Linux Audio

Check our new training course

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