Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v3.15
   1/*
   2 * raid10.c : Multiple Devices driver for Linux
   3 *
   4 * Copyright (C) 2000-2004 Neil Brown
   5 *
   6 * RAID-10 support for md.
   7 *
   8 * Base on code in raid1.c.  See raid1.c for further copyright information.
   9 *
  10 *
  11 * This program is free software; you can redistribute it and/or modify
  12 * it under the terms of the GNU General Public License as published by
  13 * the Free Software Foundation; either version 2, or (at your option)
  14 * any later version.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * (for example /usr/src/linux/COPYING); if not, write to the Free
  18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19 */
  20
  21#include <linux/slab.h>
  22#include <linux/delay.h>
  23#include <linux/blkdev.h>
  24#include <linux/module.h>
  25#include <linux/seq_file.h>
  26#include <linux/ratelimit.h>
  27#include <linux/kthread.h>
 
  28#include "md.h"
  29#include "raid10.h"
  30#include "raid0.h"
  31#include "bitmap.h"
  32
  33/*
  34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
  35 * The layout of data is defined by
  36 *    chunk_size
  37 *    raid_disks
  38 *    near_copies (stored in low byte of layout)
  39 *    far_copies (stored in second byte of layout)
  40 *    far_offset (stored in bit 16 of layout )
  41 *    use_far_sets (stored in bit 17 of layout )
 
  42 *
  43 * The data to be stored is divided into chunks using chunksize.  Each device
  44 * is divided into far_copies sections.   In each section, chunks are laid out
  45 * in a style similar to raid0, but near_copies copies of each chunk is stored
  46 * (each on a different drive).  The starting device for each section is offset
  47 * near_copies from the starting device of the previous section.  Thus there
  48 * are (near_copies * far_copies) of each chunk, and each is on a different
  49 * drive.  near_copies and far_copies must be at least one, and their product
  50 * is at most raid_disks.
  51 *
  52 * If far_offset is true, then the far_copies are handled a bit differently.
  53 * The copies are still in different stripes, but instead of being very far
  54 * apart on disk, there are adjacent stripes.
  55 *
  56 * The far and offset algorithms are handled slightly differently if
  57 * 'use_far_sets' is true.  In this case, the array's devices are grouped into
  58 * sets that are (near_copies * far_copies) in size.  The far copied stripes
  59 * are still shifted by 'near_copies' devices, but this shifting stays confined
  60 * to the set rather than the entire array.  This is done to improve the number
  61 * of device combinations that can fail without causing the array to fail.
  62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
  63 * on a device):
  64 *    A B C D    A B C D E
  65 *      ...         ...
  66 *    D A B C    E A B C D
  67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
  68 *    [A B] [C D]    [A B] [C D E]
  69 *    |...| |...|    |...| | ... |
  70 *    [B A] [D C]    [B A] [E C D]
  71 */
  72
  73/*
  74 * Number of guaranteed r10bios in case of extreme VM load:
  75 */
  76#define	NR_RAID10_BIOS 256
  77
  78/* when we get a read error on a read-only array, we redirect to another
  79 * device without failing the first device, or trying to over-write to
  80 * correct the read error.  To keep track of bad blocks on a per-bio
  81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  82 */
  83#define IO_BLOCKED ((struct bio *)1)
  84/* When we successfully write to a known bad-block, we need to remove the
  85 * bad-block marking which must be done from process context.  So we record
  86 * the success by setting devs[n].bio to IO_MADE_GOOD
  87 */
  88#define IO_MADE_GOOD ((struct bio *)2)
  89
  90#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  91
  92/* When there are this many requests queued to be written by
  93 * the raid10 thread, we become 'congested' to provide back-pressure
  94 * for writeback.
  95 */
  96static int max_queued_requests = 1024;
  97
  98static void allow_barrier(struct r10conf *conf);
  99static void lower_barrier(struct r10conf *conf);
 100static int _enough(struct r10conf *conf, int previous, int ignore);
 
 101static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
 102				int *skipped);
 103static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
 104static void end_reshape_write(struct bio *bio, int error);
 105static void end_reshape(struct r10conf *conf);
 106
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 107static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 108{
 109	struct r10conf *conf = data;
 110	int size = offsetof(struct r10bio, devs[conf->copies]);
 111
 112	/* allocate a r10bio with room for raid_disks entries in the
 113	 * bios array */
 114	return kzalloc(size, gfp_flags);
 115}
 116
 117static void r10bio_pool_free(void *r10_bio, void *data)
 118{
 119	kfree(r10_bio);
 120}
 121
 122/* Maximum size of each resync request */
 123#define RESYNC_BLOCK_SIZE (64*1024)
 124#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
 125/* amount of memory to reserve for resync requests */
 126#define RESYNC_WINDOW (1024*1024)
 127/* maximum number of concurrent requests, memory permitting */
 128#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
 
 
 129
 130/*
 131 * When performing a resync, we need to read and compare, so
 132 * we need as many pages are there are copies.
 133 * When performing a recovery, we need 2 bios, one for read,
 134 * one for write (we recover only one drive per r10buf)
 135 *
 136 */
 137static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 138{
 139	struct r10conf *conf = data;
 140	struct page *page;
 141	struct r10bio *r10_bio;
 142	struct bio *bio;
 143	int i, j;
 144	int nalloc;
 
 145
 146	r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 147	if (!r10_bio)
 148		return NULL;
 149
 150	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
 151	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 152		nalloc = conf->copies; /* resync */
 153	else
 154		nalloc = 2; /* recovery */
 155
 
 
 
 
 
 
 
 
 
 156	/*
 157	 * Allocate bios.
 158	 */
 159	for (j = nalloc ; j-- ; ) {
 160		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 161		if (!bio)
 162			goto out_free_bio;
 163		r10_bio->devs[j].bio = bio;
 164		if (!conf->have_replacement)
 165			continue;
 166		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 167		if (!bio)
 168			goto out_free_bio;
 169		r10_bio->devs[j].repl_bio = bio;
 170	}
 171	/*
 172	 * Allocate RESYNC_PAGES data pages and attach them
 173	 * where needed.
 174	 */
 175	for (j = 0 ; j < nalloc; j++) {
 176		struct bio *rbio = r10_bio->devs[j].repl_bio;
 
 
 
 
 
 
 177		bio = r10_bio->devs[j].bio;
 178		for (i = 0; i < RESYNC_PAGES; i++) {
 179			if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
 180					       &conf->mddev->recovery)) {
 181				/* we can share bv_page's during recovery
 182				 * and reshape */
 183				struct bio *rbio = r10_bio->devs[0].bio;
 184				page = rbio->bi_io_vec[i].bv_page;
 185				get_page(page);
 186			} else
 187				page = alloc_page(gfp_flags);
 188			if (unlikely(!page))
 189				goto out_free_pages;
 
 
 
 
 190
 191			bio->bi_io_vec[i].bv_page = page;
 192			if (rbio)
 193				rbio->bi_io_vec[i].bv_page = page;
 
 
 194		}
 195	}
 196
 197	return r10_bio;
 198
 199out_free_pages:
 200	for ( ; i > 0 ; i--)
 201		safe_put_page(bio->bi_io_vec[i-1].bv_page);
 202	while (j--)
 203		for (i = 0; i < RESYNC_PAGES ; i++)
 204			safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 205	j = 0;
 206out_free_bio:
 207	for ( ; j < nalloc; j++) {
 208		if (r10_bio->devs[j].bio)
 209			bio_put(r10_bio->devs[j].bio);
 210		if (r10_bio->devs[j].repl_bio)
 211			bio_put(r10_bio->devs[j].repl_bio);
 212	}
 
 
 213	r10bio_pool_free(r10_bio, conf);
 214	return NULL;
 215}
 216
 217static void r10buf_pool_free(void *__r10_bio, void *data)
 218{
 219	int i;
 220	struct r10conf *conf = data;
 221	struct r10bio *r10bio = __r10_bio;
 222	int j;
 
 223
 224	for (j=0; j < conf->copies; j++) {
 225		struct bio *bio = r10bio->devs[j].bio;
 226		if (bio) {
 227			for (i = 0; i < RESYNC_PAGES; i++) {
 228				safe_put_page(bio->bi_io_vec[i].bv_page);
 229				bio->bi_io_vec[i].bv_page = NULL;
 230			}
 231			bio_put(bio);
 232		}
 233		bio = r10bio->devs[j].repl_bio;
 234		if (bio)
 235			bio_put(bio);
 236	}
 
 
 
 
 237	r10bio_pool_free(r10bio, conf);
 238}
 239
 240static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 241{
 242	int i;
 243
 244	for (i = 0; i < conf->copies; i++) {
 245		struct bio **bio = & r10_bio->devs[i].bio;
 246		if (!BIO_SPECIAL(*bio))
 247			bio_put(*bio);
 248		*bio = NULL;
 249		bio = &r10_bio->devs[i].repl_bio;
 250		if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 251			bio_put(*bio);
 252		*bio = NULL;
 253	}
 254}
 255
 256static void free_r10bio(struct r10bio *r10_bio)
 257{
 258	struct r10conf *conf = r10_bio->mddev->private;
 259
 260	put_all_bios(conf, r10_bio);
 261	mempool_free(r10_bio, conf->r10bio_pool);
 262}
 263
 264static void put_buf(struct r10bio *r10_bio)
 265{
 266	struct r10conf *conf = r10_bio->mddev->private;
 267
 268	mempool_free(r10_bio, conf->r10buf_pool);
 269
 270	lower_barrier(conf);
 271}
 272
 273static void reschedule_retry(struct r10bio *r10_bio)
 274{
 275	unsigned long flags;
 276	struct mddev *mddev = r10_bio->mddev;
 277	struct r10conf *conf = mddev->private;
 278
 279	spin_lock_irqsave(&conf->device_lock, flags);
 280	list_add(&r10_bio->retry_list, &conf->retry_list);
 281	conf->nr_queued ++;
 282	spin_unlock_irqrestore(&conf->device_lock, flags);
 283
 284	/* wake up frozen array... */
 285	wake_up(&conf->wait_barrier);
 286
 287	md_wakeup_thread(mddev->thread);
 288}
 289
 290/*
 291 * raid_end_bio_io() is called when we have finished servicing a mirrored
 292 * operation and are ready to return a success/failure code to the buffer
 293 * cache layer.
 294 */
 295static void raid_end_bio_io(struct r10bio *r10_bio)
 296{
 297	struct bio *bio = r10_bio->master_bio;
 298	int done;
 299	struct r10conf *conf = r10_bio->mddev->private;
 300
 301	if (bio->bi_phys_segments) {
 302		unsigned long flags;
 303		spin_lock_irqsave(&conf->device_lock, flags);
 304		bio->bi_phys_segments--;
 305		done = (bio->bi_phys_segments == 0);
 306		spin_unlock_irqrestore(&conf->device_lock, flags);
 307	} else
 308		done = 1;
 309	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 310		clear_bit(BIO_UPTODATE, &bio->bi_flags);
 311	if (done) {
 312		bio_endio(bio, 0);
 313		/*
 314		 * Wake up any possible resync thread that waits for the device
 315		 * to go idle.
 316		 */
 317		allow_barrier(conf);
 318	}
 319	free_r10bio(r10_bio);
 320}
 321
 322/*
 323 * Update disk head position estimator based on IRQ completion info.
 324 */
 325static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 326{
 327	struct r10conf *conf = r10_bio->mddev->private;
 328
 329	conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 330		r10_bio->devs[slot].addr + (r10_bio->sectors);
 331}
 332
 333/*
 334 * Find the disk number which triggered given bio
 335 */
 336static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 337			 struct bio *bio, int *slotp, int *replp)
 338{
 339	int slot;
 340	int repl = 0;
 341
 342	for (slot = 0; slot < conf->copies; slot++) {
 343		if (r10_bio->devs[slot].bio == bio)
 344			break;
 345		if (r10_bio->devs[slot].repl_bio == bio) {
 346			repl = 1;
 347			break;
 348		}
 349	}
 350
 351	BUG_ON(slot == conf->copies);
 352	update_head_pos(slot, r10_bio);
 353
 354	if (slotp)
 355		*slotp = slot;
 356	if (replp)
 357		*replp = repl;
 358	return r10_bio->devs[slot].devnum;
 359}
 360
 361static void raid10_end_read_request(struct bio *bio, int error)
 362{
 363	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 364	struct r10bio *r10_bio = bio->bi_private;
 365	int slot, dev;
 366	struct md_rdev *rdev;
 367	struct r10conf *conf = r10_bio->mddev->private;
 368
 369
 370	slot = r10_bio->read_slot;
 371	dev = r10_bio->devs[slot].devnum;
 372	rdev = r10_bio->devs[slot].rdev;
 373	/*
 374	 * this branch is our 'one mirror IO has finished' event handler:
 375	 */
 376	update_head_pos(slot, r10_bio);
 377
 378	if (uptodate) {
 379		/*
 380		 * Set R10BIO_Uptodate in our master bio, so that
 381		 * we will return a good error code to the higher
 382		 * levels even if IO on some other mirrored buffer fails.
 383		 *
 384		 * The 'master' represents the composite IO operation to
 385		 * user-side. So if something waits for IO, then it will
 386		 * wait for the 'master' bio.
 387		 */
 388		set_bit(R10BIO_Uptodate, &r10_bio->state);
 389	} else {
 390		/* If all other devices that store this block have
 391		 * failed, we want to return the error upwards rather
 392		 * than fail the last device.  Here we redefine
 393		 * "uptodate" to mean "Don't want to retry"
 394		 */
 395		if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
 396			     rdev->raid_disk))
 397			uptodate = 1;
 398	}
 399	if (uptodate) {
 400		raid_end_bio_io(r10_bio);
 401		rdev_dec_pending(rdev, conf->mddev);
 402	} else {
 403		/*
 404		 * oops, read error - keep the refcount on the rdev
 405		 */
 406		char b[BDEVNAME_SIZE];
 407		printk_ratelimited(KERN_ERR
 408				   "md/raid10:%s: %s: rescheduling sector %llu\n",
 409				   mdname(conf->mddev),
 410				   bdevname(rdev->bdev, b),
 411				   (unsigned long long)r10_bio->sector);
 412		set_bit(R10BIO_ReadError, &r10_bio->state);
 413		reschedule_retry(r10_bio);
 414	}
 415}
 416
 417static void close_write(struct r10bio *r10_bio)
 418{
 419	/* clear the bitmap if all writes complete successfully */
 420	bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 421			r10_bio->sectors,
 422			!test_bit(R10BIO_Degraded, &r10_bio->state),
 423			0);
 424	md_write_end(r10_bio->mddev);
 425}
 426
 427static void one_write_done(struct r10bio *r10_bio)
 428{
 429	if (atomic_dec_and_test(&r10_bio->remaining)) {
 430		if (test_bit(R10BIO_WriteError, &r10_bio->state))
 431			reschedule_retry(r10_bio);
 432		else {
 433			close_write(r10_bio);
 434			if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 435				reschedule_retry(r10_bio);
 436			else
 437				raid_end_bio_io(r10_bio);
 438		}
 439	}
 440}
 441
 442static void raid10_end_write_request(struct bio *bio, int error)
 443{
 444	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 445	struct r10bio *r10_bio = bio->bi_private;
 446	int dev;
 447	int dec_rdev = 1;
 448	struct r10conf *conf = r10_bio->mddev->private;
 449	int slot, repl;
 450	struct md_rdev *rdev = NULL;
 
 
 
 
 451
 452	dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 453
 454	if (repl)
 455		rdev = conf->mirrors[dev].replacement;
 456	if (!rdev) {
 457		smp_rmb();
 458		repl = 0;
 459		rdev = conf->mirrors[dev].rdev;
 460	}
 461	/*
 462	 * this branch is our 'one mirror IO has finished' event handler:
 463	 */
 464	if (!uptodate) {
 465		if (repl)
 466			/* Never record new bad blocks to replacement,
 467			 * just fail it.
 468			 */
 469			md_error(rdev->mddev, rdev);
 470		else {
 471			set_bit(WriteErrorSeen,	&rdev->flags);
 472			if (!test_and_set_bit(WantReplacement, &rdev->flags))
 473				set_bit(MD_RECOVERY_NEEDED,
 474					&rdev->mddev->recovery);
 475			set_bit(R10BIO_WriteError, &r10_bio->state);
 476			dec_rdev = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 477		}
 478	} else {
 479		/*
 480		 * Set R10BIO_Uptodate in our master bio, so that
 481		 * we will return a good error code for to the higher
 482		 * levels even if IO on some other mirrored buffer fails.
 483		 *
 484		 * The 'master' represents the composite IO operation to
 485		 * user-side. So if something waits for IO, then it will
 486		 * wait for the 'master' bio.
 487		 */
 488		sector_t first_bad;
 489		int bad_sectors;
 490
 491		/*
 492		 * Do not set R10BIO_Uptodate if the current device is
 493		 * rebuilding or Faulty. This is because we cannot use
 494		 * such device for properly reading the data back (we could
 495		 * potentially use it, if the current write would have felt
 496		 * before rdev->recovery_offset, but for simplicity we don't
 497		 * check this here.
 498		 */
 499		if (test_bit(In_sync, &rdev->flags) &&
 500		    !test_bit(Faulty, &rdev->flags))
 501			set_bit(R10BIO_Uptodate, &r10_bio->state);
 502
 503		/* Maybe we can clear some bad blocks. */
 504		if (is_badblock(rdev,
 505				r10_bio->devs[slot].addr,
 506				r10_bio->sectors,
 507				&first_bad, &bad_sectors)) {
 508			bio_put(bio);
 509			if (repl)
 510				r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 511			else
 512				r10_bio->devs[slot].bio = IO_MADE_GOOD;
 513			dec_rdev = 0;
 514			set_bit(R10BIO_MadeGood, &r10_bio->state);
 515		}
 516	}
 517
 518	/*
 519	 *
 520	 * Let's see if all mirrored write operations have finished
 521	 * already.
 522	 */
 523	one_write_done(r10_bio);
 524	if (dec_rdev)
 525		rdev_dec_pending(rdev, conf->mddev);
 
 
 526}
 527
 528/*
 529 * RAID10 layout manager
 530 * As well as the chunksize and raid_disks count, there are two
 531 * parameters: near_copies and far_copies.
 532 * near_copies * far_copies must be <= raid_disks.
 533 * Normally one of these will be 1.
 534 * If both are 1, we get raid0.
 535 * If near_copies == raid_disks, we get raid1.
 536 *
 537 * Chunks are laid out in raid0 style with near_copies copies of the
 538 * first chunk, followed by near_copies copies of the next chunk and
 539 * so on.
 540 * If far_copies > 1, then after 1/far_copies of the array has been assigned
 541 * as described above, we start again with a device offset of near_copies.
 542 * So we effectively have another copy of the whole array further down all
 543 * the drives, but with blocks on different drives.
 544 * With this layout, and block is never stored twice on the one device.
 545 *
 546 * raid10_find_phys finds the sector offset of a given virtual sector
 547 * on each device that it is on.
 548 *
 549 * raid10_find_virt does the reverse mapping, from a device and a
 550 * sector offset to a virtual address
 551 */
 552
 553static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 554{
 555	int n,f;
 556	sector_t sector;
 557	sector_t chunk;
 558	sector_t stripe;
 559	int dev;
 560	int slot = 0;
 561	int last_far_set_start, last_far_set_size;
 562
 563	last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 564	last_far_set_start *= geo->far_set_size;
 565
 566	last_far_set_size = geo->far_set_size;
 567	last_far_set_size += (geo->raid_disks % geo->far_set_size);
 568
 569	/* now calculate first sector/dev */
 570	chunk = r10bio->sector >> geo->chunk_shift;
 571	sector = r10bio->sector & geo->chunk_mask;
 572
 573	chunk *= geo->near_copies;
 574	stripe = chunk;
 575	dev = sector_div(stripe, geo->raid_disks);
 576	if (geo->far_offset)
 577		stripe *= geo->far_copies;
 578
 579	sector += stripe << geo->chunk_shift;
 580
 581	/* and calculate all the others */
 582	for (n = 0; n < geo->near_copies; n++) {
 583		int d = dev;
 584		int set;
 585		sector_t s = sector;
 586		r10bio->devs[slot].devnum = d;
 587		r10bio->devs[slot].addr = s;
 588		slot++;
 589
 590		for (f = 1; f < geo->far_copies; f++) {
 591			set = d / geo->far_set_size;
 592			d += geo->near_copies;
 593
 594			if ((geo->raid_disks % geo->far_set_size) &&
 595			    (d > last_far_set_start)) {
 596				d -= last_far_set_start;
 597				d %= last_far_set_size;
 598				d += last_far_set_start;
 599			} else {
 600				d %= geo->far_set_size;
 601				d += geo->far_set_size * set;
 602			}
 603			s += geo->stride;
 604			r10bio->devs[slot].devnum = d;
 605			r10bio->devs[slot].addr = s;
 606			slot++;
 607		}
 608		dev++;
 609		if (dev >= geo->raid_disks) {
 610			dev = 0;
 611			sector += (geo->chunk_mask + 1);
 612		}
 613	}
 614}
 615
 616static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 617{
 618	struct geom *geo = &conf->geo;
 619
 620	if (conf->reshape_progress != MaxSector &&
 621	    ((r10bio->sector >= conf->reshape_progress) !=
 622	     conf->mddev->reshape_backwards)) {
 623		set_bit(R10BIO_Previous, &r10bio->state);
 624		geo = &conf->prev;
 625	} else
 626		clear_bit(R10BIO_Previous, &r10bio->state);
 627
 628	__raid10_find_phys(geo, r10bio);
 629}
 630
 631static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 632{
 633	sector_t offset, chunk, vchunk;
 634	/* Never use conf->prev as this is only called during resync
 635	 * or recovery, so reshape isn't happening
 636	 */
 637	struct geom *geo = &conf->geo;
 638	int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
 639	int far_set_size = geo->far_set_size;
 640	int last_far_set_start;
 641
 642	if (geo->raid_disks % geo->far_set_size) {
 643		last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 644		last_far_set_start *= geo->far_set_size;
 645
 646		if (dev >= last_far_set_start) {
 647			far_set_size = geo->far_set_size;
 648			far_set_size += (geo->raid_disks % geo->far_set_size);
 649			far_set_start = last_far_set_start;
 650		}
 651	}
 652
 653	offset = sector & geo->chunk_mask;
 654	if (geo->far_offset) {
 655		int fc;
 656		chunk = sector >> geo->chunk_shift;
 657		fc = sector_div(chunk, geo->far_copies);
 658		dev -= fc * geo->near_copies;
 659		if (dev < far_set_start)
 660			dev += far_set_size;
 661	} else {
 662		while (sector >= geo->stride) {
 663			sector -= geo->stride;
 664			if (dev < (geo->near_copies + far_set_start))
 665				dev += far_set_size - geo->near_copies;
 666			else
 667				dev -= geo->near_copies;
 668		}
 669		chunk = sector >> geo->chunk_shift;
 670	}
 671	vchunk = chunk * geo->raid_disks + dev;
 672	sector_div(vchunk, geo->near_copies);
 673	return (vchunk << geo->chunk_shift) + offset;
 674}
 675
 676/**
 677 *	raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 678 *	@q: request queue
 679 *	@bvm: properties of new bio
 680 *	@biovec: the request that could be merged to it.
 681 *
 682 *	Return amount of bytes we can accept at this offset
 683 *	This requires checking for end-of-chunk if near_copies != raid_disks,
 684 *	and for subordinate merge_bvec_fns if merge_check_needed.
 685 */
 686static int raid10_mergeable_bvec(struct request_queue *q,
 687				 struct bvec_merge_data *bvm,
 688				 struct bio_vec *biovec)
 689{
 690	struct mddev *mddev = q->queuedata;
 691	struct r10conf *conf = mddev->private;
 692	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 693	int max;
 694	unsigned int chunk_sectors;
 695	unsigned int bio_sectors = bvm->bi_size >> 9;
 696	struct geom *geo = &conf->geo;
 697
 698	chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
 699	if (conf->reshape_progress != MaxSector &&
 700	    ((sector >= conf->reshape_progress) !=
 701	     conf->mddev->reshape_backwards))
 702		geo = &conf->prev;
 703
 704	if (geo->near_copies < geo->raid_disks) {
 705		max = (chunk_sectors - ((sector & (chunk_sectors - 1))
 706					+ bio_sectors)) << 9;
 707		if (max < 0)
 708			/* bio_add cannot handle a negative return */
 709			max = 0;
 710		if (max <= biovec->bv_len && bio_sectors == 0)
 711			return biovec->bv_len;
 712	} else
 713		max = biovec->bv_len;
 714
 715	if (mddev->merge_check_needed) {
 716		struct {
 717			struct r10bio r10_bio;
 718			struct r10dev devs[conf->copies];
 719		} on_stack;
 720		struct r10bio *r10_bio = &on_stack.r10_bio;
 721		int s;
 722		if (conf->reshape_progress != MaxSector) {
 723			/* Cannot give any guidance during reshape */
 724			if (max <= biovec->bv_len && bio_sectors == 0)
 725				return biovec->bv_len;
 726			return 0;
 727		}
 728		r10_bio->sector = sector;
 729		raid10_find_phys(conf, r10_bio);
 730		rcu_read_lock();
 731		for (s = 0; s < conf->copies; s++) {
 732			int disk = r10_bio->devs[s].devnum;
 733			struct md_rdev *rdev = rcu_dereference(
 734				conf->mirrors[disk].rdev);
 735			if (rdev && !test_bit(Faulty, &rdev->flags)) {
 736				struct request_queue *q =
 737					bdev_get_queue(rdev->bdev);
 738				if (q->merge_bvec_fn) {
 739					bvm->bi_sector = r10_bio->devs[s].addr
 740						+ rdev->data_offset;
 741					bvm->bi_bdev = rdev->bdev;
 742					max = min(max, q->merge_bvec_fn(
 743							  q, bvm, biovec));
 744				}
 745			}
 746			rdev = rcu_dereference(conf->mirrors[disk].replacement);
 747			if (rdev && !test_bit(Faulty, &rdev->flags)) {
 748				struct request_queue *q =
 749					bdev_get_queue(rdev->bdev);
 750				if (q->merge_bvec_fn) {
 751					bvm->bi_sector = r10_bio->devs[s].addr
 752						+ rdev->data_offset;
 753					bvm->bi_bdev = rdev->bdev;
 754					max = min(max, q->merge_bvec_fn(
 755							  q, bvm, biovec));
 756				}
 757			}
 758		}
 759		rcu_read_unlock();
 760	}
 761	return max;
 762}
 763
 764/*
 765 * This routine returns the disk from which the requested read should
 766 * be done. There is a per-array 'next expected sequential IO' sector
 767 * number - if this matches on the next IO then we use the last disk.
 768 * There is also a per-disk 'last know head position' sector that is
 769 * maintained from IRQ contexts, both the normal and the resync IO
 770 * completion handlers update this position correctly. If there is no
 771 * perfect sequential match then we pick the disk whose head is closest.
 772 *
 773 * If there are 2 mirrors in the same 2 devices, performance degrades
 774 * because position is mirror, not device based.
 775 *
 776 * The rdev for the device selected will have nr_pending incremented.
 777 */
 778
 779/*
 780 * FIXME: possibly should rethink readbalancing and do it differently
 781 * depending on near_copies / far_copies geometry.
 782 */
 783static struct md_rdev *read_balance(struct r10conf *conf,
 784				    struct r10bio *r10_bio,
 785				    int *max_sectors)
 786{
 787	const sector_t this_sector = r10_bio->sector;
 788	int disk, slot;
 789	int sectors = r10_bio->sectors;
 790	int best_good_sectors;
 791	sector_t new_distance, best_dist;
 792	struct md_rdev *best_rdev, *rdev = NULL;
 793	int do_balance;
 794	int best_slot;
 795	struct geom *geo = &conf->geo;
 796
 797	raid10_find_phys(conf, r10_bio);
 798	rcu_read_lock();
 799retry:
 800	sectors = r10_bio->sectors;
 801	best_slot = -1;
 802	best_rdev = NULL;
 803	best_dist = MaxSector;
 804	best_good_sectors = 0;
 805	do_balance = 1;
 
 806	/*
 807	 * Check if we can balance. We can balance on the whole
 808	 * device if no resync is going on (recovery is ok), or below
 809	 * the resync window. We take the first readable disk when
 810	 * above the resync window.
 811	 */
 812	if (conf->mddev->recovery_cp < MaxSector
 813	    && (this_sector + sectors >= conf->next_resync))
 
 
 
 814		do_balance = 0;
 815
 816	for (slot = 0; slot < conf->copies ; slot++) {
 817		sector_t first_bad;
 818		int bad_sectors;
 819		sector_t dev_sector;
 820
 821		if (r10_bio->devs[slot].bio == IO_BLOCKED)
 822			continue;
 823		disk = r10_bio->devs[slot].devnum;
 824		rdev = rcu_dereference(conf->mirrors[disk].replacement);
 825		if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 826		    test_bit(Unmerged, &rdev->flags) ||
 827		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 828			rdev = rcu_dereference(conf->mirrors[disk].rdev);
 829		if (rdev == NULL ||
 830		    test_bit(Faulty, &rdev->flags) ||
 831		    test_bit(Unmerged, &rdev->flags))
 832			continue;
 833		if (!test_bit(In_sync, &rdev->flags) &&
 834		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 835			continue;
 836
 837		dev_sector = r10_bio->devs[slot].addr;
 838		if (is_badblock(rdev, dev_sector, sectors,
 839				&first_bad, &bad_sectors)) {
 840			if (best_dist < MaxSector)
 841				/* Already have a better slot */
 842				continue;
 843			if (first_bad <= dev_sector) {
 844				/* Cannot read here.  If this is the
 845				 * 'primary' device, then we must not read
 846				 * beyond 'bad_sectors' from another device.
 847				 */
 848				bad_sectors -= (dev_sector - first_bad);
 849				if (!do_balance && sectors > bad_sectors)
 850					sectors = bad_sectors;
 851				if (best_good_sectors > sectors)
 852					best_good_sectors = sectors;
 853			} else {
 854				sector_t good_sectors =
 855					first_bad - dev_sector;
 856				if (good_sectors > best_good_sectors) {
 857					best_good_sectors = good_sectors;
 858					best_slot = slot;
 859					best_rdev = rdev;
 860				}
 861				if (!do_balance)
 862					/* Must read from here */
 863					break;
 864			}
 865			continue;
 866		} else
 867			best_good_sectors = sectors;
 868
 869		if (!do_balance)
 870			break;
 871
 
 
 
 872		/* This optimisation is debatable, and completely destroys
 873		 * sequential read speed for 'far copies' arrays.  So only
 874		 * keep it for 'near' arrays, and review those later.
 875		 */
 876		if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 877			break;
 878
 879		/* for far > 1 always use the lowest address */
 880		if (geo->far_copies > 1)
 881			new_distance = r10_bio->devs[slot].addr;
 882		else
 883			new_distance = abs(r10_bio->devs[slot].addr -
 884					   conf->mirrors[disk].head_position);
 885		if (new_distance < best_dist) {
 886			best_dist = new_distance;
 887			best_slot = slot;
 888			best_rdev = rdev;
 889		}
 890	}
 891	if (slot >= conf->copies) {
 892		slot = best_slot;
 893		rdev = best_rdev;
 894	}
 895
 896	if (slot >= 0) {
 897		atomic_inc(&rdev->nr_pending);
 898		if (test_bit(Faulty, &rdev->flags)) {
 899			/* Cannot risk returning a device that failed
 900			 * before we inc'ed nr_pending
 901			 */
 902			rdev_dec_pending(rdev, conf->mddev);
 903			goto retry;
 904		}
 905		r10_bio->read_slot = slot;
 906	} else
 907		rdev = NULL;
 908	rcu_read_unlock();
 909	*max_sectors = best_good_sectors;
 910
 911	return rdev;
 912}
 913
 914int md_raid10_congested(struct mddev *mddev, int bits)
 915{
 916	struct r10conf *conf = mddev->private;
 917	int i, ret = 0;
 918
 919	if ((bits & (1 << BDI_async_congested)) &&
 920	    conf->pending_count >= max_queued_requests)
 921		return 1;
 922
 923	rcu_read_lock();
 924	for (i = 0;
 925	     (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
 926		     && ret == 0;
 927	     i++) {
 928		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 929		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 930			struct request_queue *q = bdev_get_queue(rdev->bdev);
 931
 932			ret |= bdi_congested(&q->backing_dev_info, bits);
 933		}
 934	}
 935	rcu_read_unlock();
 936	return ret;
 937}
 938EXPORT_SYMBOL_GPL(md_raid10_congested);
 939
 940static int raid10_congested(void *data, int bits)
 941{
 942	struct mddev *mddev = data;
 943
 944	return mddev_congested(mddev, bits) ||
 945		md_raid10_congested(mddev, bits);
 946}
 947
 948static void flush_pending_writes(struct r10conf *conf)
 949{
 950	/* Any writes that have been queued but are awaiting
 951	 * bitmap updates get flushed here.
 952	 */
 953	spin_lock_irq(&conf->device_lock);
 954
 955	if (conf->pending_bio_list.head) {
 
 956		struct bio *bio;
 
 957		bio = bio_list_get(&conf->pending_bio_list);
 958		conf->pending_count = 0;
 959		spin_unlock_irq(&conf->device_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 960		/* flush any pending bitmap writes to disk
 961		 * before proceeding w/ I/O */
 962		bitmap_unplug(conf->mddev->bitmap);
 963		wake_up(&conf->wait_barrier);
 964
 965		while (bio) { /* submit pending writes */
 966			struct bio *next = bio->bi_next;
 
 967			bio->bi_next = NULL;
 968			if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 969			    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 
 
 
 970				/* Just ignore it */
 971				bio_endio(bio, 0);
 972			else
 973				generic_make_request(bio);
 974			bio = next;
 975		}
 
 976	} else
 977		spin_unlock_irq(&conf->device_lock);
 978}
 979
 980/* Barriers....
 981 * Sometimes we need to suspend IO while we do something else,
 982 * either some resync/recovery, or reconfigure the array.
 983 * To do this we raise a 'barrier'.
 984 * The 'barrier' is a counter that can be raised multiple times
 985 * to count how many activities are happening which preclude
 986 * normal IO.
 987 * We can only raise the barrier if there is no pending IO.
 988 * i.e. if nr_pending == 0.
 989 * We choose only to raise the barrier if no-one is waiting for the
 990 * barrier to go down.  This means that as soon as an IO request
 991 * is ready, no other operations which require a barrier will start
 992 * until the IO request has had a chance.
 993 *
 994 * So: regular IO calls 'wait_barrier'.  When that returns there
 995 *    is no backgroup IO happening,  It must arrange to call
 996 *    allow_barrier when it has finished its IO.
 997 * backgroup IO calls must call raise_barrier.  Once that returns
 998 *    there is no normal IO happeing.  It must arrange to call
 999 *    lower_barrier when the particular background IO completes.
1000 */
1001
1002static void raise_barrier(struct r10conf *conf, int force)
1003{
1004	BUG_ON(force && !conf->barrier);
1005	spin_lock_irq(&conf->resync_lock);
1006
1007	/* Wait until no block IO is waiting (unless 'force') */
1008	wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1009			    conf->resync_lock);
1010
1011	/* block any new IO from starting */
1012	conf->barrier++;
1013
1014	/* Now wait for all pending IO to complete */
1015	wait_event_lock_irq(conf->wait_barrier,
1016			    !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1017			    conf->resync_lock);
1018
1019	spin_unlock_irq(&conf->resync_lock);
1020}
1021
1022static void lower_barrier(struct r10conf *conf)
1023{
1024	unsigned long flags;
1025	spin_lock_irqsave(&conf->resync_lock, flags);
1026	conf->barrier--;
1027	spin_unlock_irqrestore(&conf->resync_lock, flags);
1028	wake_up(&conf->wait_barrier);
1029}
1030
1031static void wait_barrier(struct r10conf *conf)
1032{
1033	spin_lock_irq(&conf->resync_lock);
1034	if (conf->barrier) {
1035		conf->nr_waiting++;
1036		/* Wait for the barrier to drop.
1037		 * However if there are already pending
1038		 * requests (preventing the barrier from
1039		 * rising completely), and the
1040		 * pre-process bio queue isn't empty,
1041		 * then don't wait, as we need to empty
1042		 * that queue to get the nr_pending
1043		 * count down.
1044		 */
 
1045		wait_event_lock_irq(conf->wait_barrier,
1046				    !conf->barrier ||
1047				    (conf->nr_pending &&
1048				     current->bio_list &&
1049				     !bio_list_empty(current->bio_list)),
 
1050				    conf->resync_lock);
1051		conf->nr_waiting--;
 
 
1052	}
1053	conf->nr_pending++;
1054	spin_unlock_irq(&conf->resync_lock);
1055}
1056
1057static void allow_barrier(struct r10conf *conf)
1058{
1059	unsigned long flags;
1060	spin_lock_irqsave(&conf->resync_lock, flags);
1061	conf->nr_pending--;
1062	spin_unlock_irqrestore(&conf->resync_lock, flags);
1063	wake_up(&conf->wait_barrier);
1064}
1065
1066static void freeze_array(struct r10conf *conf, int extra)
1067{
1068	/* stop syncio and normal IO and wait for everything to
1069	 * go quiet.
1070	 * We increment barrier and nr_waiting, and then
1071	 * wait until nr_pending match nr_queued+extra
1072	 * This is called in the context of one normal IO request
1073	 * that has failed. Thus any sync request that might be pending
1074	 * will be blocked by nr_pending, and we need to wait for
1075	 * pending IO requests to complete or be queued for re-try.
1076	 * Thus the number queued (nr_queued) plus this request (extra)
1077	 * must match the number of pending IOs (nr_pending) before
1078	 * we continue.
1079	 */
1080	spin_lock_irq(&conf->resync_lock);
 
1081	conf->barrier++;
1082	conf->nr_waiting++;
1083	wait_event_lock_irq_cmd(conf->wait_barrier,
1084				conf->nr_pending == conf->nr_queued+extra,
1085				conf->resync_lock,
1086				flush_pending_writes(conf));
1087
 
1088	spin_unlock_irq(&conf->resync_lock);
1089}
1090
1091static void unfreeze_array(struct r10conf *conf)
1092{
1093	/* reverse the effect of the freeze */
1094	spin_lock_irq(&conf->resync_lock);
1095	conf->barrier--;
1096	conf->nr_waiting--;
1097	wake_up(&conf->wait_barrier);
1098	spin_unlock_irq(&conf->resync_lock);
1099}
1100
1101static sector_t choose_data_offset(struct r10bio *r10_bio,
1102				   struct md_rdev *rdev)
1103{
1104	if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1105	    test_bit(R10BIO_Previous, &r10_bio->state))
1106		return rdev->data_offset;
1107	else
1108		return rdev->new_data_offset;
1109}
1110
1111struct raid10_plug_cb {
1112	struct blk_plug_cb	cb;
1113	struct bio_list		pending;
1114	int			pending_cnt;
1115};
1116
1117static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1118{
1119	struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1120						   cb);
1121	struct mddev *mddev = plug->cb.data;
1122	struct r10conf *conf = mddev->private;
1123	struct bio *bio;
1124
1125	if (from_schedule || current->bio_list) {
1126		spin_lock_irq(&conf->device_lock);
1127		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1128		conf->pending_count += plug->pending_cnt;
1129		spin_unlock_irq(&conf->device_lock);
1130		wake_up(&conf->wait_barrier);
1131		md_wakeup_thread(mddev->thread);
1132		kfree(plug);
1133		return;
1134	}
1135
1136	/* we aren't scheduling, so we can do the write-out directly. */
1137	bio = bio_list_get(&plug->pending);
1138	bitmap_unplug(mddev->bitmap);
1139	wake_up(&conf->wait_barrier);
1140
1141	while (bio) { /* submit pending writes */
1142		struct bio *next = bio->bi_next;
 
1143		bio->bi_next = NULL;
1144		if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1145		    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 
 
 
1146			/* Just ignore it */
1147			bio_endio(bio, 0);
1148		else
1149			generic_make_request(bio);
1150		bio = next;
1151	}
1152	kfree(plug);
1153}
1154
1155static void __make_request(struct mddev *mddev, struct bio *bio)
 
1156{
1157	struct r10conf *conf = mddev->private;
1158	struct r10bio *r10_bio;
1159	struct bio *read_bio;
1160	int i;
1161	const int rw = bio_data_dir(bio);
1162	const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1163	const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1164	const unsigned long do_discard = (bio->bi_rw
1165					  & (REQ_DISCARD | REQ_SECURE));
1166	const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1167	unsigned long flags;
1168	struct md_rdev *blocked_rdev;
1169	struct blk_plug_cb *cb;
1170	struct raid10_plug_cb *plug = NULL;
1171	int sectors_handled;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1172	int max_sectors;
1173	int sectors;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1174
1175	/*
1176	 * Register the new request and wait if the reconstruction
1177	 * thread has put up a bar for new requests.
1178	 * Continue immediately if no resync is active currently.
1179	 */
1180	wait_barrier(conf);
1181
1182	sectors = bio_sectors(bio);
1183	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1184	    bio->bi_iter.bi_sector < conf->reshape_progress &&
1185	    bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1186		/* IO spans the reshape position.  Need to wait for
1187		 * reshape to pass
 
1188		 */
 
1189		allow_barrier(conf);
1190		wait_event(conf->wait_barrier,
1191			   conf->reshape_progress <= bio->bi_iter.bi_sector ||
1192			   conf->reshape_progress >= bio->bi_iter.bi_sector +
1193			   sectors);
1194		wait_barrier(conf);
1195	}
 
1196	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1197	    bio_data_dir(bio) == WRITE &&
1198	    (mddev->reshape_backwards
1199	     ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1200		bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1201	     : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1202		bio->bi_iter.bi_sector < conf->reshape_progress))) {
1203		/* Need to update reshape_position in metadata */
1204		mddev->reshape_position = conf->reshape_progress;
1205		set_bit(MD_CHANGE_DEVS, &mddev->flags);
1206		set_bit(MD_CHANGE_PENDING, &mddev->flags);
1207		md_wakeup_thread(mddev->thread);
 
1208		wait_event(mddev->sb_wait,
1209			   !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1210
1211		conf->reshape_safe = mddev->reshape_position;
1212	}
1213
1214	r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1215
1216	r10_bio->master_bio = bio;
1217	r10_bio->sectors = sectors;
1218
1219	r10_bio->mddev = mddev;
1220	r10_bio->sector = bio->bi_iter.bi_sector;
1221	r10_bio->state = 0;
1222
1223	/* We might need to issue multiple reads to different
1224	 * devices if there are bad blocks around, so we keep
1225	 * track of the number of reads in bio->bi_phys_segments.
1226	 * If this is 0, there is only one r10_bio and no locking
1227	 * will be needed when the request completes.  If it is
1228	 * non-zero, then it is the number of not-completed requests.
1229	 */
1230	bio->bi_phys_segments = 0;
1231	clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1232
1233	if (rw == READ) {
1234		/*
1235		 * read balancing logic:
1236		 */
1237		struct md_rdev *rdev;
1238		int slot;
1239
1240read_again:
1241		rdev = read_balance(conf, r10_bio, &max_sectors);
1242		if (!rdev) {
1243			raid_end_bio_io(r10_bio);
1244			return;
1245		}
1246		slot = r10_bio->read_slot;
1247
1248		read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1249		bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1250			 max_sectors);
1251
1252		r10_bio->devs[slot].bio = read_bio;
1253		r10_bio->devs[slot].rdev = rdev;
1254
1255		read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1256			choose_data_offset(r10_bio, rdev);
1257		read_bio->bi_bdev = rdev->bdev;
1258		read_bio->bi_end_io = raid10_end_read_request;
1259		read_bio->bi_rw = READ | do_sync;
1260		read_bio->bi_private = r10_bio;
1261
1262		if (max_sectors < r10_bio->sectors) {
1263			/* Could not read all from this device, so we will
1264			 * need another r10_bio.
1265			 */
1266			sectors_handled = (r10_bio->sector + max_sectors
1267					   - bio->bi_iter.bi_sector);
1268			r10_bio->sectors = max_sectors;
1269			spin_lock_irq(&conf->device_lock);
1270			if (bio->bi_phys_segments == 0)
1271				bio->bi_phys_segments = 2;
1272			else
1273				bio->bi_phys_segments++;
1274			spin_unlock_irq(&conf->device_lock);
1275			/* Cannot call generic_make_request directly
1276			 * as that will be queued in __generic_make_request
1277			 * and subsequent mempool_alloc might block
1278			 * waiting for it.  so hand bio over to raid10d.
1279			 */
1280			reschedule_retry(r10_bio);
1281
1282			r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1283
1284			r10_bio->master_bio = bio;
1285			r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1286			r10_bio->state = 0;
1287			r10_bio->mddev = mddev;
1288			r10_bio->sector = bio->bi_iter.bi_sector +
1289				sectors_handled;
1290			goto read_again;
1291		} else
1292			generic_make_request(read_bio);
1293		return;
1294	}
1295
1296	/*
1297	 * WRITE:
1298	 */
1299	if (conf->pending_count >= max_queued_requests) {
1300		md_wakeup_thread(mddev->thread);
 
1301		wait_event(conf->wait_barrier,
1302			   conf->pending_count < max_queued_requests);
1303	}
1304	/* first select target devices under rcu_lock and
1305	 * inc refcount on their rdev.  Record them by setting
1306	 * bios[x] to bio
1307	 * If there are known/acknowledged bad blocks on any device
1308	 * on which we have seen a write error, we want to avoid
1309	 * writing to those blocks.  This potentially requires several
1310	 * writes to write around the bad blocks.  Each set of writes
1311	 * gets its own r10_bio with a set of bios attached.  The number
1312	 * of r10_bios is recored in bio->bi_phys_segments just as with
1313	 * the read case.
1314	 */
1315
1316	r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1317	raid10_find_phys(conf, r10_bio);
1318retry_write:
1319	blocked_rdev = NULL;
1320	rcu_read_lock();
1321	max_sectors = r10_bio->sectors;
1322
1323	for (i = 0;  i < conf->copies; i++) {
1324		int d = r10_bio->devs[i].devnum;
1325		struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1326		struct md_rdev *rrdev = rcu_dereference(
1327			conf->mirrors[d].replacement);
1328		if (rdev == rrdev)
1329			rrdev = NULL;
1330		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1331			atomic_inc(&rdev->nr_pending);
1332			blocked_rdev = rdev;
1333			break;
1334		}
1335		if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1336			atomic_inc(&rrdev->nr_pending);
1337			blocked_rdev = rrdev;
1338			break;
1339		}
1340		if (rdev && (test_bit(Faulty, &rdev->flags)
1341			     || test_bit(Unmerged, &rdev->flags)))
1342			rdev = NULL;
1343		if (rrdev && (test_bit(Faulty, &rrdev->flags)
1344			      || test_bit(Unmerged, &rrdev->flags)))
1345			rrdev = NULL;
1346
1347		r10_bio->devs[i].bio = NULL;
1348		r10_bio->devs[i].repl_bio = NULL;
1349
1350		if (!rdev && !rrdev) {
1351			set_bit(R10BIO_Degraded, &r10_bio->state);
1352			continue;
1353		}
1354		if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1355			sector_t first_bad;
1356			sector_t dev_sector = r10_bio->devs[i].addr;
1357			int bad_sectors;
1358			int is_bad;
1359
1360			is_bad = is_badblock(rdev, dev_sector,
1361					     max_sectors,
1362					     &first_bad, &bad_sectors);
1363			if (is_bad < 0) {
1364				/* Mustn't write here until the bad block
1365				 * is acknowledged
1366				 */
1367				atomic_inc(&rdev->nr_pending);
1368				set_bit(BlockedBadBlocks, &rdev->flags);
1369				blocked_rdev = rdev;
1370				break;
1371			}
1372			if (is_bad && first_bad <= dev_sector) {
1373				/* Cannot write here at all */
1374				bad_sectors -= (dev_sector - first_bad);
1375				if (bad_sectors < max_sectors)
1376					/* Mustn't write more than bad_sectors
1377					 * to other devices yet
1378					 */
1379					max_sectors = bad_sectors;
1380				/* We don't set R10BIO_Degraded as that
1381				 * only applies if the disk is missing,
1382				 * so it might be re-added, and we want to
1383				 * know to recover this chunk.
1384				 * In this case the device is here, and the
1385				 * fact that this chunk is not in-sync is
1386				 * recorded in the bad block log.
1387				 */
1388				continue;
1389			}
1390			if (is_bad) {
1391				int good_sectors = first_bad - dev_sector;
1392				if (good_sectors < max_sectors)
1393					max_sectors = good_sectors;
1394			}
1395		}
1396		if (rdev) {
1397			r10_bio->devs[i].bio = bio;
1398			atomic_inc(&rdev->nr_pending);
1399		}
1400		if (rrdev) {
1401			r10_bio->devs[i].repl_bio = bio;
1402			atomic_inc(&rrdev->nr_pending);
1403		}
1404	}
1405	rcu_read_unlock();
1406
1407	if (unlikely(blocked_rdev)) {
1408		/* Have to wait for this device to get unblocked, then retry */
1409		int j;
1410		int d;
1411
1412		for (j = 0; j < i; j++) {
1413			if (r10_bio->devs[j].bio) {
1414				d = r10_bio->devs[j].devnum;
1415				rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1416			}
1417			if (r10_bio->devs[j].repl_bio) {
1418				struct md_rdev *rdev;
1419				d = r10_bio->devs[j].devnum;
1420				rdev = conf->mirrors[d].replacement;
1421				if (!rdev) {
1422					/* Race with remove_disk */
1423					smp_mb();
1424					rdev = conf->mirrors[d].rdev;
1425				}
1426				rdev_dec_pending(rdev, mddev);
1427			}
1428		}
1429		allow_barrier(conf);
 
1430		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1431		wait_barrier(conf);
1432		goto retry_write;
1433	}
1434
1435	if (max_sectors < r10_bio->sectors) {
1436		/* We are splitting this into multiple parts, so
1437		 * we need to prepare for allocating another r10_bio.
1438		 */
1439		r10_bio->sectors = max_sectors;
1440		spin_lock_irq(&conf->device_lock);
1441		if (bio->bi_phys_segments == 0)
1442			bio->bi_phys_segments = 2;
1443		else
1444			bio->bi_phys_segments++;
1445		spin_unlock_irq(&conf->device_lock);
 
 
1446	}
1447	sectors_handled = r10_bio->sector + max_sectors -
1448		bio->bi_iter.bi_sector;
1449
1450	atomic_set(&r10_bio->remaining, 1);
1451	bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1452
1453	for (i = 0; i < conf->copies; i++) {
1454		struct bio *mbio;
1455		int d = r10_bio->devs[i].devnum;
1456		if (r10_bio->devs[i].bio) {
1457			struct md_rdev *rdev = conf->mirrors[d].rdev;
1458			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1459			bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1460				 max_sectors);
1461			r10_bio->devs[i].bio = mbio;
1462
1463			mbio->bi_iter.bi_sector	= (r10_bio->devs[i].addr+
1464					   choose_data_offset(r10_bio,
1465							      rdev));
1466			mbio->bi_bdev = rdev->bdev;
1467			mbio->bi_end_io	= raid10_end_write_request;
1468			mbio->bi_rw =
1469				WRITE | do_sync | do_fua | do_discard | do_same;
1470			mbio->bi_private = r10_bio;
1471
1472			atomic_inc(&r10_bio->remaining);
1473
1474			cb = blk_check_plugged(raid10_unplug, mddev,
1475					       sizeof(*plug));
1476			if (cb)
1477				plug = container_of(cb, struct raid10_plug_cb,
1478						    cb);
1479			else
1480				plug = NULL;
1481			spin_lock_irqsave(&conf->device_lock, flags);
1482			if (plug) {
1483				bio_list_add(&plug->pending, mbio);
1484				plug->pending_cnt++;
1485			} else {
1486				bio_list_add(&conf->pending_bio_list, mbio);
1487				conf->pending_count++;
1488			}
1489			spin_unlock_irqrestore(&conf->device_lock, flags);
1490			if (!plug)
1491				md_wakeup_thread(mddev->thread);
1492		}
1493
1494		if (r10_bio->devs[i].repl_bio) {
1495			struct md_rdev *rdev = conf->mirrors[d].replacement;
1496			if (rdev == NULL) {
1497				/* Replacement just got moved to main 'rdev' */
1498				smp_mb();
1499				rdev = conf->mirrors[d].rdev;
1500			}
1501			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1502			bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1503				 max_sectors);
1504			r10_bio->devs[i].repl_bio = mbio;
1505
1506			mbio->bi_iter.bi_sector	= (r10_bio->devs[i].addr +
1507					   choose_data_offset(
1508						   r10_bio, rdev));
1509			mbio->bi_bdev = rdev->bdev;
1510			mbio->bi_end_io	= raid10_end_write_request;
1511			mbio->bi_rw =
1512				WRITE | do_sync | do_fua | do_discard | do_same;
1513			mbio->bi_private = r10_bio;
1514
1515			atomic_inc(&r10_bio->remaining);
1516			spin_lock_irqsave(&conf->device_lock, flags);
1517			bio_list_add(&conf->pending_bio_list, mbio);
1518			conf->pending_count++;
1519			spin_unlock_irqrestore(&conf->device_lock, flags);
1520			if (!mddev_check_plugged(mddev))
1521				md_wakeup_thread(mddev->thread);
1522		}
1523	}
 
 
1524
1525	/* Don't remove the bias on 'remaining' (one_write_done) until
1526	 * after checking if we need to go around again.
1527	 */
 
1528
1529	if (sectors_handled < bio_sectors(bio)) {
1530		one_write_done(r10_bio);
1531		/* We need another r10_bio.  It has already been counted
1532		 * in bio->bi_phys_segments.
1533		 */
1534		r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1535
1536		r10_bio->master_bio = bio;
1537		r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1538
1539		r10_bio->mddev = mddev;
1540		r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1541		r10_bio->state = 0;
1542		goto retry_write;
1543	}
1544	one_write_done(r10_bio);
 
 
 
1545}
1546
1547static void make_request(struct mddev *mddev, struct bio *bio)
1548{
1549	struct r10conf *conf = mddev->private;
1550	sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1551	int chunk_sects = chunk_mask + 1;
 
1552
1553	struct bio *split;
1554
1555	if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1556		md_flush_request(mddev, bio);
1557		return;
1558	}
1559
1560	md_write_start(mddev, bio);
1561
1562
1563	do {
1564
1565		/*
1566		 * If this request crosses a chunk boundary, we need to split
1567		 * it.
1568		 */
1569		if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1570			     bio_sectors(bio) > chunk_sects
1571			     && (conf->geo.near_copies < conf->geo.raid_disks
1572				 || conf->prev.near_copies <
1573				 conf->prev.raid_disks))) {
1574			split = bio_split(bio, chunk_sects -
1575					  (bio->bi_iter.bi_sector &
1576					   (chunk_sects - 1)),
1577					  GFP_NOIO, fs_bio_set);
1578			bio_chain(split, bio);
1579		} else {
1580			split = bio;
1581		}
1582
1583		__make_request(mddev, split);
1584	} while (split != bio);
 
 
 
 
 
 
 
 
 
 
 
1585
1586	/* In case raid10d snuck in to freeze_array */
1587	wake_up(&conf->wait_barrier);
 
1588}
1589
1590static void status(struct seq_file *seq, struct mddev *mddev)
1591{
1592	struct r10conf *conf = mddev->private;
1593	int i;
1594
1595	if (conf->geo.near_copies < conf->geo.raid_disks)
1596		seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1597	if (conf->geo.near_copies > 1)
1598		seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1599	if (conf->geo.far_copies > 1) {
1600		if (conf->geo.far_offset)
1601			seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1602		else
1603			seq_printf(seq, " %d far-copies", conf->geo.far_copies);
 
 
1604	}
1605	seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1606					conf->geo.raid_disks - mddev->degraded);
1607	for (i = 0; i < conf->geo.raid_disks; i++)
1608		seq_printf(seq, "%s",
1609			      conf->mirrors[i].rdev &&
1610			      test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
 
 
1611	seq_printf(seq, "]");
1612}
1613
1614/* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1618 */
1619static int _enough(struct r10conf *conf, int previous, int ignore)
1620{
1621	int first = 0;
1622	int has_enough = 0;
1623	int disks, ncopies;
1624	if (previous) {
1625		disks = conf->prev.raid_disks;
1626		ncopies = conf->prev.near_copies;
1627	} else {
1628		disks = conf->geo.raid_disks;
1629		ncopies = conf->geo.near_copies;
1630	}
1631
1632	rcu_read_lock();
1633	do {
1634		int n = conf->copies;
1635		int cnt = 0;
1636		int this = first;
1637		while (n--) {
1638			struct md_rdev *rdev;
1639			if (this != ignore &&
1640			    (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1641			    test_bit(In_sync, &rdev->flags))
1642				cnt++;
1643			this = (this+1) % disks;
1644		}
1645		if (cnt == 0)
1646			goto out;
1647		first = (first + ncopies) % disks;
1648	} while (first != 0);
1649	has_enough = 1;
1650out:
1651	rcu_read_unlock();
1652	return has_enough;
1653}
1654
1655static int enough(struct r10conf *conf, int ignore)
1656{
1657	/* when calling 'enough', both 'prev' and 'geo' must
1658	 * be stable.
1659	 * This is ensured if ->reconfig_mutex or ->device_lock
1660	 * is held.
1661	 */
1662	return _enough(conf, 0, ignore) &&
1663		_enough(conf, 1, ignore);
1664}
1665
1666static void error(struct mddev *mddev, struct md_rdev *rdev)
1667{
1668	char b[BDEVNAME_SIZE];
1669	struct r10conf *conf = mddev->private;
1670	unsigned long flags;
1671
1672	/*
1673	 * If it is not operational, then we have already marked it as dead
1674	 * else if it is the last working disks, ignore the error, let the
1675	 * next level up know.
1676	 * else mark the drive as failed
1677	 */
1678	spin_lock_irqsave(&conf->device_lock, flags);
1679	if (test_bit(In_sync, &rdev->flags)
1680	    && !enough(conf, rdev->raid_disk)) {
1681		/*
1682		 * Don't fail the drive, just return an IO error.
1683		 */
1684		spin_unlock_irqrestore(&conf->device_lock, flags);
1685		return;
1686	}
1687	if (test_and_clear_bit(In_sync, &rdev->flags)) {
1688		mddev->degraded++;
1689			/*
1690		 * if recovery is running, make sure it aborts.
1691		 */
1692		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1693	}
1694	set_bit(Blocked, &rdev->flags);
1695	set_bit(Faulty, &rdev->flags);
1696	set_bit(MD_CHANGE_DEVS, &mddev->flags);
 
1697	spin_unlock_irqrestore(&conf->device_lock, flags);
1698	printk(KERN_ALERT
1699	       "md/raid10:%s: Disk failure on %s, disabling device.\n"
1700	       "md/raid10:%s: Operation continuing on %d devices.\n",
1701	       mdname(mddev), bdevname(rdev->bdev, b),
1702	       mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1703}
1704
1705static void print_conf(struct r10conf *conf)
1706{
1707	int i;
1708	struct raid10_info *tmp;
1709
1710	printk(KERN_DEBUG "RAID10 conf printout:\n");
1711	if (!conf) {
1712		printk(KERN_DEBUG "(!conf)\n");
1713		return;
1714	}
1715	printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1716		conf->geo.raid_disks);
1717
 
 
1718	for (i = 0; i < conf->geo.raid_disks; i++) {
1719		char b[BDEVNAME_SIZE];
1720		tmp = conf->mirrors + i;
1721		if (tmp->rdev)
1722			printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1723				i, !test_bit(In_sync, &tmp->rdev->flags),
1724			        !test_bit(Faulty, &tmp->rdev->flags),
1725				bdevname(tmp->rdev->bdev,b));
1726	}
1727}
1728
1729static void close_sync(struct r10conf *conf)
1730{
1731	wait_barrier(conf);
1732	allow_barrier(conf);
1733
1734	mempool_destroy(conf->r10buf_pool);
1735	conf->r10buf_pool = NULL;
1736}
1737
1738static int raid10_spare_active(struct mddev *mddev)
1739{
1740	int i;
1741	struct r10conf *conf = mddev->private;
1742	struct raid10_info *tmp;
1743	int count = 0;
1744	unsigned long flags;
1745
1746	/*
1747	 * Find all non-in_sync disks within the RAID10 configuration
1748	 * and mark them in_sync
1749	 */
1750	for (i = 0; i < conf->geo.raid_disks; i++) {
1751		tmp = conf->mirrors + i;
1752		if (tmp->replacement
1753		    && tmp->replacement->recovery_offset == MaxSector
1754		    && !test_bit(Faulty, &tmp->replacement->flags)
1755		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1756			/* Replacement has just become active */
1757			if (!tmp->rdev
1758			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1759				count++;
1760			if (tmp->rdev) {
1761				/* Replaced device not technically faulty,
1762				 * but we need to be sure it gets removed
1763				 * and never re-added.
1764				 */
1765				set_bit(Faulty, &tmp->rdev->flags);
1766				sysfs_notify_dirent_safe(
1767					tmp->rdev->sysfs_state);
1768			}
1769			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1770		} else if (tmp->rdev
1771			   && tmp->rdev->recovery_offset == MaxSector
1772			   && !test_bit(Faulty, &tmp->rdev->flags)
1773			   && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1774			count++;
1775			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1776		}
1777	}
1778	spin_lock_irqsave(&conf->device_lock, flags);
1779	mddev->degraded -= count;
1780	spin_unlock_irqrestore(&conf->device_lock, flags);
1781
1782	print_conf(conf);
1783	return count;
1784}
1785
1786
1787static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1788{
1789	struct r10conf *conf = mddev->private;
1790	int err = -EEXIST;
1791	int mirror;
1792	int first = 0;
1793	int last = conf->geo.raid_disks - 1;
1794	struct request_queue *q = bdev_get_queue(rdev->bdev);
1795
1796	if (mddev->recovery_cp < MaxSector)
1797		/* only hot-add to in-sync arrays, as recovery is
1798		 * very different from resync
1799		 */
1800		return -EBUSY;
1801	if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1802		return -EINVAL;
1803
 
 
 
1804	if (rdev->raid_disk >= 0)
1805		first = last = rdev->raid_disk;
1806
1807	if (q->merge_bvec_fn) {
1808		set_bit(Unmerged, &rdev->flags);
1809		mddev->merge_check_needed = 1;
1810	}
1811
1812	if (rdev->saved_raid_disk >= first &&
1813	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1814		mirror = rdev->saved_raid_disk;
1815	else
1816		mirror = first;
1817	for ( ; mirror <= last ; mirror++) {
1818		struct raid10_info *p = &conf->mirrors[mirror];
1819		if (p->recovery_disabled == mddev->recovery_disabled)
1820			continue;
1821		if (p->rdev) {
1822			if (!test_bit(WantReplacement, &p->rdev->flags) ||
1823			    p->replacement != NULL)
1824				continue;
1825			clear_bit(In_sync, &rdev->flags);
1826			set_bit(Replacement, &rdev->flags);
1827			rdev->raid_disk = mirror;
1828			err = 0;
1829			if (mddev->gendisk)
1830				disk_stack_limits(mddev->gendisk, rdev->bdev,
1831						  rdev->data_offset << 9);
1832			conf->fullsync = 1;
1833			rcu_assign_pointer(p->replacement, rdev);
1834			break;
1835		}
1836
1837		if (mddev->gendisk)
1838			disk_stack_limits(mddev->gendisk, rdev->bdev,
1839					  rdev->data_offset << 9);
1840
1841		p->head_position = 0;
1842		p->recovery_disabled = mddev->recovery_disabled - 1;
1843		rdev->raid_disk = mirror;
1844		err = 0;
1845		if (rdev->saved_raid_disk != mirror)
1846			conf->fullsync = 1;
1847		rcu_assign_pointer(p->rdev, rdev);
1848		break;
1849	}
1850	if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1851		/* Some requests might not have seen this new
1852		 * merge_bvec_fn.  We must wait for them to complete
1853		 * before merging the device fully.
1854		 * First we make sure any code which has tested
1855		 * our function has submitted the request, then
1856		 * we wait for all outstanding requests to complete.
1857		 */
1858		synchronize_sched();
1859		freeze_array(conf, 0);
1860		unfreeze_array(conf);
1861		clear_bit(Unmerged, &rdev->flags);
1862	}
1863	md_integrity_add_rdev(rdev, mddev);
1864	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1865		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1866
1867	print_conf(conf);
1868	return err;
1869}
1870
1871static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1872{
1873	struct r10conf *conf = mddev->private;
1874	int err = 0;
1875	int number = rdev->raid_disk;
1876	struct md_rdev **rdevp;
1877	struct raid10_info *p = conf->mirrors + number;
1878
1879	print_conf(conf);
1880	if (rdev == p->rdev)
1881		rdevp = &p->rdev;
1882	else if (rdev == p->replacement)
1883		rdevp = &p->replacement;
1884	else
1885		return 0;
1886
1887	if (test_bit(In_sync, &rdev->flags) ||
1888	    atomic_read(&rdev->nr_pending)) {
1889		err = -EBUSY;
1890		goto abort;
1891	}
1892	/* Only remove faulty devices if recovery
1893	 * is not possible.
1894	 */
1895	if (!test_bit(Faulty, &rdev->flags) &&
1896	    mddev->recovery_disabled != p->recovery_disabled &&
1897	    (!p->replacement || p->replacement == rdev) &&
1898	    number < conf->geo.raid_disks &&
1899	    enough(conf, -1)) {
1900		err = -EBUSY;
1901		goto abort;
1902	}
1903	*rdevp = NULL;
1904	synchronize_rcu();
1905	if (atomic_read(&rdev->nr_pending)) {
1906		/* lost the race, try later */
1907		err = -EBUSY;
1908		*rdevp = rdev;
1909		goto abort;
1910	} else if (p->replacement) {
 
 
 
1911		/* We must have just cleared 'rdev' */
1912		p->rdev = p->replacement;
1913		clear_bit(Replacement, &p->replacement->flags);
1914		smp_mb(); /* Make sure other CPUs may see both as identical
1915			   * but will never see neither -- if they are careful.
1916			   */
1917		p->replacement = NULL;
1918		clear_bit(WantReplacement, &rdev->flags);
1919	} else
1920		/* We might have just remove the Replacement as faulty
1921		 * Clear the flag just in case
1922		 */
1923		clear_bit(WantReplacement, &rdev->flags);
1924
 
1925	err = md_integrity_register(mddev);
1926
1927abort:
1928
1929	print_conf(conf);
1930	return err;
1931}
1932
1933
1934static void end_sync_read(struct bio *bio, int error)
1935{
1936	struct r10bio *r10_bio = bio->bi_private;
1937	struct r10conf *conf = r10_bio->mddev->private;
1938	int d;
1939
1940	if (bio == r10_bio->master_bio) {
1941		/* this is a reshape read */
1942		d = r10_bio->read_slot; /* really the read dev */
1943	} else
1944		d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1945
1946	if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1947		set_bit(R10BIO_Uptodate, &r10_bio->state);
1948	else
1949		/* The write handler will notice the lack of
1950		 * R10BIO_Uptodate and record any errors etc
1951		 */
1952		atomic_add(r10_bio->sectors,
1953			   &conf->mirrors[d].rdev->corrected_errors);
1954
1955	/* for reconstruct, we always reschedule after a read.
1956	 * for resync, only after all reads
1957	 */
1958	rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1959	if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1960	    atomic_dec_and_test(&r10_bio->remaining)) {
1961		/* we have read all the blocks,
1962		 * do the comparison in process context in raid10d
1963		 */
1964		reschedule_retry(r10_bio);
1965	}
1966}
1967
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1968static void end_sync_request(struct r10bio *r10_bio)
1969{
1970	struct mddev *mddev = r10_bio->mddev;
1971
1972	while (atomic_dec_and_test(&r10_bio->remaining)) {
1973		if (r10_bio->master_bio == NULL) {
1974			/* the primary of several recovery bios */
1975			sector_t s = r10_bio->sectors;
1976			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1977			    test_bit(R10BIO_WriteError, &r10_bio->state))
1978				reschedule_retry(r10_bio);
1979			else
1980				put_buf(r10_bio);
1981			md_done_sync(mddev, s, 1);
1982			break;
1983		} else {
1984			struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1985			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1986			    test_bit(R10BIO_WriteError, &r10_bio->state))
1987				reschedule_retry(r10_bio);
1988			else
1989				put_buf(r10_bio);
1990			r10_bio = r10_bio2;
1991		}
1992	}
1993}
1994
1995static void end_sync_write(struct bio *bio, int error)
1996{
1997	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1998	struct r10bio *r10_bio = bio->bi_private;
1999	struct mddev *mddev = r10_bio->mddev;
2000	struct r10conf *conf = mddev->private;
2001	int d;
2002	sector_t first_bad;
2003	int bad_sectors;
2004	int slot;
2005	int repl;
2006	struct md_rdev *rdev = NULL;
2007
2008	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2009	if (repl)
2010		rdev = conf->mirrors[d].replacement;
2011	else
2012		rdev = conf->mirrors[d].rdev;
2013
2014	if (!uptodate) {
2015		if (repl)
2016			md_error(mddev, rdev);
2017		else {
2018			set_bit(WriteErrorSeen, &rdev->flags);
2019			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2020				set_bit(MD_RECOVERY_NEEDED,
2021					&rdev->mddev->recovery);
2022			set_bit(R10BIO_WriteError, &r10_bio->state);
2023		}
2024	} else if (is_badblock(rdev,
2025			     r10_bio->devs[slot].addr,
2026			     r10_bio->sectors,
2027			     &first_bad, &bad_sectors))
2028		set_bit(R10BIO_MadeGood, &r10_bio->state);
2029
2030	rdev_dec_pending(rdev, mddev);
2031
2032	end_sync_request(r10_bio);
2033}
2034
2035/*
2036 * Note: sync and recover and handled very differently for raid10
2037 * This code is for resync.
2038 * For resync, we read through virtual addresses and read all blocks.
2039 * If there is any error, we schedule a write.  The lowest numbered
2040 * drive is authoritative.
2041 * However requests come for physical address, so we need to map.
2042 * For every physical address there are raid_disks/copies virtual addresses,
2043 * which is always are least one, but is not necessarly an integer.
2044 * This means that a physical address can span multiple chunks, so we may
2045 * have to submit multiple io requests for a single sync request.
2046 */
2047/*
2048 * We check if all blocks are in-sync and only write to blocks that
2049 * aren't in sync
2050 */
2051static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2052{
2053	struct r10conf *conf = mddev->private;
2054	int i, first;
2055	struct bio *tbio, *fbio;
2056	int vcnt;
 
2057
2058	atomic_set(&r10_bio->remaining, 1);
2059
2060	/* find the first device with a block */
2061	for (i=0; i<conf->copies; i++)
2062		if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2063			break;
2064
2065	if (i == conf->copies)
2066		goto done;
2067
2068	first = i;
2069	fbio = r10_bio->devs[i].bio;
 
 
 
2070
2071	vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2072	/* now find blocks with errors */
2073	for (i=0 ; i < conf->copies ; i++) {
2074		int  j, d;
 
 
2075
2076		tbio = r10_bio->devs[i].bio;
2077
2078		if (tbio->bi_end_io != end_sync_read)
2079			continue;
2080		if (i == first)
2081			continue;
2082		if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
 
 
 
 
2083			/* We know that the bi_io_vec layout is the same for
2084			 * both 'first' and 'i', so we just compare them.
2085			 * All vec entries are PAGE_SIZE;
2086			 */
2087			int sectors = r10_bio->sectors;
2088			for (j = 0; j < vcnt; j++) {
2089				int len = PAGE_SIZE;
2090				if (sectors < (len / 512))
2091					len = sectors * 512;
2092				if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2093					   page_address(tbio->bi_io_vec[j].bv_page),
2094					   len))
2095					break;
2096				sectors -= len/512;
2097			}
2098			if (j == vcnt)
2099				continue;
2100			atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2101			if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2102				/* Don't fix anything. */
2103				continue;
 
 
 
 
2104		}
2105		/* Ok, we need to write this bio, either to correct an
2106		 * inconsistency or to correct an unreadable block.
2107		 * First we need to fixup bv_offset, bv_len and
2108		 * bi_vecs, as the read request might have corrupted these
2109		 */
 
2110		bio_reset(tbio);
2111
2112		tbio->bi_vcnt = vcnt;
2113		tbio->bi_iter.bi_size = r10_bio->sectors << 9;
2114		tbio->bi_rw = WRITE;
2115		tbio->bi_private = r10_bio;
2116		tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2117
2118		for (j=0; j < vcnt ; j++) {
2119			tbio->bi_io_vec[j].bv_offset = 0;
2120			tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2121
2122			memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2123			       page_address(fbio->bi_io_vec[j].bv_page),
2124			       PAGE_SIZE);
2125		}
2126		tbio->bi_end_io = end_sync_write;
 
 
 
2127
2128		d = r10_bio->devs[i].devnum;
2129		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2130		atomic_inc(&r10_bio->remaining);
2131		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2132
 
 
2133		tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2134		tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2135		generic_make_request(tbio);
2136	}
2137
2138	/* Now write out to any replacement devices
2139	 * that are active
2140	 */
2141	for (i = 0; i < conf->copies; i++) {
2142		int j, d;
2143
2144		tbio = r10_bio->devs[i].repl_bio;
2145		if (!tbio || !tbio->bi_end_io)
2146			continue;
2147		if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2148		    && r10_bio->devs[i].bio != fbio)
2149			for (j = 0; j < vcnt; j++)
2150				memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2151				       page_address(fbio->bi_io_vec[j].bv_page),
2152				       PAGE_SIZE);
2153		d = r10_bio->devs[i].devnum;
2154		atomic_inc(&r10_bio->remaining);
2155		md_sync_acct(conf->mirrors[d].replacement->bdev,
2156			     bio_sectors(tbio));
2157		generic_make_request(tbio);
2158	}
2159
2160done:
2161	if (atomic_dec_and_test(&r10_bio->remaining)) {
2162		md_done_sync(mddev, r10_bio->sectors, 1);
2163		put_buf(r10_bio);
2164	}
2165}
2166
2167/*
2168 * Now for the recovery code.
2169 * Recovery happens across physical sectors.
2170 * We recover all non-is_sync drives by finding the virtual address of
2171 * each, and then choose a working drive that also has that virt address.
2172 * There is a separate r10_bio for each non-in_sync drive.
2173 * Only the first two slots are in use. The first for reading,
2174 * The second for writing.
2175 *
2176 */
2177static void fix_recovery_read_error(struct r10bio *r10_bio)
2178{
2179	/* We got a read error during recovery.
2180	 * We repeat the read in smaller page-sized sections.
2181	 * If a read succeeds, write it to the new device or record
2182	 * a bad block if we cannot.
2183	 * If a read fails, record a bad block on both old and
2184	 * new devices.
2185	 */
2186	struct mddev *mddev = r10_bio->mddev;
2187	struct r10conf *conf = mddev->private;
2188	struct bio *bio = r10_bio->devs[0].bio;
2189	sector_t sect = 0;
2190	int sectors = r10_bio->sectors;
2191	int idx = 0;
2192	int dr = r10_bio->devs[0].devnum;
2193	int dw = r10_bio->devs[1].devnum;
 
2194
2195	while (sectors) {
2196		int s = sectors;
2197		struct md_rdev *rdev;
2198		sector_t addr;
2199		int ok;
2200
2201		if (s > (PAGE_SIZE>>9))
2202			s = PAGE_SIZE >> 9;
2203
2204		rdev = conf->mirrors[dr].rdev;
2205		addr = r10_bio->devs[0].addr + sect,
2206		ok = sync_page_io(rdev,
2207				  addr,
2208				  s << 9,
2209				  bio->bi_io_vec[idx].bv_page,
2210				  READ, false);
2211		if (ok) {
2212			rdev = conf->mirrors[dw].rdev;
2213			addr = r10_bio->devs[1].addr + sect;
2214			ok = sync_page_io(rdev,
2215					  addr,
2216					  s << 9,
2217					  bio->bi_io_vec[idx].bv_page,
2218					  WRITE, false);
2219			if (!ok) {
2220				set_bit(WriteErrorSeen, &rdev->flags);
2221				if (!test_and_set_bit(WantReplacement,
2222						      &rdev->flags))
2223					set_bit(MD_RECOVERY_NEEDED,
2224						&rdev->mddev->recovery);
2225			}
2226		}
2227		if (!ok) {
2228			/* We don't worry if we cannot set a bad block -
2229			 * it really is bad so there is no loss in not
2230			 * recording it yet
2231			 */
2232			rdev_set_badblocks(rdev, addr, s, 0);
2233
2234			if (rdev != conf->mirrors[dw].rdev) {
2235				/* need bad block on destination too */
2236				struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2237				addr = r10_bio->devs[1].addr + sect;
2238				ok = rdev_set_badblocks(rdev2, addr, s, 0);
2239				if (!ok) {
2240					/* just abort the recovery */
2241					printk(KERN_NOTICE
2242					       "md/raid10:%s: recovery aborted"
2243					       " due to read error\n",
2244					       mdname(mddev));
2245
2246					conf->mirrors[dw].recovery_disabled
2247						= mddev->recovery_disabled;
2248					set_bit(MD_RECOVERY_INTR,
2249						&mddev->recovery);
2250					break;
2251				}
2252			}
2253		}
2254
2255		sectors -= s;
2256		sect += s;
2257		idx++;
2258	}
2259}
2260
2261static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2262{
2263	struct r10conf *conf = mddev->private;
2264	int d;
2265	struct bio *wbio, *wbio2;
2266
2267	if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2268		fix_recovery_read_error(r10_bio);
2269		end_sync_request(r10_bio);
2270		return;
2271	}
2272
2273	/*
2274	 * share the pages with the first bio
2275	 * and submit the write request
2276	 */
2277	d = r10_bio->devs[1].devnum;
2278	wbio = r10_bio->devs[1].bio;
2279	wbio2 = r10_bio->devs[1].repl_bio;
2280	/* Need to test wbio2->bi_end_io before we call
2281	 * generic_make_request as if the former is NULL,
2282	 * the latter is free to free wbio2.
2283	 */
2284	if (wbio2 && !wbio2->bi_end_io)
2285		wbio2 = NULL;
2286	if (wbio->bi_end_io) {
2287		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2288		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2289		generic_make_request(wbio);
2290	}
2291	if (wbio2) {
2292		atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2293		md_sync_acct(conf->mirrors[d].replacement->bdev,
2294			     bio_sectors(wbio2));
2295		generic_make_request(wbio2);
2296	}
2297}
2298
2299
2300/*
2301 * Used by fix_read_error() to decay the per rdev read_errors.
2302 * We halve the read error count for every hour that has elapsed
2303 * since the last recorded read error.
2304 *
2305 */
2306static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2307{
2308	struct timespec cur_time_mon;
2309	unsigned long hours_since_last;
2310	unsigned int read_errors = atomic_read(&rdev->read_errors);
2311
2312	ktime_get_ts(&cur_time_mon);
2313
2314	if (rdev->last_read_error.tv_sec == 0 &&
2315	    rdev->last_read_error.tv_nsec == 0) {
2316		/* first time we've seen a read error */
2317		rdev->last_read_error = cur_time_mon;
2318		return;
2319	}
2320
2321	hours_since_last = (cur_time_mon.tv_sec -
2322			    rdev->last_read_error.tv_sec) / 3600;
2323
2324	rdev->last_read_error = cur_time_mon;
2325
2326	/*
2327	 * if hours_since_last is > the number of bits in read_errors
2328	 * just set read errors to 0. We do this to avoid
2329	 * overflowing the shift of read_errors by hours_since_last.
2330	 */
2331	if (hours_since_last >= 8 * sizeof(read_errors))
2332		atomic_set(&rdev->read_errors, 0);
2333	else
2334		atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2335}
2336
2337static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2338			    int sectors, struct page *page, int rw)
2339{
2340	sector_t first_bad;
2341	int bad_sectors;
2342
2343	if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2344	    && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2345		return -1;
2346	if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2347		/* success */
2348		return 1;
2349	if (rw == WRITE) {
2350		set_bit(WriteErrorSeen, &rdev->flags);
2351		if (!test_and_set_bit(WantReplacement, &rdev->flags))
2352			set_bit(MD_RECOVERY_NEEDED,
2353				&rdev->mddev->recovery);
2354	}
2355	/* need to record an error - either for the block or the device */
2356	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2357		md_error(rdev->mddev, rdev);
2358	return 0;
2359}
2360
2361/*
2362 * This is a kernel thread which:
2363 *
2364 *	1.	Retries failed read operations on working mirrors.
2365 *	2.	Updates the raid superblock when problems encounter.
2366 *	3.	Performs writes following reads for array synchronising.
2367 */
2368
2369static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2370{
2371	int sect = 0; /* Offset from r10_bio->sector */
2372	int sectors = r10_bio->sectors;
2373	struct md_rdev*rdev;
2374	int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2375	int d = r10_bio->devs[r10_bio->read_slot].devnum;
2376
2377	/* still own a reference to this rdev, so it cannot
2378	 * have been cleared recently.
2379	 */
2380	rdev = conf->mirrors[d].rdev;
2381
2382	if (test_bit(Faulty, &rdev->flags))
2383		/* drive has already been failed, just ignore any
2384		   more fix_read_error() attempts */
2385		return;
2386
2387	check_decay_read_errors(mddev, rdev);
2388	atomic_inc(&rdev->read_errors);
2389	if (atomic_read(&rdev->read_errors) > max_read_errors) {
2390		char b[BDEVNAME_SIZE];
2391		bdevname(rdev->bdev, b);
2392
2393		printk(KERN_NOTICE
2394		       "md/raid10:%s: %s: Raid device exceeded "
2395		       "read_error threshold [cur %d:max %d]\n",
2396		       mdname(mddev), b,
2397		       atomic_read(&rdev->read_errors), max_read_errors);
2398		printk(KERN_NOTICE
2399		       "md/raid10:%s: %s: Failing raid device\n",
2400		       mdname(mddev), b);
2401		md_error(mddev, conf->mirrors[d].rdev);
2402		r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2403		return;
2404	}
2405
2406	while(sectors) {
2407		int s = sectors;
2408		int sl = r10_bio->read_slot;
2409		int success = 0;
2410		int start;
2411
2412		if (s > (PAGE_SIZE>>9))
2413			s = PAGE_SIZE >> 9;
2414
2415		rcu_read_lock();
2416		do {
2417			sector_t first_bad;
2418			int bad_sectors;
2419
2420			d = r10_bio->devs[sl].devnum;
2421			rdev = rcu_dereference(conf->mirrors[d].rdev);
2422			if (rdev &&
2423			    !test_bit(Unmerged, &rdev->flags) &&
2424			    test_bit(In_sync, &rdev->flags) &&
 
2425			    is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2426					&first_bad, &bad_sectors) == 0) {
2427				atomic_inc(&rdev->nr_pending);
2428				rcu_read_unlock();
2429				success = sync_page_io(rdev,
2430						       r10_bio->devs[sl].addr +
2431						       sect,
2432						       s<<9,
2433						       conf->tmppage, READ, false);
 
2434				rdev_dec_pending(rdev, mddev);
2435				rcu_read_lock();
2436				if (success)
2437					break;
2438			}
2439			sl++;
2440			if (sl == conf->copies)
2441				sl = 0;
2442		} while (!success && sl != r10_bio->read_slot);
2443		rcu_read_unlock();
2444
2445		if (!success) {
2446			/* Cannot read from anywhere, just mark the block
2447			 * as bad on the first device to discourage future
2448			 * reads.
2449			 */
2450			int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2451			rdev = conf->mirrors[dn].rdev;
2452
2453			if (!rdev_set_badblocks(
2454				    rdev,
2455				    r10_bio->devs[r10_bio->read_slot].addr
2456				    + sect,
2457				    s, 0)) {
2458				md_error(mddev, rdev);
2459				r10_bio->devs[r10_bio->read_slot].bio
2460					= IO_BLOCKED;
2461			}
2462			break;
2463		}
2464
2465		start = sl;
2466		/* write it back and re-read */
2467		rcu_read_lock();
2468		while (sl != r10_bio->read_slot) {
2469			char b[BDEVNAME_SIZE];
2470
2471			if (sl==0)
2472				sl = conf->copies;
2473			sl--;
2474			d = r10_bio->devs[sl].devnum;
2475			rdev = rcu_dereference(conf->mirrors[d].rdev);
2476			if (!rdev ||
2477			    test_bit(Unmerged, &rdev->flags) ||
2478			    !test_bit(In_sync, &rdev->flags))
2479				continue;
2480
2481			atomic_inc(&rdev->nr_pending);
2482			rcu_read_unlock();
2483			if (r10_sync_page_io(rdev,
2484					     r10_bio->devs[sl].addr +
2485					     sect,
2486					     s, conf->tmppage, WRITE)
2487			    == 0) {
2488				/* Well, this device is dead */
2489				printk(KERN_NOTICE
2490				       "md/raid10:%s: read correction "
2491				       "write failed"
2492				       " (%d sectors at %llu on %s)\n",
2493				       mdname(mddev), s,
2494				       (unsigned long long)(
2495					       sect +
2496					       choose_data_offset(r10_bio,
2497								  rdev)),
2498				       bdevname(rdev->bdev, b));
2499				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2500				       "drive\n",
2501				       mdname(mddev),
2502				       bdevname(rdev->bdev, b));
2503			}
2504			rdev_dec_pending(rdev, mddev);
2505			rcu_read_lock();
2506		}
2507		sl = start;
2508		while (sl != r10_bio->read_slot) {
2509			char b[BDEVNAME_SIZE];
2510
2511			if (sl==0)
2512				sl = conf->copies;
2513			sl--;
2514			d = r10_bio->devs[sl].devnum;
2515			rdev = rcu_dereference(conf->mirrors[d].rdev);
2516			if (!rdev ||
 
2517			    !test_bit(In_sync, &rdev->flags))
2518				continue;
2519
2520			atomic_inc(&rdev->nr_pending);
2521			rcu_read_unlock();
2522			switch (r10_sync_page_io(rdev,
2523					     r10_bio->devs[sl].addr +
2524					     sect,
2525					     s, conf->tmppage,
2526						 READ)) {
2527			case 0:
2528				/* Well, this device is dead */
2529				printk(KERN_NOTICE
2530				       "md/raid10:%s: unable to read back "
2531				       "corrected sectors"
2532				       " (%d sectors at %llu on %s)\n",
2533				       mdname(mddev), s,
2534				       (unsigned long long)(
2535					       sect +
2536					       choose_data_offset(r10_bio, rdev)),
2537				       bdevname(rdev->bdev, b));
2538				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2539				       "drive\n",
2540				       mdname(mddev),
2541				       bdevname(rdev->bdev, b));
2542				break;
2543			case 1:
2544				printk(KERN_INFO
2545				       "md/raid10:%s: read error corrected"
2546				       " (%d sectors at %llu on %s)\n",
2547				       mdname(mddev), s,
2548				       (unsigned long long)(
2549					       sect +
2550					       choose_data_offset(r10_bio, rdev)),
2551				       bdevname(rdev->bdev, b));
2552				atomic_add(s, &rdev->corrected_errors);
2553			}
2554
2555			rdev_dec_pending(rdev, mddev);
2556			rcu_read_lock();
2557		}
2558		rcu_read_unlock();
2559
2560		sectors -= s;
2561		sect += s;
2562	}
2563}
2564
2565static int narrow_write_error(struct r10bio *r10_bio, int i)
2566{
2567	struct bio *bio = r10_bio->master_bio;
2568	struct mddev *mddev = r10_bio->mddev;
2569	struct r10conf *conf = mddev->private;
2570	struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2571	/* bio has the data to be written to slot 'i' where
2572	 * we just recently had a write error.
2573	 * We repeatedly clone the bio and trim down to one block,
2574	 * then try the write.  Where the write fails we record
2575	 * a bad block.
2576	 * It is conceivable that the bio doesn't exactly align with
2577	 * blocks.  We must handle this.
2578	 *
2579	 * We currently own a reference to the rdev.
2580	 */
2581
2582	int block_sectors;
2583	sector_t sector;
2584	int sectors;
2585	int sect_to_write = r10_bio->sectors;
2586	int ok = 1;
2587
2588	if (rdev->badblocks.shift < 0)
2589		return 0;
2590
2591	block_sectors = 1 << rdev->badblocks.shift;
 
2592	sector = r10_bio->sector;
2593	sectors = ((r10_bio->sector + block_sectors)
2594		   & ~(sector_t)(block_sectors - 1))
2595		- sector;
2596
2597	while (sect_to_write) {
2598		struct bio *wbio;
 
2599		if (sectors > sect_to_write)
2600			sectors = sect_to_write;
2601		/* Write at 'sector' for 'sectors' */
2602		wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2603		bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2604		wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2605				   choose_data_offset(r10_bio, rdev) +
2606				   (sector - r10_bio->sector));
2607		wbio->bi_bdev = rdev->bdev;
2608		if (submit_bio_wait(WRITE, wbio) == 0)
 
 
2609			/* Failure! */
2610			ok = rdev_set_badblocks(rdev, sector,
2611						sectors, 0)
2612				&& ok;
2613
2614		bio_put(wbio);
2615		sect_to_write -= sectors;
2616		sector += sectors;
2617		sectors = block_sectors;
2618	}
2619	return ok;
2620}
2621
2622static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2623{
2624	int slot = r10_bio->read_slot;
2625	struct bio *bio;
2626	struct r10conf *conf = mddev->private;
2627	struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2628	char b[BDEVNAME_SIZE];
2629	unsigned long do_sync;
2630	int max_sectors;
2631
2632	/* we got a read error. Maybe the drive is bad.  Maybe just
2633	 * the block and we can fix it.
2634	 * We freeze all other IO, and try reading the block from
2635	 * other devices.  When we find one, we re-write
2636	 * and check it that fixes the read error.
2637	 * This is all done synchronously while the array is
2638	 * frozen.
2639	 */
2640	bio = r10_bio->devs[slot].bio;
2641	bdevname(bio->bi_bdev, b);
2642	bio_put(bio);
2643	r10_bio->devs[slot].bio = NULL;
2644
2645	if (mddev->ro == 0) {
 
 
2646		freeze_array(conf, 1);
2647		fix_read_error(conf, mddev, r10_bio);
2648		unfreeze_array(conf);
2649	} else
2650		r10_bio->devs[slot].bio = IO_BLOCKED;
2651
2652	rdev_dec_pending(rdev, mddev);
2653
2654read_more:
2655	rdev = read_balance(conf, r10_bio, &max_sectors);
2656	if (rdev == NULL) {
2657		printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2658		       " read error for block %llu\n",
2659		       mdname(mddev), b,
2660		       (unsigned long long)r10_bio->sector);
2661		raid_end_bio_io(r10_bio);
2662		return;
2663	}
2664
2665	do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2666	slot = r10_bio->read_slot;
2667	printk_ratelimited(
2668		KERN_ERR
2669		"md/raid10:%s: %s: redirecting "
2670		"sector %llu to another mirror\n",
2671		mdname(mddev),
2672		bdevname(rdev->bdev, b),
2673		(unsigned long long)r10_bio->sector);
2674	bio = bio_clone_mddev(r10_bio->master_bio,
2675			      GFP_NOIO, mddev);
2676	bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2677	r10_bio->devs[slot].bio = bio;
2678	r10_bio->devs[slot].rdev = rdev;
2679	bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2680		+ choose_data_offset(r10_bio, rdev);
2681	bio->bi_bdev = rdev->bdev;
2682	bio->bi_rw = READ | do_sync;
2683	bio->bi_private = r10_bio;
2684	bio->bi_end_io = raid10_end_read_request;
2685	if (max_sectors < r10_bio->sectors) {
2686		/* Drat - have to split this up more */
2687		struct bio *mbio = r10_bio->master_bio;
2688		int sectors_handled =
2689			r10_bio->sector + max_sectors
2690			- mbio->bi_iter.bi_sector;
2691		r10_bio->sectors = max_sectors;
2692		spin_lock_irq(&conf->device_lock);
2693		if (mbio->bi_phys_segments == 0)
2694			mbio->bi_phys_segments = 2;
2695		else
2696			mbio->bi_phys_segments++;
2697		spin_unlock_irq(&conf->device_lock);
2698		generic_make_request(bio);
2699
2700		r10_bio = mempool_alloc(conf->r10bio_pool,
2701					GFP_NOIO);
2702		r10_bio->master_bio = mbio;
2703		r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2704		r10_bio->state = 0;
2705		set_bit(R10BIO_ReadError,
2706			&r10_bio->state);
2707		r10_bio->mddev = mddev;
2708		r10_bio->sector = mbio->bi_iter.bi_sector
2709			+ sectors_handled;
2710
2711		goto read_more;
2712	} else
2713		generic_make_request(bio);
2714}
2715
2716static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2717{
2718	/* Some sort of write request has finished and it
2719	 * succeeded in writing where we thought there was a
2720	 * bad block.  So forget the bad block.
2721	 * Or possibly if failed and we need to record
2722	 * a bad block.
2723	 */
2724	int m;
2725	struct md_rdev *rdev;
2726
2727	if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2728	    test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2729		for (m = 0; m < conf->copies; m++) {
2730			int dev = r10_bio->devs[m].devnum;
2731			rdev = conf->mirrors[dev].rdev;
2732			if (r10_bio->devs[m].bio == NULL)
 
2733				continue;
2734			if (test_bit(BIO_UPTODATE,
2735				     &r10_bio->devs[m].bio->bi_flags)) {
2736				rdev_clear_badblocks(
2737					rdev,
2738					r10_bio->devs[m].addr,
2739					r10_bio->sectors, 0);
2740			} else {
2741				if (!rdev_set_badblocks(
2742					    rdev,
2743					    r10_bio->devs[m].addr,
2744					    r10_bio->sectors, 0))
2745					md_error(conf->mddev, rdev);
2746			}
2747			rdev = conf->mirrors[dev].replacement;
2748			if (r10_bio->devs[m].repl_bio == NULL)
 
2749				continue;
2750			if (test_bit(BIO_UPTODATE,
2751				     &r10_bio->devs[m].repl_bio->bi_flags)) {
2752				rdev_clear_badblocks(
2753					rdev,
2754					r10_bio->devs[m].addr,
2755					r10_bio->sectors, 0);
2756			} else {
2757				if (!rdev_set_badblocks(
2758					    rdev,
2759					    r10_bio->devs[m].addr,
2760					    r10_bio->sectors, 0))
2761					md_error(conf->mddev, rdev);
2762			}
2763		}
2764		put_buf(r10_bio);
2765	} else {
 
2766		for (m = 0; m < conf->copies; m++) {
2767			int dev = r10_bio->devs[m].devnum;
2768			struct bio *bio = r10_bio->devs[m].bio;
2769			rdev = conf->mirrors[dev].rdev;
2770			if (bio == IO_MADE_GOOD) {
2771				rdev_clear_badblocks(
2772					rdev,
2773					r10_bio->devs[m].addr,
2774					r10_bio->sectors, 0);
2775				rdev_dec_pending(rdev, conf->mddev);
2776			} else if (bio != NULL &&
2777				   !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2778				if (!narrow_write_error(r10_bio, m)) {
2779					md_error(conf->mddev, rdev);
2780					set_bit(R10BIO_Degraded,
2781						&r10_bio->state);
2782				}
2783				rdev_dec_pending(rdev, conf->mddev);
2784			}
2785			bio = r10_bio->devs[m].repl_bio;
2786			rdev = conf->mirrors[dev].replacement;
2787			if (rdev && bio == IO_MADE_GOOD) {
2788				rdev_clear_badblocks(
2789					rdev,
2790					r10_bio->devs[m].addr,
2791					r10_bio->sectors, 0);
2792				rdev_dec_pending(rdev, conf->mddev);
2793			}
2794		}
2795		if (test_bit(R10BIO_WriteError,
2796			     &r10_bio->state))
2797			close_write(r10_bio);
2798		raid_end_bio_io(r10_bio);
 
 
 
 
 
 
 
 
 
 
 
 
 
2799	}
2800}
2801
2802static void raid10d(struct md_thread *thread)
2803{
2804	struct mddev *mddev = thread->mddev;
2805	struct r10bio *r10_bio;
2806	unsigned long flags;
2807	struct r10conf *conf = mddev->private;
2808	struct list_head *head = &conf->retry_list;
2809	struct blk_plug plug;
2810
2811	md_check_recovery(mddev);
2812
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2813	blk_start_plug(&plug);
2814	for (;;) {
2815
2816		flush_pending_writes(conf);
2817
2818		spin_lock_irqsave(&conf->device_lock, flags);
2819		if (list_empty(head)) {
2820			spin_unlock_irqrestore(&conf->device_lock, flags);
2821			break;
2822		}
2823		r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2824		list_del(head->prev);
2825		conf->nr_queued--;
2826		spin_unlock_irqrestore(&conf->device_lock, flags);
2827
2828		mddev = r10_bio->mddev;
2829		conf = mddev->private;
2830		if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2831		    test_bit(R10BIO_WriteError, &r10_bio->state))
2832			handle_write_completed(conf, r10_bio);
2833		else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2834			reshape_request_write(mddev, r10_bio);
2835		else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2836			sync_request_write(mddev, r10_bio);
2837		else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2838			recovery_request_write(mddev, r10_bio);
2839		else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2840			handle_read_error(mddev, r10_bio);
2841		else {
2842			/* just a partial read to be scheduled from a
2843			 * separate context
2844			 */
2845			int slot = r10_bio->read_slot;
2846			generic_make_request(r10_bio->devs[slot].bio);
2847		}
2848
2849		cond_resched();
2850		if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2851			md_check_recovery(mddev);
2852	}
2853	blk_finish_plug(&plug);
2854}
2855
2856
2857static int init_resync(struct r10conf *conf)
2858{
2859	int buffs;
2860	int i;
2861
2862	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2863	BUG_ON(conf->r10buf_pool);
2864	conf->have_replacement = 0;
2865	for (i = 0; i < conf->geo.raid_disks; i++)
2866		if (conf->mirrors[i].replacement)
2867			conf->have_replacement = 1;
2868	conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2869	if (!conf->r10buf_pool)
2870		return -ENOMEM;
2871	conf->next_resync = 0;
2872	return 0;
2873}
2874
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2875/*
2876 * perform a "sync" on one "block"
2877 *
2878 * We need to make sure that no normal I/O request - particularly write
2879 * requests - conflict with active sync requests.
2880 *
2881 * This is achieved by tracking pending requests and a 'barrier' concept
2882 * that can be installed to exclude normal IO requests.
2883 *
2884 * Resync and recovery are handled very differently.
2885 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2886 *
2887 * For resync, we iterate over virtual addresses, read all copies,
2888 * and update if there are differences.  If only one copy is live,
2889 * skip it.
2890 * For recovery, we iterate over physical addresses, read a good
2891 * value for each non-in_sync drive, and over-write.
2892 *
2893 * So, for recovery we may have several outstanding complex requests for a
2894 * given address, one for each out-of-sync device.  We model this by allocating
2895 * a number of r10_bio structures, one for each out-of-sync device.
2896 * As we setup these structures, we collect all bio's together into a list
2897 * which we then process collectively to add pages, and then process again
2898 * to pass to generic_make_request.
2899 *
2900 * The r10_bio structures are linked using a borrowed master_bio pointer.
2901 * This link is counted in ->remaining.  When the r10_bio that points to NULL
2902 * has its remaining count decremented to 0, the whole complex operation
2903 * is complete.
2904 *
2905 */
2906
2907static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2908			     int *skipped, int go_faster)
2909{
2910	struct r10conf *conf = mddev->private;
2911	struct r10bio *r10_bio;
2912	struct bio *biolist = NULL, *bio;
2913	sector_t max_sector, nr_sectors;
2914	int i;
2915	int max_sync;
2916	sector_t sync_blocks;
2917	sector_t sectors_skipped = 0;
2918	int chunks_skipped = 0;
2919	sector_t chunk_mask = conf->geo.chunk_mask;
 
2920
2921	if (!conf->r10buf_pool)
2922		if (init_resync(conf))
2923			return 0;
2924
2925	/*
2926	 * Allow skipping a full rebuild for incremental assembly
2927	 * of a clean array, like RAID1 does.
2928	 */
2929	if (mddev->bitmap == NULL &&
2930	    mddev->recovery_cp == MaxSector &&
2931	    mddev->reshape_position == MaxSector &&
2932	    !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2933	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2934	    !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2935	    conf->fullsync == 0) {
2936		*skipped = 1;
2937		return mddev->dev_sectors - sector_nr;
2938	}
2939
2940 skipped:
2941	max_sector = mddev->dev_sectors;
2942	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2943	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2944		max_sector = mddev->resync_max_sectors;
2945	if (sector_nr >= max_sector) {
 
 
 
2946		/* If we aborted, we need to abort the
2947		 * sync on the 'current' bitmap chucks (there can
2948		 * be several when recovering multiple devices).
2949		 * as we may have started syncing it but not finished.
2950		 * We can find the current address in
2951		 * mddev->curr_resync, but for recovery,
2952		 * we need to convert that to several
2953		 * virtual addresses.
2954		 */
2955		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2956			end_reshape(conf);
 
2957			return 0;
2958		}
2959
2960		if (mddev->curr_resync < max_sector) { /* aborted */
2961			if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2962				bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2963						&sync_blocks, 1);
2964			else for (i = 0; i < conf->geo.raid_disks; i++) {
2965				sector_t sect =
2966					raid10_find_virt(conf, mddev->curr_resync, i);
2967				bitmap_end_sync(mddev->bitmap, sect,
2968						&sync_blocks, 1);
2969			}
2970		} else {
2971			/* completed sync */
2972			if ((!mddev->bitmap || conf->fullsync)
2973			    && conf->have_replacement
2974			    && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2975				/* Completed a full sync so the replacements
2976				 * are now fully recovered.
2977				 */
2978				for (i = 0; i < conf->geo.raid_disks; i++)
2979					if (conf->mirrors[i].replacement)
2980						conf->mirrors[i].replacement
2981							->recovery_offset
2982							= MaxSector;
 
 
 
2983			}
2984			conf->fullsync = 0;
2985		}
2986		bitmap_close_sync(mddev->bitmap);
2987		close_sync(conf);
2988		*skipped = 1;
2989		return sectors_skipped;
2990	}
2991
2992	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2993		return reshape_request(mddev, sector_nr, skipped);
2994
2995	if (chunks_skipped >= conf->geo.raid_disks) {
2996		/* if there has been nothing to do on any drive,
2997		 * then there is nothing to do at all..
2998		 */
2999		*skipped = 1;
3000		return (max_sector - sector_nr) + sectors_skipped;
3001	}
3002
3003	if (max_sector > mddev->resync_max)
3004		max_sector = mddev->resync_max; /* Don't do IO beyond here */
3005
3006	/* make sure whole request will fit in a chunk - if chunks
3007	 * are meaningful
3008	 */
3009	if (conf->geo.near_copies < conf->geo.raid_disks &&
3010	    max_sector > (sector_nr | chunk_mask))
3011		max_sector = (sector_nr | chunk_mask) + 1;
 
3012	/*
3013	 * If there is non-resync activity waiting for us then
3014	 * put in a delay to throttle resync.
3015	 */
3016	if (!go_faster && conf->nr_waiting)
3017		msleep_interruptible(1000);
3018
3019	/* Again, very different code for resync and recovery.
3020	 * Both must result in an r10bio with a list of bios that
3021	 * have bi_end_io, bi_sector, bi_bdev set,
3022	 * and bi_private set to the r10bio.
3023	 * For recovery, we may actually create several r10bios
3024	 * with 2 bios in each, that correspond to the bios in the main one.
3025	 * In this case, the subordinate r10bios link back through a
3026	 * borrowed master_bio pointer, and the counter in the master
3027	 * includes a ref from each subordinate.
3028	 */
3029	/* First, we decide what to do and set ->bi_end_io
3030	 * To end_sync_read if we want to read, and
3031	 * end_sync_write if we will want to write.
3032	 */
3033
3034	max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3035	if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3036		/* recovery... the complicated one */
3037		int j;
3038		r10_bio = NULL;
3039
3040		for (i = 0 ; i < conf->geo.raid_disks; i++) {
3041			int still_degraded;
3042			struct r10bio *rb2;
3043			sector_t sect;
3044			int must_sync;
3045			int any_working;
3046			struct raid10_info *mirror = &conf->mirrors[i];
 
 
 
 
 
3047
3048			if ((mirror->rdev == NULL ||
3049			     test_bit(In_sync, &mirror->rdev->flags))
3050			    &&
3051			    (mirror->replacement == NULL ||
3052			     test_bit(Faulty,
3053				      &mirror->replacement->flags)))
3054				continue;
 
3055
3056			still_degraded = 0;
3057			/* want to reconstruct this device */
3058			rb2 = r10_bio;
3059			sect = raid10_find_virt(conf, sector_nr, i);
3060			if (sect >= mddev->resync_max_sectors) {
3061				/* last stripe is not complete - don't
3062				 * try to recover this sector.
3063				 */
 
3064				continue;
3065			}
 
 
3066			/* Unless we are doing a full sync, or a replacement
3067			 * we only need to recover the block if it is set in
3068			 * the bitmap
3069			 */
3070			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3071						      &sync_blocks, 1);
3072			if (sync_blocks < max_sync)
3073				max_sync = sync_blocks;
3074			if (!must_sync &&
3075			    mirror->replacement == NULL &&
3076			    !conf->fullsync) {
3077				/* yep, skip the sync_blocks here, but don't assume
3078				 * that there will never be anything to do here
3079				 */
3080				chunks_skipped = -1;
 
3081				continue;
3082			}
 
 
 
 
3083
3084			r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
 
3085			raise_barrier(conf, rb2 != NULL);
3086			atomic_set(&r10_bio->remaining, 0);
3087
3088			r10_bio->master_bio = (struct bio*)rb2;
3089			if (rb2)
3090				atomic_inc(&rb2->remaining);
3091			r10_bio->mddev = mddev;
3092			set_bit(R10BIO_IsRecover, &r10_bio->state);
3093			r10_bio->sector = sect;
3094
3095			raid10_find_phys(conf, r10_bio);
3096
3097			/* Need to check if the array will still be
3098			 * degraded
3099			 */
3100			for (j = 0; j < conf->geo.raid_disks; j++)
3101				if (conf->mirrors[j].rdev == NULL ||
3102				    test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
 
 
3103					still_degraded = 1;
3104					break;
3105				}
 
3106
3107			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3108						      &sync_blocks, still_degraded);
3109
3110			any_working = 0;
3111			for (j=0; j<conf->copies;j++) {
3112				int k;
3113				int d = r10_bio->devs[j].devnum;
3114				sector_t from_addr, to_addr;
3115				struct md_rdev *rdev;
 
3116				sector_t sector, first_bad;
3117				int bad_sectors;
3118				if (!conf->mirrors[d].rdev ||
3119				    !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3120					continue;
3121				/* This is where we read from */
3122				any_working = 1;
3123				rdev = conf->mirrors[d].rdev;
3124				sector = r10_bio->devs[j].addr;
3125
3126				if (is_badblock(rdev, sector, max_sync,
3127						&first_bad, &bad_sectors)) {
3128					if (first_bad > sector)
3129						max_sync = first_bad - sector;
3130					else {
3131						bad_sectors -= (sector
3132								- first_bad);
3133						if (max_sync > bad_sectors)
3134							max_sync = bad_sectors;
3135						continue;
3136					}
3137				}
3138				bio = r10_bio->devs[0].bio;
3139				bio_reset(bio);
3140				bio->bi_next = biolist;
3141				biolist = bio;
3142				bio->bi_private = r10_bio;
3143				bio->bi_end_io = end_sync_read;
3144				bio->bi_rw = READ;
 
 
3145				from_addr = r10_bio->devs[j].addr;
3146				bio->bi_iter.bi_sector = from_addr +
3147					rdev->data_offset;
3148				bio->bi_bdev = rdev->bdev;
3149				atomic_inc(&rdev->nr_pending);
3150				/* and we write to 'i' (if not in_sync) */
3151
3152				for (k=0; k<conf->copies; k++)
3153					if (r10_bio->devs[k].devnum == i)
3154						break;
3155				BUG_ON(k == conf->copies);
3156				to_addr = r10_bio->devs[k].addr;
3157				r10_bio->devs[0].devnum = d;
3158				r10_bio->devs[0].addr = from_addr;
3159				r10_bio->devs[1].devnum = i;
3160				r10_bio->devs[1].addr = to_addr;
3161
3162				rdev = mirror->rdev;
3163				if (!test_bit(In_sync, &rdev->flags)) {
3164					bio = r10_bio->devs[1].bio;
3165					bio_reset(bio);
3166					bio->bi_next = biolist;
3167					biolist = bio;
3168					bio->bi_private = r10_bio;
3169					bio->bi_end_io = end_sync_write;
3170					bio->bi_rw = WRITE;
3171					bio->bi_iter.bi_sector = to_addr
3172						+ rdev->data_offset;
3173					bio->bi_bdev = rdev->bdev;
3174					atomic_inc(&r10_bio->remaining);
3175				} else
3176					r10_bio->devs[1].bio->bi_end_io = NULL;
3177
3178				/* and maybe write to replacement */
3179				bio = r10_bio->devs[1].repl_bio;
3180				if (bio)
3181					bio->bi_end_io = NULL;
3182				rdev = mirror->replacement;
3183				/* Note: if rdev != NULL, then bio
3184				 * cannot be NULL as r10buf_pool_alloc will
3185				 * have allocated it.
3186				 * So the second test here is pointless.
3187				 * But it keeps semantic-checkers happy, and
3188				 * this comment keeps human reviewers
3189				 * happy.
3190				 */
3191				if (rdev == NULL || bio == NULL ||
3192				    test_bit(Faulty, &rdev->flags))
3193					break;
3194				bio_reset(bio);
3195				bio->bi_next = biolist;
3196				biolist = bio;
3197				bio->bi_private = r10_bio;
3198				bio->bi_end_io = end_sync_write;
3199				bio->bi_rw = WRITE;
3200				bio->bi_iter.bi_sector = to_addr +
3201					rdev->data_offset;
3202				bio->bi_bdev = rdev->bdev;
3203				atomic_inc(&r10_bio->remaining);
3204				break;
3205			}
 
3206			if (j == conf->copies) {
3207				/* Cannot recover, so abort the recovery or
3208				 * record a bad block */
3209				if (any_working) {
3210					/* problem is that there are bad blocks
3211					 * on other device(s)
3212					 */
3213					int k;
3214					for (k = 0; k < conf->copies; k++)
3215						if (r10_bio->devs[k].devnum == i)
3216							break;
3217					if (!test_bit(In_sync,
3218						      &mirror->rdev->flags)
3219					    && !rdev_set_badblocks(
3220						    mirror->rdev,
3221						    r10_bio->devs[k].addr,
3222						    max_sync, 0))
3223						any_working = 0;
3224					if (mirror->replacement &&
3225					    !rdev_set_badblocks(
3226						    mirror->replacement,
3227						    r10_bio->devs[k].addr,
3228						    max_sync, 0))
3229						any_working = 0;
3230				}
3231				if (!any_working)  {
3232					if (!test_and_set_bit(MD_RECOVERY_INTR,
3233							      &mddev->recovery))
3234						printk(KERN_INFO "md/raid10:%s: insufficient "
3235						       "working devices for recovery.\n",
3236						       mdname(mddev));
3237					mirror->recovery_disabled
3238						= mddev->recovery_disabled;
3239				}
3240				put_buf(r10_bio);
3241				if (rb2)
3242					atomic_dec(&rb2->remaining);
3243				r10_bio = rb2;
 
 
 
3244				break;
3245			}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3246		}
3247		if (biolist == NULL) {
3248			while (r10_bio) {
3249				struct r10bio *rb2 = r10_bio;
3250				r10_bio = (struct r10bio*) rb2->master_bio;
3251				rb2->master_bio = NULL;
3252				put_buf(rb2);
3253			}
3254			goto giveup;
3255		}
3256	} else {
3257		/* resync. Schedule a read for every block at this virt offset */
3258		int count = 0;
3259
3260		bitmap_cond_end_sync(mddev->bitmap, sector_nr);
 
 
 
 
 
 
 
 
 
 
3261
3262		if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3263				       &sync_blocks, mddev->degraded) &&
3264		    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3265						 &mddev->recovery)) {
3266			/* We can skip this block */
3267			*skipped = 1;
3268			return sync_blocks + sectors_skipped;
3269		}
3270		if (sync_blocks < max_sync)
3271			max_sync = sync_blocks;
3272		r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
 
3273
3274		r10_bio->mddev = mddev;
3275		atomic_set(&r10_bio->remaining, 0);
3276		raise_barrier(conf, 0);
3277		conf->next_resync = sector_nr;
3278
3279		r10_bio->master_bio = NULL;
3280		r10_bio->sector = sector_nr;
3281		set_bit(R10BIO_IsSync, &r10_bio->state);
3282		raid10_find_phys(conf, r10_bio);
3283		r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3284
3285		for (i = 0; i < conf->copies; i++) {
3286			int d = r10_bio->devs[i].devnum;
3287			sector_t first_bad, sector;
3288			int bad_sectors;
 
3289
3290			if (r10_bio->devs[i].repl_bio)
3291				r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3292
3293			bio = r10_bio->devs[i].bio;
3294			bio_reset(bio);
3295			clear_bit(BIO_UPTODATE, &bio->bi_flags);
3296			if (conf->mirrors[d].rdev == NULL ||
3297			    test_bit(Faulty, &conf->mirrors[d].rdev->flags))
 
3298				continue;
 
3299			sector = r10_bio->devs[i].addr;
3300			if (is_badblock(conf->mirrors[d].rdev,
3301					sector, max_sync,
3302					&first_bad, &bad_sectors)) {
3303				if (first_bad > sector)
3304					max_sync = first_bad - sector;
3305				else {
3306					bad_sectors -= (sector - first_bad);
3307					if (max_sync > bad_sectors)
3308						max_sync = bad_sectors;
 
3309					continue;
3310				}
3311			}
3312			atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3313			atomic_inc(&r10_bio->remaining);
3314			bio->bi_next = biolist;
3315			biolist = bio;
3316			bio->bi_private = r10_bio;
3317			bio->bi_end_io = end_sync_read;
3318			bio->bi_rw = READ;
3319			bio->bi_iter.bi_sector = sector +
3320				conf->mirrors[d].rdev->data_offset;
3321			bio->bi_bdev = conf->mirrors[d].rdev->bdev;
 
3322			count++;
3323
3324			if (conf->mirrors[d].replacement == NULL ||
3325			    test_bit(Faulty,
3326				     &conf->mirrors[d].replacement->flags))
3327				continue;
 
 
3328
3329			/* Need to set up for writing to the replacement */
3330			bio = r10_bio->devs[i].repl_bio;
3331			bio_reset(bio);
3332			clear_bit(BIO_UPTODATE, &bio->bi_flags);
3333
3334			sector = r10_bio->devs[i].addr;
3335			atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3336			bio->bi_next = biolist;
3337			biolist = bio;
3338			bio->bi_private = r10_bio;
3339			bio->bi_end_io = end_sync_write;
3340			bio->bi_rw = WRITE;
3341			bio->bi_iter.bi_sector = sector +
3342				conf->mirrors[d].replacement->data_offset;
3343			bio->bi_bdev = conf->mirrors[d].replacement->bdev;
 
3344			count++;
 
3345		}
3346
3347		if (count < 2) {
3348			for (i=0; i<conf->copies; i++) {
3349				int d = r10_bio->devs[i].devnum;
3350				if (r10_bio->devs[i].bio->bi_end_io)
3351					rdev_dec_pending(conf->mirrors[d].rdev,
3352							 mddev);
3353				if (r10_bio->devs[i].repl_bio &&
3354				    r10_bio->devs[i].repl_bio->bi_end_io)
3355					rdev_dec_pending(
3356						conf->mirrors[d].replacement,
3357						mddev);
3358			}
3359			put_buf(r10_bio);
3360			biolist = NULL;
3361			goto giveup;
3362		}
3363	}
3364
3365	nr_sectors = 0;
3366	if (sector_nr + max_sync < max_sector)
3367		max_sector = sector_nr + max_sync;
3368	do {
3369		struct page *page;
3370		int len = PAGE_SIZE;
3371		if (sector_nr + (len>>9) > max_sector)
3372			len = (max_sector - sector_nr) << 9;
3373		if (len == 0)
3374			break;
3375		for (bio= biolist ; bio ; bio=bio->bi_next) {
3376			struct bio *bio2;
3377			page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3378			if (bio_add_page(bio, page, len, 0))
3379				continue;
3380
3381			/* stop here */
3382			bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3383			for (bio2 = biolist;
3384			     bio2 && bio2 != bio;
3385			     bio2 = bio2->bi_next) {
3386				/* remove last page from this bio */
3387				bio2->bi_vcnt--;
3388				bio2->bi_iter.bi_size -= len;
3389				bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3390			}
3391			goto bio_full;
3392		}
3393		nr_sectors += len>>9;
3394		sector_nr += len>>9;
3395	} while (biolist->bi_vcnt < RESYNC_PAGES);
3396 bio_full:
3397	r10_bio->sectors = nr_sectors;
3398
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3399	while (biolist) {
3400		bio = biolist;
3401		biolist = biolist->bi_next;
3402
3403		bio->bi_next = NULL;
3404		r10_bio = bio->bi_private;
3405		r10_bio->sectors = nr_sectors;
3406
3407		if (bio->bi_end_io == end_sync_read) {
3408			md_sync_acct(bio->bi_bdev, nr_sectors);
3409			set_bit(BIO_UPTODATE, &bio->bi_flags);
3410			generic_make_request(bio);
3411		}
3412	}
3413
3414	if (sectors_skipped)
3415		/* pretend they weren't skipped, it makes
3416		 * no important difference in this case
3417		 */
3418		md_done_sync(mddev, sectors_skipped, 1);
3419
3420	return sectors_skipped + nr_sectors;
3421 giveup:
3422	/* There is nowhere to write, so all non-sync
3423	 * drives must be failed or in resync, all drives
3424	 * have a bad block, so try the next chunk...
3425	 */
3426	if (sector_nr + max_sync < max_sector)
3427		max_sector = sector_nr + max_sync;
3428
3429	sectors_skipped += (max_sector - sector_nr);
3430	chunks_skipped ++;
3431	sector_nr = max_sector;
3432	goto skipped;
3433}
3434
3435static sector_t
3436raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3437{
3438	sector_t size;
3439	struct r10conf *conf = mddev->private;
3440
3441	if (!raid_disks)
3442		raid_disks = min(conf->geo.raid_disks,
3443				 conf->prev.raid_disks);
3444	if (!sectors)
3445		sectors = conf->dev_sectors;
3446
3447	size = sectors >> conf->geo.chunk_shift;
3448	sector_div(size, conf->geo.far_copies);
3449	size = size * raid_disks;
3450	sector_div(size, conf->geo.near_copies);
3451
3452	return size << conf->geo.chunk_shift;
3453}
3454
3455static void calc_sectors(struct r10conf *conf, sector_t size)
3456{
3457	/* Calculate the number of sectors-per-device that will
3458	 * actually be used, and set conf->dev_sectors and
3459	 * conf->stride
3460	 */
3461
3462	size = size >> conf->geo.chunk_shift;
3463	sector_div(size, conf->geo.far_copies);
3464	size = size * conf->geo.raid_disks;
3465	sector_div(size, conf->geo.near_copies);
3466	/* 'size' is now the number of chunks in the array */
3467	/* calculate "used chunks per device" */
3468	size = size * conf->copies;
3469
3470	/* We need to round up when dividing by raid_disks to
3471	 * get the stride size.
3472	 */
3473	size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3474
3475	conf->dev_sectors = size << conf->geo.chunk_shift;
3476
3477	if (conf->geo.far_offset)
3478		conf->geo.stride = 1 << conf->geo.chunk_shift;
3479	else {
3480		sector_div(size, conf->geo.far_copies);
3481		conf->geo.stride = size << conf->geo.chunk_shift;
3482	}
3483}
3484
3485enum geo_type {geo_new, geo_old, geo_start};
3486static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3487{
3488	int nc, fc, fo;
3489	int layout, chunk, disks;
3490	switch (new) {
3491	case geo_old:
3492		layout = mddev->layout;
3493		chunk = mddev->chunk_sectors;
3494		disks = mddev->raid_disks - mddev->delta_disks;
3495		break;
3496	case geo_new:
3497		layout = mddev->new_layout;
3498		chunk = mddev->new_chunk_sectors;
3499		disks = mddev->raid_disks;
3500		break;
3501	default: /* avoid 'may be unused' warnings */
3502	case geo_start: /* new when starting reshape - raid_disks not
3503			 * updated yet. */
3504		layout = mddev->new_layout;
3505		chunk = mddev->new_chunk_sectors;
3506		disks = mddev->raid_disks + mddev->delta_disks;
3507		break;
3508	}
3509	if (layout >> 18)
3510		return -1;
3511	if (chunk < (PAGE_SIZE >> 9) ||
3512	    !is_power_of_2(chunk))
3513		return -2;
3514	nc = layout & 255;
3515	fc = (layout >> 8) & 255;
3516	fo = layout & (1<<16);
3517	geo->raid_disks = disks;
3518	geo->near_copies = nc;
3519	geo->far_copies = fc;
3520	geo->far_offset = fo;
3521	geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3522	geo->chunk_mask = chunk - 1;
3523	geo->chunk_shift = ffz(~chunk);
3524	return nc*fc;
3525}
3526
3527static struct r10conf *setup_conf(struct mddev *mddev)
3528{
3529	struct r10conf *conf = NULL;
3530	int err = -EINVAL;
3531	struct geom geo;
3532	int copies;
3533
3534	copies = setup_geo(&geo, mddev, geo_new);
3535
3536	if (copies == -2) {
3537		printk(KERN_ERR "md/raid10:%s: chunk size must be "
3538		       "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3539		       mdname(mddev), PAGE_SIZE);
3540		goto out;
3541	}
3542
3543	if (copies < 2 || copies > mddev->raid_disks) {
3544		printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3545		       mdname(mddev), mddev->new_layout);
3546		goto out;
3547	}
3548
3549	err = -ENOMEM;
3550	conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3551	if (!conf)
3552		goto out;
3553
3554	/* FIXME calc properly */
3555	conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3556							    max(0,-mddev->delta_disks)),
3557				GFP_KERNEL);
3558	if (!conf->mirrors)
3559		goto out;
3560
3561	conf->tmppage = alloc_page(GFP_KERNEL);
3562	if (!conf->tmppage)
3563		goto out;
3564
3565	conf->geo = geo;
3566	conf->copies = copies;
3567	conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3568					   r10bio_pool_free, conf);
3569	if (!conf->r10bio_pool)
3570		goto out;
3571
 
 
 
 
3572	calc_sectors(conf, mddev->dev_sectors);
3573	if (mddev->reshape_position == MaxSector) {
3574		conf->prev = conf->geo;
3575		conf->reshape_progress = MaxSector;
3576	} else {
3577		if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3578			err = -EINVAL;
3579			goto out;
3580		}
3581		conf->reshape_progress = mddev->reshape_position;
3582		if (conf->prev.far_offset)
3583			conf->prev.stride = 1 << conf->prev.chunk_shift;
3584		else
3585			/* far_copies must be 1 */
3586			conf->prev.stride = conf->dev_sectors;
3587	}
 
3588	spin_lock_init(&conf->device_lock);
3589	INIT_LIST_HEAD(&conf->retry_list);
 
3590
3591	spin_lock_init(&conf->resync_lock);
3592	init_waitqueue_head(&conf->wait_barrier);
 
3593
3594	conf->thread = md_register_thread(raid10d, mddev, "raid10");
3595	if (!conf->thread)
3596		goto out;
3597
3598	conf->mddev = mddev;
3599	return conf;
3600
3601 out:
3602	if (err == -ENOMEM)
3603		printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3604		       mdname(mddev));
3605	if (conf) {
3606		if (conf->r10bio_pool)
3607			mempool_destroy(conf->r10bio_pool);
3608		kfree(conf->mirrors);
3609		safe_put_page(conf->tmppage);
 
 
3610		kfree(conf);
3611	}
3612	return ERR_PTR(err);
3613}
3614
3615static int run(struct mddev *mddev)
3616{
3617	struct r10conf *conf;
3618	int i, disk_idx, chunk_size;
3619	struct raid10_info *disk;
3620	struct md_rdev *rdev;
3621	sector_t size;
3622	sector_t min_offset_diff = 0;
3623	int first = 1;
3624	bool discard_supported = false;
3625
 
 
 
3626	if (mddev->private == NULL) {
3627		conf = setup_conf(mddev);
3628		if (IS_ERR(conf))
3629			return PTR_ERR(conf);
3630		mddev->private = conf;
3631	}
3632	conf = mddev->private;
3633	if (!conf)
3634		goto out;
3635
 
 
 
 
 
 
 
 
 
 
 
 
3636	mddev->thread = conf->thread;
3637	conf->thread = NULL;
3638
3639	chunk_size = mddev->chunk_sectors << 9;
3640	if (mddev->queue) {
3641		blk_queue_max_discard_sectors(mddev->queue,
3642					      mddev->chunk_sectors);
3643		blk_queue_max_write_same_sectors(mddev->queue, 0);
 
3644		blk_queue_io_min(mddev->queue, chunk_size);
3645		if (conf->geo.raid_disks % conf->geo.near_copies)
3646			blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3647		else
3648			blk_queue_io_opt(mddev->queue, chunk_size *
3649					 (conf->geo.raid_disks / conf->geo.near_copies));
3650	}
3651
3652	rdev_for_each(rdev, mddev) {
3653		long long diff;
3654		struct request_queue *q;
3655
3656		disk_idx = rdev->raid_disk;
3657		if (disk_idx < 0)
3658			continue;
3659		if (disk_idx >= conf->geo.raid_disks &&
3660		    disk_idx >= conf->prev.raid_disks)
3661			continue;
3662		disk = conf->mirrors + disk_idx;
3663
3664		if (test_bit(Replacement, &rdev->flags)) {
3665			if (disk->replacement)
3666				goto out_free_conf;
3667			disk->replacement = rdev;
3668		} else {
3669			if (disk->rdev)
3670				goto out_free_conf;
3671			disk->rdev = rdev;
3672		}
3673		q = bdev_get_queue(rdev->bdev);
3674		if (q->merge_bvec_fn)
3675			mddev->merge_check_needed = 1;
3676		diff = (rdev->new_data_offset - rdev->data_offset);
3677		if (!mddev->reshape_backwards)
3678			diff = -diff;
3679		if (diff < 0)
3680			diff = 0;
3681		if (first || diff < min_offset_diff)
3682			min_offset_diff = diff;
3683
3684		if (mddev->gendisk)
3685			disk_stack_limits(mddev->gendisk, rdev->bdev,
3686					  rdev->data_offset << 9);
3687
3688		disk->head_position = 0;
3689
3690		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3691			discard_supported = true;
 
3692	}
3693
3694	if (mddev->queue) {
3695		if (discard_supported)
3696			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3697						mddev->queue);
3698		else
3699			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3700						  mddev->queue);
3701	}
3702	/* need to check that every block has at least one working mirror */
3703	if (!enough(conf, -1)) {
3704		printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3705		       mdname(mddev));
3706		goto out_free_conf;
3707	}
3708
3709	if (conf->reshape_progress != MaxSector) {
3710		/* must ensure that shape change is supported */
3711		if (conf->geo.far_copies != 1 &&
3712		    conf->geo.far_offset == 0)
3713			goto out_free_conf;
3714		if (conf->prev.far_copies != 1 &&
3715		    conf->prev.far_offset == 0)
3716			goto out_free_conf;
3717	}
3718
3719	mddev->degraded = 0;
3720	for (i = 0;
3721	     i < conf->geo.raid_disks
3722		     || i < conf->prev.raid_disks;
3723	     i++) {
3724
3725		disk = conf->mirrors + i;
3726
3727		if (!disk->rdev && disk->replacement) {
3728			/* The replacement is all we have - use it */
3729			disk->rdev = disk->replacement;
3730			disk->replacement = NULL;
3731			clear_bit(Replacement, &disk->rdev->flags);
3732		}
3733
3734		if (!disk->rdev ||
3735		    !test_bit(In_sync, &disk->rdev->flags)) {
3736			disk->head_position = 0;
3737			mddev->degraded++;
3738			if (disk->rdev &&
3739			    disk->rdev->saved_raid_disk < 0)
3740				conf->fullsync = 1;
3741		}
3742		disk->recovery_disabled = mddev->recovery_disabled - 1;
3743	}
3744
3745	if (mddev->recovery_cp != MaxSector)
3746		printk(KERN_NOTICE "md/raid10:%s: not clean"
3747		       " -- starting background reconstruction\n",
3748		       mdname(mddev));
3749	printk(KERN_INFO
3750		"md/raid10:%s: active with %d out of %d devices\n",
3751		mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3752		conf->geo.raid_disks);
3753	/*
3754	 * Ok, everything is just fine now
3755	 */
3756	mddev->dev_sectors = conf->dev_sectors;
3757	size = raid10_size(mddev, 0, 0);
3758	md_set_array_sectors(mddev, size);
3759	mddev->resync_max_sectors = size;
 
3760
3761	if (mddev->queue) {
3762		int stripe = conf->geo.raid_disks *
3763			((mddev->chunk_sectors << 9) / PAGE_SIZE);
3764		mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3765		mddev->queue->backing_dev_info.congested_data = mddev;
3766
3767		/* Calculate max read-ahead size.
3768		 * We need to readahead at least twice a whole stripe....
3769		 * maybe...
3770		 */
3771		stripe /= conf->geo.near_copies;
3772		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3773			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3774		blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3775	}
3776
3777
3778	if (md_integrity_register(mddev))
3779		goto out_free_conf;
3780
3781	if (conf->reshape_progress != MaxSector) {
3782		unsigned long before_length, after_length;
3783
3784		before_length = ((1 << conf->prev.chunk_shift) *
3785				 conf->prev.far_copies);
3786		after_length = ((1 << conf->geo.chunk_shift) *
3787				conf->geo.far_copies);
3788
3789		if (max(before_length, after_length) > min_offset_diff) {
3790			/* This cannot work */
3791			printk("md/raid10: offset difference not enough to continue reshape\n");
3792			goto out_free_conf;
3793		}
3794		conf->offset_diff = min_offset_diff;
3795
3796		conf->reshape_safe = conf->reshape_progress;
3797		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3798		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3799		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3800		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3801		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3802							"reshape");
3803	}
3804
3805	return 0;
3806
3807out_free_conf:
3808	md_unregister_thread(&mddev->thread);
3809	if (conf->r10bio_pool)
3810		mempool_destroy(conf->r10bio_pool);
3811	safe_put_page(conf->tmppage);
3812	kfree(conf->mirrors);
3813	kfree(conf);
3814	mddev->private = NULL;
3815out:
3816	return -EIO;
3817}
3818
3819static int stop(struct mddev *mddev)
3820{
3821	struct r10conf *conf = mddev->private;
3822
3823	raise_barrier(conf, 0);
3824	lower_barrier(conf);
3825
3826	md_unregister_thread(&mddev->thread);
3827	if (mddev->queue)
3828		/* the unplug fn references 'conf'*/
3829		blk_sync_queue(mddev->queue);
3830
3831	if (conf->r10bio_pool)
3832		mempool_destroy(conf->r10bio_pool);
3833	safe_put_page(conf->tmppage);
3834	kfree(conf->mirrors);
 
 
 
 
3835	kfree(conf);
3836	mddev->private = NULL;
3837	return 0;
3838}
3839
3840static void raid10_quiesce(struct mddev *mddev, int state)
3841{
3842	struct r10conf *conf = mddev->private;
3843
3844	switch(state) {
3845	case 1:
3846		raise_barrier(conf, 0);
3847		break;
3848	case 0:
3849		lower_barrier(conf);
3850		break;
3851	}
3852}
3853
3854static int raid10_resize(struct mddev *mddev, sector_t sectors)
3855{
3856	/* Resize of 'far' arrays is not supported.
3857	 * For 'near' and 'offset' arrays we can set the
3858	 * number of sectors used to be an appropriate multiple
3859	 * of the chunk size.
3860	 * For 'offset', this is far_copies*chunksize.
3861	 * For 'near' the multiplier is the LCM of
3862	 * near_copies and raid_disks.
3863	 * So if far_copies > 1 && !far_offset, fail.
3864	 * Else find LCM(raid_disks, near_copy)*far_copies and
3865	 * multiply by chunk_size.  Then round to this number.
3866	 * This is mostly done by raid10_size()
3867	 */
3868	struct r10conf *conf = mddev->private;
3869	sector_t oldsize, size;
3870
3871	if (mddev->reshape_position != MaxSector)
3872		return -EBUSY;
3873
3874	if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3875		return -EINVAL;
3876
3877	oldsize = raid10_size(mddev, 0, 0);
3878	size = raid10_size(mddev, sectors, 0);
3879	if (mddev->external_size &&
3880	    mddev->array_sectors > size)
3881		return -EINVAL;
3882	if (mddev->bitmap) {
3883		int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3884		if (ret)
3885			return ret;
3886	}
3887	md_set_array_sectors(mddev, size);
3888	set_capacity(mddev->gendisk, mddev->array_sectors);
3889	revalidate_disk(mddev->gendisk);
3890	if (sectors > mddev->dev_sectors &&
3891	    mddev->recovery_cp > oldsize) {
3892		mddev->recovery_cp = oldsize;
3893		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3894	}
3895	calc_sectors(conf, sectors);
3896	mddev->dev_sectors = conf->dev_sectors;
3897	mddev->resync_max_sectors = size;
3898	return 0;
3899}
3900
3901static void *raid10_takeover_raid0(struct mddev *mddev)
3902{
3903	struct md_rdev *rdev;
3904	struct r10conf *conf;
3905
3906	if (mddev->degraded > 0) {
3907		printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3908		       mdname(mddev));
3909		return ERR_PTR(-EINVAL);
3910	}
 
3911
3912	/* Set new parameters */
3913	mddev->new_level = 10;
3914	/* new layout: far_copies = 1, near_copies = 2 */
3915	mddev->new_layout = (1<<8) + 2;
3916	mddev->new_chunk_sectors = mddev->chunk_sectors;
3917	mddev->delta_disks = mddev->raid_disks;
3918	mddev->raid_disks *= 2;
3919	/* make sure it will be not marked as dirty */
3920	mddev->recovery_cp = MaxSector;
 
3921
3922	conf = setup_conf(mddev);
3923	if (!IS_ERR(conf)) {
3924		rdev_for_each(rdev, mddev)
3925			if (rdev->raid_disk >= 0)
3926				rdev->new_raid_disk = rdev->raid_disk * 2;
 
 
3927		conf->barrier = 1;
3928	}
3929
3930	return conf;
3931}
3932
3933static void *raid10_takeover(struct mddev *mddev)
3934{
3935	struct r0conf *raid0_conf;
3936
3937	/* raid10 can take over:
3938	 *  raid0 - providing it has only two drives
3939	 */
3940	if (mddev->level == 0) {
3941		/* for raid0 takeover only one zone is supported */
3942		raid0_conf = mddev->private;
3943		if (raid0_conf->nr_strip_zones > 1) {
3944			printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3945			       " with more than one zone.\n",
3946			       mdname(mddev));
3947			return ERR_PTR(-EINVAL);
3948		}
3949		return raid10_takeover_raid0(mddev);
 
 
3950	}
3951	return ERR_PTR(-EINVAL);
3952}
3953
3954static int raid10_check_reshape(struct mddev *mddev)
3955{
3956	/* Called when there is a request to change
3957	 * - layout (to ->new_layout)
3958	 * - chunk size (to ->new_chunk_sectors)
3959	 * - raid_disks (by delta_disks)
3960	 * or when trying to restart a reshape that was ongoing.
3961	 *
3962	 * We need to validate the request and possibly allocate
3963	 * space if that might be an issue later.
3964	 *
3965	 * Currently we reject any reshape of a 'far' mode array,
3966	 * allow chunk size to change if new is generally acceptable,
3967	 * allow raid_disks to increase, and allow
3968	 * a switch between 'near' mode and 'offset' mode.
3969	 */
3970	struct r10conf *conf = mddev->private;
3971	struct geom geo;
3972
3973	if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3974		return -EINVAL;
3975
3976	if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3977		/* mustn't change number of copies */
3978		return -EINVAL;
3979	if (geo.far_copies > 1 && !geo.far_offset)
3980		/* Cannot switch to 'far' mode */
3981		return -EINVAL;
3982
3983	if (mddev->array_sectors & geo.chunk_mask)
3984			/* not factor of array size */
3985			return -EINVAL;
3986
3987	if (!enough(conf, -1))
3988		return -EINVAL;
3989
3990	kfree(conf->mirrors_new);
3991	conf->mirrors_new = NULL;
3992	if (mddev->delta_disks > 0) {
3993		/* allocate new 'mirrors' list */
3994		conf->mirrors_new = kzalloc(
3995			sizeof(struct raid10_info)
3996			*(mddev->raid_disks +
3997			  mddev->delta_disks),
3998			GFP_KERNEL);
3999		if (!conf->mirrors_new)
4000			return -ENOMEM;
4001	}
4002	return 0;
4003}
4004
4005/*
4006 * Need to check if array has failed when deciding whether to:
4007 *  - start an array
4008 *  - remove non-faulty devices
4009 *  - add a spare
4010 *  - allow a reshape
4011 * This determination is simple when no reshape is happening.
4012 * However if there is a reshape, we need to carefully check
4013 * both the before and after sections.
4014 * This is because some failed devices may only affect one
4015 * of the two sections, and some non-in_sync devices may
4016 * be insync in the section most affected by failed devices.
4017 */
4018static int calc_degraded(struct r10conf *conf)
4019{
4020	int degraded, degraded2;
4021	int i;
4022
4023	rcu_read_lock();
4024	degraded = 0;
4025	/* 'prev' section first */
4026	for (i = 0; i < conf->prev.raid_disks; i++) {
4027		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4028		if (!rdev || test_bit(Faulty, &rdev->flags))
4029			degraded++;
4030		else if (!test_bit(In_sync, &rdev->flags))
4031			/* When we can reduce the number of devices in
4032			 * an array, this might not contribute to
4033			 * 'degraded'.  It does now.
4034			 */
4035			degraded++;
4036	}
4037	rcu_read_unlock();
4038	if (conf->geo.raid_disks == conf->prev.raid_disks)
4039		return degraded;
4040	rcu_read_lock();
4041	degraded2 = 0;
4042	for (i = 0; i < conf->geo.raid_disks; i++) {
4043		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4044		if (!rdev || test_bit(Faulty, &rdev->flags))
4045			degraded2++;
4046		else if (!test_bit(In_sync, &rdev->flags)) {
4047			/* If reshape is increasing the number of devices,
4048			 * this section has already been recovered, so
4049			 * it doesn't contribute to degraded.
4050			 * else it does.
4051			 */
4052			if (conf->geo.raid_disks <= conf->prev.raid_disks)
4053				degraded2++;
4054		}
4055	}
4056	rcu_read_unlock();
4057	if (degraded2 > degraded)
4058		return degraded2;
4059	return degraded;
4060}
4061
4062static int raid10_start_reshape(struct mddev *mddev)
4063{
4064	/* A 'reshape' has been requested. This commits
4065	 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4066	 * This also checks if there are enough spares and adds them
4067	 * to the array.
4068	 * We currently require enough spares to make the final
4069	 * array non-degraded.  We also require that the difference
4070	 * between old and new data_offset - on each device - is
4071	 * enough that we never risk over-writing.
4072	 */
4073
4074	unsigned long before_length, after_length;
4075	sector_t min_offset_diff = 0;
4076	int first = 1;
4077	struct geom new;
4078	struct r10conf *conf = mddev->private;
4079	struct md_rdev *rdev;
4080	int spares = 0;
4081	int ret;
4082
4083	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4084		return -EBUSY;
4085
4086	if (setup_geo(&new, mddev, geo_start) != conf->copies)
4087		return -EINVAL;
4088
4089	before_length = ((1 << conf->prev.chunk_shift) *
4090			 conf->prev.far_copies);
4091	after_length = ((1 << conf->geo.chunk_shift) *
4092			conf->geo.far_copies);
4093
4094	rdev_for_each(rdev, mddev) {
4095		if (!test_bit(In_sync, &rdev->flags)
4096		    && !test_bit(Faulty, &rdev->flags))
4097			spares++;
4098		if (rdev->raid_disk >= 0) {
4099			long long diff = (rdev->new_data_offset
4100					  - rdev->data_offset);
4101			if (!mddev->reshape_backwards)
4102				diff = -diff;
4103			if (diff < 0)
4104				diff = 0;
4105			if (first || diff < min_offset_diff)
4106				min_offset_diff = diff;
 
4107		}
4108	}
4109
4110	if (max(before_length, after_length) > min_offset_diff)
4111		return -EINVAL;
4112
4113	if (spares < mddev->delta_disks)
4114		return -EINVAL;
4115
4116	conf->offset_diff = min_offset_diff;
4117	spin_lock_irq(&conf->device_lock);
4118	if (conf->mirrors_new) {
4119		memcpy(conf->mirrors_new, conf->mirrors,
4120		       sizeof(struct raid10_info)*conf->prev.raid_disks);
4121		smp_mb();
4122		kfree(conf->mirrors_old); /* FIXME and elsewhere */
4123		conf->mirrors_old = conf->mirrors;
4124		conf->mirrors = conf->mirrors_new;
4125		conf->mirrors_new = NULL;
4126	}
4127	setup_geo(&conf->geo, mddev, geo_start);
4128	smp_mb();
4129	if (mddev->reshape_backwards) {
4130		sector_t size = raid10_size(mddev, 0, 0);
4131		if (size < mddev->array_sectors) {
4132			spin_unlock_irq(&conf->device_lock);
4133			printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4134			       mdname(mddev));
4135			return -EINVAL;
4136		}
4137		mddev->resync_max_sectors = size;
4138		conf->reshape_progress = size;
4139	} else
4140		conf->reshape_progress = 0;
 
4141	spin_unlock_irq(&conf->device_lock);
4142
4143	if (mddev->delta_disks && mddev->bitmap) {
4144		ret = bitmap_resize(mddev->bitmap,
4145				    raid10_size(mddev, 0,
4146						conf->geo.raid_disks),
4147				    0, 0);
4148		if (ret)
4149			goto abort;
4150	}
4151	if (mddev->delta_disks > 0) {
4152		rdev_for_each(rdev, mddev)
4153			if (rdev->raid_disk < 0 &&
4154			    !test_bit(Faulty, &rdev->flags)) {
4155				if (raid10_add_disk(mddev, rdev) == 0) {
4156					if (rdev->raid_disk >=
4157					    conf->prev.raid_disks)
4158						set_bit(In_sync, &rdev->flags);
4159					else
4160						rdev->recovery_offset = 0;
4161
4162					if (sysfs_link_rdev(mddev, rdev))
4163						/* Failure here  is OK */;
4164				}
4165			} else if (rdev->raid_disk >= conf->prev.raid_disks
4166				   && !test_bit(Faulty, &rdev->flags)) {
4167				/* This is a spare that was manually added */
4168				set_bit(In_sync, &rdev->flags);
4169			}
4170	}
4171	/* When a reshape changes the number of devices,
4172	 * ->degraded is measured against the larger of the
4173	 * pre and  post numbers.
4174	 */
4175	spin_lock_irq(&conf->device_lock);
4176	mddev->degraded = calc_degraded(conf);
4177	spin_unlock_irq(&conf->device_lock);
4178	mddev->raid_disks = conf->geo.raid_disks;
4179	mddev->reshape_position = conf->reshape_progress;
4180	set_bit(MD_CHANGE_DEVS, &mddev->flags);
4181
4182	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4183	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
 
4184	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4185	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4186
4187	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4188						"reshape");
4189	if (!mddev->sync_thread) {
4190		ret = -EAGAIN;
4191		goto abort;
4192	}
4193	conf->reshape_checkpoint = jiffies;
4194	md_wakeup_thread(mddev->sync_thread);
4195	md_new_event(mddev);
4196	return 0;
4197
4198abort:
4199	mddev->recovery = 0;
4200	spin_lock_irq(&conf->device_lock);
4201	conf->geo = conf->prev;
4202	mddev->raid_disks = conf->geo.raid_disks;
4203	rdev_for_each(rdev, mddev)
4204		rdev->new_data_offset = rdev->data_offset;
4205	smp_wmb();
4206	conf->reshape_progress = MaxSector;
 
4207	mddev->reshape_position = MaxSector;
4208	spin_unlock_irq(&conf->device_lock);
4209	return ret;
4210}
4211
4212/* Calculate the last device-address that could contain
4213 * any block from the chunk that includes the array-address 's'
4214 * and report the next address.
4215 * i.e. the address returned will be chunk-aligned and after
4216 * any data that is in the chunk containing 's'.
4217 */
4218static sector_t last_dev_address(sector_t s, struct geom *geo)
4219{
4220	s = (s | geo->chunk_mask) + 1;
4221	s >>= geo->chunk_shift;
4222	s *= geo->near_copies;
4223	s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4224	s *= geo->far_copies;
4225	s <<= geo->chunk_shift;
4226	return s;
4227}
4228
4229/* Calculate the first device-address that could contain
4230 * any block from the chunk that includes the array-address 's'.
4231 * This too will be the start of a chunk
4232 */
4233static sector_t first_dev_address(sector_t s, struct geom *geo)
4234{
4235	s >>= geo->chunk_shift;
4236	s *= geo->near_copies;
4237	sector_div(s, geo->raid_disks);
4238	s *= geo->far_copies;
4239	s <<= geo->chunk_shift;
4240	return s;
4241}
4242
4243static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4244				int *skipped)
4245{
4246	/* We simply copy at most one chunk (smallest of old and new)
4247	 * at a time, possibly less if that exceeds RESYNC_PAGES,
4248	 * or we hit a bad block or something.
4249	 * This might mean we pause for normal IO in the middle of
4250	 * a chunk, but that is not a problem was mddev->reshape_position
4251	 * can record any location.
4252	 *
4253	 * If we will want to write to a location that isn't
4254	 * yet recorded as 'safe' (i.e. in metadata on disk) then
4255	 * we need to flush all reshape requests and update the metadata.
4256	 *
4257	 * When reshaping forwards (e.g. to more devices), we interpret
4258	 * 'safe' as the earliest block which might not have been copied
4259	 * down yet.  We divide this by previous stripe size and multiply
4260	 * by previous stripe length to get lowest device offset that we
4261	 * cannot write to yet.
4262	 * We interpret 'sector_nr' as an address that we want to write to.
4263	 * From this we use last_device_address() to find where we might
4264	 * write to, and first_device_address on the  'safe' position.
4265	 * If this 'next' write position is after the 'safe' position,
4266	 * we must update the metadata to increase the 'safe' position.
4267	 *
4268	 * When reshaping backwards, we round in the opposite direction
4269	 * and perform the reverse test:  next write position must not be
4270	 * less than current safe position.
4271	 *
4272	 * In all this the minimum difference in data offsets
4273	 * (conf->offset_diff - always positive) allows a bit of slack,
4274	 * so next can be after 'safe', but not by more than offset_disk
4275	 *
4276	 * We need to prepare all the bios here before we start any IO
4277	 * to ensure the size we choose is acceptable to all devices.
4278	 * The means one for each copy for write-out and an extra one for
4279	 * read-in.
4280	 * We store the read-in bio in ->master_bio and the others in
4281	 * ->devs[x].bio and ->devs[x].repl_bio.
4282	 */
4283	struct r10conf *conf = mddev->private;
4284	struct r10bio *r10_bio;
4285	sector_t next, safe, last;
4286	int max_sectors;
4287	int nr_sectors;
4288	int s;
4289	struct md_rdev *rdev;
4290	int need_flush = 0;
4291	struct bio *blist;
4292	struct bio *bio, *read_bio;
4293	int sectors_done = 0;
 
4294
4295	if (sector_nr == 0) {
4296		/* If restarting in the middle, skip the initial sectors */
4297		if (mddev->reshape_backwards &&
4298		    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4299			sector_nr = (raid10_size(mddev, 0, 0)
4300				     - conf->reshape_progress);
4301		} else if (!mddev->reshape_backwards &&
4302			   conf->reshape_progress > 0)
4303			sector_nr = conf->reshape_progress;
4304		if (sector_nr) {
4305			mddev->curr_resync_completed = sector_nr;
4306			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4307			*skipped = 1;
4308			return sector_nr;
4309		}
4310	}
4311
4312	/* We don't use sector_nr to track where we are up to
4313	 * as that doesn't work well for ->reshape_backwards.
4314	 * So just use ->reshape_progress.
4315	 */
4316	if (mddev->reshape_backwards) {
4317		/* 'next' is the earliest device address that we might
4318		 * write to for this chunk in the new layout
4319		 */
4320		next = first_dev_address(conf->reshape_progress - 1,
4321					 &conf->geo);
4322
4323		/* 'safe' is the last device address that we might read from
4324		 * in the old layout after a restart
4325		 */
4326		safe = last_dev_address(conf->reshape_safe - 1,
4327					&conf->prev);
4328
4329		if (next + conf->offset_diff < safe)
4330			need_flush = 1;
4331
4332		last = conf->reshape_progress - 1;
4333		sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4334					       & conf->prev.chunk_mask);
4335		if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4336			sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4337	} else {
4338		/* 'next' is after the last device address that we
4339		 * might write to for this chunk in the new layout
4340		 */
4341		next = last_dev_address(conf->reshape_progress, &conf->geo);
4342
4343		/* 'safe' is the earliest device address that we might
4344		 * read from in the old layout after a restart
4345		 */
4346		safe = first_dev_address(conf->reshape_safe, &conf->prev);
4347
4348		/* Need to update metadata if 'next' might be beyond 'safe'
4349		 * as that would possibly corrupt data
4350		 */
4351		if (next > safe + conf->offset_diff)
4352			need_flush = 1;
4353
4354		sector_nr = conf->reshape_progress;
4355		last  = sector_nr | (conf->geo.chunk_mask
4356				     & conf->prev.chunk_mask);
4357
4358		if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4359			last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4360	}
4361
4362	if (need_flush ||
4363	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4364		/* Need to update reshape_position in metadata */
4365		wait_barrier(conf);
4366		mddev->reshape_position = conf->reshape_progress;
4367		if (mddev->reshape_backwards)
4368			mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4369				- conf->reshape_progress;
4370		else
4371			mddev->curr_resync_completed = conf->reshape_progress;
4372		conf->reshape_checkpoint = jiffies;
4373		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4374		md_wakeup_thread(mddev->thread);
4375		wait_event(mddev->sb_wait, mddev->flags == 0 ||
4376			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4377		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4378			allow_barrier(conf);
4379			return sectors_done;
4380		}
4381		conf->reshape_safe = mddev->reshape_position;
4382		allow_barrier(conf);
4383	}
4384
4385read_more:
4386	/* Now schedule reads for blocks from sector_nr to last */
4387	r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
 
4388	raise_barrier(conf, sectors_done != 0);
4389	atomic_set(&r10_bio->remaining, 0);
4390	r10_bio->mddev = mddev;
4391	r10_bio->sector = sector_nr;
4392	set_bit(R10BIO_IsReshape, &r10_bio->state);
4393	r10_bio->sectors = last - sector_nr + 1;
4394	rdev = read_balance(conf, r10_bio, &max_sectors);
4395	BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4396
4397	if (!rdev) {
4398		/* Cannot read from here, so need to record bad blocks
4399		 * on all the target devices.
4400		 */
4401		// FIXME
 
4402		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4403		return sectors_done;
4404	}
4405
4406	read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4407
4408	read_bio->bi_bdev = rdev->bdev;
4409	read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4410			       + rdev->data_offset);
4411	read_bio->bi_private = r10_bio;
4412	read_bio->bi_end_io = end_sync_read;
4413	read_bio->bi_rw = READ;
4414	read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4415	read_bio->bi_flags |= 1 << BIO_UPTODATE;
4416	read_bio->bi_vcnt = 0;
4417	read_bio->bi_iter.bi_size = 0;
4418	r10_bio->master_bio = read_bio;
4419	r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4420
4421	/* Now find the locations in the new layout */
4422	__raid10_find_phys(&conf->geo, r10_bio);
4423
4424	blist = read_bio;
4425	read_bio->bi_next = NULL;
4426
 
4427	for (s = 0; s < conf->copies*2; s++) {
4428		struct bio *b;
4429		int d = r10_bio->devs[s/2].devnum;
4430		struct md_rdev *rdev2;
4431		if (s&1) {
4432			rdev2 = conf->mirrors[d].replacement;
4433			b = r10_bio->devs[s/2].repl_bio;
4434		} else {
4435			rdev2 = conf->mirrors[d].rdev;
4436			b = r10_bio->devs[s/2].bio;
4437		}
4438		if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4439			continue;
4440
4441		bio_reset(b);
4442		b->bi_bdev = rdev2->bdev;
4443		b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4444			rdev2->new_data_offset;
4445		b->bi_private = r10_bio;
4446		b->bi_end_io = end_reshape_write;
4447		b->bi_rw = WRITE;
4448		b->bi_next = blist;
4449		blist = b;
4450	}
4451
4452	/* Now add as many pages as possible to all of these bios. */
4453
4454	nr_sectors = 0;
 
4455	for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4456		struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4457		int len = (max_sectors - s) << 9;
4458		if (len > PAGE_SIZE)
4459			len = PAGE_SIZE;
4460		for (bio = blist; bio ; bio = bio->bi_next) {
4461			struct bio *bio2;
4462			if (bio_add_page(bio, page, len, 0))
4463				continue;
4464
4465			/* Didn't fit, must stop */
4466			for (bio2 = blist;
4467			     bio2 && bio2 != bio;
4468			     bio2 = bio2->bi_next) {
4469				/* Remove last page from this bio */
4470				bio2->bi_vcnt--;
4471				bio2->bi_iter.bi_size -= len;
4472				bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4473			}
4474			goto bio_full;
4475		}
4476		sector_nr += len >> 9;
4477		nr_sectors += len >> 9;
4478	}
4479bio_full:
4480	r10_bio->sectors = nr_sectors;
4481
4482	/* Now submit the read */
4483	md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4484	atomic_inc(&r10_bio->remaining);
4485	read_bio->bi_next = NULL;
4486	generic_make_request(read_bio);
4487	sector_nr += nr_sectors;
4488	sectors_done += nr_sectors;
4489	if (sector_nr <= last)
4490		goto read_more;
4491
4492	/* Now that we have done the whole section we can
4493	 * update reshape_progress
4494	 */
4495	if (mddev->reshape_backwards)
4496		conf->reshape_progress -= sectors_done;
4497	else
4498		conf->reshape_progress += sectors_done;
4499
4500	return sectors_done;
4501}
4502
4503static void end_reshape_request(struct r10bio *r10_bio);
4504static int handle_reshape_read_error(struct mddev *mddev,
4505				     struct r10bio *r10_bio);
4506static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4507{
4508	/* Reshape read completed.  Hopefully we have a block
4509	 * to write out.
4510	 * If we got a read error then we do sync 1-page reads from
4511	 * elsewhere until we find the data - or give up.
4512	 */
4513	struct r10conf *conf = mddev->private;
4514	int s;
4515
4516	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4517		if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4518			/* Reshape has been aborted */
4519			md_done_sync(mddev, r10_bio->sectors, 0);
4520			return;
4521		}
4522
4523	/* We definitely have the data in the pages, schedule the
4524	 * writes.
4525	 */
4526	atomic_set(&r10_bio->remaining, 1);
4527	for (s = 0; s < conf->copies*2; s++) {
4528		struct bio *b;
4529		int d = r10_bio->devs[s/2].devnum;
4530		struct md_rdev *rdev;
 
4531		if (s&1) {
4532			rdev = conf->mirrors[d].replacement;
4533			b = r10_bio->devs[s/2].repl_bio;
4534		} else {
4535			rdev = conf->mirrors[d].rdev;
4536			b = r10_bio->devs[s/2].bio;
4537		}
4538		if (!rdev || test_bit(Faulty, &rdev->flags))
 
4539			continue;
 
4540		atomic_inc(&rdev->nr_pending);
4541		md_sync_acct(b->bi_bdev, r10_bio->sectors);
 
4542		atomic_inc(&r10_bio->remaining);
4543		b->bi_next = NULL;
4544		generic_make_request(b);
4545	}
4546	end_reshape_request(r10_bio);
4547}
4548
4549static void end_reshape(struct r10conf *conf)
4550{
4551	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4552		return;
4553
4554	spin_lock_irq(&conf->device_lock);
4555	conf->prev = conf->geo;
4556	md_finish_reshape(conf->mddev);
4557	smp_wmb();
4558	conf->reshape_progress = MaxSector;
 
4559	spin_unlock_irq(&conf->device_lock);
4560
4561	/* read-ahead size must cover two whole stripes, which is
4562	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4563	 */
4564	if (conf->mddev->queue) {
4565		int stripe = conf->geo.raid_disks *
4566			((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4567		stripe /= conf->geo.near_copies;
4568		if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4569			conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4570	}
4571	conf->fullsync = 0;
4572}
4573
4574
4575static int handle_reshape_read_error(struct mddev *mddev,
4576				     struct r10bio *r10_bio)
4577{
4578	/* Use sync reads to get the blocks from somewhere else */
4579	int sectors = r10_bio->sectors;
4580	struct r10conf *conf = mddev->private;
4581	struct {
4582		struct r10bio r10_bio;
4583		struct r10dev devs[conf->copies];
4584	} on_stack;
4585	struct r10bio *r10b = &on_stack.r10_bio;
4586	int slot = 0;
4587	int idx = 0;
4588	struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
 
 
 
 
 
 
 
 
 
 
4589
4590	r10b->sector = r10_bio->sector;
4591	__raid10_find_phys(&conf->prev, r10b);
4592
4593	while (sectors) {
4594		int s = sectors;
4595		int success = 0;
4596		int first_slot = slot;
4597
4598		if (s > (PAGE_SIZE >> 9))
4599			s = PAGE_SIZE >> 9;
4600
 
4601		while (!success) {
4602			int d = r10b->devs[slot].devnum;
4603			struct md_rdev *rdev = conf->mirrors[d].rdev;
4604			sector_t addr;
4605			if (rdev == NULL ||
4606			    test_bit(Faulty, &rdev->flags) ||
4607			    !test_bit(In_sync, &rdev->flags))
4608				goto failed;
4609
4610			addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
 
 
4611			success = sync_page_io(rdev,
4612					       addr,
4613					       s << 9,
4614					       bvec[idx].bv_page,
4615					       READ, false);
 
 
4616			if (success)
4617				break;
4618		failed:
4619			slot++;
4620			if (slot >= conf->copies)
4621				slot = 0;
4622			if (slot == first_slot)
4623				break;
4624		}
 
4625		if (!success) {
4626			/* couldn't read this block, must give up */
4627			set_bit(MD_RECOVERY_INTR,
4628				&mddev->recovery);
 
4629			return -EIO;
4630		}
4631		sectors -= s;
4632		idx++;
4633	}
 
4634	return 0;
4635}
4636
4637static void end_reshape_write(struct bio *bio, int error)
4638{
4639	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4640	struct r10bio *r10_bio = bio->bi_private;
4641	struct mddev *mddev = r10_bio->mddev;
4642	struct r10conf *conf = mddev->private;
4643	int d;
4644	int slot;
4645	int repl;
4646	struct md_rdev *rdev = NULL;
4647
4648	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4649	if (repl)
4650		rdev = conf->mirrors[d].replacement;
4651	if (!rdev) {
4652		smp_mb();
4653		rdev = conf->mirrors[d].rdev;
4654	}
4655
4656	if (!uptodate) {
4657		/* FIXME should record badblock */
4658		md_error(mddev, rdev);
4659	}
4660
4661	rdev_dec_pending(rdev, mddev);
4662	end_reshape_request(r10_bio);
4663}
4664
4665static void end_reshape_request(struct r10bio *r10_bio)
4666{
4667	if (!atomic_dec_and_test(&r10_bio->remaining))
4668		return;
4669	md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4670	bio_put(r10_bio->master_bio);
4671	put_buf(r10_bio);
4672}
4673
4674static void raid10_finish_reshape(struct mddev *mddev)
4675{
4676	struct r10conf *conf = mddev->private;
4677
4678	if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4679		return;
4680
4681	if (mddev->delta_disks > 0) {
4682		sector_t size = raid10_size(mddev, 0, 0);
4683		md_set_array_sectors(mddev, size);
4684		if (mddev->recovery_cp > mddev->resync_max_sectors) {
4685			mddev->recovery_cp = mddev->resync_max_sectors;
4686			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4687		}
4688		mddev->resync_max_sectors = size;
4689		set_capacity(mddev->gendisk, mddev->array_sectors);
4690		revalidate_disk(mddev->gendisk);
4691	} else {
4692		int d;
 
4693		for (d = conf->geo.raid_disks ;
4694		     d < conf->geo.raid_disks - mddev->delta_disks;
4695		     d++) {
4696			struct md_rdev *rdev = conf->mirrors[d].rdev;
4697			if (rdev)
4698				clear_bit(In_sync, &rdev->flags);
4699			rdev = conf->mirrors[d].replacement;
4700			if (rdev)
4701				clear_bit(In_sync, &rdev->flags);
4702		}
 
4703	}
4704	mddev->layout = mddev->new_layout;
4705	mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4706	mddev->reshape_position = MaxSector;
4707	mddev->delta_disks = 0;
4708	mddev->reshape_backwards = 0;
4709}
4710
4711static struct md_personality raid10_personality =
4712{
4713	.name		= "raid10",
4714	.level		= 10,
4715	.owner		= THIS_MODULE,
4716	.make_request	= make_request,
4717	.run		= run,
4718	.stop		= stop,
4719	.status		= status,
4720	.error_handler	= error,
4721	.hot_add_disk	= raid10_add_disk,
4722	.hot_remove_disk= raid10_remove_disk,
4723	.spare_active	= raid10_spare_active,
4724	.sync_request	= sync_request,
4725	.quiesce	= raid10_quiesce,
4726	.size		= raid10_size,
4727	.resize		= raid10_resize,
4728	.takeover	= raid10_takeover,
4729	.check_reshape	= raid10_check_reshape,
4730	.start_reshape	= raid10_start_reshape,
4731	.finish_reshape	= raid10_finish_reshape,
 
4732};
4733
4734static int __init raid_init(void)
4735{
4736	return register_md_personality(&raid10_personality);
4737}
4738
4739static void raid_exit(void)
4740{
4741	unregister_md_personality(&raid10_personality);
4742}
4743
4744module_init(raid_init);
4745module_exit(raid_exit);
4746MODULE_LICENSE("GPL");
4747MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4748MODULE_ALIAS("md-personality-9"); /* RAID10 */
4749MODULE_ALIAS("md-raid10");
4750MODULE_ALIAS("md-level-10");
4751
4752module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
v4.17
   1/*
   2 * raid10.c : Multiple Devices driver for Linux
   3 *
   4 * Copyright (C) 2000-2004 Neil Brown
   5 *
   6 * RAID-10 support for md.
   7 *
   8 * Base on code in raid1.c.  See raid1.c for further copyright information.
   9 *
  10 *
  11 * This program is free software; you can redistribute it and/or modify
  12 * it under the terms of the GNU General Public License as published by
  13 * the Free Software Foundation; either version 2, or (at your option)
  14 * any later version.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * (for example /usr/src/linux/COPYING); if not, write to the Free
  18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19 */
  20
  21#include <linux/slab.h>
  22#include <linux/delay.h>
  23#include <linux/blkdev.h>
  24#include <linux/module.h>
  25#include <linux/seq_file.h>
  26#include <linux/ratelimit.h>
  27#include <linux/kthread.h>
  28#include <trace/events/block.h>
  29#include "md.h"
  30#include "raid10.h"
  31#include "raid0.h"
  32#include "md-bitmap.h"
  33
  34/*
  35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
  36 * The layout of data is defined by
  37 *    chunk_size
  38 *    raid_disks
  39 *    near_copies (stored in low byte of layout)
  40 *    far_copies (stored in second byte of layout)
  41 *    far_offset (stored in bit 16 of layout )
  42 *    use_far_sets (stored in bit 17 of layout )
  43 *    use_far_sets_bugfixed (stored in bit 18 of layout )
  44 *
  45 * The data to be stored is divided into chunks using chunksize.  Each device
  46 * is divided into far_copies sections.   In each section, chunks are laid out
  47 * in a style similar to raid0, but near_copies copies of each chunk is stored
  48 * (each on a different drive).  The starting device for each section is offset
  49 * near_copies from the starting device of the previous section.  Thus there
  50 * are (near_copies * far_copies) of each chunk, and each is on a different
  51 * drive.  near_copies and far_copies must be at least one, and their product
  52 * is at most raid_disks.
  53 *
  54 * If far_offset is true, then the far_copies are handled a bit differently.
  55 * The copies are still in different stripes, but instead of being very far
  56 * apart on disk, there are adjacent stripes.
  57 *
  58 * The far and offset algorithms are handled slightly differently if
  59 * 'use_far_sets' is true.  In this case, the array's devices are grouped into
  60 * sets that are (near_copies * far_copies) in size.  The far copied stripes
  61 * are still shifted by 'near_copies' devices, but this shifting stays confined
  62 * to the set rather than the entire array.  This is done to improve the number
  63 * of device combinations that can fail without causing the array to fail.
  64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
  65 * on a device):
  66 *    A B C D    A B C D E
  67 *      ...         ...
  68 *    D A B C    E A B C D
  69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
  70 *    [A B] [C D]    [A B] [C D E]
  71 *    |...| |...|    |...| | ... |
  72 *    [B A] [D C]    [B A] [E C D]
  73 */
  74
  75/*
  76 * Number of guaranteed r10bios in case of extreme VM load:
  77 */
  78#define	NR_RAID10_BIOS 256
  79
  80/* when we get a read error on a read-only array, we redirect to another
  81 * device without failing the first device, or trying to over-write to
  82 * correct the read error.  To keep track of bad blocks on a per-bio
  83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  84 */
  85#define IO_BLOCKED ((struct bio *)1)
  86/* When we successfully write to a known bad-block, we need to remove the
  87 * bad-block marking which must be done from process context.  So we record
  88 * the success by setting devs[n].bio to IO_MADE_GOOD
  89 */
  90#define IO_MADE_GOOD ((struct bio *)2)
  91
  92#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  93
  94/* When there are this many requests queued to be written by
  95 * the raid10 thread, we become 'congested' to provide back-pressure
  96 * for writeback.
  97 */
  98static int max_queued_requests = 1024;
  99
 100static void allow_barrier(struct r10conf *conf);
 101static void lower_barrier(struct r10conf *conf);
 102static int _enough(struct r10conf *conf, int previous, int ignore);
 103static int enough(struct r10conf *conf, int ignore);
 104static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
 105				int *skipped);
 106static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
 107static void end_reshape_write(struct bio *bio);
 108static void end_reshape(struct r10conf *conf);
 109
 110#define raid10_log(md, fmt, args...)				\
 111	do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
 112
 113#include "raid1-10.c"
 114
 115/*
 116 * for resync bio, r10bio pointer can be retrieved from the per-bio
 117 * 'struct resync_pages'.
 118 */
 119static inline struct r10bio *get_resync_r10bio(struct bio *bio)
 120{
 121	return get_resync_pages(bio)->raid_bio;
 122}
 123
 124static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 125{
 126	struct r10conf *conf = data;
 127	int size = offsetof(struct r10bio, devs[conf->copies]);
 128
 129	/* allocate a r10bio with room for raid_disks entries in the
 130	 * bios array */
 131	return kzalloc(size, gfp_flags);
 132}
 133
 134static void r10bio_pool_free(void *r10_bio, void *data)
 135{
 136	kfree(r10_bio);
 137}
 138
 139#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 
 
 140/* amount of memory to reserve for resync requests */
 141#define RESYNC_WINDOW (1024*1024)
 142/* maximum number of concurrent requests, memory permitting */
 143#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
 144#define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
 145#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 146
 147/*
 148 * When performing a resync, we need to read and compare, so
 149 * we need as many pages are there are copies.
 150 * When performing a recovery, we need 2 bios, one for read,
 151 * one for write (we recover only one drive per r10buf)
 152 *
 153 */
 154static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
 155{
 156	struct r10conf *conf = data;
 
 157	struct r10bio *r10_bio;
 158	struct bio *bio;
 159	int j;
 160	int nalloc, nalloc_rp;
 161	struct resync_pages *rps;
 162
 163	r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 164	if (!r10_bio)
 165		return NULL;
 166
 167	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
 168	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 169		nalloc = conf->copies; /* resync */
 170	else
 171		nalloc = 2; /* recovery */
 172
 173	/* allocate once for all bios */
 174	if (!conf->have_replacement)
 175		nalloc_rp = nalloc;
 176	else
 177		nalloc_rp = nalloc * 2;
 178	rps = kmalloc(sizeof(struct resync_pages) * nalloc_rp, gfp_flags);
 179	if (!rps)
 180		goto out_free_r10bio;
 181
 182	/*
 183	 * Allocate bios.
 184	 */
 185	for (j = nalloc ; j-- ; ) {
 186		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 187		if (!bio)
 188			goto out_free_bio;
 189		r10_bio->devs[j].bio = bio;
 190		if (!conf->have_replacement)
 191			continue;
 192		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 193		if (!bio)
 194			goto out_free_bio;
 195		r10_bio->devs[j].repl_bio = bio;
 196	}
 197	/*
 198	 * Allocate RESYNC_PAGES data pages and attach them
 199	 * where needed.
 200	 */
 201	for (j = 0; j < nalloc; j++) {
 202		struct bio *rbio = r10_bio->devs[j].repl_bio;
 203		struct resync_pages *rp, *rp_repl;
 204
 205		rp = &rps[j];
 206		if (rbio)
 207			rp_repl = &rps[nalloc + j];
 208
 209		bio = r10_bio->devs[j].bio;
 210
 211		if (!j || test_bit(MD_RECOVERY_SYNC,
 212				   &conf->mddev->recovery)) {
 213			if (resync_alloc_pages(rp, gfp_flags))
 
 
 
 
 
 
 
 214				goto out_free_pages;
 215		} else {
 216			memcpy(rp, &rps[0], sizeof(*rp));
 217			resync_get_all_pages(rp);
 218		}
 219
 220		rp->raid_bio = r10_bio;
 221		bio->bi_private = rp;
 222		if (rbio) {
 223			memcpy(rp_repl, rp, sizeof(*rp));
 224			rbio->bi_private = rp_repl;
 225		}
 226	}
 227
 228	return r10_bio;
 229
 230out_free_pages:
 231	while (--j >= 0)
 232		resync_free_pages(&rps[j * 2]);
 233
 
 
 234	j = 0;
 235out_free_bio:
 236	for ( ; j < nalloc; j++) {
 237		if (r10_bio->devs[j].bio)
 238			bio_put(r10_bio->devs[j].bio);
 239		if (r10_bio->devs[j].repl_bio)
 240			bio_put(r10_bio->devs[j].repl_bio);
 241	}
 242	kfree(rps);
 243out_free_r10bio:
 244	r10bio_pool_free(r10_bio, conf);
 245	return NULL;
 246}
 247
 248static void r10buf_pool_free(void *__r10_bio, void *data)
 249{
 
 250	struct r10conf *conf = data;
 251	struct r10bio *r10bio = __r10_bio;
 252	int j;
 253	struct resync_pages *rp = NULL;
 254
 255	for (j = conf->copies; j--; ) {
 256		struct bio *bio = r10bio->devs[j].bio;
 257
 258		rp = get_resync_pages(bio);
 259		resync_free_pages(rp);
 260		bio_put(bio);
 261
 
 
 262		bio = r10bio->devs[j].repl_bio;
 263		if (bio)
 264			bio_put(bio);
 265	}
 266
 267	/* resync pages array stored in the 1st bio's .bi_private */
 268	kfree(rp);
 269
 270	r10bio_pool_free(r10bio, conf);
 271}
 272
 273static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
 274{
 275	int i;
 276
 277	for (i = 0; i < conf->copies; i++) {
 278		struct bio **bio = & r10_bio->devs[i].bio;
 279		if (!BIO_SPECIAL(*bio))
 280			bio_put(*bio);
 281		*bio = NULL;
 282		bio = &r10_bio->devs[i].repl_bio;
 283		if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
 284			bio_put(*bio);
 285		*bio = NULL;
 286	}
 287}
 288
 289static void free_r10bio(struct r10bio *r10_bio)
 290{
 291	struct r10conf *conf = r10_bio->mddev->private;
 292
 293	put_all_bios(conf, r10_bio);
 294	mempool_free(r10_bio, conf->r10bio_pool);
 295}
 296
 297static void put_buf(struct r10bio *r10_bio)
 298{
 299	struct r10conf *conf = r10_bio->mddev->private;
 300
 301	mempool_free(r10_bio, conf->r10buf_pool);
 302
 303	lower_barrier(conf);
 304}
 305
 306static void reschedule_retry(struct r10bio *r10_bio)
 307{
 308	unsigned long flags;
 309	struct mddev *mddev = r10_bio->mddev;
 310	struct r10conf *conf = mddev->private;
 311
 312	spin_lock_irqsave(&conf->device_lock, flags);
 313	list_add(&r10_bio->retry_list, &conf->retry_list);
 314	conf->nr_queued ++;
 315	spin_unlock_irqrestore(&conf->device_lock, flags);
 316
 317	/* wake up frozen array... */
 318	wake_up(&conf->wait_barrier);
 319
 320	md_wakeup_thread(mddev->thread);
 321}
 322
 323/*
 324 * raid_end_bio_io() is called when we have finished servicing a mirrored
 325 * operation and are ready to return a success/failure code to the buffer
 326 * cache layer.
 327 */
 328static void raid_end_bio_io(struct r10bio *r10_bio)
 329{
 330	struct bio *bio = r10_bio->master_bio;
 
 331	struct r10conf *conf = r10_bio->mddev->private;
 332
 
 
 
 
 
 
 
 
 333	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
 334		bio->bi_status = BLK_STS_IOERR;
 335
 336	bio_endio(bio);
 337	/*
 338	 * Wake up any possible resync thread that waits for the device
 339	 * to go idle.
 340	 */
 341	allow_barrier(conf);
 342
 343	free_r10bio(r10_bio);
 344}
 345
 346/*
 347 * Update disk head position estimator based on IRQ completion info.
 348 */
 349static inline void update_head_pos(int slot, struct r10bio *r10_bio)
 350{
 351	struct r10conf *conf = r10_bio->mddev->private;
 352
 353	conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 354		r10_bio->devs[slot].addr + (r10_bio->sectors);
 355}
 356
 357/*
 358 * Find the disk number which triggered given bio
 359 */
 360static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
 361			 struct bio *bio, int *slotp, int *replp)
 362{
 363	int slot;
 364	int repl = 0;
 365
 366	for (slot = 0; slot < conf->copies; slot++) {
 367		if (r10_bio->devs[slot].bio == bio)
 368			break;
 369		if (r10_bio->devs[slot].repl_bio == bio) {
 370			repl = 1;
 371			break;
 372		}
 373	}
 374
 375	BUG_ON(slot == conf->copies);
 376	update_head_pos(slot, r10_bio);
 377
 378	if (slotp)
 379		*slotp = slot;
 380	if (replp)
 381		*replp = repl;
 382	return r10_bio->devs[slot].devnum;
 383}
 384
 385static void raid10_end_read_request(struct bio *bio)
 386{
 387	int uptodate = !bio->bi_status;
 388	struct r10bio *r10_bio = bio->bi_private;
 389	int slot;
 390	struct md_rdev *rdev;
 391	struct r10conf *conf = r10_bio->mddev->private;
 392
 
 393	slot = r10_bio->read_slot;
 
 394	rdev = r10_bio->devs[slot].rdev;
 395	/*
 396	 * this branch is our 'one mirror IO has finished' event handler:
 397	 */
 398	update_head_pos(slot, r10_bio);
 399
 400	if (uptodate) {
 401		/*
 402		 * Set R10BIO_Uptodate in our master bio, so that
 403		 * we will return a good error code to the higher
 404		 * levels even if IO on some other mirrored buffer fails.
 405		 *
 406		 * The 'master' represents the composite IO operation to
 407		 * user-side. So if something waits for IO, then it will
 408		 * wait for the 'master' bio.
 409		 */
 410		set_bit(R10BIO_Uptodate, &r10_bio->state);
 411	} else {
 412		/* If all other devices that store this block have
 413		 * failed, we want to return the error upwards rather
 414		 * than fail the last device.  Here we redefine
 415		 * "uptodate" to mean "Don't want to retry"
 416		 */
 417		if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
 418			     rdev->raid_disk))
 419			uptodate = 1;
 420	}
 421	if (uptodate) {
 422		raid_end_bio_io(r10_bio);
 423		rdev_dec_pending(rdev, conf->mddev);
 424	} else {
 425		/*
 426		 * oops, read error - keep the refcount on the rdev
 427		 */
 428		char b[BDEVNAME_SIZE];
 429		pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
 
 430				   mdname(conf->mddev),
 431				   bdevname(rdev->bdev, b),
 432				   (unsigned long long)r10_bio->sector);
 433		set_bit(R10BIO_ReadError, &r10_bio->state);
 434		reschedule_retry(r10_bio);
 435	}
 436}
 437
 438static void close_write(struct r10bio *r10_bio)
 439{
 440	/* clear the bitmap if all writes complete successfully */
 441	bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 442			r10_bio->sectors,
 443			!test_bit(R10BIO_Degraded, &r10_bio->state),
 444			0);
 445	md_write_end(r10_bio->mddev);
 446}
 447
 448static void one_write_done(struct r10bio *r10_bio)
 449{
 450	if (atomic_dec_and_test(&r10_bio->remaining)) {
 451		if (test_bit(R10BIO_WriteError, &r10_bio->state))
 452			reschedule_retry(r10_bio);
 453		else {
 454			close_write(r10_bio);
 455			if (test_bit(R10BIO_MadeGood, &r10_bio->state))
 456				reschedule_retry(r10_bio);
 457			else
 458				raid_end_bio_io(r10_bio);
 459		}
 460	}
 461}
 462
 463static void raid10_end_write_request(struct bio *bio)
 464{
 
 465	struct r10bio *r10_bio = bio->bi_private;
 466	int dev;
 467	int dec_rdev = 1;
 468	struct r10conf *conf = r10_bio->mddev->private;
 469	int slot, repl;
 470	struct md_rdev *rdev = NULL;
 471	struct bio *to_put = NULL;
 472	bool discard_error;
 473
 474	discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 475
 476	dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
 477
 478	if (repl)
 479		rdev = conf->mirrors[dev].replacement;
 480	if (!rdev) {
 481		smp_rmb();
 482		repl = 0;
 483		rdev = conf->mirrors[dev].rdev;
 484	}
 485	/*
 486	 * this branch is our 'one mirror IO has finished' event handler:
 487	 */
 488	if (bio->bi_status && !discard_error) {
 489		if (repl)
 490			/* Never record new bad blocks to replacement,
 491			 * just fail it.
 492			 */
 493			md_error(rdev->mddev, rdev);
 494		else {
 495			set_bit(WriteErrorSeen,	&rdev->flags);
 496			if (!test_and_set_bit(WantReplacement, &rdev->flags))
 497				set_bit(MD_RECOVERY_NEEDED,
 498					&rdev->mddev->recovery);
 499
 500			dec_rdev = 0;
 501			if (test_bit(FailFast, &rdev->flags) &&
 502			    (bio->bi_opf & MD_FAILFAST)) {
 503				md_error(rdev->mddev, rdev);
 504				if (!test_bit(Faulty, &rdev->flags))
 505					/* This is the only remaining device,
 506					 * We need to retry the write without
 507					 * FailFast
 508					 */
 509					set_bit(R10BIO_WriteError, &r10_bio->state);
 510				else {
 511					r10_bio->devs[slot].bio = NULL;
 512					to_put = bio;
 513					dec_rdev = 1;
 514				}
 515			} else
 516				set_bit(R10BIO_WriteError, &r10_bio->state);
 517		}
 518	} else {
 519		/*
 520		 * Set R10BIO_Uptodate in our master bio, so that
 521		 * we will return a good error code for to the higher
 522		 * levels even if IO on some other mirrored buffer fails.
 523		 *
 524		 * The 'master' represents the composite IO operation to
 525		 * user-side. So if something waits for IO, then it will
 526		 * wait for the 'master' bio.
 527		 */
 528		sector_t first_bad;
 529		int bad_sectors;
 530
 531		/*
 532		 * Do not set R10BIO_Uptodate if the current device is
 533		 * rebuilding or Faulty. This is because we cannot use
 534		 * such device for properly reading the data back (we could
 535		 * potentially use it, if the current write would have felt
 536		 * before rdev->recovery_offset, but for simplicity we don't
 537		 * check this here.
 538		 */
 539		if (test_bit(In_sync, &rdev->flags) &&
 540		    !test_bit(Faulty, &rdev->flags))
 541			set_bit(R10BIO_Uptodate, &r10_bio->state);
 542
 543		/* Maybe we can clear some bad blocks. */
 544		if (is_badblock(rdev,
 545				r10_bio->devs[slot].addr,
 546				r10_bio->sectors,
 547				&first_bad, &bad_sectors) && !discard_error) {
 548			bio_put(bio);
 549			if (repl)
 550				r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
 551			else
 552				r10_bio->devs[slot].bio = IO_MADE_GOOD;
 553			dec_rdev = 0;
 554			set_bit(R10BIO_MadeGood, &r10_bio->state);
 555		}
 556	}
 557
 558	/*
 559	 *
 560	 * Let's see if all mirrored write operations have finished
 561	 * already.
 562	 */
 563	one_write_done(r10_bio);
 564	if (dec_rdev)
 565		rdev_dec_pending(rdev, conf->mddev);
 566	if (to_put)
 567		bio_put(to_put);
 568}
 569
 570/*
 571 * RAID10 layout manager
 572 * As well as the chunksize and raid_disks count, there are two
 573 * parameters: near_copies and far_copies.
 574 * near_copies * far_copies must be <= raid_disks.
 575 * Normally one of these will be 1.
 576 * If both are 1, we get raid0.
 577 * If near_copies == raid_disks, we get raid1.
 578 *
 579 * Chunks are laid out in raid0 style with near_copies copies of the
 580 * first chunk, followed by near_copies copies of the next chunk and
 581 * so on.
 582 * If far_copies > 1, then after 1/far_copies of the array has been assigned
 583 * as described above, we start again with a device offset of near_copies.
 584 * So we effectively have another copy of the whole array further down all
 585 * the drives, but with blocks on different drives.
 586 * With this layout, and block is never stored twice on the one device.
 587 *
 588 * raid10_find_phys finds the sector offset of a given virtual sector
 589 * on each device that it is on.
 590 *
 591 * raid10_find_virt does the reverse mapping, from a device and a
 592 * sector offset to a virtual address
 593 */
 594
 595static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 596{
 597	int n,f;
 598	sector_t sector;
 599	sector_t chunk;
 600	sector_t stripe;
 601	int dev;
 602	int slot = 0;
 603	int last_far_set_start, last_far_set_size;
 604
 605	last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 606	last_far_set_start *= geo->far_set_size;
 607
 608	last_far_set_size = geo->far_set_size;
 609	last_far_set_size += (geo->raid_disks % geo->far_set_size);
 610
 611	/* now calculate first sector/dev */
 612	chunk = r10bio->sector >> geo->chunk_shift;
 613	sector = r10bio->sector & geo->chunk_mask;
 614
 615	chunk *= geo->near_copies;
 616	stripe = chunk;
 617	dev = sector_div(stripe, geo->raid_disks);
 618	if (geo->far_offset)
 619		stripe *= geo->far_copies;
 620
 621	sector += stripe << geo->chunk_shift;
 622
 623	/* and calculate all the others */
 624	for (n = 0; n < geo->near_copies; n++) {
 625		int d = dev;
 626		int set;
 627		sector_t s = sector;
 628		r10bio->devs[slot].devnum = d;
 629		r10bio->devs[slot].addr = s;
 630		slot++;
 631
 632		for (f = 1; f < geo->far_copies; f++) {
 633			set = d / geo->far_set_size;
 634			d += geo->near_copies;
 635
 636			if ((geo->raid_disks % geo->far_set_size) &&
 637			    (d > last_far_set_start)) {
 638				d -= last_far_set_start;
 639				d %= last_far_set_size;
 640				d += last_far_set_start;
 641			} else {
 642				d %= geo->far_set_size;
 643				d += geo->far_set_size * set;
 644			}
 645			s += geo->stride;
 646			r10bio->devs[slot].devnum = d;
 647			r10bio->devs[slot].addr = s;
 648			slot++;
 649		}
 650		dev++;
 651		if (dev >= geo->raid_disks) {
 652			dev = 0;
 653			sector += (geo->chunk_mask + 1);
 654		}
 655	}
 656}
 657
 658static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
 659{
 660	struct geom *geo = &conf->geo;
 661
 662	if (conf->reshape_progress != MaxSector &&
 663	    ((r10bio->sector >= conf->reshape_progress) !=
 664	     conf->mddev->reshape_backwards)) {
 665		set_bit(R10BIO_Previous, &r10bio->state);
 666		geo = &conf->prev;
 667	} else
 668		clear_bit(R10BIO_Previous, &r10bio->state);
 669
 670	__raid10_find_phys(geo, r10bio);
 671}
 672
 673static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 674{
 675	sector_t offset, chunk, vchunk;
 676	/* Never use conf->prev as this is only called during resync
 677	 * or recovery, so reshape isn't happening
 678	 */
 679	struct geom *geo = &conf->geo;
 680	int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
 681	int far_set_size = geo->far_set_size;
 682	int last_far_set_start;
 683
 684	if (geo->raid_disks % geo->far_set_size) {
 685		last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
 686		last_far_set_start *= geo->far_set_size;
 687
 688		if (dev >= last_far_set_start) {
 689			far_set_size = geo->far_set_size;
 690			far_set_size += (geo->raid_disks % geo->far_set_size);
 691			far_set_start = last_far_set_start;
 692		}
 693	}
 694
 695	offset = sector & geo->chunk_mask;
 696	if (geo->far_offset) {
 697		int fc;
 698		chunk = sector >> geo->chunk_shift;
 699		fc = sector_div(chunk, geo->far_copies);
 700		dev -= fc * geo->near_copies;
 701		if (dev < far_set_start)
 702			dev += far_set_size;
 703	} else {
 704		while (sector >= geo->stride) {
 705			sector -= geo->stride;
 706			if (dev < (geo->near_copies + far_set_start))
 707				dev += far_set_size - geo->near_copies;
 708			else
 709				dev -= geo->near_copies;
 710		}
 711		chunk = sector >> geo->chunk_shift;
 712	}
 713	vchunk = chunk * geo->raid_disks + dev;
 714	sector_div(vchunk, geo->near_copies);
 715	return (vchunk << geo->chunk_shift) + offset;
 716}
 717
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 718/*
 719 * This routine returns the disk from which the requested read should
 720 * be done. There is a per-array 'next expected sequential IO' sector
 721 * number - if this matches on the next IO then we use the last disk.
 722 * There is also a per-disk 'last know head position' sector that is
 723 * maintained from IRQ contexts, both the normal and the resync IO
 724 * completion handlers update this position correctly. If there is no
 725 * perfect sequential match then we pick the disk whose head is closest.
 726 *
 727 * If there are 2 mirrors in the same 2 devices, performance degrades
 728 * because position is mirror, not device based.
 729 *
 730 * The rdev for the device selected will have nr_pending incremented.
 731 */
 732
 733/*
 734 * FIXME: possibly should rethink readbalancing and do it differently
 735 * depending on near_copies / far_copies geometry.
 736 */
 737static struct md_rdev *read_balance(struct r10conf *conf,
 738				    struct r10bio *r10_bio,
 739				    int *max_sectors)
 740{
 741	const sector_t this_sector = r10_bio->sector;
 742	int disk, slot;
 743	int sectors = r10_bio->sectors;
 744	int best_good_sectors;
 745	sector_t new_distance, best_dist;
 746	struct md_rdev *best_rdev, *rdev = NULL;
 747	int do_balance;
 748	int best_slot;
 749	struct geom *geo = &conf->geo;
 750
 751	raid10_find_phys(conf, r10_bio);
 752	rcu_read_lock();
 
 
 753	best_slot = -1;
 754	best_rdev = NULL;
 755	best_dist = MaxSector;
 756	best_good_sectors = 0;
 757	do_balance = 1;
 758	clear_bit(R10BIO_FailFast, &r10_bio->state);
 759	/*
 760	 * Check if we can balance. We can balance on the whole
 761	 * device if no resync is going on (recovery is ok), or below
 762	 * the resync window. We take the first readable disk when
 763	 * above the resync window.
 764	 */
 765	if ((conf->mddev->recovery_cp < MaxSector
 766	     && (this_sector + sectors >= conf->next_resync)) ||
 767	    (mddev_is_clustered(conf->mddev) &&
 768	     md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 769					    this_sector + sectors)))
 770		do_balance = 0;
 771
 772	for (slot = 0; slot < conf->copies ; slot++) {
 773		sector_t first_bad;
 774		int bad_sectors;
 775		sector_t dev_sector;
 776
 777		if (r10_bio->devs[slot].bio == IO_BLOCKED)
 778			continue;
 779		disk = r10_bio->devs[slot].devnum;
 780		rdev = rcu_dereference(conf->mirrors[disk].replacement);
 781		if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
 
 782		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 783			rdev = rcu_dereference(conf->mirrors[disk].rdev);
 784		if (rdev == NULL ||
 785		    test_bit(Faulty, &rdev->flags))
 
 786			continue;
 787		if (!test_bit(In_sync, &rdev->flags) &&
 788		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
 789			continue;
 790
 791		dev_sector = r10_bio->devs[slot].addr;
 792		if (is_badblock(rdev, dev_sector, sectors,
 793				&first_bad, &bad_sectors)) {
 794			if (best_dist < MaxSector)
 795				/* Already have a better slot */
 796				continue;
 797			if (first_bad <= dev_sector) {
 798				/* Cannot read here.  If this is the
 799				 * 'primary' device, then we must not read
 800				 * beyond 'bad_sectors' from another device.
 801				 */
 802				bad_sectors -= (dev_sector - first_bad);
 803				if (!do_balance && sectors > bad_sectors)
 804					sectors = bad_sectors;
 805				if (best_good_sectors > sectors)
 806					best_good_sectors = sectors;
 807			} else {
 808				sector_t good_sectors =
 809					first_bad - dev_sector;
 810				if (good_sectors > best_good_sectors) {
 811					best_good_sectors = good_sectors;
 812					best_slot = slot;
 813					best_rdev = rdev;
 814				}
 815				if (!do_balance)
 816					/* Must read from here */
 817					break;
 818			}
 819			continue;
 820		} else
 821			best_good_sectors = sectors;
 822
 823		if (!do_balance)
 824			break;
 825
 826		if (best_slot >= 0)
 827			/* At least 2 disks to choose from so failfast is OK */
 828			set_bit(R10BIO_FailFast, &r10_bio->state);
 829		/* This optimisation is debatable, and completely destroys
 830		 * sequential read speed for 'far copies' arrays.  So only
 831		 * keep it for 'near' arrays, and review those later.
 832		 */
 833		if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 834			new_distance = 0;
 835
 836		/* for far > 1 always use the lowest address */
 837		else if (geo->far_copies > 1)
 838			new_distance = r10_bio->devs[slot].addr;
 839		else
 840			new_distance = abs(r10_bio->devs[slot].addr -
 841					   conf->mirrors[disk].head_position);
 842		if (new_distance < best_dist) {
 843			best_dist = new_distance;
 844			best_slot = slot;
 845			best_rdev = rdev;
 846		}
 847	}
 848	if (slot >= conf->copies) {
 849		slot = best_slot;
 850		rdev = best_rdev;
 851	}
 852
 853	if (slot >= 0) {
 854		atomic_inc(&rdev->nr_pending);
 
 
 
 
 
 
 
 855		r10_bio->read_slot = slot;
 856	} else
 857		rdev = NULL;
 858	rcu_read_unlock();
 859	*max_sectors = best_good_sectors;
 860
 861	return rdev;
 862}
 863
 864static int raid10_congested(struct mddev *mddev, int bits)
 865{
 866	struct r10conf *conf = mddev->private;
 867	int i, ret = 0;
 868
 869	if ((bits & (1 << WB_async_congested)) &&
 870	    conf->pending_count >= max_queued_requests)
 871		return 1;
 872
 873	rcu_read_lock();
 874	for (i = 0;
 875	     (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
 876		     && ret == 0;
 877	     i++) {
 878		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 879		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 880			struct request_queue *q = bdev_get_queue(rdev->bdev);
 881
 882			ret |= bdi_congested(q->backing_dev_info, bits);
 883		}
 884	}
 885	rcu_read_unlock();
 886	return ret;
 887}
 
 
 
 
 
 
 
 
 
 888
 889static void flush_pending_writes(struct r10conf *conf)
 890{
 891	/* Any writes that have been queued but are awaiting
 892	 * bitmap updates get flushed here.
 893	 */
 894	spin_lock_irq(&conf->device_lock);
 895
 896	if (conf->pending_bio_list.head) {
 897		struct blk_plug plug;
 898		struct bio *bio;
 899
 900		bio = bio_list_get(&conf->pending_bio_list);
 901		conf->pending_count = 0;
 902		spin_unlock_irq(&conf->device_lock);
 903
 904		/*
 905		 * As this is called in a wait_event() loop (see freeze_array),
 906		 * current->state might be TASK_UNINTERRUPTIBLE which will
 907		 * cause a warning when we prepare to wait again.  As it is
 908		 * rare that this path is taken, it is perfectly safe to force
 909		 * us to go around the wait_event() loop again, so the warning
 910		 * is a false-positive. Silence the warning by resetting
 911		 * thread state
 912		 */
 913		__set_current_state(TASK_RUNNING);
 914
 915		blk_start_plug(&plug);
 916		/* flush any pending bitmap writes to disk
 917		 * before proceeding w/ I/O */
 918		bitmap_unplug(conf->mddev->bitmap);
 919		wake_up(&conf->wait_barrier);
 920
 921		while (bio) { /* submit pending writes */
 922			struct bio *next = bio->bi_next;
 923			struct md_rdev *rdev = (void*)bio->bi_disk;
 924			bio->bi_next = NULL;
 925			bio_set_dev(bio, rdev->bdev);
 926			if (test_bit(Faulty, &rdev->flags)) {
 927				bio_io_error(bio);
 928			} else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
 929					    !blk_queue_discard(bio->bi_disk->queue)))
 930				/* Just ignore it */
 931				bio_endio(bio);
 932			else
 933				generic_make_request(bio);
 934			bio = next;
 935		}
 936		blk_finish_plug(&plug);
 937	} else
 938		spin_unlock_irq(&conf->device_lock);
 939}
 940
 941/* Barriers....
 942 * Sometimes we need to suspend IO while we do something else,
 943 * either some resync/recovery, or reconfigure the array.
 944 * To do this we raise a 'barrier'.
 945 * The 'barrier' is a counter that can be raised multiple times
 946 * to count how many activities are happening which preclude
 947 * normal IO.
 948 * We can only raise the barrier if there is no pending IO.
 949 * i.e. if nr_pending == 0.
 950 * We choose only to raise the barrier if no-one is waiting for the
 951 * barrier to go down.  This means that as soon as an IO request
 952 * is ready, no other operations which require a barrier will start
 953 * until the IO request has had a chance.
 954 *
 955 * So: regular IO calls 'wait_barrier'.  When that returns there
 956 *    is no backgroup IO happening,  It must arrange to call
 957 *    allow_barrier when it has finished its IO.
 958 * backgroup IO calls must call raise_barrier.  Once that returns
 959 *    there is no normal IO happeing.  It must arrange to call
 960 *    lower_barrier when the particular background IO completes.
 961 */
 962
 963static void raise_barrier(struct r10conf *conf, int force)
 964{
 965	BUG_ON(force && !conf->barrier);
 966	spin_lock_irq(&conf->resync_lock);
 967
 968	/* Wait until no block IO is waiting (unless 'force') */
 969	wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 970			    conf->resync_lock);
 971
 972	/* block any new IO from starting */
 973	conf->barrier++;
 974
 975	/* Now wait for all pending IO to complete */
 976	wait_event_lock_irq(conf->wait_barrier,
 977			    !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
 978			    conf->resync_lock);
 979
 980	spin_unlock_irq(&conf->resync_lock);
 981}
 982
 983static void lower_barrier(struct r10conf *conf)
 984{
 985	unsigned long flags;
 986	spin_lock_irqsave(&conf->resync_lock, flags);
 987	conf->barrier--;
 988	spin_unlock_irqrestore(&conf->resync_lock, flags);
 989	wake_up(&conf->wait_barrier);
 990}
 991
 992static void wait_barrier(struct r10conf *conf)
 993{
 994	spin_lock_irq(&conf->resync_lock);
 995	if (conf->barrier) {
 996		conf->nr_waiting++;
 997		/* Wait for the barrier to drop.
 998		 * However if there are already pending
 999		 * requests (preventing the barrier from
1000		 * rising completely), and the
1001		 * pre-process bio queue isn't empty,
1002		 * then don't wait, as we need to empty
1003		 * that queue to get the nr_pending
1004		 * count down.
1005		 */
1006		raid10_log(conf->mddev, "wait barrier");
1007		wait_event_lock_irq(conf->wait_barrier,
1008				    !conf->barrier ||
1009				    (atomic_read(&conf->nr_pending) &&
1010				     current->bio_list &&
1011				     (!bio_list_empty(&current->bio_list[0]) ||
1012				      !bio_list_empty(&current->bio_list[1]))),
1013				    conf->resync_lock);
1014		conf->nr_waiting--;
1015		if (!conf->nr_waiting)
1016			wake_up(&conf->wait_barrier);
1017	}
1018	atomic_inc(&conf->nr_pending);
1019	spin_unlock_irq(&conf->resync_lock);
1020}
1021
1022static void allow_barrier(struct r10conf *conf)
1023{
1024	if ((atomic_dec_and_test(&conf->nr_pending)) ||
1025			(conf->array_freeze_pending))
1026		wake_up(&conf->wait_barrier);
 
 
1027}
1028
1029static void freeze_array(struct r10conf *conf, int extra)
1030{
1031	/* stop syncio and normal IO and wait for everything to
1032	 * go quiet.
1033	 * We increment barrier and nr_waiting, and then
1034	 * wait until nr_pending match nr_queued+extra
1035	 * This is called in the context of one normal IO request
1036	 * that has failed. Thus any sync request that might be pending
1037	 * will be blocked by nr_pending, and we need to wait for
1038	 * pending IO requests to complete or be queued for re-try.
1039	 * Thus the number queued (nr_queued) plus this request (extra)
1040	 * must match the number of pending IOs (nr_pending) before
1041	 * we continue.
1042	 */
1043	spin_lock_irq(&conf->resync_lock);
1044	conf->array_freeze_pending++;
1045	conf->barrier++;
1046	conf->nr_waiting++;
1047	wait_event_lock_irq_cmd(conf->wait_barrier,
1048				atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1049				conf->resync_lock,
1050				flush_pending_writes(conf));
1051
1052	conf->array_freeze_pending--;
1053	spin_unlock_irq(&conf->resync_lock);
1054}
1055
1056static void unfreeze_array(struct r10conf *conf)
1057{
1058	/* reverse the effect of the freeze */
1059	spin_lock_irq(&conf->resync_lock);
1060	conf->barrier--;
1061	conf->nr_waiting--;
1062	wake_up(&conf->wait_barrier);
1063	spin_unlock_irq(&conf->resync_lock);
1064}
1065
1066static sector_t choose_data_offset(struct r10bio *r10_bio,
1067				   struct md_rdev *rdev)
1068{
1069	if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1070	    test_bit(R10BIO_Previous, &r10_bio->state))
1071		return rdev->data_offset;
1072	else
1073		return rdev->new_data_offset;
1074}
1075
1076struct raid10_plug_cb {
1077	struct blk_plug_cb	cb;
1078	struct bio_list		pending;
1079	int			pending_cnt;
1080};
1081
1082static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1083{
1084	struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1085						   cb);
1086	struct mddev *mddev = plug->cb.data;
1087	struct r10conf *conf = mddev->private;
1088	struct bio *bio;
1089
1090	if (from_schedule || current->bio_list) {
1091		spin_lock_irq(&conf->device_lock);
1092		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1093		conf->pending_count += plug->pending_cnt;
1094		spin_unlock_irq(&conf->device_lock);
1095		wake_up(&conf->wait_barrier);
1096		md_wakeup_thread(mddev->thread);
1097		kfree(plug);
1098		return;
1099	}
1100
1101	/* we aren't scheduling, so we can do the write-out directly. */
1102	bio = bio_list_get(&plug->pending);
1103	bitmap_unplug(mddev->bitmap);
1104	wake_up(&conf->wait_barrier);
1105
1106	while (bio) { /* submit pending writes */
1107		struct bio *next = bio->bi_next;
1108		struct md_rdev *rdev = (void*)bio->bi_disk;
1109		bio->bi_next = NULL;
1110		bio_set_dev(bio, rdev->bdev);
1111		if (test_bit(Faulty, &rdev->flags)) {
1112			bio_io_error(bio);
1113		} else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
1114				    !blk_queue_discard(bio->bi_disk->queue)))
1115			/* Just ignore it */
1116			bio_endio(bio);
1117		else
1118			generic_make_request(bio);
1119		bio = next;
1120	}
1121	kfree(plug);
1122}
1123
1124static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1125				struct r10bio *r10_bio)
1126{
1127	struct r10conf *conf = mddev->private;
 
1128	struct bio *read_bio;
1129	const int op = bio_op(bio);
1130	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1131	int max_sectors;
1132	sector_t sectors;
1133	struct md_rdev *rdev;
1134	char b[BDEVNAME_SIZE];
1135	int slot = r10_bio->read_slot;
1136	struct md_rdev *err_rdev = NULL;
1137	gfp_t gfp = GFP_NOIO;
1138
1139	if (r10_bio->devs[slot].rdev) {
1140		/*
1141		 * This is an error retry, but we cannot
1142		 * safely dereference the rdev in the r10_bio,
1143		 * we must use the one in conf.
1144		 * If it has already been disconnected (unlikely)
1145		 * we lose the device name in error messages.
1146		 */
1147		int disk;
1148		/*
1149		 * As we are blocking raid10, it is a little safer to
1150		 * use __GFP_HIGH.
1151		 */
1152		gfp = GFP_NOIO | __GFP_HIGH;
1153
1154		rcu_read_lock();
1155		disk = r10_bio->devs[slot].devnum;
1156		err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1157		if (err_rdev)
1158			bdevname(err_rdev->bdev, b);
1159		else {
1160			strcpy(b, "???");
1161			/* This never gets dereferenced */
1162			err_rdev = r10_bio->devs[slot].rdev;
1163		}
1164		rcu_read_unlock();
1165	}
1166	/*
1167	 * Register the new request and wait if the reconstruction
1168	 * thread has put up a bar for new requests.
1169	 * Continue immediately if no resync is active currently.
1170	 */
1171	wait_barrier(conf);
1172
1173	sectors = r10_bio->sectors;
1174	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1175	    bio->bi_iter.bi_sector < conf->reshape_progress &&
1176	    bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1177		/*
1178		 * IO spans the reshape position.  Need to wait for reshape to
1179		 * pass
1180		 */
1181		raid10_log(conf->mddev, "wait reshape");
1182		allow_barrier(conf);
1183		wait_event(conf->wait_barrier,
1184			   conf->reshape_progress <= bio->bi_iter.bi_sector ||
1185			   conf->reshape_progress >= bio->bi_iter.bi_sector +
1186			   sectors);
1187		wait_barrier(conf);
1188	}
1189
1190	rdev = read_balance(conf, r10_bio, &max_sectors);
1191	if (!rdev) {
1192		if (err_rdev) {
1193			pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1194					    mdname(mddev), b,
1195					    (unsigned long long)r10_bio->sector);
1196		}
1197		raid_end_bio_io(r10_bio);
1198		return;
1199	}
1200	if (err_rdev)
1201		pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1202				   mdname(mddev),
1203				   bdevname(rdev->bdev, b),
1204				   (unsigned long long)r10_bio->sector);
1205	if (max_sectors < bio_sectors(bio)) {
1206		struct bio *split = bio_split(bio, max_sectors,
1207					      gfp, conf->bio_split);
1208		bio_chain(split, bio);
1209		generic_make_request(bio);
1210		bio = split;
1211		r10_bio->master_bio = bio;
1212		r10_bio->sectors = max_sectors;
1213	}
1214	slot = r10_bio->read_slot;
1215
1216	read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1217
1218	r10_bio->devs[slot].bio = read_bio;
1219	r10_bio->devs[slot].rdev = rdev;
1220
1221	read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1222		choose_data_offset(r10_bio, rdev);
1223	bio_set_dev(read_bio, rdev->bdev);
1224	read_bio->bi_end_io = raid10_end_read_request;
1225	bio_set_op_attrs(read_bio, op, do_sync);
1226	if (test_bit(FailFast, &rdev->flags) &&
1227	    test_bit(R10BIO_FailFast, &r10_bio->state))
1228	        read_bio->bi_opf |= MD_FAILFAST;
1229	read_bio->bi_private = r10_bio;
1230
1231	if (mddev->gendisk)
1232	        trace_block_bio_remap(read_bio->bi_disk->queue,
1233	                              read_bio, disk_devt(mddev->gendisk),
1234	                              r10_bio->sector);
1235	generic_make_request(read_bio);
1236	return;
1237}
1238
1239static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1240				  struct bio *bio, bool replacement,
1241				  int n_copy)
1242{
1243	const int op = bio_op(bio);
1244	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1245	const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1246	unsigned long flags;
 
1247	struct blk_plug_cb *cb;
1248	struct raid10_plug_cb *plug = NULL;
1249	struct r10conf *conf = mddev->private;
1250	struct md_rdev *rdev;
1251	int devnum = r10_bio->devs[n_copy].devnum;
1252	struct bio *mbio;
1253
1254	if (replacement) {
1255		rdev = conf->mirrors[devnum].replacement;
1256		if (rdev == NULL) {
1257			/* Replacement just got moved to main 'rdev' */
1258			smp_mb();
1259			rdev = conf->mirrors[devnum].rdev;
1260		}
1261	} else
1262		rdev = conf->mirrors[devnum].rdev;
1263
1264	mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1265	if (replacement)
1266		r10_bio->devs[n_copy].repl_bio = mbio;
1267	else
1268		r10_bio->devs[n_copy].bio = mbio;
1269
1270	mbio->bi_iter.bi_sector	= (r10_bio->devs[n_copy].addr +
1271				   choose_data_offset(r10_bio, rdev));
1272	bio_set_dev(mbio, rdev->bdev);
1273	mbio->bi_end_io	= raid10_end_write_request;
1274	bio_set_op_attrs(mbio, op, do_sync | do_fua);
1275	if (!replacement && test_bit(FailFast,
1276				     &conf->mirrors[devnum].rdev->flags)
1277			 && enough(conf, devnum))
1278		mbio->bi_opf |= MD_FAILFAST;
1279	mbio->bi_private = r10_bio;
1280
1281	if (conf->mddev->gendisk)
1282		trace_block_bio_remap(mbio->bi_disk->queue,
1283				      mbio, disk_devt(conf->mddev->gendisk),
1284				      r10_bio->sector);
1285	/* flush_pending_writes() needs access to the rdev so...*/
1286	mbio->bi_disk = (void *)rdev;
1287
1288	atomic_inc(&r10_bio->remaining);
1289
1290	cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1291	if (cb)
1292		plug = container_of(cb, struct raid10_plug_cb, cb);
1293	else
1294		plug = NULL;
1295	if (plug) {
1296		bio_list_add(&plug->pending, mbio);
1297		plug->pending_cnt++;
1298	} else {
1299		spin_lock_irqsave(&conf->device_lock, flags);
1300		bio_list_add(&conf->pending_bio_list, mbio);
1301		conf->pending_count++;
1302		spin_unlock_irqrestore(&conf->device_lock, flags);
1303		md_wakeup_thread(mddev->thread);
1304	}
1305}
1306
1307static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1308				 struct r10bio *r10_bio)
1309{
1310	struct r10conf *conf = mddev->private;
1311	int i;
1312	struct md_rdev *blocked_rdev;
1313	sector_t sectors;
1314	int max_sectors;
1315
1316	if ((mddev_is_clustered(mddev) &&
1317	     md_cluster_ops->area_resyncing(mddev, WRITE,
1318					    bio->bi_iter.bi_sector,
1319					    bio_end_sector(bio)))) {
1320		DEFINE_WAIT(w);
1321		for (;;) {
1322			prepare_to_wait(&conf->wait_barrier,
1323					&w, TASK_IDLE);
1324			if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325				 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1326				break;
1327			schedule();
1328		}
1329		finish_wait(&conf->wait_barrier, &w);
1330	}
1331
1332	/*
1333	 * Register the new request and wait if the reconstruction
1334	 * thread has put up a bar for new requests.
1335	 * Continue immediately if no resync is active currently.
1336	 */
1337	wait_barrier(conf);
1338
1339	sectors = r10_bio->sectors;
1340	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1341	    bio->bi_iter.bi_sector < conf->reshape_progress &&
1342	    bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1343		/*
1344		 * IO spans the reshape position.  Need to wait for reshape to
1345		 * pass
1346		 */
1347		raid10_log(conf->mddev, "wait reshape");
1348		allow_barrier(conf);
1349		wait_event(conf->wait_barrier,
1350			   conf->reshape_progress <= bio->bi_iter.bi_sector ||
1351			   conf->reshape_progress >= bio->bi_iter.bi_sector +
1352			   sectors);
1353		wait_barrier(conf);
1354	}
1355
1356	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
 
1357	    (mddev->reshape_backwards
1358	     ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1359		bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1360	     : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1361		bio->bi_iter.bi_sector < conf->reshape_progress))) {
1362		/* Need to update reshape_position in metadata */
1363		mddev->reshape_position = conf->reshape_progress;
1364		set_mask_bits(&mddev->sb_flags, 0,
1365			      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1366		md_wakeup_thread(mddev->thread);
1367		raid10_log(conf->mddev, "wait reshape metadata");
1368		wait_event(mddev->sb_wait,
1369			   !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1370
1371		conf->reshape_safe = mddev->reshape_position;
1372	}
1373
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1374	if (conf->pending_count >= max_queued_requests) {
1375		md_wakeup_thread(mddev->thread);
1376		raid10_log(mddev, "wait queued");
1377		wait_event(conf->wait_barrier,
1378			   conf->pending_count < max_queued_requests);
1379	}
1380	/* first select target devices under rcu_lock and
1381	 * inc refcount on their rdev.  Record them by setting
1382	 * bios[x] to bio
1383	 * If there are known/acknowledged bad blocks on any device
1384	 * on which we have seen a write error, we want to avoid
1385	 * writing to those blocks.  This potentially requires several
1386	 * writes to write around the bad blocks.  Each set of writes
1387	 * gets its own r10_bio with a set of bios attached.
 
 
1388	 */
1389
1390	r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1391	raid10_find_phys(conf, r10_bio);
1392retry_write:
1393	blocked_rdev = NULL;
1394	rcu_read_lock();
1395	max_sectors = r10_bio->sectors;
1396
1397	for (i = 0;  i < conf->copies; i++) {
1398		int d = r10_bio->devs[i].devnum;
1399		struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1400		struct md_rdev *rrdev = rcu_dereference(
1401			conf->mirrors[d].replacement);
1402		if (rdev == rrdev)
1403			rrdev = NULL;
1404		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1405			atomic_inc(&rdev->nr_pending);
1406			blocked_rdev = rdev;
1407			break;
1408		}
1409		if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1410			atomic_inc(&rrdev->nr_pending);
1411			blocked_rdev = rrdev;
1412			break;
1413		}
1414		if (rdev && (test_bit(Faulty, &rdev->flags)))
 
1415			rdev = NULL;
1416		if (rrdev && (test_bit(Faulty, &rrdev->flags)))
 
1417			rrdev = NULL;
1418
1419		r10_bio->devs[i].bio = NULL;
1420		r10_bio->devs[i].repl_bio = NULL;
1421
1422		if (!rdev && !rrdev) {
1423			set_bit(R10BIO_Degraded, &r10_bio->state);
1424			continue;
1425		}
1426		if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1427			sector_t first_bad;
1428			sector_t dev_sector = r10_bio->devs[i].addr;
1429			int bad_sectors;
1430			int is_bad;
1431
1432			is_bad = is_badblock(rdev, dev_sector, max_sectors,
 
1433					     &first_bad, &bad_sectors);
1434			if (is_bad < 0) {
1435				/* Mustn't write here until the bad block
1436				 * is acknowledged
1437				 */
1438				atomic_inc(&rdev->nr_pending);
1439				set_bit(BlockedBadBlocks, &rdev->flags);
1440				blocked_rdev = rdev;
1441				break;
1442			}
1443			if (is_bad && first_bad <= dev_sector) {
1444				/* Cannot write here at all */
1445				bad_sectors -= (dev_sector - first_bad);
1446				if (bad_sectors < max_sectors)
1447					/* Mustn't write more than bad_sectors
1448					 * to other devices yet
1449					 */
1450					max_sectors = bad_sectors;
1451				/* We don't set R10BIO_Degraded as that
1452				 * only applies if the disk is missing,
1453				 * so it might be re-added, and we want to
1454				 * know to recover this chunk.
1455				 * In this case the device is here, and the
1456				 * fact that this chunk is not in-sync is
1457				 * recorded in the bad block log.
1458				 */
1459				continue;
1460			}
1461			if (is_bad) {
1462				int good_sectors = first_bad - dev_sector;
1463				if (good_sectors < max_sectors)
1464					max_sectors = good_sectors;
1465			}
1466		}
1467		if (rdev) {
1468			r10_bio->devs[i].bio = bio;
1469			atomic_inc(&rdev->nr_pending);
1470		}
1471		if (rrdev) {
1472			r10_bio->devs[i].repl_bio = bio;
1473			atomic_inc(&rrdev->nr_pending);
1474		}
1475	}
1476	rcu_read_unlock();
1477
1478	if (unlikely(blocked_rdev)) {
1479		/* Have to wait for this device to get unblocked, then retry */
1480		int j;
1481		int d;
1482
1483		for (j = 0; j < i; j++) {
1484			if (r10_bio->devs[j].bio) {
1485				d = r10_bio->devs[j].devnum;
1486				rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1487			}
1488			if (r10_bio->devs[j].repl_bio) {
1489				struct md_rdev *rdev;
1490				d = r10_bio->devs[j].devnum;
1491				rdev = conf->mirrors[d].replacement;
1492				if (!rdev) {
1493					/* Race with remove_disk */
1494					smp_mb();
1495					rdev = conf->mirrors[d].rdev;
1496				}
1497				rdev_dec_pending(rdev, mddev);
1498			}
1499		}
1500		allow_barrier(conf);
1501		raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1502		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1503		wait_barrier(conf);
1504		goto retry_write;
1505	}
1506
1507	if (max_sectors < r10_bio->sectors)
 
 
 
1508		r10_bio->sectors = max_sectors;
1509
1510	if (r10_bio->sectors < bio_sectors(bio)) {
1511		struct bio *split = bio_split(bio, r10_bio->sectors,
1512					      GFP_NOIO, conf->bio_split);
1513		bio_chain(split, bio);
1514		generic_make_request(bio);
1515		bio = split;
1516		r10_bio->master_bio = bio;
1517	}
 
 
1518
1519	atomic_set(&r10_bio->remaining, 1);
1520	bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1521
1522	for (i = 0; i < conf->copies; i++) {
1523		if (r10_bio->devs[i].bio)
1524			raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1525		if (r10_bio->devs[i].repl_bio)
1526			raid10_write_one_disk(mddev, r10_bio, bio, true, i);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1527	}
1528	one_write_done(r10_bio);
1529}
1530
1531static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1532{
1533	struct r10conf *conf = mddev->private;
1534	struct r10bio *r10_bio;
1535
1536	r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 
 
 
 
 
1537
1538	r10_bio->master_bio = bio;
1539	r10_bio->sectors = sectors;
1540
1541	r10_bio->mddev = mddev;
1542	r10_bio->sector = bio->bi_iter.bi_sector;
1543	r10_bio->state = 0;
1544	memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1545
1546	if (bio_data_dir(bio) == READ)
1547		raid10_read_request(mddev, bio, r10_bio);
1548	else
1549		raid10_write_request(mddev, bio, r10_bio);
1550}
1551
1552static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1553{
1554	struct r10conf *conf = mddev->private;
1555	sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1556	int chunk_sects = chunk_mask + 1;
1557	int sectors = bio_sectors(bio);
1558
1559	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
 
 
1560		md_flush_request(mddev, bio);
1561		return true;
1562	}
1563
1564	if (!md_write_start(mddev, bio))
1565		return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1566
1567	/*
1568	 * If this request crosses a chunk boundary, we need to split
1569	 * it.
1570	 */
1571	if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1572		     sectors > chunk_sects
1573		     && (conf->geo.near_copies < conf->geo.raid_disks
1574			 || conf->prev.near_copies <
1575			 conf->prev.raid_disks)))
1576		sectors = chunk_sects -
1577			(bio->bi_iter.bi_sector &
1578			 (chunk_sects - 1));
1579	__make_request(mddev, bio, sectors);
1580
1581	/* In case raid10d snuck in to freeze_array */
1582	wake_up(&conf->wait_barrier);
1583	return true;
1584}
1585
1586static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1587{
1588	struct r10conf *conf = mddev->private;
1589	int i;
1590
1591	if (conf->geo.near_copies < conf->geo.raid_disks)
1592		seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1593	if (conf->geo.near_copies > 1)
1594		seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1595	if (conf->geo.far_copies > 1) {
1596		if (conf->geo.far_offset)
1597			seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1598		else
1599			seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1600		if (conf->geo.far_set_size != conf->geo.raid_disks)
1601			seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1602	}
1603	seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1604					conf->geo.raid_disks - mddev->degraded);
1605	rcu_read_lock();
1606	for (i = 0; i < conf->geo.raid_disks; i++) {
1607		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1608		seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1609	}
1610	rcu_read_unlock();
1611	seq_printf(seq, "]");
1612}
1613
1614/* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1618 */
1619static int _enough(struct r10conf *conf, int previous, int ignore)
1620{
1621	int first = 0;
1622	int has_enough = 0;
1623	int disks, ncopies;
1624	if (previous) {
1625		disks = conf->prev.raid_disks;
1626		ncopies = conf->prev.near_copies;
1627	} else {
1628		disks = conf->geo.raid_disks;
1629		ncopies = conf->geo.near_copies;
1630	}
1631
1632	rcu_read_lock();
1633	do {
1634		int n = conf->copies;
1635		int cnt = 0;
1636		int this = first;
1637		while (n--) {
1638			struct md_rdev *rdev;
1639			if (this != ignore &&
1640			    (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1641			    test_bit(In_sync, &rdev->flags))
1642				cnt++;
1643			this = (this+1) % disks;
1644		}
1645		if (cnt == 0)
1646			goto out;
1647		first = (first + ncopies) % disks;
1648	} while (first != 0);
1649	has_enough = 1;
1650out:
1651	rcu_read_unlock();
1652	return has_enough;
1653}
1654
1655static int enough(struct r10conf *conf, int ignore)
1656{
1657	/* when calling 'enough', both 'prev' and 'geo' must
1658	 * be stable.
1659	 * This is ensured if ->reconfig_mutex or ->device_lock
1660	 * is held.
1661	 */
1662	return _enough(conf, 0, ignore) &&
1663		_enough(conf, 1, ignore);
1664}
1665
1666static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1667{
1668	char b[BDEVNAME_SIZE];
1669	struct r10conf *conf = mddev->private;
1670	unsigned long flags;
1671
1672	/*
1673	 * If it is not operational, then we have already marked it as dead
1674	 * else if it is the last working disks, ignore the error, let the
1675	 * next level up know.
1676	 * else mark the drive as failed
1677	 */
1678	spin_lock_irqsave(&conf->device_lock, flags);
1679	if (test_bit(In_sync, &rdev->flags)
1680	    && !enough(conf, rdev->raid_disk)) {
1681		/*
1682		 * Don't fail the drive, just return an IO error.
1683		 */
1684		spin_unlock_irqrestore(&conf->device_lock, flags);
1685		return;
1686	}
1687	if (test_and_clear_bit(In_sync, &rdev->flags))
1688		mddev->degraded++;
1689	/*
1690	 * If recovery is running, make sure it aborts.
1691	 */
1692	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
 
1693	set_bit(Blocked, &rdev->flags);
1694	set_bit(Faulty, &rdev->flags);
1695	set_mask_bits(&mddev->sb_flags, 0,
1696		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1697	spin_unlock_irqrestore(&conf->device_lock, flags);
1698	pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1699		"md/raid10:%s: Operation continuing on %d devices.\n",
1700		mdname(mddev), bdevname(rdev->bdev, b),
1701		mdname(mddev), conf->geo.raid_disks - mddev->degraded);
 
1702}
1703
1704static void print_conf(struct r10conf *conf)
1705{
1706	int i;
1707	struct md_rdev *rdev;
1708
1709	pr_debug("RAID10 conf printout:\n");
1710	if (!conf) {
1711		pr_debug("(!conf)\n");
1712		return;
1713	}
1714	pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1715		 conf->geo.raid_disks);
1716
1717	/* This is only called with ->reconfix_mutex held, so
1718	 * rcu protection of rdev is not needed */
1719	for (i = 0; i < conf->geo.raid_disks; i++) {
1720		char b[BDEVNAME_SIZE];
1721		rdev = conf->mirrors[i].rdev;
1722		if (rdev)
1723			pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1724				 i, !test_bit(In_sync, &rdev->flags),
1725				 !test_bit(Faulty, &rdev->flags),
1726				 bdevname(rdev->bdev,b));
1727	}
1728}
1729
1730static void close_sync(struct r10conf *conf)
1731{
1732	wait_barrier(conf);
1733	allow_barrier(conf);
1734
1735	mempool_destroy(conf->r10buf_pool);
1736	conf->r10buf_pool = NULL;
1737}
1738
1739static int raid10_spare_active(struct mddev *mddev)
1740{
1741	int i;
1742	struct r10conf *conf = mddev->private;
1743	struct raid10_info *tmp;
1744	int count = 0;
1745	unsigned long flags;
1746
1747	/*
1748	 * Find all non-in_sync disks within the RAID10 configuration
1749	 * and mark them in_sync
1750	 */
1751	for (i = 0; i < conf->geo.raid_disks; i++) {
1752		tmp = conf->mirrors + i;
1753		if (tmp->replacement
1754		    && tmp->replacement->recovery_offset == MaxSector
1755		    && !test_bit(Faulty, &tmp->replacement->flags)
1756		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1757			/* Replacement has just become active */
1758			if (!tmp->rdev
1759			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1760				count++;
1761			if (tmp->rdev) {
1762				/* Replaced device not technically faulty,
1763				 * but we need to be sure it gets removed
1764				 * and never re-added.
1765				 */
1766				set_bit(Faulty, &tmp->rdev->flags);
1767				sysfs_notify_dirent_safe(
1768					tmp->rdev->sysfs_state);
1769			}
1770			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1771		} else if (tmp->rdev
1772			   && tmp->rdev->recovery_offset == MaxSector
1773			   && !test_bit(Faulty, &tmp->rdev->flags)
1774			   && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1775			count++;
1776			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1777		}
1778	}
1779	spin_lock_irqsave(&conf->device_lock, flags);
1780	mddev->degraded -= count;
1781	spin_unlock_irqrestore(&conf->device_lock, flags);
1782
1783	print_conf(conf);
1784	return count;
1785}
1786
 
1787static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1788{
1789	struct r10conf *conf = mddev->private;
1790	int err = -EEXIST;
1791	int mirror;
1792	int first = 0;
1793	int last = conf->geo.raid_disks - 1;
 
1794
1795	if (mddev->recovery_cp < MaxSector)
1796		/* only hot-add to in-sync arrays, as recovery is
1797		 * very different from resync
1798		 */
1799		return -EBUSY;
1800	if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1801		return -EINVAL;
1802
1803	if (md_integrity_add_rdev(rdev, mddev))
1804		return -ENXIO;
1805
1806	if (rdev->raid_disk >= 0)
1807		first = last = rdev->raid_disk;
1808
 
 
 
 
 
1809	if (rdev->saved_raid_disk >= first &&
1810	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1811		mirror = rdev->saved_raid_disk;
1812	else
1813		mirror = first;
1814	for ( ; mirror <= last ; mirror++) {
1815		struct raid10_info *p = &conf->mirrors[mirror];
1816		if (p->recovery_disabled == mddev->recovery_disabled)
1817			continue;
1818		if (p->rdev) {
1819			if (!test_bit(WantReplacement, &p->rdev->flags) ||
1820			    p->replacement != NULL)
1821				continue;
1822			clear_bit(In_sync, &rdev->flags);
1823			set_bit(Replacement, &rdev->flags);
1824			rdev->raid_disk = mirror;
1825			err = 0;
1826			if (mddev->gendisk)
1827				disk_stack_limits(mddev->gendisk, rdev->bdev,
1828						  rdev->data_offset << 9);
1829			conf->fullsync = 1;
1830			rcu_assign_pointer(p->replacement, rdev);
1831			break;
1832		}
1833
1834		if (mddev->gendisk)
1835			disk_stack_limits(mddev->gendisk, rdev->bdev,
1836					  rdev->data_offset << 9);
1837
1838		p->head_position = 0;
1839		p->recovery_disabled = mddev->recovery_disabled - 1;
1840		rdev->raid_disk = mirror;
1841		err = 0;
1842		if (rdev->saved_raid_disk != mirror)
1843			conf->fullsync = 1;
1844		rcu_assign_pointer(p->rdev, rdev);
1845		break;
1846	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1847	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1848		blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1849
1850	print_conf(conf);
1851	return err;
1852}
1853
1854static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1855{
1856	struct r10conf *conf = mddev->private;
1857	int err = 0;
1858	int number = rdev->raid_disk;
1859	struct md_rdev **rdevp;
1860	struct raid10_info *p = conf->mirrors + number;
1861
1862	print_conf(conf);
1863	if (rdev == p->rdev)
1864		rdevp = &p->rdev;
1865	else if (rdev == p->replacement)
1866		rdevp = &p->replacement;
1867	else
1868		return 0;
1869
1870	if (test_bit(In_sync, &rdev->flags) ||
1871	    atomic_read(&rdev->nr_pending)) {
1872		err = -EBUSY;
1873		goto abort;
1874	}
1875	/* Only remove non-faulty devices if recovery
1876	 * is not possible.
1877	 */
1878	if (!test_bit(Faulty, &rdev->flags) &&
1879	    mddev->recovery_disabled != p->recovery_disabled &&
1880	    (!p->replacement || p->replacement == rdev) &&
1881	    number < conf->geo.raid_disks &&
1882	    enough(conf, -1)) {
1883		err = -EBUSY;
1884		goto abort;
1885	}
1886	*rdevp = NULL;
1887	if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1888		synchronize_rcu();
1889		if (atomic_read(&rdev->nr_pending)) {
1890			/* lost the race, try later */
1891			err = -EBUSY;
1892			*rdevp = rdev;
1893			goto abort;
1894		}
1895	}
1896	if (p->replacement) {
1897		/* We must have just cleared 'rdev' */
1898		p->rdev = p->replacement;
1899		clear_bit(Replacement, &p->replacement->flags);
1900		smp_mb(); /* Make sure other CPUs may see both as identical
1901			   * but will never see neither -- if they are careful.
1902			   */
1903		p->replacement = NULL;
1904	}
 
 
 
 
 
1905
1906	clear_bit(WantReplacement, &rdev->flags);
1907	err = md_integrity_register(mddev);
1908
1909abort:
1910
1911	print_conf(conf);
1912	return err;
1913}
1914
1915static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
 
1916{
 
1917	struct r10conf *conf = r10_bio->mddev->private;
 
 
 
 
 
 
 
1918
1919	if (!bio->bi_status)
1920		set_bit(R10BIO_Uptodate, &r10_bio->state);
1921	else
1922		/* The write handler will notice the lack of
1923		 * R10BIO_Uptodate and record any errors etc
1924		 */
1925		atomic_add(r10_bio->sectors,
1926			   &conf->mirrors[d].rdev->corrected_errors);
1927
1928	/* for reconstruct, we always reschedule after a read.
1929	 * for resync, only after all reads
1930	 */
1931	rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1932	if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1933	    atomic_dec_and_test(&r10_bio->remaining)) {
1934		/* we have read all the blocks,
1935		 * do the comparison in process context in raid10d
1936		 */
1937		reschedule_retry(r10_bio);
1938	}
1939}
1940
1941static void end_sync_read(struct bio *bio)
1942{
1943	struct r10bio *r10_bio = get_resync_r10bio(bio);
1944	struct r10conf *conf = r10_bio->mddev->private;
1945	int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1946
1947	__end_sync_read(r10_bio, bio, d);
1948}
1949
1950static void end_reshape_read(struct bio *bio)
1951{
1952	/* reshape read bio isn't allocated from r10buf_pool */
1953	struct r10bio *r10_bio = bio->bi_private;
1954
1955	__end_sync_read(r10_bio, bio, r10_bio->read_slot);
1956}
1957
1958static void end_sync_request(struct r10bio *r10_bio)
1959{
1960	struct mddev *mddev = r10_bio->mddev;
1961
1962	while (atomic_dec_and_test(&r10_bio->remaining)) {
1963		if (r10_bio->master_bio == NULL) {
1964			/* the primary of several recovery bios */
1965			sector_t s = r10_bio->sectors;
1966			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1967			    test_bit(R10BIO_WriteError, &r10_bio->state))
1968				reschedule_retry(r10_bio);
1969			else
1970				put_buf(r10_bio);
1971			md_done_sync(mddev, s, 1);
1972			break;
1973		} else {
1974			struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1975			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1976			    test_bit(R10BIO_WriteError, &r10_bio->state))
1977				reschedule_retry(r10_bio);
1978			else
1979				put_buf(r10_bio);
1980			r10_bio = r10_bio2;
1981		}
1982	}
1983}
1984
1985static void end_sync_write(struct bio *bio)
1986{
1987	struct r10bio *r10_bio = get_resync_r10bio(bio);
 
1988	struct mddev *mddev = r10_bio->mddev;
1989	struct r10conf *conf = mddev->private;
1990	int d;
1991	sector_t first_bad;
1992	int bad_sectors;
1993	int slot;
1994	int repl;
1995	struct md_rdev *rdev = NULL;
1996
1997	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1998	if (repl)
1999		rdev = conf->mirrors[d].replacement;
2000	else
2001		rdev = conf->mirrors[d].rdev;
2002
2003	if (bio->bi_status) {
2004		if (repl)
2005			md_error(mddev, rdev);
2006		else {
2007			set_bit(WriteErrorSeen, &rdev->flags);
2008			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2009				set_bit(MD_RECOVERY_NEEDED,
2010					&rdev->mddev->recovery);
2011			set_bit(R10BIO_WriteError, &r10_bio->state);
2012		}
2013	} else if (is_badblock(rdev,
2014			     r10_bio->devs[slot].addr,
2015			     r10_bio->sectors,
2016			     &first_bad, &bad_sectors))
2017		set_bit(R10BIO_MadeGood, &r10_bio->state);
2018
2019	rdev_dec_pending(rdev, mddev);
2020
2021	end_sync_request(r10_bio);
2022}
2023
2024/*
2025 * Note: sync and recover and handled very differently for raid10
2026 * This code is for resync.
2027 * For resync, we read through virtual addresses and read all blocks.
2028 * If there is any error, we schedule a write.  The lowest numbered
2029 * drive is authoritative.
2030 * However requests come for physical address, so we need to map.
2031 * For every physical address there are raid_disks/copies virtual addresses,
2032 * which is always are least one, but is not necessarly an integer.
2033 * This means that a physical address can span multiple chunks, so we may
2034 * have to submit multiple io requests for a single sync request.
2035 */
2036/*
2037 * We check if all blocks are in-sync and only write to blocks that
2038 * aren't in sync
2039 */
2040static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2041{
2042	struct r10conf *conf = mddev->private;
2043	int i, first;
2044	struct bio *tbio, *fbio;
2045	int vcnt;
2046	struct page **tpages, **fpages;
2047
2048	atomic_set(&r10_bio->remaining, 1);
2049
2050	/* find the first device with a block */
2051	for (i=0; i<conf->copies; i++)
2052		if (!r10_bio->devs[i].bio->bi_status)
2053			break;
2054
2055	if (i == conf->copies)
2056		goto done;
2057
2058	first = i;
2059	fbio = r10_bio->devs[i].bio;
2060	fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2061	fbio->bi_iter.bi_idx = 0;
2062	fpages = get_resync_pages(fbio)->pages;
2063
2064	vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2065	/* now find blocks with errors */
2066	for (i=0 ; i < conf->copies ; i++) {
2067		int  j, d;
2068		struct md_rdev *rdev;
2069		struct resync_pages *rp;
2070
2071		tbio = r10_bio->devs[i].bio;
2072
2073		if (tbio->bi_end_io != end_sync_read)
2074			continue;
2075		if (i == first)
2076			continue;
2077
2078		tpages = get_resync_pages(tbio)->pages;
2079		d = r10_bio->devs[i].devnum;
2080		rdev = conf->mirrors[d].rdev;
2081		if (!r10_bio->devs[i].bio->bi_status) {
2082			/* We know that the bi_io_vec layout is the same for
2083			 * both 'first' and 'i', so we just compare them.
2084			 * All vec entries are PAGE_SIZE;
2085			 */
2086			int sectors = r10_bio->sectors;
2087			for (j = 0; j < vcnt; j++) {
2088				int len = PAGE_SIZE;
2089				if (sectors < (len / 512))
2090					len = sectors * 512;
2091				if (memcmp(page_address(fpages[j]),
2092					   page_address(tpages[j]),
2093					   len))
2094					break;
2095				sectors -= len/512;
2096			}
2097			if (j == vcnt)
2098				continue;
2099			atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2100			if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2101				/* Don't fix anything. */
2102				continue;
2103		} else if (test_bit(FailFast, &rdev->flags)) {
2104			/* Just give up on this device */
2105			md_error(rdev->mddev, rdev);
2106			continue;
2107		}
2108		/* Ok, we need to write this bio, either to correct an
2109		 * inconsistency or to correct an unreadable block.
2110		 * First we need to fixup bv_offset, bv_len and
2111		 * bi_vecs, as the read request might have corrupted these
2112		 */
2113		rp = get_resync_pages(tbio);
2114		bio_reset(tbio);
2115
2116		md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
 
 
 
 
2117
2118		rp->raid_bio = r10_bio;
2119		tbio->bi_private = rp;
2120		tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
 
 
 
 
 
2121		tbio->bi_end_io = end_sync_write;
2122		bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2123
2124		bio_copy_data(tbio, fbio);
2125
 
2126		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2127		atomic_inc(&r10_bio->remaining);
2128		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2129
2130		if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2131			tbio->bi_opf |= MD_FAILFAST;
2132		tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2133		bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2134		generic_make_request(tbio);
2135	}
2136
2137	/* Now write out to any replacement devices
2138	 * that are active
2139	 */
2140	for (i = 0; i < conf->copies; i++) {
2141		int d;
2142
2143		tbio = r10_bio->devs[i].repl_bio;
2144		if (!tbio || !tbio->bi_end_io)
2145			continue;
2146		if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2147		    && r10_bio->devs[i].bio != fbio)
2148			bio_copy_data(tbio, fbio);
 
 
 
2149		d = r10_bio->devs[i].devnum;
2150		atomic_inc(&r10_bio->remaining);
2151		md_sync_acct(conf->mirrors[d].replacement->bdev,
2152			     bio_sectors(tbio));
2153		generic_make_request(tbio);
2154	}
2155
2156done:
2157	if (atomic_dec_and_test(&r10_bio->remaining)) {
2158		md_done_sync(mddev, r10_bio->sectors, 1);
2159		put_buf(r10_bio);
2160	}
2161}
2162
2163/*
2164 * Now for the recovery code.
2165 * Recovery happens across physical sectors.
2166 * We recover all non-is_sync drives by finding the virtual address of
2167 * each, and then choose a working drive that also has that virt address.
2168 * There is a separate r10_bio for each non-in_sync drive.
2169 * Only the first two slots are in use. The first for reading,
2170 * The second for writing.
2171 *
2172 */
2173static void fix_recovery_read_error(struct r10bio *r10_bio)
2174{
2175	/* We got a read error during recovery.
2176	 * We repeat the read in smaller page-sized sections.
2177	 * If a read succeeds, write it to the new device or record
2178	 * a bad block if we cannot.
2179	 * If a read fails, record a bad block on both old and
2180	 * new devices.
2181	 */
2182	struct mddev *mddev = r10_bio->mddev;
2183	struct r10conf *conf = mddev->private;
2184	struct bio *bio = r10_bio->devs[0].bio;
2185	sector_t sect = 0;
2186	int sectors = r10_bio->sectors;
2187	int idx = 0;
2188	int dr = r10_bio->devs[0].devnum;
2189	int dw = r10_bio->devs[1].devnum;
2190	struct page **pages = get_resync_pages(bio)->pages;
2191
2192	while (sectors) {
2193		int s = sectors;
2194		struct md_rdev *rdev;
2195		sector_t addr;
2196		int ok;
2197
2198		if (s > (PAGE_SIZE>>9))
2199			s = PAGE_SIZE >> 9;
2200
2201		rdev = conf->mirrors[dr].rdev;
2202		addr = r10_bio->devs[0].addr + sect,
2203		ok = sync_page_io(rdev,
2204				  addr,
2205				  s << 9,
2206				  pages[idx],
2207				  REQ_OP_READ, 0, false);
2208		if (ok) {
2209			rdev = conf->mirrors[dw].rdev;
2210			addr = r10_bio->devs[1].addr + sect;
2211			ok = sync_page_io(rdev,
2212					  addr,
2213					  s << 9,
2214					  pages[idx],
2215					  REQ_OP_WRITE, 0, false);
2216			if (!ok) {
2217				set_bit(WriteErrorSeen, &rdev->flags);
2218				if (!test_and_set_bit(WantReplacement,
2219						      &rdev->flags))
2220					set_bit(MD_RECOVERY_NEEDED,
2221						&rdev->mddev->recovery);
2222			}
2223		}
2224		if (!ok) {
2225			/* We don't worry if we cannot set a bad block -
2226			 * it really is bad so there is no loss in not
2227			 * recording it yet
2228			 */
2229			rdev_set_badblocks(rdev, addr, s, 0);
2230
2231			if (rdev != conf->mirrors[dw].rdev) {
2232				/* need bad block on destination too */
2233				struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2234				addr = r10_bio->devs[1].addr + sect;
2235				ok = rdev_set_badblocks(rdev2, addr, s, 0);
2236				if (!ok) {
2237					/* just abort the recovery */
2238					pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2239						  mdname(mddev));
 
 
2240
2241					conf->mirrors[dw].recovery_disabled
2242						= mddev->recovery_disabled;
2243					set_bit(MD_RECOVERY_INTR,
2244						&mddev->recovery);
2245					break;
2246				}
2247			}
2248		}
2249
2250		sectors -= s;
2251		sect += s;
2252		idx++;
2253	}
2254}
2255
2256static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2257{
2258	struct r10conf *conf = mddev->private;
2259	int d;
2260	struct bio *wbio, *wbio2;
2261
2262	if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2263		fix_recovery_read_error(r10_bio);
2264		end_sync_request(r10_bio);
2265		return;
2266	}
2267
2268	/*
2269	 * share the pages with the first bio
2270	 * and submit the write request
2271	 */
2272	d = r10_bio->devs[1].devnum;
2273	wbio = r10_bio->devs[1].bio;
2274	wbio2 = r10_bio->devs[1].repl_bio;
2275	/* Need to test wbio2->bi_end_io before we call
2276	 * generic_make_request as if the former is NULL,
2277	 * the latter is free to free wbio2.
2278	 */
2279	if (wbio2 && !wbio2->bi_end_io)
2280		wbio2 = NULL;
2281	if (wbio->bi_end_io) {
2282		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2283		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2284		generic_make_request(wbio);
2285	}
2286	if (wbio2) {
2287		atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2288		md_sync_acct(conf->mirrors[d].replacement->bdev,
2289			     bio_sectors(wbio2));
2290		generic_make_request(wbio2);
2291	}
2292}
2293
 
2294/*
2295 * Used by fix_read_error() to decay the per rdev read_errors.
2296 * We halve the read error count for every hour that has elapsed
2297 * since the last recorded read error.
2298 *
2299 */
2300static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2301{
2302	long cur_time_mon;
2303	unsigned long hours_since_last;
2304	unsigned int read_errors = atomic_read(&rdev->read_errors);
2305
2306	cur_time_mon = ktime_get_seconds();
2307
2308	if (rdev->last_read_error == 0) {
 
2309		/* first time we've seen a read error */
2310		rdev->last_read_error = cur_time_mon;
2311		return;
2312	}
2313
2314	hours_since_last = (long)(cur_time_mon -
2315			    rdev->last_read_error) / 3600;
2316
2317	rdev->last_read_error = cur_time_mon;
2318
2319	/*
2320	 * if hours_since_last is > the number of bits in read_errors
2321	 * just set read errors to 0. We do this to avoid
2322	 * overflowing the shift of read_errors by hours_since_last.
2323	 */
2324	if (hours_since_last >= 8 * sizeof(read_errors))
2325		atomic_set(&rdev->read_errors, 0);
2326	else
2327		atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2328}
2329
2330static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2331			    int sectors, struct page *page, int rw)
2332{
2333	sector_t first_bad;
2334	int bad_sectors;
2335
2336	if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2337	    && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2338		return -1;
2339	if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2340		/* success */
2341		return 1;
2342	if (rw == WRITE) {
2343		set_bit(WriteErrorSeen, &rdev->flags);
2344		if (!test_and_set_bit(WantReplacement, &rdev->flags))
2345			set_bit(MD_RECOVERY_NEEDED,
2346				&rdev->mddev->recovery);
2347	}
2348	/* need to record an error - either for the block or the device */
2349	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2350		md_error(rdev->mddev, rdev);
2351	return 0;
2352}
2353
2354/*
2355 * This is a kernel thread which:
2356 *
2357 *	1.	Retries failed read operations on working mirrors.
2358 *	2.	Updates the raid superblock when problems encounter.
2359 *	3.	Performs writes following reads for array synchronising.
2360 */
2361
2362static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2363{
2364	int sect = 0; /* Offset from r10_bio->sector */
2365	int sectors = r10_bio->sectors;
2366	struct md_rdev*rdev;
2367	int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2368	int d = r10_bio->devs[r10_bio->read_slot].devnum;
2369
2370	/* still own a reference to this rdev, so it cannot
2371	 * have been cleared recently.
2372	 */
2373	rdev = conf->mirrors[d].rdev;
2374
2375	if (test_bit(Faulty, &rdev->flags))
2376		/* drive has already been failed, just ignore any
2377		   more fix_read_error() attempts */
2378		return;
2379
2380	check_decay_read_errors(mddev, rdev);
2381	atomic_inc(&rdev->read_errors);
2382	if (atomic_read(&rdev->read_errors) > max_read_errors) {
2383		char b[BDEVNAME_SIZE];
2384		bdevname(rdev->bdev, b);
2385
2386		pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2387			  mdname(mddev), b,
2388			  atomic_read(&rdev->read_errors), max_read_errors);
2389		pr_notice("md/raid10:%s: %s: Failing raid device\n",
2390			  mdname(mddev), b);
2391		md_error(mddev, rdev);
 
 
 
2392		r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2393		return;
2394	}
2395
2396	while(sectors) {
2397		int s = sectors;
2398		int sl = r10_bio->read_slot;
2399		int success = 0;
2400		int start;
2401
2402		if (s > (PAGE_SIZE>>9))
2403			s = PAGE_SIZE >> 9;
2404
2405		rcu_read_lock();
2406		do {
2407			sector_t first_bad;
2408			int bad_sectors;
2409
2410			d = r10_bio->devs[sl].devnum;
2411			rdev = rcu_dereference(conf->mirrors[d].rdev);
2412			if (rdev &&
 
2413			    test_bit(In_sync, &rdev->flags) &&
2414			    !test_bit(Faulty, &rdev->flags) &&
2415			    is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2416					&first_bad, &bad_sectors) == 0) {
2417				atomic_inc(&rdev->nr_pending);
2418				rcu_read_unlock();
2419				success = sync_page_io(rdev,
2420						       r10_bio->devs[sl].addr +
2421						       sect,
2422						       s<<9,
2423						       conf->tmppage,
2424						       REQ_OP_READ, 0, false);
2425				rdev_dec_pending(rdev, mddev);
2426				rcu_read_lock();
2427				if (success)
2428					break;
2429			}
2430			sl++;
2431			if (sl == conf->copies)
2432				sl = 0;
2433		} while (!success && sl != r10_bio->read_slot);
2434		rcu_read_unlock();
2435
2436		if (!success) {
2437			/* Cannot read from anywhere, just mark the block
2438			 * as bad on the first device to discourage future
2439			 * reads.
2440			 */
2441			int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2442			rdev = conf->mirrors[dn].rdev;
2443
2444			if (!rdev_set_badblocks(
2445				    rdev,
2446				    r10_bio->devs[r10_bio->read_slot].addr
2447				    + sect,
2448				    s, 0)) {
2449				md_error(mddev, rdev);
2450				r10_bio->devs[r10_bio->read_slot].bio
2451					= IO_BLOCKED;
2452			}
2453			break;
2454		}
2455
2456		start = sl;
2457		/* write it back and re-read */
2458		rcu_read_lock();
2459		while (sl != r10_bio->read_slot) {
2460			char b[BDEVNAME_SIZE];
2461
2462			if (sl==0)
2463				sl = conf->copies;
2464			sl--;
2465			d = r10_bio->devs[sl].devnum;
2466			rdev = rcu_dereference(conf->mirrors[d].rdev);
2467			if (!rdev ||
2468			    test_bit(Faulty, &rdev->flags) ||
2469			    !test_bit(In_sync, &rdev->flags))
2470				continue;
2471
2472			atomic_inc(&rdev->nr_pending);
2473			rcu_read_unlock();
2474			if (r10_sync_page_io(rdev,
2475					     r10_bio->devs[sl].addr +
2476					     sect,
2477					     s, conf->tmppage, WRITE)
2478			    == 0) {
2479				/* Well, this device is dead */
2480				pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2481					  mdname(mddev), s,
2482					  (unsigned long long)(
2483						  sect +
2484						  choose_data_offset(r10_bio,
2485								     rdev)),
2486					  bdevname(rdev->bdev, b));
2487				pr_notice("md/raid10:%s: %s: failing drive\n",
2488					  mdname(mddev),
2489					  bdevname(rdev->bdev, b));
 
 
 
 
2490			}
2491			rdev_dec_pending(rdev, mddev);
2492			rcu_read_lock();
2493		}
2494		sl = start;
2495		while (sl != r10_bio->read_slot) {
2496			char b[BDEVNAME_SIZE];
2497
2498			if (sl==0)
2499				sl = conf->copies;
2500			sl--;
2501			d = r10_bio->devs[sl].devnum;
2502			rdev = rcu_dereference(conf->mirrors[d].rdev);
2503			if (!rdev ||
2504			    test_bit(Faulty, &rdev->flags) ||
2505			    !test_bit(In_sync, &rdev->flags))
2506				continue;
2507
2508			atomic_inc(&rdev->nr_pending);
2509			rcu_read_unlock();
2510			switch (r10_sync_page_io(rdev,
2511					     r10_bio->devs[sl].addr +
2512					     sect,
2513					     s, conf->tmppage,
2514						 READ)) {
2515			case 0:
2516				/* Well, this device is dead */
2517				pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
 
 
 
2518				       mdname(mddev), s,
2519				       (unsigned long long)(
2520					       sect +
2521					       choose_data_offset(r10_bio, rdev)),
2522				       bdevname(rdev->bdev, b));
2523				pr_notice("md/raid10:%s: %s: failing drive\n",
 
2524				       mdname(mddev),
2525				       bdevname(rdev->bdev, b));
2526				break;
2527			case 1:
2528				pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
 
 
2529				       mdname(mddev), s,
2530				       (unsigned long long)(
2531					       sect +
2532					       choose_data_offset(r10_bio, rdev)),
2533				       bdevname(rdev->bdev, b));
2534				atomic_add(s, &rdev->corrected_errors);
2535			}
2536
2537			rdev_dec_pending(rdev, mddev);
2538			rcu_read_lock();
2539		}
2540		rcu_read_unlock();
2541
2542		sectors -= s;
2543		sect += s;
2544	}
2545}
2546
2547static int narrow_write_error(struct r10bio *r10_bio, int i)
2548{
2549	struct bio *bio = r10_bio->master_bio;
2550	struct mddev *mddev = r10_bio->mddev;
2551	struct r10conf *conf = mddev->private;
2552	struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2553	/* bio has the data to be written to slot 'i' where
2554	 * we just recently had a write error.
2555	 * We repeatedly clone the bio and trim down to one block,
2556	 * then try the write.  Where the write fails we record
2557	 * a bad block.
2558	 * It is conceivable that the bio doesn't exactly align with
2559	 * blocks.  We must handle this.
2560	 *
2561	 * We currently own a reference to the rdev.
2562	 */
2563
2564	int block_sectors;
2565	sector_t sector;
2566	int sectors;
2567	int sect_to_write = r10_bio->sectors;
2568	int ok = 1;
2569
2570	if (rdev->badblocks.shift < 0)
2571		return 0;
2572
2573	block_sectors = roundup(1 << rdev->badblocks.shift,
2574				bdev_logical_block_size(rdev->bdev) >> 9);
2575	sector = r10_bio->sector;
2576	sectors = ((r10_bio->sector + block_sectors)
2577		   & ~(sector_t)(block_sectors - 1))
2578		- sector;
2579
2580	while (sect_to_write) {
2581		struct bio *wbio;
2582		sector_t wsector;
2583		if (sectors > sect_to_write)
2584			sectors = sect_to_write;
2585		/* Write at 'sector' for 'sectors' */
2586		wbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
2587		bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2588		wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2589		wbio->bi_iter.bi_sector = wsector +
2590				   choose_data_offset(r10_bio, rdev);
2591		bio_set_dev(wbio, rdev->bdev);
2592		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2593
2594		if (submit_bio_wait(wbio) < 0)
2595			/* Failure! */
2596			ok = rdev_set_badblocks(rdev, wsector,
2597						sectors, 0)
2598				&& ok;
2599
2600		bio_put(wbio);
2601		sect_to_write -= sectors;
2602		sector += sectors;
2603		sectors = block_sectors;
2604	}
2605	return ok;
2606}
2607
2608static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2609{
2610	int slot = r10_bio->read_slot;
2611	struct bio *bio;
2612	struct r10conf *conf = mddev->private;
2613	struct md_rdev *rdev = r10_bio->devs[slot].rdev;
 
 
 
2614
2615	/* we got a read error. Maybe the drive is bad.  Maybe just
2616	 * the block and we can fix it.
2617	 * We freeze all other IO, and try reading the block from
2618	 * other devices.  When we find one, we re-write
2619	 * and check it that fixes the read error.
2620	 * This is all done synchronously while the array is
2621	 * frozen.
2622	 */
2623	bio = r10_bio->devs[slot].bio;
 
2624	bio_put(bio);
2625	r10_bio->devs[slot].bio = NULL;
2626
2627	if (mddev->ro)
2628		r10_bio->devs[slot].bio = IO_BLOCKED;
2629	else if (!test_bit(FailFast, &rdev->flags)) {
2630		freeze_array(conf, 1);
2631		fix_read_error(conf, mddev, r10_bio);
2632		unfreeze_array(conf);
2633	} else
2634		md_error(mddev, rdev);
2635
2636	rdev_dec_pending(rdev, mddev);
2637	allow_barrier(conf);
2638	r10_bio->state = 0;
2639	raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2640}
2641
2642static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2643{
2644	/* Some sort of write request has finished and it
2645	 * succeeded in writing where we thought there was a
2646	 * bad block.  So forget the bad block.
2647	 * Or possibly if failed and we need to record
2648	 * a bad block.
2649	 */
2650	int m;
2651	struct md_rdev *rdev;
2652
2653	if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2654	    test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2655		for (m = 0; m < conf->copies; m++) {
2656			int dev = r10_bio->devs[m].devnum;
2657			rdev = conf->mirrors[dev].rdev;
2658			if (r10_bio->devs[m].bio == NULL ||
2659				r10_bio->devs[m].bio->bi_end_io == NULL)
2660				continue;
2661			if (!r10_bio->devs[m].bio->bi_status) {
 
2662				rdev_clear_badblocks(
2663					rdev,
2664					r10_bio->devs[m].addr,
2665					r10_bio->sectors, 0);
2666			} else {
2667				if (!rdev_set_badblocks(
2668					    rdev,
2669					    r10_bio->devs[m].addr,
2670					    r10_bio->sectors, 0))
2671					md_error(conf->mddev, rdev);
2672			}
2673			rdev = conf->mirrors[dev].replacement;
2674			if (r10_bio->devs[m].repl_bio == NULL ||
2675				r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2676				continue;
2677
2678			if (!r10_bio->devs[m].repl_bio->bi_status) {
2679				rdev_clear_badblocks(
2680					rdev,
2681					r10_bio->devs[m].addr,
2682					r10_bio->sectors, 0);
2683			} else {
2684				if (!rdev_set_badblocks(
2685					    rdev,
2686					    r10_bio->devs[m].addr,
2687					    r10_bio->sectors, 0))
2688					md_error(conf->mddev, rdev);
2689			}
2690		}
2691		put_buf(r10_bio);
2692	} else {
2693		bool fail = false;
2694		for (m = 0; m < conf->copies; m++) {
2695			int dev = r10_bio->devs[m].devnum;
2696			struct bio *bio = r10_bio->devs[m].bio;
2697			rdev = conf->mirrors[dev].rdev;
2698			if (bio == IO_MADE_GOOD) {
2699				rdev_clear_badblocks(
2700					rdev,
2701					r10_bio->devs[m].addr,
2702					r10_bio->sectors, 0);
2703				rdev_dec_pending(rdev, conf->mddev);
2704			} else if (bio != NULL && bio->bi_status) {
2705				fail = true;
2706				if (!narrow_write_error(r10_bio, m)) {
2707					md_error(conf->mddev, rdev);
2708					set_bit(R10BIO_Degraded,
2709						&r10_bio->state);
2710				}
2711				rdev_dec_pending(rdev, conf->mddev);
2712			}
2713			bio = r10_bio->devs[m].repl_bio;
2714			rdev = conf->mirrors[dev].replacement;
2715			if (rdev && bio == IO_MADE_GOOD) {
2716				rdev_clear_badblocks(
2717					rdev,
2718					r10_bio->devs[m].addr,
2719					r10_bio->sectors, 0);
2720				rdev_dec_pending(rdev, conf->mddev);
2721			}
2722		}
2723		if (fail) {
2724			spin_lock_irq(&conf->device_lock);
2725			list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2726			conf->nr_queued++;
2727			spin_unlock_irq(&conf->device_lock);
2728			/*
2729			 * In case freeze_array() is waiting for condition
2730			 * nr_pending == nr_queued + extra to be true.
2731			 */
2732			wake_up(&conf->wait_barrier);
2733			md_wakeup_thread(conf->mddev->thread);
2734		} else {
2735			if (test_bit(R10BIO_WriteError,
2736				     &r10_bio->state))
2737				close_write(r10_bio);
2738			raid_end_bio_io(r10_bio);
2739		}
2740	}
2741}
2742
2743static void raid10d(struct md_thread *thread)
2744{
2745	struct mddev *mddev = thread->mddev;
2746	struct r10bio *r10_bio;
2747	unsigned long flags;
2748	struct r10conf *conf = mddev->private;
2749	struct list_head *head = &conf->retry_list;
2750	struct blk_plug plug;
2751
2752	md_check_recovery(mddev);
2753
2754	if (!list_empty_careful(&conf->bio_end_io_list) &&
2755	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2756		LIST_HEAD(tmp);
2757		spin_lock_irqsave(&conf->device_lock, flags);
2758		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2759			while (!list_empty(&conf->bio_end_io_list)) {
2760				list_move(conf->bio_end_io_list.prev, &tmp);
2761				conf->nr_queued--;
2762			}
2763		}
2764		spin_unlock_irqrestore(&conf->device_lock, flags);
2765		while (!list_empty(&tmp)) {
2766			r10_bio = list_first_entry(&tmp, struct r10bio,
2767						   retry_list);
2768			list_del(&r10_bio->retry_list);
2769			if (mddev->degraded)
2770				set_bit(R10BIO_Degraded, &r10_bio->state);
2771
2772			if (test_bit(R10BIO_WriteError,
2773				     &r10_bio->state))
2774				close_write(r10_bio);
2775			raid_end_bio_io(r10_bio);
2776		}
2777	}
2778
2779	blk_start_plug(&plug);
2780	for (;;) {
2781
2782		flush_pending_writes(conf);
2783
2784		spin_lock_irqsave(&conf->device_lock, flags);
2785		if (list_empty(head)) {
2786			spin_unlock_irqrestore(&conf->device_lock, flags);
2787			break;
2788		}
2789		r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2790		list_del(head->prev);
2791		conf->nr_queued--;
2792		spin_unlock_irqrestore(&conf->device_lock, flags);
2793
2794		mddev = r10_bio->mddev;
2795		conf = mddev->private;
2796		if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2797		    test_bit(R10BIO_WriteError, &r10_bio->state))
2798			handle_write_completed(conf, r10_bio);
2799		else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2800			reshape_request_write(mddev, r10_bio);
2801		else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2802			sync_request_write(mddev, r10_bio);
2803		else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2804			recovery_request_write(mddev, r10_bio);
2805		else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2806			handle_read_error(mddev, r10_bio);
2807		else
2808			WARN_ON_ONCE(1);
 
 
 
 
 
2809
2810		cond_resched();
2811		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2812			md_check_recovery(mddev);
2813	}
2814	blk_finish_plug(&plug);
2815}
2816
 
2817static int init_resync(struct r10conf *conf)
2818{
2819	int buffs;
2820	int i;
2821
2822	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2823	BUG_ON(conf->r10buf_pool);
2824	conf->have_replacement = 0;
2825	for (i = 0; i < conf->geo.raid_disks; i++)
2826		if (conf->mirrors[i].replacement)
2827			conf->have_replacement = 1;
2828	conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2829	if (!conf->r10buf_pool)
2830		return -ENOMEM;
2831	conf->next_resync = 0;
2832	return 0;
2833}
2834
2835static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2836{
2837	struct r10bio *r10bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2838	struct rsync_pages *rp;
2839	struct bio *bio;
2840	int nalloc;
2841	int i;
2842
2843	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2844	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2845		nalloc = conf->copies; /* resync */
2846	else
2847		nalloc = 2; /* recovery */
2848
2849	for (i = 0; i < nalloc; i++) {
2850		bio = r10bio->devs[i].bio;
2851		rp = bio->bi_private;
2852		bio_reset(bio);
2853		bio->bi_private = rp;
2854		bio = r10bio->devs[i].repl_bio;
2855		if (bio) {
2856			rp = bio->bi_private;
2857			bio_reset(bio);
2858			bio->bi_private = rp;
2859		}
2860	}
2861	return r10bio;
2862}
2863
2864/*
2865 * Set cluster_sync_high since we need other nodes to add the
2866 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2867 */
2868static void raid10_set_cluster_sync_high(struct r10conf *conf)
2869{
2870	sector_t window_size;
2871	int extra_chunk, chunks;
2872
2873	/*
2874	 * First, here we define "stripe" as a unit which across
2875	 * all member devices one time, so we get chunks by use
2876	 * raid_disks / near_copies. Otherwise, if near_copies is
2877	 * close to raid_disks, then resync window could increases
2878	 * linearly with the increase of raid_disks, which means
2879	 * we will suspend a really large IO window while it is not
2880	 * necessary. If raid_disks is not divisible by near_copies,
2881	 * an extra chunk is needed to ensure the whole "stripe" is
2882	 * covered.
2883	 */
2884
2885	chunks = conf->geo.raid_disks / conf->geo.near_copies;
2886	if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2887		extra_chunk = 0;
2888	else
2889		extra_chunk = 1;
2890	window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2891
2892	/*
2893	 * At least use a 32M window to align with raid1's resync window
2894	 */
2895	window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2896			CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2897
2898	conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2899}
2900
2901/*
2902 * perform a "sync" on one "block"
2903 *
2904 * We need to make sure that no normal I/O request - particularly write
2905 * requests - conflict with active sync requests.
2906 *
2907 * This is achieved by tracking pending requests and a 'barrier' concept
2908 * that can be installed to exclude normal IO requests.
2909 *
2910 * Resync and recovery are handled very differently.
2911 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2912 *
2913 * For resync, we iterate over virtual addresses, read all copies,
2914 * and update if there are differences.  If only one copy is live,
2915 * skip it.
2916 * For recovery, we iterate over physical addresses, read a good
2917 * value for each non-in_sync drive, and over-write.
2918 *
2919 * So, for recovery we may have several outstanding complex requests for a
2920 * given address, one for each out-of-sync device.  We model this by allocating
2921 * a number of r10_bio structures, one for each out-of-sync device.
2922 * As we setup these structures, we collect all bio's together into a list
2923 * which we then process collectively to add pages, and then process again
2924 * to pass to generic_make_request.
2925 *
2926 * The r10_bio structures are linked using a borrowed master_bio pointer.
2927 * This link is counted in ->remaining.  When the r10_bio that points to NULL
2928 * has its remaining count decremented to 0, the whole complex operation
2929 * is complete.
2930 *
2931 */
2932
2933static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2934			     int *skipped)
2935{
2936	struct r10conf *conf = mddev->private;
2937	struct r10bio *r10_bio;
2938	struct bio *biolist = NULL, *bio;
2939	sector_t max_sector, nr_sectors;
2940	int i;
2941	int max_sync;
2942	sector_t sync_blocks;
2943	sector_t sectors_skipped = 0;
2944	int chunks_skipped = 0;
2945	sector_t chunk_mask = conf->geo.chunk_mask;
2946	int page_idx = 0;
2947
2948	if (!conf->r10buf_pool)
2949		if (init_resync(conf))
2950			return 0;
2951
2952	/*
2953	 * Allow skipping a full rebuild for incremental assembly
2954	 * of a clean array, like RAID1 does.
2955	 */
2956	if (mddev->bitmap == NULL &&
2957	    mddev->recovery_cp == MaxSector &&
2958	    mddev->reshape_position == MaxSector &&
2959	    !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2960	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2961	    !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2962	    conf->fullsync == 0) {
2963		*skipped = 1;
2964		return mddev->dev_sectors - sector_nr;
2965	}
2966
2967 skipped:
2968	max_sector = mddev->dev_sectors;
2969	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2970	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2971		max_sector = mddev->resync_max_sectors;
2972	if (sector_nr >= max_sector) {
2973		conf->cluster_sync_low = 0;
2974		conf->cluster_sync_high = 0;
2975
2976		/* If we aborted, we need to abort the
2977		 * sync on the 'current' bitmap chucks (there can
2978		 * be several when recovering multiple devices).
2979		 * as we may have started syncing it but not finished.
2980		 * We can find the current address in
2981		 * mddev->curr_resync, but for recovery,
2982		 * we need to convert that to several
2983		 * virtual addresses.
2984		 */
2985		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2986			end_reshape(conf);
2987			close_sync(conf);
2988			return 0;
2989		}
2990
2991		if (mddev->curr_resync < max_sector) { /* aborted */
2992			if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2993				bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2994						&sync_blocks, 1);
2995			else for (i = 0; i < conf->geo.raid_disks; i++) {
2996				sector_t sect =
2997					raid10_find_virt(conf, mddev->curr_resync, i);
2998				bitmap_end_sync(mddev->bitmap, sect,
2999						&sync_blocks, 1);
3000			}
3001		} else {
3002			/* completed sync */
3003			if ((!mddev->bitmap || conf->fullsync)
3004			    && conf->have_replacement
3005			    && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3006				/* Completed a full sync so the replacements
3007				 * are now fully recovered.
3008				 */
3009				rcu_read_lock();
3010				for (i = 0; i < conf->geo.raid_disks; i++) {
3011					struct md_rdev *rdev =
3012						rcu_dereference(conf->mirrors[i].replacement);
3013					if (rdev)
3014						rdev->recovery_offset = MaxSector;
3015				}
3016				rcu_read_unlock();
3017			}
3018			conf->fullsync = 0;
3019		}
3020		bitmap_close_sync(mddev->bitmap);
3021		close_sync(conf);
3022		*skipped = 1;
3023		return sectors_skipped;
3024	}
3025
3026	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3027		return reshape_request(mddev, sector_nr, skipped);
3028
3029	if (chunks_skipped >= conf->geo.raid_disks) {
3030		/* if there has been nothing to do on any drive,
3031		 * then there is nothing to do at all..
3032		 */
3033		*skipped = 1;
3034		return (max_sector - sector_nr) + sectors_skipped;
3035	}
3036
3037	if (max_sector > mddev->resync_max)
3038		max_sector = mddev->resync_max; /* Don't do IO beyond here */
3039
3040	/* make sure whole request will fit in a chunk - if chunks
3041	 * are meaningful
3042	 */
3043	if (conf->geo.near_copies < conf->geo.raid_disks &&
3044	    max_sector > (sector_nr | chunk_mask))
3045		max_sector = (sector_nr | chunk_mask) + 1;
3046
3047	/*
3048	 * If there is non-resync activity waiting for a turn, then let it
3049	 * though before starting on this new sync request.
3050	 */
3051	if (conf->nr_waiting)
3052		schedule_timeout_uninterruptible(1);
3053
3054	/* Again, very different code for resync and recovery.
3055	 * Both must result in an r10bio with a list of bios that
3056	 * have bi_end_io, bi_sector, bi_disk set,
3057	 * and bi_private set to the r10bio.
3058	 * For recovery, we may actually create several r10bios
3059	 * with 2 bios in each, that correspond to the bios in the main one.
3060	 * In this case, the subordinate r10bios link back through a
3061	 * borrowed master_bio pointer, and the counter in the master
3062	 * includes a ref from each subordinate.
3063	 */
3064	/* First, we decide what to do and set ->bi_end_io
3065	 * To end_sync_read if we want to read, and
3066	 * end_sync_write if we will want to write.
3067	 */
3068
3069	max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3070	if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3071		/* recovery... the complicated one */
3072		int j;
3073		r10_bio = NULL;
3074
3075		for (i = 0 ; i < conf->geo.raid_disks; i++) {
3076			int still_degraded;
3077			struct r10bio *rb2;
3078			sector_t sect;
3079			int must_sync;
3080			int any_working;
3081			struct raid10_info *mirror = &conf->mirrors[i];
3082			struct md_rdev *mrdev, *mreplace;
3083
3084			rcu_read_lock();
3085			mrdev = rcu_dereference(mirror->rdev);
3086			mreplace = rcu_dereference(mirror->replacement);
3087
3088			if ((mrdev == NULL ||
3089			     test_bit(Faulty, &mrdev->flags) ||
3090			     test_bit(In_sync, &mrdev->flags)) &&
3091			    (mreplace == NULL ||
3092			     test_bit(Faulty, &mreplace->flags))) {
3093				rcu_read_unlock();
3094				continue;
3095			}
3096
3097			still_degraded = 0;
3098			/* want to reconstruct this device */
3099			rb2 = r10_bio;
3100			sect = raid10_find_virt(conf, sector_nr, i);
3101			if (sect >= mddev->resync_max_sectors) {
3102				/* last stripe is not complete - don't
3103				 * try to recover this sector.
3104				 */
3105				rcu_read_unlock();
3106				continue;
3107			}
3108			if (mreplace && test_bit(Faulty, &mreplace->flags))
3109				mreplace = NULL;
3110			/* Unless we are doing a full sync, or a replacement
3111			 * we only need to recover the block if it is set in
3112			 * the bitmap
3113			 */
3114			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3115						      &sync_blocks, 1);
3116			if (sync_blocks < max_sync)
3117				max_sync = sync_blocks;
3118			if (!must_sync &&
3119			    mreplace == NULL &&
3120			    !conf->fullsync) {
3121				/* yep, skip the sync_blocks here, but don't assume
3122				 * that there will never be anything to do here
3123				 */
3124				chunks_skipped = -1;
3125				rcu_read_unlock();
3126				continue;
3127			}
3128			atomic_inc(&mrdev->nr_pending);
3129			if (mreplace)
3130				atomic_inc(&mreplace->nr_pending);
3131			rcu_read_unlock();
3132
3133			r10_bio = raid10_alloc_init_r10buf(conf);
3134			r10_bio->state = 0;
3135			raise_barrier(conf, rb2 != NULL);
3136			atomic_set(&r10_bio->remaining, 0);
3137
3138			r10_bio->master_bio = (struct bio*)rb2;
3139			if (rb2)
3140				atomic_inc(&rb2->remaining);
3141			r10_bio->mddev = mddev;
3142			set_bit(R10BIO_IsRecover, &r10_bio->state);
3143			r10_bio->sector = sect;
3144
3145			raid10_find_phys(conf, r10_bio);
3146
3147			/* Need to check if the array will still be
3148			 * degraded
3149			 */
3150			rcu_read_lock();
3151			for (j = 0; j < conf->geo.raid_disks; j++) {
3152				struct md_rdev *rdev = rcu_dereference(
3153					conf->mirrors[j].rdev);
3154				if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3155					still_degraded = 1;
3156					break;
3157				}
3158			}
3159
3160			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3161						      &sync_blocks, still_degraded);
3162
3163			any_working = 0;
3164			for (j=0; j<conf->copies;j++) {
3165				int k;
3166				int d = r10_bio->devs[j].devnum;
3167				sector_t from_addr, to_addr;
3168				struct md_rdev *rdev =
3169					rcu_dereference(conf->mirrors[d].rdev);
3170				sector_t sector, first_bad;
3171				int bad_sectors;
3172				if (!rdev ||
3173				    !test_bit(In_sync, &rdev->flags))
3174					continue;
3175				/* This is where we read from */
3176				any_working = 1;
 
3177				sector = r10_bio->devs[j].addr;
3178
3179				if (is_badblock(rdev, sector, max_sync,
3180						&first_bad, &bad_sectors)) {
3181					if (first_bad > sector)
3182						max_sync = first_bad - sector;
3183					else {
3184						bad_sectors -= (sector
3185								- first_bad);
3186						if (max_sync > bad_sectors)
3187							max_sync = bad_sectors;
3188						continue;
3189					}
3190				}
3191				bio = r10_bio->devs[0].bio;
 
3192				bio->bi_next = biolist;
3193				biolist = bio;
 
3194				bio->bi_end_io = end_sync_read;
3195				bio_set_op_attrs(bio, REQ_OP_READ, 0);
3196				if (test_bit(FailFast, &rdev->flags))
3197					bio->bi_opf |= MD_FAILFAST;
3198				from_addr = r10_bio->devs[j].addr;
3199				bio->bi_iter.bi_sector = from_addr +
3200					rdev->data_offset;
3201				bio_set_dev(bio, rdev->bdev);
3202				atomic_inc(&rdev->nr_pending);
3203				/* and we write to 'i' (if not in_sync) */
3204
3205				for (k=0; k<conf->copies; k++)
3206					if (r10_bio->devs[k].devnum == i)
3207						break;
3208				BUG_ON(k == conf->copies);
3209				to_addr = r10_bio->devs[k].addr;
3210				r10_bio->devs[0].devnum = d;
3211				r10_bio->devs[0].addr = from_addr;
3212				r10_bio->devs[1].devnum = i;
3213				r10_bio->devs[1].addr = to_addr;
3214
3215				if (!test_bit(In_sync, &mrdev->flags)) {
 
3216					bio = r10_bio->devs[1].bio;
 
3217					bio->bi_next = biolist;
3218					biolist = bio;
 
3219					bio->bi_end_io = end_sync_write;
3220					bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3221					bio->bi_iter.bi_sector = to_addr
3222						+ mrdev->data_offset;
3223					bio_set_dev(bio, mrdev->bdev);
3224					atomic_inc(&r10_bio->remaining);
3225				} else
3226					r10_bio->devs[1].bio->bi_end_io = NULL;
3227
3228				/* and maybe write to replacement */
3229				bio = r10_bio->devs[1].repl_bio;
3230				if (bio)
3231					bio->bi_end_io = NULL;
3232				/* Note: if mreplace != NULL, then bio
 
3233				 * cannot be NULL as r10buf_pool_alloc will
3234				 * have allocated it.
3235				 * So the second test here is pointless.
3236				 * But it keeps semantic-checkers happy, and
3237				 * this comment keeps human reviewers
3238				 * happy.
3239				 */
3240				if (mreplace == NULL || bio == NULL ||
3241				    test_bit(Faulty, &mreplace->flags))
3242					break;
 
3243				bio->bi_next = biolist;
3244				biolist = bio;
 
3245				bio->bi_end_io = end_sync_write;
3246				bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3247				bio->bi_iter.bi_sector = to_addr +
3248					mreplace->data_offset;
3249				bio_set_dev(bio, mreplace->bdev);
3250				atomic_inc(&r10_bio->remaining);
3251				break;
3252			}
3253			rcu_read_unlock();
3254			if (j == conf->copies) {
3255				/* Cannot recover, so abort the recovery or
3256				 * record a bad block */
3257				if (any_working) {
3258					/* problem is that there are bad blocks
3259					 * on other device(s)
3260					 */
3261					int k;
3262					for (k = 0; k < conf->copies; k++)
3263						if (r10_bio->devs[k].devnum == i)
3264							break;
3265					if (!test_bit(In_sync,
3266						      &mrdev->flags)
3267					    && !rdev_set_badblocks(
3268						    mrdev,
3269						    r10_bio->devs[k].addr,
3270						    max_sync, 0))
3271						any_working = 0;
3272					if (mreplace &&
3273					    !rdev_set_badblocks(
3274						    mreplace,
3275						    r10_bio->devs[k].addr,
3276						    max_sync, 0))
3277						any_working = 0;
3278				}
3279				if (!any_working)  {
3280					if (!test_and_set_bit(MD_RECOVERY_INTR,
3281							      &mddev->recovery))
3282						pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
 
3283						       mdname(mddev));
3284					mirror->recovery_disabled
3285						= mddev->recovery_disabled;
3286				}
3287				put_buf(r10_bio);
3288				if (rb2)
3289					atomic_dec(&rb2->remaining);
3290				r10_bio = rb2;
3291				rdev_dec_pending(mrdev, mddev);
3292				if (mreplace)
3293					rdev_dec_pending(mreplace, mddev);
3294				break;
3295			}
3296			rdev_dec_pending(mrdev, mddev);
3297			if (mreplace)
3298				rdev_dec_pending(mreplace, mddev);
3299			if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3300				/* Only want this if there is elsewhere to
3301				 * read from. 'j' is currently the first
3302				 * readable copy.
3303				 */
3304				int targets = 1;
3305				for (; j < conf->copies; j++) {
3306					int d = r10_bio->devs[j].devnum;
3307					if (conf->mirrors[d].rdev &&
3308					    test_bit(In_sync,
3309						      &conf->mirrors[d].rdev->flags))
3310						targets++;
3311				}
3312				if (targets == 1)
3313					r10_bio->devs[0].bio->bi_opf
3314						&= ~MD_FAILFAST;
3315			}
3316		}
3317		if (biolist == NULL) {
3318			while (r10_bio) {
3319				struct r10bio *rb2 = r10_bio;
3320				r10_bio = (struct r10bio*) rb2->master_bio;
3321				rb2->master_bio = NULL;
3322				put_buf(rb2);
3323			}
3324			goto giveup;
3325		}
3326	} else {
3327		/* resync. Schedule a read for every block at this virt offset */
3328		int count = 0;
3329
3330		/*
3331		 * Since curr_resync_completed could probably not update in
3332		 * time, and we will set cluster_sync_low based on it.
3333		 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3334		 * safety reason, which ensures curr_resync_completed is
3335		 * updated in bitmap_cond_end_sync.
3336		 */
3337		bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3338				     mddev_is_clustered(mddev) &&
3339				     (sector_nr + 2 * RESYNC_SECTORS >
3340				      conf->cluster_sync_high));
3341
3342		if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3343				       &sync_blocks, mddev->degraded) &&
3344		    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3345						 &mddev->recovery)) {
3346			/* We can skip this block */
3347			*skipped = 1;
3348			return sync_blocks + sectors_skipped;
3349		}
3350		if (sync_blocks < max_sync)
3351			max_sync = sync_blocks;
3352		r10_bio = raid10_alloc_init_r10buf(conf);
3353		r10_bio->state = 0;
3354
3355		r10_bio->mddev = mddev;
3356		atomic_set(&r10_bio->remaining, 0);
3357		raise_barrier(conf, 0);
3358		conf->next_resync = sector_nr;
3359
3360		r10_bio->master_bio = NULL;
3361		r10_bio->sector = sector_nr;
3362		set_bit(R10BIO_IsSync, &r10_bio->state);
3363		raid10_find_phys(conf, r10_bio);
3364		r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3365
3366		for (i = 0; i < conf->copies; i++) {
3367			int d = r10_bio->devs[i].devnum;
3368			sector_t first_bad, sector;
3369			int bad_sectors;
3370			struct md_rdev *rdev;
3371
3372			if (r10_bio->devs[i].repl_bio)
3373				r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3374
3375			bio = r10_bio->devs[i].bio;
3376			bio->bi_status = BLK_STS_IOERR;
3377			rcu_read_lock();
3378			rdev = rcu_dereference(conf->mirrors[d].rdev);
3379			if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3380				rcu_read_unlock();
3381				continue;
3382			}
3383			sector = r10_bio->devs[i].addr;
3384			if (is_badblock(rdev, sector, max_sync,
 
3385					&first_bad, &bad_sectors)) {
3386				if (first_bad > sector)
3387					max_sync = first_bad - sector;
3388				else {
3389					bad_sectors -= (sector - first_bad);
3390					if (max_sync > bad_sectors)
3391						max_sync = bad_sectors;
3392					rcu_read_unlock();
3393					continue;
3394				}
3395			}
3396			atomic_inc(&rdev->nr_pending);
3397			atomic_inc(&r10_bio->remaining);
3398			bio->bi_next = biolist;
3399			biolist = bio;
 
3400			bio->bi_end_io = end_sync_read;
3401			bio_set_op_attrs(bio, REQ_OP_READ, 0);
3402			if (test_bit(FailFast, &rdev->flags))
3403				bio->bi_opf |= MD_FAILFAST;
3404			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3405			bio_set_dev(bio, rdev->bdev);
3406			count++;
3407
3408			rdev = rcu_dereference(conf->mirrors[d].replacement);
3409			if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3410				rcu_read_unlock();
3411				continue;
3412			}
3413			atomic_inc(&rdev->nr_pending);
3414
3415			/* Need to set up for writing to the replacement */
3416			bio = r10_bio->devs[i].repl_bio;
3417			bio->bi_status = BLK_STS_IOERR;
 
3418
3419			sector = r10_bio->devs[i].addr;
 
3420			bio->bi_next = biolist;
3421			biolist = bio;
 
3422			bio->bi_end_io = end_sync_write;
3423			bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3424			if (test_bit(FailFast, &rdev->flags))
3425				bio->bi_opf |= MD_FAILFAST;
3426			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3427			bio_set_dev(bio, rdev->bdev);
3428			count++;
3429			rcu_read_unlock();
3430		}
3431
3432		if (count < 2) {
3433			for (i=0; i<conf->copies; i++) {
3434				int d = r10_bio->devs[i].devnum;
3435				if (r10_bio->devs[i].bio->bi_end_io)
3436					rdev_dec_pending(conf->mirrors[d].rdev,
3437							 mddev);
3438				if (r10_bio->devs[i].repl_bio &&
3439				    r10_bio->devs[i].repl_bio->bi_end_io)
3440					rdev_dec_pending(
3441						conf->mirrors[d].replacement,
3442						mddev);
3443			}
3444			put_buf(r10_bio);
3445			biolist = NULL;
3446			goto giveup;
3447		}
3448	}
3449
3450	nr_sectors = 0;
3451	if (sector_nr + max_sync < max_sector)
3452		max_sector = sector_nr + max_sync;
3453	do {
3454		struct page *page;
3455		int len = PAGE_SIZE;
3456		if (sector_nr + (len>>9) > max_sector)
3457			len = (max_sector - sector_nr) << 9;
3458		if (len == 0)
3459			break;
3460		for (bio= biolist ; bio ; bio=bio->bi_next) {
3461			struct resync_pages *rp = get_resync_pages(bio);
3462			page = resync_fetch_page(rp, page_idx);
3463			/*
3464			 * won't fail because the vec table is big enough
3465			 * to hold all these pages
3466			 */
3467			bio_add_page(bio, page, len, 0);
 
 
 
 
 
 
 
 
 
3468		}
3469		nr_sectors += len>>9;
3470		sector_nr += len>>9;
3471	} while (++page_idx < RESYNC_PAGES);
 
3472	r10_bio->sectors = nr_sectors;
3473
3474	if (mddev_is_clustered(mddev) &&
3475	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3476		/* It is resync not recovery */
3477		if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3478			conf->cluster_sync_low = mddev->curr_resync_completed;
3479			raid10_set_cluster_sync_high(conf);
3480			/* Send resync message */
3481			md_cluster_ops->resync_info_update(mddev,
3482						conf->cluster_sync_low,
3483						conf->cluster_sync_high);
3484		}
3485	} else if (mddev_is_clustered(mddev)) {
3486		/* This is recovery not resync */
3487		sector_t sect_va1, sect_va2;
3488		bool broadcast_msg = false;
3489
3490		for (i = 0; i < conf->geo.raid_disks; i++) {
3491			/*
3492			 * sector_nr is a device address for recovery, so we
3493			 * need translate it to array address before compare
3494			 * with cluster_sync_high.
3495			 */
3496			sect_va1 = raid10_find_virt(conf, sector_nr, i);
3497
3498			if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3499				broadcast_msg = true;
3500				/*
3501				 * curr_resync_completed is similar as
3502				 * sector_nr, so make the translation too.
3503				 */
3504				sect_va2 = raid10_find_virt(conf,
3505					mddev->curr_resync_completed, i);
3506
3507				if (conf->cluster_sync_low == 0 ||
3508				    conf->cluster_sync_low > sect_va2)
3509					conf->cluster_sync_low = sect_va2;
3510			}
3511		}
3512		if (broadcast_msg) {
3513			raid10_set_cluster_sync_high(conf);
3514			md_cluster_ops->resync_info_update(mddev,
3515						conf->cluster_sync_low,
3516						conf->cluster_sync_high);
3517		}
3518	}
3519
3520	while (biolist) {
3521		bio = biolist;
3522		biolist = biolist->bi_next;
3523
3524		bio->bi_next = NULL;
3525		r10_bio = get_resync_r10bio(bio);
3526		r10_bio->sectors = nr_sectors;
3527
3528		if (bio->bi_end_io == end_sync_read) {
3529			md_sync_acct_bio(bio, nr_sectors);
3530			bio->bi_status = 0;
3531			generic_make_request(bio);
3532		}
3533	}
3534
3535	if (sectors_skipped)
3536		/* pretend they weren't skipped, it makes
3537		 * no important difference in this case
3538		 */
3539		md_done_sync(mddev, sectors_skipped, 1);
3540
3541	return sectors_skipped + nr_sectors;
3542 giveup:
3543	/* There is nowhere to write, so all non-sync
3544	 * drives must be failed or in resync, all drives
3545	 * have a bad block, so try the next chunk...
3546	 */
3547	if (sector_nr + max_sync < max_sector)
3548		max_sector = sector_nr + max_sync;
3549
3550	sectors_skipped += (max_sector - sector_nr);
3551	chunks_skipped ++;
3552	sector_nr = max_sector;
3553	goto skipped;
3554}
3555
3556static sector_t
3557raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3558{
3559	sector_t size;
3560	struct r10conf *conf = mddev->private;
3561
3562	if (!raid_disks)
3563		raid_disks = min(conf->geo.raid_disks,
3564				 conf->prev.raid_disks);
3565	if (!sectors)
3566		sectors = conf->dev_sectors;
3567
3568	size = sectors >> conf->geo.chunk_shift;
3569	sector_div(size, conf->geo.far_copies);
3570	size = size * raid_disks;
3571	sector_div(size, conf->geo.near_copies);
3572
3573	return size << conf->geo.chunk_shift;
3574}
3575
3576static void calc_sectors(struct r10conf *conf, sector_t size)
3577{
3578	/* Calculate the number of sectors-per-device that will
3579	 * actually be used, and set conf->dev_sectors and
3580	 * conf->stride
3581	 */
3582
3583	size = size >> conf->geo.chunk_shift;
3584	sector_div(size, conf->geo.far_copies);
3585	size = size * conf->geo.raid_disks;
3586	sector_div(size, conf->geo.near_copies);
3587	/* 'size' is now the number of chunks in the array */
3588	/* calculate "used chunks per device" */
3589	size = size * conf->copies;
3590
3591	/* We need to round up when dividing by raid_disks to
3592	 * get the stride size.
3593	 */
3594	size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3595
3596	conf->dev_sectors = size << conf->geo.chunk_shift;
3597
3598	if (conf->geo.far_offset)
3599		conf->geo.stride = 1 << conf->geo.chunk_shift;
3600	else {
3601		sector_div(size, conf->geo.far_copies);
3602		conf->geo.stride = size << conf->geo.chunk_shift;
3603	}
3604}
3605
3606enum geo_type {geo_new, geo_old, geo_start};
3607static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3608{
3609	int nc, fc, fo;
3610	int layout, chunk, disks;
3611	switch (new) {
3612	case geo_old:
3613		layout = mddev->layout;
3614		chunk = mddev->chunk_sectors;
3615		disks = mddev->raid_disks - mddev->delta_disks;
3616		break;
3617	case geo_new:
3618		layout = mddev->new_layout;
3619		chunk = mddev->new_chunk_sectors;
3620		disks = mddev->raid_disks;
3621		break;
3622	default: /* avoid 'may be unused' warnings */
3623	case geo_start: /* new when starting reshape - raid_disks not
3624			 * updated yet. */
3625		layout = mddev->new_layout;
3626		chunk = mddev->new_chunk_sectors;
3627		disks = mddev->raid_disks + mddev->delta_disks;
3628		break;
3629	}
3630	if (layout >> 19)
3631		return -1;
3632	if (chunk < (PAGE_SIZE >> 9) ||
3633	    !is_power_of_2(chunk))
3634		return -2;
3635	nc = layout & 255;
3636	fc = (layout >> 8) & 255;
3637	fo = layout & (1<<16);
3638	geo->raid_disks = disks;
3639	geo->near_copies = nc;
3640	geo->far_copies = fc;
3641	geo->far_offset = fo;
3642	switch (layout >> 17) {
3643	case 0:	/* original layout.  simple but not always optimal */
3644		geo->far_set_size = disks;
3645		break;
3646	case 1: /* "improved" layout which was buggy.  Hopefully no-one is
3647		 * actually using this, but leave code here just in case.*/
3648		geo->far_set_size = disks/fc;
3649		WARN(geo->far_set_size < fc,
3650		     "This RAID10 layout does not provide data safety - please backup and create new array\n");
3651		break;
3652	case 2: /* "improved" layout fixed to match documentation */
3653		geo->far_set_size = fc * nc;
3654		break;
3655	default: /* Not a valid layout */
3656		return -1;
3657	}
3658	geo->chunk_mask = chunk - 1;
3659	geo->chunk_shift = ffz(~chunk);
3660	return nc*fc;
3661}
3662
3663static struct r10conf *setup_conf(struct mddev *mddev)
3664{
3665	struct r10conf *conf = NULL;
3666	int err = -EINVAL;
3667	struct geom geo;
3668	int copies;
3669
3670	copies = setup_geo(&geo, mddev, geo_new);
3671
3672	if (copies == -2) {
3673		pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3674			mdname(mddev), PAGE_SIZE);
 
3675		goto out;
3676	}
3677
3678	if (copies < 2 || copies > mddev->raid_disks) {
3679		pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3680			mdname(mddev), mddev->new_layout);
3681		goto out;
3682	}
3683
3684	err = -ENOMEM;
3685	conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3686	if (!conf)
3687		goto out;
3688
3689	/* FIXME calc properly */
3690	conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3691							    max(0,-mddev->delta_disks)),
3692				GFP_KERNEL);
3693	if (!conf->mirrors)
3694		goto out;
3695
3696	conf->tmppage = alloc_page(GFP_KERNEL);
3697	if (!conf->tmppage)
3698		goto out;
3699
3700	conf->geo = geo;
3701	conf->copies = copies;
3702	conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3703					   r10bio_pool_free, conf);
3704	if (!conf->r10bio_pool)
3705		goto out;
3706
3707	conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
3708	if (!conf->bio_split)
3709		goto out;
3710
3711	calc_sectors(conf, mddev->dev_sectors);
3712	if (mddev->reshape_position == MaxSector) {
3713		conf->prev = conf->geo;
3714		conf->reshape_progress = MaxSector;
3715	} else {
3716		if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3717			err = -EINVAL;
3718			goto out;
3719		}
3720		conf->reshape_progress = mddev->reshape_position;
3721		if (conf->prev.far_offset)
3722			conf->prev.stride = 1 << conf->prev.chunk_shift;
3723		else
3724			/* far_copies must be 1 */
3725			conf->prev.stride = conf->dev_sectors;
3726	}
3727	conf->reshape_safe = conf->reshape_progress;
3728	spin_lock_init(&conf->device_lock);
3729	INIT_LIST_HEAD(&conf->retry_list);
3730	INIT_LIST_HEAD(&conf->bio_end_io_list);
3731
3732	spin_lock_init(&conf->resync_lock);
3733	init_waitqueue_head(&conf->wait_barrier);
3734	atomic_set(&conf->nr_pending, 0);
3735
3736	conf->thread = md_register_thread(raid10d, mddev, "raid10");
3737	if (!conf->thread)
3738		goto out;
3739
3740	conf->mddev = mddev;
3741	return conf;
3742
3743 out:
 
 
 
3744	if (conf) {
3745		mempool_destroy(conf->r10bio_pool);
 
3746		kfree(conf->mirrors);
3747		safe_put_page(conf->tmppage);
3748		if (conf->bio_split)
3749			bioset_free(conf->bio_split);
3750		kfree(conf);
3751	}
3752	return ERR_PTR(err);
3753}
3754
3755static int raid10_run(struct mddev *mddev)
3756{
3757	struct r10conf *conf;
3758	int i, disk_idx, chunk_size;
3759	struct raid10_info *disk;
3760	struct md_rdev *rdev;
3761	sector_t size;
3762	sector_t min_offset_diff = 0;
3763	int first = 1;
3764	bool discard_supported = false;
3765
3766	if (mddev_init_writes_pending(mddev) < 0)
3767		return -ENOMEM;
3768
3769	if (mddev->private == NULL) {
3770		conf = setup_conf(mddev);
3771		if (IS_ERR(conf))
3772			return PTR_ERR(conf);
3773		mddev->private = conf;
3774	}
3775	conf = mddev->private;
3776	if (!conf)
3777		goto out;
3778
3779	if (mddev_is_clustered(conf->mddev)) {
3780		int fc, fo;
3781
3782		fc = (mddev->layout >> 8) & 255;
3783		fo = mddev->layout & (1<<16);
3784		if (fc > 1 || fo > 0) {
3785			pr_err("only near layout is supported by clustered"
3786				" raid10\n");
3787			goto out_free_conf;
3788		}
3789	}
3790
3791	mddev->thread = conf->thread;
3792	conf->thread = NULL;
3793
3794	chunk_size = mddev->chunk_sectors << 9;
3795	if (mddev->queue) {
3796		blk_queue_max_discard_sectors(mddev->queue,
3797					      mddev->chunk_sectors);
3798		blk_queue_max_write_same_sectors(mddev->queue, 0);
3799		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3800		blk_queue_io_min(mddev->queue, chunk_size);
3801		if (conf->geo.raid_disks % conf->geo.near_copies)
3802			blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3803		else
3804			blk_queue_io_opt(mddev->queue, chunk_size *
3805					 (conf->geo.raid_disks / conf->geo.near_copies));
3806	}
3807
3808	rdev_for_each(rdev, mddev) {
3809		long long diff;
 
3810
3811		disk_idx = rdev->raid_disk;
3812		if (disk_idx < 0)
3813			continue;
3814		if (disk_idx >= conf->geo.raid_disks &&
3815		    disk_idx >= conf->prev.raid_disks)
3816			continue;
3817		disk = conf->mirrors + disk_idx;
3818
3819		if (test_bit(Replacement, &rdev->flags)) {
3820			if (disk->replacement)
3821				goto out_free_conf;
3822			disk->replacement = rdev;
3823		} else {
3824			if (disk->rdev)
3825				goto out_free_conf;
3826			disk->rdev = rdev;
3827		}
 
 
 
3828		diff = (rdev->new_data_offset - rdev->data_offset);
3829		if (!mddev->reshape_backwards)
3830			diff = -diff;
3831		if (diff < 0)
3832			diff = 0;
3833		if (first || diff < min_offset_diff)
3834			min_offset_diff = diff;
3835
3836		if (mddev->gendisk)
3837			disk_stack_limits(mddev->gendisk, rdev->bdev,
3838					  rdev->data_offset << 9);
3839
3840		disk->head_position = 0;
3841
3842		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3843			discard_supported = true;
3844		first = 0;
3845	}
3846
3847	if (mddev->queue) {
3848		if (discard_supported)
3849			blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3850						mddev->queue);
3851		else
3852			blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3853						  mddev->queue);
3854	}
3855	/* need to check that every block has at least one working mirror */
3856	if (!enough(conf, -1)) {
3857		pr_err("md/raid10:%s: not enough operational mirrors.\n",
3858		       mdname(mddev));
3859		goto out_free_conf;
3860	}
3861
3862	if (conf->reshape_progress != MaxSector) {
3863		/* must ensure that shape change is supported */
3864		if (conf->geo.far_copies != 1 &&
3865		    conf->geo.far_offset == 0)
3866			goto out_free_conf;
3867		if (conf->prev.far_copies != 1 &&
3868		    conf->prev.far_offset == 0)
3869			goto out_free_conf;
3870	}
3871
3872	mddev->degraded = 0;
3873	for (i = 0;
3874	     i < conf->geo.raid_disks
3875		     || i < conf->prev.raid_disks;
3876	     i++) {
3877
3878		disk = conf->mirrors + i;
3879
3880		if (!disk->rdev && disk->replacement) {
3881			/* The replacement is all we have - use it */
3882			disk->rdev = disk->replacement;
3883			disk->replacement = NULL;
3884			clear_bit(Replacement, &disk->rdev->flags);
3885		}
3886
3887		if (!disk->rdev ||
3888		    !test_bit(In_sync, &disk->rdev->flags)) {
3889			disk->head_position = 0;
3890			mddev->degraded++;
3891			if (disk->rdev &&
3892			    disk->rdev->saved_raid_disk < 0)
3893				conf->fullsync = 1;
3894		}
3895		disk->recovery_disabled = mddev->recovery_disabled - 1;
3896	}
3897
3898	if (mddev->recovery_cp != MaxSector)
3899		pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3900			  mdname(mddev));
3901	pr_info("md/raid10:%s: active with %d out of %d devices\n",
 
 
3902		mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3903		conf->geo.raid_disks);
3904	/*
3905	 * Ok, everything is just fine now
3906	 */
3907	mddev->dev_sectors = conf->dev_sectors;
3908	size = raid10_size(mddev, 0, 0);
3909	md_set_array_sectors(mddev, size);
3910	mddev->resync_max_sectors = size;
3911	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3912
3913	if (mddev->queue) {
3914		int stripe = conf->geo.raid_disks *
3915			((mddev->chunk_sectors << 9) / PAGE_SIZE);
 
 
3916
3917		/* Calculate max read-ahead size.
3918		 * We need to readahead at least twice a whole stripe....
3919		 * maybe...
3920		 */
3921		stripe /= conf->geo.near_copies;
3922		if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3923			mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
 
3924	}
3925
 
3926	if (md_integrity_register(mddev))
3927		goto out_free_conf;
3928
3929	if (conf->reshape_progress != MaxSector) {
3930		unsigned long before_length, after_length;
3931
3932		before_length = ((1 << conf->prev.chunk_shift) *
3933				 conf->prev.far_copies);
3934		after_length = ((1 << conf->geo.chunk_shift) *
3935				conf->geo.far_copies);
3936
3937		if (max(before_length, after_length) > min_offset_diff) {
3938			/* This cannot work */
3939			pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3940			goto out_free_conf;
3941		}
3942		conf->offset_diff = min_offset_diff;
3943
 
3944		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3945		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3946		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3947		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3948		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3949							"reshape");
3950	}
3951
3952	return 0;
3953
3954out_free_conf:
3955	md_unregister_thread(&mddev->thread);
3956	mempool_destroy(conf->r10bio_pool);
 
3957	safe_put_page(conf->tmppage);
3958	kfree(conf->mirrors);
3959	kfree(conf);
3960	mddev->private = NULL;
3961out:
3962	return -EIO;
3963}
3964
3965static void raid10_free(struct mddev *mddev, void *priv)
3966{
3967	struct r10conf *conf = priv;
 
 
 
3968
3969	mempool_destroy(conf->r10bio_pool);
 
 
 
 
 
 
3970	safe_put_page(conf->tmppage);
3971	kfree(conf->mirrors);
3972	kfree(conf->mirrors_old);
3973	kfree(conf->mirrors_new);
3974	if (conf->bio_split)
3975		bioset_free(conf->bio_split);
3976	kfree(conf);
 
 
3977}
3978
3979static void raid10_quiesce(struct mddev *mddev, int quiesce)
3980{
3981	struct r10conf *conf = mddev->private;
3982
3983	if (quiesce)
 
3984		raise_barrier(conf, 0);
3985	else
 
3986		lower_barrier(conf);
 
 
3987}
3988
3989static int raid10_resize(struct mddev *mddev, sector_t sectors)
3990{
3991	/* Resize of 'far' arrays is not supported.
3992	 * For 'near' and 'offset' arrays we can set the
3993	 * number of sectors used to be an appropriate multiple
3994	 * of the chunk size.
3995	 * For 'offset', this is far_copies*chunksize.
3996	 * For 'near' the multiplier is the LCM of
3997	 * near_copies and raid_disks.
3998	 * So if far_copies > 1 && !far_offset, fail.
3999	 * Else find LCM(raid_disks, near_copy)*far_copies and
4000	 * multiply by chunk_size.  Then round to this number.
4001	 * This is mostly done by raid10_size()
4002	 */
4003	struct r10conf *conf = mddev->private;
4004	sector_t oldsize, size;
4005
4006	if (mddev->reshape_position != MaxSector)
4007		return -EBUSY;
4008
4009	if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4010		return -EINVAL;
4011
4012	oldsize = raid10_size(mddev, 0, 0);
4013	size = raid10_size(mddev, sectors, 0);
4014	if (mddev->external_size &&
4015	    mddev->array_sectors > size)
4016		return -EINVAL;
4017	if (mddev->bitmap) {
4018		int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
4019		if (ret)
4020			return ret;
4021	}
4022	md_set_array_sectors(mddev, size);
 
 
4023	if (sectors > mddev->dev_sectors &&
4024	    mddev->recovery_cp > oldsize) {
4025		mddev->recovery_cp = oldsize;
4026		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4027	}
4028	calc_sectors(conf, sectors);
4029	mddev->dev_sectors = conf->dev_sectors;
4030	mddev->resync_max_sectors = size;
4031	return 0;
4032}
4033
4034static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4035{
4036	struct md_rdev *rdev;
4037	struct r10conf *conf;
4038
4039	if (mddev->degraded > 0) {
4040		pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4041			mdname(mddev));
4042		return ERR_PTR(-EINVAL);
4043	}
4044	sector_div(size, devs);
4045
4046	/* Set new parameters */
4047	mddev->new_level = 10;
4048	/* new layout: far_copies = 1, near_copies = 2 */
4049	mddev->new_layout = (1<<8) + 2;
4050	mddev->new_chunk_sectors = mddev->chunk_sectors;
4051	mddev->delta_disks = mddev->raid_disks;
4052	mddev->raid_disks *= 2;
4053	/* make sure it will be not marked as dirty */
4054	mddev->recovery_cp = MaxSector;
4055	mddev->dev_sectors = size;
4056
4057	conf = setup_conf(mddev);
4058	if (!IS_ERR(conf)) {
4059		rdev_for_each(rdev, mddev)
4060			if (rdev->raid_disk >= 0) {
4061				rdev->new_raid_disk = rdev->raid_disk * 2;
4062				rdev->sectors = size;
4063			}
4064		conf->barrier = 1;
4065	}
4066
4067	return conf;
4068}
4069
4070static void *raid10_takeover(struct mddev *mddev)
4071{
4072	struct r0conf *raid0_conf;
4073
4074	/* raid10 can take over:
4075	 *  raid0 - providing it has only two drives
4076	 */
4077	if (mddev->level == 0) {
4078		/* for raid0 takeover only one zone is supported */
4079		raid0_conf = mddev->private;
4080		if (raid0_conf->nr_strip_zones > 1) {
4081			pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4082				mdname(mddev));
 
4083			return ERR_PTR(-EINVAL);
4084		}
4085		return raid10_takeover_raid0(mddev,
4086			raid0_conf->strip_zone->zone_end,
4087			raid0_conf->strip_zone->nb_dev);
4088	}
4089	return ERR_PTR(-EINVAL);
4090}
4091
4092static int raid10_check_reshape(struct mddev *mddev)
4093{
4094	/* Called when there is a request to change
4095	 * - layout (to ->new_layout)
4096	 * - chunk size (to ->new_chunk_sectors)
4097	 * - raid_disks (by delta_disks)
4098	 * or when trying to restart a reshape that was ongoing.
4099	 *
4100	 * We need to validate the request and possibly allocate
4101	 * space if that might be an issue later.
4102	 *
4103	 * Currently we reject any reshape of a 'far' mode array,
4104	 * allow chunk size to change if new is generally acceptable,
4105	 * allow raid_disks to increase, and allow
4106	 * a switch between 'near' mode and 'offset' mode.
4107	 */
4108	struct r10conf *conf = mddev->private;
4109	struct geom geo;
4110
4111	if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4112		return -EINVAL;
4113
4114	if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4115		/* mustn't change number of copies */
4116		return -EINVAL;
4117	if (geo.far_copies > 1 && !geo.far_offset)
4118		/* Cannot switch to 'far' mode */
4119		return -EINVAL;
4120
4121	if (mddev->array_sectors & geo.chunk_mask)
4122			/* not factor of array size */
4123			return -EINVAL;
4124
4125	if (!enough(conf, -1))
4126		return -EINVAL;
4127
4128	kfree(conf->mirrors_new);
4129	conf->mirrors_new = NULL;
4130	if (mddev->delta_disks > 0) {
4131		/* allocate new 'mirrors' list */
4132		conf->mirrors_new = kzalloc(
4133			sizeof(struct raid10_info)
4134			*(mddev->raid_disks +
4135			  mddev->delta_disks),
4136			GFP_KERNEL);
4137		if (!conf->mirrors_new)
4138			return -ENOMEM;
4139	}
4140	return 0;
4141}
4142
4143/*
4144 * Need to check if array has failed when deciding whether to:
4145 *  - start an array
4146 *  - remove non-faulty devices
4147 *  - add a spare
4148 *  - allow a reshape
4149 * This determination is simple when no reshape is happening.
4150 * However if there is a reshape, we need to carefully check
4151 * both the before and after sections.
4152 * This is because some failed devices may only affect one
4153 * of the two sections, and some non-in_sync devices may
4154 * be insync in the section most affected by failed devices.
4155 */
4156static int calc_degraded(struct r10conf *conf)
4157{
4158	int degraded, degraded2;
4159	int i;
4160
4161	rcu_read_lock();
4162	degraded = 0;
4163	/* 'prev' section first */
4164	for (i = 0; i < conf->prev.raid_disks; i++) {
4165		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4166		if (!rdev || test_bit(Faulty, &rdev->flags))
4167			degraded++;
4168		else if (!test_bit(In_sync, &rdev->flags))
4169			/* When we can reduce the number of devices in
4170			 * an array, this might not contribute to
4171			 * 'degraded'.  It does now.
4172			 */
4173			degraded++;
4174	}
4175	rcu_read_unlock();
4176	if (conf->geo.raid_disks == conf->prev.raid_disks)
4177		return degraded;
4178	rcu_read_lock();
4179	degraded2 = 0;
4180	for (i = 0; i < conf->geo.raid_disks; i++) {
4181		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4182		if (!rdev || test_bit(Faulty, &rdev->flags))
4183			degraded2++;
4184		else if (!test_bit(In_sync, &rdev->flags)) {
4185			/* If reshape is increasing the number of devices,
4186			 * this section has already been recovered, so
4187			 * it doesn't contribute to degraded.
4188			 * else it does.
4189			 */
4190			if (conf->geo.raid_disks <= conf->prev.raid_disks)
4191				degraded2++;
4192		}
4193	}
4194	rcu_read_unlock();
4195	if (degraded2 > degraded)
4196		return degraded2;
4197	return degraded;
4198}
4199
4200static int raid10_start_reshape(struct mddev *mddev)
4201{
4202	/* A 'reshape' has been requested. This commits
4203	 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4204	 * This also checks if there are enough spares and adds them
4205	 * to the array.
4206	 * We currently require enough spares to make the final
4207	 * array non-degraded.  We also require that the difference
4208	 * between old and new data_offset - on each device - is
4209	 * enough that we never risk over-writing.
4210	 */
4211
4212	unsigned long before_length, after_length;
4213	sector_t min_offset_diff = 0;
4214	int first = 1;
4215	struct geom new;
4216	struct r10conf *conf = mddev->private;
4217	struct md_rdev *rdev;
4218	int spares = 0;
4219	int ret;
4220
4221	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4222		return -EBUSY;
4223
4224	if (setup_geo(&new, mddev, geo_start) != conf->copies)
4225		return -EINVAL;
4226
4227	before_length = ((1 << conf->prev.chunk_shift) *
4228			 conf->prev.far_copies);
4229	after_length = ((1 << conf->geo.chunk_shift) *
4230			conf->geo.far_copies);
4231
4232	rdev_for_each(rdev, mddev) {
4233		if (!test_bit(In_sync, &rdev->flags)
4234		    && !test_bit(Faulty, &rdev->flags))
4235			spares++;
4236		if (rdev->raid_disk >= 0) {
4237			long long diff = (rdev->new_data_offset
4238					  - rdev->data_offset);
4239			if (!mddev->reshape_backwards)
4240				diff = -diff;
4241			if (diff < 0)
4242				diff = 0;
4243			if (first || diff < min_offset_diff)
4244				min_offset_diff = diff;
4245			first = 0;
4246		}
4247	}
4248
4249	if (max(before_length, after_length) > min_offset_diff)
4250		return -EINVAL;
4251
4252	if (spares < mddev->delta_disks)
4253		return -EINVAL;
4254
4255	conf->offset_diff = min_offset_diff;
4256	spin_lock_irq(&conf->device_lock);
4257	if (conf->mirrors_new) {
4258		memcpy(conf->mirrors_new, conf->mirrors,
4259		       sizeof(struct raid10_info)*conf->prev.raid_disks);
4260		smp_mb();
4261		kfree(conf->mirrors_old);
4262		conf->mirrors_old = conf->mirrors;
4263		conf->mirrors = conf->mirrors_new;
4264		conf->mirrors_new = NULL;
4265	}
4266	setup_geo(&conf->geo, mddev, geo_start);
4267	smp_mb();
4268	if (mddev->reshape_backwards) {
4269		sector_t size = raid10_size(mddev, 0, 0);
4270		if (size < mddev->array_sectors) {
4271			spin_unlock_irq(&conf->device_lock);
4272			pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4273				mdname(mddev));
4274			return -EINVAL;
4275		}
4276		mddev->resync_max_sectors = size;
4277		conf->reshape_progress = size;
4278	} else
4279		conf->reshape_progress = 0;
4280	conf->reshape_safe = conf->reshape_progress;
4281	spin_unlock_irq(&conf->device_lock);
4282
4283	if (mddev->delta_disks && mddev->bitmap) {
4284		ret = bitmap_resize(mddev->bitmap,
4285				    raid10_size(mddev, 0,
4286						conf->geo.raid_disks),
4287				    0, 0);
4288		if (ret)
4289			goto abort;
4290	}
4291	if (mddev->delta_disks > 0) {
4292		rdev_for_each(rdev, mddev)
4293			if (rdev->raid_disk < 0 &&
4294			    !test_bit(Faulty, &rdev->flags)) {
4295				if (raid10_add_disk(mddev, rdev) == 0) {
4296					if (rdev->raid_disk >=
4297					    conf->prev.raid_disks)
4298						set_bit(In_sync, &rdev->flags);
4299					else
4300						rdev->recovery_offset = 0;
4301
4302					if (sysfs_link_rdev(mddev, rdev))
4303						/* Failure here  is OK */;
4304				}
4305			} else if (rdev->raid_disk >= conf->prev.raid_disks
4306				   && !test_bit(Faulty, &rdev->flags)) {
4307				/* This is a spare that was manually added */
4308				set_bit(In_sync, &rdev->flags);
4309			}
4310	}
4311	/* When a reshape changes the number of devices,
4312	 * ->degraded is measured against the larger of the
4313	 * pre and  post numbers.
4314	 */
4315	spin_lock_irq(&conf->device_lock);
4316	mddev->degraded = calc_degraded(conf);
4317	spin_unlock_irq(&conf->device_lock);
4318	mddev->raid_disks = conf->geo.raid_disks;
4319	mddev->reshape_position = conf->reshape_progress;
4320	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4321
4322	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4323	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4324	clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4325	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4326	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4327
4328	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4329						"reshape");
4330	if (!mddev->sync_thread) {
4331		ret = -EAGAIN;
4332		goto abort;
4333	}
4334	conf->reshape_checkpoint = jiffies;
4335	md_wakeup_thread(mddev->sync_thread);
4336	md_new_event(mddev);
4337	return 0;
4338
4339abort:
4340	mddev->recovery = 0;
4341	spin_lock_irq(&conf->device_lock);
4342	conf->geo = conf->prev;
4343	mddev->raid_disks = conf->geo.raid_disks;
4344	rdev_for_each(rdev, mddev)
4345		rdev->new_data_offset = rdev->data_offset;
4346	smp_wmb();
4347	conf->reshape_progress = MaxSector;
4348	conf->reshape_safe = MaxSector;
4349	mddev->reshape_position = MaxSector;
4350	spin_unlock_irq(&conf->device_lock);
4351	return ret;
4352}
4353
4354/* Calculate the last device-address that could contain
4355 * any block from the chunk that includes the array-address 's'
4356 * and report the next address.
4357 * i.e. the address returned will be chunk-aligned and after
4358 * any data that is in the chunk containing 's'.
4359 */
4360static sector_t last_dev_address(sector_t s, struct geom *geo)
4361{
4362	s = (s | geo->chunk_mask) + 1;
4363	s >>= geo->chunk_shift;
4364	s *= geo->near_copies;
4365	s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4366	s *= geo->far_copies;
4367	s <<= geo->chunk_shift;
4368	return s;
4369}
4370
4371/* Calculate the first device-address that could contain
4372 * any block from the chunk that includes the array-address 's'.
4373 * This too will be the start of a chunk
4374 */
4375static sector_t first_dev_address(sector_t s, struct geom *geo)
4376{
4377	s >>= geo->chunk_shift;
4378	s *= geo->near_copies;
4379	sector_div(s, geo->raid_disks);
4380	s *= geo->far_copies;
4381	s <<= geo->chunk_shift;
4382	return s;
4383}
4384
4385static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4386				int *skipped)
4387{
4388	/* We simply copy at most one chunk (smallest of old and new)
4389	 * at a time, possibly less if that exceeds RESYNC_PAGES,
4390	 * or we hit a bad block or something.
4391	 * This might mean we pause for normal IO in the middle of
4392	 * a chunk, but that is not a problem as mddev->reshape_position
4393	 * can record any location.
4394	 *
4395	 * If we will want to write to a location that isn't
4396	 * yet recorded as 'safe' (i.e. in metadata on disk) then
4397	 * we need to flush all reshape requests and update the metadata.
4398	 *
4399	 * When reshaping forwards (e.g. to more devices), we interpret
4400	 * 'safe' as the earliest block which might not have been copied
4401	 * down yet.  We divide this by previous stripe size and multiply
4402	 * by previous stripe length to get lowest device offset that we
4403	 * cannot write to yet.
4404	 * We interpret 'sector_nr' as an address that we want to write to.
4405	 * From this we use last_device_address() to find where we might
4406	 * write to, and first_device_address on the  'safe' position.
4407	 * If this 'next' write position is after the 'safe' position,
4408	 * we must update the metadata to increase the 'safe' position.
4409	 *
4410	 * When reshaping backwards, we round in the opposite direction
4411	 * and perform the reverse test:  next write position must not be
4412	 * less than current safe position.
4413	 *
4414	 * In all this the minimum difference in data offsets
4415	 * (conf->offset_diff - always positive) allows a bit of slack,
4416	 * so next can be after 'safe', but not by more than offset_diff
4417	 *
4418	 * We need to prepare all the bios here before we start any IO
4419	 * to ensure the size we choose is acceptable to all devices.
4420	 * The means one for each copy for write-out and an extra one for
4421	 * read-in.
4422	 * We store the read-in bio in ->master_bio and the others in
4423	 * ->devs[x].bio and ->devs[x].repl_bio.
4424	 */
4425	struct r10conf *conf = mddev->private;
4426	struct r10bio *r10_bio;
4427	sector_t next, safe, last;
4428	int max_sectors;
4429	int nr_sectors;
4430	int s;
4431	struct md_rdev *rdev;
4432	int need_flush = 0;
4433	struct bio *blist;
4434	struct bio *bio, *read_bio;
4435	int sectors_done = 0;
4436	struct page **pages;
4437
4438	if (sector_nr == 0) {
4439		/* If restarting in the middle, skip the initial sectors */
4440		if (mddev->reshape_backwards &&
4441		    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4442			sector_nr = (raid10_size(mddev, 0, 0)
4443				     - conf->reshape_progress);
4444		} else if (!mddev->reshape_backwards &&
4445			   conf->reshape_progress > 0)
4446			sector_nr = conf->reshape_progress;
4447		if (sector_nr) {
4448			mddev->curr_resync_completed = sector_nr;
4449			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4450			*skipped = 1;
4451			return sector_nr;
4452		}
4453	}
4454
4455	/* We don't use sector_nr to track where we are up to
4456	 * as that doesn't work well for ->reshape_backwards.
4457	 * So just use ->reshape_progress.
4458	 */
4459	if (mddev->reshape_backwards) {
4460		/* 'next' is the earliest device address that we might
4461		 * write to for this chunk in the new layout
4462		 */
4463		next = first_dev_address(conf->reshape_progress - 1,
4464					 &conf->geo);
4465
4466		/* 'safe' is the last device address that we might read from
4467		 * in the old layout after a restart
4468		 */
4469		safe = last_dev_address(conf->reshape_safe - 1,
4470					&conf->prev);
4471
4472		if (next + conf->offset_diff < safe)
4473			need_flush = 1;
4474
4475		last = conf->reshape_progress - 1;
4476		sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4477					       & conf->prev.chunk_mask);
4478		if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4479			sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4480	} else {
4481		/* 'next' is after the last device address that we
4482		 * might write to for this chunk in the new layout
4483		 */
4484		next = last_dev_address(conf->reshape_progress, &conf->geo);
4485
4486		/* 'safe' is the earliest device address that we might
4487		 * read from in the old layout after a restart
4488		 */
4489		safe = first_dev_address(conf->reshape_safe, &conf->prev);
4490
4491		/* Need to update metadata if 'next' might be beyond 'safe'
4492		 * as that would possibly corrupt data
4493		 */
4494		if (next > safe + conf->offset_diff)
4495			need_flush = 1;
4496
4497		sector_nr = conf->reshape_progress;
4498		last  = sector_nr | (conf->geo.chunk_mask
4499				     & conf->prev.chunk_mask);
4500
4501		if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4502			last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4503	}
4504
4505	if (need_flush ||
4506	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4507		/* Need to update reshape_position in metadata */
4508		wait_barrier(conf);
4509		mddev->reshape_position = conf->reshape_progress;
4510		if (mddev->reshape_backwards)
4511			mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4512				- conf->reshape_progress;
4513		else
4514			mddev->curr_resync_completed = conf->reshape_progress;
4515		conf->reshape_checkpoint = jiffies;
4516		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4517		md_wakeup_thread(mddev->thread);
4518		wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4519			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4520		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4521			allow_barrier(conf);
4522			return sectors_done;
4523		}
4524		conf->reshape_safe = mddev->reshape_position;
4525		allow_barrier(conf);
4526	}
4527
4528read_more:
4529	/* Now schedule reads for blocks from sector_nr to last */
4530	r10_bio = raid10_alloc_init_r10buf(conf);
4531	r10_bio->state = 0;
4532	raise_barrier(conf, sectors_done != 0);
4533	atomic_set(&r10_bio->remaining, 0);
4534	r10_bio->mddev = mddev;
4535	r10_bio->sector = sector_nr;
4536	set_bit(R10BIO_IsReshape, &r10_bio->state);
4537	r10_bio->sectors = last - sector_nr + 1;
4538	rdev = read_balance(conf, r10_bio, &max_sectors);
4539	BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4540
4541	if (!rdev) {
4542		/* Cannot read from here, so need to record bad blocks
4543		 * on all the target devices.
4544		 */
4545		// FIXME
4546		mempool_free(r10_bio, conf->r10buf_pool);
4547		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4548		return sectors_done;
4549	}
4550
4551	read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4552
4553	bio_set_dev(read_bio, rdev->bdev);
4554	read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4555			       + rdev->data_offset);
4556	read_bio->bi_private = r10_bio;
4557	read_bio->bi_end_io = end_reshape_read;
4558	bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4559	read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4560	read_bio->bi_status = 0;
4561	read_bio->bi_vcnt = 0;
4562	read_bio->bi_iter.bi_size = 0;
4563	r10_bio->master_bio = read_bio;
4564	r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4565
4566	/* Now find the locations in the new layout */
4567	__raid10_find_phys(&conf->geo, r10_bio);
4568
4569	blist = read_bio;
4570	read_bio->bi_next = NULL;
4571
4572	rcu_read_lock();
4573	for (s = 0; s < conf->copies*2; s++) {
4574		struct bio *b;
4575		int d = r10_bio->devs[s/2].devnum;
4576		struct md_rdev *rdev2;
4577		if (s&1) {
4578			rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4579			b = r10_bio->devs[s/2].repl_bio;
4580		} else {
4581			rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4582			b = r10_bio->devs[s/2].bio;
4583		}
4584		if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4585			continue;
4586
4587		bio_set_dev(b, rdev2->bdev);
 
4588		b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4589			rdev2->new_data_offset;
 
4590		b->bi_end_io = end_reshape_write;
4591		bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4592		b->bi_next = blist;
4593		blist = b;
4594	}
4595
4596	/* Now add as many pages as possible to all of these bios. */
4597
4598	nr_sectors = 0;
4599	pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4600	for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4601		struct page *page = pages[s / (PAGE_SIZE >> 9)];
4602		int len = (max_sectors - s) << 9;
4603		if (len > PAGE_SIZE)
4604			len = PAGE_SIZE;
4605		for (bio = blist; bio ; bio = bio->bi_next) {
4606			/*
4607			 * won't fail because the vec table is big enough
4608			 * to hold all these pages
4609			 */
4610			bio_add_page(bio, page, len, 0);
 
 
 
 
 
 
 
 
 
4611		}
4612		sector_nr += len >> 9;
4613		nr_sectors += len >> 9;
4614	}
4615	rcu_read_unlock();
4616	r10_bio->sectors = nr_sectors;
4617
4618	/* Now submit the read */
4619	md_sync_acct_bio(read_bio, r10_bio->sectors);
4620	atomic_inc(&r10_bio->remaining);
4621	read_bio->bi_next = NULL;
4622	generic_make_request(read_bio);
4623	sector_nr += nr_sectors;
4624	sectors_done += nr_sectors;
4625	if (sector_nr <= last)
4626		goto read_more;
4627
4628	/* Now that we have done the whole section we can
4629	 * update reshape_progress
4630	 */
4631	if (mddev->reshape_backwards)
4632		conf->reshape_progress -= sectors_done;
4633	else
4634		conf->reshape_progress += sectors_done;
4635
4636	return sectors_done;
4637}
4638
4639static void end_reshape_request(struct r10bio *r10_bio);
4640static int handle_reshape_read_error(struct mddev *mddev,
4641				     struct r10bio *r10_bio);
4642static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4643{
4644	/* Reshape read completed.  Hopefully we have a block
4645	 * to write out.
4646	 * If we got a read error then we do sync 1-page reads from
4647	 * elsewhere until we find the data - or give up.
4648	 */
4649	struct r10conf *conf = mddev->private;
4650	int s;
4651
4652	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4653		if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4654			/* Reshape has been aborted */
4655			md_done_sync(mddev, r10_bio->sectors, 0);
4656			return;
4657		}
4658
4659	/* We definitely have the data in the pages, schedule the
4660	 * writes.
4661	 */
4662	atomic_set(&r10_bio->remaining, 1);
4663	for (s = 0; s < conf->copies*2; s++) {
4664		struct bio *b;
4665		int d = r10_bio->devs[s/2].devnum;
4666		struct md_rdev *rdev;
4667		rcu_read_lock();
4668		if (s&1) {
4669			rdev = rcu_dereference(conf->mirrors[d].replacement);
4670			b = r10_bio->devs[s/2].repl_bio;
4671		} else {
4672			rdev = rcu_dereference(conf->mirrors[d].rdev);
4673			b = r10_bio->devs[s/2].bio;
4674		}
4675		if (!rdev || test_bit(Faulty, &rdev->flags)) {
4676			rcu_read_unlock();
4677			continue;
4678		}
4679		atomic_inc(&rdev->nr_pending);
4680		rcu_read_unlock();
4681		md_sync_acct_bio(b, r10_bio->sectors);
4682		atomic_inc(&r10_bio->remaining);
4683		b->bi_next = NULL;
4684		generic_make_request(b);
4685	}
4686	end_reshape_request(r10_bio);
4687}
4688
4689static void end_reshape(struct r10conf *conf)
4690{
4691	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4692		return;
4693
4694	spin_lock_irq(&conf->device_lock);
4695	conf->prev = conf->geo;
4696	md_finish_reshape(conf->mddev);
4697	smp_wmb();
4698	conf->reshape_progress = MaxSector;
4699	conf->reshape_safe = MaxSector;
4700	spin_unlock_irq(&conf->device_lock);
4701
4702	/* read-ahead size must cover two whole stripes, which is
4703	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4704	 */
4705	if (conf->mddev->queue) {
4706		int stripe = conf->geo.raid_disks *
4707			((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4708		stripe /= conf->geo.near_copies;
4709		if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4710			conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4711	}
4712	conf->fullsync = 0;
4713}
4714
 
4715static int handle_reshape_read_error(struct mddev *mddev,
4716				     struct r10bio *r10_bio)
4717{
4718	/* Use sync reads to get the blocks from somewhere else */
4719	int sectors = r10_bio->sectors;
4720	struct r10conf *conf = mddev->private;
4721	struct r10bio *r10b;
 
 
 
 
4722	int slot = 0;
4723	int idx = 0;
4724	struct page **pages;
4725
4726	r10b = kmalloc(sizeof(*r10b) +
4727	       sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4728	if (!r10b) {
4729		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4730		return -ENOMEM;
4731	}
4732
4733	/* reshape IOs share pages from .devs[0].bio */
4734	pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4735
4736	r10b->sector = r10_bio->sector;
4737	__raid10_find_phys(&conf->prev, r10b);
4738
4739	while (sectors) {
4740		int s = sectors;
4741		int success = 0;
4742		int first_slot = slot;
4743
4744		if (s > (PAGE_SIZE >> 9))
4745			s = PAGE_SIZE >> 9;
4746
4747		rcu_read_lock();
4748		while (!success) {
4749			int d = r10b->devs[slot].devnum;
4750			struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4751			sector_t addr;
4752			if (rdev == NULL ||
4753			    test_bit(Faulty, &rdev->flags) ||
4754			    !test_bit(In_sync, &rdev->flags))
4755				goto failed;
4756
4757			addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4758			atomic_inc(&rdev->nr_pending);
4759			rcu_read_unlock();
4760			success = sync_page_io(rdev,
4761					       addr,
4762					       s << 9,
4763					       pages[idx],
4764					       REQ_OP_READ, 0, false);
4765			rdev_dec_pending(rdev, mddev);
4766			rcu_read_lock();
4767			if (success)
4768				break;
4769		failed:
4770			slot++;
4771			if (slot >= conf->copies)
4772				slot = 0;
4773			if (slot == first_slot)
4774				break;
4775		}
4776		rcu_read_unlock();
4777		if (!success) {
4778			/* couldn't read this block, must give up */
4779			set_bit(MD_RECOVERY_INTR,
4780				&mddev->recovery);
4781			kfree(r10b);
4782			return -EIO;
4783		}
4784		sectors -= s;
4785		idx++;
4786	}
4787	kfree(r10b);
4788	return 0;
4789}
4790
4791static void end_reshape_write(struct bio *bio)
4792{
4793	struct r10bio *r10_bio = get_resync_r10bio(bio);
 
4794	struct mddev *mddev = r10_bio->mddev;
4795	struct r10conf *conf = mddev->private;
4796	int d;
4797	int slot;
4798	int repl;
4799	struct md_rdev *rdev = NULL;
4800
4801	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4802	if (repl)
4803		rdev = conf->mirrors[d].replacement;
4804	if (!rdev) {
4805		smp_mb();
4806		rdev = conf->mirrors[d].rdev;
4807	}
4808
4809	if (bio->bi_status) {
4810		/* FIXME should record badblock */
4811		md_error(mddev, rdev);
4812	}
4813
4814	rdev_dec_pending(rdev, mddev);
4815	end_reshape_request(r10_bio);
4816}
4817
4818static void end_reshape_request(struct r10bio *r10_bio)
4819{
4820	if (!atomic_dec_and_test(&r10_bio->remaining))
4821		return;
4822	md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4823	bio_put(r10_bio->master_bio);
4824	put_buf(r10_bio);
4825}
4826
4827static void raid10_finish_reshape(struct mddev *mddev)
4828{
4829	struct r10conf *conf = mddev->private;
4830
4831	if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4832		return;
4833
4834	if (mddev->delta_disks > 0) {
 
 
4835		if (mddev->recovery_cp > mddev->resync_max_sectors) {
4836			mddev->recovery_cp = mddev->resync_max_sectors;
4837			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4838		}
4839		mddev->resync_max_sectors = mddev->array_sectors;
 
 
4840	} else {
4841		int d;
4842		rcu_read_lock();
4843		for (d = conf->geo.raid_disks ;
4844		     d < conf->geo.raid_disks - mddev->delta_disks;
4845		     d++) {
4846			struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4847			if (rdev)
4848				clear_bit(In_sync, &rdev->flags);
4849			rdev = rcu_dereference(conf->mirrors[d].replacement);
4850			if (rdev)
4851				clear_bit(In_sync, &rdev->flags);
4852		}
4853		rcu_read_unlock();
4854	}
4855	mddev->layout = mddev->new_layout;
4856	mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4857	mddev->reshape_position = MaxSector;
4858	mddev->delta_disks = 0;
4859	mddev->reshape_backwards = 0;
4860}
4861
4862static struct md_personality raid10_personality =
4863{
4864	.name		= "raid10",
4865	.level		= 10,
4866	.owner		= THIS_MODULE,
4867	.make_request	= raid10_make_request,
4868	.run		= raid10_run,
4869	.free		= raid10_free,
4870	.status		= raid10_status,
4871	.error_handler	= raid10_error,
4872	.hot_add_disk	= raid10_add_disk,
4873	.hot_remove_disk= raid10_remove_disk,
4874	.spare_active	= raid10_spare_active,
4875	.sync_request	= raid10_sync_request,
4876	.quiesce	= raid10_quiesce,
4877	.size		= raid10_size,
4878	.resize		= raid10_resize,
4879	.takeover	= raid10_takeover,
4880	.check_reshape	= raid10_check_reshape,
4881	.start_reshape	= raid10_start_reshape,
4882	.finish_reshape	= raid10_finish_reshape,
4883	.congested	= raid10_congested,
4884};
4885
4886static int __init raid_init(void)
4887{
4888	return register_md_personality(&raid10_personality);
4889}
4890
4891static void raid_exit(void)
4892{
4893	unregister_md_personality(&raid10_personality);
4894}
4895
4896module_init(raid_init);
4897module_exit(raid_exit);
4898MODULE_LICENSE("GPL");
4899MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4900MODULE_ALIAS("md-personality-9"); /* RAID10 */
4901MODULE_ALIAS("md-raid10");
4902MODULE_ALIAS("md-level-10");
4903
4904module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);