plug(&plug);
2574}
2575
2576static int init_resync(struct r1conf *conf)
2577{
2578	int buffs;
2579
2580	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2581	BUG_ON(mempool_initialized(&conf->r1buf_pool));
2582
2583	return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2584			    r1buf_pool_free, conf->poolinfo);
2585}
2586
2587static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2588{
2589	struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2590	struct resync_pages *rps;
2591	struct bio *bio;
2592	int i;
2593
2594	for (i = conf->poolinfo->raid_disks; i--; ) {
2595		bio = r1bio->bios[i];
2596		rps = bio->bi_private;
2597		bio_reset(bio);
2598		bio->bi_private = rps;
2599	}
2600	r1bio->master_bio = NULL;
2601	return r1bio;
2602}
2603
2604/*
2605 * perform a "sync" on one "block"
2606 *
2607 * We need to make sure that no normal I/O request - particularly write
2608 * requests - conflict with active sync requests.
2609 *
2610 * This is achieved by tracking pending requests and a 'barrier' concept
2611 * that can be installed to exclude normal IO requests.
2612 */
2613
2614static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2615				   int *skipped)
2616{
2617	struct r1conf *conf = mddev->private;
2618	struct r1bio *r1_bio;
2619	struct bio *bio;
2620	sector_t max_sector, nr_sectors;
2621	int disk = -1;
2622	int i;
2623	int wonly = -1;
2624	int write_targets = 0, read_targets = 0;
2625	sector_t sync_blocks;
2626	int still_degraded = 0;
2627	int good_sectors = RESYNC_SECTORS;
2628	int min_bad = 0; /* number of sectors that are bad in all devices */
2629	int idx = sector_to_idx(sector_nr);
2630	int page_idx = 0;
2631
2632	if (!mempool_initialized(&conf->r1buf_pool))
2633		if (init_resync(conf))
2634			return 0;
2635
2636	max_sector = mddev->dev_sectors;
2637	if (sector_nr >= max_sector) {
2638		/* If we aborted, we need to abort the
2639		 * sync on the 'current' bitmap chunk (there will
2640		 * only be one in raid1 resync.
2641		 * We can find the current addess in mddev->curr_resync
2642		 */
2643		if (mddev->curr_resync < max_sector) /* aborted */
2644			md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2645					   &sync_blocks, 1);
2646		else /* completed sync */
2647			conf->fullsync = 0;
2648
2649		md_bitmap_close_sync(mddev->bitmap);
2650		close_sync(conf);
2651
2652		if (mddev_is_clustered(mddev)) {
2653			conf->cluster_sync_low = 0;
2654			conf->cluster_sync_high = 0;
2655		}
2656		return 0;
2657	}
2658
2659	if (mddev->bitmap == NULL &&
2660	    mddev->recovery_cp == MaxSector &&
2661	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2662	    conf->fullsync == 0) {
2663		*skipped = 1;
2664		return max_sector - sector_nr;
2665	}
2666	/* before building a request, check if we can skip these blocks..
2667	 * This call the bitmap_start_sync doesn't actually record anything
2668	 */
2669	if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2670	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2671		/* We can skip this block, and probably several more */
2672		*skipped = 1;
2673		return sync_blocks;
2674	}
2675
2676	/*
2677	 * If there is non-resync activity waiting for a turn, then let it
2678	 * though before starting on this new sync request.
2679	 */
2680	if (atomic_read(&conf->nr_waiting[idx]))
2681		schedule_timeout_uninterruptible(1);
2682
2683	/* we are incrementing sector_nr below. To be safe, we check against
2684	 * sector_nr + two times RESYNC_SECTORS
2685	 */
2686
2687	md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2688		mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
 
2689
2690
2691	if (raise_barrier(conf, sector_nr))
2692		return 0;
2693
2694	r1_bio = raid1_alloc_init_r1buf(conf);
2695
2696	rcu_read_lock();
2697	/*
2698	 * If we get a correctably read error during resync or recovery,
2699	 * we might want to read from a different device.  So we
2700	 * flag all drives that could conceivably be read from for READ,
2701	 * and any others (which will be non-In_sync devices) for WRITE.
2702	 * If a read fails, we try reading from something else for which READ
2703	 * is OK.
2704	 */
2705
2706	r1_bio->mddev = mddev;
2707	r1_bio->sector = sector_nr;
2708	r1_bio->state = 0;
2709	set_bit(R1BIO_IsSync, &r1_bio->state);
2710	/* make sure good_sectors won't go across barrier unit boundary */
2711	good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2712
2713	for (i = 0; i < conf->raid_disks * 2; i++) {
2714		struct md_rdev *rdev;
2715		bio = r1_bio->bios[i];
 
2716
2717		rdev = rcu_dereference(conf->mirrors[i].rdev);
2718		if (rdev == NULL ||
2719		    test_bit(Faulty, &rdev->flags)) {
2720			if (i < conf->raid_disks)
2721				still_degraded = 1;
2722		} else if (!test_bit(In_sync, &rdev->flags)) {
2723			bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2724			bio->bi_end_io = end_sync_write;
2725			write_targets ++;
2726		} else {
2727			/* may need to read from here */
2728			sector_t first_bad = MaxSector;
2729			int bad_sectors;
2730
2731			if (is_badblock(rdev, sector_nr, good_sectors,
2732					&first_bad, &bad_sectors)) {
2733				if (first_bad > sector_nr)
2734					good_sectors = first_bad - sector_nr;
2735				else {
2736					bad_sectors -= (sector_nr - first_bad);
2737					if (min_bad == 0 ||
2738					    min_bad > bad_sectors)
2739						min_bad = bad_sectors;
2740				}
2741			}
2742			if (sector_nr < first_bad) {
2743				if (test_bit(WriteMostly, &rdev->flags)) {
2744					if (wonly < 0)
2745						wonly = i;
2746				} else {
2747					if (disk < 0)
2748						disk = i;
2749				}
2750				bio_set_op_attrs(bio, REQ_OP_READ, 0);
2751				bio->bi_end_io = end_sync_read;
2752				read_targets++;
2753			} else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2754				test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2755				!test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2756				/*
2757				 * The device is suitable for reading (InSync),
2758				 * but has bad block(s) here. Let's try to correct them,
2759				 * if we are doing resync or repair. Otherwise, leave
2760				 * this device alone for this sync request.
2761				 */
2762				bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2763				bio->bi_end_io = end_sync_write;
2764				write_targets++;
2765			}
2766		}
2767		if (rdev && bio->bi_end_io) {
2768			atomic_inc(&rdev->nr_pending);
2769			bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2770			bio_set_dev(bio, rdev->bdev);
2771			if (test_bit(FailFast, &rdev->flags))
2772				bio->bi_opf |= MD_FAILFAST;
2773		}
2774	}
2775	rcu_read_unlock();
2776	if (disk < 0)
2777		disk = wonly;
2778	r1_bio->read_disk = disk;
2779
2780	if (read_targets == 0 && min_bad > 0) {
2781		/* These sectors are bad on all InSync devices, so we
2782		 * need to mark them bad on all write targets
2783		 */
2784		int ok = 1;
2785		for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2786			if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2787				struct md_rdev *rdev = conf->mirrors[i].rdev;
2788				ok = rdev_set_badblocks(rdev, sector_nr,
2789							min_bad, 0
2790					) && ok;
2791			}
2792		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2793		*skipped = 1;
2794		put_buf(r1_bio);
2795
2796		if (!ok) {
2797			/* Cannot record the badblocks, so need to
2798			 * abort the resync.
2799			 * If there are multiple read targets, could just
2800			 * fail the really bad ones ???
2801			 */
2802			conf->recovery_disabled = mddev->recovery_disabled;
2803			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2804			return 0;
2805		} else
2806			return min_bad;
2807
2808	}
2809	if (min_bad > 0 && min_bad < good_sectors) {
2810		/* only resync enough to reach the next bad->good
2811		 * transition */
2812		good_sectors = min_bad;
2813	}
2814
2815	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2816		/* extra read targets are also write targets */
2817		write_targets += read_targets-1;
2818
2819	if (write_targets == 0 || read_targets == 0) {
2820		/* There is nowhere to write, so all non-sync
2821		 * drives must be failed - so we are finished
2822		 */
2823		sector_t rv;
2824		if (min_bad > 0)
2825			max_sector = sector_nr + min_bad;
2826		rv = max_sector - sector_nr;
2827		*skipped = 1;
2828		put_buf(r1_bio);
2829		return rv;
2830	}
2831
2832	if (max_sector > mddev->resync_max)
2833		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2834	if (max_sector > sector_nr + good_sectors)
2835		max_sector = sector_nr + good_sectors;
2836	nr_sectors = 0;
2837	sync_blocks = 0;
2838	do {
2839		struct page *page;
2840		int len = PAGE_SIZE;
2841		if (sector_nr + (len>>9) > max_sector)
2842			len = (max_sector - sector_nr) << 9;
2843		if (len == 0)
2844			break;
2845		if (sync_blocks == 0) {
2846			if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2847						  &sync_blocks, still_degraded) &&
2848			    !conf->fullsync &&
2849			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2850				break;
2851			if ((len >> 9) > sync_blocks)
2852				len = sync_blocks<<9;
2853		}
2854
2855		for (i = 0 ; i < conf->raid_disks * 2; i++) {
2856			struct resync_pages *rp;
2857
2858			bio = r1_bio->bios[i];
2859			rp = get_resync_pages(bio);
2860			if (bio->bi_end_io) {
2861				page = resync_fetch_page(rp, page_idx);
2862
2863				/*
2864				 * won't fail because the vec table is big
2865				 * enough to hold all these pages
2866				 */
2867				bio_add_page(bio, page, len, 0);
 
 
 
 
 
 
 
 
 
2868			}
2869		}
2870		nr_sectors += len>>9;
2871		sector_nr += len>>9;
2872		sync_blocks -= (len>>9);
2873	} while (++page_idx < RESYNC_PAGES);
2874
2875	r1_bio->sectors = nr_sectors;
2876
2877	if (mddev_is_clustered(mddev) &&
2878			conf->cluster_sync_high < sector_nr + nr_sectors) {
2879		conf->cluster_sync_low = mddev->curr_resync_completed;
2880		conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2881		/* Send resync message */
2882		md_cluster_ops->resync_info_update(mddev,
2883				conf->cluster_sync_low,
2884				conf->cluster_sync_high);
2885	}
2886
2887	/* For a user-requested sync, we read all readable devices and do a
2888	 * compare
2889	 */
2890	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2891		atomic_set(&r1_bio->remaining, read_targets);
2892		for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2893			bio = r1_bio->bios[i];
2894			if (bio->bi_end_io == end_sync_read) {
2895				read_targets--;
2896				md_sync_acct_bio(bio, nr_sectors);
2897				if (read_targets == 1)
2898					bio->bi_opf &= ~MD_FAILFAST;
2899				submit_bio_noacct(bio);
2900			}
2901		}
2902	} else {
2903		atomic_set(&r1_bio->remaining, 1);
2904		bio = r1_bio->bios[r1_bio->read_disk];
2905		md_sync_acct_bio(bio, nr_sectors);
2906		if (read_targets == 1)
2907			bio->bi_opf &= ~MD_FAILFAST;
2908		submit_bio_noacct(bio);
2909	}
2910	return nr_sectors;
2911}
2912
2913static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2914{
2915	if (sectors)
2916		return sectors;
2917
2918	return mddev->dev_sectors;
2919}
2920
2921static struct r1conf *setup_conf(struct mddev *mddev)
2922{
2923	struct r1conf *conf;
2924	int i;
2925	struct raid1_info *disk;
2926	struct md_rdev *rdev;
2927	int err = -ENOMEM;
2928
2929	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2930	if (!conf)
2931		goto abort;
2932
2933	conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2934				   sizeof(atomic_t), GFP_KERNEL);
2935	if (!conf->nr_pending)
2936		goto abort;
2937
2938	conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2939				   sizeof(atomic_t), GFP_KERNEL);
2940	if (!conf->nr_waiting)
2941		goto abort;
2942
2943	conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2944				  sizeof(atomic_t), GFP_KERNEL);
2945	if (!conf->nr_queued)
2946		goto abort;
2947
2948	conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2949				sizeof(atomic_t), GFP_KERNEL);
2950	if (!conf->barrier)
2951		goto abort;
2952
2953	conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2954					    mddev->raid_disks, 2),
2955				GFP_KERNEL);
2956	if (!conf->mirrors)
2957		goto abort;
2958
2959	conf->tmppage = alloc_page(GFP_KERNEL);
2960	if (!conf->tmppage)
2961		goto abort;
2962
2963	conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2964	if (!conf->poolinfo)
2965		goto abort;
2966	conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2967	err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2968			   rbio_pool_free, conf->poolinfo);
2969	if (err)
2970		goto abort;
2971
2972	err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2973	if (err)
2974		goto abort;
2975
2976	conf->poolinfo->mddev = mddev;
2977
2978	err = -EINVAL;
2979	spin_lock_init(&conf->device_lock);
2980	rdev_for_each(rdev, mddev) {
 
2981		int disk_idx = rdev->raid_disk;
2982		if (disk_idx >= mddev->raid_disks
2983		    || disk_idx < 0)
2984			continue;
2985		if (test_bit(Replacement, &rdev->flags))
2986			disk = conf->mirrors + mddev->raid_disks + disk_idx;
2987		else
2988			disk = conf->mirrors + disk_idx;
2989
2990		if (disk->rdev)
2991			goto abort;
2992		disk->rdev = rdev;
 
 
2993		disk->head_position = 0;
2994		disk->seq_start = MaxSector;
2995	}
2996	conf->raid_disks = mddev->raid_disks;
2997	conf->mddev = mddev;
2998	INIT_LIST_HEAD(&conf->retry_list);
2999	INIT_LIST_HEAD(&conf->bio_end_io_list);
3000
3001	spin_lock_init(&conf->resync_lock);
3002	init_waitqueue_head(&conf->wait_barrier);
3003
3004	bio_list_init(&conf->pending_bio_list);
3005	conf->pending_count = 0;
3006	conf->recovery_disabled = mddev->recovery_disabled - 1;
3007
 
 
 
3008	err = -EIO;
3009	for (i = 0; i < conf->raid_disks * 2; i++) {
3010
3011		disk = conf->mirrors + i;
3012
3013		if (i < conf->raid_disks &&
3014		    disk[conf->raid_disks].rdev) {
3015			/* This slot has a replacement. */
3016			if (!disk->rdev) {
3017				/* No original, just make the replacement
3018				 * a recovering spare
3019				 */
3020				disk->rdev =
3021					disk[conf->raid_disks].rdev;
3022				disk[conf->raid_disks].rdev = NULL;
3023			} else if (!test_bit(In_sync, &disk->rdev->flags))
3024				/* Original is not in_sync - bad */
3025				goto abort;
3026		}
3027
3028		if (!disk->rdev ||
3029		    !test_bit(In_sync, &disk->rdev->flags)) {
3030			disk->head_position = 0;
3031			if (disk->rdev &&
3032			    (disk->rdev->saved_raid_disk < 0))
3033				conf->fullsync = 1;
3034		}
3035	}
3036
3037	err = -ENOMEM;
3038	conf->thread = md_register_thread(raid1d, mddev, "raid1");
3039	if (!conf->thread)
 
 
 
3040		goto abort;
 
3041
3042	return conf;
3043
3044 abort:
3045	if (conf) {
3046		mempool_exit(&conf->r1bio_pool);
3047		kfree(conf->mirrors);
3048		safe_put_page(conf->tmppage);
3049		kfree(conf->poolinfo);
3050		kfree(conf->nr_pending);
3051		kfree(conf->nr_waiting);
3052		kfree(conf->nr_queued);
3053		kfree(conf->barrier);
3054		bioset_exit(&conf->bio_split);
3055		kfree(conf);
3056	}
3057	return ERR_PTR(err);
3058}
3059
3060static void raid1_free(struct mddev *mddev, void *priv);
3061static int raid1_run(struct mddev *mddev)
3062{
3063	struct r1conf *conf;
3064	int i;
3065	struct md_rdev *rdev;
3066	int ret;
3067	bool discard_supported = false;
3068
3069	if (mddev->level != 1) {
3070		pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3071			mdname(mddev), mddev->level);
3072		return -EIO;
3073	}
3074	if (mddev->reshape_position != MaxSector) {
3075		pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3076			mdname(mddev));
3077		return -EIO;
3078	}
3079	if (mddev_init_writes_pending(mddev) < 0)
3080		return -ENOMEM;
3081	/*
3082	 * copy the already verified devices into our private RAID1
3083	 * bookkeeping area. [whatever we allocate in run(),
3084	 * should be freed in raid1_free()]
3085	 */
3086	if (mddev->private == NULL)
3087		conf = setup_conf(mddev);
3088	else
3089		conf = mddev->private;
3090
3091	if (IS_ERR(conf))
3092		return PTR_ERR(conf);
3093
3094	if (mddev->queue) {
3095		blk_queue_max_write_same_sectors(mddev->queue, 0);
3096		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3097	}
3098
3099	rdev_for_each(rdev, mddev) {
3100		if (!mddev->gendisk)
3101			continue;
3102		disk_stack_limits(mddev->gendisk, rdev->bdev,
3103				  rdev->data_offset << 9);
3104		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3105			discard_supported = true;
3106	}
3107
3108	mddev->degraded = 0;
3109	for (i = 0; i < conf->raid_disks; i++)
3110		if (conf->mirrors[i].rdev == NULL ||
3111		    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3112		    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3113			mddev->degraded++;
3114	/*
3115	 * RAID1 needs at least one disk in active
3116	 */
3117	if (conf->raid_disks - mddev->degraded < 1) {
3118		ret = -EINVAL;
3119		goto abort;
3120	}
3121
3122	if (conf->raid_disks - mddev->degraded == 1)
3123		mddev->recovery_cp = MaxSector;
3124
3125	if (mddev->recovery_cp != MaxSector)
3126		pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3127			mdname(mddev));
3128	pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
 
 
3129		mdname(mddev), mddev->raid_disks - mddev->degraded,
3130		mddev->raid_disks);
3131
3132	/*
3133	 * Ok, everything is just fine now
3134	 */
3135	mddev->thread = conf->thread;
3136	conf->thread = NULL;
3137	mddev->private = conf;
3138	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3139
3140	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3141
3142	if (mddev->queue) {
3143		if (discard_supported)
3144			blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3145						mddev->queue);
3146		else
3147			blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3148						  mddev->queue);
3149	}
3150
3151	ret = md_integrity_register(mddev);
3152	if (ret) {
3153		md_unregister_thread(&mddev->thread);
3154		goto abort;
3155	}
3156	return 0;
3157
3158abort:
3159	raid1_free(mddev, conf);
3160	return ret;
3161}
3162
3163static void raid1_free(struct mddev *mddev, void *priv)
3164{
3165	struct r1conf *conf = priv;
3166
3167	mempool_exit(&conf->r1bio_pool);
3168	kfree(conf->mirrors);
3169	safe_put_page(conf->tmppage);
3170	kfree(conf->poolinfo);
3171	kfree(conf->nr_pending);
3172	kfree(conf->nr_waiting);
3173	kfree(conf->nr_queued);
3174	kfree(conf->barrier);
3175	bioset_exit(&conf->bio_split);
3176	kfree(conf);
3177}
3178
3179static int raid1_resize(struct mddev *mddev, sector_t sectors)
3180{
3181	/* no resync is happening, and there is enough space
3182	 * on all devices, so we can resize.
3183	 * We need to make sure resync covers any new space.
3184	 * If the array is shrinking we should possibly wait until
3185	 * any io in the removed space completes, but it hardly seems
3186	 * worth it.
3187	 */
3188	sector_t newsize = raid1_size(mddev, sectors, 0);
3189	if (mddev->external_size &&
3190	    mddev->array_sectors > newsize)
3191		return -EINVAL;
3192	if (mddev->bitmap) {
3193		int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3194		if (ret)
3195			return ret;
3196	}
3197	md_set_array_sectors(mddev, newsize);
 
 
3198	if (sectors > mddev->dev_sectors &&
3199	    mddev->recovery_cp > mddev->dev_sectors) {
3200		mddev->recovery_cp = mddev->dev_sectors;
3201		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3202	}
3203	mddev->dev_sectors = sectors;
3204	mddev->resync_max_sectors = sectors;
3205	return 0;
3206}
3207
3208static int raid1_reshape(struct mddev *mddev)
3209{
3210	/* We need to:
3211	 * 1/ resize the r1bio_pool
3212	 * 2/ resize conf->mirrors
3213	 *
3214	 * We allocate a new r1bio_pool if we can.
3215	 * Then raise a device barrier and wait until all IO stops.
3216	 * Then resize conf->mirrors and swap in the new r1bio pool.
3217	 *
3218	 * At the same time, we "pack" the devices so that all the missing
3219	 * devices have the higher raid_disk numbers.
3220	 */
3221	mempool_t newpool, oldpool;
3222	struct pool_info *newpoolinfo;
3223	struct raid1_info *newmirrors;
3224	struct r1conf *conf = mddev->private;
3225	int cnt, raid_disks;
3226	unsigned long flags;
3227	int d, d2;
3228	int ret;
3229
3230	memset(&newpool, 0, sizeof(newpool));
3231	memset(&oldpool, 0, sizeof(oldpool));
3232
3233	/* Cannot change chunk_size, layout, or level */
3234	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3235	    mddev->layout != mddev->new_layout ||
3236	    mddev->level != mddev->new_level) {
3237		mddev->new_chunk_sectors = mddev->chunk_sectors;
3238		mddev->new_layout = mddev->layout;
3239		mddev->new_level = mddev->level;
3240		return -EINVAL;
3241	}
3242
3243	if (!mddev_is_clustered(mddev))
3244		md_allow_write(mddev);
 
 
 
3245
3246	raid_disks = mddev->raid_disks + mddev->delta_disks;
3247
3248	if (raid_disks < conf->raid_disks) {
3249		cnt=0;
3250		for (d= 0; d < conf->raid_disks; d++)
3251			if (conf->mirrors[d].rdev)
3252				cnt++;
3253		if (cnt > raid_disks)
3254			return -EBUSY;
3255	}
3256
3257	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3258	if (!newpoolinfo)
3259		return -ENOMEM;
3260	newpoolinfo->mddev = mddev;
3261	newpoolinfo->raid_disks = raid_disks * 2;
3262
3263	ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3264			   rbio_pool_free, newpoolinfo);
3265	if (ret) {
3266		kfree(newpoolinfo);
3267		return ret;
3268	}
3269	newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3270					 raid_disks, 2),
3271			     GFP_KERNEL);
3272	if (!newmirrors) {
3273		kfree(newpoolinfo);
3274		mempool_exit(&newpool);
3275		return -ENOMEM;
3276	}
3277
3278	freeze_array(conf, 0);
3279
3280	/* ok, everything is stopped */
3281	oldpool = conf->r1bio_pool;
3282	conf->r1bio_pool = newpool;
3283
3284	for (d = d2 = 0; d < conf->raid_disks; d++) {
3285		struct md_rdev *rdev = conf->mirrors[d].rdev;
3286		if (rdev && rdev->raid_disk != d2) {
3287			sysfs_unlink_rdev(mddev, rdev);
3288			rdev->raid_disk = d2;
3289			sysfs_unlink_rdev(mddev, rdev);
3290			if (sysfs_link_rdev(mddev, rdev))
3291				pr_warn("md/raid1:%s: cannot register rd%d\n",
3292					mdname(mddev), rdev->raid_disk);
 
3293		}
3294		if (rdev)
3295			newmirrors[d2++].rdev = rdev;
3296	}
3297	kfree(conf->mirrors);
3298	conf->mirrors = newmirrors;
3299	kfree(conf->poolinfo);
3300	conf->poolinfo = newpoolinfo;
3301
3302	spin_lock_irqsave(&conf->device_lock, flags);
3303	mddev->degraded += (raid_disks - conf->raid_disks);
3304	spin_unlock_irqrestore(&conf->device_lock, flags);
3305	conf->raid_disks = mddev->raid_disks = raid_disks;
3306	mddev->delta_disks = 0;
3307
3308	unfreeze_array(conf);
3309
3310	set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3311	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3312	md_wakeup_thread(mddev->thread);
3313
3314	mempool_exit(&oldpool);
3315	return 0;
3316}
3317
3318static void raid1_quiesce(struct mddev *mddev, int quiesce)
3319{
3320	struct r1conf *conf = mddev->private;
3321
3322	if (quiesce)
 
 
 
 
3323		freeze_array(conf, 0);
3324	else
 
3325		unfreeze_array(conf);
 
 
3326}
3327
3328static void *raid1_takeover(struct mddev *mddev)
3329{
3330	/* raid1 can take over:
3331	 *  raid5 with 2 devices, any layout or chunk size
3332	 */
3333	if (mddev->level == 5 && mddev->raid_disks == 2) {
3334		struct r1conf *conf;
3335		mddev->new_level = 1;
3336		mddev->new_layout = 0;
3337		mddev->new_chunk_sectors = 0;
3338		conf = setup_conf(mddev);
3339		if (!IS_ERR(conf)) {
3340			/* Array must appear to be quiesced */
3341			conf->array_frozen = 1;
3342			mddev_clear_unsupported_flags(mddev,
3343				UNSUPPORTED_MDDEV_FLAGS);
3344		}
3345		return conf;
3346	}
3347	return ERR_PTR(-EINVAL);
3348}
3349
3350static struct md_personality raid1_personality =
3351{
3352	.name		= "raid1",
3353	.level		= 1,
3354	.owner		= THIS_MODULE,
3355	.make_request	= raid1_make_request,
3356	.run		= raid1_run,
3357	.free		= raid1_free,
3358	.status		= raid1_status,
3359	.error_handler	= raid1_error,
3360	.hot_add_disk	= raid1_add_disk,
3361	.hot_remove_disk= raid1_remove_disk,
3362	.spare_active	= raid1_spare_active,
3363	.sync_request	= raid1_sync_request,
3364	.resize		= raid1_resize,
3365	.size		= raid1_size,
3366	.check_reshape	= raid1_reshape,
3367	.quiesce	= raid1_quiesce,
3368	.takeover	= raid1_takeover,
 
3369};
3370
3371static int __init raid_init(void)
3372{
3373	return register_md_personality(&raid1_personality);
3374}
3375
3376static void raid_exit(void)
3377{
3378	unregister_md_personality(&raid1_personality);
3379}
3380
3381module_init(raid_init);
3382module_exit(raid_exit);
3383MODULE_LICENSE("GPL");
3384MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3385MODULE_ALIAS("md-personality-3"); /* RAID1 */
3386MODULE_ALIAS("md-raid1");
3387MODULE_ALIAS("md-level-1");
3388
3389module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);