Linux Audio

Check our new training course

Loading...
v3.5.6
 
   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 there are this many requests queue to be written by
  50 * the raid1 thread, we become 'congested' to provide back-pressure
  51 * for writeback.
  52 */
  53static int max_queued_requests = 1024;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  54
  55static void allow_barrier(struct r1conf *conf);
  56static void lower_barrier(struct r1conf *conf);
 
 
 
 
 
 
  57
  58static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  59{
  60	struct pool_info *pi = data;
  61	int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  62
  63	/* allocate a r1bio with room for raid_disks entries in the bios array */
  64	return kzalloc(size, gfp_flags);
  65}
  66
  67static void r1bio_pool_free(void *r1_bio, void *data)
  68{
  69	kfree(r1_bio);
  70}
  71
  72#define RESYNC_BLOCK_SIZE (64*1024)
  73//#define RESYNC_BLOCK_SIZE PAGE_SIZE
  74#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  75#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  76#define RESYNC_WINDOW (2048*1024)
 
 
  77
  78static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  79{
  80	struct pool_info *pi = data;
  81	struct page *page;
  82	struct r1bio *r1_bio;
  83	struct bio *bio;
  84	int i, j;
 
 
  85
  86	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  87	if (!r1_bio)
  88		return NULL;
  89
 
 
 
 
 
  90	/*
  91	 * Allocate bios : 1 for reading, n-1 for writing
  92	 */
  93	for (j = pi->raid_disks ; j-- ; ) {
  94		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
  95		if (!bio)
  96			goto out_free_bio;
  97		r1_bio->bios[j] = bio;
  98	}
  99	/*
 100	 * Allocate RESYNC_PAGES data pages and attach them to
 101	 * the first bio.
 102	 * If this is a user-requested check/repair, allocate
 103	 * RESYNC_PAGES for each bio.
 104	 */
 105	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 106		j = pi->raid_disks;
 107	else
 108		j = 1;
 109	while(j--) {
 
 
 110		bio = r1_bio->bios[j];
 111		for (i = 0; i < RESYNC_PAGES; i++) {
 112			page = alloc_page(gfp_flags);
 113			if (unlikely(!page))
 114				goto out_free_pages;
 115
 116			bio->bi_io_vec[i].bv_page = page;
 117			bio->bi_vcnt = i+1;
 
 
 
 
 118		}
 119	}
 120	/* If not user-requests, copy the page pointers to all bios */
 121	if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 122		for (i=0; i<RESYNC_PAGES ; i++)
 123			for (j=1; j<pi->raid_disks; j++)
 124				r1_bio->bios[j]->bi_io_vec[i].bv_page =
 125					r1_bio->bios[0]->bi_io_vec[i].bv_page;
 126	}
 127
 128	r1_bio->master_bio = NULL;
 129
 130	return r1_bio;
 131
 132out_free_pages:
 133	for (j=0 ; j < pi->raid_disks; j++)
 134		for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 135			put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 136	j = -1;
 137out_free_bio:
 138	while (++j < pi->raid_disks)
 139		bio_put(r1_bio->bios[j]);
 140	r1bio_pool_free(r1_bio, data);
 
 
 
 141	return NULL;
 142}
 143
 144static void r1buf_pool_free(void *__r1_bio, void *data)
 145{
 146	struct pool_info *pi = data;
 147	int i,j;
 148	struct r1bio *r1bio = __r1_bio;
 
 149
 150	for (i = 0; i < RESYNC_PAGES; i++)
 151		for (j = pi->raid_disks; j-- ;) {
 152			if (j == 0 ||
 153			    r1bio->bios[j]->bi_io_vec[i].bv_page !=
 154			    r1bio->bios[0]->bi_io_vec[i].bv_page)
 155				safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 156		}
 157	for (i=0 ; i < pi->raid_disks; i++)
 158		bio_put(r1bio->bios[i]);
 
 159
 160	r1bio_pool_free(r1bio, data);
 
 
 
 161}
 162
 163static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 164{
 165	int i;
 166
 167	for (i = 0; i < conf->raid_disks * 2; i++) {
 168		struct bio **bio = r1_bio->bios + i;
 169		if (!BIO_SPECIAL(*bio))
 170			bio_put(*bio);
 171		*bio = NULL;
 172	}
 173}
 174
 175static void free_r1bio(struct r1bio *r1_bio)
 176{
 177	struct r1conf *conf = r1_bio->mddev->private;
 178
 179	put_all_bios(conf, r1_bio);
 180	mempool_free(r1_bio, conf->r1bio_pool);
 181}
 182
 183static void put_buf(struct r1bio *r1_bio)
 184{
 185	struct r1conf *conf = r1_bio->mddev->private;
 
 186	int i;
 187
 188	for (i = 0; i < conf->raid_disks * 2; i++) {
 189		struct bio *bio = r1_bio->bios[i];
 190		if (bio->bi_end_io)
 191			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 192	}
 193
 194	mempool_free(r1_bio, conf->r1buf_pool);
 195
 196	lower_barrier(conf);
 197}
 198
 199static void reschedule_retry(struct r1bio *r1_bio)
 200{
 201	unsigned long flags;
 202	struct mddev *mddev = r1_bio->mddev;
 203	struct r1conf *conf = mddev->private;
 
 204
 
 205	spin_lock_irqsave(&conf->device_lock, flags);
 206	list_add(&r1_bio->retry_list, &conf->retry_list);
 207	conf->nr_queued ++;
 208	spin_unlock_irqrestore(&conf->device_lock, flags);
 209
 210	wake_up(&conf->wait_barrier);
 211	md_wakeup_thread(mddev->thread);
 212}
 213
 214/*
 215 * raid_end_bio_io() is called when we have finished servicing a mirrored
 216 * operation and are ready to return a success/failure code to the buffer
 217 * cache layer.
 218 */
 219static void call_bio_endio(struct r1bio *r1_bio)
 220{
 221	struct bio *bio = r1_bio->master_bio;
 222	int done;
 223	struct r1conf *conf = r1_bio->mddev->private;
 224
 225	if (bio->bi_phys_segments) {
 226		unsigned long flags;
 227		spin_lock_irqsave(&conf->device_lock, flags);
 228		bio->bi_phys_segments--;
 229		done = (bio->bi_phys_segments == 0);
 230		spin_unlock_irqrestore(&conf->device_lock, flags);
 231	} else
 232		done = 1;
 233
 234	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 235		clear_bit(BIO_UPTODATE, &bio->bi_flags);
 236	if (done) {
 237		bio_endio(bio, 0);
 238		/*
 239		 * Wake up any possible resync thread that waits for the device
 240		 * to go idle.
 241		 */
 242		allow_barrier(conf);
 243	}
 244}
 245
 246static void raid_end_bio_io(struct r1bio *r1_bio)
 247{
 248	struct bio *bio = r1_bio->master_bio;
 249
 250	/* if nobody has done the final endio yet, do it now */
 251	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 252		pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 253			 (bio_data_dir(bio) == WRITE) ? "write" : "read",
 254			 (unsigned long long) bio->bi_sector,
 255			 (unsigned long long) bio->bi_sector +
 256			 (bio->bi_size >> 9) - 1);
 257
 258		call_bio_endio(r1_bio);
 259	}
 260	free_r1bio(r1_bio);
 261}
 262
 263/*
 264 * Update disk head position estimator based on IRQ completion info.
 265 */
 266static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 267{
 268	struct r1conf *conf = r1_bio->mddev->private;
 269
 270	conf->mirrors[disk].head_position =
 271		r1_bio->sector + (r1_bio->sectors);
 272}
 273
 274/*
 275 * Find the disk number which triggered given bio
 276 */
 277static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 278{
 279	int mirror;
 280	struct r1conf *conf = r1_bio->mddev->private;
 281	int raid_disks = conf->raid_disks;
 282
 283	for (mirror = 0; mirror < raid_disks * 2; mirror++)
 284		if (r1_bio->bios[mirror] == bio)
 285			break;
 286
 287	BUG_ON(mirror == raid_disks * 2);
 288	update_head_pos(mirror, r1_bio);
 289
 290	return mirror;
 291}
 292
 293static void raid1_end_read_request(struct bio *bio, int error)
 294{
 295	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 296	struct r1bio *r1_bio = bio->bi_private;
 297	int mirror;
 298	struct r1conf *conf = r1_bio->mddev->private;
 
 299
 300	mirror = r1_bio->read_disk;
 301	/*
 302	 * this branch is our 'one mirror IO has finished' event handler:
 303	 */
 304	update_head_pos(mirror, r1_bio);
 305
 306	if (uptodate)
 307		set_bit(R1BIO_Uptodate, &r1_bio->state);
 
 
 
 
 
 308	else {
 309		/* If all other devices have failed, we want to return
 310		 * the error upwards rather than fail the last device.
 311		 * Here we redefine "uptodate" to mean "Don't want to retry"
 312		 */
 313		unsigned long flags;
 314		spin_lock_irqsave(&conf->device_lock, flags);
 315		if (r1_bio->mddev->degraded == conf->raid_disks ||
 316		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 317		     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 318			uptodate = 1;
 319		spin_unlock_irqrestore(&conf->device_lock, flags);
 320	}
 321
 322	if (uptodate)
 323		raid_end_bio_io(r1_bio);
 324	else {
 
 325		/*
 326		 * oops, read error:
 327		 */
 328		char b[BDEVNAME_SIZE];
 329		printk_ratelimited(
 330			KERN_ERR "md/raid1:%s: %s: "
 331			"rescheduling sector %llu\n",
 332			mdname(conf->mddev),
 333			bdevname(conf->mirrors[mirror].rdev->bdev,
 334				 b),
 335			(unsigned long long)r1_bio->sector);
 336		set_bit(R1BIO_ReadError, &r1_bio->state);
 337		reschedule_retry(r1_bio);
 
 338	}
 339
 340	rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 341}
 342
 343static void close_write(struct r1bio *r1_bio)
 344{
 345	/* it really is the end of this request */
 346	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 347		/* free extra copy of the data pages */
 348		int i = r1_bio->behind_page_count;
 349		while (i--)
 350			safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 351		kfree(r1_bio->behind_bvecs);
 352		r1_bio->behind_bvecs = NULL;
 353	}
 354	/* clear the bitmap if all writes complete successfully */
 355	bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 356			r1_bio->sectors,
 357			!test_bit(R1BIO_Degraded, &r1_bio->state),
 358			test_bit(R1BIO_BehindIO, &r1_bio->state));
 359	md_write_end(r1_bio->mddev);
 360}
 361
 362static void r1_bio_write_done(struct r1bio *r1_bio)
 363{
 364	if (!atomic_dec_and_test(&r1_bio->remaining))
 365		return;
 366
 367	if (test_bit(R1BIO_WriteError, &r1_bio->state))
 368		reschedule_retry(r1_bio);
 369	else {
 370		close_write(r1_bio);
 371		if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 372			reschedule_retry(r1_bio);
 373		else
 374			raid_end_bio_io(r1_bio);
 375	}
 376}
 377
 378static void raid1_end_write_request(struct bio *bio, int error)
 379{
 380	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 381	struct r1bio *r1_bio = bio->bi_private;
 382	int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 383	struct r1conf *conf = r1_bio->mddev->private;
 384	struct bio *to_put = NULL;
 
 
 
 385
 386	mirror = find_bio_disk(r1_bio, bio);
 387
 388	/*
 389	 * 'one mirror IO has finished' event handler:
 390	 */
 391	if (!uptodate) {
 392		set_bit(WriteErrorSeen,
 393			&conf->mirrors[mirror].rdev->flags);
 394		if (!test_and_set_bit(WantReplacement,
 395				      &conf->mirrors[mirror].rdev->flags))
 396			set_bit(MD_RECOVERY_NEEDED, &
 397				conf->mddev->recovery);
 398
 399		set_bit(R1BIO_WriteError, &r1_bio->state);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 400	} else {
 401		/*
 402		 * Set R1BIO_Uptodate in our master bio, so that we
 403		 * will return a good error code for to the higher
 404		 * levels even if IO on some other mirrored buffer
 405		 * fails.
 406		 *
 407		 * The 'master' represents the composite IO operation
 408		 * to user-side. So if something waits for IO, then it
 409		 * will wait for the 'master' bio.
 410		 */
 411		sector_t first_bad;
 412		int bad_sectors;
 413
 414		r1_bio->bios[mirror] = NULL;
 415		to_put = bio;
 416		set_bit(R1BIO_Uptodate, &r1_bio->state);
 
 
 
 
 
 
 
 
 
 
 417
 418		/* Maybe we can clear some bad blocks. */
 419		if (is_badblock(conf->mirrors[mirror].rdev,
 420				r1_bio->sector, r1_bio->sectors,
 421				&first_bad, &bad_sectors)) {
 422			r1_bio->bios[mirror] = IO_MADE_GOOD;
 423			set_bit(R1BIO_MadeGood, &r1_bio->state);
 424		}
 425	}
 426
 427	if (behind) {
 428		if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 
 
 
 
 
 
 429			atomic_dec(&r1_bio->behind_remaining);
 430
 431		/*
 432		 * In behind mode, we ACK the master bio once the I/O
 433		 * has safely reached all non-writemostly
 434		 * disks. Setting the Returned bit ensures that this
 435		 * gets done only once -- we don't ever want to return
 436		 * -EIO here, instead we'll wait
 437		 */
 438		if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 439		    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 440			/* Maybe we can return now */
 441			if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 442				struct bio *mbio = r1_bio->master_bio;
 443				pr_debug("raid1: behind end write sectors"
 444					 " %llu-%llu\n",
 445					 (unsigned long long) mbio->bi_sector,
 446					 (unsigned long long) mbio->bi_sector +
 447					 (mbio->bi_size >> 9) - 1);
 448				call_bio_endio(r1_bio);
 449			}
 450		}
 451	}
 452	if (r1_bio->bios[mirror] == NULL)
 453		rdev_dec_pending(conf->mirrors[mirror].rdev,
 454				 conf->mddev);
 455
 456	/*
 457	 * Let's see if all mirrored write operations have finished
 458	 * already.
 459	 */
 460	r1_bio_write_done(r1_bio);
 461
 462	if (to_put)
 463		bio_put(to_put);
 464}
 465
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 466
 467/*
 468 * This routine returns the disk from which the requested read should
 469 * be done. There is a per-array 'next expected sequential IO' sector
 470 * number - if this matches on the next IO then we use the last disk.
 471 * There is also a per-disk 'last know head position' sector that is
 472 * maintained from IRQ contexts, both the normal and the resync IO
 473 * completion handlers update this position correctly. If there is no
 474 * perfect sequential match then we pick the disk whose head is closest.
 475 *
 476 * If there are 2 mirrors in the same 2 devices, performance degrades
 477 * because position is mirror, not device based.
 478 *
 479 * The rdev for the device selected will have nr_pending incremented.
 480 */
 481static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 482{
 483	const sector_t this_sector = r1_bio->sector;
 484	int sectors;
 485	int best_good_sectors;
 486	int start_disk;
 487	int best_disk;
 488	int i;
 489	sector_t best_dist;
 
 490	struct md_rdev *rdev;
 491	int choose_first;
 
 492
 493	rcu_read_lock();
 494	/*
 495	 * Check if we can balance. We can balance on the whole
 496	 * device if no resync is going on, or below the resync window.
 497	 * We take the first readable disk when above the resync window.
 498	 */
 499 retry:
 500	sectors = r1_bio->sectors;
 501	best_disk = -1;
 
 502	best_dist = MaxSector;
 
 
 503	best_good_sectors = 0;
 504
 505	if (conf->mddev->recovery_cp < MaxSector &&
 506	    (this_sector + sectors >= conf->next_resync)) {
 
 
 
 
 
 507		choose_first = 1;
 508		start_disk = 0;
 509	} else {
 510		choose_first = 0;
 511		start_disk = conf->last_used;
 512	}
 513
 514	for (i = 0 ; i < conf->raid_disks * 2 ; i++) {
 515		sector_t dist;
 516		sector_t first_bad;
 517		int bad_sectors;
 518
 519		int disk = start_disk + i;
 520		if (disk >= conf->raid_disks * 2)
 521			disk -= conf->raid_disks * 2;
 522
 523		rdev = rcu_dereference(conf->mirrors[disk].rdev);
 524		if (r1_bio->bios[disk] == IO_BLOCKED
 525		    || rdev == NULL
 526		    || test_bit(Unmerged, &rdev->flags)
 527		    || test_bit(Faulty, &rdev->flags))
 528			continue;
 529		if (!test_bit(In_sync, &rdev->flags) &&
 530		    rdev->recovery_offset < this_sector + sectors)
 531			continue;
 532		if (test_bit(WriteMostly, &rdev->flags)) {
 533			/* Don't balance among write-mostly, just
 534			 * use the first as a last resort */
 535			if (best_disk < 0) {
 536				if (is_badblock(rdev, this_sector, sectors,
 537						&first_bad, &bad_sectors)) {
 538					if (first_bad < this_sector)
 539						/* Cannot use this */
 540						continue;
 541					best_good_sectors = first_bad - this_sector;
 542				} else
 543					best_good_sectors = sectors;
 544				best_disk = disk;
 
 545			}
 546			continue;
 547		}
 548		/* This is a reasonable device to use.  It might
 549		 * even be best.
 550		 */
 551		if (is_badblock(rdev, this_sector, sectors,
 552				&first_bad, &bad_sectors)) {
 553			if (best_dist < MaxSector)
 554				/* already have a better device */
 555				continue;
 556			if (first_bad <= this_sector) {
 557				/* cannot read here. If this is the 'primary'
 558				 * device, then we must not read beyond
 559				 * bad_sectors from another device..
 560				 */
 561				bad_sectors -= (this_sector - first_bad);
 562				if (choose_first && sectors > bad_sectors)
 563					sectors = bad_sectors;
 564				if (best_good_sectors > sectors)
 565					best_good_sectors = sectors;
 566
 567			} else {
 568				sector_t good_sectors = first_bad - this_sector;
 569				if (good_sectors > best_good_sectors) {
 570					best_good_sectors = good_sectors;
 571					best_disk = disk;
 572				}
 573				if (choose_first)
 574					break;
 575			}
 576			continue;
 577		} else
 
 
 578			best_good_sectors = sectors;
 
 579
 
 
 
 
 
 
 
 580		dist = abs(this_sector - conf->mirrors[disk].head_position);
 581		if (choose_first
 582		    /* Don't change to another disk for sequential reads */
 583		    || conf->next_seq_sect == this_sector
 584		    || dist == 0
 585		    /* If device is idle, use it */
 586		    || atomic_read(&rdev->nr_pending) == 0) {
 
 
 
 
 587			best_disk = disk;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 588			break;
 589		}
 
 
 
 
 
 
 
 
 
 590		if (dist < best_dist) {
 591			best_dist = dist;
 592			best_disk = disk;
 593		}
 594	}
 595
 
 
 
 
 
 
 
 
 
 
 
 
 
 596	if (best_disk >= 0) {
 597		rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 598		if (!rdev)
 599			goto retry;
 600		atomic_inc(&rdev->nr_pending);
 601		if (test_bit(Faulty, &rdev->flags)) {
 602			/* cannot risk returning a device that failed
 603			 * before we inc'ed nr_pending
 604			 */
 605			rdev_dec_pending(rdev, conf->mddev);
 606			goto retry;
 607		}
 608		sectors = best_good_sectors;
 609		conf->next_seq_sect = this_sector + sectors;
 610		conf->last_used = best_disk;
 
 
 
 611	}
 612	rcu_read_unlock();
 613	*max_sectors = sectors;
 614
 615	return best_disk;
 616}
 617
 618static int raid1_mergeable_bvec(struct request_queue *q,
 619				struct bvec_merge_data *bvm,
 620				struct bio_vec *biovec)
 621{
 622	struct mddev *mddev = q->queuedata;
 623	struct r1conf *conf = mddev->private;
 624	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 625	int max = biovec->bv_len;
 626
 627	if (mddev->merge_check_needed) {
 628		int disk;
 629		rcu_read_lock();
 630		for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 631			struct md_rdev *rdev = rcu_dereference(
 632				conf->mirrors[disk].rdev);
 633			if (rdev && !test_bit(Faulty, &rdev->flags)) {
 634				struct request_queue *q =
 635					bdev_get_queue(rdev->bdev);
 636				if (q->merge_bvec_fn) {
 637					bvm->bi_sector = sector +
 638						rdev->data_offset;
 639					bvm->bi_bdev = rdev->bdev;
 640					max = min(max, q->merge_bvec_fn(
 641							  q, bvm, biovec));
 642				}
 643			}
 644		}
 645		rcu_read_unlock();
 646	}
 647	return max;
 648
 649}
 650
 651int md_raid1_congested(struct mddev *mddev, int bits)
 652{
 653	struct r1conf *conf = mddev->private;
 654	int i, ret = 0;
 655
 656	if ((bits & (1 << BDI_async_congested)) &&
 657	    conf->pending_count >= max_queued_requests)
 658		return 1;
 659
 660	rcu_read_lock();
 661	for (i = 0; i < conf->raid_disks * 2; i++) {
 662		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 663		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 664			struct request_queue *q = bdev_get_queue(rdev->bdev);
 665
 666			BUG_ON(!q);
 667
 668			/* Note the '|| 1' - when read_balance prefers
 669			 * non-congested targets, it can be removed
 670			 */
 671			if ((bits & (1<<BDI_async_congested)) || 1)
 672				ret |= bdi_congested(&q->backing_dev_info, bits);
 673			else
 674				ret &= bdi_congested(&q->backing_dev_info, bits);
 675		}
 676	}
 677	rcu_read_unlock();
 678	return ret;
 679}
 680EXPORT_SYMBOL_GPL(md_raid1_congested);
 681
 682static int raid1_congested(void *data, int bits)
 683{
 684	struct mddev *mddev = data;
 
 
 685
 686	return mddev_congested(mddev, bits) ||
 687		md_raid1_congested(mddev, bits);
 
 
 
 
 
 
 
 
 
 
 
 
 
 688}
 689
 690static void flush_pending_writes(struct r1conf *conf)
 691{
 692	/* Any writes that have been queued but are awaiting
 693	 * bitmap updates get flushed here.
 694	 */
 695	spin_lock_irq(&conf->device_lock);
 696
 697	if (conf->pending_bio_list.head) {
 
 698		struct bio *bio;
 
 699		bio = bio_list_get(&conf->pending_bio_list);
 700		conf->pending_count = 0;
 701		spin_unlock_irq(&conf->device_lock);
 702		/* flush any pending bitmap writes to
 703		 * disk before proceeding w/ I/O */
 704		bitmap_unplug(conf->mddev->bitmap);
 705		wake_up(&conf->wait_barrier);
 706
 707		while (bio) { /* submit pending writes */
 708			struct bio *next = bio->bi_next;
 709			bio->bi_next = NULL;
 710			generic_make_request(bio);
 711			bio = next;
 712		}
 
 
 
 
 
 
 
 713	} else
 714		spin_unlock_irq(&conf->device_lock);
 715}
 716
 717/* Barriers....
 718 * Sometimes we need to suspend IO while we do something else,
 719 * either some resync/recovery, or reconfigure the array.
 720 * To do this we raise a 'barrier'.
 721 * The 'barrier' is a counter that can be raised multiple times
 722 * to count how many activities are happening which preclude
 723 * normal IO.
 724 * We can only raise the barrier if there is no pending IO.
 725 * i.e. if nr_pending == 0.
 726 * We choose only to raise the barrier if no-one is waiting for the
 727 * barrier to go down.  This means that as soon as an IO request
 728 * is ready, no other operations which require a barrier will start
 729 * until the IO request has had a chance.
 730 *
 731 * So: regular IO calls 'wait_barrier'.  When that returns there
 732 *    is no backgroup IO happening,  It must arrange to call
 733 *    allow_barrier when it has finished its IO.
 734 * backgroup IO calls must call raise_barrier.  Once that returns
 735 *    there is no normal IO happeing.  It must arrange to call
 736 *    lower_barrier when the particular background IO completes.
 
 
 
 737 */
 738#define RESYNC_DEPTH 32
 739
 740static void raise_barrier(struct r1conf *conf)
 741{
 
 
 742	spin_lock_irq(&conf->resync_lock);
 743
 744	/* Wait until no block IO is waiting */
 745	wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 746			    conf->resync_lock, );
 
 747
 748	/* block any new IO from starting */
 749	conf->barrier++;
 750
 751	/* Now wait for all pending IO to complete */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 752	wait_event_lock_irq(conf->wait_barrier,
 753			    !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 754			    conf->resync_lock, );
 
 
 
 
 
 
 
 
 
 
 755
 
 756	spin_unlock_irq(&conf->resync_lock);
 
 
 757}
 758
 759static void lower_barrier(struct r1conf *conf)
 760{
 761	unsigned long flags;
 762	BUG_ON(conf->barrier <= 0);
 763	spin_lock_irqsave(&conf->resync_lock, flags);
 764	conf->barrier--;
 765	spin_unlock_irqrestore(&conf->resync_lock, flags);
 
 766	wake_up(&conf->wait_barrier);
 767}
 768
 769static void wait_barrier(struct r1conf *conf)
 770{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 771	spin_lock_irq(&conf->resync_lock);
 772	if (conf->barrier) {
 773		conf->nr_waiting++;
 774		/* Wait for the barrier to drop.
 775		 * However if there are already pending
 776		 * requests (preventing the barrier from
 777		 * rising completely), and the
 778		 * pre-process bio queue isn't empty,
 779		 * then don't wait, as we need to empty
 780		 * that queue to get the nr_pending
 781		 * count down.
 782		 */
 783		wait_event_lock_irq(conf->wait_barrier,
 784				    !conf->barrier ||
 785				    (conf->nr_pending &&
 786				     current->bio_list &&
 787				     !bio_list_empty(current->bio_list)),
 788				    conf->resync_lock,
 789			);
 790		conf->nr_waiting--;
 791	}
 792	conf->nr_pending++;
 793	spin_unlock_irq(&conf->resync_lock);
 794}
 795
 796static void allow_barrier(struct r1conf *conf)
 797{
 798	unsigned long flags;
 799	spin_lock_irqsave(&conf->resync_lock, flags);
 800	conf->nr_pending--;
 801	spin_unlock_irqrestore(&conf->resync_lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 802	wake_up(&conf->wait_barrier);
 
 
 
 
 
 
 
 803}
 804
 805static void freeze_array(struct r1conf *conf)
 806{
 807	/* stop syncio and normal IO and wait for everything to
 808	 * go quite.
 809	 * We increment barrier and nr_waiting, and then
 810	 * wait until nr_pending match nr_queued+1
 811	 * This is called in the context of one normal IO request
 812	 * that has failed. Thus any sync request that might be pending
 813	 * will be blocked by nr_pending, and we need to wait for
 814	 * pending IO requests to complete or be queued for re-try.
 815	 * Thus the number queued (nr_queued) plus this request (1)
 816	 * must match the number of pending IOs (nr_pending) before
 817	 * we continue.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 818	 */
 819	spin_lock_irq(&conf->resync_lock);
 820	conf->barrier++;
 821	conf->nr_waiting++;
 822	wait_event_lock_irq(conf->wait_barrier,
 823			    conf->nr_pending == conf->nr_queued+1,
 824			    conf->resync_lock,
 825			    flush_pending_writes(conf));
 
 826	spin_unlock_irq(&conf->resync_lock);
 827}
 828static void unfreeze_array(struct r1conf *conf)
 829{
 830	/* reverse the effect of the freeze */
 831	spin_lock_irq(&conf->resync_lock);
 832	conf->barrier--;
 833	conf->nr_waiting--;
 834	wake_up(&conf->wait_barrier);
 835	spin_unlock_irq(&conf->resync_lock);
 
 836}
 837
 838
 839/* duplicate the data pages for behind I/O 
 840 */
 841static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 842{
 843	int i;
 844	struct bio_vec *bvec;
 845	struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 846					GFP_NOIO);
 847	if (unlikely(!bvecs))
 
 
 848		return;
 849
 850	bio_for_each_segment(bvec, bio, i) {
 851		bvecs[i] = *bvec;
 852		bvecs[i].bv_page = alloc_page(GFP_NOIO);
 853		if (unlikely(!bvecs[i].bv_page))
 854			goto do_sync_io;
 855		memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 856		       kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 857		kunmap(bvecs[i].bv_page);
 858		kunmap(bvec->bv_page);
 
 
 
 
 
 
 
 
 
 
 
 859	}
 860	r1_bio->behind_bvecs = bvecs;
 861	r1_bio->behind_page_count = bio->bi_vcnt;
 
 
 862	set_bit(R1BIO_BehindIO, &r1_bio->state);
 
 863	return;
 864
 865do_sync_io:
 866	for (i = 0; i < bio->bi_vcnt; i++)
 867		if (bvecs[i].bv_page)
 868			put_page(bvecs[i].bv_page);
 869	kfree(bvecs);
 870	pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 871}
 872
 873static void make_request(struct mddev *mddev, struct bio * bio)
 
 874{
 875	struct r1conf *conf = mddev->private;
 876	struct mirror_info *mirror;
 877	struct r1bio *r1_bio;
 
 
 
 
 
 
 
 
 
 
 
 
 
 878	struct bio *read_bio;
 879	int i, disks;
 880	struct bitmap *bitmap;
 881	unsigned long flags;
 882	const int rw = bio_data_dir(bio);
 883	const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 884	const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 885	struct md_rdev *blocked_rdev;
 886	int first_clone;
 887	int sectors_handled;
 888	int max_sectors;
 
 
 
 889
 890	/*
 891	 * Register the new request and wait if the reconstruction
 892	 * thread has put up a bar for new requests.
 893	 * Continue immediately if no resync is active currently.
 894	 */
 
 895
 896	md_write_start(mddev, bio); /* wait on superblock update early */
 897
 898	if (bio_data_dir(bio) == WRITE &&
 899	    bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 900	    bio->bi_sector < mddev->suspend_hi) {
 901		/* As the suspend_* range is controlled by
 902		 * userspace, we want an interruptible
 903		 * wait.
 904		 */
 905		DEFINE_WAIT(w);
 906		for (;;) {
 907			flush_signals(current);
 908			prepare_to_wait(&conf->wait_barrier,
 909					&w, TASK_INTERRUPTIBLE);
 910			if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 911			    bio->bi_sector >= mddev->suspend_hi)
 912				break;
 913			schedule();
 914		}
 915		finish_wait(&conf->wait_barrier, &w);
 916	}
 917
 918	wait_barrier(conf);
 
 
 
 
 919
 920	bitmap = mddev->bitmap;
 
 
 
 
 921
 922	/*
 923	 * make_request() can abort the operation when READA is being
 924	 * used and no empty request is available.
 925	 *
 926	 */
 927	r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 928
 929	r1_bio->master_bio = bio;
 930	r1_bio->sectors = bio->bi_size >> 9;
 931	r1_bio->state = 0;
 932	r1_bio->mddev = mddev;
 933	r1_bio->sector = bio->bi_sector;
 
 
 
 
 
 
 
 934
 935	/* We might need to issue multiple reads to different
 936	 * devices if there are bad blocks around, so we keep
 937	 * track of the number of reads in bio->bi_phys_segments.
 938	 * If this is 0, there is only one r1_bio and no locking
 939	 * will be needed when requests complete.  If it is
 940	 * non-zero, then it is the number of not-completed requests.
 941	 */
 942	bio->bi_phys_segments = 0;
 943	clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 944
 945	if (rw == READ) {
 
 946		/*
 947		 * read balancing logic:
 
 948		 */
 949		int rdisk;
 
 
 
 950
 951read_again:
 952		rdisk = read_balance(conf, r1_bio, &max_sectors);
 
 
 
 
 
 
 
 953
 954		if (rdisk < 0) {
 955			/* couldn't find anywhere to read from */
 956			raid_end_bio_io(r1_bio);
 957			return;
 958		}
 959		mirror = conf->mirrors + rdisk;
 960
 961		if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 962		    bitmap) {
 963			/* Reading from a write-mostly device must
 964			 * take care not to over-take any writes
 965			 * that are 'behind'
 966			 */
 967			wait_event(bitmap->behind_wait,
 968				   atomic_read(&bitmap->behind_writes) == 0);
 969		}
 970		r1_bio->read_disk = rdisk;
 971
 972		read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 973		md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
 974			    max_sectors);
 975
 976		r1_bio->bios[rdisk] = read_bio;
 977
 978		read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 979		read_bio->bi_bdev = mirror->rdev->bdev;
 980		read_bio->bi_end_io = raid1_end_read_request;
 981		read_bio->bi_rw = READ | do_sync;
 982		read_bio->bi_private = r1_bio;
 983
 984		if (max_sectors < r1_bio->sectors) {
 985			/* could not read all from this device, so we will
 986			 * need another r1_bio.
 987			 */
 988
 989			sectors_handled = (r1_bio->sector + max_sectors
 990					   - bio->bi_sector);
 991			r1_bio->sectors = max_sectors;
 992			spin_lock_irq(&conf->device_lock);
 993			if (bio->bi_phys_segments == 0)
 994				bio->bi_phys_segments = 2;
 995			else
 996				bio->bi_phys_segments++;
 997			spin_unlock_irq(&conf->device_lock);
 998			/* Cannot call generic_make_request directly
 999			 * as that will be queued in __make_request
1000			 * and subsequent mempool_alloc might block waiting
1001			 * for it.  So hand bio over to raid1d.
1002			 */
1003			reschedule_retry(r1_bio);
1004
1005			r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 
 
1006
1007			r1_bio->master_bio = bio;
1008			r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1009			r1_bio->state = 0;
1010			r1_bio->mddev = mddev;
1011			r1_bio->sector = bio->bi_sector + sectors_handled;
1012			goto read_again;
1013		} else
1014			generic_make_request(read_bio);
1015		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1016	}
1017
1018	/*
1019	 * WRITE:
 
 
1020	 */
 
 
 
 
 
1021	if (conf->pending_count >= max_queued_requests) {
1022		md_wakeup_thread(mddev->thread);
 
1023		wait_event(conf->wait_barrier,
1024			   conf->pending_count < max_queued_requests);
1025	}
1026	/* first select target devices under rcu_lock and
1027	 * inc refcount on their rdev.  Record them by setting
1028	 * bios[x] to bio
1029	 * If there are known/acknowledged bad blocks on any device on
1030	 * which we have seen a write error, we want to avoid writing those
1031	 * blocks.
1032	 * This potentially requires several writes to write around
1033	 * the bad blocks.  Each set of writes gets it's own r1bio
1034	 * with a set of bios attached.
1035	 */
1036
1037	disks = conf->raid_disks * 2;
1038 retry_write:
1039	blocked_rdev = NULL;
1040	rcu_read_lock();
1041	max_sectors = r1_bio->sectors;
1042	for (i = 0;  i < disks; i++) {
1043		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1044		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1045			atomic_inc(&rdev->nr_pending);
1046			blocked_rdev = rdev;
1047			break;
1048		}
1049		r1_bio->bios[i] = NULL;
1050		if (!rdev || test_bit(Faulty, &rdev->flags)
1051		    || test_bit(Unmerged, &rdev->flags)) {
1052			if (i < conf->raid_disks)
1053				set_bit(R1BIO_Degraded, &r1_bio->state);
1054			continue;
1055		}
1056
1057		atomic_inc(&rdev->nr_pending);
1058		if (test_bit(WriteErrorSeen, &rdev->flags)) {
1059			sector_t first_bad;
1060			int bad_sectors;
1061			int is_bad;
1062
1063			is_bad = is_badblock(rdev, r1_bio->sector,
1064					     max_sectors,
1065					     &first_bad, &bad_sectors);
1066			if (is_bad < 0) {
1067				/* mustn't write here until the bad block is
1068				 * acknowledged*/
1069				set_bit(BlockedBadBlocks, &rdev->flags);
1070				blocked_rdev = rdev;
1071				break;
1072			}
1073			if (is_bad && first_bad <= r1_bio->sector) {
1074				/* Cannot write here at all */
1075				bad_sectors -= (r1_bio->sector - first_bad);
1076				if (bad_sectors < max_sectors)
1077					/* mustn't write more than bad_sectors
1078					 * to other devices yet
1079					 */
1080					max_sectors = bad_sectors;
1081				rdev_dec_pending(rdev, mddev);
1082				/* We don't set R1BIO_Degraded as that
1083				 * only applies if the disk is
1084				 * missing, so it might be re-added,
1085				 * and we want to know to recover this
1086				 * chunk.
1087				 * In this case the device is here,
1088				 * and the fact that this chunk is not
1089				 * in-sync is recorded in the bad
1090				 * block log
1091				 */
1092				continue;
1093			}
1094			if (is_bad) {
1095				int good_sectors = first_bad - r1_bio->sector;
1096				if (good_sectors < max_sectors)
1097					max_sectors = good_sectors;
1098			}
1099		}
1100		r1_bio->bios[i] = bio;
1101	}
1102	rcu_read_unlock();
1103
1104	if (unlikely(blocked_rdev)) {
1105		/* Wait for this device to become unblocked */
1106		int j;
1107
1108		for (j = 0; j < i; j++)
1109			if (r1_bio->bios[j])
1110				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1111		r1_bio->state = 0;
1112		allow_barrier(conf);
 
1113		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1114		wait_barrier(conf);
1115		goto retry_write;
1116	}
1117
1118	if (max_sectors < r1_bio->sectors) {
1119		/* We are splitting this write into multiple parts, so
1120		 * we need to prepare for allocating another r1_bio.
1121		 */
 
 
 
1122		r1_bio->sectors = max_sectors;
1123		spin_lock_irq(&conf->device_lock);
1124		if (bio->bi_phys_segments == 0)
1125			bio->bi_phys_segments = 2;
1126		else
1127			bio->bi_phys_segments++;
1128		spin_unlock_irq(&conf->device_lock);
1129	}
1130	sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1131
1132	atomic_set(&r1_bio->remaining, 1);
1133	atomic_set(&r1_bio->behind_remaining, 0);
1134
1135	first_clone = 1;
 
1136	for (i = 0; i < disks; i++) {
1137		struct bio *mbio;
1138		if (!r1_bio->bios[i])
1139			continue;
1140
1141		mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1142		md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1143
1144		if (first_clone) {
1145			/* do behind I/O ?
1146			 * Not if there are too many, or cannot
1147			 * allocate memory, or a reader on WriteMostly
1148			 * is waiting for behind writes to flush */
1149			if (bitmap &&
1150			    (atomic_read(&bitmap->behind_writes)
1151			     < mddev->bitmap_info.max_write_behind) &&
1152			    !waitqueue_active(&bitmap->behind_wait))
1153				alloc_behind_pages(mbio, r1_bio);
 
1154
1155			bitmap_startwrite(bitmap, r1_bio->sector,
1156					  r1_bio->sectors,
1157					  test_bit(R1BIO_BehindIO,
1158						   &r1_bio->state));
1159			first_clone = 0;
1160		}
1161		if (r1_bio->behind_bvecs) {
1162			struct bio_vec *bvec;
1163			int j;
1164
1165			/* Yes, I really want the '__' version so that
1166			 * we clear any unused pointer in the io_vec, rather
1167			 * than leave them unchanged.  This is important
1168			 * because when we come to free the pages, we won't
1169			 * know the original bi_idx, so we just free
1170			 * them all
1171			 */
1172			__bio_for_each_segment(bvec, mbio, j, 0)
1173				bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1174			if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 
 
 
 
1175				atomic_inc(&r1_bio->behind_remaining);
1176		}
1177
1178		r1_bio->bios[i] = mbio;
1179
1180		mbio->bi_sector	= (r1_bio->sector +
1181				   conf->mirrors[i].rdev->data_offset);
1182		mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1183		mbio->bi_end_io	= raid1_end_write_request;
1184		mbio->bi_rw = WRITE | do_flush_fua | do_sync;
 
 
 
 
1185		mbio->bi_private = r1_bio;
1186
1187		atomic_inc(&r1_bio->remaining);
1188		spin_lock_irqsave(&conf->device_lock, flags);
1189		bio_list_add(&conf->pending_bio_list, mbio);
1190		conf->pending_count++;
1191		spin_unlock_irqrestore(&conf->device_lock, flags);
1192		if (!mddev_check_plugged(mddev))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1193			md_wakeup_thread(mddev->thread);
1194	}
1195	/* Mustn't call r1_bio_write_done before this next test,
1196	 * as it could result in the bio being freed.
1197	 */
1198	if (sectors_handled < (bio->bi_size >> 9)) {
1199		r1_bio_write_done(r1_bio);
1200		/* We need another r1_bio.  It has already been counted
1201		 * in bio->bi_phys_segments
1202		 */
1203		r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1204		r1_bio->master_bio = bio;
1205		r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1206		r1_bio->state = 0;
1207		r1_bio->mddev = mddev;
1208		r1_bio->sector = bio->bi_sector + sectors_handled;
1209		goto retry_write;
1210	}
1211
1212	r1_bio_write_done(r1_bio);
1213
1214	/* In case raid1d snuck in to freeze_array */
1215	wake_up(&conf->wait_barrier);
1216}
1217
1218static void status(struct seq_file *seq, struct mddev *mddev)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1219{
1220	struct r1conf *conf = mddev->private;
1221	int i;
1222
1223	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1224		   conf->raid_disks - mddev->degraded);
1225	rcu_read_lock();
1226	for (i = 0; i < conf->raid_disks; i++) {
1227		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1228		seq_printf(seq, "%s",
1229			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1230	}
1231	rcu_read_unlock();
1232	seq_printf(seq, "]");
1233}
1234
1235
1236static void error(struct mddev *mddev, struct md_rdev *rdev)
1237{
1238	char b[BDEVNAME_SIZE];
1239	struct r1conf *conf = mddev->private;
 
1240
1241	/*
1242	 * If it is not operational, then we have already marked it as dead
1243	 * else if it is the last working disks, ignore the error, let the
1244	 * next level up know.
1245	 * else mark the drive as failed
1246	 */
1247	if (test_bit(In_sync, &rdev->flags)
 
1248	    && (conf->raid_disks - mddev->degraded) == 1) {
1249		/*
1250		 * Don't fail the drive, act as though we were just a
1251		 * normal single drive.
1252		 * However don't try a recovery from this drive as
1253		 * it is very likely to fail.
1254		 */
1255		conf->recovery_disabled = mddev->recovery_disabled;
 
1256		return;
1257	}
1258	set_bit(Blocked, &rdev->flags);
1259	if (test_and_clear_bit(In_sync, &rdev->flags)) {
1260		unsigned long flags;
1261		spin_lock_irqsave(&conf->device_lock, flags);
1262		mddev->degraded++;
1263		set_bit(Faulty, &rdev->flags);
1264		spin_unlock_irqrestore(&conf->device_lock, flags);
1265		/*
1266		 * if recovery is running, make sure it aborts.
1267		 */
1268		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1269	} else
1270		set_bit(Faulty, &rdev->flags);
1271	set_bit(MD_CHANGE_DEVS, &mddev->flags);
1272	printk(KERN_ALERT
1273	       "md/raid1:%s: Disk failure on %s, disabling device.\n"
1274	       "md/raid1:%s: Operation continuing on %d devices.\n",
1275	       mdname(mddev), bdevname(rdev->bdev, b),
1276	       mdname(mddev), conf->raid_disks - mddev->degraded);
1277}
1278
1279static void print_conf(struct r1conf *conf)
1280{
1281	int i;
1282
1283	printk(KERN_DEBUG "RAID1 conf printout:\n");
1284	if (!conf) {
1285		printk(KERN_DEBUG "(!conf)\n");
1286		return;
1287	}
1288	printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1289		conf->raid_disks);
1290
1291	rcu_read_lock();
1292	for (i = 0; i < conf->raid_disks; i++) {
1293		char b[BDEVNAME_SIZE];
1294		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1295		if (rdev)
1296			printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1297			       i, !test_bit(In_sync, &rdev->flags),
1298			       !test_bit(Faulty, &rdev->flags),
1299			       bdevname(rdev->bdev,b));
1300	}
1301	rcu_read_unlock();
1302}
1303
1304static void close_sync(struct r1conf *conf)
1305{
1306	wait_barrier(conf);
1307	allow_barrier(conf);
 
 
 
 
1308
1309	mempool_destroy(conf->r1buf_pool);
1310	conf->r1buf_pool = NULL;
1311}
1312
1313static int raid1_spare_active(struct mddev *mddev)
1314{
1315	int i;
1316	struct r1conf *conf = mddev->private;
1317	int count = 0;
1318	unsigned long flags;
1319
1320	/*
1321	 * Find all failed disks within the RAID1 configuration 
1322	 * and mark them readable.
1323	 * Called under mddev lock, so rcu protection not needed.
 
 
1324	 */
 
1325	for (i = 0; i < conf->raid_disks; i++) {
1326		struct md_rdev *rdev = conf->mirrors[i].rdev;
1327		struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1328		if (repl
 
1329		    && repl->recovery_offset == MaxSector
1330		    && !test_bit(Faulty, &repl->flags)
1331		    && !test_and_set_bit(In_sync, &repl->flags)) {
1332			/* replacement has just become active */
1333			if (!rdev ||
1334			    !test_and_clear_bit(In_sync, &rdev->flags))
1335				count++;
1336			if (rdev) {
1337				/* Replaced device not technically
1338				 * faulty, but we need to be sure
1339				 * it gets removed and never re-added
1340				 */
1341				set_bit(Faulty, &rdev->flags);
1342				sysfs_notify_dirent_safe(
1343					rdev->sysfs_state);
1344			}
1345		}
1346		if (rdev
 
1347		    && !test_bit(Faulty, &rdev->flags)
1348		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1349			count++;
1350			sysfs_notify_dirent_safe(rdev->sysfs_state);
1351		}
1352	}
1353	spin_lock_irqsave(&conf->device_lock, flags);
1354	mddev->degraded -= count;
1355	spin_unlock_irqrestore(&conf->device_lock, flags);
1356
1357	print_conf(conf);
1358	return count;
1359}
1360
1361
1362static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1363{
1364	struct r1conf *conf = mddev->private;
1365	int err = -EEXIST;
1366	int mirror = 0;
1367	struct mirror_info *p;
1368	int first = 0;
1369	int last = conf->raid_disks - 1;
1370	struct request_queue *q = bdev_get_queue(rdev->bdev);
1371
1372	if (mddev->recovery_disabled == conf->recovery_disabled)
1373		return -EBUSY;
1374
 
 
 
1375	if (rdev->raid_disk >= 0)
1376		first = last = rdev->raid_disk;
1377
1378	if (q->merge_bvec_fn) {
1379		set_bit(Unmerged, &rdev->flags);
1380		mddev->merge_check_needed = 1;
1381	}
 
 
 
 
 
1382
1383	for (mirror = first; mirror <= last; mirror++) {
1384		p = conf->mirrors+mirror;
1385		if (!p->rdev) {
1386
1387			disk_stack_limits(mddev->gendisk, rdev->bdev,
1388					  rdev->data_offset << 9);
1389
1390			p->head_position = 0;
1391			rdev->raid_disk = mirror;
1392			err = 0;
1393			/* As all devices are equivalent, we don't need a full recovery
1394			 * if this was recently any drive of the array
1395			 */
1396			if (rdev->saved_raid_disk < 0)
1397				conf->fullsync = 1;
1398			rcu_assign_pointer(p->rdev, rdev);
1399			break;
1400		}
1401		if (test_bit(WantReplacement, &p->rdev->flags) &&
1402		    p[conf->raid_disks].rdev == NULL) {
1403			/* Add this device as a replacement */
1404			clear_bit(In_sync, &rdev->flags);
1405			set_bit(Replacement, &rdev->flags);
1406			rdev->raid_disk = mirror;
1407			err = 0;
1408			conf->fullsync = 1;
1409			rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1410			break;
1411		}
1412	}
1413	if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1414		/* Some requests might not have seen this new
1415		 * merge_bvec_fn.  We must wait for them to complete
1416		 * before merging the device fully.
1417		 * First we make sure any code which has tested
1418		 * our function has submitted the request, then
1419		 * we wait for all outstanding requests to complete.
1420		 */
1421		synchronize_sched();
1422		raise_barrier(conf);
1423		lower_barrier(conf);
1424		clear_bit(Unmerged, &rdev->flags);
1425	}
1426	md_integrity_add_rdev(rdev, mddev);
1427	print_conf(conf);
1428	return err;
1429}
1430
1431static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1432{
1433	struct r1conf *conf = mddev->private;
1434	int err = 0;
1435	int number = rdev->raid_disk;
1436	struct mirror_info *p = conf->mirrors+ number;
1437
1438	if (rdev != p->rdev)
1439		p = conf->mirrors + conf->raid_disks + number;
1440
1441	print_conf(conf);
1442	if (rdev == p->rdev) {
1443		if (test_bit(In_sync, &rdev->flags) ||
1444		    atomic_read(&rdev->nr_pending)) {
1445			err = -EBUSY;
1446			goto abort;
1447		}
1448		/* Only remove non-faulty devices if recovery
1449		 * is not possible.
1450		 */
1451		if (!test_bit(Faulty, &rdev->flags) &&
1452		    mddev->recovery_disabled != conf->recovery_disabled &&
1453		    mddev->degraded < conf->raid_disks) {
1454			err = -EBUSY;
1455			goto abort;
1456		}
1457		p->rdev = NULL;
1458		synchronize_rcu();
1459		if (atomic_read(&rdev->nr_pending)) {
1460			/* lost the race, try later */
1461			err = -EBUSY;
1462			p->rdev = rdev;
1463			goto abort;
1464		} else if (conf->mirrors[conf->raid_disks + number].rdev) {
 
 
 
1465			/* We just removed a device that is being replaced.
1466			 * Move down the replacement.  We drain all IO before
1467			 * doing this to avoid confusion.
1468			 */
1469			struct md_rdev *repl =
1470				conf->mirrors[conf->raid_disks + number].rdev;
1471			raise_barrier(conf);
 
 
 
 
 
 
 
 
 
 
 
1472			clear_bit(Replacement, &repl->flags);
1473			p->rdev = repl;
1474			conf->mirrors[conf->raid_disks + number].rdev = NULL;
1475			lower_barrier(conf);
1476			clear_bit(WantReplacement, &rdev->flags);
1477		} else
1478			clear_bit(WantReplacement, &rdev->flags);
1479		err = md_integrity_register(mddev);
1480	}
1481abort:
1482
1483	print_conf(conf);
1484	return err;
1485}
1486
1487
1488static void end_sync_read(struct bio *bio, int error)
1489{
1490	struct r1bio *r1_bio = bio->bi_private;
1491
1492	update_head_pos(r1_bio->read_disk, r1_bio);
1493
1494	/*
1495	 * we have read a block, now it needs to be re-written,
1496	 * or re-read if the read failed.
1497	 * We don't do much here, just schedule handling by raid1d
1498	 */
1499	if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1500		set_bit(R1BIO_Uptodate, &r1_bio->state);
1501
1502	if (atomic_dec_and_test(&r1_bio->remaining))
1503		reschedule_retry(r1_bio);
1504}
1505
1506static void end_sync_write(struct bio *bio, int error)
1507{
1508	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1509	struct r1bio *r1_bio = bio->bi_private;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1510	struct mddev *mddev = r1_bio->mddev;
1511	struct r1conf *conf = mddev->private;
1512	int mirror=0;
1513	sector_t first_bad;
1514	int bad_sectors;
1515
1516	mirror = find_bio_disk(r1_bio, bio);
1517
1518	if (!uptodate) {
1519		sector_t sync_blocks = 0;
1520		sector_t s = r1_bio->sector;
1521		long sectors_to_go = r1_bio->sectors;
1522		/* make sure these bits doesn't get cleared. */
1523		do {
1524			bitmap_end_sync(mddev->bitmap, s,
1525					&sync_blocks, 1);
1526			s += sync_blocks;
1527			sectors_to_go -= sync_blocks;
1528		} while (sectors_to_go > 0);
1529		set_bit(WriteErrorSeen,
1530			&conf->mirrors[mirror].rdev->flags);
1531		if (!test_and_set_bit(WantReplacement,
1532				      &conf->mirrors[mirror].rdev->flags))
1533			set_bit(MD_RECOVERY_NEEDED, &
1534				mddev->recovery);
1535		set_bit(R1BIO_WriteError, &r1_bio->state);
1536	} else if (is_badblock(conf->mirrors[mirror].rdev,
1537			       r1_bio->sector,
1538			       r1_bio->sectors,
1539			       &first_bad, &bad_sectors) &&
1540		   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1541				r1_bio->sector,
1542				r1_bio->sectors,
1543				&first_bad, &bad_sectors)
1544		)
1545		set_bit(R1BIO_MadeGood, &r1_bio->state);
1546
1547	if (atomic_dec_and_test(&r1_bio->remaining)) {
1548		int s = r1_bio->sectors;
1549		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1550		    test_bit(R1BIO_WriteError, &r1_bio->state))
1551			reschedule_retry(r1_bio);
1552		else {
1553			put_buf(r1_bio);
1554			md_done_sync(mddev, s, uptodate);
1555		}
1556	}
1557}
1558
1559static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1560			    int sectors, struct page *page, int rw)
1561{
1562	if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1563		/* success */
1564		return 1;
1565	if (rw == WRITE) {
1566		set_bit(WriteErrorSeen, &rdev->flags);
1567		if (!test_and_set_bit(WantReplacement,
1568				      &rdev->flags))
1569			set_bit(MD_RECOVERY_NEEDED, &
1570				rdev->mddev->recovery);
1571	}
1572	/* need to record an error - either for the block or the device */
1573	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1574		md_error(rdev->mddev, rdev);
1575	return 0;
1576}
1577
1578static int fix_sync_read_error(struct r1bio *r1_bio)
1579{
1580	/* Try some synchronous reads of other devices to get
1581	 * good data, much like with normal read errors.  Only
1582	 * read into the pages we already have so we don't
1583	 * need to re-issue the read request.
1584	 * We don't need to freeze the array, because being in an
1585	 * active sync request, there is no normal IO, and
1586	 * no overlapping syncs.
1587	 * We don't need to check is_badblock() again as we
1588	 * made sure that anything with a bad block in range
1589	 * will have bi_end_io clear.
1590	 */
1591	struct mddev *mddev = r1_bio->mddev;
1592	struct r1conf *conf = mddev->private;
1593	struct bio *bio = r1_bio->bios[r1_bio->read_disk];
 
1594	sector_t sect = r1_bio->sector;
1595	int sectors = r1_bio->sectors;
1596	int idx = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
1597
1598	while(sectors) {
1599		int s = sectors;
1600		int d = r1_bio->read_disk;
1601		int success = 0;
1602		struct md_rdev *rdev;
1603		int start;
1604
1605		if (s > (PAGE_SIZE>>9))
1606			s = PAGE_SIZE >> 9;
1607		do {
1608			if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1609				/* No rcu protection needed here devices
1610				 * can only be removed when no resync is
1611				 * active, and resync is currently active
1612				 */
1613				rdev = conf->mirrors[d].rdev;
1614				if (sync_page_io(rdev, sect, s<<9,
1615						 bio->bi_io_vec[idx].bv_page,
1616						 READ, false)) {
1617					success = 1;
1618					break;
1619				}
1620			}
1621			d++;
1622			if (d == conf->raid_disks * 2)
1623				d = 0;
1624		} while (!success && d != r1_bio->read_disk);
1625
1626		if (!success) {
1627			char b[BDEVNAME_SIZE];
1628			int abort = 0;
1629			/* Cannot read from anywhere, this block is lost.
1630			 * Record a bad block on each device.  If that doesn't
1631			 * work just disable and interrupt the recovery.
1632			 * Don't fail devices as that won't really help.
1633			 */
1634			printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1635			       " for block %llu\n",
1636			       mdname(mddev),
1637			       bdevname(bio->bi_bdev, b),
1638			       (unsigned long long)r1_bio->sector);
1639			for (d = 0; d < conf->raid_disks * 2; d++) {
1640				rdev = conf->mirrors[d].rdev;
1641				if (!rdev || test_bit(Faulty, &rdev->flags))
1642					continue;
1643				if (!rdev_set_badblocks(rdev, sect, s, 0))
1644					abort = 1;
1645			}
1646			if (abort) {
1647				conf->recovery_disabled =
1648					mddev->recovery_disabled;
1649				set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1650				md_done_sync(mddev, r1_bio->sectors, 0);
1651				put_buf(r1_bio);
1652				return 0;
1653			}
1654			/* Try next page */
1655			sectors -= s;
1656			sect += s;
1657			idx++;
1658			continue;
1659		}
1660
1661		start = d;
1662		/* write it back and re-read */
1663		while (d != r1_bio->read_disk) {
1664			if (d == 0)
1665				d = conf->raid_disks * 2;
1666			d--;
1667			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1668				continue;
1669			rdev = conf->mirrors[d].rdev;
1670			if (r1_sync_page_io(rdev, sect, s,
1671					    bio->bi_io_vec[idx].bv_page,
1672					    WRITE) == 0) {
1673				r1_bio->bios[d]->bi_end_io = NULL;
1674				rdev_dec_pending(rdev, mddev);
1675			}
1676		}
1677		d = start;
1678		while (d != r1_bio->read_disk) {
1679			if (d == 0)
1680				d = conf->raid_disks * 2;
1681			d--;
1682			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1683				continue;
1684			rdev = conf->mirrors[d].rdev;
1685			if (r1_sync_page_io(rdev, sect, s,
1686					    bio->bi_io_vec[idx].bv_page,
1687					    READ) != 0)
1688				atomic_add(s, &rdev->corrected_errors);
1689		}
1690		sectors -= s;
1691		sect += s;
1692		idx ++;
1693	}
1694	set_bit(R1BIO_Uptodate, &r1_bio->state);
1695	set_bit(BIO_UPTODATE, &bio->bi_flags);
1696	return 1;
1697}
1698
1699static int process_checks(struct r1bio *r1_bio)
1700{
1701	/* We have read all readable devices.  If we haven't
1702	 * got the block, then there is no hope left.
1703	 * If we have, then we want to do a comparison
1704	 * and skip the write if everything is the same.
1705	 * If any blocks failed to read, then we need to
1706	 * attempt an over-write
1707	 */
1708	struct mddev *mddev = r1_bio->mddev;
1709	struct r1conf *conf = mddev->private;
1710	int primary;
1711	int i;
1712	int vcnt;
1713
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1714	for (primary = 0; primary < conf->raid_disks * 2; primary++)
1715		if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1716		    test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1717			r1_bio->bios[primary]->bi_end_io = NULL;
1718			rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1719			break;
1720		}
1721	r1_bio->read_disk = primary;
1722	vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1723	for (i = 0; i < conf->raid_disks * 2; i++) {
1724		int j;
1725		struct bio *pbio = r1_bio->bios[primary];
1726		struct bio *sbio = r1_bio->bios[i];
1727		int size;
 
 
 
 
 
1728
1729		if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1730			continue;
 
 
 
 
 
1731
1732		if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1733			for (j = vcnt; j-- ; ) {
1734				struct page *p, *s;
1735				p = pbio->bi_io_vec[j].bv_page;
1736				s = sbio->bi_io_vec[j].bv_page;
1737				if (memcmp(page_address(p),
1738					   page_address(s),
1739					   sbio->bi_io_vec[j].bv_len))
1740					break;
1741			}
1742		} else
1743			j = 0;
1744		if (j >= 0)
1745			mddev->resync_mismatches += r1_bio->sectors;
1746		if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1747			      && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1748			/* No need to write to this device. */
1749			sbio->bi_end_io = NULL;
1750			rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1751			continue;
1752		}
1753		/* fixup the bio for reuse */
1754		sbio->bi_vcnt = vcnt;
1755		sbio->bi_size = r1_bio->sectors << 9;
1756		sbio->bi_idx = 0;
1757		sbio->bi_phys_segments = 0;
1758		sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1759		sbio->bi_flags |= 1 << BIO_UPTODATE;
1760		sbio->bi_next = NULL;
1761		sbio->bi_sector = r1_bio->sector +
1762			conf->mirrors[i].rdev->data_offset;
1763		sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1764		size = sbio->bi_size;
1765		for (j = 0; j < vcnt ; j++) {
1766			struct bio_vec *bi;
1767			bi = &sbio->bi_io_vec[j];
1768			bi->bv_offset = 0;
1769			if (size > PAGE_SIZE)
1770				bi->bv_len = PAGE_SIZE;
1771			else
1772				bi->bv_len = size;
1773			size -= PAGE_SIZE;
1774			memcpy(page_address(bi->bv_page),
1775			       page_address(pbio->bi_io_vec[j].bv_page),
1776			       PAGE_SIZE);
1777		}
1778	}
1779	return 0;
1780}
1781
1782static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1783{
1784	struct r1conf *conf = mddev->private;
1785	int i;
1786	int disks = conf->raid_disks * 2;
1787	struct bio *bio, *wbio;
1788
1789	bio = r1_bio->bios[r1_bio->read_disk];
1790
1791	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1792		/* ouch - failed to read all of that. */
1793		if (!fix_sync_read_error(r1_bio))
1794			return;
1795
1796	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1797		if (process_checks(r1_bio) < 0)
1798			return;
1799	/*
1800	 * schedule writes
1801	 */
1802	atomic_set(&r1_bio->remaining, 1);
1803	for (i = 0; i < disks ; i++) {
1804		wbio = r1_bio->bios[i];
1805		if (wbio->bi_end_io == NULL ||
1806		    (wbio->bi_end_io == end_sync_read &&
1807		     (i == r1_bio->read_disk ||
1808		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1809			continue;
 
 
 
 
 
 
 
 
1810
1811		wbio->bi_rw = WRITE;
1812		wbio->bi_end_io = end_sync_write;
1813		atomic_inc(&r1_bio->remaining);
1814		md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1815
1816		generic_make_request(wbio);
1817	}
1818
1819	if (atomic_dec_and_test(&r1_bio->remaining)) {
1820		/* if we're here, all write(s) have completed, so clean up */
1821		int s = r1_bio->sectors;
1822		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1823		    test_bit(R1BIO_WriteError, &r1_bio->state))
1824			reschedule_retry(r1_bio);
1825		else {
1826			put_buf(r1_bio);
1827			md_done_sync(mddev, s, 1);
1828		}
1829	}
1830}
1831
1832/*
1833 * This is a kernel thread which:
1834 *
1835 *	1.	Retries failed read operations on working mirrors.
1836 *	2.	Updates the raid superblock when problems encounter.
1837 *	3.	Performs writes following reads for array synchronising.
1838 */
1839
1840static void fix_read_error(struct r1conf *conf, int read_disk,
1841			   sector_t sect, int sectors)
1842{
1843	struct mddev *mddev = conf->mddev;
1844	while(sectors) {
1845		int s = sectors;
1846		int d = read_disk;
1847		int success = 0;
1848		int start;
1849		struct md_rdev *rdev;
1850
1851		if (s > (PAGE_SIZE>>9))
1852			s = PAGE_SIZE >> 9;
1853
1854		do {
1855			/* Note: no rcu protection needed here
1856			 * as this is synchronous in the raid1d thread
1857			 * which is the thread that might remove
1858			 * a device.  If raid1d ever becomes multi-threaded....
1859			 */
1860			sector_t first_bad;
1861			int bad_sectors;
1862
1863			rdev = conf->mirrors[d].rdev;
 
1864			if (rdev &&
1865			    (test_bit(In_sync, &rdev->flags) ||
1866			     (!test_bit(Faulty, &rdev->flags) &&
1867			      rdev->recovery_offset >= sect + s)) &&
1868			    is_badblock(rdev, sect, s,
1869					&first_bad, &bad_sectors) == 0 &&
1870			    sync_page_io(rdev, sect, s<<9,
1871					 conf->tmppage, READ, false))
1872				success = 1;
1873			else {
1874				d++;
1875				if (d == conf->raid_disks * 2)
1876					d = 0;
1877			}
 
 
 
 
 
1878		} while (!success && d != read_disk);
1879
1880		if (!success) {
1881			/* Cannot read from anywhere - mark it bad */
1882			struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1883			if (!rdev_set_badblocks(rdev, sect, s, 0))
1884				md_error(mddev, rdev);
1885			break;
1886		}
1887		/* write it back and re-read */
1888		start = d;
1889		while (d != read_disk) {
1890			if (d==0)
1891				d = conf->raid_disks * 2;
1892			d--;
1893			rdev = conf->mirrors[d].rdev;
 
1894			if (rdev &&
1895			    test_bit(In_sync, &rdev->flags))
 
 
1896				r1_sync_page_io(rdev, sect, s,
1897						conf->tmppage, WRITE);
 
 
 
1898		}
1899		d = start;
1900		while (d != read_disk) {
1901			char b[BDEVNAME_SIZE];
1902			if (d==0)
1903				d = conf->raid_disks * 2;
1904			d--;
1905			rdev = conf->mirrors[d].rdev;
 
1906			if (rdev &&
1907			    test_bit(In_sync, &rdev->flags)) {
 
 
1908				if (r1_sync_page_io(rdev, sect, s,
1909						    conf->tmppage, READ)) {
1910					atomic_add(s, &rdev->corrected_errors);
1911					printk(KERN_INFO
1912					       "md/raid1:%s: read error corrected "
1913					       "(%d sectors at %llu on %s)\n",
1914					       mdname(mddev), s,
1915					       (unsigned long long)(sect +
1916					           rdev->data_offset),
1917					       bdevname(rdev->bdev, b));
1918				}
1919			}
 
 
1920		}
1921		sectors -= s;
1922		sect += s;
1923	}
1924}
1925
1926static void bi_complete(struct bio *bio, int error)
1927{
1928	complete((struct completion *)bio->bi_private);
1929}
1930
1931static int submit_bio_wait(int rw, struct bio *bio)
1932{
1933	struct completion event;
1934	rw |= REQ_SYNC;
1935
1936	init_completion(&event);
1937	bio->bi_private = &event;
1938	bio->bi_end_io = bi_complete;
1939	submit_bio(rw, bio);
1940	wait_for_completion(&event);
1941
1942	return test_bit(BIO_UPTODATE, &bio->bi_flags);
1943}
1944
1945static int narrow_write_error(struct r1bio *r1_bio, int i)
1946{
1947	struct mddev *mddev = r1_bio->mddev;
1948	struct r1conf *conf = mddev->private;
1949	struct md_rdev *rdev = conf->mirrors[i].rdev;
1950	int vcnt, idx;
1951	struct bio_vec *vec;
1952
1953	/* bio has the data to be written to device 'i' where
1954	 * we just recently had a write error.
1955	 * We repeatedly clone the bio and trim down to one block,
1956	 * then try the write.  Where the write fails we record
1957	 * a bad block.
1958	 * It is conceivable that the bio doesn't exactly align with
1959	 * blocks.  We must handle this somehow.
1960	 *
1961	 * We currently own a reference on the rdev.
1962	 */
1963
1964	int block_sectors;
1965	sector_t sector;
1966	int sectors;
1967	int sect_to_write = r1_bio->sectors;
1968	int ok = 1;
1969
1970	if (rdev->badblocks.shift < 0)
1971		return 0;
1972
1973	block_sectors = 1 << rdev->badblocks.shift;
 
1974	sector = r1_bio->sector;
1975	sectors = ((sector + block_sectors)
1976		   & ~(sector_t)(block_sectors - 1))
1977		- sector;
1978
1979	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1980		vcnt = r1_bio->behind_page_count;
1981		vec = r1_bio->behind_bvecs;
1982		idx = 0;
1983		while (vec[idx].bv_page == NULL)
1984			idx++;
1985	} else {
1986		vcnt = r1_bio->master_bio->bi_vcnt;
1987		vec = r1_bio->master_bio->bi_io_vec;
1988		idx = r1_bio->master_bio->bi_idx;
1989	}
1990	while (sect_to_write) {
1991		struct bio *wbio;
1992		if (sectors > sect_to_write)
1993			sectors = sect_to_write;
1994		/* Write at 'sector' for 'sectors'*/
1995
1996		wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1997		memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1998		wbio->bi_sector = r1_bio->sector;
1999		wbio->bi_rw = WRITE;
2000		wbio->bi_vcnt = vcnt;
2001		wbio->bi_size = r1_bio->sectors << 9;
2002		wbio->bi_idx = idx;
2003
2004		md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2005		wbio->bi_sector += rdev->data_offset;
2006		wbio->bi_bdev = rdev->bdev;
2007		if (submit_bio_wait(WRITE, wbio) == 0)
 
 
 
 
 
 
2008			/* failure! */
2009			ok = rdev_set_badblocks(rdev, sector,
2010						sectors, 0)
2011				&& ok;
2012
2013		bio_put(wbio);
2014		sect_to_write -= sectors;
2015		sector += sectors;
2016		sectors = block_sectors;
2017	}
2018	return ok;
2019}
2020
2021static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2022{
2023	int m;
2024	int s = r1_bio->sectors;
2025	for (m = 0; m < conf->raid_disks * 2 ; m++) {
2026		struct md_rdev *rdev = conf->mirrors[m].rdev;
2027		struct bio *bio = r1_bio->bios[m];
2028		if (bio->bi_end_io == NULL)
2029			continue;
2030		if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2031		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2032			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2033		}
2034		if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2035		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2036			if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2037				md_error(conf->mddev, rdev);
2038		}
2039	}
2040	put_buf(r1_bio);
2041	md_done_sync(conf->mddev, s, 1);
2042}
2043
2044static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2045{
2046	int m;
 
 
2047	for (m = 0; m < conf->raid_disks * 2 ; m++)
2048		if (r1_bio->bios[m] == IO_MADE_GOOD) {
2049			struct md_rdev *rdev = conf->mirrors[m].rdev;
2050			rdev_clear_badblocks(rdev,
2051					     r1_bio->sector,
2052					     r1_bio->sectors, 0);
2053			rdev_dec_pending(rdev, conf->mddev);
2054		} else if (r1_bio->bios[m] != NULL) {
2055			/* This drive got a write error.  We need to
2056			 * narrow down and record precise write
2057			 * errors.
2058			 */
 
2059			if (!narrow_write_error(r1_bio, m)) {
2060				md_error(conf->mddev,
2061					 conf->mirrors[m].rdev);
2062				/* an I/O failed, we can't clear the bitmap */
2063				set_bit(R1BIO_Degraded, &r1_bio->state);
2064			}
2065			rdev_dec_pending(conf->mirrors[m].rdev,
2066					 conf->mddev);
2067		}
2068	if (test_bit(R1BIO_WriteError, &r1_bio->state))
2069		close_write(r1_bio);
2070	raid_end_bio_io(r1_bio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2071}
2072
2073static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2074{
2075	int disk;
2076	int max_sectors;
2077	struct mddev *mddev = conf->mddev;
2078	struct bio *bio;
2079	char b[BDEVNAME_SIZE];
2080	struct md_rdev *rdev;
2081
2082	clear_bit(R1BIO_ReadError, &r1_bio->state);
2083	/* we got a read error. Maybe the drive is bad.  Maybe just
2084	 * the block and we can fix it.
2085	 * We freeze all other IO, and try reading the block from
2086	 * other devices.  When we find one, we re-write
2087	 * and check it that fixes the read error.
2088	 * This is all done synchronously while the array is
2089	 * frozen
2090	 */
2091	if (mddev->ro == 0) {
2092		freeze_array(conf);
 
 
 
 
 
 
 
2093		fix_read_error(conf, r1_bio->read_disk,
2094			       r1_bio->sector, r1_bio->sectors);
2095		unfreeze_array(conf);
2096	} else
2097		md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2098
2099	bio = r1_bio->bios[r1_bio->read_disk];
2100	bdevname(bio->bi_bdev, b);
2101read_more:
2102	disk = read_balance(conf, r1_bio, &max_sectors);
2103	if (disk == -1) {
2104		printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2105		       " read error for block %llu\n",
2106		       mdname(mddev), b, (unsigned long long)r1_bio->sector);
2107		raid_end_bio_io(r1_bio);
2108	} else {
2109		const unsigned long do_sync
2110			= r1_bio->master_bio->bi_rw & REQ_SYNC;
2111		if (bio) {
2112			r1_bio->bios[r1_bio->read_disk] =
2113				mddev->ro ? IO_BLOCKED : NULL;
2114			bio_put(bio);
2115		}
2116		r1_bio->read_disk = disk;
2117		bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2118		md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2119		r1_bio->bios[r1_bio->read_disk] = bio;
2120		rdev = conf->mirrors[disk].rdev;
2121		printk_ratelimited(KERN_ERR
2122				   "md/raid1:%s: redirecting sector %llu"
2123				   " to other mirror: %s\n",
2124				   mdname(mddev),
2125				   (unsigned long long)r1_bio->sector,
2126				   bdevname(rdev->bdev, b));
2127		bio->bi_sector = r1_bio->sector + rdev->data_offset;
2128		bio->bi_bdev = rdev->bdev;
2129		bio->bi_end_io = raid1_end_read_request;
2130		bio->bi_rw = READ | do_sync;
2131		bio->bi_private = r1_bio;
2132		if (max_sectors < r1_bio->sectors) {
2133			/* Drat - have to split this up more */
2134			struct bio *mbio = r1_bio->master_bio;
2135			int sectors_handled = (r1_bio->sector + max_sectors
2136					       - mbio->bi_sector);
2137			r1_bio->sectors = max_sectors;
2138			spin_lock_irq(&conf->device_lock);
2139			if (mbio->bi_phys_segments == 0)
2140				mbio->bi_phys_segments = 2;
2141			else
2142				mbio->bi_phys_segments++;
2143			spin_unlock_irq(&conf->device_lock);
2144			generic_make_request(bio);
2145			bio = NULL;
2146
2147			r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2148
2149			r1_bio->master_bio = mbio;
2150			r1_bio->sectors = (mbio->bi_size >> 9)
2151					  - sectors_handled;
2152			r1_bio->state = 0;
2153			set_bit(R1BIO_ReadError, &r1_bio->state);
2154			r1_bio->mddev = mddev;
2155			r1_bio->sector = mbio->bi_sector + sectors_handled;
2156
2157			goto read_more;
2158		} else
2159			generic_make_request(bio);
2160	}
2161}
2162
2163static void raid1d(struct mddev *mddev)
2164{
 
2165	struct r1bio *r1_bio;
2166	unsigned long flags;
2167	struct r1conf *conf = mddev->private;
2168	struct list_head *head = &conf->retry_list;
2169	struct blk_plug plug;
 
2170
2171	md_check_recovery(mddev);
2172
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2173	blk_start_plug(&plug);
2174	for (;;) {
2175
2176		if (atomic_read(&mddev->plug_cnt) == 0)
2177			flush_pending_writes(conf);
2178
2179		spin_lock_irqsave(&conf->device_lock, flags);
2180		if (list_empty(head)) {
2181			spin_unlock_irqrestore(&conf->device_lock, flags);
2182			break;
2183		}
2184		r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2185		list_del(head->prev);
2186		conf->nr_queued--;
 
2187		spin_unlock_irqrestore(&conf->device_lock, flags);
2188
2189		mddev = r1_bio->mddev;
2190		conf = mddev->private;
2191		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2192			if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2193			    test_bit(R1BIO_WriteError, &r1_bio->state))
2194				handle_sync_write_finished(conf, r1_bio);
2195			else
2196				sync_request_write(mddev, r1_bio);
2197		} else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2198			   test_bit(R1BIO_WriteError, &r1_bio->state))
2199			handle_write_finished(conf, r1_bio);
2200		else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2201			handle_read_error(conf, r1_bio);
2202		else
2203			/* just a partial read to be scheduled from separate
2204			 * context
2205			 */
2206			generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2207
2208		cond_resched();
2209		if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2210			md_check_recovery(mddev);
2211	}
2212	blk_finish_plug(&plug);
2213}
2214
2215
2216static int init_resync(struct r1conf *conf)
2217{
2218	int buffs;
2219
2220	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2221	BUG_ON(conf->r1buf_pool);
2222	conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2223					  conf->poolinfo);
2224	if (!conf->r1buf_pool)
2225		return -ENOMEM;
2226	conf->next_resync = 0;
2227	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2228}
2229
2230/*
2231 * perform a "sync" on one "block"
2232 *
2233 * We need to make sure that no normal I/O request - particularly write
2234 * requests - conflict with active sync requests.
2235 *
2236 * This is achieved by tracking pending requests and a 'barrier' concept
2237 * that can be installed to exclude normal IO requests.
2238 */
2239
2240static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
 
2241{
2242	struct r1conf *conf = mddev->private;
2243	struct r1bio *r1_bio;
2244	struct bio *bio;
2245	sector_t max_sector, nr_sectors;
2246	int disk = -1;
2247	int i;
2248	int wonly = -1;
2249	int write_targets = 0, read_targets = 0;
2250	sector_t sync_blocks;
2251	int still_degraded = 0;
2252	int good_sectors = RESYNC_SECTORS;
2253	int min_bad = 0; /* number of sectors that are bad in all devices */
 
 
2254
2255	if (!conf->r1buf_pool)
2256		if (init_resync(conf))
2257			return 0;
2258
2259	max_sector = mddev->dev_sectors;
2260	if (sector_nr >= max_sector) {
2261		/* If we aborted, we need to abort the
2262		 * sync on the 'current' bitmap chunk (there will
2263		 * only be one in raid1 resync.
2264		 * We can find the current addess in mddev->curr_resync
2265		 */
2266		if (mddev->curr_resync < max_sector) /* aborted */
2267			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2268						&sync_blocks, 1);
2269		else /* completed sync */
2270			conf->fullsync = 0;
2271
2272		bitmap_close_sync(mddev->bitmap);
2273		close_sync(conf);
 
 
 
 
 
2274		return 0;
2275	}
2276
2277	if (mddev->bitmap == NULL &&
2278	    mddev->recovery_cp == MaxSector &&
2279	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2280	    conf->fullsync == 0) {
2281		*skipped = 1;
2282		return max_sector - sector_nr;
2283	}
2284	/* before building a request, check if we can skip these blocks..
2285	 * This call the bitmap_start_sync doesn't actually record anything
2286	 */
2287	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2288	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2289		/* We can skip this block, and probably several more */
2290		*skipped = 1;
2291		return sync_blocks;
2292	}
 
2293	/*
2294	 * If there is non-resync activity waiting for a turn,
2295	 * and resync is going fast enough,
2296	 * then let it though before starting on this new sync request.
2297	 */
2298	if (!go_faster && conf->nr_waiting)
2299		msleep_interruptible(1000);
2300
2301	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2302	r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2303	raise_barrier(conf);
 
 
 
2304
2305	conf->next_resync = sector_nr;
 
 
 
2306
2307	rcu_read_lock();
2308	/*
2309	 * If we get a correctably read error during resync or recovery,
2310	 * we might want to read from a different device.  So we
2311	 * flag all drives that could conceivably be read from for READ,
2312	 * and any others (which will be non-In_sync devices) for WRITE.
2313	 * If a read fails, we try reading from something else for which READ
2314	 * is OK.
2315	 */
2316
2317	r1_bio->mddev = mddev;
2318	r1_bio->sector = sector_nr;
2319	r1_bio->state = 0;
2320	set_bit(R1BIO_IsSync, &r1_bio->state);
 
 
2321
2322	for (i = 0; i < conf->raid_disks * 2; i++) {
2323		struct md_rdev *rdev;
2324		bio = r1_bio->bios[i];
2325
2326		/* take from bio_init */
2327		bio->bi_next = NULL;
2328		bio->bi_flags &= ~(BIO_POOL_MASK-1);
2329		bio->bi_flags |= 1 << BIO_UPTODATE;
2330		bio->bi_rw = READ;
2331		bio->bi_vcnt = 0;
2332		bio->bi_idx = 0;
2333		bio->bi_phys_segments = 0;
2334		bio->bi_size = 0;
2335		bio->bi_end_io = NULL;
2336		bio->bi_private = NULL;
2337
2338		rdev = rcu_dereference(conf->mirrors[i].rdev);
2339		if (rdev == NULL ||
2340		    test_bit(Faulty, &rdev->flags)) {
2341			if (i < conf->raid_disks)
2342				still_degraded = 1;
2343		} else if (!test_bit(In_sync, &rdev->flags)) {
2344			bio->bi_rw = WRITE;
2345			bio->bi_end_io = end_sync_write;
2346			write_targets ++;
2347		} else {
2348			/* may need to read from here */
2349			sector_t first_bad = MaxSector;
2350			int bad_sectors;
2351
2352			if (is_badblock(rdev, sector_nr, good_sectors,
2353					&first_bad, &bad_sectors)) {
2354				if (first_bad > sector_nr)
2355					good_sectors = first_bad - sector_nr;
2356				else {
2357					bad_sectors -= (sector_nr - first_bad);
2358					if (min_bad == 0 ||
2359					    min_bad > bad_sectors)
2360						min_bad = bad_sectors;
2361				}
2362			}
2363			if (sector_nr < first_bad) {
2364				if (test_bit(WriteMostly, &rdev->flags)) {
2365					if (wonly < 0)
2366						wonly = i;
2367				} else {
2368					if (disk < 0)
2369						disk = i;
2370				}
2371				bio->bi_rw = READ;
2372				bio->bi_end_io = end_sync_read;
2373				read_targets++;
 
 
 
 
 
 
 
 
 
 
 
 
2374			}
2375		}
2376		if (bio->bi_end_io) {
2377			atomic_inc(&rdev->nr_pending);
2378			bio->bi_sector = sector_nr + rdev->data_offset;
2379			bio->bi_bdev = rdev->bdev;
2380			bio->bi_private = r1_bio;
 
2381		}
2382	}
2383	rcu_read_unlock();
2384	if (disk < 0)
2385		disk = wonly;
2386	r1_bio->read_disk = disk;
2387
2388	if (read_targets == 0 && min_bad > 0) {
2389		/* These sectors are bad on all InSync devices, so we
2390		 * need to mark them bad on all write targets
2391		 */
2392		int ok = 1;
2393		for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2394			if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2395				struct md_rdev *rdev = conf->mirrors[i].rdev;
2396				ok = rdev_set_badblocks(rdev, sector_nr,
2397							min_bad, 0
2398					) && ok;
2399			}
2400		set_bit(MD_CHANGE_DEVS, &mddev->flags);
2401		*skipped = 1;
2402		put_buf(r1_bio);
2403
2404		if (!ok) {
2405			/* Cannot record the badblocks, so need to
2406			 * abort the resync.
2407			 * If there are multiple read targets, could just
2408			 * fail the really bad ones ???
2409			 */
2410			conf->recovery_disabled = mddev->recovery_disabled;
2411			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2412			return 0;
2413		} else
2414			return min_bad;
2415
2416	}
2417	if (min_bad > 0 && min_bad < good_sectors) {
2418		/* only resync enough to reach the next bad->good
2419		 * transition */
2420		good_sectors = min_bad;
2421	}
2422
2423	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2424		/* extra read targets are also write targets */
2425		write_targets += read_targets-1;
2426
2427	if (write_targets == 0 || read_targets == 0) {
2428		/* There is nowhere to write, so all non-sync
2429		 * drives must be failed - so we are finished
2430		 */
2431		sector_t rv;
2432		if (min_bad > 0)
2433			max_sector = sector_nr + min_bad;
2434		rv = max_sector - sector_nr;
2435		*skipped = 1;
2436		put_buf(r1_bio);
2437		return rv;
2438	}
2439
2440	if (max_sector > mddev->resync_max)
2441		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2442	if (max_sector > sector_nr + good_sectors)
2443		max_sector = sector_nr + good_sectors;
2444	nr_sectors = 0;
2445	sync_blocks = 0;
2446	do {
2447		struct page *page;
2448		int len = PAGE_SIZE;
2449		if (sector_nr + (len>>9) > max_sector)
2450			len = (max_sector - sector_nr) << 9;
2451		if (len == 0)
2452			break;
2453		if (sync_blocks == 0) {
2454			if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2455					       &sync_blocks, still_degraded) &&
2456			    !conf->fullsync &&
2457			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2458				break;
2459			BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2460			if ((len >> 9) > sync_blocks)
2461				len = sync_blocks<<9;
2462		}
2463
2464		for (i = 0 ; i < conf->raid_disks * 2; i++) {
 
 
2465			bio = r1_bio->bios[i];
 
2466			if (bio->bi_end_io) {
2467				page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2468				if (bio_add_page(bio, page, len, 0) == 0) {
2469					/* stop here */
2470					bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2471					while (i > 0) {
2472						i--;
2473						bio = r1_bio->bios[i];
2474						if (bio->bi_end_io==NULL)
2475							continue;
2476						/* remove last page from this bio */
2477						bio->bi_vcnt--;
2478						bio->bi_size -= len;
2479						bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2480					}
2481					goto bio_full;
2482				}
2483			}
2484		}
2485		nr_sectors += len>>9;
2486		sector_nr += len>>9;
2487		sync_blocks -= (len>>9);
2488	} while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2489 bio_full:
2490	r1_bio->sectors = nr_sectors;
2491
 
 
 
 
 
 
 
 
 
 
2492	/* For a user-requested sync, we read all readable devices and do a
2493	 * compare
2494	 */
2495	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2496		atomic_set(&r1_bio->remaining, read_targets);
2497		for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2498			bio = r1_bio->bios[i];
2499			if (bio->bi_end_io == end_sync_read) {
2500				read_targets--;
2501				md_sync_acct(bio->bi_bdev, nr_sectors);
 
 
2502				generic_make_request(bio);
2503			}
2504		}
2505	} else {
2506		atomic_set(&r1_bio->remaining, 1);
2507		bio = r1_bio->bios[r1_bio->read_disk];
2508		md_sync_acct(bio->bi_bdev, nr_sectors);
 
 
2509		generic_make_request(bio);
2510
2511	}
2512	return nr_sectors;
2513}
2514
2515static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2516{
2517	if (sectors)
2518		return sectors;
2519
2520	return mddev->dev_sectors;
2521}
2522
2523static struct r1conf *setup_conf(struct mddev *mddev)
2524{
2525	struct r1conf *conf;
2526	int i;
2527	struct mirror_info *disk;
2528	struct md_rdev *rdev;
2529	int err = -ENOMEM;
2530
2531	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2532	if (!conf)
2533		goto abort;
2534
2535	conf->mirrors = kzalloc(sizeof(struct mirror_info)
2536				* mddev->raid_disks * 2,
2537				 GFP_KERNEL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2538	if (!conf->mirrors)
2539		goto abort;
2540
2541	conf->tmppage = alloc_page(GFP_KERNEL);
2542	if (!conf->tmppage)
2543		goto abort;
2544
2545	conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2546	if (!conf->poolinfo)
2547		goto abort;
2548	conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2549	conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2550					  r1bio_pool_free,
2551					  conf->poolinfo);
2552	if (!conf->r1bio_pool)
 
 
 
2553		goto abort;
2554
2555	conf->poolinfo->mddev = mddev;
2556
2557	err = -EINVAL;
2558	spin_lock_init(&conf->device_lock);
2559	rdev_for_each(rdev, mddev) {
2560		struct request_queue *q;
2561		int disk_idx = rdev->raid_disk;
2562		if (disk_idx >= mddev->raid_disks
2563		    || disk_idx < 0)
2564			continue;
2565		if (test_bit(Replacement, &rdev->flags))
2566			disk = conf->mirrors + conf->raid_disks + disk_idx;
2567		else
2568			disk = conf->mirrors + disk_idx;
2569
2570		if (disk->rdev)
2571			goto abort;
2572		disk->rdev = rdev;
2573		q = bdev_get_queue(rdev->bdev);
2574		if (q->merge_bvec_fn)
2575			mddev->merge_check_needed = 1;
2576
2577		disk->head_position = 0;
 
2578	}
2579	conf->raid_disks = mddev->raid_disks;
2580	conf->mddev = mddev;
2581	INIT_LIST_HEAD(&conf->retry_list);
 
2582
2583	spin_lock_init(&conf->resync_lock);
2584	init_waitqueue_head(&conf->wait_barrier);
2585
2586	bio_list_init(&conf->pending_bio_list);
2587	conf->pending_count = 0;
2588	conf->recovery_disabled = mddev->recovery_disabled - 1;
2589
2590	err = -EIO;
2591	conf->last_used = -1;
2592	for (i = 0; i < conf->raid_disks * 2; i++) {
2593
2594		disk = conf->mirrors + i;
2595
2596		if (i < conf->raid_disks &&
2597		    disk[conf->raid_disks].rdev) {
2598			/* This slot has a replacement. */
2599			if (!disk->rdev) {
2600				/* No original, just make the replacement
2601				 * a recovering spare
2602				 */
2603				disk->rdev =
2604					disk[conf->raid_disks].rdev;
2605				disk[conf->raid_disks].rdev = NULL;
2606			} else if (!test_bit(In_sync, &disk->rdev->flags))
2607				/* Original is not in_sync - bad */
2608				goto abort;
2609		}
2610
2611		if (!disk->rdev ||
2612		    !test_bit(In_sync, &disk->rdev->flags)) {
2613			disk->head_position = 0;
2614			if (disk->rdev &&
2615			    (disk->rdev->saved_raid_disk < 0))
2616				conf->fullsync = 1;
2617		} else if (conf->last_used < 0)
2618			/*
2619			 * The first working device is used as a
2620			 * starting point to read balancing.
2621			 */
2622			conf->last_used = i;
2623	}
2624
2625	if (conf->last_used < 0) {
2626		printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2627		       mdname(mddev));
2628		goto abort;
2629	}
2630	err = -ENOMEM;
2631	conf->thread = md_register_thread(raid1d, mddev, "raid1");
2632	if (!conf->thread) {
2633		printk(KERN_ERR
2634		       "md/raid1:%s: couldn't allocate thread\n",
2635		       mdname(mddev));
2636		goto abort;
2637	}
2638
2639	return conf;
2640
2641 abort:
2642	if (conf) {
2643		if (conf->r1bio_pool)
2644			mempool_destroy(conf->r1bio_pool);
2645		kfree(conf->mirrors);
2646		safe_put_page(conf->tmppage);
2647		kfree(conf->poolinfo);
 
 
 
 
 
2648		kfree(conf);
2649	}
2650	return ERR_PTR(err);
2651}
2652
2653static int stop(struct mddev *mddev);
2654static int run(struct mddev *mddev)
2655{
2656	struct r1conf *conf;
2657	int i;
2658	struct md_rdev *rdev;
2659	int ret;
 
2660
2661	if (mddev->level != 1) {
2662		printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2663		       mdname(mddev), mddev->level);
2664		return -EIO;
2665	}
2666	if (mddev->reshape_position != MaxSector) {
2667		printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2668		       mdname(mddev));
2669		return -EIO;
2670	}
 
 
2671	/*
2672	 * copy the already verified devices into our private RAID1
2673	 * bookkeeping area. [whatever we allocate in run(),
2674	 * should be freed in stop()]
2675	 */
2676	if (mddev->private == NULL)
2677		conf = setup_conf(mddev);
2678	else
2679		conf = mddev->private;
2680
2681	if (IS_ERR(conf))
2682		return PTR_ERR(conf);
2683
 
 
 
 
 
2684	rdev_for_each(rdev, mddev) {
2685		if (!mddev->gendisk)
2686			continue;
2687		disk_stack_limits(mddev->gendisk, rdev->bdev,
2688				  rdev->data_offset << 9);
 
 
2689	}
2690
2691	mddev->degraded = 0;
2692	for (i=0; i < conf->raid_disks; i++)
2693		if (conf->mirrors[i].rdev == NULL ||
2694		    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2695		    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2696			mddev->degraded++;
 
 
 
 
 
 
 
2697
2698	if (conf->raid_disks - mddev->degraded == 1)
2699		mddev->recovery_cp = MaxSector;
2700
2701	if (mddev->recovery_cp != MaxSector)
2702		printk(KERN_NOTICE "md/raid1:%s: not clean"
2703		       " -- starting background reconstruction\n",
2704		       mdname(mddev));
2705	printk(KERN_INFO 
2706		"md/raid1:%s: active with %d out of %d mirrors\n",
2707		mdname(mddev), mddev->raid_disks - mddev->degraded, 
2708		mddev->raid_disks);
2709
2710	/*
2711	 * Ok, everything is just fine now
2712	 */
2713	mddev->thread = conf->thread;
2714	conf->thread = NULL;
2715	mddev->private = conf;
 
2716
2717	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2718
2719	if (mddev->queue) {
2720		mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2721		mddev->queue->backing_dev_info.congested_data = mddev;
2722		blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
 
 
 
2723	}
2724
2725	ret =  md_integrity_register(mddev);
2726	if (ret)
2727		stop(mddev);
 
 
 
 
 
 
2728	return ret;
2729}
2730
2731static int stop(struct mddev *mddev)
2732{
2733	struct r1conf *conf = mddev->private;
2734	struct bitmap *bitmap = mddev->bitmap;
2735
2736	/* wait for behind writes to complete */
2737	if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2738		printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2739		       mdname(mddev));
2740		/* need to kick something here to make sure I/O goes? */
2741		wait_event(bitmap->behind_wait,
2742			   atomic_read(&bitmap->behind_writes) == 0);
2743	}
2744
2745	raise_barrier(conf);
2746	lower_barrier(conf);
2747
2748	md_unregister_thread(&mddev->thread);
2749	if (conf->r1bio_pool)
2750		mempool_destroy(conf->r1bio_pool);
2751	kfree(conf->mirrors);
 
2752	kfree(conf->poolinfo);
 
 
 
 
 
2753	kfree(conf);
2754	mddev->private = NULL;
2755	return 0;
2756}
2757
2758static int raid1_resize(struct mddev *mddev, sector_t sectors)
2759{
2760	/* no resync is happening, and there is enough space
2761	 * on all devices, so we can resize.
2762	 * We need to make sure resync covers any new space.
2763	 * If the array is shrinking we should possibly wait until
2764	 * any io in the removed space completes, but it hardly seems
2765	 * worth it.
2766	 */
2767	sector_t newsize = raid1_size(mddev, sectors, 0);
2768	if (mddev->external_size &&
2769	    mddev->array_sectors > newsize)
2770		return -EINVAL;
2771	if (mddev->bitmap) {
2772		int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2773		if (ret)
2774			return ret;
2775	}
2776	md_set_array_sectors(mddev, newsize);
2777	set_capacity(mddev->gendisk, mddev->array_sectors);
2778	revalidate_disk(mddev->gendisk);
2779	if (sectors > mddev->dev_sectors &&
2780	    mddev->recovery_cp > mddev->dev_sectors) {
2781		mddev->recovery_cp = mddev->dev_sectors;
2782		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2783	}
2784	mddev->dev_sectors = sectors;
2785	mddev->resync_max_sectors = sectors;
2786	return 0;
2787}
2788
2789static int raid1_reshape(struct mddev *mddev)
2790{
2791	/* We need to:
2792	 * 1/ resize the r1bio_pool
2793	 * 2/ resize conf->mirrors
2794	 *
2795	 * We allocate a new r1bio_pool if we can.
2796	 * Then raise a device barrier and wait until all IO stops.
2797	 * Then resize conf->mirrors and swap in the new r1bio pool.
2798	 *
2799	 * At the same time, we "pack" the devices so that all the missing
2800	 * devices have the higher raid_disk numbers.
2801	 */
2802	mempool_t *newpool, *oldpool;
2803	struct pool_info *newpoolinfo;
2804	struct mirror_info *newmirrors;
2805	struct r1conf *conf = mddev->private;
2806	int cnt, raid_disks;
2807	unsigned long flags;
2808	int d, d2, err;
 
 
 
 
2809
2810	/* Cannot change chunk_size, layout, or level */
2811	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2812	    mddev->layout != mddev->new_layout ||
2813	    mddev->level != mddev->new_level) {
2814		mddev->new_chunk_sectors = mddev->chunk_sectors;
2815		mddev->new_layout = mddev->layout;
2816		mddev->new_level = mddev->level;
2817		return -EINVAL;
2818	}
2819
2820	err = md_allow_write(mddev);
2821	if (err)
2822		return err;
2823
2824	raid_disks = mddev->raid_disks + mddev->delta_disks;
2825
2826	if (raid_disks < conf->raid_disks) {
2827		cnt=0;
2828		for (d= 0; d < conf->raid_disks; d++)
2829			if (conf->mirrors[d].rdev)
2830				cnt++;
2831		if (cnt > raid_disks)
2832			return -EBUSY;
2833	}
2834
2835	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2836	if (!newpoolinfo)
2837		return -ENOMEM;
2838	newpoolinfo->mddev = mddev;
2839	newpoolinfo->raid_disks = raid_disks * 2;
2840
2841	newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2842				 r1bio_pool_free, newpoolinfo);
2843	if (!newpool) {
2844		kfree(newpoolinfo);
2845		return -ENOMEM;
2846	}
2847	newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks * 2,
 
2848			     GFP_KERNEL);
2849	if (!newmirrors) {
2850		kfree(newpoolinfo);
2851		mempool_destroy(newpool);
2852		return -ENOMEM;
2853	}
2854
2855	raise_barrier(conf);
2856
2857	/* ok, everything is stopped */
2858	oldpool = conf->r1bio_pool;
2859	conf->r1bio_pool = newpool;
2860
2861	for (d = d2 = 0; d < conf->raid_disks; d++) {
2862		struct md_rdev *rdev = conf->mirrors[d].rdev;
2863		if (rdev && rdev->raid_disk != d2) {
2864			sysfs_unlink_rdev(mddev, rdev);
2865			rdev->raid_disk = d2;
2866			sysfs_unlink_rdev(mddev, rdev);
2867			if (sysfs_link_rdev(mddev, rdev))
2868				printk(KERN_WARNING
2869				       "md/raid1:%s: cannot register rd%d\n",
2870				       mdname(mddev), rdev->raid_disk);
2871		}
2872		if (rdev)
2873			newmirrors[d2++].rdev = rdev;
2874	}
2875	kfree(conf->mirrors);
2876	conf->mirrors = newmirrors;
2877	kfree(conf->poolinfo);
2878	conf->poolinfo = newpoolinfo;
2879
2880	spin_lock_irqsave(&conf->device_lock, flags);
2881	mddev->degraded += (raid_disks - conf->raid_disks);
2882	spin_unlock_irqrestore(&conf->device_lock, flags);
2883	conf->raid_disks = mddev->raid_disks = raid_disks;
2884	mddev->delta_disks = 0;
2885
2886	conf->last_used = 0; /* just make sure it is in-range */
2887	lower_barrier(conf);
2888
 
2889	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2890	md_wakeup_thread(mddev->thread);
2891
2892	mempool_destroy(oldpool);
2893	return 0;
2894}
2895
2896static void raid1_quiesce(struct mddev *mddev, int state)
2897{
2898	struct r1conf *conf = mddev->private;
2899
2900	switch(state) {
2901	case 2: /* wake for suspend */
2902		wake_up(&conf->wait_barrier);
2903		break;
2904	case 1:
2905		raise_barrier(conf);
2906		break;
2907	case 0:
2908		lower_barrier(conf);
2909		break;
2910	}
2911}
2912
2913static void *raid1_takeover(struct mddev *mddev)
2914{
2915	/* raid1 can take over:
2916	 *  raid5 with 2 devices, any layout or chunk size
2917	 */
2918	if (mddev->level == 5 && mddev->raid_disks == 2) {
2919		struct r1conf *conf;
2920		mddev->new_level = 1;
2921		mddev->new_layout = 0;
2922		mddev->new_chunk_sectors = 0;
2923		conf = setup_conf(mddev);
2924		if (!IS_ERR(conf))
2925			conf->barrier = 1;
 
 
 
 
2926		return conf;
2927	}
2928	return ERR_PTR(-EINVAL);
2929}
2930
2931static struct md_personality raid1_personality =
2932{
2933	.name		= "raid1",
2934	.level		= 1,
2935	.owner		= THIS_MODULE,
2936	.make_request	= make_request,
2937	.run		= run,
2938	.stop		= stop,
2939	.status		= status,
2940	.error_handler	= error,
2941	.hot_add_disk	= raid1_add_disk,
2942	.hot_remove_disk= raid1_remove_disk,
2943	.spare_active	= raid1_spare_active,
2944	.sync_request	= sync_request,
2945	.resize		= raid1_resize,
2946	.size		= raid1_size,
2947	.check_reshape	= raid1_reshape,
2948	.quiesce	= raid1_quiesce,
2949	.takeover	= raid1_takeover,
 
2950};
2951
2952static int __init raid_init(void)
2953{
2954	return register_md_personality(&raid1_personality);
2955}
2956
2957static void raid_exit(void)
2958{
2959	unregister_md_personality(&raid1_personality);
2960}
2961
2962module_init(raid_init);
2963module_exit(raid_exit);
2964MODULE_LICENSE("GPL");
2965MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2966MODULE_ALIAS("md-personality-3"); /* RAID1 */
2967MODULE_ALIAS("md-raid1");
2968MODULE_ALIAS("md-level-1");
2969
2970module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
v5.4
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * raid1.c : Multiple Devices driver for Linux
   4 *
   5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6 *
   7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8 *
   9 * RAID-1 management functions.
  10 *
  11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  12 *
  13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  15 *
  16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  17 * bitmapped intelligence in resync:
  18 *
  19 *      - bitmap marked during normal i/o
  20 *      - bitmap used to skip nondirty blocks during sync
  21 *
  22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  23 * - persistent bitmap code
 
 
 
 
 
 
 
 
 
  24 */
  25
  26#include <linux/slab.h>
  27#include <linux/delay.h>
  28#include <linux/blkdev.h>
  29#include <linux/module.h>
  30#include <linux/seq_file.h>
  31#include <linux/ratelimit.h>
  32
  33#include <trace/events/block.h>
  34
  35#include "md.h"
  36#include "raid1.h"
  37#include "md-bitmap.h"
  38
  39#define UNSUPPORTED_MDDEV_FLAGS		\
  40	((1L << MD_HAS_JOURNAL) |	\
  41	 (1L << MD_JOURNAL_CLEAN) |	\
  42	 (1L << MD_HAS_PPL) |		\
  43	 (1L << MD_HAS_MULTIPLE_PPLS))
  44
  45static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  46static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  47
  48#define raid1_log(md, fmt, args...)				\
  49	do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  50
  51#include "raid1-10.c"
  52
  53static int check_and_add_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
  54{
  55	struct wb_info *wi, *temp_wi;
  56	unsigned long flags;
  57	int ret = 0;
  58	struct mddev *mddev = rdev->mddev;
  59
  60	wi = mempool_alloc(mddev->wb_info_pool, GFP_NOIO);
  61
  62	spin_lock_irqsave(&rdev->wb_list_lock, flags);
  63	list_for_each_entry(temp_wi, &rdev->wb_list, list) {
  64		/* collision happened */
  65		if (hi > temp_wi->lo && lo < temp_wi->hi) {
  66			ret = -EBUSY;
  67			break;
  68		}
  69	}
  70
  71	if (!ret) {
  72		wi->lo = lo;
  73		wi->hi = hi;
  74		list_add(&wi->list, &rdev->wb_list);
  75	} else
  76		mempool_free(wi, mddev->wb_info_pool);
  77	spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
  78
  79	return ret;
  80}
  81
  82static void remove_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
  83{
  84	struct wb_info *wi;
  85	unsigned long flags;
  86	int found = 0;
  87	struct mddev *mddev = rdev->mddev;
  88
  89	spin_lock_irqsave(&rdev->wb_list_lock, flags);
  90	list_for_each_entry(wi, &rdev->wb_list, list)
  91		if (hi == wi->hi && lo == wi->lo) {
  92			list_del(&wi->list);
  93			mempool_free(wi, mddev->wb_info_pool);
  94			found = 1;
  95			break;
  96		}
  97
  98	if (!found)
  99		WARN(1, "The write behind IO is not recorded\n");
 100	spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
 101	wake_up(&rdev->wb_io_wait);
 102}
 103
 104/*
 105 * for resync bio, r1bio pointer can be retrieved from the per-bio
 106 * 'struct resync_pages'.
 107 */
 108static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 109{
 110	return get_resync_pages(bio)->raid_bio;
 111}
 112
 113static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 114{
 115	struct pool_info *pi = data;
 116	int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 117
 118	/* allocate a r1bio with room for raid_disks entries in the bios array */
 119	return kzalloc(size, gfp_flags);
 120}
 121
 122#define RESYNC_DEPTH 32
 
 
 
 
 
 
 123#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 124#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 125#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 126#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 127#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 128
 129static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 130{
 131	struct pool_info *pi = data;
 
 132	struct r1bio *r1_bio;
 133	struct bio *bio;
 134	int need_pages;
 135	int j;
 136	struct resync_pages *rps;
 137
 138	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 139	if (!r1_bio)
 140		return NULL;
 141
 142	rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 143			    gfp_flags);
 144	if (!rps)
 145		goto out_free_r1bio;
 146
 147	/*
 148	 * Allocate bios : 1 for reading, n-1 for writing
 149	 */
 150	for (j = pi->raid_disks ; j-- ; ) {
 151		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 152		if (!bio)
 153			goto out_free_bio;
 154		r1_bio->bios[j] = bio;
 155	}
 156	/*
 157	 * Allocate RESYNC_PAGES data pages and attach them to
 158	 * the first bio.
 159	 * If this is a user-requested check/repair, allocate
 160	 * RESYNC_PAGES for each bio.
 161	 */
 162	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 163		need_pages = pi->raid_disks;
 164	else
 165		need_pages = 1;
 166	for (j = 0; j < pi->raid_disks; j++) {
 167		struct resync_pages *rp = &rps[j];
 168
 169		bio = r1_bio->bios[j];
 
 
 
 
 170
 171		if (j < need_pages) {
 172			if (resync_alloc_pages(rp, gfp_flags))
 173				goto out_free_pages;
 174		} else {
 175			memcpy(rp, &rps[0], sizeof(*rp));
 176			resync_get_all_pages(rp);
 177		}
 178
 179		rp->raid_bio = r1_bio;
 180		bio->bi_private = rp;
 
 
 
 
 181	}
 182
 183	r1_bio->master_bio = NULL;
 184
 185	return r1_bio;
 186
 187out_free_pages:
 188	while (--j >= 0)
 189		resync_free_pages(&rps[j]);
 190
 
 191out_free_bio:
 192	while (++j < pi->raid_disks)
 193		bio_put(r1_bio->bios[j]);
 194	kfree(rps);
 195
 196out_free_r1bio:
 197	rbio_pool_free(r1_bio, data);
 198	return NULL;
 199}
 200
 201static void r1buf_pool_free(void *__r1_bio, void *data)
 202{
 203	struct pool_info *pi = data;
 204	int i;
 205	struct r1bio *r1bio = __r1_bio;
 206	struct resync_pages *rp = NULL;
 207
 208	for (i = pi->raid_disks; i--; ) {
 209		rp = get_resync_pages(r1bio->bios[i]);
 210		resync_free_pages(rp);
 
 
 
 
 
 211		bio_put(r1bio->bios[i]);
 212	}
 213
 214	/* resync pages array stored in the 1st bio's .bi_private */
 215	kfree(rp);
 216
 217	rbio_pool_free(r1bio, data);
 218}
 219
 220static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 221{
 222	int i;
 223
 224	for (i = 0; i < conf->raid_disks * 2; i++) {
 225		struct bio **bio = r1_bio->bios + i;
 226		if (!BIO_SPECIAL(*bio))
 227			bio_put(*bio);
 228		*bio = NULL;
 229	}
 230}
 231
 232static void free_r1bio(struct r1bio *r1_bio)
 233{
 234	struct r1conf *conf = r1_bio->mddev->private;
 235
 236	put_all_bios(conf, r1_bio);
 237	mempool_free(r1_bio, &conf->r1bio_pool);
 238}
 239
 240static void put_buf(struct r1bio *r1_bio)
 241{
 242	struct r1conf *conf = r1_bio->mddev->private;
 243	sector_t sect = r1_bio->sector;
 244	int i;
 245
 246	for (i = 0; i < conf->raid_disks * 2; i++) {
 247		struct bio *bio = r1_bio->bios[i];
 248		if (bio->bi_end_io)
 249			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 250	}
 251
 252	mempool_free(r1_bio, &conf->r1buf_pool);
 253
 254	lower_barrier(conf, sect);
 255}
 256
 257static void reschedule_retry(struct r1bio *r1_bio)
 258{
 259	unsigned long flags;
 260	struct mddev *mddev = r1_bio->mddev;
 261	struct r1conf *conf = mddev->private;
 262	int idx;
 263
 264	idx = sector_to_idx(r1_bio->sector);
 265	spin_lock_irqsave(&conf->device_lock, flags);
 266	list_add(&r1_bio->retry_list, &conf->retry_list);
 267	atomic_inc(&conf->nr_queued[idx]);
 268	spin_unlock_irqrestore(&conf->device_lock, flags);
 269
 270	wake_up(&conf->wait_barrier);
 271	md_wakeup_thread(mddev->thread);
 272}
 273
 274/*
 275 * raid_end_bio_io() is called when we have finished servicing a mirrored
 276 * operation and are ready to return a success/failure code to the buffer
 277 * cache layer.
 278 */
 279static void call_bio_endio(struct r1bio *r1_bio)
 280{
 281	struct bio *bio = r1_bio->master_bio;
 
 282	struct r1conf *conf = r1_bio->mddev->private;
 283
 
 
 
 
 
 
 
 
 
 284	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 285		bio->bi_status = BLK_STS_IOERR;
 286
 287	bio_endio(bio);
 288	/*
 289	 * Wake up any possible resync thread that waits for the device
 290	 * to go idle.
 291	 */
 292	allow_barrier(conf, r1_bio->sector);
 
 293}
 294
 295static void raid_end_bio_io(struct r1bio *r1_bio)
 296{
 297	struct bio *bio = r1_bio->master_bio;
 298
 299	/* if nobody has done the final endio yet, do it now */
 300	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 301		pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 302			 (bio_data_dir(bio) == WRITE) ? "write" : "read",
 303			 (unsigned long long) bio->bi_iter.bi_sector,
 304			 (unsigned long long) bio_end_sector(bio) - 1);
 
 305
 306		call_bio_endio(r1_bio);
 307	}
 308	free_r1bio(r1_bio);
 309}
 310
 311/*
 312 * Update disk head position estimator based on IRQ completion info.
 313 */
 314static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 315{
 316	struct r1conf *conf = r1_bio->mddev->private;
 317
 318	conf->mirrors[disk].head_position =
 319		r1_bio->sector + (r1_bio->sectors);
 320}
 321
 322/*
 323 * Find the disk number which triggered given bio
 324 */
 325static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 326{
 327	int mirror;
 328	struct r1conf *conf = r1_bio->mddev->private;
 329	int raid_disks = conf->raid_disks;
 330
 331	for (mirror = 0; mirror < raid_disks * 2; mirror++)
 332		if (r1_bio->bios[mirror] == bio)
 333			break;
 334
 335	BUG_ON(mirror == raid_disks * 2);
 336	update_head_pos(mirror, r1_bio);
 337
 338	return mirror;
 339}
 340
 341static void raid1_end_read_request(struct bio *bio)
 342{
 343	int uptodate = !bio->bi_status;
 344	struct r1bio *r1_bio = bio->bi_private;
 
 345	struct r1conf *conf = r1_bio->mddev->private;
 346	struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 347
 
 348	/*
 349	 * this branch is our 'one mirror IO has finished' event handler:
 350	 */
 351	update_head_pos(r1_bio->read_disk, r1_bio);
 352
 353	if (uptodate)
 354		set_bit(R1BIO_Uptodate, &r1_bio->state);
 355	else if (test_bit(FailFast, &rdev->flags) &&
 356		 test_bit(R1BIO_FailFast, &r1_bio->state))
 357		/* This was a fail-fast read so we definitely
 358		 * want to retry */
 359		;
 360	else {
 361		/* If all other devices have failed, we want to return
 362		 * the error upwards rather than fail the last device.
 363		 * Here we redefine "uptodate" to mean "Don't want to retry"
 364		 */
 365		unsigned long flags;
 366		spin_lock_irqsave(&conf->device_lock, flags);
 367		if (r1_bio->mddev->degraded == conf->raid_disks ||
 368		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 369		     test_bit(In_sync, &rdev->flags)))
 370			uptodate = 1;
 371		spin_unlock_irqrestore(&conf->device_lock, flags);
 372	}
 373
 374	if (uptodate) {
 375		raid_end_bio_io(r1_bio);
 376		rdev_dec_pending(rdev, conf->mddev);
 377	} else {
 378		/*
 379		 * oops, read error:
 380		 */
 381		char b[BDEVNAME_SIZE];
 382		pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 383				   mdname(conf->mddev),
 384				   bdevname(rdev->bdev, b),
 385				   (unsigned long long)r1_bio->sector);
 
 
 
 386		set_bit(R1BIO_ReadError, &r1_bio->state);
 387		reschedule_retry(r1_bio);
 388		/* don't drop the reference on read_disk yet */
 389	}
 
 
 390}
 391
 392static void close_write(struct r1bio *r1_bio)
 393{
 394	/* it really is the end of this request */
 395	if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 396		bio_free_pages(r1_bio->behind_master_bio);
 397		bio_put(r1_bio->behind_master_bio);
 398		r1_bio->behind_master_bio = NULL;
 
 
 
 399	}
 400	/* clear the bitmap if all writes complete successfully */
 401	md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 402			   r1_bio->sectors,
 403			   !test_bit(R1BIO_Degraded, &r1_bio->state),
 404			   test_bit(R1BIO_BehindIO, &r1_bio->state));
 405	md_write_end(r1_bio->mddev);
 406}
 407
 408static void r1_bio_write_done(struct r1bio *r1_bio)
 409{
 410	if (!atomic_dec_and_test(&r1_bio->remaining))
 411		return;
 412
 413	if (test_bit(R1BIO_WriteError, &r1_bio->state))
 414		reschedule_retry(r1_bio);
 415	else {
 416		close_write(r1_bio);
 417		if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 418			reschedule_retry(r1_bio);
 419		else
 420			raid_end_bio_io(r1_bio);
 421	}
 422}
 423
 424static void raid1_end_write_request(struct bio *bio)
 425{
 
 426	struct r1bio *r1_bio = bio->bi_private;
 427	int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 428	struct r1conf *conf = r1_bio->mddev->private;
 429	struct bio *to_put = NULL;
 430	int mirror = find_bio_disk(r1_bio, bio);
 431	struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 432	bool discard_error;
 433
 434	discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 435
 436	/*
 437	 * 'one mirror IO has finished' event handler:
 438	 */
 439	if (bio->bi_status && !discard_error) {
 440		set_bit(WriteErrorSeen,	&rdev->flags);
 441		if (!test_and_set_bit(WantReplacement, &rdev->flags))
 
 
 442			set_bit(MD_RECOVERY_NEEDED, &
 443				conf->mddev->recovery);
 444
 445		if (test_bit(FailFast, &rdev->flags) &&
 446		    (bio->bi_opf & MD_FAILFAST) &&
 447		    /* We never try FailFast to WriteMostly devices */
 448		    !test_bit(WriteMostly, &rdev->flags)) {
 449			md_error(r1_bio->mddev, rdev);
 450		}
 451
 452		/*
 453		 * When the device is faulty, it is not necessary to
 454		 * handle write error.
 455		 * For failfast, this is the only remaining device,
 456		 * We need to retry the write without FailFast.
 457		 */
 458		if (!test_bit(Faulty, &rdev->flags))
 459			set_bit(R1BIO_WriteError, &r1_bio->state);
 460		else {
 461			/* Finished with this branch */
 462			r1_bio->bios[mirror] = NULL;
 463			to_put = bio;
 464		}
 465	} else {
 466		/*
 467		 * Set R1BIO_Uptodate in our master bio, so that we
 468		 * will return a good error code for to the higher
 469		 * levels even if IO on some other mirrored buffer
 470		 * fails.
 471		 *
 472		 * The 'master' represents the composite IO operation
 473		 * to user-side. So if something waits for IO, then it
 474		 * will wait for the 'master' bio.
 475		 */
 476		sector_t first_bad;
 477		int bad_sectors;
 478
 479		r1_bio->bios[mirror] = NULL;
 480		to_put = bio;
 481		/*
 482		 * Do not set R1BIO_Uptodate if the current device is
 483		 * rebuilding or Faulty. This is because we cannot use
 484		 * such device for properly reading the data back (we could
 485		 * potentially use it, if the current write would have felt
 486		 * before rdev->recovery_offset, but for simplicity we don't
 487		 * check this here.
 488		 */
 489		if (test_bit(In_sync, &rdev->flags) &&
 490		    !test_bit(Faulty, &rdev->flags))
 491			set_bit(R1BIO_Uptodate, &r1_bio->state);
 492
 493		/* Maybe we can clear some bad blocks. */
 494		if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 495				&first_bad, &bad_sectors) && !discard_error) {
 
 496			r1_bio->bios[mirror] = IO_MADE_GOOD;
 497			set_bit(R1BIO_MadeGood, &r1_bio->state);
 498		}
 499	}
 500
 501	if (behind) {
 502		if (test_bit(WBCollisionCheck, &rdev->flags)) {
 503			sector_t lo = r1_bio->sector;
 504			sector_t hi = r1_bio->sector + r1_bio->sectors;
 505
 506			remove_wb(rdev, lo, hi);
 507		}
 508		if (test_bit(WriteMostly, &rdev->flags))
 509			atomic_dec(&r1_bio->behind_remaining);
 510
 511		/*
 512		 * In behind mode, we ACK the master bio once the I/O
 513		 * has safely reached all non-writemostly
 514		 * disks. Setting the Returned bit ensures that this
 515		 * gets done only once -- we don't ever want to return
 516		 * -EIO here, instead we'll wait
 517		 */
 518		if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 519		    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 520			/* Maybe we can return now */
 521			if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 522				struct bio *mbio = r1_bio->master_bio;
 523				pr_debug("raid1: behind end write sectors"
 524					 " %llu-%llu\n",
 525					 (unsigned long long) mbio->bi_iter.bi_sector,
 526					 (unsigned long long) bio_end_sector(mbio) - 1);
 
 527				call_bio_endio(r1_bio);
 528			}
 529		}
 530	}
 531	if (r1_bio->bios[mirror] == NULL)
 532		rdev_dec_pending(rdev, conf->mddev);
 
 533
 534	/*
 535	 * Let's see if all mirrored write operations have finished
 536	 * already.
 537	 */
 538	r1_bio_write_done(r1_bio);
 539
 540	if (to_put)
 541		bio_put(to_put);
 542}
 543
 544static sector_t align_to_barrier_unit_end(sector_t start_sector,
 545					  sector_t sectors)
 546{
 547	sector_t len;
 548
 549	WARN_ON(sectors == 0);
 550	/*
 551	 * len is the number of sectors from start_sector to end of the
 552	 * barrier unit which start_sector belongs to.
 553	 */
 554	len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 555	      start_sector;
 556
 557	if (len > sectors)
 558		len = sectors;
 559
 560	return len;
 561}
 562
 563/*
 564 * This routine returns the disk from which the requested read should
 565 * be done. There is a per-array 'next expected sequential IO' sector
 566 * number - if this matches on the next IO then we use the last disk.
 567 * There is also a per-disk 'last know head position' sector that is
 568 * maintained from IRQ contexts, both the normal and the resync IO
 569 * completion handlers update this position correctly. If there is no
 570 * perfect sequential match then we pick the disk whose head is closest.
 571 *
 572 * If there are 2 mirrors in the same 2 devices, performance degrades
 573 * because position is mirror, not device based.
 574 *
 575 * The rdev for the device selected will have nr_pending incremented.
 576 */
 577static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 578{
 579	const sector_t this_sector = r1_bio->sector;
 580	int sectors;
 581	int best_good_sectors;
 582	int best_disk, best_dist_disk, best_pending_disk;
 583	int has_nonrot_disk;
 584	int disk;
 585	sector_t best_dist;
 586	unsigned int min_pending;
 587	struct md_rdev *rdev;
 588	int choose_first;
 589	int choose_next_idle;
 590
 591	rcu_read_lock();
 592	/*
 593	 * Check if we can balance. We can balance on the whole
 594	 * device if no resync is going on, or below the resync window.
 595	 * We take the first readable disk when above the resync window.
 596	 */
 597 retry:
 598	sectors = r1_bio->sectors;
 599	best_disk = -1;
 600	best_dist_disk = -1;
 601	best_dist = MaxSector;
 602	best_pending_disk = -1;
 603	min_pending = UINT_MAX;
 604	best_good_sectors = 0;
 605	has_nonrot_disk = 0;
 606	choose_next_idle = 0;
 607	clear_bit(R1BIO_FailFast, &r1_bio->state);
 608
 609	if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 610	    (mddev_is_clustered(conf->mddev) &&
 611	    md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 612		    this_sector + sectors)))
 613		choose_first = 1;
 614	else
 
 615		choose_first = 0;
 
 
 616
 617	for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 618		sector_t dist;
 619		sector_t first_bad;
 620		int bad_sectors;
 621		unsigned int pending;
 622		bool nonrot;
 
 
 623
 624		rdev = rcu_dereference(conf->mirrors[disk].rdev);
 625		if (r1_bio->bios[disk] == IO_BLOCKED
 626		    || rdev == NULL
 
 627		    || test_bit(Faulty, &rdev->flags))
 628			continue;
 629		if (!test_bit(In_sync, &rdev->flags) &&
 630		    rdev->recovery_offset < this_sector + sectors)
 631			continue;
 632		if (test_bit(WriteMostly, &rdev->flags)) {
 633			/* Don't balance among write-mostly, just
 634			 * use the first as a last resort */
 635			if (best_dist_disk < 0) {
 636				if (is_badblock(rdev, this_sector, sectors,
 637						&first_bad, &bad_sectors)) {
 638					if (first_bad <= this_sector)
 639						/* Cannot use this */
 640						continue;
 641					best_good_sectors = first_bad - this_sector;
 642				} else
 643					best_good_sectors = sectors;
 644				best_dist_disk = disk;
 645				best_pending_disk = disk;
 646			}
 647			continue;
 648		}
 649		/* This is a reasonable device to use.  It might
 650		 * even be best.
 651		 */
 652		if (is_badblock(rdev, this_sector, sectors,
 653				&first_bad, &bad_sectors)) {
 654			if (best_dist < MaxSector)
 655				/* already have a better device */
 656				continue;
 657			if (first_bad <= this_sector) {
 658				/* cannot read here. If this is the 'primary'
 659				 * device, then we must not read beyond
 660				 * bad_sectors from another device..
 661				 */
 662				bad_sectors -= (this_sector - first_bad);
 663				if (choose_first && sectors > bad_sectors)
 664					sectors = bad_sectors;
 665				if (best_good_sectors > sectors)
 666					best_good_sectors = sectors;
 667
 668			} else {
 669				sector_t good_sectors = first_bad - this_sector;
 670				if (good_sectors > best_good_sectors) {
 671					best_good_sectors = good_sectors;
 672					best_disk = disk;
 673				}
 674				if (choose_first)
 675					break;
 676			}
 677			continue;
 678		} else {
 679			if ((sectors > best_good_sectors) && (best_disk >= 0))
 680				best_disk = -1;
 681			best_good_sectors = sectors;
 682		}
 683
 684		if (best_disk >= 0)
 685			/* At least two disks to choose from so failfast is OK */
 686			set_bit(R1BIO_FailFast, &r1_bio->state);
 687
 688		nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 689		has_nonrot_disk |= nonrot;
 690		pending = atomic_read(&rdev->nr_pending);
 691		dist = abs(this_sector - conf->mirrors[disk].head_position);
 692		if (choose_first) {
 693			best_disk = disk;
 694			break;
 695		}
 696		/* Don't change to another disk for sequential reads */
 697		if (conf->mirrors[disk].next_seq_sect == this_sector
 698		    || dist == 0) {
 699			int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 700			struct raid1_info *mirror = &conf->mirrors[disk];
 701
 702			best_disk = disk;
 703			/*
 704			 * If buffered sequential IO size exceeds optimal
 705			 * iosize, check if there is idle disk. If yes, choose
 706			 * the idle disk. read_balance could already choose an
 707			 * idle disk before noticing it's a sequential IO in
 708			 * this disk. This doesn't matter because this disk
 709			 * will idle, next time it will be utilized after the
 710			 * first disk has IO size exceeds optimal iosize. In
 711			 * this way, iosize of the first disk will be optimal
 712			 * iosize at least. iosize of the second disk might be
 713			 * small, but not a big deal since when the second disk
 714			 * starts IO, the first disk is likely still busy.
 715			 */
 716			if (nonrot && opt_iosize > 0 &&
 717			    mirror->seq_start != MaxSector &&
 718			    mirror->next_seq_sect > opt_iosize &&
 719			    mirror->next_seq_sect - opt_iosize >=
 720			    mirror->seq_start) {
 721				choose_next_idle = 1;
 722				continue;
 723			}
 724			break;
 725		}
 726
 727		if (choose_next_idle)
 728			continue;
 729
 730		if (min_pending > pending) {
 731			min_pending = pending;
 732			best_pending_disk = disk;
 733		}
 734
 735		if (dist < best_dist) {
 736			best_dist = dist;
 737			best_dist_disk = disk;
 738		}
 739	}
 740
 741	/*
 742	 * If all disks are rotational, choose the closest disk. If any disk is
 743	 * non-rotational, choose the disk with less pending request even the
 744	 * disk is rotational, which might/might not be optimal for raids with
 745	 * mixed ratation/non-rotational disks depending on workload.
 746	 */
 747	if (best_disk == -1) {
 748		if (has_nonrot_disk || min_pending == 0)
 749			best_disk = best_pending_disk;
 750		else
 751			best_disk = best_dist_disk;
 752	}
 753
 754	if (best_disk >= 0) {
 755		rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 756		if (!rdev)
 757			goto retry;
 758		atomic_inc(&rdev->nr_pending);
 
 
 
 
 
 
 
 759		sectors = best_good_sectors;
 760
 761		if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 762			conf->mirrors[best_disk].seq_start = this_sector;
 763
 764		conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 765	}
 766	rcu_read_unlock();
 767	*max_sectors = sectors;
 768
 769	return best_disk;
 770}
 771
 772static int raid1_congested(struct mddev *mddev, int bits)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 773{
 774	struct r1conf *conf = mddev->private;
 775	int i, ret = 0;
 776
 777	if ((bits & (1 << WB_async_congested)) &&
 778	    conf->pending_count >= max_queued_requests)
 779		return 1;
 780
 781	rcu_read_lock();
 782	for (i = 0; i < conf->raid_disks * 2; i++) {
 783		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 784		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 785			struct request_queue *q = bdev_get_queue(rdev->bdev);
 786
 787			BUG_ON(!q);
 788
 789			/* Note the '|| 1' - when read_balance prefers
 790			 * non-congested targets, it can be removed
 791			 */
 792			if ((bits & (1 << WB_async_congested)) || 1)
 793				ret |= bdi_congested(q->backing_dev_info, bits);
 794			else
 795				ret &= bdi_congested(q->backing_dev_info, bits);
 796		}
 797	}
 798	rcu_read_unlock();
 799	return ret;
 800}
 
 801
 802static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 803{
 804	/* flush any pending bitmap writes to disk before proceeding w/ I/O */
 805	md_bitmap_unplug(conf->mddev->bitmap);
 806	wake_up(&conf->wait_barrier);
 807
 808	while (bio) { /* submit pending writes */
 809		struct bio *next = bio->bi_next;
 810		struct md_rdev *rdev = (void *)bio->bi_disk;
 811		bio->bi_next = NULL;
 812		bio_set_dev(bio, rdev->bdev);
 813		if (test_bit(Faulty, &rdev->flags)) {
 814			bio_io_error(bio);
 815		} else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 816				    !blk_queue_discard(bio->bi_disk->queue)))
 817			/* Just ignore it */
 818			bio_endio(bio);
 819		else
 820			generic_make_request(bio);
 821		bio = next;
 822	}
 823}
 824
 825static void flush_pending_writes(struct r1conf *conf)
 826{
 827	/* Any writes that have been queued but are awaiting
 828	 * bitmap updates get flushed here.
 829	 */
 830	spin_lock_irq(&conf->device_lock);
 831
 832	if (conf->pending_bio_list.head) {
 833		struct blk_plug plug;
 834		struct bio *bio;
 835
 836		bio = bio_list_get(&conf->pending_bio_list);
 837		conf->pending_count = 0;
 838		spin_unlock_irq(&conf->device_lock);
 
 
 
 
 839
 840		/*
 841		 * As this is called in a wait_event() loop (see freeze_array),
 842		 * current->state might be TASK_UNINTERRUPTIBLE which will
 843		 * cause a warning when we prepare to wait again.  As it is
 844		 * rare that this path is taken, it is perfectly safe to force
 845		 * us to go around the wait_event() loop again, so the warning
 846		 * is a false-positive.  Silence the warning by resetting
 847		 * thread state
 848		 */
 849		__set_current_state(TASK_RUNNING);
 850		blk_start_plug(&plug);
 851		flush_bio_list(conf, bio);
 852		blk_finish_plug(&plug);
 853	} else
 854		spin_unlock_irq(&conf->device_lock);
 855}
 856
 857/* Barriers....
 858 * Sometimes we need to suspend IO while we do something else,
 859 * either some resync/recovery, or reconfigure the array.
 860 * To do this we raise a 'barrier'.
 861 * The 'barrier' is a counter that can be raised multiple times
 862 * to count how many activities are happening which preclude
 863 * normal IO.
 864 * We can only raise the barrier if there is no pending IO.
 865 * i.e. if nr_pending == 0.
 866 * We choose only to raise the barrier if no-one is waiting for the
 867 * barrier to go down.  This means that as soon as an IO request
 868 * is ready, no other operations which require a barrier will start
 869 * until the IO request has had a chance.
 870 *
 871 * So: regular IO calls 'wait_barrier'.  When that returns there
 872 *    is no backgroup IO happening,  It must arrange to call
 873 *    allow_barrier when it has finished its IO.
 874 * backgroup IO calls must call raise_barrier.  Once that returns
 875 *    there is no normal IO happeing.  It must arrange to call
 876 *    lower_barrier when the particular background IO completes.
 877 *
 878 * If resync/recovery is interrupted, returns -EINTR;
 879 * Otherwise, returns 0.
 880 */
 881static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 
 
 882{
 883	int idx = sector_to_idx(sector_nr);
 884
 885	spin_lock_irq(&conf->resync_lock);
 886
 887	/* Wait until no block IO is waiting */
 888	wait_event_lock_irq(conf->wait_barrier,
 889			    !atomic_read(&conf->nr_waiting[idx]),
 890			    conf->resync_lock);
 891
 892	/* block any new IO from starting */
 893	atomic_inc(&conf->barrier[idx]);
 894	/*
 895	 * In raise_barrier() we firstly increase conf->barrier[idx] then
 896	 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 897	 * increase conf->nr_pending[idx] then check conf->barrier[idx].
 898	 * A memory barrier here to make sure conf->nr_pending[idx] won't
 899	 * be fetched before conf->barrier[idx] is increased. Otherwise
 900	 * there will be a race between raise_barrier() and _wait_barrier().
 901	 */
 902	smp_mb__after_atomic();
 903
 904	/* For these conditions we must wait:
 905	 * A: while the array is in frozen state
 906	 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 907	 *    existing in corresponding I/O barrier bucket.
 908	 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 909	 *    max resync count which allowed on current I/O barrier bucket.
 910	 */
 911	wait_event_lock_irq(conf->wait_barrier,
 912			    (!conf->array_frozen &&
 913			     !atomic_read(&conf->nr_pending[idx]) &&
 914			     atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 915				test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 916			    conf->resync_lock);
 917
 918	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 919		atomic_dec(&conf->barrier[idx]);
 920		spin_unlock_irq(&conf->resync_lock);
 921		wake_up(&conf->wait_barrier);
 922		return -EINTR;
 923	}
 924
 925	atomic_inc(&conf->nr_sync_pending);
 926	spin_unlock_irq(&conf->resync_lock);
 927
 928	return 0;
 929}
 930
 931static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 932{
 933	int idx = sector_to_idx(sector_nr);
 934
 935	BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 936
 937	atomic_dec(&conf->barrier[idx]);
 938	atomic_dec(&conf->nr_sync_pending);
 939	wake_up(&conf->wait_barrier);
 940}
 941
 942static void _wait_barrier(struct r1conf *conf, int idx)
 943{
 944	/*
 945	 * We need to increase conf->nr_pending[idx] very early here,
 946	 * then raise_barrier() can be blocked when it waits for
 947	 * conf->nr_pending[idx] to be 0. Then we can avoid holding
 948	 * conf->resync_lock when there is no barrier raised in same
 949	 * barrier unit bucket. Also if the array is frozen, I/O
 950	 * should be blocked until array is unfrozen.
 951	 */
 952	atomic_inc(&conf->nr_pending[idx]);
 953	/*
 954	 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 955	 * check conf->barrier[idx]. In raise_barrier() we firstly increase
 956	 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 957	 * barrier is necessary here to make sure conf->barrier[idx] won't be
 958	 * fetched before conf->nr_pending[idx] is increased. Otherwise there
 959	 * will be a race between _wait_barrier() and raise_barrier().
 960	 */
 961	smp_mb__after_atomic();
 962
 963	/*
 964	 * Don't worry about checking two atomic_t variables at same time
 965	 * here. If during we check conf->barrier[idx], the array is
 966	 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 967	 * 0, it is safe to return and make the I/O continue. Because the
 968	 * array is frozen, all I/O returned here will eventually complete
 969	 * or be queued, no race will happen. See code comment in
 970	 * frozen_array().
 971	 */
 972	if (!READ_ONCE(conf->array_frozen) &&
 973	    !atomic_read(&conf->barrier[idx]))
 974		return;
 975
 976	/*
 977	 * After holding conf->resync_lock, conf->nr_pending[idx]
 978	 * should be decreased before waiting for barrier to drop.
 979	 * Otherwise, we may encounter a race condition because
 980	 * raise_barrer() might be waiting for conf->nr_pending[idx]
 981	 * to be 0 at same time.
 982	 */
 983	spin_lock_irq(&conf->resync_lock);
 984	atomic_inc(&conf->nr_waiting[idx]);
 985	atomic_dec(&conf->nr_pending[idx]);
 986	/*
 987	 * In case freeze_array() is waiting for
 988	 * get_unqueued_pending() == extra
 989	 */
 990	wake_up(&conf->wait_barrier);
 991	/* Wait for the barrier in same barrier unit bucket to drop. */
 992	wait_event_lock_irq(conf->wait_barrier,
 993			    !conf->array_frozen &&
 994			     !atomic_read(&conf->barrier[idx]),
 995			    conf->resync_lock);
 996	atomic_inc(&conf->nr_pending[idx]);
 997	atomic_dec(&conf->nr_waiting[idx]);
 
 
 
 
 
 
 
 998	spin_unlock_irq(&conf->resync_lock);
 999}
1000
1001static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
1002{
1003	int idx = sector_to_idx(sector_nr);
1004
1005	/*
1006	 * Very similar to _wait_barrier(). The difference is, for read
1007	 * I/O we don't need wait for sync I/O, but if the whole array
1008	 * is frozen, the read I/O still has to wait until the array is
1009	 * unfrozen. Since there is no ordering requirement with
1010	 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1011	 */
1012	atomic_inc(&conf->nr_pending[idx]);
1013
1014	if (!READ_ONCE(conf->array_frozen))
1015		return;
1016
1017	spin_lock_irq(&conf->resync_lock);
1018	atomic_inc(&conf->nr_waiting[idx]);
1019	atomic_dec(&conf->nr_pending[idx]);
1020	/*
1021	 * In case freeze_array() is waiting for
1022	 * get_unqueued_pending() == extra
1023	 */
1024	wake_up(&conf->wait_barrier);
1025	/* Wait for array to be unfrozen */
1026	wait_event_lock_irq(conf->wait_barrier,
1027			    !conf->array_frozen,
1028			    conf->resync_lock);
1029	atomic_inc(&conf->nr_pending[idx]);
1030	atomic_dec(&conf->nr_waiting[idx]);
1031	spin_unlock_irq(&conf->resync_lock);
1032}
1033
1034static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1035{
1036	int idx = sector_to_idx(sector_nr);
1037
1038	_wait_barrier(conf, idx);
1039}
1040
1041static void _allow_barrier(struct r1conf *conf, int idx)
1042{
1043	atomic_dec(&conf->nr_pending[idx]);
1044	wake_up(&conf->wait_barrier);
1045}
1046
1047static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1048{
1049	int idx = sector_to_idx(sector_nr);
1050
1051	_allow_barrier(conf, idx);
1052}
1053
1054/* conf->resync_lock should be held */
1055static int get_unqueued_pending(struct r1conf *conf)
1056{
1057	int idx, ret;
1058
1059	ret = atomic_read(&conf->nr_sync_pending);
1060	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1061		ret += atomic_read(&conf->nr_pending[idx]) -
1062			atomic_read(&conf->nr_queued[idx]);
1063
1064	return ret;
1065}
1066
1067static void freeze_array(struct r1conf *conf, int extra)
1068{
1069	/* Stop sync I/O and normal I/O and wait for everything to
1070	 * go quiet.
1071	 * This is called in two situations:
1072	 * 1) management command handlers (reshape, remove disk, quiesce).
1073	 * 2) one normal I/O request failed.
1074
1075	 * After array_frozen is set to 1, new sync IO will be blocked at
1076	 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1077	 * or wait_read_barrier(). The flying I/Os will either complete or be
1078	 * queued. When everything goes quite, there are only queued I/Os left.
1079
1080	 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1081	 * barrier bucket index which this I/O request hits. When all sync and
1082	 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1083	 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1084	 * in handle_read_error(), we may call freeze_array() before trying to
1085	 * fix the read error. In this case, the error read I/O is not queued,
1086	 * so get_unqueued_pending() == 1.
1087	 *
1088	 * Therefore before this function returns, we need to wait until
1089	 * get_unqueued_pendings(conf) gets equal to extra. For
1090	 * normal I/O context, extra is 1, in rested situations extra is 0.
1091	 */
1092	spin_lock_irq(&conf->resync_lock);
1093	conf->array_frozen = 1;
1094	raid1_log(conf->mddev, "wait freeze");
1095	wait_event_lock_irq_cmd(
1096		conf->wait_barrier,
1097		get_unqueued_pending(conf) == extra,
1098		conf->resync_lock,
1099		flush_pending_writes(conf));
1100	spin_unlock_irq(&conf->resync_lock);
1101}
1102static void unfreeze_array(struct r1conf *conf)
1103{
1104	/* reverse the effect of the freeze */
1105	spin_lock_irq(&conf->resync_lock);
1106	conf->array_frozen = 0;
 
 
1107	spin_unlock_irq(&conf->resync_lock);
1108	wake_up(&conf->wait_barrier);
1109}
1110
1111static void alloc_behind_master_bio(struct r1bio *r1_bio,
1112					   struct bio *bio)
 
 
1113{
1114	int size = bio->bi_iter.bi_size;
1115	unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1116	int i = 0;
1117	struct bio *behind_bio = NULL;
1118
1119	behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1120	if (!behind_bio)
1121		return;
1122
1123	/* discard op, we don't support writezero/writesame yet */
1124	if (!bio_has_data(bio)) {
1125		behind_bio->bi_iter.bi_size = size;
1126		goto skip_copy;
1127	}
1128
1129	behind_bio->bi_write_hint = bio->bi_write_hint;
1130
1131	while (i < vcnt && size) {
1132		struct page *page;
1133		int len = min_t(int, PAGE_SIZE, size);
1134
1135		page = alloc_page(GFP_NOIO);
1136		if (unlikely(!page))
1137			goto free_pages;
1138
1139		bio_add_page(behind_bio, page, len, 0);
1140
1141		size -= len;
1142		i++;
1143	}
1144
1145	bio_copy_data(behind_bio, bio);
1146skip_copy:
1147	r1_bio->behind_master_bio = behind_bio;
1148	set_bit(R1BIO_BehindIO, &r1_bio->state);
1149
1150	return;
1151
1152free_pages:
1153	pr_debug("%dB behind alloc failed, doing sync I/O\n",
1154		 bio->bi_iter.bi_size);
1155	bio_free_pages(behind_bio);
1156	bio_put(behind_bio);
1157}
1158
1159struct raid1_plug_cb {
1160	struct blk_plug_cb	cb;
1161	struct bio_list		pending;
1162	int			pending_cnt;
1163};
1164
1165static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1166{
1167	struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1168						  cb);
1169	struct mddev *mddev = plug->cb.data;
1170	struct r1conf *conf = mddev->private;
1171	struct bio *bio;
1172
1173	if (from_schedule || current->bio_list) {
1174		spin_lock_irq(&conf->device_lock);
1175		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1176		conf->pending_count += plug->pending_cnt;
1177		spin_unlock_irq(&conf->device_lock);
1178		wake_up(&conf->wait_barrier);
1179		md_wakeup_thread(mddev->thread);
1180		kfree(plug);
1181		return;
1182	}
1183
1184	/* we aren't scheduling, so we can do the write-out directly. */
1185	bio = bio_list_get(&plug->pending);
1186	flush_bio_list(conf, bio);
1187	kfree(plug);
1188}
1189
1190static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1191{
1192	r1_bio->master_bio = bio;
1193	r1_bio->sectors = bio_sectors(bio);
1194	r1_bio->state = 0;
1195	r1_bio->mddev = mddev;
1196	r1_bio->sector = bio->bi_iter.bi_sector;
1197}
1198
1199static inline struct r1bio *
1200alloc_r1bio(struct mddev *mddev, struct bio *bio)
1201{
1202	struct r1conf *conf = mddev->private;
 
1203	struct r1bio *r1_bio;
1204
1205	r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1206	/* Ensure no bio records IO_BLOCKED */
1207	memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1208	init_r1bio(r1_bio, mddev, bio);
1209	return r1_bio;
1210}
1211
1212static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1213			       int max_read_sectors, struct r1bio *r1_bio)
1214{
1215	struct r1conf *conf = mddev->private;
1216	struct raid1_info *mirror;
1217	struct bio *read_bio;
1218	struct bitmap *bitmap = mddev->bitmap;
1219	const int op = bio_op(bio);
1220	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
 
 
 
 
 
 
1221	int max_sectors;
1222	int rdisk;
1223	bool print_msg = !!r1_bio;
1224	char b[BDEVNAME_SIZE];
1225
1226	/*
1227	 * If r1_bio is set, we are blocking the raid1d thread
1228	 * so there is a tiny risk of deadlock.  So ask for
1229	 * emergency memory if needed.
1230	 */
1231	gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1232
1233	if (print_msg) {
1234		/* Need to get the block device name carefully */
1235		struct md_rdev *rdev;
1236		rcu_read_lock();
1237		rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1238		if (rdev)
1239			bdevname(rdev->bdev, b);
1240		else
1241			strcpy(b, "???");
1242		rcu_read_unlock();
 
 
 
 
 
 
 
 
 
 
1243	}
1244
1245	/*
1246	 * Still need barrier for READ in case that whole
1247	 * array is frozen.
1248	 */
1249	wait_read_barrier(conf, bio->bi_iter.bi_sector);
1250
1251	if (!r1_bio)
1252		r1_bio = alloc_r1bio(mddev, bio);
1253	else
1254		init_r1bio(r1_bio, mddev, bio);
1255	r1_bio->sectors = max_read_sectors;
1256
1257	/*
1258	 * make_request() can abort the operation when read-ahead is being
1259	 * used and no empty request is available.
 
1260	 */
1261	rdisk = read_balance(conf, r1_bio, &max_sectors);
1262
1263	if (rdisk < 0) {
1264		/* couldn't find anywhere to read from */
1265		if (print_msg) {
1266			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1267					    mdname(mddev),
1268					    b,
1269					    (unsigned long long)r1_bio->sector);
1270		}
1271		raid_end_bio_io(r1_bio);
1272		return;
1273	}
1274	mirror = conf->mirrors + rdisk;
1275
1276	if (print_msg)
1277		pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1278				    mdname(mddev),
1279				    (unsigned long long)r1_bio->sector,
1280				    bdevname(mirror->rdev->bdev, b));
 
 
 
 
1281
1282	if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1283	    bitmap) {
1284		/*
1285		 * Reading from a write-mostly device must take care not to
1286		 * over-take any writes that are 'behind'
1287		 */
1288		raid1_log(mddev, "wait behind writes");
1289		wait_event(bitmap->behind_wait,
1290			   atomic_read(&bitmap->behind_writes) == 0);
1291	}
1292
1293	if (max_sectors < bio_sectors(bio)) {
1294		struct bio *split = bio_split(bio, max_sectors,
1295					      gfp, &conf->bio_split);
1296		bio_chain(split, bio);
1297		generic_make_request(bio);
1298		bio = split;
1299		r1_bio->master_bio = bio;
1300		r1_bio->sectors = max_sectors;
1301	}
1302
1303	r1_bio->read_disk = rdisk;
 
 
 
 
 
1304
1305	read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
 
 
 
 
 
 
 
 
 
1306
1307	r1_bio->bios[rdisk] = read_bio;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1308
1309	read_bio->bi_iter.bi_sector = r1_bio->sector +
1310		mirror->rdev->data_offset;
1311	bio_set_dev(read_bio, mirror->rdev->bdev);
1312	read_bio->bi_end_io = raid1_end_read_request;
1313	bio_set_op_attrs(read_bio, op, do_sync);
1314	if (test_bit(FailFast, &mirror->rdev->flags) &&
1315	    test_bit(R1BIO_FailFast, &r1_bio->state))
1316	        read_bio->bi_opf |= MD_FAILFAST;
1317	read_bio->bi_private = r1_bio;
 
 
 
 
 
 
1318
1319	if (mddev->gendisk)
1320	        trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1321				disk_devt(mddev->gendisk), r1_bio->sector);
1322
1323	generic_make_request(read_bio);
1324}
1325
1326static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1327				int max_write_sectors)
1328{
1329	struct r1conf *conf = mddev->private;
1330	struct r1bio *r1_bio;
1331	int i, disks;
1332	struct bitmap *bitmap = mddev->bitmap;
1333	unsigned long flags;
1334	struct md_rdev *blocked_rdev;
1335	struct blk_plug_cb *cb;
1336	struct raid1_plug_cb *plug = NULL;
1337	int first_clone;
1338	int max_sectors;
1339
1340	if (mddev_is_clustered(mddev) &&
1341	     md_cluster_ops->area_resyncing(mddev, WRITE,
1342		     bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1343
1344		DEFINE_WAIT(w);
1345		for (;;) {
1346			prepare_to_wait(&conf->wait_barrier,
1347					&w, TASK_IDLE);
1348			if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1349							bio->bi_iter.bi_sector,
1350							bio_end_sector(bio)))
1351				break;
1352			schedule();
1353		}
1354		finish_wait(&conf->wait_barrier, &w);
1355	}
1356
1357	/*
1358	 * Register the new request and wait if the reconstruction
1359	 * thread has put up a bar for new requests.
1360	 * Continue immediately if no resync is active currently.
1361	 */
1362	wait_barrier(conf, bio->bi_iter.bi_sector);
1363
1364	r1_bio = alloc_r1bio(mddev, bio);
1365	r1_bio->sectors = max_write_sectors;
1366
1367	if (conf->pending_count >= max_queued_requests) {
1368		md_wakeup_thread(mddev->thread);
1369		raid1_log(mddev, "wait queued");
1370		wait_event(conf->wait_barrier,
1371			   conf->pending_count < max_queued_requests);
1372	}
1373	/* first select target devices under rcu_lock and
1374	 * inc refcount on their rdev.  Record them by setting
1375	 * bios[x] to bio
1376	 * If there are known/acknowledged bad blocks on any device on
1377	 * which we have seen a write error, we want to avoid writing those
1378	 * blocks.
1379	 * This potentially requires several writes to write around
1380	 * the bad blocks.  Each set of writes gets it's own r1bio
1381	 * with a set of bios attached.
1382	 */
1383
1384	disks = conf->raid_disks * 2;
1385 retry_write:
1386	blocked_rdev = NULL;
1387	rcu_read_lock();
1388	max_sectors = r1_bio->sectors;
1389	for (i = 0;  i < disks; i++) {
1390		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1391		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1392			atomic_inc(&rdev->nr_pending);
1393			blocked_rdev = rdev;
1394			break;
1395		}
1396		r1_bio->bios[i] = NULL;
1397		if (!rdev || test_bit(Faulty, &rdev->flags)) {
 
1398			if (i < conf->raid_disks)
1399				set_bit(R1BIO_Degraded, &r1_bio->state);
1400			continue;
1401		}
1402
1403		atomic_inc(&rdev->nr_pending);
1404		if (test_bit(WriteErrorSeen, &rdev->flags)) {
1405			sector_t first_bad;
1406			int bad_sectors;
1407			int is_bad;
1408
1409			is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
 
1410					     &first_bad, &bad_sectors);
1411			if (is_bad < 0) {
1412				/* mustn't write here until the bad block is
1413				 * acknowledged*/
1414				set_bit(BlockedBadBlocks, &rdev->flags);
1415				blocked_rdev = rdev;
1416				break;
1417			}
1418			if (is_bad && first_bad <= r1_bio->sector) {
1419				/* Cannot write here at all */
1420				bad_sectors -= (r1_bio->sector - first_bad);
1421				if (bad_sectors < max_sectors)
1422					/* mustn't write more than bad_sectors
1423					 * to other devices yet
1424					 */
1425					max_sectors = bad_sectors;
1426				rdev_dec_pending(rdev, mddev);
1427				/* We don't set R1BIO_Degraded as that
1428				 * only applies if the disk is
1429				 * missing, so it might be re-added,
1430				 * and we want to know to recover this
1431				 * chunk.
1432				 * In this case the device is here,
1433				 * and the fact that this chunk is not
1434				 * in-sync is recorded in the bad
1435				 * block log
1436				 */
1437				continue;
1438			}
1439			if (is_bad) {
1440				int good_sectors = first_bad - r1_bio->sector;
1441				if (good_sectors < max_sectors)
1442					max_sectors = good_sectors;
1443			}
1444		}
1445		r1_bio->bios[i] = bio;
1446	}
1447	rcu_read_unlock();
1448
1449	if (unlikely(blocked_rdev)) {
1450		/* Wait for this device to become unblocked */
1451		int j;
1452
1453		for (j = 0; j < i; j++)
1454			if (r1_bio->bios[j])
1455				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1456		r1_bio->state = 0;
1457		allow_barrier(conf, bio->bi_iter.bi_sector);
1458		raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1459		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1460		wait_barrier(conf, bio->bi_iter.bi_sector);
1461		goto retry_write;
1462	}
1463
1464	if (max_sectors < bio_sectors(bio)) {
1465		struct bio *split = bio_split(bio, max_sectors,
1466					      GFP_NOIO, &conf->bio_split);
1467		bio_chain(split, bio);
1468		generic_make_request(bio);
1469		bio = split;
1470		r1_bio->master_bio = bio;
1471		r1_bio->sectors = max_sectors;
 
 
 
 
 
 
1472	}
 
1473
1474	atomic_set(&r1_bio->remaining, 1);
1475	atomic_set(&r1_bio->behind_remaining, 0);
1476
1477	first_clone = 1;
1478
1479	for (i = 0; i < disks; i++) {
1480		struct bio *mbio = NULL;
1481		if (!r1_bio->bios[i])
1482			continue;
1483
 
 
 
1484		if (first_clone) {
1485			/* do behind I/O ?
1486			 * Not if there are too many, or cannot
1487			 * allocate memory, or a reader on WriteMostly
1488			 * is waiting for behind writes to flush */
1489			if (bitmap &&
1490			    (atomic_read(&bitmap->behind_writes)
1491			     < mddev->bitmap_info.max_write_behind) &&
1492			    !waitqueue_active(&bitmap->behind_wait)) {
1493				alloc_behind_master_bio(r1_bio, bio);
1494			}
1495
1496			md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1497					     test_bit(R1BIO_BehindIO, &r1_bio->state));
 
 
1498			first_clone = 0;
1499		}
1500
1501		if (r1_bio->behind_master_bio)
1502			mbio = bio_clone_fast(r1_bio->behind_master_bio,
1503					      GFP_NOIO, &mddev->bio_set);
1504		else
1505			mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1506
1507		if (r1_bio->behind_master_bio) {
1508			struct md_rdev *rdev = conf->mirrors[i].rdev;
1509
1510			if (test_bit(WBCollisionCheck, &rdev->flags)) {
1511				sector_t lo = r1_bio->sector;
1512				sector_t hi = r1_bio->sector + r1_bio->sectors;
1513
1514				wait_event(rdev->wb_io_wait,
1515					   check_and_add_wb(rdev, lo, hi) == 0);
1516			}
1517			if (test_bit(WriteMostly, &rdev->flags))
1518				atomic_inc(&r1_bio->behind_remaining);
1519		}
1520
1521		r1_bio->bios[i] = mbio;
1522
1523		mbio->bi_iter.bi_sector	= (r1_bio->sector +
1524				   conf->mirrors[i].rdev->data_offset);
1525		bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1526		mbio->bi_end_io	= raid1_end_write_request;
1527		mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1528		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1529		    !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1530		    conf->raid_disks - mddev->degraded > 1)
1531			mbio->bi_opf |= MD_FAILFAST;
1532		mbio->bi_private = r1_bio;
1533
1534		atomic_inc(&r1_bio->remaining);
1535
1536		if (mddev->gendisk)
1537			trace_block_bio_remap(mbio->bi_disk->queue,
1538					      mbio, disk_devt(mddev->gendisk),
1539					      r1_bio->sector);
1540		/* flush_pending_writes() needs access to the rdev so...*/
1541		mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1542
1543		cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1544		if (cb)
1545			plug = container_of(cb, struct raid1_plug_cb, cb);
1546		else
1547			plug = NULL;
1548		if (plug) {
1549			bio_list_add(&plug->pending, mbio);
1550			plug->pending_cnt++;
1551		} else {
1552			spin_lock_irqsave(&conf->device_lock, flags);
1553			bio_list_add(&conf->pending_bio_list, mbio);
1554			conf->pending_count++;
1555			spin_unlock_irqrestore(&conf->device_lock, flags);
1556			md_wakeup_thread(mddev->thread);
1557		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1558	}
1559
1560	r1_bio_write_done(r1_bio);
1561
1562	/* In case raid1d snuck in to freeze_array */
1563	wake_up(&conf->wait_barrier);
1564}
1565
1566static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1567{
1568	sector_t sectors;
1569
1570	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1571		md_flush_request(mddev, bio);
1572		return true;
1573	}
1574
1575	/*
1576	 * There is a limit to the maximum size, but
1577	 * the read/write handler might find a lower limit
1578	 * due to bad blocks.  To avoid multiple splits,
1579	 * we pass the maximum number of sectors down
1580	 * and let the lower level perform the split.
1581	 */
1582	sectors = align_to_barrier_unit_end(
1583		bio->bi_iter.bi_sector, bio_sectors(bio));
1584
1585	if (bio_data_dir(bio) == READ)
1586		raid1_read_request(mddev, bio, sectors, NULL);
1587	else {
1588		if (!md_write_start(mddev,bio))
1589			return false;
1590		raid1_write_request(mddev, bio, sectors);
1591	}
1592	return true;
1593}
1594
1595static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1596{
1597	struct r1conf *conf = mddev->private;
1598	int i;
1599
1600	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1601		   conf->raid_disks - mddev->degraded);
1602	rcu_read_lock();
1603	for (i = 0; i < conf->raid_disks; i++) {
1604		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1605		seq_printf(seq, "%s",
1606			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1607	}
1608	rcu_read_unlock();
1609	seq_printf(seq, "]");
1610}
1611
1612static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
 
1613{
1614	char b[BDEVNAME_SIZE];
1615	struct r1conf *conf = mddev->private;
1616	unsigned long flags;
1617
1618	/*
1619	 * If it is not operational, then we have already marked it as dead
1620	 * else if it is the last working disks with "fail_last_dev == false",
1621	 * ignore the error, let the next level up know.
1622	 * else mark the drive as failed
1623	 */
1624	spin_lock_irqsave(&conf->device_lock, flags);
1625	if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1626	    && (conf->raid_disks - mddev->degraded) == 1) {
1627		/*
1628		 * Don't fail the drive, act as though we were just a
1629		 * normal single drive.
1630		 * However don't try a recovery from this drive as
1631		 * it is very likely to fail.
1632		 */
1633		conf->recovery_disabled = mddev->recovery_disabled;
1634		spin_unlock_irqrestore(&conf->device_lock, flags);
1635		return;
1636	}
1637	set_bit(Blocked, &rdev->flags);
1638	if (test_and_clear_bit(In_sync, &rdev->flags))
 
 
1639		mddev->degraded++;
1640	set_bit(Faulty, &rdev->flags);
1641	spin_unlock_irqrestore(&conf->device_lock, flags);
1642	/*
1643	 * if recovery is running, make sure it aborts.
1644	 */
1645	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1646	set_mask_bits(&mddev->sb_flags, 0,
1647		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1648	pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1649		"md/raid1:%s: Operation continuing on %d devices.\n",
1650		mdname(mddev), bdevname(rdev->bdev, b),
1651		mdname(mddev), conf->raid_disks - mddev->degraded);
 
 
1652}
1653
1654static void print_conf(struct r1conf *conf)
1655{
1656	int i;
1657
1658	pr_debug("RAID1 conf printout:\n");
1659	if (!conf) {
1660		pr_debug("(!conf)\n");
1661		return;
1662	}
1663	pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1664		 conf->raid_disks);
1665
1666	rcu_read_lock();
1667	for (i = 0; i < conf->raid_disks; i++) {
1668		char b[BDEVNAME_SIZE];
1669		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1670		if (rdev)
1671			pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1672				 i, !test_bit(In_sync, &rdev->flags),
1673				 !test_bit(Faulty, &rdev->flags),
1674				 bdevname(rdev->bdev,b));
1675	}
1676	rcu_read_unlock();
1677}
1678
1679static void close_sync(struct r1conf *conf)
1680{
1681	int idx;
1682
1683	for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1684		_wait_barrier(conf, idx);
1685		_allow_barrier(conf, idx);
1686	}
1687
1688	mempool_exit(&conf->r1buf_pool);
 
1689}
1690
1691static int raid1_spare_active(struct mddev *mddev)
1692{
1693	int i;
1694	struct r1conf *conf = mddev->private;
1695	int count = 0;
1696	unsigned long flags;
1697
1698	/*
1699	 * Find all failed disks within the RAID1 configuration
1700	 * and mark them readable.
1701	 * Called under mddev lock, so rcu protection not needed.
1702	 * device_lock used to avoid races with raid1_end_read_request
1703	 * which expects 'In_sync' flags and ->degraded to be consistent.
1704	 */
1705	spin_lock_irqsave(&conf->device_lock, flags);
1706	for (i = 0; i < conf->raid_disks; i++) {
1707		struct md_rdev *rdev = conf->mirrors[i].rdev;
1708		struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1709		if (repl
1710		    && !test_bit(Candidate, &repl->flags)
1711		    && repl->recovery_offset == MaxSector
1712		    && !test_bit(Faulty, &repl->flags)
1713		    && !test_and_set_bit(In_sync, &repl->flags)) {
1714			/* replacement has just become active */
1715			if (!rdev ||
1716			    !test_and_clear_bit(In_sync, &rdev->flags))
1717				count++;
1718			if (rdev) {
1719				/* Replaced device not technically
1720				 * faulty, but we need to be sure
1721				 * it gets removed and never re-added
1722				 */
1723				set_bit(Faulty, &rdev->flags);
1724				sysfs_notify_dirent_safe(
1725					rdev->sysfs_state);
1726			}
1727		}
1728		if (rdev
1729		    && rdev->recovery_offset == MaxSector
1730		    && !test_bit(Faulty, &rdev->flags)
1731		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1732			count++;
1733			sysfs_notify_dirent_safe(rdev->sysfs_state);
1734		}
1735	}
 
1736	mddev->degraded -= count;
1737	spin_unlock_irqrestore(&conf->device_lock, flags);
1738
1739	print_conf(conf);
1740	return count;
1741}
1742
 
1743static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1744{
1745	struct r1conf *conf = mddev->private;
1746	int err = -EEXIST;
1747	int mirror = 0;
1748	struct raid1_info *p;
1749	int first = 0;
1750	int last = conf->raid_disks - 1;
 
1751
1752	if (mddev->recovery_disabled == conf->recovery_disabled)
1753		return -EBUSY;
1754
1755	if (md_integrity_add_rdev(rdev, mddev))
1756		return -ENXIO;
1757
1758	if (rdev->raid_disk >= 0)
1759		first = last = rdev->raid_disk;
1760
1761	/*
1762	 * find the disk ... but prefer rdev->saved_raid_disk
1763	 * if possible.
1764	 */
1765	if (rdev->saved_raid_disk >= 0 &&
1766	    rdev->saved_raid_disk >= first &&
1767	    rdev->saved_raid_disk < conf->raid_disks &&
1768	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1769		first = last = rdev->saved_raid_disk;
1770
1771	for (mirror = first; mirror <= last; mirror++) {
1772		p = conf->mirrors + mirror;
1773		if (!p->rdev) {
1774			if (mddev->gendisk)
1775				disk_stack_limits(mddev->gendisk, rdev->bdev,
1776						  rdev->data_offset << 9);
1777
1778			p->head_position = 0;
1779			rdev->raid_disk = mirror;
1780			err = 0;
1781			/* As all devices are equivalent, we don't need a full recovery
1782			 * if this was recently any drive of the array
1783			 */
1784			if (rdev->saved_raid_disk < 0)
1785				conf->fullsync = 1;
1786			rcu_assign_pointer(p->rdev, rdev);
1787			break;
1788		}
1789		if (test_bit(WantReplacement, &p->rdev->flags) &&
1790		    p[conf->raid_disks].rdev == NULL) {
1791			/* Add this device as a replacement */
1792			clear_bit(In_sync, &rdev->flags);
1793			set_bit(Replacement, &rdev->flags);
1794			rdev->raid_disk = mirror;
1795			err = 0;
1796			conf->fullsync = 1;
1797			rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1798			break;
1799		}
1800	}
1801	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1802		blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
 
 
 
 
 
 
 
 
 
 
 
 
1803	print_conf(conf);
1804	return err;
1805}
1806
1807static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1808{
1809	struct r1conf *conf = mddev->private;
1810	int err = 0;
1811	int number = rdev->raid_disk;
1812	struct raid1_info *p = conf->mirrors + number;
1813
1814	if (rdev != p->rdev)
1815		p = conf->mirrors + conf->raid_disks + number;
1816
1817	print_conf(conf);
1818	if (rdev == p->rdev) {
1819		if (test_bit(In_sync, &rdev->flags) ||
1820		    atomic_read(&rdev->nr_pending)) {
1821			err = -EBUSY;
1822			goto abort;
1823		}
1824		/* Only remove non-faulty devices if recovery
1825		 * is not possible.
1826		 */
1827		if (!test_bit(Faulty, &rdev->flags) &&
1828		    mddev->recovery_disabled != conf->recovery_disabled &&
1829		    mddev->degraded < conf->raid_disks) {
1830			err = -EBUSY;
1831			goto abort;
1832		}
1833		p->rdev = NULL;
1834		if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1835			synchronize_rcu();
1836			if (atomic_read(&rdev->nr_pending)) {
1837				/* lost the race, try later */
1838				err = -EBUSY;
1839				p->rdev = rdev;
1840				goto abort;
1841			}
1842		}
1843		if (conf->mirrors[conf->raid_disks + number].rdev) {
1844			/* We just removed a device that is being replaced.
1845			 * Move down the replacement.  We drain all IO before
1846			 * doing this to avoid confusion.
1847			 */
1848			struct md_rdev *repl =
1849				conf->mirrors[conf->raid_disks + number].rdev;
1850			freeze_array(conf, 0);
1851			if (atomic_read(&repl->nr_pending)) {
1852				/* It means that some queued IO of retry_list
1853				 * hold repl. Thus, we cannot set replacement
1854				 * as NULL, avoiding rdev NULL pointer
1855				 * dereference in sync_request_write and
1856				 * handle_write_finished.
1857				 */
1858				err = -EBUSY;
1859				unfreeze_array(conf);
1860				goto abort;
1861			}
1862			clear_bit(Replacement, &repl->flags);
1863			p->rdev = repl;
1864			conf->mirrors[conf->raid_disks + number].rdev = NULL;
1865			unfreeze_array(conf);
1866		}
1867
1868		clear_bit(WantReplacement, &rdev->flags);
1869		err = md_integrity_register(mddev);
1870	}
1871abort:
1872
1873	print_conf(conf);
1874	return err;
1875}
1876
1877static void end_sync_read(struct bio *bio)
 
1878{
1879	struct r1bio *r1_bio = get_resync_r1bio(bio);
1880
1881	update_head_pos(r1_bio->read_disk, r1_bio);
1882
1883	/*
1884	 * we have read a block, now it needs to be re-written,
1885	 * or re-read if the read failed.
1886	 * We don't do much here, just schedule handling by raid1d
1887	 */
1888	if (!bio->bi_status)
1889		set_bit(R1BIO_Uptodate, &r1_bio->state);
1890
1891	if (atomic_dec_and_test(&r1_bio->remaining))
1892		reschedule_retry(r1_bio);
1893}
1894
1895static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1896{
1897	sector_t sync_blocks = 0;
1898	sector_t s = r1_bio->sector;
1899	long sectors_to_go = r1_bio->sectors;
1900
1901	/* make sure these bits don't get cleared. */
1902	do {
1903		md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1904		s += sync_blocks;
1905		sectors_to_go -= sync_blocks;
1906	} while (sectors_to_go > 0);
1907}
1908
1909static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1910{
1911	if (atomic_dec_and_test(&r1_bio->remaining)) {
1912		struct mddev *mddev = r1_bio->mddev;
1913		int s = r1_bio->sectors;
1914
1915		if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1916		    test_bit(R1BIO_WriteError, &r1_bio->state))
1917			reschedule_retry(r1_bio);
1918		else {
1919			put_buf(r1_bio);
1920			md_done_sync(mddev, s, uptodate);
1921		}
1922	}
1923}
1924
1925static void end_sync_write(struct bio *bio)
1926{
1927	int uptodate = !bio->bi_status;
1928	struct r1bio *r1_bio = get_resync_r1bio(bio);
1929	struct mddev *mddev = r1_bio->mddev;
1930	struct r1conf *conf = mddev->private;
 
1931	sector_t first_bad;
1932	int bad_sectors;
1933	struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
 
1934
1935	if (!uptodate) {
1936		abort_sync_write(mddev, r1_bio);
1937		set_bit(WriteErrorSeen, &rdev->flags);
1938		if (!test_and_set_bit(WantReplacement, &rdev->flags))
 
 
 
 
 
 
 
 
 
 
 
1939			set_bit(MD_RECOVERY_NEEDED, &
1940				mddev->recovery);
1941		set_bit(R1BIO_WriteError, &r1_bio->state);
1942	} else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 
 
1943			       &first_bad, &bad_sectors) &&
1944		   !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1945				r1_bio->sector,
1946				r1_bio->sectors,
1947				&first_bad, &bad_sectors)
1948		)
1949		set_bit(R1BIO_MadeGood, &r1_bio->state);
1950
1951	put_sync_write_buf(r1_bio, uptodate);
 
 
 
 
 
 
 
 
 
1952}
1953
1954static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1955			    int sectors, struct page *page, int rw)
1956{
1957	if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1958		/* success */
1959		return 1;
1960	if (rw == WRITE) {
1961		set_bit(WriteErrorSeen, &rdev->flags);
1962		if (!test_and_set_bit(WantReplacement,
1963				      &rdev->flags))
1964			set_bit(MD_RECOVERY_NEEDED, &
1965				rdev->mddev->recovery);
1966	}
1967	/* need to record an error - either for the block or the device */
1968	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1969		md_error(rdev->mddev, rdev);
1970	return 0;
1971}
1972
1973static int fix_sync_read_error(struct r1bio *r1_bio)
1974{
1975	/* Try some synchronous reads of other devices to get
1976	 * good data, much like with normal read errors.  Only
1977	 * read into the pages we already have so we don't
1978	 * need to re-issue the read request.
1979	 * We don't need to freeze the array, because being in an
1980	 * active sync request, there is no normal IO, and
1981	 * no overlapping syncs.
1982	 * We don't need to check is_badblock() again as we
1983	 * made sure that anything with a bad block in range
1984	 * will have bi_end_io clear.
1985	 */
1986	struct mddev *mddev = r1_bio->mddev;
1987	struct r1conf *conf = mddev->private;
1988	struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1989	struct page **pages = get_resync_pages(bio)->pages;
1990	sector_t sect = r1_bio->sector;
1991	int sectors = r1_bio->sectors;
1992	int idx = 0;
1993	struct md_rdev *rdev;
1994
1995	rdev = conf->mirrors[r1_bio->read_disk].rdev;
1996	if (test_bit(FailFast, &rdev->flags)) {
1997		/* Don't try recovering from here - just fail it
1998		 * ... unless it is the last working device of course */
1999		md_error(mddev, rdev);
2000		if (test_bit(Faulty, &rdev->flags))
2001			/* Don't try to read from here, but make sure
2002			 * put_buf does it's thing
2003			 */
2004			bio->bi_end_io = end_sync_write;
2005	}
2006
2007	while(sectors) {
2008		int s = sectors;
2009		int d = r1_bio->read_disk;
2010		int success = 0;
 
2011		int start;
2012
2013		if (s > (PAGE_SIZE>>9))
2014			s = PAGE_SIZE >> 9;
2015		do {
2016			if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2017				/* No rcu protection needed here devices
2018				 * can only be removed when no resync is
2019				 * active, and resync is currently active
2020				 */
2021				rdev = conf->mirrors[d].rdev;
2022				if (sync_page_io(rdev, sect, s<<9,
2023						 pages[idx],
2024						 REQ_OP_READ, 0, false)) {
2025					success = 1;
2026					break;
2027				}
2028			}
2029			d++;
2030			if (d == conf->raid_disks * 2)
2031				d = 0;
2032		} while (!success && d != r1_bio->read_disk);
2033
2034		if (!success) {
2035			char b[BDEVNAME_SIZE];
2036			int abort = 0;
2037			/* Cannot read from anywhere, this block is lost.
2038			 * Record a bad block on each device.  If that doesn't
2039			 * work just disable and interrupt the recovery.
2040			 * Don't fail devices as that won't really help.
2041			 */
2042			pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2043					    mdname(mddev), bio_devname(bio, b),
2044					    (unsigned long long)r1_bio->sector);
 
 
2045			for (d = 0; d < conf->raid_disks * 2; d++) {
2046				rdev = conf->mirrors[d].rdev;
2047				if (!rdev || test_bit(Faulty, &rdev->flags))
2048					continue;
2049				if (!rdev_set_badblocks(rdev, sect, s, 0))
2050					abort = 1;
2051			}
2052			if (abort) {
2053				conf->recovery_disabled =
2054					mddev->recovery_disabled;
2055				set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2056				md_done_sync(mddev, r1_bio->sectors, 0);
2057				put_buf(r1_bio);
2058				return 0;
2059			}
2060			/* Try next page */
2061			sectors -= s;
2062			sect += s;
2063			idx++;
2064			continue;
2065		}
2066
2067		start = d;
2068		/* write it back and re-read */
2069		while (d != r1_bio->read_disk) {
2070			if (d == 0)
2071				d = conf->raid_disks * 2;
2072			d--;
2073			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2074				continue;
2075			rdev = conf->mirrors[d].rdev;
2076			if (r1_sync_page_io(rdev, sect, s,
2077					    pages[idx],
2078					    WRITE) == 0) {
2079				r1_bio->bios[d]->bi_end_io = NULL;
2080				rdev_dec_pending(rdev, mddev);
2081			}
2082		}
2083		d = start;
2084		while (d != r1_bio->read_disk) {
2085			if (d == 0)
2086				d = conf->raid_disks * 2;
2087			d--;
2088			if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2089				continue;
2090			rdev = conf->mirrors[d].rdev;
2091			if (r1_sync_page_io(rdev, sect, s,
2092					    pages[idx],
2093					    READ) != 0)
2094				atomic_add(s, &rdev->corrected_errors);
2095		}
2096		sectors -= s;
2097		sect += s;
2098		idx ++;
2099	}
2100	set_bit(R1BIO_Uptodate, &r1_bio->state);
2101	bio->bi_status = 0;
2102	return 1;
2103}
2104
2105static void process_checks(struct r1bio *r1_bio)
2106{
2107	/* We have read all readable devices.  If we haven't
2108	 * got the block, then there is no hope left.
2109	 * If we have, then we want to do a comparison
2110	 * and skip the write if everything is the same.
2111	 * If any blocks failed to read, then we need to
2112	 * attempt an over-write
2113	 */
2114	struct mddev *mddev = r1_bio->mddev;
2115	struct r1conf *conf = mddev->private;
2116	int primary;
2117	int i;
2118	int vcnt;
2119
2120	/* Fix variable parts of all bios */
2121	vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2122	for (i = 0; i < conf->raid_disks * 2; i++) {
2123		blk_status_t status;
2124		struct bio *b = r1_bio->bios[i];
2125		struct resync_pages *rp = get_resync_pages(b);
2126		if (b->bi_end_io != end_sync_read)
2127			continue;
2128		/* fixup the bio for reuse, but preserve errno */
2129		status = b->bi_status;
2130		bio_reset(b);
2131		b->bi_status = status;
2132		b->bi_iter.bi_sector = r1_bio->sector +
2133			conf->mirrors[i].rdev->data_offset;
2134		bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2135		b->bi_end_io = end_sync_read;
2136		rp->raid_bio = r1_bio;
2137		b->bi_private = rp;
2138
2139		/* initialize bvec table again */
2140		md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2141	}
2142	for (primary = 0; primary < conf->raid_disks * 2; primary++)
2143		if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2144		    !r1_bio->bios[primary]->bi_status) {
2145			r1_bio->bios[primary]->bi_end_io = NULL;
2146			rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2147			break;
2148		}
2149	r1_bio->read_disk = primary;
 
2150	for (i = 0; i < conf->raid_disks * 2; i++) {
2151		int j = 0;
2152		struct bio *pbio = r1_bio->bios[primary];
2153		struct bio *sbio = r1_bio->bios[i];
2154		blk_status_t status = sbio->bi_status;
2155		struct page **ppages = get_resync_pages(pbio)->pages;
2156		struct page **spages = get_resync_pages(sbio)->pages;
2157		struct bio_vec *bi;
2158		int page_len[RESYNC_PAGES] = { 0 };
2159		struct bvec_iter_all iter_all;
2160
2161		if (sbio->bi_end_io != end_sync_read)
2162			continue;
2163		/* Now we can 'fixup' the error value */
2164		sbio->bi_status = 0;
2165
2166		bio_for_each_segment_all(bi, sbio, iter_all)
2167			page_len[j++] = bi->bv_len;
2168
2169		if (!status) {
2170			for (j = vcnt; j-- ; ) {
2171				if (memcmp(page_address(ppages[j]),
2172					   page_address(spages[j]),
2173					   page_len[j]))
 
 
 
2174					break;
2175			}
2176		} else
2177			j = 0;
2178		if (j >= 0)
2179			atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2180		if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2181			      && !status)) {
2182			/* No need to write to this device. */
2183			sbio->bi_end_io = NULL;
2184			rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2185			continue;
2186		}
2187
2188		bio_copy_data(sbio, pbio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2189	}
 
2190}
2191
2192static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2193{
2194	struct r1conf *conf = mddev->private;
2195	int i;
2196	int disks = conf->raid_disks * 2;
2197	struct bio *wbio;
 
 
2198
2199	if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2200		/* ouch - failed to read all of that. */
2201		if (!fix_sync_read_error(r1_bio))
2202			return;
2203
2204	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2205		process_checks(r1_bio);
2206
2207	/*
2208	 * schedule writes
2209	 */
2210	atomic_set(&r1_bio->remaining, 1);
2211	for (i = 0; i < disks ; i++) {
2212		wbio = r1_bio->bios[i];
2213		if (wbio->bi_end_io == NULL ||
2214		    (wbio->bi_end_io == end_sync_read &&
2215		     (i == r1_bio->read_disk ||
2216		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2217			continue;
2218		if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2219			abort_sync_write(mddev, r1_bio);
2220			continue;
2221		}
2222
2223		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2224		if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2225			wbio->bi_opf |= MD_FAILFAST;
2226
 
2227		wbio->bi_end_io = end_sync_write;
2228		atomic_inc(&r1_bio->remaining);
2229		md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2230
2231		generic_make_request(wbio);
2232	}
2233
2234	put_sync_write_buf(r1_bio, 1);
 
 
 
 
 
 
 
 
 
 
2235}
2236
2237/*
2238 * This is a kernel thread which:
2239 *
2240 *	1.	Retries failed read operations on working mirrors.
2241 *	2.	Updates the raid superblock when problems encounter.
2242 *	3.	Performs writes following reads for array synchronising.
2243 */
2244
2245static void fix_read_error(struct r1conf *conf, int read_disk,
2246			   sector_t sect, int sectors)
2247{
2248	struct mddev *mddev = conf->mddev;
2249	while(sectors) {
2250		int s = sectors;
2251		int d = read_disk;
2252		int success = 0;
2253		int start;
2254		struct md_rdev *rdev;
2255
2256		if (s > (PAGE_SIZE>>9))
2257			s = PAGE_SIZE >> 9;
2258
2259		do {
 
 
 
 
 
2260			sector_t first_bad;
2261			int bad_sectors;
2262
2263			rcu_read_lock();
2264			rdev = rcu_dereference(conf->mirrors[d].rdev);
2265			if (rdev &&
2266			    (test_bit(In_sync, &rdev->flags) ||
2267			     (!test_bit(Faulty, &rdev->flags) &&
2268			      rdev->recovery_offset >= sect + s)) &&
2269			    is_badblock(rdev, sect, s,
2270					&first_bad, &bad_sectors) == 0) {
2271				atomic_inc(&rdev->nr_pending);
2272				rcu_read_unlock();
2273				if (sync_page_io(rdev, sect, s<<9,
2274					 conf->tmppage, REQ_OP_READ, 0, false))
2275					success = 1;
2276				rdev_dec_pending(rdev, mddev);
2277				if (success)
2278					break;
2279			} else
2280				rcu_read_unlock();
2281			d++;
2282			if (d == conf->raid_disks * 2)
2283				d = 0;
2284		} while (!success && d != read_disk);
2285
2286		if (!success) {
2287			/* Cannot read from anywhere - mark it bad */
2288			struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2289			if (!rdev_set_badblocks(rdev, sect, s, 0))
2290				md_error(mddev, rdev);
2291			break;
2292		}
2293		/* write it back and re-read */
2294		start = d;
2295		while (d != read_disk) {
2296			if (d==0)
2297				d = conf->raid_disks * 2;
2298			d--;
2299			rcu_read_lock();
2300			rdev = rcu_dereference(conf->mirrors[d].rdev);
2301			if (rdev &&
2302			    !test_bit(Faulty, &rdev->flags)) {
2303				atomic_inc(&rdev->nr_pending);
2304				rcu_read_unlock();
2305				r1_sync_page_io(rdev, sect, s,
2306						conf->tmppage, WRITE);
2307				rdev_dec_pending(rdev, mddev);
2308			} else
2309				rcu_read_unlock();
2310		}
2311		d = start;
2312		while (d != read_disk) {
2313			char b[BDEVNAME_SIZE];
2314			if (d==0)
2315				d = conf->raid_disks * 2;
2316			d--;
2317			rcu_read_lock();
2318			rdev = rcu_dereference(conf->mirrors[d].rdev);
2319			if (rdev &&
2320			    !test_bit(Faulty, &rdev->flags)) {
2321				atomic_inc(&rdev->nr_pending);
2322				rcu_read_unlock();
2323				if (r1_sync_page_io(rdev, sect, s,
2324						    conf->tmppage, READ)) {
2325					atomic_add(s, &rdev->corrected_errors);
2326					pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2327						mdname(mddev), s,
2328						(unsigned long long)(sect +
2329								     rdev->data_offset),
2330						bdevname(rdev->bdev, b));
 
 
2331				}
2332				rdev_dec_pending(rdev, mddev);
2333			} else
2334				rcu_read_unlock();
2335		}
2336		sectors -= s;
2337		sect += s;
2338	}
2339}
2340
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2341static int narrow_write_error(struct r1bio *r1_bio, int i)
2342{
2343	struct mddev *mddev = r1_bio->mddev;
2344	struct r1conf *conf = mddev->private;
2345	struct md_rdev *rdev = conf->mirrors[i].rdev;
 
 
2346
2347	/* bio has the data to be written to device 'i' where
2348	 * we just recently had a write error.
2349	 * We repeatedly clone the bio and trim down to one block,
2350	 * then try the write.  Where the write fails we record
2351	 * a bad block.
2352	 * It is conceivable that the bio doesn't exactly align with
2353	 * blocks.  We must handle this somehow.
2354	 *
2355	 * We currently own a reference on the rdev.
2356	 */
2357
2358	int block_sectors;
2359	sector_t sector;
2360	int sectors;
2361	int sect_to_write = r1_bio->sectors;
2362	int ok = 1;
2363
2364	if (rdev->badblocks.shift < 0)
2365		return 0;
2366
2367	block_sectors = roundup(1 << rdev->badblocks.shift,
2368				bdev_logical_block_size(rdev->bdev) >> 9);
2369	sector = r1_bio->sector;
2370	sectors = ((sector + block_sectors)
2371		   & ~(sector_t)(block_sectors - 1))
2372		- sector;
2373
 
 
 
 
 
 
 
 
 
 
 
2374	while (sect_to_write) {
2375		struct bio *wbio;
2376		if (sectors > sect_to_write)
2377			sectors = sect_to_write;
2378		/* Write at 'sector' for 'sectors'*/
2379
2380		if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2381			wbio = bio_clone_fast(r1_bio->behind_master_bio,
2382					      GFP_NOIO,
2383					      &mddev->bio_set);
2384		} else {
2385			wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2386					      &mddev->bio_set);
2387		}
2388
2389		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2390		wbio->bi_iter.bi_sector = r1_bio->sector;
2391		wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2392
2393		bio_trim(wbio, sector - r1_bio->sector, sectors);
2394		wbio->bi_iter.bi_sector += rdev->data_offset;
2395		bio_set_dev(wbio, rdev->bdev);
2396
2397		if (submit_bio_wait(wbio) < 0)
2398			/* failure! */
2399			ok = rdev_set_badblocks(rdev, sector,
2400						sectors, 0)
2401				&& ok;
2402
2403		bio_put(wbio);
2404		sect_to_write -= sectors;
2405		sector += sectors;
2406		sectors = block_sectors;
2407	}
2408	return ok;
2409}
2410
2411static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2412{
2413	int m;
2414	int s = r1_bio->sectors;
2415	for (m = 0; m < conf->raid_disks * 2 ; m++) {
2416		struct md_rdev *rdev = conf->mirrors[m].rdev;
2417		struct bio *bio = r1_bio->bios[m];
2418		if (bio->bi_end_io == NULL)
2419			continue;
2420		if (!bio->bi_status &&
2421		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2422			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2423		}
2424		if (bio->bi_status &&
2425		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
2426			if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2427				md_error(conf->mddev, rdev);
2428		}
2429	}
2430	put_buf(r1_bio);
2431	md_done_sync(conf->mddev, s, 1);
2432}
2433
2434static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2435{
2436	int m, idx;
2437	bool fail = false;
2438
2439	for (m = 0; m < conf->raid_disks * 2 ; m++)
2440		if (r1_bio->bios[m] == IO_MADE_GOOD) {
2441			struct md_rdev *rdev = conf->mirrors[m].rdev;
2442			rdev_clear_badblocks(rdev,
2443					     r1_bio->sector,
2444					     r1_bio->sectors, 0);
2445			rdev_dec_pending(rdev, conf->mddev);
2446		} else if (r1_bio->bios[m] != NULL) {
2447			/* This drive got a write error.  We need to
2448			 * narrow down and record precise write
2449			 * errors.
2450			 */
2451			fail = true;
2452			if (!narrow_write_error(r1_bio, m)) {
2453				md_error(conf->mddev,
2454					 conf->mirrors[m].rdev);
2455				/* an I/O failed, we can't clear the bitmap */
2456				set_bit(R1BIO_Degraded, &r1_bio->state);
2457			}
2458			rdev_dec_pending(conf->mirrors[m].rdev,
2459					 conf->mddev);
2460		}
2461	if (fail) {
2462		spin_lock_irq(&conf->device_lock);
2463		list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2464		idx = sector_to_idx(r1_bio->sector);
2465		atomic_inc(&conf->nr_queued[idx]);
2466		spin_unlock_irq(&conf->device_lock);
2467		/*
2468		 * In case freeze_array() is waiting for condition
2469		 * get_unqueued_pending() == extra to be true.
2470		 */
2471		wake_up(&conf->wait_barrier);
2472		md_wakeup_thread(conf->mddev->thread);
2473	} else {
2474		if (test_bit(R1BIO_WriteError, &r1_bio->state))
2475			close_write(r1_bio);
2476		raid_end_bio_io(r1_bio);
2477	}
2478}
2479
2480static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2481{
 
 
2482	struct mddev *mddev = conf->mddev;
2483	struct bio *bio;
 
2484	struct md_rdev *rdev;
2485
2486	clear_bit(R1BIO_ReadError, &r1_bio->state);
2487	/* we got a read error. Maybe the drive is bad.  Maybe just
2488	 * the block and we can fix it.
2489	 * We freeze all other IO, and try reading the block from
2490	 * other devices.  When we find one, we re-write
2491	 * and check it that fixes the read error.
2492	 * This is all done synchronously while the array is
2493	 * frozen
2494	 */
2495
2496	bio = r1_bio->bios[r1_bio->read_disk];
2497	bio_put(bio);
2498	r1_bio->bios[r1_bio->read_disk] = NULL;
2499
2500	rdev = conf->mirrors[r1_bio->read_disk].rdev;
2501	if (mddev->ro == 0
2502	    && !test_bit(FailFast, &rdev->flags)) {
2503		freeze_array(conf, 1);
2504		fix_read_error(conf, r1_bio->read_disk,
2505			       r1_bio->sector, r1_bio->sectors);
2506		unfreeze_array(conf);
2507	} else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2508		md_error(mddev, rdev);
 
 
 
 
 
 
 
 
 
 
2509	} else {
2510		r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2511	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2512
2513	rdev_dec_pending(rdev, conf->mddev);
2514	allow_barrier(conf, r1_bio->sector);
2515	bio = r1_bio->master_bio;
 
 
 
 
2516
2517	/* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2518	r1_bio->state = 0;
2519	raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
 
2520}
2521
2522static void raid1d(struct md_thread *thread)
2523{
2524	struct mddev *mddev = thread->mddev;
2525	struct r1bio *r1_bio;
2526	unsigned long flags;
2527	struct r1conf *conf = mddev->private;
2528	struct list_head *head = &conf->retry_list;
2529	struct blk_plug plug;
2530	int idx;
2531
2532	md_check_recovery(mddev);
2533
2534	if (!list_empty_careful(&conf->bio_end_io_list) &&
2535	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2536		LIST_HEAD(tmp);
2537		spin_lock_irqsave(&conf->device_lock, flags);
2538		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2539			list_splice_init(&conf->bio_end_io_list, &tmp);
2540		spin_unlock_irqrestore(&conf->device_lock, flags);
2541		while (!list_empty(&tmp)) {
2542			r1_bio = list_first_entry(&tmp, struct r1bio,
2543						  retry_list);
2544			list_del(&r1_bio->retry_list);
2545			idx = sector_to_idx(r1_bio->sector);
2546			atomic_dec(&conf->nr_queued[idx]);
2547			if (mddev->degraded)
2548				set_bit(R1BIO_Degraded, &r1_bio->state);
2549			if (test_bit(R1BIO_WriteError, &r1_bio->state))
2550				close_write(r1_bio);
2551			raid_end_bio_io(r1_bio);
2552		}
2553	}
2554
2555	blk_start_plug(&plug);
2556	for (;;) {
2557
2558		flush_pending_writes(conf);
 
2559
2560		spin_lock_irqsave(&conf->device_lock, flags);
2561		if (list_empty(head)) {
2562			spin_unlock_irqrestore(&conf->device_lock, flags);
2563			break;
2564		}
2565		r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2566		list_del(head->prev);
2567		idx = sector_to_idx(r1_bio->sector);
2568		atomic_dec(&conf->nr_queued[idx]);
2569		spin_unlock_irqrestore(&conf->device_lock, flags);
2570
2571		mddev = r1_bio->mddev;
2572		conf = mddev->private;
2573		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2574			if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2575			    test_bit(R1BIO_WriteError, &r1_bio->state))
2576				handle_sync_write_finished(conf, r1_bio);
2577			else
2578				sync_request_write(mddev, r1_bio);
2579		} else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2580			   test_bit(R1BIO_WriteError, &r1_bio->state))
2581			handle_write_finished(conf, r1_bio);
2582		else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2583			handle_read_error(conf, r1_bio);
2584		else
2585			WARN_ON_ONCE(1);
 
 
 
2586
2587		cond_resched();
2588		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2589			md_check_recovery(mddev);
2590	}
2591	blk_finish_plug(&plug);
2592}
2593
 
2594static int init_resync(struct r1conf *conf)
2595{
2596	int buffs;
2597
2598	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2599	BUG_ON(mempool_initialized(&conf->r1buf_pool));
2600
2601	return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2602			    r1buf_pool_free, conf->poolinfo);
2603}
2604
2605static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2606{
2607	struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2608	struct resync_pages *rps;
2609	struct bio *bio;
2610	int i;
2611
2612	for (i = conf->poolinfo->raid_disks; i--; ) {
2613		bio = r1bio->bios[i];
2614		rps = bio->bi_private;
2615		bio_reset(bio);
2616		bio->bi_private = rps;
2617	}
2618	r1bio->master_bio = NULL;
2619	return r1bio;
2620}
2621
2622/*
2623 * perform a "sync" on one "block"
2624 *
2625 * We need to make sure that no normal I/O request - particularly write
2626 * requests - conflict with active sync requests.
2627 *
2628 * This is achieved by tracking pending requests and a 'barrier' concept
2629 * that can be installed to exclude normal IO requests.
2630 */
2631
2632static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2633				   int *skipped)
2634{
2635	struct r1conf *conf = mddev->private;
2636	struct r1bio *r1_bio;
2637	struct bio *bio;
2638	sector_t max_sector, nr_sectors;
2639	int disk = -1;
2640	int i;
2641	int wonly = -1;
2642	int write_targets = 0, read_targets = 0;
2643	sector_t sync_blocks;
2644	int still_degraded = 0;
2645	int good_sectors = RESYNC_SECTORS;
2646	int min_bad = 0; /* number of sectors that are bad in all devices */
2647	int idx = sector_to_idx(sector_nr);
2648	int page_idx = 0;
2649
2650	if (!mempool_initialized(&conf->r1buf_pool))
2651		if (init_resync(conf))
2652			return 0;
2653
2654	max_sector = mddev->dev_sectors;
2655	if (sector_nr >= max_sector) {
2656		/* If we aborted, we need to abort the
2657		 * sync on the 'current' bitmap chunk (there will
2658		 * only be one in raid1 resync.
2659		 * We can find the current addess in mddev->curr_resync
2660		 */
2661		if (mddev->curr_resync < max_sector) /* aborted */
2662			md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2663					   &sync_blocks, 1);
2664		else /* completed sync */
2665			conf->fullsync = 0;
2666
2667		md_bitmap_close_sync(mddev->bitmap);
2668		close_sync(conf);
2669
2670		if (mddev_is_clustered(mddev)) {
2671			conf->cluster_sync_low = 0;
2672			conf->cluster_sync_high = 0;
2673		}
2674		return 0;
2675	}
2676
2677	if (mddev->bitmap == NULL &&
2678	    mddev->recovery_cp == MaxSector &&
2679	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2680	    conf->fullsync == 0) {
2681		*skipped = 1;
2682		return max_sector - sector_nr;
2683	}
2684	/* before building a request, check if we can skip these blocks..
2685	 * This call the bitmap_start_sync doesn't actually record anything
2686	 */
2687	if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2688	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2689		/* We can skip this block, and probably several more */
2690		*skipped = 1;
2691		return sync_blocks;
2692	}
2693
2694	/*
2695	 * If there is non-resync activity waiting for a turn, then let it
2696	 * though before starting on this new sync request.
2697	 */
2698	if (atomic_read(&conf->nr_waiting[idx]))
2699		schedule_timeout_uninterruptible(1);
2700
2701	/* we are incrementing sector_nr below. To be safe, we check against
2702	 * sector_nr + two times RESYNC_SECTORS
2703	 */
2704
2705	md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2706		mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2707
2708
2709	if (raise_barrier(conf, sector_nr))
2710		return 0;
2711
2712	r1_bio = raid1_alloc_init_r1buf(conf);
2713
2714	rcu_read_lock();
2715	/*
2716	 * If we get a correctably read error during resync or recovery,
2717	 * we might want to read from a different device.  So we
2718	 * flag all drives that could conceivably be read from for READ,
2719	 * and any others (which will be non-In_sync devices) for WRITE.
2720	 * If a read fails, we try reading from something else for which READ
2721	 * is OK.
2722	 */
2723
2724	r1_bio->mddev = mddev;
2725	r1_bio->sector = sector_nr;
2726	r1_bio->state = 0;
2727	set_bit(R1BIO_IsSync, &r1_bio->state);
2728	/* make sure good_sectors won't go across barrier unit boundary */
2729	good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2730
2731	for (i = 0; i < conf->raid_disks * 2; i++) {
2732		struct md_rdev *rdev;
2733		bio = r1_bio->bios[i];
2734
 
 
 
 
 
 
 
 
 
 
 
 
2735		rdev = rcu_dereference(conf->mirrors[i].rdev);
2736		if (rdev == NULL ||
2737		    test_bit(Faulty, &rdev->flags)) {
2738			if (i < conf->raid_disks)
2739				still_degraded = 1;
2740		} else if (!test_bit(In_sync, &rdev->flags)) {
2741			bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2742			bio->bi_end_io = end_sync_write;
2743			write_targets ++;
2744		} else {
2745			/* may need to read from here */
2746			sector_t first_bad = MaxSector;
2747			int bad_sectors;
2748
2749			if (is_badblock(rdev, sector_nr, good_sectors,
2750					&first_bad, &bad_sectors)) {
2751				if (first_bad > sector_nr)
2752					good_sectors = first_bad - sector_nr;
2753				else {
2754					bad_sectors -= (sector_nr - first_bad);
2755					if (min_bad == 0 ||
2756					    min_bad > bad_sectors)
2757						min_bad = bad_sectors;
2758				}
2759			}
2760			if (sector_nr < first_bad) {
2761				if (test_bit(WriteMostly, &rdev->flags)) {
2762					if (wonly < 0)
2763						wonly = i;
2764				} else {
2765					if (disk < 0)
2766						disk = i;
2767				}
2768				bio_set_op_attrs(bio, REQ_OP_READ, 0);
2769				bio->bi_end_io = end_sync_read;
2770				read_targets++;
2771			} else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2772				test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2773				!test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2774				/*
2775				 * The device is suitable for reading (InSync),
2776				 * but has bad block(s) here. Let's try to correct them,
2777				 * if we are doing resync or repair. Otherwise, leave
2778				 * this device alone for this sync request.
2779				 */
2780				bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2781				bio->bi_end_io = end_sync_write;
2782				write_targets++;
2783			}
2784		}
2785		if (bio->bi_end_io) {
2786			atomic_inc(&rdev->nr_pending);
2787			bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2788			bio_set_dev(bio, rdev->bdev);
2789			if (test_bit(FailFast, &rdev->flags))
2790				bio->bi_opf |= MD_FAILFAST;
2791		}
2792	}
2793	rcu_read_unlock();
2794	if (disk < 0)
2795		disk = wonly;
2796	r1_bio->read_disk = disk;
2797
2798	if (read_targets == 0 && min_bad > 0) {
2799		/* These sectors are bad on all InSync devices, so we
2800		 * need to mark them bad on all write targets
2801		 */
2802		int ok = 1;
2803		for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2804			if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2805				struct md_rdev *rdev = conf->mirrors[i].rdev;
2806				ok = rdev_set_badblocks(rdev, sector_nr,
2807							min_bad, 0
2808					) && ok;
2809			}
2810		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2811		*skipped = 1;
2812		put_buf(r1_bio);
2813
2814		if (!ok) {
2815			/* Cannot record the badblocks, so need to
2816			 * abort the resync.
2817			 * If there are multiple read targets, could just
2818			 * fail the really bad ones ???
2819			 */
2820			conf->recovery_disabled = mddev->recovery_disabled;
2821			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2822			return 0;
2823		} else
2824			return min_bad;
2825
2826	}
2827	if (min_bad > 0 && min_bad < good_sectors) {
2828		/* only resync enough to reach the next bad->good
2829		 * transition */
2830		good_sectors = min_bad;
2831	}
2832
2833	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2834		/* extra read targets are also write targets */
2835		write_targets += read_targets-1;
2836
2837	if (write_targets == 0 || read_targets == 0) {
2838		/* There is nowhere to write, so all non-sync
2839		 * drives must be failed - so we are finished
2840		 */
2841		sector_t rv;
2842		if (min_bad > 0)
2843			max_sector = sector_nr + min_bad;
2844		rv = max_sector - sector_nr;
2845		*skipped = 1;
2846		put_buf(r1_bio);
2847		return rv;
2848	}
2849
2850	if (max_sector > mddev->resync_max)
2851		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2852	if (max_sector > sector_nr + good_sectors)
2853		max_sector = sector_nr + good_sectors;
2854	nr_sectors = 0;
2855	sync_blocks = 0;
2856	do {
2857		struct page *page;
2858		int len = PAGE_SIZE;
2859		if (sector_nr + (len>>9) > max_sector)
2860			len = (max_sector - sector_nr) << 9;
2861		if (len == 0)
2862			break;
2863		if (sync_blocks == 0) {
2864			if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2865						  &sync_blocks, still_degraded) &&
2866			    !conf->fullsync &&
2867			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2868				break;
 
2869			if ((len >> 9) > sync_blocks)
2870				len = sync_blocks<<9;
2871		}
2872
2873		for (i = 0 ; i < conf->raid_disks * 2; i++) {
2874			struct resync_pages *rp;
2875
2876			bio = r1_bio->bios[i];
2877			rp = get_resync_pages(bio);
2878			if (bio->bi_end_io) {
2879				page = resync_fetch_page(rp, page_idx);
2880
2881				/*
2882				 * won't fail because the vec table is big
2883				 * enough to hold all these pages
2884				 */
2885				bio_add_page(bio, page, len, 0);
 
 
 
 
 
 
 
 
 
2886			}
2887		}
2888		nr_sectors += len>>9;
2889		sector_nr += len>>9;
2890		sync_blocks -= (len>>9);
2891	} while (++page_idx < RESYNC_PAGES);
2892
2893	r1_bio->sectors = nr_sectors;
2894
2895	if (mddev_is_clustered(mddev) &&
2896			conf->cluster_sync_high < sector_nr + nr_sectors) {
2897		conf->cluster_sync_low = mddev->curr_resync_completed;
2898		conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2899		/* Send resync message */
2900		md_cluster_ops->resync_info_update(mddev,
2901				conf->cluster_sync_low,
2902				conf->cluster_sync_high);
2903	}
2904
2905	/* For a user-requested sync, we read all readable devices and do a
2906	 * compare
2907	 */
2908	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2909		atomic_set(&r1_bio->remaining, read_targets);
2910		for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2911			bio = r1_bio->bios[i];
2912			if (bio->bi_end_io == end_sync_read) {
2913				read_targets--;
2914				md_sync_acct_bio(bio, nr_sectors);
2915				if (read_targets == 1)
2916					bio->bi_opf &= ~MD_FAILFAST;
2917				generic_make_request(bio);
2918			}
2919		}
2920	} else {
2921		atomic_set(&r1_bio->remaining, 1);
2922		bio = r1_bio->bios[r1_bio->read_disk];
2923		md_sync_acct_bio(bio, nr_sectors);
2924		if (read_targets == 1)
2925			bio->bi_opf &= ~MD_FAILFAST;
2926		generic_make_request(bio);
 
2927	}
2928	return nr_sectors;
2929}
2930
2931static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2932{
2933	if (sectors)
2934		return sectors;
2935
2936	return mddev->dev_sectors;
2937}
2938
2939static struct r1conf *setup_conf(struct mddev *mddev)
2940{
2941	struct r1conf *conf;
2942	int i;
2943	struct raid1_info *disk;
2944	struct md_rdev *rdev;
2945	int err = -ENOMEM;
2946
2947	conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2948	if (!conf)
2949		goto abort;
2950
2951	conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2952				   sizeof(atomic_t), GFP_KERNEL);
2953	if (!conf->nr_pending)
2954		goto abort;
2955
2956	conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2957				   sizeof(atomic_t), GFP_KERNEL);
2958	if (!conf->nr_waiting)
2959		goto abort;
2960
2961	conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2962				  sizeof(atomic_t), GFP_KERNEL);
2963	if (!conf->nr_queued)
2964		goto abort;
2965
2966	conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2967				sizeof(atomic_t), GFP_KERNEL);
2968	if (!conf->barrier)
2969		goto abort;
2970
2971	conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2972					    mddev->raid_disks, 2),
2973				GFP_KERNEL);
2974	if (!conf->mirrors)
2975		goto abort;
2976
2977	conf->tmppage = alloc_page(GFP_KERNEL);
2978	if (!conf->tmppage)
2979		goto abort;
2980
2981	conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2982	if (!conf->poolinfo)
2983		goto abort;
2984	conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2985	err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2986			   rbio_pool_free, conf->poolinfo);
2987	if (err)
2988		goto abort;
2989
2990	err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2991	if (err)
2992		goto abort;
2993
2994	conf->poolinfo->mddev = mddev;
2995
2996	err = -EINVAL;
2997	spin_lock_init(&conf->device_lock);
2998	rdev_for_each(rdev, mddev) {
 
2999		int disk_idx = rdev->raid_disk;
3000		if (disk_idx >= mddev->raid_disks
3001		    || disk_idx < 0)
3002			continue;
3003		if (test_bit(Replacement, &rdev->flags))
3004			disk = conf->mirrors + mddev->raid_disks + disk_idx;
3005		else
3006			disk = conf->mirrors + disk_idx;
3007
3008		if (disk->rdev)
3009			goto abort;
3010		disk->rdev = rdev;
 
 
 
 
3011		disk->head_position = 0;
3012		disk->seq_start = MaxSector;
3013	}
3014	conf->raid_disks = mddev->raid_disks;
3015	conf->mddev = mddev;
3016	INIT_LIST_HEAD(&conf->retry_list);
3017	INIT_LIST_HEAD(&conf->bio_end_io_list);
3018
3019	spin_lock_init(&conf->resync_lock);
3020	init_waitqueue_head(&conf->wait_barrier);
3021
3022	bio_list_init(&conf->pending_bio_list);
3023	conf->pending_count = 0;
3024	conf->recovery_disabled = mddev->recovery_disabled - 1;
3025
3026	err = -EIO;
 
3027	for (i = 0; i < conf->raid_disks * 2; i++) {
3028
3029		disk = conf->mirrors + i;
3030
3031		if (i < conf->raid_disks &&
3032		    disk[conf->raid_disks].rdev) {
3033			/* This slot has a replacement. */
3034			if (!disk->rdev) {
3035				/* No original, just make the replacement
3036				 * a recovering spare
3037				 */
3038				disk->rdev =
3039					disk[conf->raid_disks].rdev;
3040				disk[conf->raid_disks].rdev = NULL;
3041			} else if (!test_bit(In_sync, &disk->rdev->flags))
3042				/* Original is not in_sync - bad */
3043				goto abort;
3044		}
3045
3046		if (!disk->rdev ||
3047		    !test_bit(In_sync, &disk->rdev->flags)) {
3048			disk->head_position = 0;
3049			if (disk->rdev &&
3050			    (disk->rdev->saved_raid_disk < 0))
3051				conf->fullsync = 1;
3052		}
 
 
 
 
 
3053	}
3054
 
 
 
 
 
3055	err = -ENOMEM;
3056	conf->thread = md_register_thread(raid1d, mddev, "raid1");
3057	if (!conf->thread)
 
 
 
3058		goto abort;
 
3059
3060	return conf;
3061
3062 abort:
3063	if (conf) {
3064		mempool_exit(&conf->r1bio_pool);
 
3065		kfree(conf->mirrors);
3066		safe_put_page(conf->tmppage);
3067		kfree(conf->poolinfo);
3068		kfree(conf->nr_pending);
3069		kfree(conf->nr_waiting);
3070		kfree(conf->nr_queued);
3071		kfree(conf->barrier);
3072		bioset_exit(&conf->bio_split);
3073		kfree(conf);
3074	}
3075	return ERR_PTR(err);
3076}
3077
3078static void raid1_free(struct mddev *mddev, void *priv);
3079static int raid1_run(struct mddev *mddev)
3080{
3081	struct r1conf *conf;
3082	int i;
3083	struct md_rdev *rdev;
3084	int ret;
3085	bool discard_supported = false;
3086
3087	if (mddev->level != 1) {
3088		pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3089			mdname(mddev), mddev->level);
3090		return -EIO;
3091	}
3092	if (mddev->reshape_position != MaxSector) {
3093		pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3094			mdname(mddev));
3095		return -EIO;
3096	}
3097	if (mddev_init_writes_pending(mddev) < 0)
3098		return -ENOMEM;
3099	/*
3100	 * copy the already verified devices into our private RAID1
3101	 * bookkeeping area. [whatever we allocate in run(),
3102	 * should be freed in raid1_free()]
3103	 */
3104	if (mddev->private == NULL)
3105		conf = setup_conf(mddev);
3106	else
3107		conf = mddev->private;
3108
3109	if (IS_ERR(conf))
3110		return PTR_ERR(conf);
3111
3112	if (mddev->queue) {
3113		blk_queue_max_write_same_sectors(mddev->queue, 0);
3114		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3115	}
3116
3117	rdev_for_each(rdev, mddev) {
3118		if (!mddev->gendisk)
3119			continue;
3120		disk_stack_limits(mddev->gendisk, rdev->bdev,
3121				  rdev->data_offset << 9);
3122		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3123			discard_supported = true;
3124	}
3125
3126	mddev->degraded = 0;
3127	for (i = 0; i < conf->raid_disks; i++)
3128		if (conf->mirrors[i].rdev == NULL ||
3129		    !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3130		    test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3131			mddev->degraded++;
3132	/*
3133	 * RAID1 needs at least one disk in active
3134	 */
3135	if (conf->raid_disks - mddev->degraded < 1) {
3136		ret = -EINVAL;
3137		goto abort;
3138	}
3139
3140	if (conf->raid_disks - mddev->degraded == 1)
3141		mddev->recovery_cp = MaxSector;
3142
3143	if (mddev->recovery_cp != MaxSector)
3144		pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3145			mdname(mddev));
3146	pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3147		mdname(mddev), mddev->raid_disks - mddev->degraded,
 
 
3148		mddev->raid_disks);
3149
3150	/*
3151	 * Ok, everything is just fine now
3152	 */
3153	mddev->thread = conf->thread;
3154	conf->thread = NULL;
3155	mddev->private = conf;
3156	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3157
3158	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3159
3160	if (mddev->queue) {
3161		if (discard_supported)
3162			blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3163						mddev->queue);
3164		else
3165			blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3166						  mddev->queue);
3167	}
3168
3169	ret = md_integrity_register(mddev);
3170	if (ret) {
3171		md_unregister_thread(&mddev->thread);
3172		goto abort;
3173	}
3174	return 0;
3175
3176abort:
3177	raid1_free(mddev, conf);
3178	return ret;
3179}
3180
3181static void raid1_free(struct mddev *mddev, void *priv)
3182{
3183	struct r1conf *conf = priv;
 
 
 
 
 
 
 
 
 
 
3184
3185	mempool_exit(&conf->r1bio_pool);
 
 
 
 
 
3186	kfree(conf->mirrors);
3187	safe_put_page(conf->tmppage);
3188	kfree(conf->poolinfo);
3189	kfree(conf->nr_pending);
3190	kfree(conf->nr_waiting);
3191	kfree(conf->nr_queued);
3192	kfree(conf->barrier);
3193	bioset_exit(&conf->bio_split);
3194	kfree(conf);
 
 
3195}
3196
3197static int raid1_resize(struct mddev *mddev, sector_t sectors)
3198{
3199	/* no resync is happening, and there is enough space
3200	 * on all devices, so we can resize.
3201	 * We need to make sure resync covers any new space.
3202	 * If the array is shrinking we should possibly wait until
3203	 * any io in the removed space completes, but it hardly seems
3204	 * worth it.
3205	 */
3206	sector_t newsize = raid1_size(mddev, sectors, 0);
3207	if (mddev->external_size &&
3208	    mddev->array_sectors > newsize)
3209		return -EINVAL;
3210	if (mddev->bitmap) {
3211		int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3212		if (ret)
3213			return ret;
3214	}
3215	md_set_array_sectors(mddev, newsize);
 
 
3216	if (sectors > mddev->dev_sectors &&
3217	    mddev->recovery_cp > mddev->dev_sectors) {
3218		mddev->recovery_cp = mddev->dev_sectors;
3219		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3220	}
3221	mddev->dev_sectors = sectors;
3222	mddev->resync_max_sectors = sectors;
3223	return 0;
3224}
3225
3226static int raid1_reshape(struct mddev *mddev)
3227{
3228	/* We need to:
3229	 * 1/ resize the r1bio_pool
3230	 * 2/ resize conf->mirrors
3231	 *
3232	 * We allocate a new r1bio_pool if we can.
3233	 * Then raise a device barrier and wait until all IO stops.
3234	 * Then resize conf->mirrors and swap in the new r1bio pool.
3235	 *
3236	 * At the same time, we "pack" the devices so that all the missing
3237	 * devices have the higher raid_disk numbers.
3238	 */
3239	mempool_t newpool, oldpool;
3240	struct pool_info *newpoolinfo;
3241	struct raid1_info *newmirrors;
3242	struct r1conf *conf = mddev->private;
3243	int cnt, raid_disks;
3244	unsigned long flags;
3245	int d, d2;
3246	int ret;
3247
3248	memset(&newpool, 0, sizeof(newpool));
3249	memset(&oldpool, 0, sizeof(oldpool));
3250
3251	/* Cannot change chunk_size, layout, or level */
3252	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3253	    mddev->layout != mddev->new_layout ||
3254	    mddev->level != mddev->new_level) {
3255		mddev->new_chunk_sectors = mddev->chunk_sectors;
3256		mddev->new_layout = mddev->layout;
3257		mddev->new_level = mddev->level;
3258		return -EINVAL;
3259	}
3260
3261	if (!mddev_is_clustered(mddev))
3262		md_allow_write(mddev);
 
3263
3264	raid_disks = mddev->raid_disks + mddev->delta_disks;
3265
3266	if (raid_disks < conf->raid_disks) {
3267		cnt=0;
3268		for (d= 0; d < conf->raid_disks; d++)
3269			if (conf->mirrors[d].rdev)
3270				cnt++;
3271		if (cnt > raid_disks)
3272			return -EBUSY;
3273	}
3274
3275	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3276	if (!newpoolinfo)
3277		return -ENOMEM;
3278	newpoolinfo->mddev = mddev;
3279	newpoolinfo->raid_disks = raid_disks * 2;
3280
3281	ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3282			   rbio_pool_free, newpoolinfo);
3283	if (ret) {
3284		kfree(newpoolinfo);
3285		return ret;
3286	}
3287	newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3288					 raid_disks, 2),
3289			     GFP_KERNEL);
3290	if (!newmirrors) {
3291		kfree(newpoolinfo);
3292		mempool_exit(&newpool);
3293		return -ENOMEM;
3294	}
3295
3296	freeze_array(conf, 0);
3297
3298	/* ok, everything is stopped */
3299	oldpool = conf->r1bio_pool;
3300	conf->r1bio_pool = newpool;
3301
3302	for (d = d2 = 0; d < conf->raid_disks; d++) {
3303		struct md_rdev *rdev = conf->mirrors[d].rdev;
3304		if (rdev && rdev->raid_disk != d2) {
3305			sysfs_unlink_rdev(mddev, rdev);
3306			rdev->raid_disk = d2;
3307			sysfs_unlink_rdev(mddev, rdev);
3308			if (sysfs_link_rdev(mddev, rdev))
3309				pr_warn("md/raid1:%s: cannot register rd%d\n",
3310					mdname(mddev), rdev->raid_disk);
 
3311		}
3312		if (rdev)
3313			newmirrors[d2++].rdev = rdev;
3314	}
3315	kfree(conf->mirrors);
3316	conf->mirrors = newmirrors;
3317	kfree(conf->poolinfo);
3318	conf->poolinfo = newpoolinfo;
3319
3320	spin_lock_irqsave(&conf->device_lock, flags);
3321	mddev->degraded += (raid_disks - conf->raid_disks);
3322	spin_unlock_irqrestore(&conf->device_lock, flags);
3323	conf->raid_disks = mddev->raid_disks = raid_disks;
3324	mddev->delta_disks = 0;
3325
3326	unfreeze_array(conf);
 
3327
3328	set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3329	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3330	md_wakeup_thread(mddev->thread);
3331
3332	mempool_exit(&oldpool);
3333	return 0;
3334}
3335
3336static void raid1_quiesce(struct mddev *mddev, int quiesce)
3337{
3338	struct r1conf *conf = mddev->private;
3339
3340	if (quiesce)
3341		freeze_array(conf, 0);
3342	else
3343		unfreeze_array(conf);
 
 
 
 
 
 
 
3344}
3345
3346static void *raid1_takeover(struct mddev *mddev)
3347{
3348	/* raid1 can take over:
3349	 *  raid5 with 2 devices, any layout or chunk size
3350	 */
3351	if (mddev->level == 5 && mddev->raid_disks == 2) {
3352		struct r1conf *conf;
3353		mddev->new_level = 1;
3354		mddev->new_layout = 0;
3355		mddev->new_chunk_sectors = 0;
3356		conf = setup_conf(mddev);
3357		if (!IS_ERR(conf)) {
3358			/* Array must appear to be quiesced */
3359			conf->array_frozen = 1;
3360			mddev_clear_unsupported_flags(mddev,
3361				UNSUPPORTED_MDDEV_FLAGS);
3362		}
3363		return conf;
3364	}
3365	return ERR_PTR(-EINVAL);
3366}
3367
3368static struct md_personality raid1_personality =
3369{
3370	.name		= "raid1",
3371	.level		= 1,
3372	.owner		= THIS_MODULE,
3373	.make_request	= raid1_make_request,
3374	.run		= raid1_run,
3375	.free		= raid1_free,
3376	.status		= raid1_status,
3377	.error_handler	= raid1_error,
3378	.hot_add_disk	= raid1_add_disk,
3379	.hot_remove_disk= raid1_remove_disk,
3380	.spare_active	= raid1_spare_active,
3381	.sync_request	= raid1_sync_request,
3382	.resize		= raid1_resize,
3383	.size		= raid1_size,
3384	.check_reshape	= raid1_reshape,
3385	.quiesce	= raid1_quiesce,
3386	.takeover	= raid1_takeover,
3387	.congested	= raid1_congested,
3388};
3389
3390static int __init raid_init(void)
3391{
3392	return register_md_personality(&raid1_personality);
3393}
3394
3395static void raid_exit(void)
3396{
3397	unregister_md_personality(&raid1_personality);
3398}
3399
3400module_init(raid_init);
3401module_exit(raid_exit);
3402MODULE_LICENSE("GPL");
3403MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3404MODULE_ALIAS("md-personality-3"); /* RAID1 */
3405MODULE_ALIAS("md-raid1");
3406MODULE_ALIAS("md-level-1");
3407
3408module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);