raid5.c

   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * raid5.c : Multiple Devices driver for Linux
   4 *	   Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   5 *	   Copyright (C) 1999, 2000 Ingo Molnar
   6 *	   Copyright (C) 2002, 2003 H. Peter Anvin
   7 *
   8 * RAID-4/5/6 management functions.
   9 * Thanks to Penguin Computing for making the RAID-6 development possible
  10 * by donating a test server!
  11 */
  12
  13/*
  14 * BITMAP UNPLUGGING:
  15 *
  16 * The sequencing for updating the bitmap reliably is a little
  17 * subtle (and I got it wrong the first time) so it deserves some
  18 * explanation.
  19 *
  20 * We group bitmap updates into batches.  Each batch has a number.
  21 * We may write out several batches at once, but that isn't very important.
  22 * conf->seq_write is the number of the last batch successfully written.
  23 * conf->seq_flush is the number of the last batch that was closed to
  24 *    new additions.
  25 * When we discover that we will need to write to any block in a stripe
  26 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  27 * the number of the batch it will be in. This is seq_flush+1.
  28 * When we are ready to do a write, if that batch hasn't been written yet,
  29 *   we plug the array and queue the stripe for later.
  30 * When an unplug happens, we increment bm_flush, thus closing the current
  31 *   batch.
  32 * When we notice that bm_flush > bm_write, we write out all pending updates
  33 * to the bitmap, and advance bm_write to where bm_flush was.
  34 * This may occasionally write a bit out twice, but is sure never to
  35 * miss any bits.
  36 */
  37
  38#include <linux/blkdev.h>
  39#include <linux/kthread.h>
  40#include <linux/raid/pq.h>
  41#include <linux/async_tx.h>
  42#include <linux/module.h>
  43#include <linux/async.h>
  44#include <linux/seq_file.h>
  45#include <linux/cpu.h>
  46#include <linux/slab.h>
  47#include <linux/ratelimit.h>
  48#include <linux/nodemask.h>
  49
  50#include <trace/events/block.h>
  51#include <linux/list_sort.h>
  52
  53#include "md.h"
  54#include "raid5.h"
  55#include "raid0.h"
  56#include "md-bitmap.h"
  57#include "raid5-log.h"
  58
  59#define UNSUPPORTED_MDDEV_FLAGS	(1L << MD_FAILFAST_SUPPORTED)
  60
  61#define cpu_to_group(cpu) cpu_to_node(cpu)
  62#define ANY_GROUP NUMA_NO_NODE
  63
  64#define RAID5_MAX_REQ_STRIPES 256
  65
  66static bool devices_handle_discard_safely = false;
  67module_param(devices_handle_discard_safely, bool, 0644);
  68MODULE_PARM_DESC(devices_handle_discard_safely,
  69		 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  70static struct workqueue_struct *raid5_wq;
  71
  72static void raid5_quiesce(struct mddev *mddev, int quiesce);
  73
  74static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  75{
  76	int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
  77	return &conf->stripe_hashtbl[hash];
  78}
  79
  80static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
  81{
  82	return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
  83}
  84
  85static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  86	__acquires(&conf->device_lock)
  87{
  88	spin_lock_irq(conf->hash_locks + hash);
  89	spin_lock(&conf->device_lock);
  90}
  91
  92static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  93	__releases(&conf->device_lock)
  94{
  95	spin_unlock(&conf->device_lock);
  96	spin_unlock_irq(conf->hash_locks + hash);
  97}
  98
  99static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 100	__acquires(&conf->device_lock)
 101{
 102	int i;
 103	spin_lock_irq(conf->hash_locks);
 104	for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 105		spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 106	spin_lock(&conf->device_lock);
 107}
 108
 109static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 110	__releases(&conf->device_lock)
 111{
 112	int i;
 113	spin_unlock(&conf->device_lock);
 114	for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 115		spin_unlock(conf->hash_locks + i);
 116	spin_unlock_irq(conf->hash_locks);
 117}
 118
 119/* Find first data disk in a raid6 stripe */
 120static inline int raid6_d0(struct stripe_head *sh)
 121{
 122	if (sh->ddf_layout)
 123		/* ddf always start from first device */
 124		return 0;
 125	/* md starts just after Q block */
 126	if (sh->qd_idx == sh->disks - 1)
 127		return 0;
 128	else
 129		return sh->qd_idx + 1;
 130}
 131static inline int raid6_next_disk(int disk, int raid_disks)
 132{
 133	disk++;
 134	return (disk < raid_disks) ? disk : 0;
 135}
 136
 137/* When walking through the disks in a raid5, starting at raid6_d0,
 138 * We need to map each disk to a 'slot', where the data disks are slot
 139 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 140 * is raid_disks-1.  This help does that mapping.
 141 */
 142static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 143			     int *count, int syndrome_disks)
 144{
 145	int slot = *count;
 146
 147	if (sh->ddf_layout)
 148		(*count)++;
 149	if (idx == sh->pd_idx)
 150		return syndrome_disks;
 151	if (idx == sh->qd_idx)
 152		return syndrome_disks + 1;
 153	if (!sh->ddf_layout)
 154		(*count)++;
 155	return slot;
 156}
 157
 158static void print_raid5_conf(struct r5conf *conf);
 159
 160static int stripe_operations_active(struct stripe_head *sh)
 161{
 162	return sh->check_state || sh->reconstruct_state ||
 163	       test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 164	       test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 165}
 166
 167static bool stripe_is_lowprio(struct stripe_head *sh)
 168{
 169	return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 170		test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 171	       !test_bit(STRIPE_R5C_CACHING, &sh->state);
 172}
 173
 174static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 175	__must_hold(&sh->raid_conf->device_lock)
 176{
 177	struct r5conf *conf = sh->raid_conf;
 178	struct r5worker_group *group;
 179	int thread_cnt;
 180	int i, cpu = sh->cpu;
 181
 182	if (!cpu_online(cpu)) {
 183		cpu = cpumask_any(cpu_online_mask);
 184		sh->cpu = cpu;
 185	}
 186
 187	if (list_empty(&sh->lru)) {
 188		struct r5worker_group *group;
 189		group = conf->worker_groups + cpu_to_group(cpu);
 190		if (stripe_is_lowprio(sh))
 191			list_add_tail(&sh->lru, &group->loprio_list);
 192		else
 193			list_add_tail(&sh->lru, &group->handle_list);
 194		group->stripes_cnt++;
 195		sh->group = group;
 196	}
 197
 198	if (conf->worker_cnt_per_group == 0) {
 199		md_wakeup_thread(conf->mddev->thread);
 200		return;
 201	}
 202
 203	group = conf->worker_groups + cpu_to_group(sh->cpu);
 204
 205	group->workers[0].working = true;
 206	/* at least one worker should run to avoid race */
 207	queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 208
 209	thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 210	/* wakeup more workers */
 211	for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 212		if (group->workers[i].working == false) {
 213			group->workers[i].working = true;
 214			queue_work_on(sh->cpu, raid5_wq,
 215				      &group->workers[i].work);
 216			thread_cnt--;
 217		}
 218	}
 219}
 220
 221static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 222			      struct list_head *temp_inactive_list)
 223	__must_hold(&conf->device_lock)
 224{
 225	int i;
 226	int injournal = 0;	/* number of date pages with R5_InJournal */
 227
 228	BUG_ON(!list_empty(&sh->lru));
 229	BUG_ON(atomic_read(&conf->active_stripes)==0);
 230
 231	if (r5c_is_writeback(conf->log))
 232		for (i = sh->disks; i--; )
 233			if (test_bit(R5_InJournal, &sh->dev[i].flags))
 234				injournal++;
 235	/*
 236	 * In the following cases, the stripe cannot be released to cached
 237	 * lists. Therefore, we make the stripe write out and set
 238	 * STRIPE_HANDLE:
 239	 *   1. when quiesce in r5c write back;
 240	 *   2. when resync is requested fot the stripe.
 241	 */
 242	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 243	    (conf->quiesce && r5c_is_writeback(conf->log) &&
 244	     !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 245		if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 246			r5c_make_stripe_write_out(sh);
 247		set_bit(STRIPE_HANDLE, &sh->state);
 248	}
 249
 250	if (test_bit(STRIPE_HANDLE, &sh->state)) {
 251		if (test_bit(STRIPE_DELAYED, &sh->state) &&
 252		    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 253			list_add_tail(&sh->lru, &conf->delayed_list);
 254		else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 255			   sh->bm_seq - conf->seq_write > 0)
 256			list_add_tail(&sh->lru, &conf->bitmap_list);
 257		else {
 258			clear_bit(STRIPE_DELAYED, &sh->state);
 259			clear_bit(STRIPE_BIT_DELAY, &sh->state);
 260			if (conf->worker_cnt_per_group == 0) {
 261				if (stripe_is_lowprio(sh))
 262					list_add_tail(&sh->lru,
 263							&conf->loprio_list);
 264				else
 265					list_add_tail(&sh->lru,
 266							&conf->handle_list);
 267			} else {
 268				raid5_wakeup_stripe_thread(sh);
 269				return;
 270			}
 271		}
 272		md_wakeup_thread(conf->mddev->thread);
 273	} else {
 274		BUG_ON(stripe_operations_active(sh));
 275		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 276			if (atomic_dec_return(&conf->preread_active_stripes)
 277			    < IO_THRESHOLD)
 278				md_wakeup_thread(conf->mddev->thread);
 279		atomic_dec(&conf->active_stripes);
 280		if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 281			if (!r5c_is_writeback(conf->log))
 282				list_add_tail(&sh->lru, temp_inactive_list);
 283			else {
 284				WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 285				if (injournal == 0)
 286					list_add_tail(&sh->lru, temp_inactive_list);
 287				else if (injournal == conf->raid_disks - conf->max_degraded) {
 288					/* full stripe */
 289					if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 290						atomic_inc(&conf->r5c_cached_full_stripes);
 291					if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 292						atomic_dec(&conf->r5c_cached_partial_stripes);
 293					list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 294					r5c_check_cached_full_stripe(conf);
 295				} else
 296					/*
 297					 * STRIPE_R5C_PARTIAL_STRIPE is set in
 298					 * r5c_try_caching_write(). No need to
 299					 * set it again.
 300					 */
 301					list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 302			}
 303		}
 304	}
 305}
 306
 307static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 308			     struct list_head *temp_inactive_list)
 309	__must_hold(&conf->device_lock)
 310{
 311	if (atomic_dec_and_test(&sh->count))
 312		do_release_stripe(conf, sh, temp_inactive_list);
 313}
 314
 315/*
 316 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 317 *
 318 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 319 * given time. Adding stripes only takes device lock, while deleting stripes
 320 * only takes hash lock.
 321 */
 322static void release_inactive_stripe_list(struct r5conf *conf,
 323					 struct list_head *temp_inactive_list,
 324					 int hash)
 325{
 326	int size;
 327	bool do_wakeup = false;
 328	unsigned long flags;
 329
 330	if (hash == NR_STRIPE_HASH_LOCKS) {
 331		size = NR_STRIPE_HASH_LOCKS;
 332		hash = NR_STRIPE_HASH_LOCKS - 1;
 333	} else
 334		size = 1;
 335	while (size) {
 336		struct list_head *list = &temp_inactive_list[size - 1];
 337
 338		/*
 339		 * We don't hold any lock here yet, raid5_get_active_stripe() might
 340		 * remove stripes from the list
 341		 */
 342		if (!list_empty_careful(list)) {
 343			spin_lock_irqsave(conf->hash_locks + hash, flags);
 344			if (list_empty(conf->inactive_list + hash) &&
 345			    !list_empty(list))
 346				atomic_dec(&conf->empty_inactive_list_nr);
 347			list_splice_tail_init(list, conf->inactive_list + hash);
 348			do_wakeup = true;
 349			spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 350		}
 351		size--;
 352		hash--;
 353	}
 354
 355	if (do_wakeup) {
 356		wake_up(&conf->wait_for_stripe);
 357		if (atomic_read(&conf->active_stripes) == 0)
 358			wake_up(&conf->wait_for_quiescent);
 359		if (conf->retry_read_aligned)
 360			md_wakeup_thread(conf->mddev->thread);
 361	}
 362}
 363
 364static int release_stripe_list(struct r5conf *conf,
 365			       struct list_head *temp_inactive_list)
 366	__must_hold(&conf->device_lock)
 367{
 368	struct stripe_head *sh, *t;
 369	int count = 0;
 370	struct llist_node *head;
 371
 372	head = llist_del_all(&conf->released_stripes);
 373	head = llist_reverse_order(head);
 374	llist_for_each_entry_safe(sh, t, head, release_list) {
 375		int hash;
 376
 377		/* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 378		smp_mb();
 379		clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 380		/*
 381		 * Don't worry the bit is set here, because if the bit is set
 382		 * again, the count is always > 1. This is true for
 383		 * STRIPE_ON_UNPLUG_LIST bit too.
 384		 */
 385		hash = sh->hash_lock_index;
 386		__release_stripe(conf, sh, &temp_inactive_list[hash]);
 387		count++;
 388	}
 389
 390	return count;
 391}
 392
 393void raid5_release_stripe(struct stripe_head *sh)
 394{
 395	struct r5conf *conf = sh->raid_conf;
 396	unsigned long flags;
 397	struct list_head list;
 398	int hash;
 399	bool wakeup;
 400
 401	/* Avoid release_list until the last reference.
 402	 */
 403	if (atomic_add_unless(&sh->count, -1, 1))
 404		return;
 405
 406	if (unlikely(!conf->mddev->thread) ||
 407		test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 408		goto slow_path;
 409	wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 410	if (wakeup)
 411		md_wakeup_thread(conf->mddev->thread);
 412	return;
 413slow_path:
 414	/* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 415	if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 416		INIT_LIST_HEAD(&list);
 417		hash = sh->hash_lock_index;
 418		do_release_stripe(conf, sh, &list);
 419		spin_unlock_irqrestore(&conf->device_lock, flags);
 420		release_inactive_stripe_list(conf, &list, hash);
 421	}
 422}
 423
 424static inline void remove_hash(struct stripe_head *sh)
 425{
 426	pr_debug("remove_hash(), stripe %llu\n",
 427		(unsigned long long)sh->sector);
 428
 429	hlist_del_init(&sh->hash);
 430}
 431
 432static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 433{
 434	struct hlist_head *hp = stripe_hash(conf, sh->sector);
 435
 436	pr_debug("insert_hash(), stripe %llu\n",
 437		(unsigned long long)sh->sector);
 438
 439	hlist_add_head(&sh->hash, hp);
 440}
 441
 442/* find an idle stripe, make sure it is unhashed, and return it. */
 443static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 444{
 445	struct stripe_head *sh = NULL;
 446	struct list_head *first;
 447
 448	if (list_empty(conf->inactive_list + hash))
 449		goto out;
 450	first = (conf->inactive_list + hash)->next;
 451	sh = list_entry(first, struct stripe_head, lru);
 452	list_del_init(first);
 453	remove_hash(sh);
 454	atomic_inc(&conf->active_stripes);
 455	BUG_ON(hash != sh->hash_lock_index);
 456	if (list_empty(conf->inactive_list + hash))
 457		atomic_inc(&conf->empty_inactive_list_nr);
 458out:
 459	return sh;
 460}
 461
 462#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
 463static void free_stripe_pages(struct stripe_head *sh)
 464{
 465	int i;
 466	struct page *p;
 467
 468	/* Have not allocate page pool */
 469	if (!sh->pages)
 470		return;
 471
 472	for (i = 0; i < sh->nr_pages; i++) {
 473		p = sh->pages[i];
 474		if (p)
 475			put_page(p);
 476		sh->pages[i] = NULL;
 477	}
 478}
 479
 480static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
 481{
 482	int i;
 483	struct page *p;
 484
 485	for (i = 0; i < sh->nr_pages; i++) {
 486		/* The page have allocated. */
 487		if (sh->pages[i])
 488			continue;
 489
 490		p = alloc_page(gfp);
 491		if (!p) {
 492			free_stripe_pages(sh);
 493			return -ENOMEM;
 494		}
 495		sh->pages[i] = p;
 496	}
 497	return 0;
 498}
 499
 500static int
 501init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
 502{
 503	int nr_pages, cnt;
 504
 505	if (sh->pages)
 506		return 0;
 507
 508	/* Each of the sh->dev[i] need one conf->stripe_size */
 509	cnt = PAGE_SIZE / conf->stripe_size;
 510	nr_pages = (disks + cnt - 1) / cnt;
 511
 512	sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 513	if (!sh->pages)
 514		return -ENOMEM;
 515	sh->nr_pages = nr_pages;
 516	sh->stripes_per_page = cnt;
 517	return 0;
 518}
 519#endif
 520
 521static void shrink_buffers(struct stripe_head *sh)
 522{
 523	int i;
 524	int num = sh->raid_conf->pool_size;
 525
 526#if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 527	for (i = 0; i < num ; i++) {
 528		struct page *p;
 529
 530		WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 531		p = sh->dev[i].page;
 532		if (!p)
 533			continue;
 534		sh->dev[i].page = NULL;
 535		put_page(p);
 536	}
 537#else
 538	for (i = 0; i < num; i++)
 539		sh->dev[i].page = NULL;
 540	free_stripe_pages(sh); /* Free pages */
 541#endif
 542}
 543
 544static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 545{
 546	int i;
 547	int num = sh->raid_conf->pool_size;
 548
 549#if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 550	for (i = 0; i < num; i++) {
 551		struct page *page;
 552
 553		if (!(page = alloc_page(gfp))) {
 554			return 1;
 555		}
 556		sh->dev[i].page = page;
 557		sh->dev[i].orig_page = page;
 558		sh->dev[i].offset = 0;
 559	}
 560#else
 561	if (alloc_stripe_pages(sh, gfp))
 562		return -ENOMEM;
 563
 564	for (i = 0; i < num; i++) {
 565		sh->dev[i].page = raid5_get_dev_page(sh, i);
 566		sh->dev[i].orig_page = sh->dev[i].page;
 567		sh->dev[i].offset = raid5_get_page_offset(sh, i);
 568	}
 569#endif
 570	return 0;
 571}
 572
 573static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 574			    struct stripe_head *sh);
 575
 576static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 577{
 578	struct r5conf *conf = sh->raid_conf;
 579	int i, seq;
 580
 581	BUG_ON(atomic_read(&sh->count) != 0);
 582	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 583	BUG_ON(stripe_operations_active(sh));
 584	BUG_ON(sh->batch_head);
 585
 586	pr_debug("init_stripe called, stripe %llu\n",
 587		(unsigned long long)sector);
 588retry:
 589	seq = read_seqcount_begin(&conf->gen_lock);
 590	sh->generation = conf->generation - previous;
 591	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 592	sh->sector = sector;
 593	stripe_set_idx(sector, conf, previous, sh);
 594	sh->state = 0;
 595
 596	for (i = sh->disks; i--; ) {
 597		struct r5dev *dev = &sh->dev[i];
 598
 599		if (dev->toread || dev->read || dev->towrite || dev->written ||
 600		    test_bit(R5_LOCKED, &dev->flags)) {
 601			pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 602			       (unsigned long long)sh->sector, i, dev->toread,
 603			       dev->read, dev->towrite, dev->written,
 604			       test_bit(R5_LOCKED, &dev->flags));
 605			WARN_ON(1);
 606		}
 607		dev->flags = 0;
 608		dev->sector = raid5_compute_blocknr(sh, i, previous);
 609	}
 610	if (read_seqcount_retry(&conf->gen_lock, seq))
 611		goto retry;
 612	sh->overwrite_disks = 0;
 613	insert_hash(conf, sh);
 614	sh->cpu = smp_processor_id();
 615	set_bit(STRIPE_BATCH_READY, &sh->state);
 616}
 617
 618static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 619					 short generation)
 620{
 621	struct stripe_head *sh;
 622
 623	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 624	hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 625		if (sh->sector == sector && sh->generation == generation)
 626			return sh;
 627	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 628	return NULL;
 629}
 630
 631static struct stripe_head *find_get_stripe(struct r5conf *conf,
 632		sector_t sector, short generation, int hash)
 633{
 634	int inc_empty_inactive_list_flag;
 635	struct stripe_head *sh;
 636
 637	sh = __find_stripe(conf, sector, generation);
 638	if (!sh)
 639		return NULL;
 640
 641	if (atomic_inc_not_zero(&sh->count))
 642		return sh;
 643
 644	/*
 645	 * Slow path. The reference count is zero which means the stripe must
 646	 * be on a list (sh->lru). Must remove the stripe from the list that
 647	 * references it with the device_lock held.
 648	 */
 649
 650	spin_lock(&conf->device_lock);
 651	if (!atomic_read(&sh->count)) {
 652		if (!test_bit(STRIPE_HANDLE, &sh->state))
 653			atomic_inc(&conf->active_stripes);
 654		BUG_ON(list_empty(&sh->lru) &&
 655		       !test_bit(STRIPE_EXPANDING, &sh->state));
 656		inc_empty_inactive_list_flag = 0;
 657		if (!list_empty(conf->inactive_list + hash))
 658			inc_empty_inactive_list_flag = 1;
 659		list_del_init(&sh->lru);
 660		if (list_empty(conf->inactive_list + hash) &&
 661		    inc_empty_inactive_list_flag)
 662			atomic_inc(&conf->empty_inactive_list_nr);
 663		if (sh->group) {
 664			sh->group->stripes_cnt--;
 665			sh->group = NULL;
 666		}
 667	}
 668	atomic_inc(&sh->count);
 669	spin_unlock(&conf->device_lock);
 670
 671	return sh;
 672}
 673
 674/*
 675 * Need to check if array has failed when deciding whether to:
 676 *  - start an array
 677 *  - remove non-faulty devices
 678 *  - add a spare
 679 *  - allow a reshape
 680 * This determination is simple when no reshape is happening.
 681 * However if there is a reshape, we need to carefully check
 682 * both the before and after sections.
 683 * This is because some failed devices may only affect one
 684 * of the two sections, and some non-in_sync devices may
 685 * be insync in the section most affected by failed devices.
 686 *
 687 * Most calls to this function hold &conf->device_lock. Calls
 688 * in raid5_run() do not require the lock as no other threads
 689 * have been started yet.
 690 */
 691int raid5_calc_degraded(struct r5conf *conf)
 692{
 693	int degraded, degraded2;
 694	int i;
 695
 696	degraded = 0;
 697	for (i = 0; i < conf->previous_raid_disks; i++) {
 698		struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
 699
 700		if (rdev && test_bit(Faulty, &rdev->flags))
 701			rdev = READ_ONCE(conf->disks[i].replacement);
 702		if (!rdev || test_bit(Faulty, &rdev->flags))
 703			degraded++;
 704		else if (test_bit(In_sync, &rdev->flags))
 705			;
 706		else
 707			/* not in-sync or faulty.
 708			 * If the reshape increases the number of devices,
 709			 * this is being recovered by the reshape, so
 710			 * this 'previous' section is not in_sync.
 711			 * If the number of devices is being reduced however,
 712			 * the device can only be part of the array if
 713			 * we are reverting a reshape, so this section will
 714			 * be in-sync.
 715			 */
 716			if (conf->raid_disks >= conf->previous_raid_disks)
 717				degraded++;
 718	}
 719	if (conf->raid_disks == conf->previous_raid_disks)
 720		return degraded;
 721	degraded2 = 0;
 722	for (i = 0; i < conf->raid_disks; i++) {
 723		struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
 724
 725		if (rdev && test_bit(Faulty, &rdev->flags))
 726			rdev = READ_ONCE(conf->disks[i].replacement);
 727		if (!rdev || test_bit(Faulty, &rdev->flags))
 728			degraded2++;
 729		else if (test_bit(In_sync, &rdev->flags))
 730			;
 731		else
 732			/* not in-sync or faulty.
 733			 * If reshape increases the number of devices, this
 734			 * section has already been recovered, else it
 735			 * almost certainly hasn't.
 736			 */
 737			if (conf->raid_disks <= conf->previous_raid_disks)
 738				degraded2++;
 739	}
 740	if (degraded2 > degraded)
 741		return degraded2;
 742	return degraded;
 743}
 744
 745static bool has_failed(struct r5conf *conf)
 746{
 747	int degraded = conf->mddev->degraded;
 748
 749	if (test_bit(MD_BROKEN, &conf->mddev->flags))
 750		return true;
 751
 752	if (conf->mddev->reshape_position != MaxSector)
 753		degraded = raid5_calc_degraded(conf);
 754
 755	return degraded > conf->max_degraded;
 756}
 757
 758enum stripe_result {
 759	STRIPE_SUCCESS = 0,
 760	STRIPE_RETRY,
 761	STRIPE_SCHEDULE_AND_RETRY,
 762	STRIPE_FAIL,
 763	STRIPE_WAIT_RESHAPE,
 764};
 765
 766struct stripe_request_ctx {
 767	/* a reference to the last stripe_head for batching */
 768	struct stripe_head *batch_last;
 769
 770	/* first sector in the request */
 771	sector_t first_sector;
 772
 773	/* last sector in the request */
 774	sector_t last_sector;
 775
 776	/*
 777	 * bitmap to track stripe sectors that have been added to stripes
 778	 * add one to account for unaligned requests
 779	 */
 780	DECLARE_BITMAP(sectors_to_do, RAID5_MAX_REQ_STRIPES + 1);
 781
 782	/* the request had REQ_PREFLUSH, cleared after the first stripe_head */
 783	bool do_flush;
 784};
 785
 786/*
 787 * Block until another thread clears R5_INACTIVE_BLOCKED or
 788 * there are fewer than 3/4 the maximum number of active stripes
 789 * and there is an inactive stripe available.
 790 */
 791static bool is_inactive_blocked(struct r5conf *conf, int hash)
 792{
 793	if (list_empty(conf->inactive_list + hash))
 794		return false;
 795
 796	if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
 797		return true;
 798
 799	return (atomic_read(&conf->active_stripes) <
 800		(conf->max_nr_stripes * 3 / 4));
 801}
 802
 803struct stripe_head *raid5_get_active_stripe(struct r5conf *conf,
 804		struct stripe_request_ctx *ctx, sector_t sector,
 805		unsigned int flags)
 806{
 807	struct stripe_head *sh;
 808	int hash = stripe_hash_locks_hash(conf, sector);
 809	int previous = !!(flags & R5_GAS_PREVIOUS);
 810
 811	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 812
 813	spin_lock_irq(conf->hash_locks + hash);
 814
 815	for (;;) {
 816		if (!(flags & R5_GAS_NOQUIESCE) && conf->quiesce) {
 817			/*
 818			 * Must release the reference to batch_last before
 819			 * waiting, on quiesce, otherwise the batch_last will
 820			 * hold a reference to a stripe and raid5_quiesce()
 821			 * will deadlock waiting for active_stripes to go to
 822			 * zero.
 823			 */
 824			if (ctx && ctx->batch_last) {
 825				raid5_release_stripe(ctx->batch_last);
 826				ctx->batch_last = NULL;
 827			}
 828
 829			wait_event_lock_irq(conf->wait_for_quiescent,
 830					    !conf->quiesce,
 831					    *(conf->hash_locks + hash));
 832		}
 833
 834		sh = find_get_stripe(conf, sector, conf->generation - previous,
 835				     hash);
 836		if (sh)
 837			break;
 838
 839		if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 840			sh = get_free_stripe(conf, hash);
 841			if (sh) {
 842				r5c_check_stripe_cache_usage(conf);
 843				init_stripe(sh, sector, previous);
 844				atomic_inc(&sh->count);
 845				break;
 846			}
 847
 848			if (!test_bit(R5_DID_ALLOC, &conf->cache_state))
 849				set_bit(R5_ALLOC_MORE, &conf->cache_state);
 850		}
 851
 852		if (flags & R5_GAS_NOBLOCK)
 853			break;
 854
 855		set_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 856		r5l_wake_reclaim(conf->log, 0);
 857
 858		/* release batch_last before wait to avoid risk of deadlock */
 859		if (ctx && ctx->batch_last) {
 860			raid5_release_stripe(ctx->batch_last);
 861			ctx->batch_last = NULL;
 862		}
 863
 864		wait_event_lock_irq(conf->wait_for_stripe,
 865				    is_inactive_blocked(conf, hash),
 866				    *(conf->hash_locks + hash));
 867		clear_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 868	}
 869
 870	spin_unlock_irq(conf->hash_locks + hash);
 871	return sh;
 872}
 873
 874static bool is_full_stripe_write(struct stripe_head *sh)
 875{
 876	BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 877	return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 878}
 879
 880static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 881		__acquires(&sh1->stripe_lock)
 882		__acquires(&sh2->stripe_lock)
 883{
 884	if (sh1 > sh2) {
 885		spin_lock_irq(&sh2->stripe_lock);
 886		spin_lock_nested(&sh1->stripe_lock, 1);
 887	} else {
 888		spin_lock_irq(&sh1->stripe_lock);
 889		spin_lock_nested(&sh2->stripe_lock, 1);
 890	}
 891}
 892
 893static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 894		__releases(&sh1->stripe_lock)
 895		__releases(&sh2->stripe_lock)
 896{
 897	spin_unlock(&sh1->stripe_lock);
 898	spin_unlock_irq(&sh2->stripe_lock);
 899}
 900
 901/* Only freshly new full stripe normal write stripe can be added to a batch list */
 902static bool stripe_can_batch(struct stripe_head *sh)
 903{
 904	struct r5conf *conf = sh->raid_conf;
 905
 906	if (raid5_has_log(conf) || raid5_has_ppl(conf))
 907		return false;
 908	return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 909	       is_full_stripe_write(sh);
 910}
 911
 912/* we only do back search */
 913static void stripe_add_to_batch_list(struct r5conf *conf,
 914		struct stripe_head *sh, struct stripe_head *last_sh)
 915{
 916	struct stripe_head *head;
 917	sector_t head_sector, tmp_sec;
 918	int hash;
 919	int dd_idx;
 920
 921	/* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 922	tmp_sec = sh->sector;
 923	if (!sector_div(tmp_sec, conf->chunk_sectors))
 924		return;
 925	head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
 926
 927	if (last_sh && head_sector == last_sh->sector) {
 928		head = last_sh;
 929		atomic_inc(&head->count);
 930	} else {
 931		hash = stripe_hash_locks_hash(conf, head_sector);
 932		spin_lock_irq(conf->hash_locks + hash);
 933		head = find_get_stripe(conf, head_sector, conf->generation,
 934				       hash);
 935		spin_unlock_irq(conf->hash_locks + hash);
 936		if (!head)
 937			return;
 938		if (!stripe_can_batch(head))
 939			goto out;
 940	}
 941
 942	lock_two_stripes(head, sh);
 943	/* clear_batch_ready clear the flag */
 944	if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 945		goto unlock_out;
 946
 947	if (sh->batch_head)
 948		goto unlock_out;
 949
 950	dd_idx = 0;
 951	while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 952		dd_idx++;
 953	if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 954	    bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 955		goto unlock_out;
 956
 957	if (head->batch_head) {
 958		spin_lock(&head->batch_head->batch_lock);
 959		/* This batch list is already running */
 960		if (!stripe_can_batch(head)) {
 961			spin_unlock(&head->batch_head->batch_lock);
 962			goto unlock_out;
 963		}
 964		/*
 965		 * We must assign batch_head of this stripe within the
 966		 * batch_lock, otherwise clear_batch_ready of batch head
 967		 * stripe could clear BATCH_READY bit of this stripe and
 968		 * this stripe->batch_head doesn't get assigned, which
 969		 * could confuse clear_batch_ready for this stripe
 970		 */
 971		sh->batch_head = head->batch_head;
 972
 973		/*
 974		 * at this point, head's BATCH_READY could be cleared, but we
 975		 * can still add the stripe to batch list
 976		 */
 977		list_add(&sh->batch_list, &head->batch_list);
 978		spin_unlock(&head->batch_head->batch_lock);
 979	} else {
 980		head->batch_head = head;
 981		sh->batch_head = head->batch_head;
 982		spin_lock(&head->batch_lock);
 983		list_add_tail(&sh->batch_list, &head->batch_list);
 984		spin_unlock(&head->batch_lock);
 985	}
 986
 987	if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 988		if (atomic_dec_return(&conf->preread_active_stripes)
 989		    < IO_THRESHOLD)
 990			md_wakeup_thread(conf->mddev->thread);
 991
 992	if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 993		int seq = sh->bm_seq;
 994		if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 995		    sh->batch_head->bm_seq > seq)
 996			seq = sh->batch_head->bm_seq;
 997		set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 998		sh->batch_head->bm_seq = seq;
 999	}
1000
1001	atomic_inc(&sh->count);
1002unlock_out:
1003	unlock_two_stripes(head, sh);
1004out:
1005	raid5_release_stripe(head);
1006}
1007
1008/* Determine if 'data_offset' or 'new_data_offset' should be used
1009 * in this stripe_head.
1010 */
1011static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
1012{
1013	sector_t progress = conf->reshape_progress;
1014	/* Need a memory barrier to make sure we see the value
1015	 * of conf->generation, or ->data_offset that was set before
1016	 * reshape_progress was updated.
1017	 */
1018	smp_rmb();
1019	if (progress == MaxSector)
1020		return 0;
1021	if (sh->generation == conf->generation - 1)
1022		return 0;
1023	/* We are in a reshape, and this is a new-generation stripe,
1024	 * so use new_data_offset.
1025	 */
1026	return 1;
1027}
1028
1029static void dispatch_bio_list(struct bio_list *tmp)
1030{
1031	struct bio *bio;
1032
1033	while ((bio = bio_list_pop(tmp)))
1034		submit_bio_noacct(bio);
1035}
1036
1037static int cmp_stripe(void *priv, const struct list_head *a,
1038		      const struct list_head *b)
1039{
1040	const struct r5pending_data *da = list_entry(a,
1041				struct r5pending_data, sibling);
1042	const struct r5pending_data *db = list_entry(b,
1043				struct r5pending_data, sibling);
1044	if (da->sector > db->sector)
1045		return 1;
1046	if (da->sector < db->sector)
1047		return -1;
1048	return 0;
1049}
1050
1051static void dispatch_defer_bios(struct r5conf *conf, int target,
1052				struct bio_list *list)
1053{
1054	struct r5pending_data *data;
1055	struct list_head *first, *next = NULL;
1056	int cnt = 0;
1057
1058	if (conf->pending_data_cnt == 0)
1059		return;
1060
1061	list_sort(NULL, &conf->pending_list, cmp_stripe);
1062
1063	first = conf->pending_list.next;
1064
1065	/* temporarily move the head */
1066	if (conf->next_pending_data)
1067		list_move_tail(&conf->pending_list,
1068				&conf->next_pending_data->sibling);
1069
1070	while (!list_empty(&conf->pending_list)) {
1071		data = list_first_entry(&conf->pending_list,
1072			struct r5pending_data, sibling);
1073		if (&data->sibling == first)
1074			first = data->sibling.next;
1075		next = data->sibling.next;
1076
1077		bio_list_merge(list, &data->bios);
1078		list_move(&data->sibling, &conf->free_list);
1079		cnt++;
1080		if (cnt >= target)
1081			break;
1082	}
1083	conf->pending_data_cnt -= cnt;
1084	BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1085
1086	if (next != &conf->pending_list)
1087		conf->next_pending_data = list_entry(next,
1088				struct r5pending_data, sibling);
1089	else
1090		conf->next_pending_data = NULL;
1091	/* list isn't empty */
1092	if (first != &conf->pending_list)
1093		list_move_tail(&conf->pending_list, first);
1094}
1095
1096static void flush_deferred_bios(struct r5conf *conf)
1097{
1098	struct bio_list tmp = BIO_EMPTY_LIST;
1099
1100	if (conf->pending_data_cnt == 0)
1101		return;
1102
1103	spin_lock(&conf->pending_bios_lock);
1104	dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1105	BUG_ON(conf->pending_data_cnt != 0);
1106	spin_unlock(&conf->pending_bios_lock);
1107
1108	dispatch_bio_list(&tmp);
1109}
1110
1111static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1112				struct bio_list *bios)
1113{
1114	struct bio_list tmp = BIO_EMPTY_LIST;
1115	struct r5pending_data *ent;
1116
1117	spin_lock(&conf->pending_bios_lock);
1118	ent = list_first_entry(&conf->free_list, struct r5pending_data,
1119							sibling);
1120	list_move_tail(&ent->sibling, &conf->pending_list);
1121	ent->sector = sector;
1122	bio_list_init(&ent->bios);
1123	bio_list_merge(&ent->bios, bios);
1124	conf->pending_data_cnt++;
1125	if (conf->pending_data_cnt >= PENDING_IO_MAX)
1126		dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1127
1128	spin_unlock(&conf->pending_bios_lock);
1129
1130	dispatch_bio_list(&tmp);
1131}
1132
1133static void
1134raid5_end_read_request(struct bio *bi);
1135static void
1136raid5_end_write_request(struct bio *bi);
1137
1138static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1139{
1140	struct r5conf *conf = sh->raid_conf;
1141	int i, disks = sh->disks;
1142	struct stripe_head *head_sh = sh;
1143	struct bio_list pending_bios = BIO_EMPTY_LIST;
1144	struct r5dev *dev;
1145	bool should_defer;
1146
1147	might_sleep();
1148
1149	if (log_stripe(sh, s) == 0)
1150		return;
1151
1152	should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1153
1154	for (i = disks; i--; ) {
1155		enum req_op op;
1156		blk_opf_t op_flags = 0;
1157		int replace_only = 0;
1158		struct bio *bi, *rbi;
1159		struct md_rdev *rdev, *rrdev = NULL;
1160
1161		sh = head_sh;
1162		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1163			op = REQ_OP_WRITE;
1164			if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1165				op_flags = REQ_FUA;
1166			if (test_bit(R5_Discard, &sh->dev[i].flags))
1167				op = REQ_OP_DISCARD;
1168		} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1169			op = REQ_OP_READ;
1170		else if (test_and_clear_bit(R5_WantReplace,
1171					    &sh->dev[i].flags)) {
1172			op = REQ_OP_WRITE;
1173			replace_only = 1;
1174		} else
1175			continue;
1176		if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1177			op_flags |= REQ_SYNC;
1178
1179again:
1180		dev = &sh->dev[i];
1181		bi = &dev->req;
1182		rbi = &dev->rreq; /* For writing to replacement */
1183
1184		rdev = conf->disks[i].rdev;
1185		rrdev = conf->disks[i].replacement;
1186		if (op_is_write(op)) {
1187			if (replace_only)
1188				rdev = NULL;
1189			if (rdev == rrdev)
1190				/* We raced and saw duplicates */
1191				rrdev = NULL;
1192		} else {
1193			if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1194				rdev = rrdev;
1195			rrdev = NULL;
1196		}
1197
1198		if (rdev && test_bit(Faulty, &rdev->flags))
1199			rdev = NULL;
1200		if (rdev)
1201			atomic_inc(&rdev->nr_pending);
1202		if (rrdev && test_bit(Faulty, &rrdev->flags))
1203			rrdev = NULL;
1204		if (rrdev)
1205			atomic_inc(&rrdev->nr_pending);
1206
1207		/* We have already checked bad blocks for reads.  Now
1208		 * need to check for writes.  We never accept write errors
1209		 * on the replacement, so we don't to check rrdev.
1210		 */
1211		while (op_is_write(op) && rdev &&
1212		       test_bit(WriteErrorSeen, &rdev->flags)) {
1213			int bad = rdev_has_badblock(rdev, sh->sector,
1214						    RAID5_STRIPE_SECTORS(conf));
1215			if (!bad)
1216				break;
1217
1218			if (bad < 0) {
1219				set_bit(BlockedBadBlocks, &rdev->flags);
1220				if (!conf->mddev->external &&
1221				    conf->mddev->sb_flags) {
1222					/* It is very unlikely, but we might
1223					 * still need to write out the
1224					 * bad block log - better give it
1225					 * a chance*/
1226					md_check_recovery(conf->mddev);
1227				}
1228				/*
1229				 * Because md_wait_for_blocked_rdev
1230				 * will dec nr_pending, we must
1231				 * increment it first.
1232				 */
1233				atomic_inc(&rdev->nr_pending);
1234				md_wait_for_blocked_rdev(rdev, conf->mddev);
1235			} else {
1236				/* Acknowledged bad block - skip the write */
1237				rdev_dec_pending(rdev, conf->mddev);
1238				rdev = NULL;
1239			}
1240		}
1241
1242		if (rdev) {
1243			set_bit(STRIPE_IO_STARTED, &sh->state);
1244
1245			bio_init(bi, rdev->bdev, &dev->vec, 1, op | op_flags);
1246			bi->bi_end_io = op_is_write(op)
1247				? raid5_end_write_request
1248				: raid5_end_read_request;
1249			bi->bi_private = sh;
1250
1251			pr_debug("%s: for %llu schedule op %d on disc %d\n",
1252				__func__, (unsigned long long)sh->sector,
1253				bi->bi_opf, i);
1254			atomic_inc(&sh->count);
1255			if (sh != head_sh)
1256				atomic_inc(&head_sh->count);
1257			if (use_new_offset(conf, sh))
1258				bi->bi_iter.bi_sector = (sh->sector
1259						 + rdev->new_data_offset);
1260			else
1261				bi->bi_iter.bi_sector = (sh->sector
1262						 + rdev->data_offset);
1263			if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1264				bi->bi_opf |= REQ_NOMERGE;
1265
1266			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1267				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1268
1269			if (!op_is_write(op) &&
1270			    test_bit(R5_InJournal, &sh->dev[i].flags))
1271				/*
1272				 * issuing read for a page in journal, this
1273				 * must be preparing for prexor in rmw; read
1274				 * the data into orig_page
1275				 */
1276				sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1277			else
1278				sh->dev[i].vec.bv_page = sh->dev[i].page;
1279			bi->bi_vcnt = 1;
1280			bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1281			bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1282			bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1283			/*
1284			 * If this is discard request, set bi_vcnt 0. We don't
1285			 * want to confuse SCSI because SCSI will replace payload
1286			 */
1287			if (op == REQ_OP_DISCARD)
1288				bi->bi_vcnt = 0;
1289			if (rrdev)
1290				set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1291
1292			mddev_trace_remap(conf->mddev, bi, sh->dev[i].sector);
1293			if (should_defer && op_is_write(op))
1294				bio_list_add(&pending_bios, bi);
1295			else
1296				submit_bio_noacct(bi);
1297		}
1298		if (rrdev) {
1299			set_bit(STRIPE_IO_STARTED, &sh->state);
1300
1301			bio_init(rbi, rrdev->bdev, &dev->rvec, 1, op | op_flags);
1302			BUG_ON(!op_is_write(op));
1303			rbi->bi_end_io = raid5_end_write_request;
1304			rbi->bi_private = sh;
1305
1306			pr_debug("%s: for %llu schedule op %d on "
1307				 "replacement disc %d\n",
1308				__func__, (unsigned long long)sh->sector,
1309				rbi->bi_opf, i);
1310			atomic_inc(&sh->count);
1311			if (sh != head_sh)
1312				atomic_inc(&head_sh->count);
1313			if (use_new_offset(conf, sh))
1314				rbi->bi_iter.bi_sector = (sh->sector
1315						  + rrdev->new_data_offset);
1316			else
1317				rbi->bi_iter.bi_sector = (sh->sector
1318						  + rrdev->data_offset);
1319			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1320				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1321			sh->dev[i].rvec.bv_page = sh->dev[i].page;
1322			rbi->bi_vcnt = 1;
1323			rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1324			rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1325			rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1326			/*
1327			 * If this is discard request, set bi_vcnt 0. We don't
1328			 * want to confuse SCSI because SCSI will replace payload
1329			 */
1330			if (op == REQ_OP_DISCARD)
1331				rbi->bi_vcnt = 0;
1332			mddev_trace_remap(conf->mddev, rbi, sh->dev[i].sector);
1333			if (should_defer && op_is_write(op))
1334				bio_list_add(&pending_bios, rbi);
1335			else
1336				submit_bio_noacct(rbi);
1337		}
1338		if (!rdev && !rrdev) {
1339			pr_debug("skip op %d on disc %d for sector %llu\n",
1340				bi->bi_opf, i, (unsigned long long)sh->sector);
1341			clear_bit(R5_LOCKED, &sh->dev[i].flags);
1342			set_bit(STRIPE_HANDLE, &sh->state);
1343		}
1344
1345		if (!head_sh->batch_head)
1346			continue;
1347		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1348				      batch_list);
1349		if (sh != head_sh)
1350			goto again;
1351	}
1352
1353	if (should_defer && !bio_list_empty(&pending_bios))
1354		defer_issue_bios(conf, head_sh->sector, &pending_bios);
1355}
1356
1357static struct dma_async_tx_descriptor *
1358async_copy_data(int frombio, struct bio *bio, struct page **page,
1359	unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1360	struct stripe_head *sh, int no_skipcopy)
1361{
1362	struct bio_vec bvl;
1363	struct bvec_iter iter;
1364	struct page *bio_page;
1365	int page_offset;
1366	struct async_submit_ctl submit;
1367	enum async_tx_flags flags = 0;
1368	struct r5conf *conf = sh->raid_conf;
1369
1370	if (bio->bi_iter.bi_sector >= sector)
1371		page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1372	else
1373		page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1374
1375	if (frombio)
1376		flags |= ASYNC_TX_FENCE;
1377	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1378
1379	bio_for_each_segment(bvl, bio, iter) {
1380		int len = bvl.bv_len;
1381		int clen;
1382		int b_offset = 0;
1383
1384		if (page_offset < 0) {
1385			b_offset = -page_offset;
1386			page_offset += b_offset;
1387			len -= b_offset;
1388		}
1389
1390		if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1391			clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1392		else
1393			clen = len;
1394
1395		if (clen > 0) {
1396			b_offset += bvl.bv_offset;
1397			bio_page = bvl.bv_page;
1398			if (frombio) {
1399				if (conf->skip_copy &&
1400				    b_offset == 0 && page_offset == 0 &&
1401				    clen == RAID5_STRIPE_SIZE(conf) &&
1402				    !no_skipcopy)
1403					*page = bio_page;
1404				else
1405					tx = async_memcpy(*page, bio_page, page_offset + poff,
1406						  b_offset, clen, &submit);
1407			} else
1408				tx = async_memcpy(bio_page, *page, b_offset,
1409						  page_offset + poff, clen, &submit);
1410		}
1411		/* chain the operations */
1412		submit.depend_tx = tx;
1413
1414		if (clen < len) /* hit end of page */
1415			break;
1416		page_offset +=  len;
1417	}
1418
1419	return tx;
1420}
1421
1422static void ops_complete_biofill(void *stripe_head_ref)
1423{
1424	struct stripe_head *sh = stripe_head_ref;
1425	int i;
1426	struct r5conf *conf = sh->raid_conf;
1427
1428	pr_debug("%s: stripe %llu\n", __func__,
1429		(unsigned long long)sh->sector);
1430
1431	/* clear completed biofills */
1432	for (i = sh->disks; i--; ) {
1433		struct r5dev *dev = &sh->dev[i];
1434
1435		/* acknowledge completion of a biofill operation */
1436		/* and check if we need to reply to a read request,
1437		 * new R5_Wantfill requests are held off until
1438		 * !STRIPE_BIOFILL_RUN
1439		 */
1440		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1441			struct bio *rbi, *rbi2;
1442
1443			BUG_ON(!dev->read);
1444			rbi = dev->read;
1445			dev->read = NULL;
1446			while (rbi && rbi->bi_iter.bi_sector <
1447				dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1448				rbi2 = r5_next_bio(conf, rbi, dev->sector);
1449				bio_endio(rbi);
1450				rbi = rbi2;
1451			}
1452		}
1453	}
1454	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1455
1456	set_bit(STRIPE_HANDLE, &sh->state);
1457	raid5_release_stripe(sh);
1458}
1459
1460static void ops_run_biofill(struct stripe_head *sh)
1461{
1462	struct dma_async_tx_descriptor *tx = NULL;
1463	struct async_submit_ctl submit;
1464	int i;
1465	struct r5conf *conf = sh->raid_conf;
1466
1467	BUG_ON(sh->batch_head);
1468	pr_debug("%s: stripe %llu\n", __func__,
1469		(unsigned long long)sh->sector);
1470
1471	for (i = sh->disks; i--; ) {
1472		struct r5dev *dev = &sh->dev[i];
1473		if (test_bit(R5_Wantfill, &dev->flags)) {
1474			struct bio *rbi;
1475			spin_lock_irq(&sh->stripe_lock);
1476			dev->read = rbi = dev->toread;
1477			dev->toread = NULL;
1478			spin_unlock_irq(&sh->stripe_lock);
1479			while (rbi && rbi->bi_iter.bi_sector <
1480				dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1481				tx = async_copy_data(0, rbi, &dev->page,
1482						     dev->offset,
1483						     dev->sector, tx, sh, 0);
1484				rbi = r5_next_bio(conf, rbi, dev->sector);
1485			}
1486		}
1487	}
1488
1489	atomic_inc(&sh->count);
1490	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1491	async_trigger_callback(&submit);
1492}
1493
1494static void mark_target_uptodate(struct stripe_head *sh, int target)
1495{
1496	struct r5dev *tgt;
1497
1498	if (target < 0)
1499		return;
1500
1501	tgt = &sh->dev[target];
1502	set_bit(R5_UPTODATE, &tgt->flags);
1503	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1504	clear_bit(R5_Wantcompute, &tgt->flags);
1505}
1506
1507static void ops_complete_compute(void *stripe_head_ref)
1508{
1509	struct stripe_head *sh = stripe_head_ref;
1510
1511	pr_debug("%s: stripe %llu\n", __func__,
1512		(unsigned long long)sh->sector);
1513
1514	/* mark the computed target(s) as uptodate */
1515	mark_target_uptodate(sh, sh->ops.target);
1516	mark_target_uptodate(sh, sh->ops.target2);
1517
1518	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1519	if (sh->check_state == check_state_compute_run)
1520		sh->check_state = check_state_compute_result;
1521	set_bit(STRIPE_HANDLE, &sh->state);
1522	raid5_release_stripe(sh);
1523}
1524
1525/* return a pointer to the address conversion region of the scribble buffer */
1526static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1527{
1528	return percpu->scribble + i * percpu->scribble_obj_size;
1529}
1530
1531/* return a pointer to the address conversion region of the scribble buffer */
1532static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1533				 struct raid5_percpu *percpu, int i)
1534{
1535	return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1536}
1537
1538/*
1539 * Return a pointer to record offset address.
1540 */
1541static unsigned int *
1542to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1543{
1544	return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1545}
1546
1547static struct dma_async_tx_descriptor *
1548ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1549{
1550	int disks = sh->disks;
1551	struct page **xor_srcs = to_addr_page(percpu, 0);
1552	unsigned int *off_srcs = to_addr_offs(sh, percpu);
1553	int target = sh->ops.target;
1554	struct r5dev *tgt = &sh->dev[target];
1555	struct page *xor_dest = tgt->page;
1556	unsigned int off_dest = tgt->offset;
1557	int count = 0;
1558	struct dma_async_tx_descriptor *tx;
1559	struct async_submit_ctl submit;
1560	int i;
1561
1562	BUG_ON(sh->batch_head);
1563
1564	pr_debug("%s: stripe %llu block: %d\n",
1565		__func__, (unsigned long long)sh->sector, target);
1566	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1567
1568	for (i = disks; i--; ) {
1569		if (i != target) {
1570			off_srcs[count] = sh->dev[i].offset;
1571			xor_srcs[count++] = sh->dev[i].page;
1572		}
1573	}
1574
1575	atomic_inc(&sh->count);
1576
1577	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1578			  ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1579	if (unlikely(count == 1))
1580		tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1581				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1582	else
1583		tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1584				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1585
1586	return tx;
1587}
1588
1589/* set_syndrome_sources - populate source buffers for gen_syndrome
1590 * @srcs - (struct page *) array of size sh->disks
1591 * @offs - (unsigned int) array of offset for each page
1592 * @sh - stripe_head to parse
1593 *
1594 * Populates srcs in proper layout order for the stripe and returns the
1595 * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1596 * destination buffer is recorded in srcs[count] and the Q destination
1597 * is recorded in srcs[count+1]].
1598 */
1599static int set_syndrome_sources(struct page **srcs,
1600				unsigned int *offs,
1601				struct stripe_head *sh,
1602				int srctype)
1603{
1604	int disks = sh->disks;
1605	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1606	int d0_idx = raid6_d0(sh);
1607	int count;
1608	int i;
1609
1610	for (i = 0; i < disks; i++)
1611		srcs[i] = NULL;
1612
1613	count = 0;
1614	i = d0_idx;
1615	do {
1616		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1617		struct r5dev *dev = &sh->dev[i];
1618
1619		if (i == sh->qd_idx || i == sh->pd_idx ||
1620		    (srctype == SYNDROME_SRC_ALL) ||
1621		    (srctype == SYNDROME_SRC_WANT_DRAIN &&
1622		     (test_bit(R5_Wantdrain, &dev->flags) ||
1623		      test_bit(R5_InJournal, &dev->flags))) ||
1624		    (srctype == SYNDROME_SRC_WRITTEN &&
1625		     (dev->written ||
1626		      test_bit(R5_InJournal, &dev->flags)))) {
1627			if (test_bit(R5_InJournal, &dev->flags))
1628				srcs[slot] = sh->dev[i].orig_page;
1629			else
1630				srcs[slot] = sh->dev[i].page;
1631			/*
1632			 * For R5_InJournal, PAGE_SIZE must be 4KB and will
1633			 * not shared page. In that case, dev[i].offset
1634			 * is 0.
1635			 */
1636			offs[slot] = sh->dev[i].offset;
1637		}
1638		i = raid6_next_disk(i, disks);
1639	} while (i != d0_idx);
1640
1641	return syndrome_disks;
1642}
1643
1644static struct dma_async_tx_descriptor *
1645ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1646{
1647	int disks = sh->disks;
1648	struct page **blocks = to_addr_page(percpu, 0);
1649	unsigned int *offs = to_addr_offs(sh, percpu);
1650	int target;
1651	int qd_idx = sh->qd_idx;
1652	struct dma_async_tx_descriptor *tx;
1653	struct async_submit_ctl submit;
1654	struct r5dev *tgt;
1655	struct page *dest;
1656	unsigned int dest_off;
1657	int i;
1658	int count;
1659
1660	BUG_ON(sh->batch_head);
1661	if (sh->ops.target < 0)
1662		target = sh->ops.target2;
1663	else if (sh->ops.target2 < 0)
1664		target = sh->ops.target;
1665	else
1666		/* we should only have one valid target */
1667		BUG();
1668	BUG_ON(target < 0);
1669	pr_debug("%s: stripe %llu block: %d\n",
1670		__func__, (unsigned long long)sh->sector, target);
1671
1672	tgt = &sh->dev[target];
1673	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1674	dest = tgt->page;
1675	dest_off = tgt->offset;
1676
1677	atomic_inc(&sh->count);
1678
1679	if (target == qd_idx) {
1680		count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1681		blocks[count] = NULL; /* regenerating p is not necessary */
1682		BUG_ON(blocks[count+1] != dest); /* q should already be set */
1683		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1684				  ops_complete_compute, sh,
1685				  to_addr_conv(sh, percpu, 0));
1686		tx = async_gen_syndrome(blocks, offs, count+2,
1687				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1688	} else {
1689		/* Compute any data- or p-drive using XOR */
1690		count = 0;
1691		for (i = disks; i-- ; ) {
1692			if (i == target || i == qd_idx)
1693				continue;
1694			offs[count] = sh->dev[i].offset;
1695			blocks[count++] = sh->dev[i].page;
1696		}
1697
1698		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1699				  NULL, ops_complete_compute, sh,
1700				  to_addr_conv(sh, percpu, 0));
1701		tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1702				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1703	}
1704
1705	return tx;
1706}
1707
1708static struct dma_async_tx_descriptor *
1709ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1710{
1711	int i, count, disks = sh->disks;
1712	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1713	int d0_idx = raid6_d0(sh);
1714	int faila = -1, failb = -1;
1715	int target = sh->ops.target;
1716	int target2 = sh->ops.target2;
1717	struct r5dev *tgt = &sh->dev[target];
1718	struct r5dev *tgt2 = &sh->dev[target2];
1719	struct dma_async_tx_descriptor *tx;
1720	struct page **blocks = to_addr_page(percpu, 0);
1721	unsigned int *offs = to_addr_offs(sh, percpu);
1722	struct async_submit_ctl submit;
1723
1724	BUG_ON(sh->batch_head);
1725	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1726		 __func__, (unsigned long long)sh->sector, target, target2);
1727	BUG_ON(target < 0 || target2 < 0);
1728	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1729	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1730
1731	/* we need to open-code set_syndrome_sources to handle the
1732	 * slot number conversion for 'faila' and 'failb'
1733	 */
1734	for (i = 0; i < disks ; i++) {
1735		offs[i] = 0;
1736		blocks[i] = NULL;
1737	}
1738	count = 0;
1739	i = d0_idx;
1740	do {
1741		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1742
1743		offs[slot] = sh->dev[i].offset;
1744		blocks[slot] = sh->dev[i].page;
1745
1746		if (i == target)
1747			faila = slot;
1748		if (i == target2)
1749			failb = slot;
1750		i = raid6_next_disk(i, disks);
1751	} while (i != d0_idx);
1752
1753	BUG_ON(faila == failb);
1754	if (failb < faila)
1755		swap(faila, failb);
1756	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1757		 __func__, (unsigned long long)sh->sector, faila, failb);
1758
1759	atomic_inc(&sh->count);
1760
1761	if (failb == syndrome_disks+1) {
1762		/* Q disk is one of the missing disks */
1763		if (faila == syndrome_disks) {
1764			/* Missing P+Q, just recompute */
1765			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1766					  ops_complete_compute, sh,
1767					  to_addr_conv(sh, percpu, 0));
1768			return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1769						  RAID5_STRIPE_SIZE(sh->raid_conf),
1770						  &submit);
1771		} else {
1772			struct page *dest;
1773			unsigned int dest_off;
1774			int data_target;
1775			int qd_idx = sh->qd_idx;
1776
1777			/* Missing D+Q: recompute D from P, then recompute Q */
1778			if (target == qd_idx)
1779				data_target = target2;
1780			else
1781				data_target = target;
1782
1783			count = 0;
1784			for (i = disks; i-- ; ) {
1785				if (i == data_target || i == qd_idx)
1786					continue;
1787				offs[count] = sh->dev[i].offset;
1788				blocks[count++] = sh->dev[i].page;
1789			}
1790			dest = sh->dev[data_target].page;
1791			dest_off = sh->dev[data_target].offset;
1792			init_async_submit(&submit,
1793					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1794					  NULL, NULL, NULL,
1795					  to_addr_conv(sh, percpu, 0));
1796			tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1797				       RAID5_STRIPE_SIZE(sh->raid_conf),
1798				       &submit);
1799
1800			count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1801			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1802					  ops_complete_compute, sh,
1803					  to_addr_conv(sh, percpu, 0));
1804			return async_gen_syndrome(blocks, offs, count+2,
1805						  RAID5_STRIPE_SIZE(sh->raid_conf),
1806						  &submit);
1807		}
1808	} else {
1809		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1810				  ops_complete_compute, sh,
1811				  to_addr_conv(sh, percpu, 0));
1812		if (failb == syndrome_disks) {
1813			/* We're missing D+P. */
1814			return async_raid6_datap_recov(syndrome_disks+2,
1815						RAID5_STRIPE_SIZE(sh->raid_conf),
1816						faila,
1817						blocks, offs, &submit);
1818		} else {
1819			/* We're missing D+D. */
1820			return async_raid6_2data_recov(syndrome_disks+2,
1821						RAID5_STRIPE_SIZE(sh->raid_conf),
1822						faila, failb,
1823						blocks, offs, &submit);
1824		}
1825	}
1826}
1827
1828static void ops_complete_prexor(void *stripe_head_ref)
1829{
1830	struct stripe_head *sh = stripe_head_ref;
1831
1832	pr_debug("%s: stripe %llu\n", __func__,
1833		(unsigned long long)sh->sector);
1834
1835	if (r5c_is_writeback(sh->raid_conf->log))
1836		/*
1837		 * raid5-cache write back uses orig_page during prexor.
1838		 * After prexor, it is time to free orig_page
1839		 */
1840		r5c_release_extra_page(sh);
1841}
1842
1843static struct dma_async_tx_descriptor *
1844ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1845		struct dma_async_tx_descriptor *tx)
1846{
1847	int disks = sh->disks;
1848	struct page **xor_srcs = to_addr_page(percpu, 0);
1849	unsigned int *off_srcs = to_addr_offs(sh, percpu);
1850	int count = 0, pd_idx = sh->pd_idx, i;
1851	struct async_submit_ctl submit;
1852
1853	/* existing parity data subtracted */
1854	unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1855	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1856
1857	BUG_ON(sh->batch_head);
1858	pr_debug("%s: stripe %llu\n", __func__,
1859		(unsigned long long)sh->sector);
1860
1861	for (i = disks; i--; ) {
1862		struct r5dev *dev = &sh->dev[i];
1863		/* Only process blocks that are known to be uptodate */
1864		if (test_bit(R5_InJournal, &dev->flags)) {
1865			/*
1866			 * For this case, PAGE_SIZE must be equal to 4KB and
1867			 * page offset is zero.
1868			 */
1869			off_srcs[count] = dev->offset;
1870			xor_srcs[count++] = dev->orig_page;
1871		} else if (test_bit(R5_Wantdrain, &dev->flags)) {
1872			off_srcs[count] = dev->offset;
1873			xor_srcs[count++] = dev->page;
1874		}
1875	}
1876
1877	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1878			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1879	tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1880			RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1881
1882	return tx;
1883}
1884
1885static struct dma_async_tx_descriptor *
1886ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1887		struct dma_async_tx_descriptor *tx)
1888{
1889	struct page **blocks = to_addr_page(percpu, 0);
1890	unsigned int *offs = to_addr_offs(sh, percpu);
1891	int count;
1892	struct async_submit_ctl submit;
1893
1894	pr_debug("%s: stripe %llu\n", __func__,
1895		(unsigned long long)sh->sector);
1896
1897	count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1898
1899	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1900			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1901	tx = async_gen_syndrome(blocks, offs, count+2,
1902			RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1903
1904	return tx;
1905}
1906
1907static struct dma_async_tx_descriptor *
1908ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1909{
1910	struct r5conf *conf = sh->raid_conf;
1911	int disks = sh->disks;
1912	int i;
1913	struct stripe_head *head_sh = sh;
1914
1915	pr_debug("%s: stripe %llu\n", __func__,
1916		(unsigned long long)sh->sector);
1917
1918	for (i = disks; i--; ) {
1919		struct r5dev *dev;
1920		struct bio *chosen;
1921
1922		sh = head_sh;
1923		if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1924			struct bio *wbi;
1925
1926again:
1927			dev = &sh->dev[i];
1928			/*
1929			 * clear R5_InJournal, so when rewriting a page in
1930			 * journal, it is not skipped by r5l_log_stripe()
1931			 */
1932			clear_bit(R5_InJournal, &dev->flags);
1933			spin_lock_irq(&sh->stripe_lock);
1934			chosen = dev->towrite;
1935			dev->towrite = NULL;
1936			sh->overwrite_disks = 0;
1937			BUG_ON(dev->written);
1938			wbi = dev->written = chosen;
1939			spin_unlock_irq(&sh->stripe_lock);
1940			WARN_ON(dev->page != dev->orig_page);
1941
1942			while (wbi && wbi->bi_iter.bi_sector <
1943				dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1944				if (wbi->bi_opf & REQ_FUA)
1945					set_bit(R5_WantFUA, &dev->flags);
1946				if (wbi->bi_opf & REQ_SYNC)
1947					set_bit(R5_SyncIO, &dev->flags);
1948				if (bio_op(wbi) == REQ_OP_DISCARD)
1949					set_bit(R5_Discard, &dev->flags);
1950				else {
1951					tx = async_copy_data(1, wbi, &dev->page,
1952							     dev->offset,
1953							     dev->sector, tx, sh,
1954							     r5c_is_writeback(conf->log));
1955					if (dev->page != dev->orig_page &&
1956					    !r5c_is_writeback(conf->log)) {
1957						set_bit(R5_SkipCopy, &dev->flags);
1958						clear_bit(R5_UPTODATE, &dev->flags);
1959						clear_bit(R5_OVERWRITE, &dev->flags);
1960					}
1961				}
1962				wbi = r5_next_bio(conf, wbi, dev->sector);
1963			}
1964
1965			if (head_sh->batch_head) {
1966				sh = list_first_entry(&sh->batch_list,
1967						      struct stripe_head,
1968						      batch_list);
1969				if (sh == head_sh)
1970					continue;
1971				goto again;
1972			}
1973		}
1974	}
1975
1976	return tx;
1977}
1978
1979static void ops_complete_reconstruct(void *stripe_head_ref)
1980{
1981	struct stripe_head *sh = stripe_head_ref;
1982	int disks = sh->disks;
1983	int pd_idx = sh->pd_idx;
1984	int qd_idx = sh->qd_idx;
1985	int i;
1986	bool fua = false, sync = false, discard = false;
1987
1988	pr_debug("%s: stripe %llu\n", __func__,
1989		(unsigned long long)sh->sector);
1990
1991	for (i = disks; i--; ) {
1992		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1993		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1994		discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1995	}
1996
1997	for (i = disks; i--; ) {
1998		struct r5dev *dev = &sh->dev[i];
1999
2000		if (dev->written || i == pd_idx || i == qd_idx) {
2001			if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
2002				set_bit(R5_UPTODATE, &dev->flags);
2003				if (test_bit(STRIPE_EXPAND_READY, &sh->state))
2004					set_bit(R5_Expanded, &dev->flags);
2005			}
2006			if (fua)
2007				set_bit(R5_WantFUA, &dev->flags);
2008			if (sync)
2009				set_bit(R5_SyncIO, &dev->flags);
2010		}
2011	}
2012
2013	if (sh->reconstruct_state == reconstruct_state_drain_run)
2014		sh->reconstruct_state = reconstruct_state_drain_result;
2015	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
2016		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
2017	else {
2018		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
2019		sh->reconstruct_state = reconstruct_state_result;
2020	}
2021
2022	set_bit(STRIPE_HANDLE, &sh->state);
2023	raid5_release_stripe(sh);
2024}
2025
2026static void
2027ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
2028		     struct dma_async_tx_descriptor *tx)
2029{
2030	int disks = sh->disks;
2031	struct page **xor_srcs;
2032	unsigned int *off_srcs;
2033	struct async_submit_ctl submit;
2034	int count, pd_idx = sh->pd_idx, i;
2035	struct page *xor_dest;
2036	unsigned int off_dest;
2037	int prexor = 0;
2038	unsigned long flags;
2039	int j = 0;
2040	struct stripe_head *head_sh = sh;
2041	int last_stripe;
2042
2043	pr_debug("%s: stripe %llu\n", __func__,
2044		(unsigned long long)sh->sector);
2045
2046	for (i = 0; i < sh->disks; i++) {
2047		if (pd_idx == i)
2048			continue;
2049		if (!test_bit(R5_Discard, &sh->dev[i].flags))
2050			break;
2051	}
2052	if (i >= sh->disks) {
2053		atomic_inc(&sh->count);
2054		set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2055		ops_complete_reconstruct(sh);
2056		return;
2057	}
2058again:
2059	count = 0;
2060	xor_srcs = to_addr_page(percpu, j);
2061	off_srcs = to_addr_offs(sh, percpu);
2062	/* check if prexor is active which means only process blocks
2063	 * that are part of a read-modify-write (written)
2064	 */
2065	if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2066		prexor = 1;
2067		off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2068		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2069		for (i = disks; i--; ) {
2070			struct r5dev *dev = &sh->dev[i];
2071			if (head_sh->dev[i].written ||
2072			    test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2073				off_srcs[count] = dev->offset;
2074				xor_srcs[count++] = dev->page;
2075			}
2076		}
2077	} else {
2078		xor_dest = sh->dev[pd_idx].page;
2079		off_dest = sh->dev[pd_idx].offset;
2080		for (i = disks; i--; ) {
2081			struct r5dev *dev = &sh->dev[i];
2082			if (i != pd_idx) {
2083				off_srcs[count] = dev->offset;
2084				xor_srcs[count++] = dev->page;
2085			}
2086		}
2087	}
2088
2089	/* 1/ if we prexor'd then the dest is reused as a source
2090	 * 2/ if we did not prexor then we are redoing the parity
2091	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2092	 * for the synchronous xor case
2093	 */
2094	last_stripe = !head_sh->batch_head ||
2095		list_first_entry(&sh->batch_list,
2096				 struct stripe_head, batch_list) == head_sh;
2097	if (last_stripe) {
2098		flags = ASYNC_TX_ACK |
2099			(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2100
2101		atomic_inc(&head_sh->count);
2102		init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2103				  to_addr_conv(sh, percpu, j));
2104	} else {
2105		flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2106		init_async_submit(&submit, flags, tx, NULL, NULL,
2107				  to_addr_conv(sh, percpu, j));
2108	}
2109
2110	if (unlikely(count == 1))
2111		tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2112				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2113	else
2114		tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2115				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2116	if (!last_stripe) {
2117		j++;
2118		sh = list_first_entry(&sh->batch_list, struct stripe_head,
2119				      batch_list);
2120		goto again;
2121	}
2122}
2123
2124static void
2125ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2126		     struct dma_async_tx_descriptor *tx)
2127{
2128	struct async_submit_ctl submit;
2129	struct page **blocks;
2130	unsigned int *offs;
2131	int count, i, j = 0;
2132	struct stripe_head *head_sh = sh;
2133	int last_stripe;
2134	int synflags;
2135	unsigned long txflags;
2136
2137	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2138
2139	for (i = 0; i < sh->disks; i++) {
2140		if (sh->pd_idx == i || sh->qd_idx == i)
2141			continue;
2142		if (!test_bit(R5_Discard, &sh->dev[i].flags))
2143			break;
2144	}
2145	if (i >= sh->disks) {
2146		atomic_inc(&sh->count);
2147		set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2148		set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2149		ops_complete_reconstruct(sh);
2150		return;
2151	}
2152
2153again:
2154	blocks = to_addr_page(percpu, j);
2155	offs = to_addr_offs(sh, percpu);
2156
2157	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2158		synflags = SYNDROME_SRC_WRITTEN;
2159		txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2160	} else {
2161		synflags = SYNDROME_SRC_ALL;
2162		txflags = ASYNC_TX_ACK;
2163	}
2164
2165	count = set_syndrome_sources(blocks, offs, sh, synflags);
2166	last_stripe = !head_sh->batch_head ||
2167		list_first_entry(&sh->batch_list,
2168				 struct stripe_head, batch_list) == head_sh;
2169
2170	if (last_stripe) {
2171		atomic_inc(&head_sh->count);
2172		init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2173				  head_sh, to_addr_conv(sh, percpu, j));
2174	} else
2175		init_async_submit(&submit, 0, tx, NULL, NULL,
2176				  to_addr_conv(sh, percpu, j));
2177	tx = async_gen_syndrome(blocks, offs, count+2,
2178			RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2179	if (!last_stripe) {
2180		j++;
2181		sh = list_first_entry(&sh->batch_list, struct stripe_head,
2182				      batch_list);
2183		goto again;
2184	}
2185}
2186
2187static void ops_complete_check(void *stripe_head_ref)
2188{
2189	struct stripe_head *sh = stripe_head_ref;
2190
2191	pr_debug("%s: stripe %llu\n", __func__,
2192		(unsigned long long)sh->sector);
2193
2194	sh->check_state = check_state_check_result;
2195	set_bit(STRIPE_HANDLE, &sh->state);
2196	raid5_release_stripe(sh);
2197}
2198
2199static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2200{
2201	int disks = sh->disks;
2202	int pd_idx = sh->pd_idx;
2203	int qd_idx = sh->qd_idx;
2204	struct page *xor_dest;
2205	unsigned int off_dest;
2206	struct page **xor_srcs = to_addr_page(percpu, 0);
2207	unsigned int *off_srcs = to_addr_offs(sh, percpu);
2208	struct dma_async_tx_descriptor *tx;
2209	struct async_submit_ctl submit;
2210	int count;
2211	int i;
2212
2213	pr_debug("%s: stripe %llu\n", __func__,
2214		(unsigned long long)sh->sector);
2215
2216	BUG_ON(sh->batch_head);
2217	count = 0;
2218	xor_dest = sh->dev[pd_idx].page;
2219	off_dest = sh->dev[pd_idx].offset;
2220	off_srcs[count] = off_dest;
2221	xor_srcs[count++] = xor_dest;
2222	for (i = disks; i--; ) {
2223		if (i == pd_idx || i == qd_idx)
2224			continue;
2225		off_srcs[count] = sh->dev[i].offset;
2226		xor_srcs[count++] = sh->dev[i].page;
2227	}
2228
2229	init_async_submit(&submit, 0, NULL, NULL, NULL,
2230			  to_addr_conv(sh, percpu, 0));
2231	tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2232			   RAID5_STRIPE_SIZE(sh->raid_conf),
2233			   &sh->ops.zero_sum_result, &submit);
2234
2235	atomic_inc(&sh->count);
2236	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2237	tx = async_trigger_callback(&submit);
2238}
2239
2240static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2241{
2242	struct page **srcs = to_addr_page(percpu, 0);
2243	unsigned int *offs = to_addr_offs(sh, percpu);
2244	struct async_submit_ctl submit;
2245	int count;
2246
2247	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2248		(unsigned long long)sh->sector, checkp);
2249
2250	BUG_ON(sh->batch_head);
2251	count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2252	if (!checkp)
2253		srcs[count] = NULL;
2254
2255	atomic_inc(&sh->count);
2256	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2257			  sh, to_addr_conv(sh, percpu, 0));
2258	async_syndrome_val(srcs, offs, count+2,
2259			   RAID5_STRIPE_SIZE(sh->raid_conf),
2260			   &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2261}
2262
2263static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2264{
2265	int overlap_clear = 0, i, disks = sh->disks;
2266	struct dma_async_tx_descriptor *tx = NULL;
2267	struct r5conf *conf = sh->raid_conf;
2268	int level = conf->level;
2269	struct raid5_percpu *percpu;
2270
2271	local_lock(&conf->percpu->lock);
2272	percpu = this_cpu_ptr(conf->percpu);
2273	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2274		ops_run_biofill(sh);
2275		overlap_clear++;
2276	}
2277
2278	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2279		if (level < 6)
2280			tx = ops_run_compute5(sh, percpu);
2281		else {
2282			if (sh->ops.target2 < 0 || sh->ops.target < 0)
2283				tx = ops_run_compute6_1(sh, percpu);
2284			else
2285				tx = ops_run_compute6_2(sh, percpu);
2286		}
2287		/* terminate the chain if reconstruct is not set to be run */
2288		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2289			async_tx_ack(tx);
2290	}
2291
2292	if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2293		if (level < 6)
2294			tx = ops_run_prexor5(sh, percpu, tx);
2295		else
2296			tx = ops_run_prexor6(sh, percpu, tx);
2297	}
2298
2299	if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2300		tx = ops_run_partial_parity(sh, percpu, tx);
2301
2302	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2303		tx = ops_run_biodrain(sh, tx);
2304		overlap_clear++;
2305	}
2306
2307	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2308		if (level < 6)
2309			ops_run_reconstruct5(sh, percpu, tx);
2310		else
2311			ops_run_reconstruct6(sh, percpu, tx);
2312	}
2313
2314	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2315		if (sh->check_state == check_state_run)
2316			ops_run_check_p(sh, percpu);
2317		else if (sh->check_state == check_state_run_q)
2318			ops_run_check_pq(sh, percpu, 0);
2319		else if (sh->check_state == check_state_run_pq)
2320			ops_run_check_pq(sh, percpu, 1);
2321		else
2322			BUG();
2323	}
2324
2325	if (overlap_clear && !sh->batch_head) {
2326		for (i = disks; i--; ) {
2327			struct r5dev *dev = &sh->dev[i];
2328			if (test_and_clear_bit(R5_Overlap, &dev->flags))
2329				wake_up_bit(&dev->flags, R5_Overlap);
2330		}
2331	}
2332	local_unlock(&conf->percpu->lock);
2333}
2334
2335static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2336{
2337#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2338	kfree(sh->pages);
2339#endif
2340	if (sh->ppl_page)
2341		__free_page(sh->ppl_page);
2342	kmem_cache_free(sc, sh);
2343}
2344
2345static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2346	int disks, struct r5conf *conf)
2347{
2348	struct stripe_head *sh;
2349
2350	sh = kmem_cache_zalloc(sc, gfp);
2351	if (sh) {
2352		spin_lock_init(&sh->stripe_lock);
2353		spin_lock_init(&sh->batch_lock);
2354		INIT_LIST_HEAD(&sh->batch_list);
2355		INIT_LIST_HEAD(&sh->lru);
2356		INIT_LIST_HEAD(&sh->r5c);
2357		INIT_LIST_HEAD(&sh->log_list);
2358		atomic_set(&sh->count, 1);
2359		sh->raid_conf = conf;
2360		sh->log_start = MaxSector;
2361
2362		if (raid5_has_ppl(conf)) {
2363			sh->ppl_page = alloc_page(gfp);
2364			if (!sh->ppl_page) {
2365				free_stripe(sc, sh);
2366				return NULL;
2367			}
2368		}
2369#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2370		if (init_stripe_shared_pages(sh, conf, disks)) {
2371			free_stripe(sc, sh);
2372			return NULL;
2373		}
2374#endif
2375	}
2376	return sh;
2377}
2378static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2379{
2380	struct stripe_head *sh;
2381
2382	sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2383	if (!sh)
2384		return 0;
2385
2386	if (grow_buffers(sh, gfp)) {
2387		shrink_buffers(sh);
2388		free_stripe(conf->slab_cache, sh);
2389		return 0;
2390	}
2391	sh->hash_lock_index =
2392		conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2393	/* we just created an active stripe so... */
2394	atomic_inc(&conf->active_stripes);
2395
2396	raid5_release_stripe(sh);
2397	WRITE_ONCE(conf->max_nr_stripes, conf->max_nr_stripes + 1);
2398	return 1;
2399}
2400
2401static int grow_stripes(struct r5conf *conf, int num)
2402{
2403	struct kmem_cache *sc;
2404	size_t namelen = sizeof(conf->cache_name[0]);
2405	int devs = max(conf->raid_disks, conf->previous_raid_disks);
2406
2407	if (mddev_is_dm(conf->mddev))
2408		snprintf(conf->cache_name[0], namelen,
2409			"raid%d-%p", conf->level, conf->mddev);
2410	else
2411		snprintf(conf->cache_name[0], namelen,
2412			"raid%d-%s", conf->level, mdname(conf->mddev));
2413	snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2414
2415	conf->active_name = 0;
2416	sc = kmem_cache_create(conf->cache_name[conf->active_name],
2417			       struct_size_t(struct stripe_head, dev, devs),
2418			       0, 0, NULL);
2419	if (!sc)
2420		return 1;
2421	conf->slab_cache = sc;
2422	conf->pool_size = devs;
2423	while (num--)
2424		if (!grow_one_stripe(conf, GFP_KERNEL))
2425			return 1;
2426
2427	return 0;
2428}
2429
2430/**
2431 * scribble_alloc - allocate percpu scribble buffer for required size
2432 *		    of the scribble region
2433 * @percpu: from for_each_present_cpu() of the caller
2434 * @num: total number of disks in the array
2435 * @cnt: scribble objs count for required size of the scribble region
2436 *
2437 * The scribble buffer size must be enough to contain:
2438 * 1/ a struct page pointer for each device in the array +2
2439 * 2/ room to convert each entry in (1) to its corresponding dma
2440 *    (dma_map_page()) or page (page_address()) address.
2441 *
2442 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2443 * calculate over all devices (not just the data blocks), using zeros in place
2444 * of the P and Q blocks.
2445 */
2446static int scribble_alloc(struct raid5_percpu *percpu,
2447			  int num, int cnt)
2448{
2449	size_t obj_size =
2450		sizeof(struct page *) * (num + 2) +
2451		sizeof(addr_conv_t) * (num + 2) +
2452		sizeof(unsigned int) * (num + 2);
2453	void *scribble;
2454
2455	/*
2456	 * If here is in raid array suspend context, it is in memalloc noio
2457	 * context as well, there is no potential recursive memory reclaim
2458	 * I/Os with the GFP_KERNEL flag.
2459	 */
2460	scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2461	if (!scribble)
2462		return -ENOMEM;
2463
2464	kvfree(percpu->scribble);
2465
2466	percpu->scribble = scribble;
2467	percpu->scribble_obj_size = obj_size;
2468	return 0;
2469}
2470
2471static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2472{
2473	unsigned long cpu;
2474	int err = 0;
2475
2476	/* Never shrink. */
2477	if (conf->scribble_disks >= new_disks &&
2478	    conf->scribble_sectors >= new_sectors)
2479		return 0;
2480
2481	raid5_quiesce(conf->mddev, true);
2482	cpus_read_lock();
2483
2484	for_each_present_cpu(cpu) {
2485		struct raid5_percpu *percpu;
2486
2487		percpu = per_cpu_ptr(conf->percpu, cpu);
2488		err = scribble_alloc(percpu, new_disks,
2489				     new_sectors / RAID5_STRIPE_SECTORS(conf));
2490		if (err)
2491			break;
2492	}
2493
2494	cpus_read_unlock();
2495	raid5_quiesce(conf->mddev, false);
2496
2497	if (!err) {
2498		conf->scribble_disks = new_disks;
2499		conf->scribble_sectors = new_sectors;
2500	}
2501	return err;
2502}
2503
2504static int resize_stripes(struct r5conf *conf, int newsize)
2505{
2506	/* Make all the stripes able to hold 'newsize' devices.
2507	 * New slots in each stripe get 'page' set to a new page.
2508	 *
2509	 * This happens in stages:
2510	 * 1/ create a new kmem_cache and allocate the required number of
2511	 *    stripe_heads.
2512	 * 2/ gather all the old stripe_heads and transfer the pages across
2513	 *    to the new stripe_heads.  This will have the side effect of
2514	 *    freezing the array as once all stripe_heads have been collected,
2515	 *    no IO will be possible.  Old stripe heads are freed once their
2516	 *    pages have been transferred over, and the old kmem_cache is
2517	 *    freed when all stripes are done.
2518	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2519	 *    we simple return a failure status - no need to clean anything up.
2520	 * 4/ allocate new pages for the new slots in the new stripe_heads.
2521	 *    If this fails, we don't bother trying the shrink the
2522	 *    stripe_heads down again, we just leave them as they are.
2523	 *    As each stripe_head is processed the new one is released into
2524	 *    active service.
2525	 *
2526	 * Once step2 is started, we cannot afford to wait for a write,
2527	 * so we use GFP_NOIO allocations.
2528	 */
2529	struct stripe_head *osh, *nsh;
2530	LIST_HEAD(newstripes);
2531	struct disk_info *ndisks;
2532	int err = 0;
2533	struct kmem_cache *sc;
2534	int i;
2535	int hash, cnt;
2536
2537	md_allow_write(conf->mddev);
2538
2539	/* Step 1 */
2540	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2541			       struct_size_t(struct stripe_head, dev, newsize),
2542			       0, 0, NULL);
2543	if (!sc)
2544		return -ENOMEM;
2545
2546	/* Need to ensure auto-resizing doesn't interfere */
2547	mutex_lock(&conf->cache_size_mutex);
2548
2549	for (i = conf->max_nr_stripes; i; i--) {
2550		nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2551		if (!nsh)
2552			break;
2553
2554		list_add(&nsh->lru, &newstripes);
2555	}
2556	if (i) {
2557		/* didn't get enough, give up */
2558		while (!list_empty(&newstripes)) {
2559			nsh = list_entry(newstripes.next, struct stripe_head, lru);
2560			list_del(&nsh->lru);
2561			free_stripe(sc, nsh);
2562		}
2563		kmem_cache_destroy(sc);
2564		mutex_unlock(&conf->cache_size_mutex);
2565		return -ENOMEM;
2566	}
2567	/* Step 2 - Must use GFP_NOIO now.
2568	 * OK, we have enough stripes, start collecting inactive
2569	 * stripes and copying them over
2570	 */
2571	hash = 0;
2572	cnt = 0;
2573	list_for_each_entry(nsh, &newstripes, lru) {
2574		lock_device_hash_lock(conf, hash);
2575		wait_event_cmd(conf->wait_for_stripe,
2576				    !list_empty(conf->inactive_list + hash),
2577				    unlock_device_hash_lock(conf, hash),
2578				    lock_device_hash_lock(conf, hash));
2579		osh = get_free_stripe(conf, hash);
2580		unlock_device_hash_lock(conf, hash);
2581
2582#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2583	for (i = 0; i < osh->nr_pages; i++) {
2584		nsh->pages[i] = osh->pages[i];
2585		osh->pages[i] = NULL;
2586	}
2587#endif
2588		for(i=0; i<conf->pool_size; i++) {
2589			nsh->dev[i].page = osh->dev[i].page;
2590			nsh->dev[i].orig_page = osh->dev[i].page;
2591			nsh->dev[i].offset = osh->dev[i].offset;
2592		}
2593		nsh->hash_lock_index = hash;
2594		free_stripe(conf->slab_cache, osh);
2595		cnt++;
2596		if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2597		    !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2598			hash++;
2599			cnt = 0;
2600		}
2601	}
2602	kmem_cache_destroy(conf->slab_cache);
2603
2604	/* Step 3.
2605	 * At this point, we are holding all the stripes so the array
2606	 * is completely stalled, so now is a good time to resize
2607	 * conf->disks and the scribble region
2608	 */
2609	ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2610	if (ndisks) {
2611		for (i = 0; i < conf->pool_size; i++)
2612			ndisks[i] = conf->disks[i];
2613
2614		for (i = conf->pool_size; i < newsize; i++) {
2615			ndisks[i].extra_page = alloc_page(GFP_NOIO);
2616			if (!ndisks[i].extra_page)
2617				err = -ENOMEM;
2618		}
2619
2620		if (err) {
2621			for (i = conf->pool_size; i < newsize; i++)
2622				if (ndisks[i].extra_page)
2623					put_page(ndisks[i].extra_page);
2624			kfree(ndisks);
2625		} else {
2626			kfree(conf->disks);
2627			conf->disks = ndisks;
2628		}
2629	} else
2630		err = -ENOMEM;
2631
2632	conf->slab_cache = sc;
2633	conf->active_name = 1-conf->active_name;
2634
2635	/* Step 4, return new stripes to service */
2636	while(!list_empty(&newstripes)) {
2637		nsh = list_entry(newstripes.next, struct stripe_head, lru);
2638		list_del_init(&nsh->lru);
2639
2640#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2641		for (i = 0; i < nsh->nr_pages; i++) {
2642			if (nsh->pages[i])
2643				continue;
2644			nsh->pages[i] = alloc_page(GFP_NOIO);
2645			if (!nsh->pages[i])
2646				err = -ENOMEM;
2647		}
2648
2649		for (i = conf->raid_disks; i < newsize; i++) {
2650			if (nsh->dev[i].page)
2651				continue;
2652			nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2653			nsh->dev[i].orig_page = nsh->dev[i].page;
2654			nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2655		}
2656#else
2657		for (i=conf->raid_disks; i < newsize; i++)
2658			if (nsh->dev[i].page == NULL) {
2659				struct page *p = alloc_page(GFP_NOIO);
2660				nsh->dev[i].page = p;
2661				nsh->dev[i].orig_page = p;
2662				nsh->dev[i].offset = 0;
2663				if (!p)
2664					err = -ENOMEM;
2665			}
2666#endif
2667		raid5_release_stripe(nsh);
2668	}
2669	/* critical section pass, GFP_NOIO no longer needed */
2670
2671	if (!err)
2672		conf->pool_size = newsize;
2673	mutex_unlock(&conf->cache_size_mutex);
2674
2675	return err;
2676}
2677
2678static int drop_one_stripe(struct r5conf *conf)
2679{
2680	struct stripe_head *sh;
2681	int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2682
2683	spin_lock_irq(conf->hash_locks + hash);
2684	sh = get_free_stripe(conf, hash);
2685	spin_unlock_irq(conf->hash_locks + hash);
2686	if (!sh)
2687		return 0;
2688	BUG_ON(atomic_read(&sh->count));
2689	shrink_buffers(sh);
2690	free_stripe(conf->slab_cache, sh);
2691	atomic_dec(&conf->active_stripes);
2692	WRITE_ONCE(conf->max_nr_stripes, conf->max_nr_stripes - 1);
2693	return 1;
2694}
2695
2696static void shrink_stripes(struct r5conf *conf)
2697{
2698	while (conf->max_nr_stripes &&
2699	       drop_one_stripe(conf))
2700		;
2701
2702	kmem_cache_destroy(conf->slab_cache);
2703	conf->slab_cache = NULL;
2704}
2705
2706static void raid5_end_read_request(struct bio * bi)
2707{
2708	struct stripe_head *sh = bi->bi_private;
2709	struct r5conf *conf = sh->raid_conf;
2710	int disks = sh->disks, i;
2711	struct md_rdev *rdev = NULL;
2712	sector_t s;
2713
2714	for (i=0 ; i<disks; i++)
2715		if (bi == &sh->dev[i].req)
2716			break;
2717
2718	pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2719		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2720		bi->bi_status);
2721	if (i == disks) {
2722		BUG();
2723		return;
2724	}
2725	if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2726		/* If replacement finished while this request was outstanding,
2727		 * 'replacement' might be NULL already.
2728		 * In that case it moved down to 'rdev'.
2729		 * rdev is not removed until all requests are finished.
2730		 */
2731		rdev = conf->disks[i].replacement;
2732	if (!rdev)
2733		rdev = conf->disks[i].rdev;
2734
2735	if (use_new_offset(conf, sh))
2736		s = sh->sector + rdev->new_data_offset;
2737	else
2738		s = sh->sector + rdev->data_offset;
2739	if (!bi->bi_status) {
2740		set_bit(R5_UPTODATE, &sh->dev[i].flags);
2741		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2742			/* Note that this cannot happen on a
2743			 * replacement device.  We just fail those on
2744			 * any error
2745			 */
2746			pr_info_ratelimited(
2747				"md/raid:%s: read error corrected (%lu sectors at %llu on %pg)\n",
2748				mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2749				(unsigned long long)s,
2750				rdev->bdev);
2751			atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2752			clear_bit(R5_ReadError, &sh->dev[i].flags);
2753			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2754		} else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2755			clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2756
2757		if (test_bit(R5_InJournal, &sh->dev[i].flags))
2758			/*
2759			 * end read for a page in journal, this
2760			 * must be preparing for prexor in rmw
2761			 */
2762			set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2763
2764		if (atomic_read(&rdev->read_errors))
2765			atomic_set(&rdev->read_errors, 0);
2766	} else {
2767		int retry = 0;
2768		int set_bad = 0;
2769
2770		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2771		if (!(bi->bi_status == BLK_STS_PROTECTION))
2772			atomic_inc(&rdev->read_errors);
2773		if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2774			pr_warn_ratelimited(
2775				"md/raid:%s: read error on replacement device (sector %llu on %pg).\n",
2776				mdname(conf->mddev),
2777				(unsigned long long)s,
2778				rdev->bdev);
2779		else if (conf->mddev->degraded >= conf->max_degraded) {
2780			set_bad = 1;
2781			pr_warn_ratelimited(
2782				"md/raid:%s: read error not correctable (sector %llu on %pg).\n",
2783				mdname(conf->mddev),
2784				(unsigned long long)s,
2785				rdev->bdev);
2786		} else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2787			/* Oh, no!!! */
2788			set_bad = 1;
2789			pr_warn_ratelimited(
2790				"md/raid:%s: read error NOT corrected!! (sector %llu on %pg).\n",
2791				mdname(conf->mddev),
2792				(unsigned long long)s,
2793				rdev->bdev);
2794		} else if (atomic_read(&rdev->read_errors)
2795			 > conf->max_nr_stripes) {
2796			if (!test_bit(Faulty, &rdev->flags)) {
2797				pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2798				    mdname(conf->mddev),
2799				    atomic_read(&rdev->read_errors),
2800				    conf->max_nr_stripes);
2801				pr_warn("md/raid:%s: Too many read errors, failing device %pg.\n",
2802				    mdname(conf->mddev), rdev->bdev);
2803			}
2804		} else
2805			retry = 1;
2806		if (set_bad && test_bit(In_sync, &rdev->flags)
2807		    && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2808			retry = 1;
2809		if (retry)
2810			if (sh->qd_idx >= 0 && sh->pd_idx == i)
2811				set_bit(R5_ReadError, &sh->dev[i].flags);
2812			else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2813				set_bit(R5_ReadError, &sh->dev[i].flags);
2814				clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2815			} else
2816				set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2817		else {
2818			clear_bit(R5_ReadError, &sh->dev[i].flags);
2819			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2820			if (!(set_bad
2821			      && test_bit(In_sync, &rdev->flags)
2822			      && rdev_set_badblocks(
2823				      rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2824				md_error(conf->mddev, rdev);
2825		}
2826	}
2827	rdev_dec_pending(rdev, conf->mddev);
2828	bio_uninit(bi);
2829	clear_bit(R5_LOCKED, &sh->dev[i].flags);
2830	set_bit(STRIPE_HANDLE, &sh->state);
2831	raid5_release_stripe(sh);
2832}
2833
2834static void raid5_end_write_request(struct bio *bi)
2835{
2836	struct stripe_head *sh = bi->bi_private;
2837	struct r5conf *conf = sh->raid_conf;
2838	int disks = sh->disks, i;
2839	struct md_rdev *rdev;
2840	int replacement = 0;
2841
2842	for (i = 0 ; i < disks; i++) {
2843		if (bi == &sh->dev[i].req) {
2844			rdev = conf->disks[i].rdev;
2845			break;
2846		}
2847		if (bi == &sh->dev[i].rreq) {
2848			rdev = conf->disks[i].replacement;
2849			if (rdev)
2850				replacement = 1;
2851			else
2852				/* rdev was removed and 'replacement'
2853				 * replaced it.  rdev is not removed
2854				 * until all requests are finished.
2855				 */
2856				rdev = conf->disks[i].rdev;
2857			break;
2858		}
2859	}
2860	pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2861		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2862		bi->bi_status);
2863	if (i == disks) {
2864		BUG();
2865		return;
2866	}
2867
2868	if (replacement) {
2869		if (bi->bi_status)
2870			md_error(conf->mddev, rdev);
2871		else if (rdev_has_badblock(rdev, sh->sector,
2872					   RAID5_STRIPE_SECTORS(conf)))
2873			set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2874	} else {
2875		if (bi->bi_status) {
2876			set_bit(WriteErrorSeen, &rdev->flags);
2877			set_bit(R5_WriteError, &sh->dev[i].flags);
2878			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2879				set_bit(MD_RECOVERY_NEEDED,
2880					&rdev->mddev->recovery);
2881		} else if (rdev_has_badblock(rdev, sh->sector,
2882					     RAID5_STRIPE_SECTORS(conf))) {
2883			set_bit(R5_MadeGood, &sh->dev[i].flags);
2884			if (test_bit(R5_ReadError, &sh->dev[i].flags))
2885				/* That was a successful write so make
2886				 * sure it looks like we already did
2887				 * a re-write.
2888				 */
2889				set_bit(R5_ReWrite, &sh->dev[i].flags);
2890		}
2891	}
2892	rdev_dec_pending(rdev, conf->mddev);
2893
2894	if (sh->batch_head && bi->bi_status && !replacement)
2895		set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2896
2897	bio_uninit(bi);
2898	if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2899		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2900	set_bit(STRIPE_HANDLE, &sh->state);
2901
2902	if (sh->batch_head && sh != sh->batch_head)
2903		raid5_release_stripe(sh->batch_head);
2904	raid5_release_stripe(sh);
2905}
2906
2907static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2908{
2909	struct r5conf *conf = mddev->private;
2910	unsigned long flags;
2911	pr_debug("raid456: error called\n");
2912
2913	pr_crit("md/raid:%s: Disk failure on %pg, disabling device.\n",
2914		mdname(mddev), rdev->bdev);
2915
2916	spin_lock_irqsave(&conf->device_lock, flags);
2917	set_bit(Faulty, &rdev->flags);
2918	clear_bit(In_sync, &rdev->flags);
2919	mddev->degraded = raid5_calc_degraded(conf);
2920
2921	if (has_failed(conf)) {
2922		set_bit(MD_BROKEN, &conf->mddev->flags);
2923		conf->recovery_disabled = mddev->recovery_disabled;
2924
2925		pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2926			mdname(mddev), mddev->degraded, conf->raid_disks);
2927	} else {
2928		pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2929			mdname(mddev), conf->raid_disks - mddev->degraded);
2930	}
2931
2932	spin_unlock_irqrestore(&conf->device_lock, flags);
2933	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2934
2935	set_bit(Blocked, &rdev->flags);
2936	set_mask_bits(&mddev->sb_flags, 0,
2937		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2938	r5c_update_on_rdev_error(mddev, rdev);
2939}
2940
2941/*
2942 * Input: a 'big' sector number,
2943 * Output: index of the data and parity disk, and the sector # in them.
2944 */
2945sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2946			      int previous, int *dd_idx,
2947			      struct stripe_head *sh)
2948{
2949	sector_t stripe, stripe2;
2950	sector_t chunk_number;
2951	unsigned int chunk_offset;
2952	int pd_idx, qd_idx;
2953	int ddf_layout = 0;
2954	sector_t new_sector;
2955	int algorithm = previous ? conf->prev_algo
2956				 : conf->algorithm;
2957	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2958					 : conf->chunk_sectors;
2959	int raid_disks = previous ? conf->previous_raid_disks
2960				  : conf->raid_disks;
2961	int data_disks = raid_disks - conf->max_degraded;
2962
2963	/* First compute the information on this sector */
2964
2965	/*
2966	 * Compute the chunk number and the sector offset inside the chunk
2967	 */
2968	chunk_offset = sector_div(r_sector, sectors_per_chunk);
2969	chunk_number = r_sector;
2970
2971	/*
2972	 * Compute the stripe number
2973	 */
2974	stripe = chunk_number;
2975	*dd_idx = sector_div(stripe, data_disks);
2976	stripe2 = stripe;
2977	/*
2978	 * Select the parity disk based on the user selected algorithm.
2979	 */
2980	pd_idx = qd_idx = -1;
2981	switch(conf->level) {
2982	case 4:
2983		pd_idx = data_disks;
2984		break;
2985	case 5:
2986		switch (algorithm) {
2987		case ALGORITHM_LEFT_ASYMMETRIC:
2988			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2989			if (*dd_idx >= pd_idx)
2990				(*dd_idx)++;
2991			break;
2992		case ALGORITHM_RIGHT_ASYMMETRIC:
2993			pd_idx = sector_div(stripe2, raid_disks);
2994			if (*dd_idx >= pd_idx)
2995				(*dd_idx)++;
2996			break;
2997		case ALGORITHM_LEFT_SYMMETRIC:
2998			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2999			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3000			break;
3001		case ALGORITHM_RIGHT_SYMMETRIC:
3002			pd_idx = sector_div(stripe2, raid_disks);
3003			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3004			break;
3005		case ALGORITHM_PARITY_0:
3006			pd_idx = 0;
3007			(*dd_idx)++;
3008			break;
3009		case ALGORITHM_PARITY_N:
3010			pd_idx = data_disks;
3011			break;
3012		default:
3013			BUG();
3014		}
3015		break;
3016	case 6:
3017
3018		switch (algorithm) {
3019		case ALGORITHM_LEFT_ASYMMETRIC:
3020			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3021			qd_idx = pd_idx + 1;
3022			if (pd_idx == raid_disks-1) {
3023				(*dd_idx)++;	/* Q D D D P */
3024				qd_idx = 0;
3025			} else if (*dd_idx >= pd_idx)
3026				(*dd_idx) += 2; /* D D P Q D */
3027			break;
3028		case ALGORITHM_RIGHT_ASYMMETRIC:
3029			pd_idx = sector_div(stripe2, raid_disks);
3030			qd_idx = pd_idx + 1;
3031			if (pd_idx == raid_disks-1) {
3032				(*dd_idx)++;	/* Q D D D P */
3033				qd_idx = 0;
3034			} else if (*dd_idx >= pd_idx)
3035				(*dd_idx) += 2; /* D D P Q D */
3036			break;
3037		case ALGORITHM_LEFT_SYMMETRIC:
3038			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3039			qd_idx = (pd_idx + 1) % raid_disks;
3040			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3041			break;
3042		case ALGORITHM_RIGHT_SYMMETRIC:
3043			pd_idx = sector_div(stripe2, raid_disks);
3044			qd_idx = (pd_idx + 1) % raid_disks;
3045			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3046			break;
3047
3048		case ALGORITHM_PARITY_0:
3049			pd_idx = 0;
3050			qd_idx = 1;
3051			(*dd_idx) += 2;
3052			break;
3053		case ALGORITHM_PARITY_N:
3054			pd_idx = data_disks;
3055			qd_idx = data_disks + 1;
3056			break;
3057
3058		case ALGORITHM_ROTATING_ZERO_RESTART:
3059			/* Exactly the same as RIGHT_ASYMMETRIC, but or
3060			 * of blocks for computing Q is different.
3061			 */
3062			pd_idx = sector_div(stripe2, raid_disks);
3063			qd_idx = pd_idx + 1;
3064			if (pd_idx == raid_disks-1) {
3065				(*dd_idx)++;	/* Q D D D P */
3066				qd_idx = 0;
3067			} else if (*dd_idx >= pd_idx)
3068				(*dd_idx) += 2; /* D D P Q D */
3069			ddf_layout = 1;
3070			break;
3071
3072		case ALGORITHM_ROTATING_N_RESTART:
3073			/* Same a left_asymmetric, by first stripe is
3074			 * D D D P Q  rather than
3075			 * Q D D D P
3076			 */
3077			stripe2 += 1;
3078			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3079			qd_idx = pd_idx + 1;
3080			if (pd_idx == raid_disks-1) {
3081				(*dd_idx)++;	/* Q D D D P */
3082				qd_idx = 0;
3083			} else if (*dd_idx >= pd_idx)
3084				(*dd_idx) += 2; /* D D P Q D */
3085			ddf_layout = 1;
3086			break;
3087
3088		case ALGORITHM_ROTATING_N_CONTINUE:
3089			/* Same as left_symmetric but Q is before P */
3090			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3091			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3092			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3093			ddf_layout = 1;
3094			break;
3095
3096		case ALGORITHM_LEFT_ASYMMETRIC_6:
3097			/* RAID5 left_asymmetric, with Q on last device */
3098			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3099			if (*dd_idx >= pd_idx)
3100				(*dd_idx)++;
3101			qd_idx = raid_disks - 1;
3102			break;
3103
3104		case ALGORITHM_RIGHT_ASYMMETRIC_6:
3105			pd_idx = sector_div(stripe2, raid_disks-1);
3106			if (*dd_idx >= pd_idx)
3107				(*dd_idx)++;
3108			qd_idx = raid_disks - 1;
3109			break;
3110
3111		case ALGORITHM_LEFT_SYMMETRIC_6:
3112			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3113			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3114			qd_idx = raid_disks - 1;
3115			break;
3116
3117		case ALGORITHM_RIGHT_SYMMETRIC_6:
3118			pd_idx = sector_div(stripe2, raid_disks-1);
3119			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3120			qd_idx = raid_disks - 1;
3121			break;
3122
3123		case ALGORITHM_PARITY_0_6:
3124			pd_idx = 0;
3125			(*dd_idx)++;
3126			qd_idx = raid_disks - 1;
3127			break;
3128
3129		default:
3130			BUG();
3131		}
3132		break;
3133	}
3134
3135	if (sh) {
3136		sh->pd_idx = pd_idx;
3137		sh->qd_idx = qd_idx;
3138		sh->ddf_layout = ddf_layout;
3139	}
3140	/*
3141	 * Finally, compute the new sector number
3142	 */
3143	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3144	return new_sector;
3145}
3146
3147sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3148{
3149	struct r5conf *conf = sh->raid_conf;
3150	int raid_disks = sh->disks;
3151	int data_disks = raid_disks - conf->max_degraded;
3152	sector_t new_sector = sh->sector, check;
3153	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3154					 : conf->chunk_sectors;
3155	int algorithm = previous ? conf->prev_algo
3156				 : conf->algorithm;
3157	sector_t stripe;
3158	int chunk_offset;
3159	sector_t chunk_number;
3160	int dummy1, dd_idx = i;
3161	sector_t r_sector;
3162	struct stripe_head sh2;
3163
3164	chunk_offset = sector_div(new_sector, sectors_per_chunk);
3165	stripe = new_sector;
3166
3167	if (i == sh->pd_idx)
3168		return 0;
3169	switch(conf->level) {
3170	case 4: break;
3171	case 5:
3172		switch (algorithm) {
3173		case ALGORITHM_LEFT_ASYMMETRIC:
3174		case ALGORITHM_RIGHT_ASYMMETRIC:
3175			if (i > sh->pd_idx)
3176				i--;
3177			break;
3178		case ALGORITHM_LEFT_SYMMETRIC:
3179		case ALGORITHM_RIGHT_SYMMETRIC:
3180			if (i < sh->pd_idx)
3181				i += raid_disks;
3182			i -= (sh->pd_idx + 1);
3183			break;
3184		case ALGORITHM_PARITY_0:
3185			i -= 1;
3186			break;
3187		case ALGORITHM_PARITY_N:
3188			break;
3189		default:
3190			BUG();
3191		}
3192		break;
3193	case 6:
3194		if (i == sh->qd_idx)
3195			return 0; /* It is the Q disk */
3196		switch (algorithm) {
3197		case ALGORITHM_LEFT_ASYMMETRIC:
3198		case ALGORITHM_RIGHT_ASYMMETRIC:
3199		case ALGORITHM_ROTATING_ZERO_RESTART:
3200		case ALGORITHM_ROTATING_N_RESTART:
3201			if (sh->pd_idx == raid_disks-1)
3202				i--;	/* Q D D D P */
3203			else if (i > sh->pd_idx)
3204				i -= 2; /* D D P Q D */
3205			break;
3206		case ALGORITHM_LEFT_SYMMETRIC:
3207		case ALGORITHM_RIGHT_SYMMETRIC:
3208			if (sh->pd_idx == raid_disks-1)
3209				i--; /* Q D D D P */
3210			else {
3211				/* D D P Q D */
3212				if (i < sh->pd_idx)
3213					i += raid_disks;
3214				i -= (sh->pd_idx + 2);
3215			}
3216			break;
3217		case ALGORITHM_PARITY_0:
3218			i -= 2;
3219			break;
3220		case ALGORITHM_PARITY_N:
3221			break;
3222		case ALGORITHM_ROTATING_N_CONTINUE:
3223			/* Like left_symmetric, but P is before Q */
3224			if (sh->pd_idx == 0)
3225				i--;	/* P D D D Q */
3226			else {
3227				/* D D Q P D */
3228				if (i < sh->pd_idx)
3229					i += raid_disks;
3230				i -= (sh->pd_idx + 1);
3231			}
3232			break;
3233		case ALGORITHM_LEFT_ASYMMETRIC_6:
3234		case ALGORITHM_RIGHT_ASYMMETRIC_6:
3235			if (i > sh->pd_idx)
3236				i--;
3237			break;
3238		case ALGORITHM_LEFT_SYMMETRIC_6:
3239		case ALGORITHM_RIGHT_SYMMETRIC_6:
3240			if (i < sh->pd_idx)
3241				i += data_disks + 1;
3242			i -= (sh->pd_idx + 1);
3243			break;
3244		case ALGORITHM_PARITY_0_6:
3245			i -= 1;
3246			break;
3247		default:
3248			BUG();
3249		}
3250		break;
3251	}
3252
3253	chunk_number = stripe * data_disks + i;
3254	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3255
3256	check = raid5_compute_sector(conf, r_sector,
3257				     previous, &dummy1, &sh2);
3258	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3259		|| sh2.qd_idx != sh->qd_idx) {
3260		pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3261			mdname(conf->mddev));
3262		return 0;
3263	}
3264	return r_sector;
3265}
3266
3267/*
3268 * There are cases where we want handle_stripe_dirtying() and
3269 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3270 *
3271 * This function checks whether we want to delay the towrite. Specifically,
3272 * we delay the towrite when:
3273 *
3274 *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3275 *      stripe has data in journal (for other devices).
3276 *
3277 *      In this case, when reading data for the non-overwrite dev, it is
3278 *      necessary to handle complex rmw of write back cache (prexor with
3279 *      orig_page, and xor with page). To keep read path simple, we would
3280 *      like to flush data in journal to RAID disks first, so complex rmw
3281 *      is handled in the write patch (handle_stripe_dirtying).
3282 *
3283 *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3284 *
3285 *      It is important to be able to flush all stripes in raid5-cache.
3286 *      Therefore, we need reserve some space on the journal device for
3287 *      these flushes. If flush operation includes pending writes to the
3288 *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3289 *      for the flush out. If we exclude these pending writes from flush
3290 *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3291 *      Therefore, excluding pending writes in these cases enables more
3292 *      efficient use of the journal device.
3293 *
3294 *      Note: To make sure the stripe makes progress, we only delay
3295 *      towrite for stripes with data already in journal (injournal > 0).
3296 *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3297 *      no_space_stripes list.
3298 *
3299 *   3. during journal failure
3300 *      In journal failure, we try to flush all cached data to raid disks
3301 *      based on data in stripe cache. The array is read-only to upper
3302 *      layers, so we would skip all pending writes.
3303 *
3304 */
3305static inline bool delay_towrite(struct r5conf *conf,
3306				 struct r5dev *dev,
3307				 struct stripe_head_state *s)
3308{
3309	/* case 1 above */
3310	if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3311	    !test_bit(R5_Insync, &dev->flags) && s->injournal)
3312		return true;
3313	/* case 2 above */
3314	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3315	    s->injournal > 0)
3316		return true;
3317	/* case 3 above */
3318	if (s->log_failed && s->injournal)
3319		return true;
3320	return false;
3321}
3322
3323static void
3324schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3325			 int rcw, int expand)
3326{
3327	int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3328	struct r5conf *conf = sh->raid_conf;
3329	int level = conf->level;
3330
3331	if (rcw) {
3332		/*
3333		 * In some cases, handle_stripe_dirtying initially decided to
3334		 * run rmw and allocates extra page for prexor. However, rcw is
3335		 * cheaper later on. We need to free the extra page now,
3336		 * because we won't be able to do that in ops_complete_prexor().
3337		 */
3338		r5c_release_extra_page(sh);
3339
3340		for (i = disks; i--; ) {
3341			struct r5dev *dev = &sh->dev[i];
3342
3343			if (dev->towrite && !delay_towrite(conf, dev, s)) {
3344				set_bit(R5_LOCKED, &dev->flags);
3345				set_bit(R5_Wantdrain, &dev->flags);
3346				if (!expand)
3347					clear_bit(R5_UPTODATE, &dev->flags);
3348				s->locked++;
3349			} else if (test_bit(R5_InJournal, &dev->flags)) {
3350				set_bit(R5_LOCKED, &dev->flags);
3351				s->locked++;
3352			}
3353		}
3354		/* if we are not expanding this is a proper write request, and
3355		 * there will be bios with new data to be drained into the
3356		 * stripe cache
3357		 */
3358		if (!expand) {
3359			if (!s->locked)
3360				/* False alarm, nothing to do */
3361				return;
3362			sh->reconstruct_state = reconstruct_state_drain_run;
3363			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3364		} else
3365			sh->reconstruct_state = reconstruct_state_run;
3366
3367		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3368
3369		if (s->locked + conf->max_degraded == disks)
3370			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3371				atomic_inc(&conf->pending_full_writes);
3372	} else {
3373		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3374			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3375		BUG_ON(level == 6 &&
3376			(!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3377			   test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3378
3379		for (i = disks; i--; ) {
3380			struct r5dev *dev = &sh->dev[i];
3381			if (i == pd_idx || i == qd_idx)
3382				continue;
3383
3384			if (dev->towrite &&
3385			    (test_bit(R5_UPTODATE, &dev->flags) ||
3386			     test_bit(R5_Wantcompute, &dev->flags))) {
3387				set_bit(R5_Wantdrain, &dev->flags);
3388				set_bit(R5_LOCKED, &dev->flags);
3389				clear_bit(R5_UPTODATE, &dev->flags);
3390				s->locked++;
3391			} else if (test_bit(R5_InJournal, &dev->flags)) {
3392				set_bit(R5_LOCKED, &dev->flags);
3393				s->locked++;
3394			}
3395		}
3396		if (!s->locked)
3397			/* False alarm - nothing to do */
3398			return;
3399		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3400		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3401		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3402		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3403	}
3404
3405	/* keep the parity disk(s) locked while asynchronous operations
3406	 * are in flight
3407	 */
3408	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3409	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3410	s->locked++;
3411
3412	if (level == 6) {
3413		int qd_idx = sh->qd_idx;
3414		struct r5dev *dev = &sh->dev[qd_idx];
3415
3416		set_bit(R5_LOCKED, &dev->flags);
3417		clear_bit(R5_UPTODATE, &dev->flags);
3418		s->locked++;
3419	}
3420
3421	if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3422	    test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3423	    !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3424	    test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3425		set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3426
3427	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3428		__func__, (unsigned long long)sh->sector,
3429		s->locked, s->ops_request);
3430}
3431
3432static bool stripe_bio_overlaps(struct stripe_head *sh, struct bio *bi,
3433				int dd_idx, int forwrite)
3434{
3435	struct r5conf *conf = sh->raid_conf;
3436	struct bio **bip;
3437
3438	pr_debug("checking bi b#%llu to stripe s#%llu\n",
3439		 bi->bi_iter.bi_sector, sh->sector);
3440
3441	/* Don't allow new IO added to stripes in batch list */
3442	if (sh->batch_head)
3443		return true;
3444
3445	if (forwrite)
3446		bip = &sh->dev[dd_idx].towrite;
3447	else
3448		bip = &sh->dev[dd_idx].toread;
3449
3450	while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3451		if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3452			return true;
3453		bip = &(*bip)->bi_next;
3454	}
3455
3456	if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3457		return true;
3458
3459	if (forwrite && raid5_has_ppl(conf)) {
3460		/*
3461		 * With PPL only writes to consecutive data chunks within a
3462		 * stripe are allowed because for a single stripe_head we can
3463		 * only have one PPL entry at a time, which describes one data
3464		 * range. Not really an overlap, but R5_Overlap can be
3465		 * used to handle this.
3466		 */
3467		sector_t sector;
3468		sector_t first = 0;
3469		sector_t last = 0;
3470		int count = 0;
3471		int i;
3472
3473		for (i = 0; i < sh->disks; i++) {
3474			if (i != sh->pd_idx &&
3475			    (i == dd_idx || sh->dev[i].towrite)) {
3476				sector = sh->dev[i].sector;
3477				if (count == 0 || sector < first)
3478					first = sector;
3479				if (sector > last)
3480					last = sector;
3481				count++;
3482			}
3483		}
3484
3485		if (first + conf->chunk_sectors * (count - 1) != last)
3486			return true;
3487	}
3488
3489	return false;
3490}
3491
3492static void __add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3493			     int dd_idx, int forwrite, int previous)
3494{
3495	struct r5conf *conf = sh->raid_conf;
3496	struct bio **bip;
3497	int firstwrite = 0;
3498
3499	if (forwrite) {
3500		bip = &sh->dev[dd_idx].towrite;
3501		if (!*bip)
3502			firstwrite = 1;
3503	} else {
3504		bip = &sh->dev[dd_idx].toread;
3505	}
3506
3507	while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector)
3508		bip = &(*bip)->bi_next;
3509
3510	if (!forwrite || previous)
3511		clear_bit(STRIPE_BATCH_READY, &sh->state);
3512
3513	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3514	if (*bip)
3515		bi->bi_next = *bip;
3516	*bip = bi;
3517	bio_inc_remaining(bi);
3518	md_write_inc(conf->mddev, bi);
3519
3520	if (forwrite) {
3521		/* check if page is covered */
3522		sector_t sector = sh->dev[dd_idx].sector;
3523		for (bi=sh->dev[dd_idx].towrite;
3524		     sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3525			     bi && bi->bi_iter.bi_sector <= sector;
3526		     bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3527			if (bio_end_sector(bi) >= sector)
3528				sector = bio_end_sector(bi);
3529		}
3530		if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3531			if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3532				sh->overwrite_disks++;
3533	}
3534
3535	pr_debug("added bi b#%llu to stripe s#%llu, disk %d, logical %llu\n",
3536		 (*bip)->bi_iter.bi_sector, sh->sector, dd_idx,
3537		 sh->dev[dd_idx].sector);
3538
3539	if (conf->mddev->bitmap && firstwrite && !sh->batch_head) {
3540		sh->bm_seq = conf->seq_flush+1;
3541		set_bit(STRIPE_BIT_DELAY, &sh->state);
3542	}
3543}
3544
3545/*
3546 * Each stripe/dev can have one or more bios attached.
3547 * toread/towrite point to the first in a chain.
3548 * The bi_next chain must be in order.
3549 */
3550static bool add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3551			   int dd_idx, int forwrite, int previous)
3552{
3553	spin_lock_irq(&sh->stripe_lock);
3554
3555	if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
3556		set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3557		spin_unlock_irq(&sh->stripe_lock);
3558		return false;
3559	}
3560
3561	__add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
3562	spin_unlock_irq(&sh->stripe_lock);
3563	return true;
3564}
3565
3566static void end_reshape(struct r5conf *conf);
3567
3568static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3569			    struct stripe_head *sh)
3570{
3571	int sectors_per_chunk =
3572		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3573	int dd_idx;
3574	int chunk_offset = sector_div(stripe, sectors_per_chunk);
3575	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3576
3577	raid5_compute_sector(conf,
3578			     stripe * (disks - conf->max_degraded)
3579			     *sectors_per_chunk + chunk_offset,
3580			     previous,
3581			     &dd_idx, sh);
3582}
3583
3584static void
3585handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3586		     struct stripe_head_state *s, int disks)
3587{
3588	int i;
3589	BUG_ON(sh->batch_head);
3590	for (i = disks; i--; ) {
3591		struct bio *bi;
3592
3593		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3594			struct md_rdev *rdev = conf->disks[i].rdev;
3595
3596			if (rdev && test_bit(In_sync, &rdev->flags) &&
3597			    !test_bit(Faulty, &rdev->flags))
3598				atomic_inc(&rdev->nr_pending);
3599			else
3600				rdev = NULL;
3601			if (rdev) {
3602				if (!rdev_set_badblocks(
3603					    rdev,
3604					    sh->sector,
3605					    RAID5_STRIPE_SECTORS(conf), 0))
3606					md_error(conf->mddev, rdev);
3607				rdev_dec_pending(rdev, conf->mddev);
3608			}
3609		}
3610		spin_lock_irq(&sh->stripe_lock);
3611		/* fail all writes first */
3612		bi = sh->dev[i].towrite;
3613		sh->dev[i].towrite = NULL;
3614		sh->overwrite_disks = 0;
3615		spin_unlock_irq(&sh->stripe_lock);
3616
3617		log_stripe_write_finished(sh);
3618
3619		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3620			wake_up_bit(&sh->dev[i].flags, R5_Overlap);
3621
3622		while (bi && bi->bi_iter.bi_sector <
3623			sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3624			struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3625
3626			md_write_end(conf->mddev);
3627			bio_io_error(bi);
3628			bi = nextbi;
3629		}
3630		/* and fail all 'written' */
3631		bi = sh->dev[i].written;
3632		sh->dev[i].written = NULL;
3633		if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3634			WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3635			sh->dev[i].page = sh->dev[i].orig_page;
3636		}
3637
3638		while (bi && bi->bi_iter.bi_sector <
3639		       sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3640			struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3641
3642			md_write_end(conf->mddev);
3643			bio_io_error(bi);
3644			bi = bi2;
3645		}
3646
3647		/* fail any reads if this device is non-operational and
3648		 * the data has not reached the cache yet.
3649		 */
3650		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3651		    s->failed > conf->max_degraded &&
3652		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3653		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
3654			spin_lock_irq(&sh->stripe_lock);
3655			bi = sh->dev[i].toread;
3656			sh->dev[i].toread = NULL;
3657			spin_unlock_irq(&sh->stripe_lock);
3658			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3659				wake_up_bit(&sh->dev[i].flags, R5_Overlap);
3660			if (bi)
3661				s->to_read--;
3662			while (bi && bi->bi_iter.bi_sector <
3663			       sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3664				struct bio *nextbi =
3665					r5_next_bio(conf, bi, sh->dev[i].sector);
3666
3667				bio_io_error(bi);
3668				bi = nextbi;
3669			}
3670		}
3671		/* If we were in the middle of a write the parity block might
3672		 * still be locked - so just clear all R5_LOCKED flags
3673		 */
3674		clear_bit(R5_LOCKED, &sh->dev[i].flags);
3675	}
3676	s->to_write = 0;
3677	s->written = 0;
3678
3679	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3680		if (atomic_dec_and_test(&conf->pending_full_writes))
3681			md_wakeup_thread(conf->mddev->thread);
3682}
3683
3684static void
3685handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3686		   struct stripe_head_state *s)
3687{
3688	int abort = 0;
3689	int i;
3690
3691	BUG_ON(sh->batch_head);
3692	clear_bit(STRIPE_SYNCING, &sh->state);
3693	if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3694		wake_up_bit(&sh->dev[sh->pd_idx].flags, R5_Overlap);
3695	s->syncing = 0;
3696	s->replacing = 0;
3697	/* There is nothing more to do for sync/check/repair.
3698	 * Don't even need to abort as that is handled elsewhere
3699	 * if needed, and not always wanted e.g. if there is a known
3700	 * bad block here.
3701	 * For recover/replace we need to record a bad block on all
3702	 * non-sync devices, or abort the recovery
3703	 */
3704	if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3705		/* During recovery devices cannot be removed, so
3706		 * locking and refcounting of rdevs is not needed
3707		 */
3708		for (i = 0; i < conf->raid_disks; i++) {
3709			struct md_rdev *rdev = conf->disks[i].rdev;
3710
3711			if (rdev
3712			    && !test_bit(Faulty, &rdev->flags)
3713			    && !test_bit(In_sync, &rdev->flags)
3714			    && !rdev_set_badblocks(rdev, sh->sector,
3715						   RAID5_STRIPE_SECTORS(conf), 0))
3716				abort = 1;
3717			rdev = conf->disks[i].replacement;
3718
3719			if (rdev
3720			    && !test_bit(Faulty, &rdev->flags)
3721			    && !test_bit(In_sync, &rdev->flags)
3722			    && !rdev_set_badblocks(rdev, sh->sector,
3723						   RAID5_STRIPE_SECTORS(conf), 0))
3724				abort = 1;
3725		}
3726		if (abort)
3727			conf->recovery_disabled =
3728				conf->mddev->recovery_disabled;
3729	}
3730	md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3731}
3732
3733static int want_replace(struct stripe_head *sh, int disk_idx)
3734{
3735	struct md_rdev *rdev;
3736	int rv = 0;
3737
3738	rdev = sh->raid_conf->disks[disk_idx].replacement;
3739	if (rdev
3740	    && !test_bit(Faulty, &rdev->flags)
3741	    && !test_bit(In_sync, &rdev->flags)
3742	    && (rdev->recovery_offset <= sh->sector
3743		|| rdev->mddev->resync_offset <= sh->sector))
3744		rv = 1;
3745	return rv;
3746}
3747
3748static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3749			   int disk_idx, int disks)
3750{
3751	struct r5dev *dev = &sh->dev[disk_idx];
3752	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3753				  &sh->dev[s->failed_num[1]] };
3754	struct mddev *mddev = sh->raid_conf->mddev;
3755	bool force_rcw = false;
3756	int i;
3757
3758	if (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW ||
3759	    (mddev->bitmap_ops && mddev->bitmap_ops->blocks_synced &&
3760	     !mddev->bitmap_ops->blocks_synced(mddev, sh->sector)))
3761		force_rcw = true;
3762
3763	if (test_bit(R5_LOCKED, &dev->flags) ||
3764	    test_bit(R5_UPTODATE, &dev->flags))
3765		/* No point reading this as we already have it or have
3766		 * decided to get it.
3767		 */
3768		return 0;
3769
3770	if (dev->toread ||
3771	    (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3772		/* We need this block to directly satisfy a request */
3773		return 1;
3774
3775	if (s->syncing || s->expanding ||
3776	    (s->replacing && want_replace(sh, disk_idx)))
3777		/* When syncing, or expanding we read everything.
3778		 * When replacing, we need the replaced block.
3779		 */
3780		return 1;
3781
3782	if ((s->failed >= 1 && fdev[0]->toread) ||
3783	    (s->failed >= 2 && fdev[1]->toread))
3784		/* If we want to read from a failed device, then
3785		 * we need to actually read every other device.
3786		 */
3787		return 1;
3788
3789	/* Sometimes neither read-modify-write nor reconstruct-write
3790	 * cycles can work.  In those cases we read every block we
3791	 * can.  Then the parity-update is certain to have enough to
3792	 * work with.
3793	 * This can only be a problem when we need to write something,
3794	 * and some device has failed.  If either of those tests
3795	 * fail we need look no further.
3796	 */
3797	if (!s->failed || !s->to_write)
3798		return 0;
3799
3800	if (test_bit(R5_Insync, &dev->flags) &&
3801	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3802		/* Pre-reads at not permitted until after short delay
3803		 * to gather multiple requests.  However if this
3804		 * device is no Insync, the block could only be computed
3805		 * and there is no need to delay that.
3806		 */
3807		return 0;
3808
3809	for (i = 0; i < s->failed && i < 2; i++) {
3810		if (fdev[i]->towrite &&
3811		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3812		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3813			/* If we have a partial write to a failed
3814			 * device, then we will need to reconstruct
3815			 * the content of that device, so all other
3816			 * devices must be read.
3817			 */
3818			return 1;
3819
3820		if (s->failed >= 2 &&
3821		    (fdev[i]->towrite ||
3822		     s->failed_num[i] == sh->pd_idx ||
3823		     s->failed_num[i] == sh->qd_idx) &&
3824		    !test_bit(R5_UPTODATE, &fdev[i]->flags))
3825			/* In max degraded raid6, If the failed disk is P, Q,
3826			 * or we want to read the failed disk, we need to do
3827			 * reconstruct-write.
3828			 */
3829			force_rcw = true;
3830	}
3831
3832	/* If we are forced to do a reconstruct-write, because parity
3833	 * cannot be trusted and we are currently recovering it, there
3834	 * is extra need to be careful.
3835	 * If one of the devices that we would need to read, because
3836	 * it is not being overwritten (and maybe not written at all)
3837	 * is missing/faulty, then we need to read everything we can.
3838	 */
3839	if (!force_rcw &&
3840	    sh->sector < sh->raid_conf->mddev->resync_offset)
3841		/* reconstruct-write isn't being forced */
3842		return 0;
3843	for (i = 0; i < s->failed && i < 2; i++) {
3844		if (s->failed_num[i] != sh->pd_idx &&
3845		    s->failed_num[i] != sh->qd_idx &&
3846		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3847		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3848			return 1;
3849	}
3850
3851	return 0;
3852}
3853
3854/* fetch_block - checks the given member device to see if its data needs
3855 * to be read or computed to satisfy a request.
3856 *
3857 * Returns 1 when no more member devices need to be checked, otherwise returns
3858 * 0 to tell the loop in handle_stripe_fill to continue
3859 */
3860static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3861		       int disk_idx, int disks)
3862{
3863	struct r5dev *dev = &sh->dev[disk_idx];
3864
3865	/* is the data in this block needed, and can we get it? */
3866	if (need_this_block(sh, s, disk_idx, disks)) {
3867		/* we would like to get this block, possibly by computing it,
3868		 * otherwise read it if the backing disk is insync
3869		 */
3870		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3871		BUG_ON(test_bit(R5_Wantread, &dev->flags));
3872		BUG_ON(sh->batch_head);
3873
3874		/*
3875		 * In the raid6 case if the only non-uptodate disk is P
3876		 * then we already trusted P to compute the other failed
3877		 * drives. It is safe to compute rather than re-read P.
3878		 * In other cases we only compute blocks from failed
3879		 * devices, otherwise check/repair might fail to detect
3880		 * a real inconsistency.
3881		 */
3882
3883		if ((s->uptodate == disks - 1) &&
3884		    ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3885		    (s->failed && (disk_idx == s->failed_num[0] ||
3886				   disk_idx == s->failed_num[1])))) {
3887			/* have disk failed, and we're requested to fetch it;
3888			 * do compute it
3889			 */
3890			pr_debug("Computing stripe %llu block %d\n",
3891			       (unsigned long long)sh->sector, disk_idx);
3892			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3893			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3894			set_bit(R5_Wantcompute, &dev->flags);
3895			sh->ops.target = disk_idx;
3896			sh->ops.target2 = -1; /* no 2nd target */
3897			s->req_compute = 1;
3898			/* Careful: from this point on 'uptodate' is in the eye
3899			 * of raid_run_ops which services 'compute' operations
3900			 * before writes. R5_Wantcompute flags a block that will
3901			 * be R5_UPTODATE by the time it is needed for a
3902			 * subsequent operation.
3903			 */
3904			s->uptodate++;
3905			return 1;
3906		} else if (s->uptodate == disks-2 && s->failed >= 2) {
3907			/* Computing 2-failure is *very* expensive; only
3908			 * do it if failed >= 2
3909			 */
3910			int other;
3911			for (other = disks; other--; ) {
3912				if (other == disk_idx)
3913					continue;
3914				if (!test_bit(R5_UPTODATE,
3915				      &sh->dev[other].flags))
3916					break;
3917			}
3918			BUG_ON(other < 0);
3919			pr_debug("Computing stripe %llu blocks %d,%d\n",
3920			       (unsigned long long)sh->sector,
3921			       disk_idx, other);
3922			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3923			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3924			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3925			set_bit(R5_Wantcompute, &sh->dev[other].flags);
3926			sh->ops.target = disk_idx;
3927			sh->ops.target2 = other;
3928			s->uptodate += 2;
3929			s->req_compute = 1;
3930			return 1;
3931		} else if (test_bit(R5_Insync, &dev->flags)) {
3932			set_bit(R5_LOCKED, &dev->flags);
3933			set_bit(R5_Wantread, &dev->flags);
3934			s->locked++;
3935			pr_debug("Reading block %d (sync=%d)\n",
3936				disk_idx, s->syncing);
3937		}
3938	}
3939
3940	return 0;
3941}
3942
3943/*
3944 * handle_stripe_fill - read or compute data to satisfy pending requests.
3945 */
3946static void handle_stripe_fill(struct stripe_head *sh,
3947			       struct stripe_head_state *s,
3948			       int disks)
3949{
3950	int i;
3951
3952	/* look for blocks to read/compute, skip this if a compute
3953	 * is already in flight, or if the stripe contents are in the
3954	 * midst of changing due to a write
3955	 */
3956	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3957	    !sh->reconstruct_state) {
3958
3959		/*
3960		 * For degraded stripe with data in journal, do not handle
3961		 * read requests yet, instead, flush the stripe to raid
3962		 * disks first, this avoids handling complex rmw of write
3963		 * back cache (prexor with orig_page, and then xor with
3964		 * page) in the read path
3965		 */
3966		if (s->to_read && s->injournal && s->failed) {
3967			if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3968				r5c_make_stripe_write_out(sh);
3969			goto out;
3970		}
3971
3972		for (i = disks; i--; )
3973			if (fetch_block(sh, s, i, disks))
3974				break;
3975	}
3976out:
3977	set_bit(STRIPE_HANDLE, &sh->state);
3978}
3979
3980static void break_stripe_batch_list(struct stripe_head *head_sh,
3981				    unsigned long handle_flags);
3982/* handle_stripe_clean_event
3983 * any written block on an uptodate or failed drive can be returned.
3984 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3985 * never LOCKED, so we don't need to test 'failed' directly.
3986 */
3987static void handle_stripe_clean_event(struct r5conf *conf,
3988	struct stripe_head *sh, int disks)
3989{
3990	int i;
3991	struct r5dev *dev;
3992	int discard_pending = 0;
3993	struct stripe_head *head_sh = sh;
3994	bool do_endio = false;
3995
3996	for (i = disks; i--; )
3997		if (sh->dev[i].written) {
3998			dev = &sh->dev[i];
3999			if (!test_bit(R5_LOCKED, &dev->flags) &&
4000			    (test_bit(R5_UPTODATE, &dev->flags) ||
4001			     test_bit(R5_Discard, &dev->flags) ||
4002			     test_bit(R5_SkipCopy, &dev->flags))) {
4003				/* We can return any write requests */
4004				struct bio *wbi, *wbi2;
4005				pr_debug("Return write for disc %d\n", i);
4006				if (test_and_clear_bit(R5_Discard, &dev->flags))
4007					clear_bit(R5_UPTODATE, &dev->flags);
4008				if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
4009					WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
4010				}
4011				do_endio = true;
4012
4013returnbi:
4014				dev->page = dev->orig_page;
4015				wbi = dev->written;
4016				dev->written = NULL;
4017				while (wbi && wbi->bi_iter.bi_sector <
4018					dev->sector + RAID5_STRIPE_SECTORS(conf)) {
4019					wbi2 = r5_next_bio(conf, wbi, dev->sector);
4020					md_write_end(conf->mddev);
4021					bio_endio(wbi);
4022					wbi = wbi2;
4023				}
4024
4025				if (head_sh->batch_head) {
4026					sh = list_first_entry(&sh->batch_list,
4027							      struct stripe_head,
4028							      batch_list);
4029					if (sh != head_sh) {
4030						dev = &sh->dev[i];
4031						goto returnbi;
4032					}
4033				}
4034				sh = head_sh;
4035				dev = &sh->dev[i];
4036			} else if (test_bit(R5_Discard, &dev->flags))
4037				discard_pending = 1;
4038		}
4039
4040	log_stripe_write_finished(sh);
4041
4042	if (!discard_pending &&
4043	    test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4044		int hash;
4045		clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4046		clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4047		if (sh->qd_idx >= 0) {
4048			clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4049			clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4050		}
4051		/* now that discard is done we can proceed with any sync */
4052		clear_bit(STRIPE_DISCARD, &sh->state);
4053		/*
4054		 * SCSI discard will change some bio fields and the stripe has
4055		 * no updated data, so remove it from hash list and the stripe
4056		 * will be reinitialized
4057		 */
4058unhash:
4059		hash = sh->hash_lock_index;
4060		spin_lock_irq(conf->hash_locks + hash);
4061		remove_hash(sh);
4062		spin_unlock_irq(conf->hash_locks + hash);
4063		if (head_sh->batch_head) {
4064			sh = list_first_entry(&sh->batch_list,
4065					      struct stripe_head, batch_list);
4066			if (sh != head_sh)
4067					goto unhash;
4068		}
4069		sh = head_sh;
4070
4071		if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4072			set_bit(STRIPE_HANDLE, &sh->state);
4073
4074	}
4075
4076	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4077		if (atomic_dec_and_test(&conf->pending_full_writes))
4078			md_wakeup_thread(conf->mddev->thread);
4079
4080	if (head_sh->batch_head && do_endio)
4081		break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4082}
4083
4084/*
4085 * For RMW in write back cache, we need extra page in prexor to store the
4086 * old data. This page is stored in dev->orig_page.
4087 *
4088 * This function checks whether we have data for prexor. The exact logic
4089 * is:
4090 *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4091 */
4092static inline bool uptodate_for_rmw(struct r5dev *dev)
4093{
4094	return (test_bit(R5_UPTODATE, &dev->flags)) &&
4095		(!test_bit(R5_InJournal, &dev->flags) ||
4096		 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4097}
4098
4099static int handle_stripe_dirtying(struct r5conf *conf,
4100				  struct stripe_head *sh,
4101				  struct stripe_head_state *s,
4102				  int disks)
4103{
4104	int rmw = 0, rcw = 0, i;
4105	struct mddev *mddev = conf->mddev;
4106	sector_t resync_offset = mddev->resync_offset;
4107
4108	/* Check whether resync is now happening or should start.
4109	 * If yes, then the array is dirty (after unclean shutdown or
4110	 * initial creation), so parity in some stripes might be inconsistent.
4111	 * In this case, we need to always do reconstruct-write, to ensure
4112	 * that in case of drive failure or read-error correction, we
4113	 * generate correct data from the parity.
4114	 */
4115	if (conf->rmw_level == PARITY_DISABLE_RMW ||
4116	    (resync_offset < MaxSector && sh->sector >= resync_offset &&
4117	     s->failed == 0)) {
4118		/* Calculate the real rcw later - for now make it
4119		 * look like rcw is cheaper
4120		 */
4121		rcw = 1; rmw = 2;
4122		pr_debug("force RCW rmw_level=%u, resync_offset=%llu sh->sector=%llu\n",
4123			 conf->rmw_level, (unsigned long long)resync_offset,
4124			 (unsigned long long)sh->sector);
4125	} else if (mddev->bitmap_ops && mddev->bitmap_ops->blocks_synced &&
4126		   !mddev->bitmap_ops->blocks_synced(mddev, sh->sector)) {
4127		/* The initial recover is not done, must read everything */
4128		rcw = 1; rmw = 2;
4129		pr_debug("force RCW by lazy recovery, sh->sector=%llu\n",
4130			 sh->sector);
4131	} else for (i = disks; i--; ) {
4132		/* would I have to read this buffer for read_modify_write */
4133		struct r5dev *dev = &sh->dev[i];
4134		if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4135		     i == sh->pd_idx || i == sh->qd_idx ||
4136		     test_bit(R5_InJournal, &dev->flags)) &&
4137		    !test_bit(R5_LOCKED, &dev->flags) &&
4138		    !(uptodate_for_rmw(dev) ||
4139		      test_bit(R5_Wantcompute, &dev->flags))) {
4140			if (test_bit(R5_Insync, &dev->flags))
4141				rmw++;
4142			else
4143				rmw += 2*disks;  /* cannot read it */
4144		}
4145		/* Would I have to read this buffer for reconstruct_write */
4146		if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4147		    i != sh->pd_idx && i != sh->qd_idx &&
4148		    !test_bit(R5_LOCKED, &dev->flags) &&
4149		    !(test_bit(R5_UPTODATE, &dev->flags) ||
4150		      test_bit(R5_Wantcompute, &dev->flags))) {
4151			if (test_bit(R5_Insync, &dev->flags))
4152				rcw++;
4153			else
4154				rcw += 2*disks;
4155		}
4156	}
4157
4158	pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4159		 (unsigned long long)sh->sector, sh->state, rmw, rcw);
4160	set_bit(STRIPE_HANDLE, &sh->state);
4161	if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4162		/* prefer read-modify-write, but need to get some data */
4163		mddev_add_trace_msg(mddev, "raid5 rmw %llu %d",
4164				sh->sector, rmw);
4165
4166		for (i = disks; i--; ) {
4167			struct r5dev *dev = &sh->dev[i];
4168			if (test_bit(R5_InJournal, &dev->flags) &&
4169			    dev->page == dev->orig_page &&
4170			    !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4171				/* alloc page for prexor */
4172				struct page *p = alloc_page(GFP_NOIO);
4173
4174				if (p) {
4175					dev->orig_page = p;
4176					continue;
4177				}
4178
4179				/*
4180				 * alloc_page() failed, try use
4181				 * disk_info->extra_page
4182				 */
4183				if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4184						      &conf->cache_state)) {
4185					r5c_use_extra_page(sh);
4186					break;
4187				}
4188
4189				/* extra_page in use, add to delayed_list */
4190				set_bit(STRIPE_DELAYED, &sh->state);
4191				s->waiting_extra_page = 1;
4192				return -EAGAIN;
4193			}
4194		}
4195
4196		for (i = disks; i--; ) {
4197			struct r5dev *dev = &sh->dev[i];
4198			if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4199			     i == sh->pd_idx || i == sh->qd_idx ||
4200			     test_bit(R5_InJournal, &dev->flags)) &&
4201			    !test_bit(R5_LOCKED, &dev->flags) &&
4202			    !(uptodate_for_rmw(dev) ||
4203			      test_bit(R5_Wantcompute, &dev->flags)) &&
4204			    test_bit(R5_Insync, &dev->flags)) {
4205				if (test_bit(STRIPE_PREREAD_ACTIVE,
4206					     &sh->state)) {
4207					pr_debug("Read_old block %d for r-m-w\n",
4208						 i);
4209					set_bit(R5_LOCKED, &dev->flags);
4210					set_bit(R5_Wantread, &dev->flags);
4211					s->locked++;
4212				} else
4213					set_bit(STRIPE_DELAYED, &sh->state);
4214			}
4215		}
4216	}
4217	if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4218		/* want reconstruct write, but need to get some data */
4219		int qread =0;
4220		rcw = 0;
4221		for (i = disks; i--; ) {
4222			struct r5dev *dev = &sh->dev[i];
4223			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4224			    i != sh->pd_idx && i != sh->qd_idx &&
4225			    !test_bit(R5_LOCKED, &dev->flags) &&
4226			    !(test_bit(R5_UPTODATE, &dev->flags) ||
4227			      test_bit(R5_Wantcompute, &dev->flags))) {
4228				rcw++;
4229				if (test_bit(R5_Insync, &dev->flags) &&
4230				    test_bit(STRIPE_PREREAD_ACTIVE,
4231					     &sh->state)) {
4232					pr_debug("Read_old block "
4233						"%d for Reconstruct\n", i);
4234					set_bit(R5_LOCKED, &dev->flags);
4235					set_bit(R5_Wantread, &dev->flags);
4236					s->locked++;
4237					qread++;
4238				} else
4239					set_bit(STRIPE_DELAYED, &sh->state);
4240			}
4241		}
4242		if (rcw && !mddev_is_dm(mddev))
4243			blk_add_trace_msg(mddev->gendisk->queue,
4244				"raid5 rcw %llu %d %d %d",
4245				(unsigned long long)sh->sector, rcw, qread,
4246				test_bit(STRIPE_DELAYED, &sh->state));
4247	}
4248
4249	if (rcw > disks && rmw > disks &&
4250	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4251		set_bit(STRIPE_DELAYED, &sh->state);
4252
4253	/* now if nothing is locked, and if we have enough data,
4254	 * we can start a write request
4255	 */
4256	/* since handle_stripe can be called at any time we need to handle the
4257	 * case where a compute block operation has been submitted and then a
4258	 * subsequent call wants to start a write request.  raid_run_ops only
4259	 * handles the case where compute block and reconstruct are requested
4260	 * simultaneously.  If this is not the case then new writes need to be
4261	 * held off until the compute completes.
4262	 */
4263	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4264	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4265	     !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4266		schedule_reconstruction(sh, s, rcw == 0, 0);
4267	return 0;
4268}
4269
4270static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4271				struct stripe_head_state *s, int disks)
4272{
4273	struct r5dev *dev = NULL;
4274
4275	BUG_ON(sh->batch_head);
4276	set_bit(STRIPE_HANDLE, &sh->state);
4277
4278	switch (sh->check_state) {
4279	case check_state_idle:
4280		/* start a new check operation if there are no failures */
4281		if (s->failed == 0) {
4282			BUG_ON(s->uptodate != disks);
4283			sh->check_state = check_state_run;
4284			set_bit(STRIPE_OP_CHECK, &s->ops_request);
4285			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4286			s->uptodate--;
4287			break;
4288		}
4289		dev = &sh->dev[s->failed_num[0]];
4290		fallthrough;
4291	case check_state_compute_result:
4292		sh->check_state = check_state_idle;
4293		if (!dev)
4294			dev = &sh->dev[sh->pd_idx];
4295
4296		/* check that a write has not made the stripe insync */
4297		if (test_bit(STRIPE_INSYNC, &sh->state))
4298			break;
4299
4300		/* either failed parity check, or recovery is happening */
4301		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4302		BUG_ON(s->uptodate != disks);
4303
4304		set_bit(R5_LOCKED, &dev->flags);
4305		s->locked++;
4306		set_bit(R5_Wantwrite, &dev->flags);
4307
4308		set_bit(STRIPE_INSYNC, &sh->state);
4309		break;
4310	case check_state_run:
4311		break; /* we will be called again upon completion */
4312	case check_state_check_result:
4313		sh->check_state = check_state_idle;
4314
4315		/* if a failure occurred during the check operation, leave
4316		 * STRIPE_INSYNC not set and let the stripe be handled again
4317		 */
4318		if (s->failed)
4319			break;
4320
4321		/* handle a successful check operation, if parity is correct
4322		 * we are done.  Otherwise update the mismatch count and repair
4323		 * parity if !MD_RECOVERY_CHECK
4324		 */
4325		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4326			/* parity is correct (on disc,
4327			 * not in buffer any more)
4328			 */
4329			set_bit(STRIPE_INSYNC, &sh->state);
4330		else {
4331			atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4332			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4333				/* don't try to repair!! */
4334				set_bit(STRIPE_INSYNC, &sh->state);
4335				pr_warn_ratelimited("%s: mismatch sector in range "
4336						    "%llu-%llu\n", mdname(conf->mddev),
4337						    (unsigned long long) sh->sector,
4338						    (unsigned long long) sh->sector +
4339						    RAID5_STRIPE_SECTORS(conf));
4340			} else {
4341				sh->check_state = check_state_compute_run;
4342				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4343				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4344				set_bit(R5_Wantcompute,
4345					&sh->dev[sh->pd_idx].flags);
4346				sh->ops.target = sh->pd_idx;
4347				sh->ops.target2 = -1;
4348				s->uptodate++;
4349			}
4350		}
4351		break;
4352	case check_state_compute_run:
4353		break;
4354	default:
4355		pr_err("%s: unknown check_state: %d sector: %llu\n",
4356		       __func__, sh->check_state,
4357		       (unsigned long long) sh->sector);
4358		BUG();
4359	}
4360}
4361
4362static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4363				  struct stripe_head_state *s,
4364				  int disks)
4365{
4366	int pd_idx = sh->pd_idx;
4367	int qd_idx = sh->qd_idx;
4368	struct r5dev *dev;
4369
4370	BUG_ON(sh->batch_head);
4371	set_bit(STRIPE_HANDLE, &sh->state);
4372
4373	BUG_ON(s->failed > 2);
4374
4375	/* Want to check and possibly repair P and Q.
4376	 * However there could be one 'failed' device, in which
4377	 * case we can only check one of them, possibly using the
4378	 * other to generate missing data
4379	 */
4380
4381	switch (sh->check_state) {
4382	case check_state_idle:
4383		/* start a new check operation if there are < 2 failures */
4384		if (s->failed == s->q_failed) {
4385			/* The only possible failed device holds Q, so it
4386			 * makes sense to check P (If anything else were failed,
4387			 * we would have used P to recreate it).
4388			 */
4389			sh->check_state = check_state_run;
4390		}
4391		if (!s->q_failed && s->failed < 2) {
4392			/* Q is not failed, and we didn't use it to generate
4393			 * anything, so it makes sense to check it
4394			 */
4395			if (sh->check_state == check_state_run)
4396				sh->check_state = check_state_run_pq;
4397			else
4398				sh->check_state = check_state_run_q;
4399		}
4400
4401		/* discard potentially stale zero_sum_result */
4402		sh->ops.zero_sum_result = 0;
4403
4404		if (sh->check_state == check_state_run) {
4405			/* async_xor_zero_sum destroys the contents of P */
4406			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4407			s->uptodate--;
4408		}
4409		if (sh->check_state >= check_state_run &&
4410		    sh->check_state <= check_state_run_pq) {
4411			/* async_syndrome_zero_sum preserves P and Q, so
4412			 * no need to mark them !uptodate here
4413			 */
4414			set_bit(STRIPE_OP_CHECK, &s->ops_request);
4415			break;
4416		}
4417
4418		/* we have 2-disk failure */
4419		BUG_ON(s->failed != 2);
4420		fallthrough;
4421	case check_state_compute_result:
4422		sh->check_state = check_state_idle;
4423
4424		/* check that a write has not made the stripe insync */
4425		if (test_bit(STRIPE_INSYNC, &sh->state))
4426			break;
4427
4428		/* now write out any block on a failed drive,
4429		 * or P or Q if they were recomputed
4430		 */
4431		dev = NULL;
4432		if (s->failed == 2) {
4433			dev = &sh->dev[s->failed_num[1]];
4434			s->locked++;
4435			set_bit(R5_LOCKED, &dev->flags);
4436			set_bit(R5_Wantwrite, &dev->flags);
4437		}
4438		if (s->failed >= 1) {
4439			dev = &sh->dev[s->failed_num[0]];
4440			s->locked++;
4441			set_bit(R5_LOCKED, &dev->flags);
4442			set_bit(R5_Wantwrite, &dev->flags);
4443		}
4444		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4445			dev = &sh->dev[pd_idx];
4446			s->locked++;
4447			set_bit(R5_LOCKED, &dev->flags);
4448			set_bit(R5_Wantwrite, &dev->flags);
4449		}
4450		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4451			dev = &sh->dev[qd_idx];
4452			s->locked++;
4453			set_bit(R5_LOCKED, &dev->flags);
4454			set_bit(R5_Wantwrite, &dev->flags);
4455		}
4456		if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4457			      "%s: disk%td not up to date\n",
4458			      mdname(conf->mddev),
4459			      dev - (struct r5dev *) &sh->dev)) {
4460			clear_bit(R5_LOCKED, &dev->flags);
4461			clear_bit(R5_Wantwrite, &dev->flags);
4462			s->locked--;
4463		}
4464
4465		set_bit(STRIPE_INSYNC, &sh->state);
4466		break;
4467	case check_state_run:
4468	case check_state_run_q:
4469	case check_state_run_pq:
4470		break; /* we will be called again upon completion */
4471	case check_state_check_result:
4472		sh->check_state = check_state_idle;
4473
4474		/* handle a successful check operation, if parity is correct
4475		 * we are done.  Otherwise update the mismatch count and repair
4476		 * parity if !MD_RECOVERY_CHECK
4477		 */
4478		if (sh->ops.zero_sum_result == 0) {
4479			/* both parities are correct */
4480			if (!s->failed)
4481				set_bit(STRIPE_INSYNC, &sh->state);
4482			else {
4483				/* in contrast to the raid5 case we can validate
4484				 * parity, but still have a failure to write
4485				 * back
4486				 */
4487				sh->check_state = check_state_compute_result;
4488				/* Returning at this point means that we may go
4489				 * off and bring p and/or q uptodate again so
4490				 * we make sure to check zero_sum_result again
4491				 * to verify if p or q need writeback
4492				 */
4493			}
4494		} else {
4495			atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4496			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4497				/* don't try to repair!! */
4498				set_bit(STRIPE_INSYNC, &sh->state);
4499				pr_warn_ratelimited("%s: mismatch sector in range "
4500						    "%llu-%llu\n", mdname(conf->mddev),
4501						    (unsigned long long) sh->sector,
4502						    (unsigned long long) sh->sector +
4503						    RAID5_STRIPE_SECTORS(conf));
4504			} else {
4505				int *target = &sh->ops.target;
4506
4507				sh->ops.target = -1;
4508				sh->ops.target2 = -1;
4509				sh->check_state = check_state_compute_run;
4510				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4511				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4512				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4513					set_bit(R5_Wantcompute,
4514						&sh->dev[pd_idx].flags);
4515					*target = pd_idx;
4516					target = &sh->ops.target2;
4517					s->uptodate++;
4518				}
4519				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4520					set_bit(R5_Wantcompute,
4521						&sh->dev[qd_idx].flags);
4522					*target = qd_idx;
4523					s->uptodate++;
4524				}
4525			}
4526		}
4527		break;
4528	case check_state_compute_run:
4529		break;
4530	default:
4531		pr_warn("%s: unknown check_state: %d sector: %llu\n",
4532			__func__, sh->check_state,
4533			(unsigned long long) sh->sector);
4534		BUG();
4535	}
4536}
4537
4538static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4539{
4540	int i;
4541
4542	/* We have read all the blocks in this stripe and now we need to
4543	 * copy some of them into a target stripe for expand.
4544	 */
4545	struct dma_async_tx_descriptor *tx = NULL;
4546	BUG_ON(sh->batch_head);
4547	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4548	for (i = 0; i < sh->disks; i++)
4549		if (i != sh->pd_idx && i != sh->qd_idx) {
4550			int dd_idx, j;
4551			struct stripe_head *sh2;
4552			struct async_submit_ctl submit;
4553
4554			sector_t bn = raid5_compute_blocknr(sh, i, 1);
4555			sector_t s = raid5_compute_sector(conf, bn, 0,
4556							  &dd_idx, NULL);
4557			sh2 = raid5_get_active_stripe(conf, NULL, s,
4558				R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
4559			if (sh2 == NULL)
4560				/* so far only the early blocks of this stripe
4561				 * have been requested.  When later blocks
4562				 * get requested, we will try again
4563				 */
4564				continue;
4565			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4566			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4567				/* must have already done this block */
4568				raid5_release_stripe(sh2);
4569				continue;
4570			}
4571
4572			/* place all the copies on one channel */
4573			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4574			tx = async_memcpy(sh2->dev[dd_idx].page,
4575					  sh->dev[i].page, sh2->dev[dd_idx].offset,
4576					  sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4577					  &submit);
4578
4579			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4580			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4581			for (j = 0; j < conf->raid_disks; j++)
4582				if (j != sh2->pd_idx &&
4583				    j != sh2->qd_idx &&
4584				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
4585					break;
4586			if (j == conf->raid_disks) {
4587				set_bit(STRIPE_EXPAND_READY, &sh2->state);
4588				set_bit(STRIPE_HANDLE, &sh2->state);
4589			}
4590			raid5_release_stripe(sh2);
4591
4592		}
4593	/* done submitting copies, wait for them to complete */
4594	async_tx_quiesce(&tx);
4595}
4596
4597/*
4598 * handle_stripe - do things to a stripe.
4599 *
4600 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4601 * state of various bits to see what needs to be done.
4602 * Possible results:
4603 *    return some read requests which now have data
4604 *    return some write requests which are safely on storage
4605 *    schedule a read on some buffers
4606 *    schedule a write of some buffers
4607 *    return confirmation of parity correctness
4608 *
4609 */
4610
4611static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4612{
4613	struct r5conf *conf = sh->raid_conf;
4614	int disks = sh->disks;
4615	struct r5dev *dev;
4616	int i;
4617	int do_recovery = 0;
4618
4619	memset(s, 0, sizeof(*s));
4620
4621	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4622	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4623	s->failed_num[0] = -1;
4624	s->failed_num[1] = -1;
4625	s->log_failed = r5l_log_disk_error(conf);
4626
4627	/* Now to look around and see what can be done */
4628	for (i=disks; i--; ) {
4629		struct md_rdev *rdev;
4630		int is_bad = 0;
4631
4632		dev = &sh->dev[i];
4633
4634		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4635			 i, dev->flags,
4636			 dev->toread, dev->towrite, dev->written);
4637		/* maybe we can reply to a read
4638		 *
4639		 * new wantfill requests are only permitted while
4640		 * ops_complete_biofill is guaranteed to be inactive
4641		 */
4642		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4643		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4644			set_bit(R5_Wantfill, &dev->flags);
4645
4646		/* now count some things */
4647		if (test_bit(R5_LOCKED, &dev->flags))
4648			s->locked++;
4649		if (test_bit(R5_UPTODATE, &dev->flags))
4650			s->uptodate++;
4651		if (test_bit(R5_Wantcompute, &dev->flags)) {
4652			s->compute++;
4653			BUG_ON(s->compute > 2);
4654		}
4655
4656		if (test_bit(R5_Wantfill, &dev->flags))
4657			s->to_fill++;
4658		else if (dev->toread)
4659			s->to_read++;
4660		if (dev->towrite) {
4661			s->to_write++;
4662			if (!test_bit(R5_OVERWRITE, &dev->flags))
4663				s->non_overwrite++;
4664		}
4665		if (dev->written)
4666			s->written++;
4667		/* Prefer to use the replacement for reads, but only
4668		 * if it is recovered enough and has no bad blocks.
4669		 */
4670		rdev = conf->disks[i].replacement;
4671		if (rdev && !test_bit(Faulty, &rdev->flags) &&
4672		    rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4673		    !rdev_has_badblock(rdev, sh->sector,
4674				       RAID5_STRIPE_SECTORS(conf)))
4675			set_bit(R5_ReadRepl, &dev->flags);
4676		else {
4677			if (rdev && !test_bit(Faulty, &rdev->flags))
4678				set_bit(R5_NeedReplace, &dev->flags);
4679			else
4680				clear_bit(R5_NeedReplace, &dev->flags);
4681			rdev = conf->disks[i].rdev;
4682			clear_bit(R5_ReadRepl, &dev->flags);
4683		}
4684		if (rdev && test_bit(Faulty, &rdev->flags))
4685			rdev = NULL;
4686		if (rdev) {
4687			is_bad = rdev_has_badblock(rdev, sh->sector,
4688						   RAID5_STRIPE_SECTORS(conf));
4689			if (s->blocked_rdev == NULL) {
4690				if (is_bad < 0)
4691					set_bit(BlockedBadBlocks, &rdev->flags);
4692				if (rdev_blocked(rdev)) {
4693					s->blocked_rdev = rdev;
4694					atomic_inc(&rdev->nr_pending);
4695				}
4696			}
4697		}
4698		clear_bit(R5_Insync, &dev->flags);
4699		if (!rdev)
4700			/* Not in-sync */;
4701		else if (is_bad) {
4702			/* also not in-sync */
4703			if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4704			    test_bit(R5_UPTODATE, &dev->flags)) {
4705				/* treat as in-sync, but with a read error
4706				 * which we can now try to correct
4707				 */
4708				set_bit(R5_Insync, &dev->flags);
4709				set_bit(R5_ReadError, &dev->flags);
4710			}
4711		} else if (test_bit(In_sync, &rdev->flags))
4712			set_bit(R5_Insync, &dev->flags);
4713		else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <=
4714			 rdev->recovery_offset) {
4715			/*
4716			 * in sync if:
4717			 *  - normal IO, or
4718			 *  - resync IO that is not lazy recovery
4719			 *
4720			 * For lazy recovery, we have to mark the rdev without
4721			 * In_sync as failed, to build initial xor data.
4722			 */
4723			if (!test_bit(STRIPE_SYNCING, &sh->state) ||
4724			    !test_bit(MD_RECOVERY_LAZY_RECOVER,
4725				      &conf->mddev->recovery))
4726				set_bit(R5_Insync, &dev->flags);
4727		} else if (test_bit(R5_UPTODATE, &dev->flags) &&
4728			 test_bit(R5_Expanded, &dev->flags))
4729			/* If we've reshaped into here, we assume it is Insync.
4730			 * We will shortly update recovery_offset to make
4731			 * it official.
4732			 */
4733			set_bit(R5_Insync, &dev->flags);
4734
4735		if (test_bit(R5_WriteError, &dev->flags)) {
4736			/* This flag does not apply to '.replacement'
4737			 * only to .rdev, so make sure to check that*/
4738			struct md_rdev *rdev2 = conf->disks[i].rdev;
4739
4740			if (rdev2 == rdev)
4741				clear_bit(R5_Insync, &dev->flags);
4742			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4743				s->handle_bad_blocks = 1;
4744				atomic_inc(&rdev2->nr_pending);
4745			} else
4746				clear_bit(R5_WriteError, &dev->flags);
4747		}
4748		if (test_bit(R5_MadeGood, &dev->flags)) {
4749			/* This flag does not apply to '.replacement'
4750			 * only to .rdev, so make sure to check that*/
4751			struct md_rdev *rdev2 = conf->disks[i].rdev;
4752
4753			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4754				s->handle_bad_blocks = 1;
4755				atomic_inc(&rdev2->nr_pending);
4756			} else
4757				clear_bit(R5_MadeGood, &dev->flags);
4758		}
4759		if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4760			struct md_rdev *rdev2 = conf->disks[i].replacement;
4761
4762			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4763				s->handle_bad_blocks = 1;
4764				atomic_inc(&rdev2->nr_pending);
4765			} else
4766				clear_bit(R5_MadeGoodRepl, &dev->flags);
4767		}
4768		if (!test_bit(R5_Insync, &dev->flags)) {
4769			/* The ReadError flag will just be confusing now */
4770			clear_bit(R5_ReadError, &dev->flags);
4771			clear_bit(R5_ReWrite, &dev->flags);
4772		}
4773		if (test_bit(R5_ReadError, &dev->flags))
4774			clear_bit(R5_Insync, &dev->flags);
4775		if (!test_bit(R5_Insync, &dev->flags)) {
4776			if (s->failed < 2)
4777				s->failed_num[s->failed] = i;
4778			s->failed++;
4779			if (rdev && !test_bit(Faulty, &rdev->flags))
4780				do_recovery = 1;
4781			else if (!rdev) {
4782				rdev = conf->disks[i].replacement;
4783				if (rdev && !test_bit(Faulty, &rdev->flags))
4784					do_recovery = 1;
4785			}
4786		}
4787
4788		if (test_bit(R5_InJournal, &dev->flags))
4789			s->injournal++;
4790		if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4791			s->just_cached++;
4792	}
4793	if (test_bit(STRIPE_SYNCING, &sh->state)) {
4794		/* If there is a failed device being replaced,
4795		 *     we must be recovering.
4796		 * else if we are after resync_offset, we must be syncing
4797		 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4798		 * else we can only be replacing
4799		 * sync and recovery both need to read all devices, and so
4800		 * use the same flag.
4801		 */
4802		if (do_recovery ||
4803		    sh->sector >= conf->mddev->resync_offset ||
4804		    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4805			s->syncing = 1;
4806		else
4807			s->replacing = 1;
4808	}
4809}
4810
4811/*
4812 * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4813 * a head which can now be handled.
4814 */
4815static int clear_batch_ready(struct stripe_head *sh)
4816{
4817	struct stripe_head *tmp;
4818	if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4819		return (sh->batch_head && sh->batch_head != sh);
4820	spin_lock(&sh->stripe_lock);
4821	if (!sh->batch_head) {
4822		spin_unlock(&sh->stripe_lock);
4823		return 0;
4824	}
4825
4826	/*
4827	 * this stripe could be added to a batch list before we check
4828	 * BATCH_READY, skips it
4829	 */
4830	if (sh->batch_head != sh) {
4831		spin_unlock(&sh->stripe_lock);
4832		return 1;
4833	}
4834	spin_lock(&sh->batch_lock);
4835	list_for_each_entry(tmp, &sh->batch_list, batch_list)
4836		clear_bit(STRIPE_BATCH_READY, &tmp->state);
4837	spin_unlock(&sh->batch_lock);
4838	spin_unlock(&sh->stripe_lock);
4839
4840	/*
4841	 * BATCH_READY is cleared, no new stripes can be added.
4842	 * batch_list can be accessed without lock
4843	 */
4844	return 0;
4845}
4846
4847static void break_stripe_batch_list(struct stripe_head *head_sh,
4848				    unsigned long handle_flags)
4849{
4850	struct stripe_head *sh, *next;
4851	int i;
4852
4853	list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4854
4855		list_del_init(&sh->batch_list);
4856
4857		WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4858					  (1 << STRIPE_SYNCING) |
4859					  (1 << STRIPE_REPLACED) |
4860					  (1 << STRIPE_DELAYED) |
4861					  (1 << STRIPE_BIT_DELAY) |
4862					  (1 << STRIPE_FULL_WRITE) |
4863					  (1 << STRIPE_BIOFILL_RUN) |
4864					  (1 << STRIPE_COMPUTE_RUN)  |
4865					  (1 << STRIPE_DISCARD) |
4866					  (1 << STRIPE_BATCH_READY) |
4867					  (1 << STRIPE_BATCH_ERR)),
4868			"stripe state: %lx\n", sh->state);
4869		WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4870					      (1 << STRIPE_REPLACED)),
4871			"head stripe state: %lx\n", head_sh->state);
4872
4873		set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4874					    (1 << STRIPE_PREREAD_ACTIVE) |
4875					    (1 << STRIPE_ON_UNPLUG_LIST)),
4876			      head_sh->state & (1 << STRIPE_INSYNC));
4877
4878		sh->check_state = head_sh->check_state;
4879		sh->reconstruct_state = head_sh->reconstruct_state;
4880		spin_lock_irq(&sh->stripe_lock);
4881		sh->batch_head = NULL;
4882		spin_unlock_irq(&sh->stripe_lock);
4883		for (i = 0; i < sh->disks; i++) {
4884			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4885				wake_up_bit(&sh->dev[i].flags, R5_Overlap);
4886			sh->dev[i].flags = head_sh->dev[i].flags &
4887				(~((1 << R5_WriteError) | (1 << R5_Overlap)));
4888		}
4889		if (handle_flags == 0 ||
4890		    sh->state & handle_flags)
4891			set_bit(STRIPE_HANDLE, &sh->state);
4892		raid5_release_stripe(sh);
4893	}
4894	spin_lock_irq(&head_sh->stripe_lock);
4895	head_sh->batch_head = NULL;
4896	spin_unlock_irq(&head_sh->stripe_lock);
4897	for (i = 0; i < head_sh->disks; i++)
4898		if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4899			wake_up_bit(&head_sh->dev[i].flags, R5_Overlap);
4900	if (head_sh->state & handle_flags)
4901		set_bit(STRIPE_HANDLE, &head_sh->state);
4902}
4903
4904static void handle_stripe(struct stripe_head *sh)
4905{
4906	struct stripe_head_state s;
4907	struct r5conf *conf = sh->raid_conf;
4908	int i;
4909	int prexor;
4910	int disks = sh->disks;
4911	struct r5dev *pdev, *qdev;
4912
4913	clear_bit(STRIPE_HANDLE, &sh->state);
4914
4915	/*
4916	 * handle_stripe should not continue handle the batched stripe, only
4917	 * the head of batch list or lone stripe can continue. Otherwise we
4918	 * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4919	 * is set for the batched stripe.
4920	 */
4921	if (clear_batch_ready(sh))
4922		return;
4923
4924	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4925		/* already being handled, ensure it gets handled
4926		 * again when current action finishes */
4927		set_bit(STRIPE_HANDLE, &sh->state);
4928		return;
4929	}
4930
4931	if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4932		break_stripe_batch_list(sh, 0);
4933
4934	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4935		spin_lock(&sh->stripe_lock);
4936		/*
4937		 * Cannot process 'sync' concurrently with 'discard'.
4938		 * Flush data in r5cache before 'sync'.
4939		 */
4940		if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4941		    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4942		    !test_bit(STRIPE_DISCARD, &sh->state) &&
4943		    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4944			set_bit(STRIPE_SYNCING, &sh->state);
4945			clear_bit(STRIPE_INSYNC, &sh->state);
4946			clear_bit(STRIPE_REPLACED, &sh->state);
4947		}
4948		spin_unlock(&sh->stripe_lock);
4949	}
4950	clear_bit(STRIPE_DELAYED, &sh->state);
4951
4952	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4953		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4954	       (unsigned long long)sh->sector, sh->state,
4955	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4956	       sh->check_state, sh->reconstruct_state);
4957
4958	analyse_stripe(sh, &s);
4959
4960	if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4961		goto finish;
4962
4963	if (s.handle_bad_blocks ||
4964	    (md_is_rdwr(conf->mddev) &&
4965	     test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags))) {
4966		set_bit(STRIPE_HANDLE, &sh->state);
4967		goto finish;
4968	}
4969
4970	if (unlikely(s.blocked_rdev)) {
4971		if (s.syncing || s.expanding || s.expanded ||
4972		    s.replacing || s.to_write || s.written) {
4973			set_bit(STRIPE_HANDLE, &sh->state);
4974			goto finish;
4975		}
4976		/* There is nothing for the blocked_rdev to block */
4977		rdev_dec_pending(s.blocked_rdev, conf->mddev);
4978		s.blocked_rdev = NULL;
4979	}
4980
4981	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4982		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4983		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4984	}
4985
4986	pr_debug("locked=%d uptodate=%d to_read=%d"
4987	       " to_write=%d failed=%d failed_num=%d,%d\n",
4988	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4989	       s.failed_num[0], s.failed_num[1]);
4990	/*
4991	 * check if the array has lost more than max_degraded devices and,
4992	 * if so, some requests might need to be failed.
4993	 *
4994	 * When journal device failed (log_failed), we will only process
4995	 * the stripe if there is data need write to raid disks
4996	 */
4997	if (s.failed > conf->max_degraded ||
4998	    (s.log_failed && s.injournal == 0)) {
4999		sh->check_state = 0;
5000		sh->reconstruct_state = 0;
5001		break_stripe_batch_list(sh, 0);
5002		if (s.to_read+s.to_write+s.written)
5003			handle_failed_stripe(conf, sh, &s, disks);
5004		if (s.syncing + s.replacing)
5005			handle_failed_sync(conf, sh, &s);
5006	}
5007
5008	/* Now we check to see if any write operations have recently
5009	 * completed
5010	 */
5011	prexor = 0;
5012	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
5013		prexor = 1;
5014	if (sh->reconstruct_state == reconstruct_state_drain_result ||
5015	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
5016		sh->reconstruct_state = reconstruct_state_idle;
5017
5018		/* All the 'written' buffers and the parity block are ready to
5019		 * be written back to disk
5020		 */
5021		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
5022		       !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
5023		BUG_ON(sh->qd_idx >= 0 &&
5024		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
5025		       !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
5026		for (i = disks; i--; ) {
5027			struct r5dev *dev = &sh->dev[i];
5028			if (test_bit(R5_LOCKED, &dev->flags) &&
5029				(i == sh->pd_idx || i == sh->qd_idx ||
5030				 dev->written || test_bit(R5_InJournal,
5031							  &dev->flags))) {
5032				pr_debug("Writing block %d\n", i);
5033				set_bit(R5_Wantwrite, &dev->flags);
5034				if (prexor)
5035					continue;
5036				if (s.failed > 1)
5037					continue;
5038				if (!test_bit(R5_Insync, &dev->flags) ||
5039				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
5040				     s.failed == 0))
5041					set_bit(STRIPE_INSYNC, &sh->state);
5042			}
5043		}
5044		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5045			s.dec_preread_active = 1;
5046	}
5047
5048	/*
5049	 * might be able to return some write requests if the parity blocks
5050	 * are safe, or on a failed drive
5051	 */
5052	pdev = &sh->dev[sh->pd_idx];
5053	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5054		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5055	qdev = &sh->dev[sh->qd_idx];
5056	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5057		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5058		|| conf->level < 6;
5059
5060	if (s.written &&
5061	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5062			     && !test_bit(R5_LOCKED, &pdev->flags)
5063			     && (test_bit(R5_UPTODATE, &pdev->flags) ||
5064				 test_bit(R5_Discard, &pdev->flags))))) &&
5065	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5066			     && !test_bit(R5_LOCKED, &qdev->flags)
5067			     && (test_bit(R5_UPTODATE, &qdev->flags) ||
5068				 test_bit(R5_Discard, &qdev->flags))))))
5069		handle_stripe_clean_event(conf, sh, disks);
5070
5071	if (s.just_cached)
5072		r5c_handle_cached_data_endio(conf, sh, disks);
5073	log_stripe_write_finished(sh);
5074
5075	/* Now we might consider reading some blocks, either to check/generate
5076	 * parity, or to satisfy requests
5077	 * or to load a block that is being partially written.
5078	 */
5079	if (s.to_read || s.non_overwrite
5080	    || (s.to_write && s.failed)
5081	    || (s.syncing && (s.uptodate + s.compute < disks))
5082	    || s.replacing
5083	    || s.expanding)
5084		handle_stripe_fill(sh, &s, disks);
5085
5086	/*
5087	 * When the stripe finishes full journal write cycle (write to journal
5088	 * and raid disk), this is the clean up procedure so it is ready for
5089	 * next operation.
5090	 */
5091	r5c_finish_stripe_write_out(conf, sh, &s);
5092
5093	/*
5094	 * Now to consider new write requests, cache write back and what else,
5095	 * if anything should be read.  We do not handle new writes when:
5096	 * 1/ A 'write' operation (copy+xor) is already in flight.
5097	 * 2/ A 'check' operation is in flight, as it may clobber the parity
5098	 *    block.
5099	 * 3/ A r5c cache log write is in flight.
5100	 */
5101
5102	if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5103		if (!r5c_is_writeback(conf->log)) {
5104			if (s.to_write)
5105				handle_stripe_dirtying(conf, sh, &s, disks);
5106		} else { /* write back cache */
5107			int ret = 0;
5108
5109			/* First, try handle writes in caching phase */
5110			if (s.to_write)
5111				ret = r5c_try_caching_write(conf, sh, &s,
5112							    disks);
5113			/*
5114			 * If caching phase failed: ret == -EAGAIN
5115			 *    OR
5116			 * stripe under reclaim: !caching && injournal
5117			 *
5118			 * fall back to handle_stripe_dirtying()
5119			 */
5120			if (ret == -EAGAIN ||
5121			    /* stripe under reclaim: !caching && injournal */
5122			    (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5123			     s.injournal > 0)) {
5124				ret = handle_stripe_dirtying(conf, sh, &s,
5125							     disks);
5126				if (ret == -EAGAIN)
5127					goto finish;
5128			}
5129		}
5130	}
5131
5132	/* maybe we need to check and possibly fix the parity for this stripe
5133	 * Any reads will already have been scheduled, so we just see if enough
5134	 * data is available.  The parity check is held off while parity
5135	 * dependent operations are in flight.
5136	 */
5137	if (sh->check_state ||
5138	    (s.syncing && s.locked == 0 &&
5139	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5140	     !test_bit(STRIPE_INSYNC, &sh->state))) {
5141		if (conf->level == 6)
5142			handle_parity_checks6(conf, sh, &s, disks);
5143		else
5144			handle_parity_checks5(conf, sh, &s, disks);
5145	}
5146
5147	if ((s.replacing || s.syncing) && s.locked == 0
5148	    && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5149	    && !test_bit(STRIPE_REPLACED, &sh->state)) {
5150		/* Write out to replacement devices where possible */
5151		for (i = 0; i < conf->raid_disks; i++)
5152			if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5153				WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5154				set_bit(R5_WantReplace, &sh->dev[i].flags);
5155				set_bit(R5_LOCKED, &sh->dev[i].flags);
5156				s.locked++;
5157			}
5158		if (s.replacing)
5159			set_bit(STRIPE_INSYNC, &sh->state);
5160		set_bit(STRIPE_REPLACED, &sh->state);
5161	}
5162	if ((s.syncing || s.replacing) && s.locked == 0 &&
5163	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5164	    test_bit(STRIPE_INSYNC, &sh->state)) {
5165		md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5166		clear_bit(STRIPE_SYNCING, &sh->state);
5167		if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5168			wake_up_bit(&sh->dev[sh->pd_idx].flags, R5_Overlap);
5169	}
5170
5171	/* If the failed drives are just a ReadError, then we might need
5172	 * to progress the repair/check process
5173	 */
5174	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5175		for (i = 0; i < s.failed; i++) {
5176			struct r5dev *dev = &sh->dev[s.failed_num[i]];
5177			if (test_bit(R5_ReadError, &dev->flags)
5178			    && !test_bit(R5_LOCKED, &dev->flags)
5179			    && test_bit(R5_UPTODATE, &dev->flags)
5180				) {
5181				if (!test_bit(R5_ReWrite, &dev->flags)) {
5182					set_bit(R5_Wantwrite, &dev->flags);
5183					set_bit(R5_ReWrite, &dev->flags);
5184				} else
5185					/* let's read it back */
5186					set_bit(R5_Wantread, &dev->flags);
5187				set_bit(R5_LOCKED, &dev->flags);
5188				s.locked++;
5189			}
5190		}
5191
5192	/* Finish reconstruct operations initiated by the expansion process */
5193	if (sh->reconstruct_state == reconstruct_state_result) {
5194		struct stripe_head *sh_src
5195			= raid5_get_active_stripe(conf, NULL, sh->sector,
5196					R5_GAS_PREVIOUS | R5_GAS_NOBLOCK |
5197					R5_GAS_NOQUIESCE);
5198		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5199			/* sh cannot be written until sh_src has been read.
5200			 * so arrange for sh to be delayed a little
5201			 */
5202			set_bit(STRIPE_DELAYED, &sh->state);
5203			set_bit(STRIPE_HANDLE, &sh->state);
5204			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5205					      &sh_src->state))
5206				atomic_inc(&conf->preread_active_stripes);
5207			raid5_release_stripe(sh_src);
5208			goto finish;
5209		}
5210		if (sh_src)
5211			raid5_release_stripe(sh_src);
5212
5213		sh->reconstruct_state = reconstruct_state_idle;
5214		clear_bit(STRIPE_EXPANDING, &sh->state);
5215		for (i = conf->raid_disks; i--; ) {
5216			set_bit(R5_Wantwrite, &sh->dev[i].flags);
5217			set_bit(R5_LOCKED, &sh->dev[i].flags);
5218			s.locked++;
5219		}
5220	}
5221
5222	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5223	    !sh->reconstruct_state) {
5224		/* Need to write out all blocks after computing parity */
5225		sh->disks = conf->raid_disks;
5226		stripe_set_idx(sh->sector, conf, 0, sh);
5227		schedule_reconstruction(sh, &s, 1, 1);
5228	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5229		clear_bit(STRIPE_EXPAND_READY, &sh->state);
5230		atomic_dec(&conf->reshape_stripes);
5231		wake_up(&conf->wait_for_reshape);
5232		md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5233	}
5234
5235	if (s.expanding && s.locked == 0 &&
5236	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5237		handle_stripe_expansion(conf, sh);
5238
5239finish:
5240	/* wait for this device to become unblocked */
5241	if (unlikely(s.blocked_rdev)) {
5242		if (conf->mddev->external)
5243			md_wait_for_blocked_rdev(s.blocked_rdev,
5244						 conf->mddev);
5245		else
5246			/* Internal metadata will immediately
5247			 * be written by raid5d, so we don't
5248			 * need to wait here.
5249			 */
5250			rdev_dec_pending(s.blocked_rdev,
5251					 conf->mddev);
5252	}
5253
5254	if (s.handle_bad_blocks)
5255		for (i = disks; i--; ) {
5256			struct md_rdev *rdev;
5257			struct r5dev *dev = &sh->dev[i];
5258			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5259				/* We own a safe reference to the rdev */
5260				rdev = conf->disks[i].rdev;
5261				if (!rdev_set_badblocks(rdev, sh->sector,
5262							RAID5_STRIPE_SECTORS(conf), 0))
5263					md_error(conf->mddev, rdev);
5264				rdev_dec_pending(rdev, conf->mddev);
5265			}
5266			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5267				rdev = conf->disks[i].rdev;
5268				rdev_clear_badblocks(rdev, sh->sector,
5269						     RAID5_STRIPE_SECTORS(conf), 0);
5270				rdev_dec_pending(rdev, conf->mddev);
5271			}
5272			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5273				rdev = conf->disks[i].replacement;
5274				if (!rdev)
5275					/* rdev have been moved down */
5276					rdev = conf->disks[i].rdev;
5277				rdev_clear_badblocks(rdev, sh->sector,
5278						     RAID5_STRIPE_SECTORS(conf), 0);
5279				rdev_dec_pending(rdev, conf->mddev);
5280			}
5281		}
5282
5283	if (s.ops_request)
5284		raid_run_ops(sh, s.ops_request);
5285
5286	ops_run_io(sh, &s);
5287
5288	if (s.dec_preread_active) {
5289		/* We delay this until after ops_run_io so that if make_request
5290		 * is waiting on a flush, it won't continue until the writes
5291		 * have actually been submitted.
5292		 */
5293		atomic_dec(&conf->preread_active_stripes);
5294		if (atomic_read(&conf->preread_active_stripes) <
5295		    IO_THRESHOLD)
5296			md_wakeup_thread(conf->mddev->thread);
5297	}
5298
5299	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5300}
5301
5302static void raid5_activate_delayed(struct r5conf *conf)
5303	__must_hold(&conf->device_lock)
5304{
5305	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5306		while (!list_empty(&conf->delayed_list)) {
5307			struct list_head *l = conf->delayed_list.next;
5308			struct stripe_head *sh;
5309			sh = list_entry(l, struct stripe_head, lru);
5310			list_del_init(l);
5311			clear_bit(STRIPE_DELAYED, &sh->state);
5312			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5313				atomic_inc(&conf->preread_active_stripes);
5314			list_add_tail(&sh->lru, &conf->hold_list);
5315			raid5_wakeup_stripe_thread(sh);
5316		}
5317	}
5318}
5319
5320static void activate_bit_delay(struct r5conf *conf,
5321		struct list_head *temp_inactive_list)
5322	__must_hold(&conf->device_lock)
5323{
5324	struct list_head head;
5325	list_add(&head, &conf->bitmap_list);
5326	list_del_init(&conf->bitmap_list);
5327	while (!list_empty(&head)) {
5328		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5329		int hash;
5330		list_del_init(&sh->lru);
5331		atomic_inc(&sh->count);
5332		hash = sh->hash_lock_index;
5333		__release_stripe(conf, sh, &temp_inactive_list[hash]);
5334	}
5335}
5336
5337static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5338{
5339	struct r5conf *conf = mddev->private;
5340	sector_t sector = bio->bi_iter.bi_sector;
5341	unsigned int chunk_sectors;
5342	unsigned int bio_sectors = bio_sectors(bio);
5343
5344	chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5345	return  chunk_sectors >=
5346		((sector & (chunk_sectors - 1)) + bio_sectors);
5347}
5348
5349/*
5350 *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5351 *  later sampled by raid5d.
5352 */
5353static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5354{
5355	unsigned long flags;
5356
5357	spin_lock_irqsave(&conf->device_lock, flags);
5358
5359	bi->bi_next = conf->retry_read_aligned_list;
5360	conf->retry_read_aligned_list = bi;
5361
5362	spin_unlock_irqrestore(&conf->device_lock, flags);
5363	md_wakeup_thread(conf->mddev->thread);
5364}
5365
5366static struct bio *remove_bio_from_retry(struct r5conf *conf,
5367					 unsigned int *offset)
5368{
5369	struct bio *bi;
5370
5371	bi = conf->retry_read_aligned;
5372	if (bi) {
5373		*offset = conf->retry_read_offset;
5374		conf->retry_read_aligned = NULL;
5375		return bi;
5376	}
5377	bi = conf->retry_read_aligned_list;
5378	if(bi) {
5379		conf->retry_read_aligned_list = bi->bi_next;
5380		bi->bi_next = NULL;
5381		*offset = 0;
5382	}
5383
5384	return bi;
5385}
5386
5387/*
5388 *  The "raid5_align_endio" should check if the read succeeded and if it
5389 *  did, call bio_endio on the original bio (having bio_put the new bio
5390 *  first).
5391 *  If the read failed..
5392 */
5393static void raid5_align_endio(struct bio *bi)
5394{
5395	struct bio *raid_bi = bi->bi_private;
5396	struct md_rdev *rdev = (void *)raid_bi->bi_next;
5397	struct mddev *mddev = rdev->mddev;
5398	struct r5conf *conf = mddev->private;
5399	blk_status_t error = bi->bi_status;
5400
5401	bio_put(bi);
5402	raid_bi->bi_next = NULL;
5403	rdev_dec_pending(rdev, conf->mddev);
5404
5405	if (!error) {
5406		bio_endio(raid_bi);
5407		if (atomic_dec_and_test(&conf->active_aligned_reads))
5408			wake_up(&conf->wait_for_quiescent);
5409		return;
5410	}
5411
5412	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5413
5414	add_bio_to_retry(raid_bi, conf);
5415}
5416
5417static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5418{
5419	struct r5conf *conf = mddev->private;
5420	struct bio *align_bio;
5421	struct md_rdev *rdev;
5422	sector_t sector, end_sector;
5423	int dd_idx;
5424	bool did_inc;
5425
5426	if (!in_chunk_boundary(mddev, raid_bio)) {
5427		pr_debug("%s: non aligned\n", __func__);
5428		return 0;
5429	}
5430
5431	sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5432				      &dd_idx, NULL);
5433	end_sector = sector + bio_sectors(raid_bio);
5434
5435	if (r5c_big_stripe_cached(conf, sector))
5436		return 0;
5437
5438	rdev = conf->disks[dd_idx].replacement;
5439	if (!rdev || test_bit(Faulty, &rdev->flags) ||
5440	    rdev->recovery_offset < end_sector) {
5441		rdev = conf->disks[dd_idx].rdev;
5442		if (!rdev)
5443			return 0;
5444		if (test_bit(Faulty, &rdev->flags) ||
5445		    !(test_bit(In_sync, &rdev->flags) ||
5446		      rdev->recovery_offset >= end_sector))
5447			return 0;
5448	}
5449
5450	atomic_inc(&rdev->nr_pending);
5451
5452	if (rdev_has_badblock(rdev, sector, bio_sectors(raid_bio))) {
5453		rdev_dec_pending(rdev, mddev);
5454		return 0;
5455	}
5456
5457	md_account_bio(mddev, &raid_bio);
5458	raid_bio->bi_next = (void *)rdev;
5459
5460	align_bio = bio_alloc_clone(rdev->bdev, raid_bio, GFP_NOIO,
5461				    &mddev->bio_set);
5462	align_bio->bi_end_io = raid5_align_endio;
5463	align_bio->bi_private = raid_bio;
5464	align_bio->bi_iter.bi_sector = sector;
5465
5466	/* No reshape active, so we can trust rdev->data_offset */
5467	align_bio->bi_iter.bi_sector += rdev->data_offset;
5468
5469	did_inc = false;
5470	if (conf->quiesce == 0) {
5471		atomic_inc(&conf->active_aligned_reads);
5472		did_inc = true;
5473	}
5474	/* need a memory barrier to detect the race with raid5_quiesce() */
5475	if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5476		/* quiesce is in progress, so we need to undo io activation and wait
5477		 * for it to finish
5478		 */
5479		if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5480			wake_up(&conf->wait_for_quiescent);
5481		spin_lock_irq(&conf->device_lock);
5482		wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5483				    conf->device_lock);
5484		atomic_inc(&conf->active_aligned_reads);
5485		spin_unlock_irq(&conf->device_lock);
5486	}
5487
5488	mddev_trace_remap(mddev, align_bio, raid_bio->bi_iter.bi_sector);
5489	submit_bio_noacct(align_bio);
5490	return 1;
5491}
5492
5493static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5494{
5495	sector_t sector = raid_bio->bi_iter.bi_sector;
5496	unsigned chunk_sects = mddev->chunk_sectors;
5497	unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5498
5499	if (sectors < bio_sectors(raid_bio)) {
5500		struct r5conf *conf = mddev->private;
5501
5502		raid_bio = bio_submit_split_bioset(raid_bio, sectors,
5503						   &conf->bio_split);
5504		if (!raid_bio)
5505			return NULL;
5506	}
5507
5508	if (!raid5_read_one_chunk(mddev, raid_bio))
5509		return raid_bio;
5510
5511	return NULL;
5512}
5513
5514/* __get_priority_stripe - get the next stripe to process
5515 *
5516 * Full stripe writes are allowed to pass preread active stripes up until
5517 * the bypass_threshold is exceeded.  In general the bypass_count
5518 * increments when the handle_list is handled before the hold_list; however, it
5519 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5520 * stripe with in flight i/o.  The bypass_count will be reset when the
5521 * head of the hold_list has changed, i.e. the head was promoted to the
5522 * handle_list.
5523 */
5524static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5525	__must_hold(&conf->device_lock)
5526{
5527	struct stripe_head *sh, *tmp;
5528	struct list_head *handle_list = NULL;
5529	struct r5worker_group *wg;
5530	bool second_try = !r5c_is_writeback(conf->log) &&
5531		!r5l_log_disk_error(conf);
5532	bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5533		r5l_log_disk_error(conf);
5534
5535again:
5536	wg = NULL;
5537	sh = NULL;
5538	if (conf->worker_cnt_per_group == 0) {
5539		handle_list = try_loprio ? &conf->loprio_list :
5540					&conf->handle_list;
5541	} else if (group != ANY_GROUP) {
5542		handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5543				&conf->worker_groups[group].handle_list;
5544		wg = &conf->worker_groups[group];
5545	} else {
5546		int i;
5547		for (i = 0; i < conf->group_cnt; i++) {
5548			handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5549				&conf->worker_groups[i].handle_list;
5550			wg = &conf->worker_groups[i];
5551			if (!list_empty(handle_list))
5552				break;
5553		}
5554	}
5555
5556	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5557		  __func__,
5558		  list_empty(handle_list) ? "empty" : "busy",
5559		  list_empty(&conf->hold_list) ? "empty" : "busy",
5560		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
5561
5562	if (!list_empty(handle_list)) {
5563		sh = list_entry(handle_list->next, typeof(*sh), lru);
5564
5565		if (list_empty(&conf->hold_list))
5566			conf->bypass_count = 0;
5567		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5568			if (conf->hold_list.next == conf->last_hold)
5569				conf->bypass_count++;
5570			else {
5571				conf->last_hold = conf->hold_list.next;
5572				conf->bypass_count -= conf->bypass_threshold;
5573				if (conf->bypass_count < 0)
5574					conf->bypass_count = 0;
5575			}
5576		}
5577	} else if (!list_empty(&conf->hold_list) &&
5578		   ((conf->bypass_threshold &&
5579		     conf->bypass_count > conf->bypass_threshold) ||
5580		    atomic_read(&conf->pending_full_writes) == 0)) {
5581
5582		list_for_each_entry(tmp, &conf->hold_list,  lru) {
5583			if (conf->worker_cnt_per_group == 0 ||
5584			    group == ANY_GROUP ||
5585			    !cpu_online(tmp->cpu) ||
5586			    cpu_to_group(tmp->cpu) == group) {
5587				sh = tmp;
5588				break;
5589			}
5590		}
5591
5592		if (sh) {
5593			conf->bypass_count -= conf->bypass_threshold;
5594			if (conf->bypass_count < 0)
5595				conf->bypass_count = 0;
5596		}
5597		wg = NULL;
5598	}
5599
5600	if (!sh) {
5601		if (second_try)
5602			return NULL;
5603		second_try = true;
5604		try_loprio = !try_loprio;
5605		goto again;
5606	}
5607
5608	if (wg) {
5609		wg->stripes_cnt--;
5610		sh->group = NULL;
5611	}
5612	list_del_init(&sh->lru);
5613	BUG_ON(atomic_inc_return(&sh->count) != 1);
5614	return sh;
5615}
5616
5617struct raid5_plug_cb {
5618	struct blk_plug_cb	cb;
5619	struct list_head	list;
5620	struct list_head	temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5621};
5622
5623static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5624{
5625	struct raid5_plug_cb *cb = container_of(
5626		blk_cb, struct raid5_plug_cb, cb);
5627	struct stripe_head *sh;
5628	struct mddev *mddev = cb->cb.data;
5629	struct r5conf *conf = mddev->private;
5630	int cnt = 0;
5631	int hash;
5632
5633	if (cb->list.next && !list_empty(&cb->list)) {
5634		spin_lock_irq(&conf->device_lock);
5635		while (!list_empty(&cb->list)) {
5636			sh = list_first_entry(&cb->list, struct stripe_head, lru);
5637			list_del_init(&sh->lru);
5638			/*
5639			 * avoid race release_stripe_plug() sees
5640			 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5641			 * is still in our list
5642			 */
5643			smp_mb__before_atomic();
5644			clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5645			/*
5646			 * STRIPE_ON_RELEASE_LIST could be set here. In that
5647			 * case, the count is always > 1 here
5648			 */
5649			hash = sh->hash_lock_index;
5650			__release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5651			cnt++;
5652		}
5653		spin_unlock_irq(&conf->device_lock);
5654	}
5655	release_inactive_stripe_list(conf, cb->temp_inactive_list,
5656				     NR_STRIPE_HASH_LOCKS);
5657	if (!mddev_is_dm(mddev))
5658		trace_block_unplug(mddev->gendisk->queue, cnt, !from_schedule);
5659	kfree(cb);
5660}
5661
5662static void release_stripe_plug(struct mddev *mddev,
5663				struct stripe_head *sh)
5664{
5665	struct blk_plug_cb *blk_cb = blk_check_plugged(
5666		raid5_unplug, mddev,
5667		sizeof(struct raid5_plug_cb));
5668	struct raid5_plug_cb *cb;
5669
5670	if (!blk_cb) {
5671		raid5_release_stripe(sh);
5672		return;
5673	}
5674
5675	cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5676
5677	if (cb->list.next == NULL) {
5678		int i;
5679		INIT_LIST_HEAD(&cb->list);
5680		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5681			INIT_LIST_HEAD(cb->temp_inactive_list + i);
5682	}
5683
5684	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5685		list_add_tail(&sh->lru, &cb->list);
5686	else
5687		raid5_release_stripe(sh);
5688}
5689
5690static void make_discard_request(struct mddev *mddev, struct bio *bi)
5691{
5692	struct r5conf *conf = mddev->private;
5693	sector_t logical_sector, last_sector;
5694	struct stripe_head *sh;
5695	int stripe_sectors;
5696
5697	/* We need to handle this when io_uring supports discard/trim */
5698	if (WARN_ON_ONCE(bi->bi_opf & REQ_NOWAIT))
5699		return;
5700
5701	if (mddev->reshape_position != MaxSector)
5702		/* Skip discard while reshape is happening */
5703		return;
5704
5705	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5706	last_sector = bio_end_sector(bi);
5707
5708	bi->bi_next = NULL;
5709
5710	stripe_sectors = conf->chunk_sectors *
5711		(conf->raid_disks - conf->max_degraded);
5712	logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5713					       stripe_sectors);
5714	sector_div(last_sector, stripe_sectors);
5715
5716	logical_sector *= conf->chunk_sectors;
5717	last_sector *= conf->chunk_sectors;
5718
5719	for (; logical_sector < last_sector;
5720	     logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5721		DEFINE_WAIT(w);
5722		int d;
5723	again:
5724		sh = raid5_get_active_stripe(conf, NULL, logical_sector, 0);
5725		set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5726		if (test_bit(STRIPE_SYNCING, &sh->state)) {
5727			raid5_release_stripe(sh);
5728			wait_on_bit(&sh->dev[sh->pd_idx].flags, R5_Overlap,
5729				    TASK_UNINTERRUPTIBLE);
5730			goto again;
5731		}
5732		clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5733		spin_lock_irq(&sh->stripe_lock);
5734		for (d = 0; d < conf->raid_disks; d++) {
5735			if (d == sh->pd_idx || d == sh->qd_idx)
5736				continue;
5737			if (sh->dev[d].towrite || sh->dev[d].toread) {
5738				set_bit(R5_Overlap, &sh->dev[d].flags);
5739				spin_unlock_irq(&sh->stripe_lock);
5740				raid5_release_stripe(sh);
5741				wait_on_bit(&sh->dev[d].flags, R5_Overlap,
5742					    TASK_UNINTERRUPTIBLE);
5743				goto again;
5744			}
5745		}
5746		set_bit(STRIPE_DISCARD, &sh->state);
5747		sh->overwrite_disks = 0;
5748		for (d = 0; d < conf->raid_disks; d++) {
5749			if (d == sh->pd_idx || d == sh->qd_idx)
5750				continue;
5751			sh->dev[d].towrite = bi;
5752			set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5753			bio_inc_remaining(bi);
5754			md_write_inc(mddev, bi);
5755			sh->overwrite_disks++;
5756		}
5757		spin_unlock_irq(&sh->stripe_lock);
5758		if (conf->mddev->bitmap) {
5759			sh->bm_seq = conf->seq_flush + 1;
5760			set_bit(STRIPE_BIT_DELAY, &sh->state);
5761		}
5762
5763		set_bit(STRIPE_HANDLE, &sh->state);
5764		clear_bit(STRIPE_DELAYED, &sh->state);
5765		if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5766			atomic_inc(&conf->preread_active_stripes);
5767		release_stripe_plug(mddev, sh);
5768	}
5769
5770	bio_endio(bi);
5771}
5772
5773static bool ahead_of_reshape(struct mddev *mddev, sector_t sector,
5774			     sector_t reshape_sector)
5775{
5776	return mddev->reshape_backwards ? sector < reshape_sector :
5777					  sector >= reshape_sector;
5778}
5779
5780static bool range_ahead_of_reshape(struct mddev *mddev, sector_t min,
5781				   sector_t max, sector_t reshape_sector)
5782{
5783	return mddev->reshape_backwards ? max < reshape_sector :
5784					  min >= reshape_sector;
5785}
5786
5787static bool stripe_ahead_of_reshape(struct mddev *mddev, struct r5conf *conf,
5788				    struct stripe_head *sh)
5789{
5790	sector_t max_sector = 0, min_sector = MaxSector;
5791	bool ret = false;
5792	int dd_idx;
5793
5794	for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5795		if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5796			continue;
5797
5798		min_sector = min(min_sector, sh->dev[dd_idx].sector);
5799		max_sector = max(max_sector, sh->dev[dd_idx].sector);
5800	}
5801
5802	spin_lock_irq(&conf->device_lock);
5803
5804	if (!range_ahead_of_reshape(mddev, min_sector, max_sector,
5805				     conf->reshape_progress))
5806		/* mismatch, need to try again */
5807		ret = true;
5808
5809	spin_unlock_irq(&conf->device_lock);
5810
5811	return ret;
5812}
5813
5814static int add_all_stripe_bios(struct r5conf *conf,
5815		struct stripe_request_ctx *ctx, struct stripe_head *sh,
5816		struct bio *bi, int forwrite, int previous)
5817{
5818	int dd_idx;
5819
5820	spin_lock_irq(&sh->stripe_lock);
5821
5822	for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5823		struct r5dev *dev = &sh->dev[dd_idx];
5824
5825		if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5826			continue;
5827
5828		if (dev->sector < ctx->first_sector ||
5829		    dev->sector >= ctx->last_sector)
5830			continue;
5831
5832		if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
5833			set_bit(R5_Overlap, &dev->flags);
5834			spin_unlock_irq(&sh->stripe_lock);
5835			raid5_release_stripe(sh);
5836			/* release batch_last before wait to avoid risk of deadlock */
5837			if (ctx->batch_last) {
5838				raid5_release_stripe(ctx->batch_last);
5839				ctx->batch_last = NULL;
5840			}
5841			md_wakeup_thread(conf->mddev->thread);
5842			wait_on_bit(&dev->flags, R5_Overlap, TASK_UNINTERRUPTIBLE);
5843			return 0;
5844		}
5845	}
5846
5847	for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5848		struct r5dev *dev = &sh->dev[dd_idx];
5849
5850		if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5851			continue;
5852
5853		if (dev->sector < ctx->first_sector ||
5854		    dev->sector >= ctx->last_sector)
5855			continue;
5856
5857		__add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
5858		clear_bit((dev->sector - ctx->first_sector) >>
5859			  RAID5_STRIPE_SHIFT(conf), ctx->sectors_to_do);
5860	}
5861
5862	spin_unlock_irq(&sh->stripe_lock);
5863	return 1;
5864}
5865
5866enum reshape_loc {
5867	LOC_NO_RESHAPE,
5868	LOC_AHEAD_OF_RESHAPE,
5869	LOC_INSIDE_RESHAPE,
5870	LOC_BEHIND_RESHAPE,
5871};
5872
5873static enum reshape_loc get_reshape_loc(struct mddev *mddev,
5874		struct r5conf *conf, sector_t logical_sector)
5875{
5876	sector_t reshape_progress, reshape_safe;
5877
5878	if (likely(conf->reshape_progress == MaxSector))
5879		return LOC_NO_RESHAPE;
5880	/*
5881	 * Spinlock is needed as reshape_progress may be
5882	 * 64bit on a 32bit platform, and so it might be
5883	 * possible to see a half-updated value
5884	 * Of course reshape_progress could change after
5885	 * the lock is dropped, so once we get a reference
5886	 * to the stripe that we think it is, we will have
5887	 * to check again.
5888	 */
5889	spin_lock_irq(&conf->device_lock);
5890	reshape_progress = conf->reshape_progress;
5891	reshape_safe = conf->reshape_safe;
5892	spin_unlock_irq(&conf->device_lock);
5893	if (reshape_progress == MaxSector)
5894		return LOC_NO_RESHAPE;
5895	if (ahead_of_reshape(mddev, logical_sector, reshape_progress))
5896		return LOC_AHEAD_OF_RESHAPE;
5897	if (ahead_of_reshape(mddev, logical_sector, reshape_safe))
5898		return LOC_INSIDE_RESHAPE;
5899	return LOC_BEHIND_RESHAPE;
5900}
5901
5902static void raid5_bitmap_sector(struct mddev *mddev, sector_t *offset,
5903				unsigned long *sectors)
5904{
5905	struct r5conf *conf = mddev->private;
5906	sector_t start = *offset;
5907	sector_t end = start + *sectors;
5908	sector_t prev_start = start;
5909	sector_t prev_end = end;
5910	int sectors_per_chunk;
5911	enum reshape_loc loc;
5912	int dd_idx;
5913
5914	sectors_per_chunk = conf->chunk_sectors *
5915		(conf->raid_disks - conf->max_degraded);
5916	start = round_down(start, sectors_per_chunk);
5917	end = round_up(end, sectors_per_chunk);
5918
5919	start = raid5_compute_sector(conf, start, 0, &dd_idx, NULL);
5920	end = raid5_compute_sector(conf, end, 0, &dd_idx, NULL);
5921
5922	/*
5923	 * For LOC_INSIDE_RESHAPE, this IO will wait for reshape to make
5924	 * progress, hence it's the same as LOC_BEHIND_RESHAPE.
5925	 */
5926	loc = get_reshape_loc(mddev, conf, prev_start);
5927	if (likely(loc != LOC_AHEAD_OF_RESHAPE)) {
5928		*offset = start;
5929		*sectors = end - start;
5930		return;
5931	}
5932
5933	sectors_per_chunk = conf->prev_chunk_sectors *
5934		(conf->previous_raid_disks - conf->max_degraded);
5935	prev_start = round_down(prev_start, sectors_per_chunk);
5936	prev_end = round_down(prev_end, sectors_per_chunk);
5937
5938	prev_start = raid5_compute_sector(conf, prev_start, 1, &dd_idx, NULL);
5939	prev_end = raid5_compute_sector(conf, prev_end, 1, &dd_idx, NULL);
5940
5941	/*
5942	 * for LOC_AHEAD_OF_RESHAPE, reshape can make progress before this IO
5943	 * is handled in make_stripe_request(), we can't know this here hence
5944	 * we set bits for both.
5945	 */
5946	*offset = min(start, prev_start);
5947	*sectors = max(end, prev_end) - *offset;
5948}
5949
5950static enum stripe_result make_stripe_request(struct mddev *mddev,
5951		struct r5conf *conf, struct stripe_request_ctx *ctx,
5952		sector_t logical_sector, struct bio *bi)
5953{
5954	const int rw = bio_data_dir(bi);
5955	enum stripe_result ret;
5956	struct stripe_head *sh;
5957	enum reshape_loc loc;
5958	sector_t new_sector;
5959	int previous = 0, flags = 0;
5960	int seq, dd_idx;
5961
5962	seq = read_seqcount_begin(&conf->gen_lock);
5963	loc = get_reshape_loc(mddev, conf, logical_sector);
5964	if (loc == LOC_INSIDE_RESHAPE) {
5965		ret = STRIPE_SCHEDULE_AND_RETRY;
5966		goto out;
5967	}
5968	if (loc == LOC_AHEAD_OF_RESHAPE)
5969		previous = 1;
5970
5971	new_sector = raid5_compute_sector(conf, logical_sector, previous,
5972					  &dd_idx, NULL);
5973	pr_debug("raid456: %s, sector %llu logical %llu\n", __func__,
5974		 new_sector, logical_sector);
5975
5976	if (previous)
5977		flags |= R5_GAS_PREVIOUS;
5978	if (bi->bi_opf & REQ_RAHEAD)
5979		flags |= R5_GAS_NOBLOCK;
5980	sh = raid5_get_active_stripe(conf, ctx, new_sector, flags);
5981	if (unlikely(!sh)) {
5982		/* cannot get stripe, just give-up */
5983		bi->bi_status = BLK_STS_IOERR;
5984		return STRIPE_FAIL;
5985	}
5986
5987	if (unlikely(previous) &&
5988	    stripe_ahead_of_reshape(mddev, conf, sh)) {
5989		/*
5990		 * Expansion moved on while waiting for a stripe.
5991		 * Expansion could still move past after this
5992		 * test, but as we are holding a reference to
5993		 * 'sh', we know that if that happens,
5994		 *  STRIPE_EXPANDING will get set and the expansion
5995		 * won't proceed until we finish with the stripe.
5996		 */
5997		ret = STRIPE_SCHEDULE_AND_RETRY;
5998		goto out_release;
5999	}
6000
6001	if (read_seqcount_retry(&conf->gen_lock, seq)) {
6002		/* Might have got the wrong stripe_head by accident */
6003		ret = STRIPE_RETRY;
6004		goto out_release;
6005	}
6006
6007	if (test_bit(STRIPE_EXPANDING, &sh->state)) {
6008		md_wakeup_thread(mddev->thread);
6009		ret = STRIPE_SCHEDULE_AND_RETRY;
6010		goto out_release;
6011	}
6012
6013	if (!add_all_stripe_bios(conf, ctx, sh, bi, rw, previous)) {
6014		ret = STRIPE_RETRY;
6015		goto out;
6016	}
6017
6018	if (stripe_can_batch(sh)) {
6019		stripe_add_to_batch_list(conf, sh, ctx->batch_last);
6020		if (ctx->batch_last)
6021			raid5_release_stripe(ctx->batch_last);
6022		atomic_inc(&sh->count);
6023		ctx->batch_last = sh;
6024	}
6025
6026	if (ctx->do_flush) {
6027		set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
6028		/* we only need flush for one stripe */
6029		ctx->do_flush = false;
6030	}
6031
6032	set_bit(STRIPE_HANDLE, &sh->state);
6033	clear_bit(STRIPE_DELAYED, &sh->state);
6034	if ((!sh->batch_head || sh == sh->batch_head) &&
6035	    (bi->bi_opf & REQ_SYNC) &&
6036	    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
6037		atomic_inc(&conf->preread_active_stripes);
6038
6039	release_stripe_plug(mddev, sh);
6040	return STRIPE_SUCCESS;
6041
6042out_release:
6043	raid5_release_stripe(sh);
6044out:
6045	if (ret == STRIPE_SCHEDULE_AND_RETRY && reshape_interrupted(mddev)) {
6046		bi->bi_status = BLK_STS_RESOURCE;
6047		ret = STRIPE_WAIT_RESHAPE;
6048		pr_err_ratelimited("dm-raid456: io across reshape position while reshape can't make progress");
6049	}
6050	return ret;
6051}
6052
6053/*
6054 * If the bio covers multiple data disks, find sector within the bio that has
6055 * the lowest chunk offset in the first chunk.
6056 */
6057static sector_t raid5_bio_lowest_chunk_sector(struct r5conf *conf,
6058					      struct bio *bi)
6059{
6060	int sectors_per_chunk = conf->chunk_sectors;
6061	int raid_disks = conf->raid_disks;
6062	int dd_idx;
6063	struct stripe_head sh;
6064	unsigned int chunk_offset;
6065	sector_t r_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6066	sector_t sector;
6067
6068	/* We pass in fake stripe_head to get back parity disk numbers */
6069	sector = raid5_compute_sector(conf, r_sector, 0, &dd_idx, &sh);
6070	chunk_offset = sector_div(sector, sectors_per_chunk);
6071	if (sectors_per_chunk - chunk_offset >= bio_sectors(bi))
6072		return r_sector;
6073	/*
6074	 * Bio crosses to the next data disk. Check whether it's in the same
6075	 * chunk.
6076	 */
6077	dd_idx++;
6078	while (dd_idx == sh.pd_idx || dd_idx == sh.qd_idx)
6079		dd_idx++;
6080	if (dd_idx >= raid_disks)
6081		return r_sector;
6082	return r_sector + sectors_per_chunk - chunk_offset;
6083}
6084
6085static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
6086{
6087	DEFINE_WAIT_FUNC(wait, woken_wake_function);
6088	bool on_wq;
6089	struct r5conf *conf = mddev->private;
6090	sector_t logical_sector;
6091	struct stripe_request_ctx ctx = {};
6092	const int rw = bio_data_dir(bi);
6093	enum stripe_result res;
6094	int s, stripe_cnt;
6095
6096	if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
6097		int ret = log_handle_flush_request(conf, bi);
6098
6099		if (ret == 0)
6100			return true;
6101		if (ret == -ENODEV) {
6102			if (md_flush_request(mddev, bi))
6103				return true;
6104		}
6105		/* ret == -EAGAIN, fallback */
6106		/*
6107		 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
6108		 * we need to flush journal device
6109		 */
6110		ctx.do_flush = bi->bi_opf & REQ_PREFLUSH;
6111	}
6112
6113	md_write_start(mddev, bi);
6114	/*
6115	 * If array is degraded, better not do chunk aligned read because
6116	 * later we might have to read it again in order to reconstruct
6117	 * data on failed drives.
6118	 */
6119	if (rw == READ && mddev->degraded == 0 &&
6120	    mddev->reshape_position == MaxSector) {
6121		bi = chunk_aligned_read(mddev, bi);
6122		if (!bi)
6123			return true;
6124	}
6125
6126	if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
6127		make_discard_request(mddev, bi);
6128		md_write_end(mddev);
6129		return true;
6130	}
6131
6132	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6133	ctx.first_sector = logical_sector;
6134	ctx.last_sector = bio_end_sector(bi);
6135	bi->bi_next = NULL;
6136
6137	stripe_cnt = DIV_ROUND_UP_SECTOR_T(ctx.last_sector - logical_sector,
6138					   RAID5_STRIPE_SECTORS(conf));
6139	bitmap_set(ctx.sectors_to_do, 0, stripe_cnt);
6140
6141	pr_debug("raid456: %s, logical %llu to %llu\n", __func__,
6142		 bi->bi_iter.bi_sector, ctx.last_sector);
6143
6144	/* Bail out if conflicts with reshape and REQ_NOWAIT is set */
6145	if ((bi->bi_opf & REQ_NOWAIT) &&
6146	    get_reshape_loc(mddev, conf, logical_sector) == LOC_INSIDE_RESHAPE) {
6147		bio_wouldblock_error(bi);
6148		if (rw == WRITE)
6149			md_write_end(mddev);
6150		return true;
6151	}
6152	md_account_bio(mddev, &bi);
6153
6154	/*
6155	 * Lets start with the stripe with the lowest chunk offset in the first
6156	 * chunk. That has the best chances of creating IOs adjacent to
6157	 * previous IOs in case of sequential IO and thus creates the most
6158	 * sequential IO pattern. We don't bother with the optimization when
6159	 * reshaping as the performance benefit is not worth the complexity.
6160	 */
6161	if (likely(conf->reshape_progress == MaxSector)) {
6162		logical_sector = raid5_bio_lowest_chunk_sector(conf, bi);
6163		on_wq = false;
6164	} else {
6165		add_wait_queue(&conf->wait_for_reshape, &wait);
6166		on_wq = true;
6167	}
6168	s = (logical_sector - ctx.first_sector) >> RAID5_STRIPE_SHIFT(conf);
6169
6170	while (1) {
6171		res = make_stripe_request(mddev, conf, &ctx, logical_sector,
6172					  bi);
6173		if (res == STRIPE_FAIL || res == STRIPE_WAIT_RESHAPE)
6174			break;
6175
6176		if (res == STRIPE_RETRY)
6177			continue;
6178
6179		if (res == STRIPE_SCHEDULE_AND_RETRY) {
6180			WARN_ON_ONCE(!on_wq);
6181			/*
6182			 * Must release the reference to batch_last before
6183			 * scheduling and waiting for work to be done,
6184			 * otherwise the batch_last stripe head could prevent
6185			 * raid5_activate_delayed() from making progress
6186			 * and thus deadlocking.
6187			 */
6188			if (ctx.batch_last) {
6189				raid5_release_stripe(ctx.batch_last);
6190				ctx.batch_last = NULL;
6191			}
6192
6193			wait_woken(&wait, TASK_UNINTERRUPTIBLE,
6194				   MAX_SCHEDULE_TIMEOUT);
6195			continue;
6196		}
6197
6198		s = find_next_bit_wrap(ctx.sectors_to_do, stripe_cnt, s);
6199		if (s == stripe_cnt)
6200			break;
6201
6202		logical_sector = ctx.first_sector +
6203			(s << RAID5_STRIPE_SHIFT(conf));
6204	}
6205	if (unlikely(on_wq))
6206		remove_wait_queue(&conf->wait_for_reshape, &wait);
6207
6208	if (ctx.batch_last)
6209		raid5_release_stripe(ctx.batch_last);
6210
6211	if (rw == WRITE)
6212		md_write_end(mddev);
6213	if (res == STRIPE_WAIT_RESHAPE) {
6214		md_free_cloned_bio(bi);
6215		return false;
6216	}
6217
6218	bio_endio(bi);
6219	return true;
6220}
6221
6222static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
6223
6224static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
6225{
6226	/* reshaping is quite different to recovery/resync so it is
6227	 * handled quite separately ... here.
6228	 *
6229	 * On each call to sync_request, we gather one chunk worth of
6230	 * destination stripes and flag them as expanding.
6231	 * Then we find all the source stripes and request reads.
6232	 * As the reads complete, handle_stripe will copy the data
6233	 * into the destination stripe and release that stripe.
6234	 */
6235	struct r5conf *conf = mddev->private;
6236	struct stripe_head *sh;
6237	struct md_rdev *rdev;
6238	sector_t first_sector, last_sector;
6239	int raid_disks = conf->previous_raid_disks;
6240	int data_disks = raid_disks - conf->max_degraded;
6241	int new_data_disks = conf->raid_disks - conf->max_degraded;
6242	int i;
6243	int dd_idx;
6244	sector_t writepos, readpos, safepos;
6245	sector_t stripe_addr;
6246	int reshape_sectors;
6247	struct list_head stripes;
6248	sector_t retn;
6249
6250	if (sector_nr == 0) {
6251		/* If restarting in the middle, skip the initial sectors */
6252		if (mddev->reshape_backwards &&
6253		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
6254			sector_nr = raid5_size(mddev, 0, 0)
6255				- conf->reshape_progress;
6256		} else if (mddev->reshape_backwards &&
6257			   conf->reshape_progress == MaxSector) {
6258			/* shouldn't happen, but just in case, finish up.*/
6259			sector_nr = MaxSector;
6260		} else if (!mddev->reshape_backwards &&
6261			   conf->reshape_progress > 0)
6262			sector_nr = conf->reshape_progress;
6263		sector_div(sector_nr, new_data_disks);
6264		if (sector_nr) {
6265			mddev->curr_resync_completed = sector_nr;
6266			sysfs_notify_dirent_safe(mddev->sysfs_completed);
6267			*skipped = 1;
6268			retn = sector_nr;
6269			goto finish;
6270		}
6271	}
6272
6273	/* We need to process a full chunk at a time.
6274	 * If old and new chunk sizes differ, we need to process the
6275	 * largest of these
6276	 */
6277
6278	reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
6279
6280	/* We update the metadata at least every 10 seconds, or when
6281	 * the data about to be copied would over-write the source of
6282	 * the data at the front of the range.  i.e. one new_stripe
6283	 * along from reshape_progress new_maps to after where
6284	 * reshape_safe old_maps to
6285	 */
6286	writepos = conf->reshape_progress;
6287	sector_div(writepos, new_data_disks);
6288	readpos = conf->reshape_progress;
6289	sector_div(readpos, data_disks);
6290	safepos = conf->reshape_safe;
6291	sector_div(safepos, data_disks);
6292	if (mddev->reshape_backwards) {
6293		if (WARN_ON(writepos < reshape_sectors))
6294			return MaxSector;
6295
6296		writepos -= reshape_sectors;
6297		readpos += reshape_sectors;
6298		safepos += reshape_sectors;
6299	} else {
6300		writepos += reshape_sectors;
6301		/* readpos and safepos are worst-case calculations.
6302		 * A negative number is overly pessimistic, and causes
6303		 * obvious problems for unsigned storage.  So clip to 0.
6304		 */
6305		readpos -= min_t(sector_t, reshape_sectors, readpos);
6306		safepos -= min_t(sector_t, reshape_sectors, safepos);
6307	}
6308
6309	/* Having calculated the 'writepos' possibly use it
6310	 * to set 'stripe_addr' which is where we will write to.
6311	 */
6312	if (mddev->reshape_backwards) {
6313		if (WARN_ON(conf->reshape_progress == 0))
6314			return MaxSector;
6315
6316		stripe_addr = writepos;
6317		if (WARN_ON((mddev->dev_sectors &
6318		    ~((sector_t)reshape_sectors - 1)) -
6319		    reshape_sectors - stripe_addr != sector_nr))
6320			return MaxSector;
6321	} else {
6322		if (WARN_ON(writepos != sector_nr + reshape_sectors))
6323			return MaxSector;
6324
6325		stripe_addr = sector_nr;
6326	}
6327
6328	/* 'writepos' is the most advanced device address we might write.
6329	 * 'readpos' is the least advanced device address we might read.
6330	 * 'safepos' is the least address recorded in the metadata as having
6331	 *     been reshaped.
6332	 * If there is a min_offset_diff, these are adjusted either by
6333	 * increasing the safepos/readpos if diff is negative, or
6334	 * increasing writepos if diff is positive.
6335	 * If 'readpos' is then behind 'writepos', there is no way that we can
6336	 * ensure safety in the face of a crash - that must be done by userspace
6337	 * making a backup of the data.  So in that case there is no particular
6338	 * rush to update metadata.
6339	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
6340	 * update the metadata to advance 'safepos' to match 'readpos' so that
6341	 * we can be safe in the event of a crash.
6342	 * So we insist on updating metadata if safepos is behind writepos and
6343	 * readpos is beyond writepos.
6344	 * In any case, update the metadata every 10 seconds.
6345	 * Maybe that number should be configurable, but I'm not sure it is
6346	 * worth it.... maybe it could be a multiple of safemode_delay???
6347	 */
6348	if (conf->min_offset_diff < 0) {
6349		safepos += -conf->min_offset_diff;
6350		readpos += -conf->min_offset_diff;
6351	} else
6352		writepos += conf->min_offset_diff;
6353
6354	if ((mddev->reshape_backwards
6355	     ? (safepos > writepos && readpos < writepos)
6356	     : (safepos < writepos && readpos > writepos)) ||
6357	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6358		/* Cannot proceed until we've updated the superblock... */
6359		wait_event(conf->wait_for_reshape,
6360			   atomic_read(&conf->reshape_stripes)==0
6361			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6362		if (atomic_read(&conf->reshape_stripes) != 0)
6363			return 0;
6364		mddev->reshape_position = conf->reshape_progress;
6365		mddev->curr_resync_completed = sector_nr;
6366		if (!mddev->reshape_backwards)
6367			/* Can update recovery_offset */
6368			rdev_for_each(rdev, mddev)
6369				if (rdev->raid_disk >= 0 &&
6370				    !test_bit(Journal, &rdev->flags) &&
6371				    !test_bit(In_sync, &rdev->flags) &&
6372				    rdev->recovery_offset < sector_nr)
6373					rdev->recovery_offset = sector_nr;
6374
6375		conf->reshape_checkpoint = jiffies;
6376		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6377		md_wakeup_thread(mddev->thread);
6378		wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6379			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6380		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6381			return 0;
6382		spin_lock_irq(&conf->device_lock);
6383		conf->reshape_safe = mddev->reshape_position;
6384		spin_unlock_irq(&conf->device_lock);
6385		wake_up(&conf->wait_for_reshape);
6386		sysfs_notify_dirent_safe(mddev->sysfs_completed);
6387	}
6388
6389	INIT_LIST_HEAD(&stripes);
6390	for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6391		int j;
6392		int skipped_disk = 0;
6393		sh = raid5_get_active_stripe(conf, NULL, stripe_addr+i,
6394					     R5_GAS_NOQUIESCE);
6395		set_bit(STRIPE_EXPANDING, &sh->state);
6396		atomic_inc(&conf->reshape_stripes);
6397		/* If any of this stripe is beyond the end of the old
6398		 * array, then we need to zero those blocks
6399		 */
6400		for (j=sh->disks; j--;) {
6401			sector_t s;
6402			if (j == sh->pd_idx)
6403				continue;
6404			if (conf->level == 6 &&
6405			    j == sh->qd_idx)
6406				continue;
6407			s = raid5_compute_blocknr(sh, j, 0);
6408			if (s < raid5_size(mddev, 0, 0)) {
6409				skipped_disk = 1;
6410				continue;
6411			}
6412			memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6413			set_bit(R5_Expanded, &sh->dev[j].flags);
6414			set_bit(R5_UPTODATE, &sh->dev[j].flags);
6415		}
6416		if (!skipped_disk) {
6417			set_bit(STRIPE_EXPAND_READY, &sh->state);
6418			set_bit(STRIPE_HANDLE, &sh->state);
6419		}
6420		list_add(&sh->lru, &stripes);
6421	}
6422	spin_lock_irq(&conf->device_lock);
6423	if (mddev->reshape_backwards)
6424		conf->reshape_progress -= reshape_sectors * new_data_disks;
6425	else
6426		conf->reshape_progress += reshape_sectors * new_data_disks;
6427	spin_unlock_irq(&conf->device_lock);
6428	/* Ok, those stripe are ready. We can start scheduling
6429	 * reads on the source stripes.
6430	 * The source stripes are determined by mapping the first and last
6431	 * block on the destination stripes.
6432	 */
6433	first_sector =
6434		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6435				     1, &dd_idx, NULL);
6436	last_sector =
6437		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6438					    * new_data_disks - 1),
6439				     1, &dd_idx, NULL);
6440	if (last_sector >= mddev->dev_sectors)
6441		last_sector = mddev->dev_sectors - 1;
6442	while (first_sector <= last_sector) {
6443		sh = raid5_get_active_stripe(conf, NULL, first_sector,
6444				R5_GAS_PREVIOUS | R5_GAS_NOQUIESCE);
6445		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6446		set_bit(STRIPE_HANDLE, &sh->state);
6447		raid5_release_stripe(sh);
6448		first_sector += RAID5_STRIPE_SECTORS(conf);
6449	}
6450	/* Now that the sources are clearly marked, we can release
6451	 * the destination stripes
6452	 */
6453	while (!list_empty(&stripes)) {
6454		sh = list_entry(stripes.next, struct stripe_head, lru);
6455		list_del_init(&sh->lru);
6456		raid5_release_stripe(sh);
6457	}
6458	/* If this takes us to the resync_max point where we have to pause,
6459	 * then we need to write out the superblock.
6460	 */
6461	sector_nr += reshape_sectors;
6462	retn = reshape_sectors;
6463finish:
6464	if (mddev->curr_resync_completed > mddev->resync_max ||
6465	    (sector_nr - mddev->curr_resync_completed) * 2
6466	    >= mddev->resync_max - mddev->curr_resync_completed) {
6467		/* Cannot proceed until we've updated the superblock... */
6468		wait_event(conf->wait_for_reshape,
6469			   atomic_read(&conf->reshape_stripes) == 0
6470			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6471		if (atomic_read(&conf->reshape_stripes) != 0)
6472			goto ret;
6473		mddev->reshape_position = conf->reshape_progress;
6474		mddev->curr_resync_completed = sector_nr;
6475		if (!mddev->reshape_backwards)
6476			/* Can update recovery_offset */
6477			rdev_for_each(rdev, mddev)
6478				if (rdev->raid_disk >= 0 &&
6479				    !test_bit(Journal, &rdev->flags) &&
6480				    !test_bit(In_sync, &rdev->flags) &&
6481				    rdev->recovery_offset < sector_nr)
6482					rdev->recovery_offset = sector_nr;
6483		conf->reshape_checkpoint = jiffies;
6484		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6485		md_wakeup_thread(mddev->thread);
6486		wait_event(mddev->sb_wait,
6487			   !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6488			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6489		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6490			goto ret;
6491		spin_lock_irq(&conf->device_lock);
6492		conf->reshape_safe = mddev->reshape_position;
6493		spin_unlock_irq(&conf->device_lock);
6494		wake_up(&conf->wait_for_reshape);
6495		sysfs_notify_dirent_safe(mddev->sysfs_completed);
6496	}
6497ret:
6498	return retn;
6499}
6500
6501static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6502					  sector_t max_sector, int *skipped)
6503{
6504	struct r5conf *conf = mddev->private;
6505	struct stripe_head *sh;
6506	sector_t sync_blocks;
6507	bool still_degraded = false;
6508	int i;
6509
6510	if (sector_nr >= max_sector) {
6511		/* just being told to finish up .. nothing much to do */
6512
6513		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6514			end_reshape(conf);
6515			return 0;
6516		}
6517
6518		if (mddev->curr_resync < max_sector) /* aborted */
6519			md_bitmap_end_sync(mddev, mddev->curr_resync,
6520					   &sync_blocks);
6521		else /* completed sync */
6522			conf->fullsync = 0;
6523		if (md_bitmap_enabled(mddev, false))
6524			mddev->bitmap_ops->close_sync(mddev);
6525
6526		return 0;
6527	}
6528
6529	/* Allow raid5_quiesce to complete */
6530	wait_event(conf->wait_for_reshape, conf->quiesce != 2);
6531
6532	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6533		return reshape_request(mddev, sector_nr, skipped);
6534
6535	/* No need to check resync_max as we never do more than one
6536	 * stripe, and as resync_max will always be on a chunk boundary,
6537	 * if the check in md_do_sync didn't fire, there is no chance
6538	 * of overstepping resync_max here
6539	 */
6540
6541	/* if there is too many failed drives and we are trying
6542	 * to resync, then assert that we are finished, because there is
6543	 * nothing we can do.
6544	 */
6545	if (mddev->degraded >= conf->max_degraded &&
6546	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6547		sector_t rv = mddev->dev_sectors - sector_nr;
6548		*skipped = 1;
6549		return rv;
6550	}
6551	if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6552	    !conf->fullsync &&
6553	    !md_bitmap_start_sync(mddev, sector_nr, &sync_blocks, true) &&
6554	    sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6555		/* we can skip this block, and probably more */
6556		do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6557		*skipped = 1;
6558		/* keep things rounded to whole stripes */
6559		return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6560	}
6561
6562	if (md_bitmap_enabled(mddev, false))
6563		mddev->bitmap_ops->cond_end_sync(mddev, sector_nr, false);
6564
6565	sh = raid5_get_active_stripe(conf, NULL, sector_nr,
6566				     R5_GAS_NOBLOCK);
6567	if (sh == NULL) {
6568		sh = raid5_get_active_stripe(conf, NULL, sector_nr, 0);
6569		/* make sure we don't swamp the stripe cache if someone else
6570		 * is trying to get access
6571		 */
6572		schedule_timeout_uninterruptible(1);
6573	}
6574	/* Need to check if array will still be degraded after recovery/resync
6575	 * Note in case of > 1 drive failures it's possible we're rebuilding
6576	 * one drive while leaving another faulty drive in array.
6577	 */
6578	for (i = 0; i < conf->raid_disks; i++) {
6579		struct md_rdev *rdev = conf->disks[i].rdev;
6580
6581		if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6582			still_degraded = true;
6583	}
6584
6585	md_bitmap_start_sync(mddev, sector_nr, &sync_blocks, still_degraded);
6586	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6587	set_bit(STRIPE_HANDLE, &sh->state);
6588
6589	raid5_release_stripe(sh);
6590
6591	return RAID5_STRIPE_SECTORS(conf);
6592}
6593
6594static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6595			       unsigned int offset)
6596{
6597	/* We may not be able to submit a whole bio at once as there
6598	 * may not be enough stripe_heads available.
6599	 * We cannot pre-allocate enough stripe_heads as we may need
6600	 * more than exist in the cache (if we allow ever large chunks).
6601	 * So we do one stripe head at a time and record in
6602	 * ->bi_hw_segments how many have been done.
6603	 *
6604	 * We *know* that this entire raid_bio is in one chunk, so
6605	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6606	 */
6607	struct stripe_head *sh;
6608	int dd_idx;
6609	sector_t sector, logical_sector, last_sector;
6610	int scnt = 0;
6611	int handled = 0;
6612
6613	logical_sector = raid_bio->bi_iter.bi_sector &
6614		~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6615	sector = raid5_compute_sector(conf, logical_sector,
6616				      0, &dd_idx, NULL);
6617	last_sector = bio_end_sector(raid_bio);
6618
6619	for (; logical_sector < last_sector;
6620	     logical_sector += RAID5_STRIPE_SECTORS(conf),
6621		     sector += RAID5_STRIPE_SECTORS(conf),
6622		     scnt++) {
6623
6624		if (scnt < offset)
6625			/* already done this stripe */
6626			continue;
6627
6628		sh = raid5_get_active_stripe(conf, NULL, sector,
6629				R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
6630		if (!sh) {
6631			/* failed to get a stripe - must wait */
6632			conf->retry_read_aligned = raid_bio;
6633			conf->retry_read_offset = scnt;
6634			return handled;
6635		}
6636
6637		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6638			raid5_release_stripe(sh);
6639			conf->retry_read_aligned = raid_bio;
6640			conf->retry_read_offset = scnt;
6641			return handled;
6642		}
6643
6644		set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6645		handle_stripe(sh);
6646		raid5_release_stripe(sh);
6647		handled++;
6648	}
6649
6650	bio_endio(raid_bio);
6651
6652	if (atomic_dec_and_test(&conf->active_aligned_reads))
6653		wake_up(&conf->wait_for_quiescent);
6654	return handled;
6655}
6656
6657static int handle_active_stripes(struct r5conf *conf, int group,
6658				 struct r5worker *worker,
6659				 struct list_head *temp_inactive_list)
6660		__must_hold(&conf->device_lock)
6661{
6662	struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6663	int i, batch_size = 0, hash;
6664	bool release_inactive = false;
6665
6666	while (batch_size < MAX_STRIPE_BATCH &&
6667			(sh = __get_priority_stripe(conf, group)) != NULL)
6668		batch[batch_size++] = sh;
6669
6670	if (batch_size == 0) {
6671		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6672			if (!list_empty(temp_inactive_list + i))
6673				break;
6674		if (i == NR_STRIPE_HASH_LOCKS) {
6675			spin_unlock_irq(&conf->device_lock);
6676			log_flush_stripe_to_raid(conf);
6677			spin_lock_irq(&conf->device_lock);
6678			return batch_size;
6679		}
6680		release_inactive = true;
6681	}
6682	spin_unlock_irq(&conf->device_lock);
6683
6684	release_inactive_stripe_list(conf, temp_inactive_list,
6685				     NR_STRIPE_HASH_LOCKS);
6686
6687	r5l_flush_stripe_to_raid(conf->log);
6688	if (release_inactive) {
6689		spin_lock_irq(&conf->device_lock);
6690		return 0;
6691	}
6692
6693	for (i = 0; i < batch_size; i++)
6694		handle_stripe(batch[i]);
6695	log_write_stripe_run(conf);
6696
6697	cond_resched();
6698
6699	spin_lock_irq(&conf->device_lock);
6700	for (i = 0; i < batch_size; i++) {
6701		hash = batch[i]->hash_lock_index;
6702		__release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6703	}
6704	return batch_size;
6705}
6706
6707static void raid5_do_work(struct work_struct *work)
6708{
6709	struct r5worker *worker = container_of(work, struct r5worker, work);
6710	struct r5worker_group *group = worker->group;
6711	struct r5conf *conf = group->conf;
6712	struct mddev *mddev = conf->mddev;
6713	int group_id = group - conf->worker_groups;
6714	int handled;
6715	struct blk_plug plug;
6716
6717	pr_debug("+++ raid5worker active\n");
6718
6719	blk_start_plug(&plug);
6720	handled = 0;
6721	spin_lock_irq(&conf->device_lock);
6722	while (1) {
6723		int batch_size, released;
6724
6725		released = release_stripe_list(conf, worker->temp_inactive_list);
6726
6727		batch_size = handle_active_stripes(conf, group_id, worker,
6728						   worker->temp_inactive_list);
6729		worker->working = false;
6730		if (!batch_size && !released)
6731			break;
6732		handled += batch_size;
6733		wait_event_lock_irq(mddev->sb_wait,
6734			!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6735			conf->device_lock);
6736	}
6737	pr_debug("%d stripes handled\n", handled);
6738
6739	spin_unlock_irq(&conf->device_lock);
6740
6741	flush_deferred_bios(conf);
6742
6743	r5l_flush_stripe_to_raid(conf->log);
6744
6745	async_tx_issue_pending_all();
6746	blk_finish_plug(&plug);
6747
6748	pr_debug("--- raid5worker inactive\n");
6749}
6750
6751/*
6752 * This is our raid5 kernel thread.
6753 *
6754 * We scan the hash table for stripes which can be handled now.
6755 * During the scan, completed stripes are saved for us by the interrupt
6756 * handler, so that they will not have to wait for our next wakeup.
6757 */
6758static void raid5d(struct md_thread *thread)
6759{
6760	struct mddev *mddev = thread->mddev;
6761	struct r5conf *conf = mddev->private;
6762	int handled;
6763	struct blk_plug plug;
6764
6765	pr_debug("+++ raid5d active\n");
6766
6767	md_check_recovery(mddev);
6768
6769	blk_start_plug(&plug);
6770	handled = 0;
6771	spin_lock_irq(&conf->device_lock);
6772	while (1) {
6773		struct bio *bio;
6774		int batch_size, released;
6775		unsigned int offset;
6776
6777		if (md_is_rdwr(mddev) &&
6778		    test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
6779			break;
6780
6781		released = release_stripe_list(conf, conf->temp_inactive_list);
6782		if (released)
6783			clear_bit(R5_DID_ALLOC, &conf->cache_state);
6784
6785		if (
6786		    !list_empty(&conf->bitmap_list)) {
6787			/* Now is a good time to flush some bitmap updates */
6788			conf->seq_flush++;
6789			spin_unlock_irq(&conf->device_lock);
6790			if (md_bitmap_enabled(mddev, true))
6791				mddev->bitmap_ops->unplug(mddev, true);
6792			spin_lock_irq(&conf->device_lock);
6793			conf->seq_write = conf->seq_flush;
6794			activate_bit_delay(conf, conf->temp_inactive_list);
6795		}
6796		raid5_activate_delayed(conf);
6797
6798		while ((bio = remove_bio_from_retry(conf, &offset))) {
6799			int ok;
6800			spin_unlock_irq(&conf->device_lock);
6801			ok = retry_aligned_read(conf, bio, offset);
6802			spin_lock_irq(&conf->device_lock);
6803			if (!ok)
6804				break;
6805			handled++;
6806		}
6807
6808		batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6809						   conf->temp_inactive_list);
6810		if (!batch_size && !released)
6811			break;
6812		handled += batch_size;
6813
6814		if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6815			spin_unlock_irq(&conf->device_lock);
6816			md_check_recovery(mddev);
6817			spin_lock_irq(&conf->device_lock);
6818		}
6819	}
6820	pr_debug("%d stripes handled\n", handled);
6821
6822	spin_unlock_irq(&conf->device_lock);
6823	if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6824	    mutex_trylock(&conf->cache_size_mutex)) {
6825		grow_one_stripe(conf, __GFP_NOWARN);
6826		/* Set flag even if allocation failed.  This helps
6827		 * slow down allocation requests when mem is short
6828		 */
6829		set_bit(R5_DID_ALLOC, &conf->cache_state);
6830		mutex_unlock(&conf->cache_size_mutex);
6831	}
6832
6833	flush_deferred_bios(conf);
6834
6835	r5l_flush_stripe_to_raid(conf->log);
6836
6837	async_tx_issue_pending_all();
6838	blk_finish_plug(&plug);
6839
6840	pr_debug("--- raid5d inactive\n");
6841}
6842
6843static ssize_t
6844raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6845{
6846	struct r5conf *conf;
6847	int ret = 0;
6848	spin_lock(&mddev->lock);
6849	conf = mddev->private;
6850	if (conf)
6851		ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6852	spin_unlock(&mddev->lock);
6853	return ret;
6854}
6855
6856int
6857raid5_set_cache_size(struct mddev *mddev, int size)
6858{
6859	int result = 0;
6860	struct r5conf *conf = mddev->private;
6861
6862	if (size <= 16 || size > 32768)
6863		return -EINVAL;
6864
6865	WRITE_ONCE(conf->min_nr_stripes, size);
6866	mutex_lock(&conf->cache_size_mutex);
6867	while (size < conf->max_nr_stripes &&
6868	       drop_one_stripe(conf))
6869		;
6870	mutex_unlock(&conf->cache_size_mutex);
6871
6872	md_allow_write(mddev);
6873
6874	mutex_lock(&conf->cache_size_mutex);
6875	while (size > conf->max_nr_stripes)
6876		if (!grow_one_stripe(conf, GFP_KERNEL)) {
6877			WRITE_ONCE(conf->min_nr_stripes, conf->max_nr_stripes);
6878			result = -ENOMEM;
6879			break;
6880		}
6881	mutex_unlock(&conf->cache_size_mutex);
6882
6883	return result;
6884}
6885EXPORT_SYMBOL(raid5_set_cache_size);
6886
6887static ssize_t
6888raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6889{
6890	struct r5conf *conf;
6891	unsigned long new;
6892	int err;
6893
6894	if (len >= PAGE_SIZE)
6895		return -EINVAL;
6896	if (kstrtoul(page, 10, &new))
6897		return -EINVAL;
6898	err = mddev_lock(mddev);
6899	if (err)
6900		return err;
6901	conf = mddev->private;
6902	if (!conf)
6903		err = -ENODEV;
6904	else
6905		err = raid5_set_cache_size(mddev, new);
6906	mddev_unlock(mddev);
6907
6908	return err ?: len;
6909}
6910
6911static struct md_sysfs_entry
6912raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6913				raid5_show_stripe_cache_size,
6914				raid5_store_stripe_cache_size);
6915
6916static ssize_t
6917raid5_show_rmw_level(struct mddev  *mddev, char *page)
6918{
6919	struct r5conf *conf = mddev->private;
6920	if (conf)
6921		return sprintf(page, "%d\n", conf->rmw_level);
6922	else
6923		return 0;
6924}
6925
6926static ssize_t
6927raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6928{
6929	struct r5conf *conf = mddev->private;
6930	unsigned long new;
6931
6932	if (!conf)
6933		return -ENODEV;
6934
6935	if (len >= PAGE_SIZE)
6936		return -EINVAL;
6937
6938	if (kstrtoul(page, 10, &new))
6939		return -EINVAL;
6940
6941	if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6942		return -EINVAL;
6943
6944	if (new != PARITY_DISABLE_RMW &&
6945	    new != PARITY_ENABLE_RMW &&
6946	    new != PARITY_PREFER_RMW)
6947		return -EINVAL;
6948
6949	conf->rmw_level = new;
6950	return len;
6951}
6952
6953static struct md_sysfs_entry
6954raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6955			 raid5_show_rmw_level,
6956			 raid5_store_rmw_level);
6957
6958static ssize_t
6959raid5_show_stripe_size(struct mddev  *mddev, char *page)
6960{
6961	struct r5conf *conf;
6962	int ret = 0;
6963
6964	spin_lock(&mddev->lock);
6965	conf = mddev->private;
6966	if (conf)
6967		ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6968	spin_unlock(&mddev->lock);
6969	return ret;
6970}
6971
6972#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6973static ssize_t
6974raid5_store_stripe_size(struct mddev  *mddev, const char *page, size_t len)
6975{
6976	struct r5conf *conf;
6977	unsigned long new;
6978	int err;
6979	int size;
6980
6981	if (len >= PAGE_SIZE)
6982		return -EINVAL;
6983	if (kstrtoul(page, 10, &new))
6984		return -EINVAL;
6985
6986	/*
6987	 * The value should not be bigger than PAGE_SIZE. It requires to
6988	 * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6989	 * of two.
6990	 */
6991	if (new % DEFAULT_STRIPE_SIZE != 0 ||
6992			new > PAGE_SIZE || new == 0 ||
6993			new != roundup_pow_of_two(new))
6994		return -EINVAL;
6995
6996	err = mddev_suspend_and_lock(mddev);
6997	if (err)
6998		return err;
6999
7000	conf = mddev->private;
7001	if (!conf) {
7002		err = -ENODEV;
7003		goto out_unlock;
7004	}
7005
7006	if (new == conf->stripe_size)
7007		goto out_unlock;
7008
7009	pr_debug("md/raid: change stripe_size from %lu to %lu\n",
7010			conf->stripe_size, new);
7011
7012	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
7013	    mddev->reshape_position != MaxSector || mddev->sysfs_active) {
7014		err = -EBUSY;
7015		goto out_unlock;
7016	}
7017
7018	mutex_lock(&conf->cache_size_mutex);
7019	size = conf->max_nr_stripes;
7020
7021	shrink_stripes(conf);
7022
7023	conf->stripe_size = new;
7024	conf->stripe_shift = ilog2(new) - 9;
7025	conf->stripe_sectors = new >> 9;
7026	if (grow_stripes(conf, size)) {
7027		pr_warn("md/raid:%s: couldn't allocate buffers\n",
7028				mdname(mddev));
7029		err = -ENOMEM;
7030	}
7031	mutex_unlock(&conf->cache_size_mutex);
7032
7033out_unlock:
7034	mddev_unlock_and_resume(mddev);
7035	return err ?: len;
7036}
7037
7038static struct md_sysfs_entry
7039raid5_stripe_size = __ATTR(stripe_size, 0644,
7040			 raid5_show_stripe_size,
7041			 raid5_store_stripe_size);
7042#else
7043static struct md_sysfs_entry
7044raid5_stripe_size = __ATTR(stripe_size, 0444,
7045			 raid5_show_stripe_size,
7046			 NULL);
7047#endif
7048
7049static ssize_t
7050raid5_show_preread_threshold(struct mddev *mddev, char *page)
7051{
7052	struct r5conf *conf;
7053	int ret = 0;
7054	spin_lock(&mddev->lock);
7055	conf = mddev->private;
7056	if (conf)
7057		ret = sprintf(page, "%d\n", conf->bypass_threshold);
7058	spin_unlock(&mddev->lock);
7059	return ret;
7060}
7061
7062static ssize_t
7063raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
7064{
7065	struct r5conf *conf;
7066	unsigned long new;
7067	int err;
7068
7069	if (len >= PAGE_SIZE)
7070		return -EINVAL;
7071	if (kstrtoul(page, 10, &new))
7072		return -EINVAL;
7073
7074	err = mddev_lock(mddev);
7075	if (err)
7076		return err;
7077	conf = mddev->private;
7078	if (!conf)
7079		err = -ENODEV;
7080	else if (new > conf->min_nr_stripes)
7081		err = -EINVAL;
7082	else
7083		conf->bypass_threshold = new;
7084	mddev_unlock(mddev);
7085	return err ?: len;
7086}
7087
7088static struct md_sysfs_entry
7089raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
7090					S_IRUGO | S_IWUSR,
7091					raid5_show_preread_threshold,
7092					raid5_store_preread_threshold);
7093
7094static ssize_t
7095raid5_show_skip_copy(struct mddev *mddev, char *page)
7096{
7097	struct r5conf *conf;
7098	int ret = 0;
7099	spin_lock(&mddev->lock);
7100	conf = mddev->private;
7101	if (conf)
7102		ret = sprintf(page, "%d\n", conf->skip_copy);
7103	spin_unlock(&mddev->lock);
7104	return ret;
7105}
7106
7107static ssize_t
7108raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
7109{
7110	struct r5conf *conf;
7111	unsigned long new;
7112	int err;
7113
7114	if (len >= PAGE_SIZE)
7115		return -EINVAL;
7116	if (kstrtoul(page, 10, &new))
7117		return -EINVAL;
7118	new = !!new;
7119
7120	err = mddev_suspend_and_lock(mddev);
7121	if (err)
7122		return err;
7123	conf = mddev->private;
7124	if (!conf)
7125		err = -ENODEV;
7126	else if (new != conf->skip_copy) {
7127		struct request_queue *q = mddev->gendisk->queue;
7128		struct queue_limits lim = queue_limits_start_update(q);
7129
7130		conf->skip_copy = new;
7131		if (new)
7132			lim.features |= BLK_FEAT_STABLE_WRITES;
7133		else
7134			lim.features &= ~BLK_FEAT_STABLE_WRITES;
7135		err = queue_limits_commit_update(q, &lim);
7136	}
7137	mddev_unlock_and_resume(mddev);
7138	return err ?: len;
7139}
7140
7141static struct md_sysfs_entry
7142raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
7143					raid5_show_skip_copy,
7144					raid5_store_skip_copy);
7145
7146static ssize_t
7147stripe_cache_active_show(struct mddev *mddev, char *page)
7148{
7149	struct r5conf *conf = mddev->private;
7150	if (conf)
7151		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
7152	else
7153		return 0;
7154}
7155
7156static struct md_sysfs_entry
7157raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
7158
7159static ssize_t
7160raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
7161{
7162	struct r5conf *conf;
7163	int ret = 0;
7164	spin_lock(&mddev->lock);
7165	conf = mddev->private;
7166	if (conf)
7167		ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
7168	spin_unlock(&mddev->lock);
7169	return ret;
7170}
7171
7172static int alloc_thread_groups(struct r5conf *conf, int cnt,
7173			       int *group_cnt,
7174			       struct r5worker_group **worker_groups);
7175static ssize_t
7176raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
7177{
7178	struct r5conf *conf;
7179	unsigned int new;
7180	int err;
7181	struct r5worker_group *new_groups, *old_groups;
7182	int group_cnt;
7183
7184	if (len >= PAGE_SIZE)
7185		return -EINVAL;
7186	if (kstrtouint(page, 10, &new))
7187		return -EINVAL;
7188	/* 8192 should be big enough */
7189	if (new > 8192)
7190		return -EINVAL;
7191
7192	err = mddev_suspend_and_lock(mddev);
7193	if (err)
7194		return err;
7195	conf = mddev->private;
7196	if (!conf) {
7197		mddev_unlock_and_resume(mddev);
7198		return -ENODEV;
7199	}
7200	raid5_quiesce(mddev, true);
7201
7202	if (new != conf->worker_cnt_per_group) {
7203		old_groups = conf->worker_groups;
7204		if (old_groups)
7205			flush_workqueue(raid5_wq);
7206
7207		err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
7208		if (!err) {
7209			spin_lock_irq(&conf->device_lock);
7210			conf->group_cnt = group_cnt;
7211			conf->worker_cnt_per_group = new;
7212			conf->worker_groups = new_groups;
7213			spin_unlock_irq(&conf->device_lock);
7214
7215			if (old_groups)
7216				kfree(old_groups[0].workers);
7217			kfree(old_groups);
7218		}
7219	}
7220
7221	raid5_quiesce(mddev, false);
7222	mddev_unlock_and_resume(mddev);
7223
7224	return err ?: len;
7225}
7226
7227static struct md_sysfs_entry
7228raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
7229				raid5_show_group_thread_cnt,
7230				raid5_store_group_thread_cnt);
7231
7232static struct attribute *raid5_attrs[] =  {
7233	&raid5_stripecache_size.attr,
7234	&raid5_stripecache_active.attr,
7235	&raid5_preread_bypass_threshold.attr,
7236	&raid5_group_thread_cnt.attr,
7237	&raid5_skip_copy.attr,
7238	&raid5_rmw_level.attr,
7239	&raid5_stripe_size.attr,
7240	&r5c_journal_mode.attr,
7241	&ppl_write_hint.attr,
7242	NULL,
7243};
7244static const struct attribute_group raid5_attrs_group = {
7245	.name = NULL,
7246	.attrs = raid5_attrs,
7247};
7248
7249static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
7250			       struct r5worker_group **worker_groups)
7251{
7252	int i, j, k;
7253	ssize_t size;
7254	struct r5worker *workers;
7255
7256	if (cnt == 0) {
7257		*group_cnt = 0;
7258		*worker_groups = NULL;
7259		return 0;
7260	}
7261	*group_cnt = num_possible_nodes();
7262	size = sizeof(struct r5worker) * cnt;
7263	workers = kcalloc(size, *group_cnt, GFP_NOIO);
7264	*worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
7265				 GFP_NOIO);
7266	if (!*worker_groups || !workers) {
7267		kfree(workers);
7268		kfree(*worker_groups);
7269		return -ENOMEM;
7270	}
7271
7272	for (i = 0; i < *group_cnt; i++) {
7273		struct r5worker_group *group;
7274
7275		group = &(*worker_groups)[i];
7276		INIT_LIST_HEAD(&group->handle_list);
7277		INIT_LIST_HEAD(&group->loprio_list);
7278		group->conf = conf;
7279		group->workers = workers + i * cnt;
7280
7281		for (j = 0; j < cnt; j++) {
7282			struct r5worker *worker = group->workers + j;
7283			worker->group = group;
7284			INIT_WORK(&worker->work, raid5_do_work);
7285
7286			for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
7287				INIT_LIST_HEAD(worker->temp_inactive_list + k);
7288		}
7289	}
7290
7291	return 0;
7292}
7293
7294static void free_thread_groups(struct r5conf *conf)
7295{
7296	if (conf->worker_groups)
7297		kfree(conf->worker_groups[0].workers);
7298	kfree(conf->worker_groups);
7299	conf->worker_groups = NULL;
7300}
7301
7302static sector_t
7303raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7304{
7305	struct r5conf *conf = mddev->private;
7306
7307	if (!sectors)
7308		sectors = mddev->dev_sectors;
7309	if (!raid_disks)
7310		/* size is defined by the smallest of previous and new size */
7311		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7312
7313	sectors &= ~((sector_t)conf->chunk_sectors - 1);
7314	sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7315	return sectors * (raid_disks - conf->max_degraded);
7316}
7317
7318static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7319{
7320	safe_put_page(percpu->spare_page);
7321	percpu->spare_page = NULL;
7322	kvfree(percpu->scribble);
7323	percpu->scribble = NULL;
7324}
7325
7326static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7327{
7328	if (conf->level == 6 && !percpu->spare_page) {
7329		percpu->spare_page = alloc_page(GFP_KERNEL);
7330		if (!percpu->spare_page)
7331			return -ENOMEM;
7332	}
7333
7334	if (scribble_alloc(percpu,
7335			   max(conf->raid_disks,
7336			       conf->previous_raid_disks),
7337			   max(conf->chunk_sectors,
7338			       conf->prev_chunk_sectors)
7339			   / RAID5_STRIPE_SECTORS(conf))) {
7340		free_scratch_buffer(conf, percpu);
7341		return -ENOMEM;
7342	}
7343
7344	local_lock_init(&percpu->lock);
7345	return 0;
7346}
7347
7348static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7349{
7350	struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7351
7352	free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7353	return 0;
7354}
7355
7356static void raid5_free_percpu(struct r5conf *conf)
7357{
7358	if (!conf->percpu)
7359		return;
7360
7361	cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7362	free_percpu(conf->percpu);
7363}
7364
7365static void free_conf(struct r5conf *conf)
7366{
7367	int i;
7368
7369	log_exit(conf);
7370
7371	shrinker_free(conf->shrinker);
7372	free_thread_groups(conf);
7373	shrink_stripes(conf);
7374	raid5_free_percpu(conf);
7375	for (i = 0; i < conf->pool_size; i++)
7376		if (conf->disks[i].extra_page)
7377			put_page(conf->disks[i].extra_page);
7378	kfree(conf->disks);
7379	bioset_exit(&conf->bio_split);
7380	kfree(conf->stripe_hashtbl);
7381	kfree(conf->pending_data);
7382	kfree(conf);
7383}
7384
7385static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7386{
7387	struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7388	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7389
7390	if (alloc_scratch_buffer(conf, percpu)) {
7391		pr_warn("%s: failed memory allocation for cpu%u\n",
7392			__func__, cpu);
7393		return -ENOMEM;
7394	}
7395	return 0;
7396}
7397
7398static int raid5_alloc_percpu(struct r5conf *conf)
7399{
7400	int err = 0;
7401
7402	conf->percpu = alloc_percpu(struct raid5_percpu);
7403	if (!conf->percpu)
7404		return -ENOMEM;
7405
7406	err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7407	if (!err) {
7408		conf->scribble_disks = max(conf->raid_disks,
7409			conf->previous_raid_disks);
7410		conf->scribble_sectors = max(conf->chunk_sectors,
7411			conf->prev_chunk_sectors);
7412	}
7413	return err;
7414}
7415
7416static unsigned long raid5_cache_scan(struct shrinker *shrink,
7417				      struct shrink_control *sc)
7418{
7419	struct r5conf *conf = shrink->private_data;
7420	unsigned long ret = SHRINK_STOP;
7421
7422	if (mutex_trylock(&conf->cache_size_mutex)) {
7423		ret= 0;
7424		while (ret < sc->nr_to_scan &&
7425		       conf->max_nr_stripes > conf->min_nr_stripes) {
7426			if (drop_one_stripe(conf) == 0) {
7427				ret = SHRINK_STOP;
7428				break;
7429			}
7430			ret++;
7431		}
7432		mutex_unlock(&conf->cache_size_mutex);
7433	}
7434	return ret;
7435}
7436
7437static unsigned long raid5_cache_count(struct shrinker *shrink,
7438				       struct shrink_control *sc)
7439{
7440	struct r5conf *conf = shrink->private_data;
7441	int max_stripes = READ_ONCE(conf->max_nr_stripes);
7442	int min_stripes = READ_ONCE(conf->min_nr_stripes);
7443
7444	if (max_stripes < min_stripes)
7445		/* unlikely, but not impossible */
7446		return 0;
7447	return max_stripes - min_stripes;
7448}
7449
7450static struct r5conf *setup_conf(struct mddev *mddev)
7451{
7452	struct r5conf *conf;
7453	int raid_disk, memory, max_disks;
7454	struct md_rdev *rdev;
7455	struct disk_info *disk;
7456	char pers_name[6];
7457	int i;
7458	int group_cnt;
7459	struct r5worker_group *new_group;
7460	int ret = -ENOMEM;
7461
7462	if (mddev->new_level != 5
7463	    && mddev->new_level != 4
7464	    && mddev->new_level != 6) {
7465		pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7466			mdname(mddev), mddev->new_level);
7467		return ERR_PTR(-EIO);
7468	}
7469	if ((mddev->new_level == 5
7470	     && !algorithm_valid_raid5(mddev->new_layout)) ||
7471	    (mddev->new_level == 6
7472	     && !algorithm_valid_raid6(mddev->new_layout))) {
7473		pr_warn("md/raid:%s: layout %d not supported\n",
7474			mdname(mddev), mddev->new_layout);
7475		return ERR_PTR(-EIO);
7476	}
7477	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7478		pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7479			mdname(mddev), mddev->raid_disks);
7480		return ERR_PTR(-EINVAL);
7481	}
7482
7483	if (!mddev->new_chunk_sectors ||
7484	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7485	    !is_power_of_2(mddev->new_chunk_sectors)) {
7486		pr_warn("md/raid:%s: invalid chunk size %d\n",
7487			mdname(mddev), mddev->new_chunk_sectors << 9);
7488		return ERR_PTR(-EINVAL);
7489	}
7490
7491	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7492	if (conf == NULL)
7493		goto abort;
7494
7495#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7496	conf->stripe_size = DEFAULT_STRIPE_SIZE;
7497	conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7498	conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7499#endif
7500	INIT_LIST_HEAD(&conf->free_list);
7501	INIT_LIST_HEAD(&conf->pending_list);
7502	conf->pending_data = kcalloc(PENDING_IO_MAX,
7503				     sizeof(struct r5pending_data),
7504				     GFP_KERNEL);
7505	if (!conf->pending_data)
7506		goto abort;
7507	for (i = 0; i < PENDING_IO_MAX; i++)
7508		list_add(&conf->pending_data[i].sibling, &conf->free_list);
7509	/* Don't enable multi-threading by default*/
7510	if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7511		conf->group_cnt = group_cnt;
7512		conf->worker_cnt_per_group = 0;
7513		conf->worker_groups = new_group;
7514	} else
7515		goto abort;
7516	spin_lock_init(&conf->device_lock);
7517	seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7518	mutex_init(&conf->cache_size_mutex);
7519
7520	init_waitqueue_head(&conf->wait_for_quiescent);
7521	init_waitqueue_head(&conf->wait_for_stripe);
7522	init_waitqueue_head(&conf->wait_for_reshape);
7523	INIT_LIST_HEAD(&conf->handle_list);
7524	INIT_LIST_HEAD(&conf->loprio_list);
7525	INIT_LIST_HEAD(&conf->hold_list);
7526	INIT_LIST_HEAD(&conf->delayed_list);
7527	INIT_LIST_HEAD(&conf->bitmap_list);
7528	init_llist_head(&conf->released_stripes);
7529	atomic_set(&conf->active_stripes, 0);
7530	atomic_set(&conf->preread_active_stripes, 0);
7531	atomic_set(&conf->active_aligned_reads, 0);
7532	spin_lock_init(&conf->pending_bios_lock);
7533	conf->batch_bio_dispatch = true;
7534	rdev_for_each(rdev, mddev) {
7535		if (test_bit(Journal, &rdev->flags))
7536			continue;
7537		if (bdev_nonrot(rdev->bdev)) {
7538			conf->batch_bio_dispatch = false;
7539			break;
7540		}
7541	}
7542
7543	conf->bypass_threshold = BYPASS_THRESHOLD;
7544	conf->recovery_disabled = mddev->recovery_disabled - 1;
7545
7546	conf->raid_disks = mddev->raid_disks;
7547	if (mddev->reshape_position == MaxSector)
7548		conf->previous_raid_disks = mddev->raid_disks;
7549	else
7550		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7551	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7552
7553	conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7554			      GFP_KERNEL);
7555
7556	if (!conf->disks)
7557		goto abort;
7558
7559	for (i = 0; i < max_disks; i++) {
7560		conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7561		if (!conf->disks[i].extra_page)
7562			goto abort;
7563	}
7564
7565	ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7566	if (ret)
7567		goto abort;
7568	conf->mddev = mddev;
7569
7570	ret = -ENOMEM;
7571	conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL);
7572	if (!conf->stripe_hashtbl)
7573		goto abort;
7574
7575	/* We init hash_locks[0] separately to that it can be used
7576	 * as the reference lock in the spin_lock_nest_lock() call
7577	 * in lock_all_device_hash_locks_irq in order to convince
7578	 * lockdep that we know what we are doing.
7579	 */
7580	spin_lock_init(conf->hash_locks);
7581	for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7582		spin_lock_init(conf->hash_locks + i);
7583
7584	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7585		INIT_LIST_HEAD(conf->inactive_list + i);
7586
7587	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7588		INIT_LIST_HEAD(conf->temp_inactive_list + i);
7589
7590	atomic_set(&conf->r5c_cached_full_stripes, 0);
7591	INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7592	atomic_set(&conf->r5c_cached_partial_stripes, 0);
7593	INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7594	atomic_set(&conf->r5c_flushing_full_stripes, 0);
7595	atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7596
7597	conf->level = mddev->new_level;
7598	conf->chunk_sectors = mddev->new_chunk_sectors;
7599	ret = raid5_alloc_percpu(conf);
7600	if (ret)
7601		goto abort;
7602
7603	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7604
7605	ret = -EIO;
7606	rdev_for_each(rdev, mddev) {
7607		raid_disk = rdev->raid_disk;
7608		if (raid_disk >= max_disks
7609		    || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7610			continue;
7611		disk = conf->disks + raid_disk;
7612
7613		if (test_bit(Replacement, &rdev->flags)) {
7614			if (disk->replacement)
7615				goto abort;
7616			disk->replacement = rdev;
7617		} else {
7618			if (disk->rdev)
7619				goto abort;
7620			disk->rdev = rdev;
7621		}
7622
7623		if (test_bit(In_sync, &rdev->flags)) {
7624			pr_info("md/raid:%s: device %pg operational as raid disk %d\n",
7625				mdname(mddev), rdev->bdev, raid_disk);
7626		} else if (rdev->saved_raid_disk != raid_disk)
7627			/* Cannot rely on bitmap to complete recovery */
7628			conf->fullsync = 1;
7629	}
7630
7631	conf->level = mddev->new_level;
7632	if (conf->level == 6) {
7633		conf->max_degraded = 2;
7634		if (raid6_call.xor_syndrome)
7635			conf->rmw_level = PARITY_ENABLE_RMW;
7636		else
7637			conf->rmw_level = PARITY_DISABLE_RMW;
7638	} else {
7639		conf->max_degraded = 1;
7640		conf->rmw_level = PARITY_ENABLE_RMW;
7641	}
7642	conf->algorithm = mddev->new_layout;
7643	conf->reshape_progress = mddev->reshape_position;
7644	if (conf->reshape_progress != MaxSector) {
7645		conf->prev_chunk_sectors = mddev->chunk_sectors;
7646		conf->prev_algo = mddev->layout;
7647	} else {
7648		conf->prev_chunk_sectors = conf->chunk_sectors;
7649		conf->prev_algo = conf->algorithm;
7650	}
7651
7652	conf->min_nr_stripes = NR_STRIPES;
7653	if (mddev->reshape_position != MaxSector) {
7654		int stripes = max_t(int,
7655			((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7656			((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7657		conf->min_nr_stripes = max(NR_STRIPES, stripes);
7658		if (conf->min_nr_stripes != NR_STRIPES)
7659			pr_info("md/raid:%s: force stripe size %d for reshape\n",
7660				mdname(mddev), conf->min_nr_stripes);
7661	}
7662	memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7663		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7664	atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7665	if (grow_stripes(conf, conf->min_nr_stripes)) {
7666		pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7667			mdname(mddev), memory);
7668		ret = -ENOMEM;
7669		goto abort;
7670	} else
7671		pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7672	/*
7673	 * Losing a stripe head costs more than the time to refill it,
7674	 * it reduces the queue depth and so can hurt throughput.
7675	 * So set it rather large, scaled by number of devices.
7676	 */
7677	conf->shrinker = shrinker_alloc(0, "md-raid5:%s", mdname(mddev));
7678	if (!conf->shrinker) {
7679		ret = -ENOMEM;
7680		pr_warn("md/raid:%s: couldn't allocate shrinker.\n",
7681			mdname(mddev));
7682		goto abort;
7683	}
7684
7685	conf->shrinker->seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7686	conf->shrinker->scan_objects = raid5_cache_scan;
7687	conf->shrinker->count_objects = raid5_cache_count;
7688	conf->shrinker->batch = 128;
7689	conf->shrinker->private_data = conf;
7690
7691	shrinker_register(conf->shrinker);
7692
7693	sprintf(pers_name, "raid%d", mddev->new_level);
7694	rcu_assign_pointer(conf->thread,
7695			   md_register_thread(raid5d, mddev, pers_name));
7696	if (!conf->thread) {
7697		pr_warn("md/raid:%s: couldn't allocate thread.\n",
7698			mdname(mddev));
7699		ret = -ENOMEM;
7700		goto abort;
7701	}
7702
7703	return conf;
7704
7705 abort:
7706	if (conf)
7707		free_conf(conf);
7708	return ERR_PTR(ret);
7709}
7710
7711static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7712{
7713	switch (algo) {
7714	case ALGORITHM_PARITY_0:
7715		if (raid_disk < max_degraded)
7716			return 1;
7717		break;
7718	case ALGORITHM_PARITY_N:
7719		if (raid_disk >= raid_disks - max_degraded)
7720			return 1;
7721		break;
7722	case ALGORITHM_PARITY_0_6:
7723		if (raid_disk == 0 ||
7724		    raid_disk == raid_disks - 1)
7725			return 1;
7726		break;
7727	case ALGORITHM_LEFT_ASYMMETRIC_6:
7728	case ALGORITHM_RIGHT_ASYMMETRIC_6:
7729	case ALGORITHM_LEFT_SYMMETRIC_6:
7730	case ALGORITHM_RIGHT_SYMMETRIC_6:
7731		if (raid_disk == raid_disks - 1)
7732			return 1;
7733	}
7734	return 0;
7735}
7736
7737static int raid5_set_limits(struct mddev *mddev)
7738{
7739	struct r5conf *conf = mddev->private;
7740	struct queue_limits lim;
7741	int data_disks, stripe;
7742	struct md_rdev *rdev;
7743
7744	/*
7745	 * The read-ahead size must cover two whole stripes, which is
7746	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices.
7747	 */
7748	data_disks = conf->previous_raid_disks - conf->max_degraded;
7749
7750	/*
7751	 * We can only discard a whole stripe. It doesn't make sense to
7752	 * discard data disk but write parity disk
7753	 */
7754	stripe = roundup_pow_of_two(data_disks * (mddev->chunk_sectors << 9));
7755
7756	md_init_stacking_limits(&lim);
7757	lim.logical_block_size = mddev->logical_block_size;
7758	lim.io_min = mddev->chunk_sectors << 9;
7759	lim.io_opt = lim.io_min * (conf->raid_disks - conf->max_degraded);
7760	lim.features |= BLK_FEAT_RAID_PARTIAL_STRIPES_EXPENSIVE;
7761	lim.discard_granularity = stripe;
7762	lim.max_write_zeroes_sectors = 0;
7763	lim.max_hw_wzeroes_unmap_sectors = 0;
7764	mddev_stack_rdev_limits(mddev, &lim, 0);
7765	rdev_for_each(rdev, mddev)
7766		queue_limits_stack_bdev(&lim, rdev->bdev, rdev->new_data_offset,
7767				mddev->gendisk->disk_name);
7768
7769	/*
7770	 * Zeroing is required for discard, otherwise data could be lost.
7771	 *
7772	 * Consider a scenario: discard a stripe (the stripe could be
7773	 * inconsistent if discard_zeroes_data is 0); write one disk of the
7774	 * stripe (the stripe could be inconsistent again depending on which
7775	 * disks are used to calculate parity); the disk is broken; The stripe
7776	 * data of this disk is lost.
7777	 *
7778	 * We only allow DISCARD if the sysadmin has confirmed that only safe
7779	 * devices are in use by setting a module parameter.  A better idea
7780	 * might be to turn DISCARD into WRITE_ZEROES requests, as that is
7781	 * required to be safe.
7782	 */
7783	if (!devices_handle_discard_safely ||
7784	    lim.max_discard_sectors < (stripe >> 9) ||
7785	    lim.discard_granularity < stripe)
7786		lim.max_hw_discard_sectors = 0;
7787
7788	/*
7789	 * Requests require having a bitmap for each stripe.
7790	 * Limit the max sectors based on this.
7791	 */
7792	lim.max_hw_sectors = RAID5_MAX_REQ_STRIPES << RAID5_STRIPE_SHIFT(conf);
7793
7794	/* No restrictions on the number of segments in the request */
7795	lim.max_segments = USHRT_MAX;
7796
7797	return queue_limits_set(mddev->gendisk->queue, &lim);
7798}
7799
7800static int raid5_run(struct mddev *mddev)
7801{
7802	struct r5conf *conf;
7803	int dirty_parity_disks = 0;
7804	struct md_rdev *rdev;
7805	struct md_rdev *journal_dev = NULL;
7806	sector_t reshape_offset = 0;
7807	int i;
7808	long long min_offset_diff = 0;
7809	int first = 1;
7810	int ret = -EIO;
7811
7812	if (mddev->resync_offset != MaxSector)
7813		pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7814			  mdname(mddev));
7815
7816	rdev_for_each(rdev, mddev) {
7817		long long diff;
7818
7819		if (test_bit(Journal, &rdev->flags)) {
7820			journal_dev = rdev;
7821			continue;
7822		}
7823		if (rdev->raid_disk < 0)
7824			continue;
7825		diff = (rdev->new_data_offset - rdev->data_offset);
7826		if (first) {
7827			min_offset_diff = diff;
7828			first = 0;
7829		} else if (mddev->reshape_backwards &&
7830			 diff < min_offset_diff)
7831			min_offset_diff = diff;
7832		else if (!mddev->reshape_backwards &&
7833			 diff > min_offset_diff)
7834			min_offset_diff = diff;
7835	}
7836
7837	if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7838	    (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7839		pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7840			  mdname(mddev));
7841		return -EINVAL;
7842	}
7843
7844	if (mddev->reshape_position != MaxSector) {
7845		/* Check that we can continue the reshape.
7846		 * Difficulties arise if the stripe we would write to
7847		 * next is at or after the stripe we would read from next.
7848		 * For a reshape that changes the number of devices, this
7849		 * is only possible for a very short time, and mdadm makes
7850		 * sure that time appears to have past before assembling
7851		 * the array.  So we fail if that time hasn't passed.
7852		 * For a reshape that keeps the number of devices the same
7853		 * mdadm must be monitoring the reshape can keeping the
7854		 * critical areas read-only and backed up.  It will start
7855		 * the array in read-only mode, so we check for that.
7856		 */
7857		sector_t here_new, here_old;
7858		int old_disks;
7859		int max_degraded = (mddev->level == 6 ? 2 : 1);
7860		int chunk_sectors;
7861		int new_data_disks;
7862
7863		if (journal_dev) {
7864			pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7865				mdname(mddev));
7866			return -EINVAL;
7867		}
7868
7869		if (mddev->new_level != mddev->level) {
7870			pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7871				mdname(mddev));
7872			return -EINVAL;
7873		}
7874		old_disks = mddev->raid_disks - mddev->delta_disks;
7875		/* reshape_position must be on a new-stripe boundary, and one
7876		 * further up in new geometry must map after here in old
7877		 * geometry.
7878		 * If the chunk sizes are different, then as we perform reshape
7879		 * in units of the largest of the two, reshape_position needs
7880		 * be a multiple of the largest chunk size times new data disks.
7881		 */
7882		here_new = mddev->reshape_position;
7883		chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7884		new_data_disks = mddev->raid_disks - max_degraded;
7885		if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7886			pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7887				mdname(mddev));
7888			return -EINVAL;
7889		}
7890		reshape_offset = here_new * chunk_sectors;
7891		/* here_new is the stripe we will write to */
7892		here_old = mddev->reshape_position;
7893		sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7894		/* here_old is the first stripe that we might need to read
7895		 * from */
7896		if (mddev->delta_disks == 0) {
7897			/* We cannot be sure it is safe to start an in-place
7898			 * reshape.  It is only safe if user-space is monitoring
7899			 * and taking constant backups.
7900			 * mdadm always starts a situation like this in
7901			 * readonly mode so it can take control before
7902			 * allowing any writes.  So just check for that.
7903			 */
7904			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7905			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
7906				/* not really in-place - so OK */;
7907			else if (mddev->ro == 0) {
7908				pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7909					mdname(mddev));
7910				return -EINVAL;
7911			}
7912		} else if (mddev->reshape_backwards
7913		    ? (here_new * chunk_sectors + min_offset_diff <=
7914		       here_old * chunk_sectors)
7915		    : (here_new * chunk_sectors >=
7916		       here_old * chunk_sectors + (-min_offset_diff))) {
7917			/* Reading from the same stripe as writing to - bad */
7918			pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7919				mdname(mddev));
7920			return -EINVAL;
7921		}
7922		pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7923		/* OK, we should be able to continue; */
7924	} else {
7925		BUG_ON(mddev->level != mddev->new_level);
7926		BUG_ON(mddev->layout != mddev->new_layout);
7927		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7928		BUG_ON(mddev->delta_disks != 0);
7929	}
7930
7931	if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7932	    test_bit(MD_HAS_PPL, &mddev->flags)) {
7933		pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7934			mdname(mddev));
7935		clear_bit(MD_HAS_PPL, &mddev->flags);
7936		clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7937	}
7938
7939	if (mddev->private == NULL)
7940		conf = setup_conf(mddev);
7941	else
7942		conf = mddev->private;
7943
7944	if (IS_ERR(conf))
7945		return PTR_ERR(conf);
7946
7947	if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7948		if (!journal_dev) {
7949			pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7950				mdname(mddev));
7951			mddev->ro = 1;
7952			set_disk_ro(mddev->gendisk, 1);
7953		} else if (mddev->resync_offset == MaxSector)
7954			set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7955	}
7956
7957	conf->min_offset_diff = min_offset_diff;
7958	rcu_assign_pointer(mddev->thread, conf->thread);
7959	rcu_assign_pointer(conf->thread, NULL);
7960	mddev->private = conf;
7961
7962	for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7963	     i++) {
7964		rdev = conf->disks[i].rdev;
7965		if (!rdev)
7966			continue;
7967		if (conf->disks[i].replacement &&
7968		    conf->reshape_progress != MaxSector) {
7969			/* replacements and reshape simply do not mix. */
7970			pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7971			goto abort;
7972		}
7973		if (test_bit(In_sync, &rdev->flags))
7974			continue;
7975		/* This disc is not fully in-sync.  However if it
7976		 * just stored parity (beyond the recovery_offset),
7977		 * when we don't need to be concerned about the
7978		 * array being dirty.
7979		 * When reshape goes 'backwards', we never have
7980		 * partially completed devices, so we only need
7981		 * to worry about reshape going forwards.
7982		 */
7983		/* Hack because v0.91 doesn't store recovery_offset properly. */
7984		if (mddev->major_version == 0 &&
7985		    mddev->minor_version > 90)
7986			rdev->recovery_offset = reshape_offset;
7987
7988		if (rdev->recovery_offset < reshape_offset) {
7989			/* We need to check old and new layout */
7990			if (!only_parity(rdev->raid_disk,
7991					 conf->algorithm,
7992					 conf->raid_disks,
7993					 conf->max_degraded))
7994				continue;
7995		}
7996		if (!only_parity(rdev->raid_disk,
7997				 conf->prev_algo,
7998				 conf->previous_raid_disks,
7999				 conf->max_degraded))
8000			continue;
8001		dirty_parity_disks++;
8002	}
8003
8004	/*
8005	 * 0 for a fully functional array, 1 or 2 for a degraded array.
8006	 */
8007	mddev->degraded = raid5_calc_degraded(conf);
8008
8009	if (has_failed(conf)) {
8010		pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
8011			mdname(mddev), mddev->degraded, conf->raid_disks);
8012		goto abort;
8013	}
8014
8015	/* device size must be a multiple of chunk size */
8016	mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
8017	mddev->resync_max_sectors = mddev->dev_sectors;
8018
8019	if (mddev->degraded > dirty_parity_disks &&
8020	    mddev->resync_offset != MaxSector) {
8021		if (test_bit(MD_HAS_PPL, &mddev->flags))
8022			pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
8023				mdname(mddev));
8024		else if (mddev->ok_start_degraded)
8025			pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
8026				mdname(mddev));
8027		else {
8028			pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
8029				mdname(mddev));
8030			goto abort;
8031		}
8032	}
8033
8034	pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
8035		mdname(mddev), conf->level,
8036		mddev->raid_disks-mddev->degraded, mddev->raid_disks,
8037		mddev->new_layout);
8038
8039	print_raid5_conf(conf);
8040
8041	if (conf->reshape_progress != MaxSector) {
8042		conf->reshape_safe = conf->reshape_progress;
8043		atomic_set(&conf->reshape_stripes, 0);
8044		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8045		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8046		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8047		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8048	}
8049
8050	/* Ok, everything is just fine now */
8051	if (mddev->to_remove == &raid5_attrs_group)
8052		mddev->to_remove = NULL;
8053	else if (mddev->kobj.sd &&
8054	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
8055		pr_warn("raid5: failed to create sysfs attributes for %s\n",
8056			mdname(mddev));
8057	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
8058
8059	if (!mddev_is_dm(mddev)) {
8060		ret = raid5_set_limits(mddev);
8061		if (ret)
8062			goto abort;
8063	}
8064
8065	ret = log_init(conf, journal_dev, raid5_has_ppl(conf));
8066	if (ret)
8067		goto abort;
8068
8069	return 0;
8070abort:
8071	md_unregister_thread(mddev, &mddev->thread);
8072	print_raid5_conf(conf);
8073	free_conf(conf);
8074	mddev->private = NULL;
8075	pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
8076	return ret;
8077}
8078
8079static void raid5_free(struct mddev *mddev, void *priv)
8080{
8081	struct r5conf *conf = priv;
8082
8083	free_conf(conf);
8084	mddev->to_remove = &raid5_attrs_group;
8085}
8086
8087static void raid5_status(struct seq_file *seq, struct mddev *mddev)
8088{
8089	struct r5conf *conf = mddev->private;
8090	int i;
8091
8092	lockdep_assert_held(&mddev->lock);
8093
8094	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
8095		conf->chunk_sectors / 2, mddev->layout);
8096	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
8097	for (i = 0; i < conf->raid_disks; i++) {
8098		struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
8099
8100		seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
8101	}
8102	seq_printf (seq, "]");
8103}
8104
8105static void print_raid5_conf(struct r5conf *conf)
8106{
8107	struct md_rdev *rdev;
8108	int i;
8109
8110	pr_debug("RAID conf printout:\n");
8111	if (!conf) {
8112		pr_debug("(conf==NULL)\n");
8113		return;
8114	}
8115	pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
8116	       conf->raid_disks,
8117	       conf->raid_disks - conf->mddev->degraded);
8118
8119	for (i = 0; i < conf->raid_disks; i++) {
8120		rdev = conf->disks[i].rdev;
8121		if (rdev)
8122			pr_debug(" disk %d, o:%d, dev:%pg\n",
8123			       i, !test_bit(Faulty, &rdev->flags),
8124			       rdev->bdev);
8125	}
8126}
8127
8128static int raid5_spare_active(struct mddev *mddev)
8129{
8130	int i;
8131	struct r5conf *conf = mddev->private;
8132	struct md_rdev *rdev, *replacement;
8133	int count = 0;
8134	unsigned long flags;
8135
8136	for (i = 0; i < conf->raid_disks; i++) {
8137		rdev = conf->disks[i].rdev;
8138		replacement = conf->disks[i].replacement;
8139		if (replacement
8140		    && replacement->recovery_offset == MaxSector
8141		    && !test_bit(Faulty, &replacement->flags)
8142		    && !test_and_set_bit(In_sync, &replacement->flags)) {
8143			/* Replacement has just become active. */
8144			if (!rdev
8145			    || !test_and_clear_bit(In_sync, &rdev->flags))
8146				count++;
8147			if (rdev) {
8148				/* Replaced device not technically faulty,
8149				 * but we need to be sure it gets removed
8150				 * and never re-added.
8151				 */
8152				set_bit(Faulty, &rdev->flags);
8153				sysfs_notify_dirent_safe(
8154					rdev->sysfs_state);
8155			}
8156			sysfs_notify_dirent_safe(replacement->sysfs_state);
8157		} else if (rdev
8158		    && rdev->recovery_offset == MaxSector
8159		    && !test_bit(Faulty, &rdev->flags)
8160		    && !test_and_set_bit(In_sync, &rdev->flags)) {
8161			count++;
8162			sysfs_notify_dirent_safe(rdev->sysfs_state);
8163		}
8164	}
8165	spin_lock_irqsave(&conf->device_lock, flags);
8166	mddev->degraded = raid5_calc_degraded(conf);
8167	spin_unlock_irqrestore(&conf->device_lock, flags);
8168	print_raid5_conf(conf);
8169	return count;
8170}
8171
8172static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
8173{
8174	struct r5conf *conf = mddev->private;
8175	int err = 0;
8176	int number = rdev->raid_disk;
8177	struct md_rdev **rdevp;
8178	struct disk_info *p;
8179	struct md_rdev *tmp;
8180
8181	print_raid5_conf(conf);
8182	if (test_bit(Journal, &rdev->flags) && conf->log) {
8183		/*
8184		 * we can't wait pending write here, as this is called in
8185		 * raid5d, wait will deadlock.
8186		 * neilb: there is no locking about new writes here,
8187		 * so this cannot be safe.
8188		 */
8189		if (atomic_read(&conf->active_stripes) ||
8190		    atomic_read(&conf->r5c_cached_full_stripes) ||
8191		    atomic_read(&conf->r5c_cached_partial_stripes)) {
8192			return -EBUSY;
8193		}
8194		log_exit(conf);
8195		return 0;
8196	}
8197	if (unlikely(number >= conf->pool_size))
8198		return 0;
8199	p = conf->disks + number;
8200	if (rdev == p->rdev)
8201		rdevp = &p->rdev;
8202	else if (rdev == p->replacement)
8203		rdevp = &p->replacement;
8204	else
8205		return 0;
8206
8207	if (number >= conf->raid_disks &&
8208	    conf->reshape_progress == MaxSector)
8209		clear_bit(In_sync, &rdev->flags);
8210
8211	if (test_bit(In_sync, &rdev->flags) ||
8212	    atomic_read(&rdev->nr_pending)) {
8213		err = -EBUSY;
8214		goto abort;
8215	}
8216	/* Only remove non-faulty devices if recovery
8217	 * isn't possible.
8218	 */
8219	if (!test_bit(Faulty, &rdev->flags) &&
8220	    mddev->recovery_disabled != conf->recovery_disabled &&
8221	    !has_failed(conf) &&
8222	    (!p->replacement || p->replacement == rdev) &&
8223	    number < conf->raid_disks) {
8224		err = -EBUSY;
8225		goto abort;
8226	}
8227	WRITE_ONCE(*rdevp, NULL);
8228	if (!err) {
8229		err = log_modify(conf, rdev, false);
8230		if (err)
8231			goto abort;
8232	}
8233
8234	tmp = p->replacement;
8235	if (tmp) {
8236		/* We must have just cleared 'rdev' */
8237		WRITE_ONCE(p->rdev, tmp);
8238		clear_bit(Replacement, &tmp->flags);
8239		WRITE_ONCE(p->replacement, NULL);
8240
8241		if (!err)
8242			err = log_modify(conf, tmp, true);
8243	}
8244
8245	clear_bit(WantReplacement, &rdev->flags);
8246abort:
8247
8248	print_raid5_conf(conf);
8249	return err;
8250}
8251
8252static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
8253{
8254	struct r5conf *conf = mddev->private;
8255	int ret, err = -EEXIST;
8256	int disk;
8257	struct disk_info *p;
8258	struct md_rdev *tmp;
8259	int first = 0;
8260	int last = conf->raid_disks - 1;
8261
8262	if (test_bit(Journal, &rdev->flags)) {
8263		if (conf->log)
8264			return -EBUSY;
8265
8266		rdev->raid_disk = 0;
8267		/*
8268		 * The array is in readonly mode if journal is missing, so no
8269		 * write requests running. We should be safe
8270		 */
8271		ret = log_init(conf, rdev, false);
8272		if (ret)
8273			return ret;
8274
8275		ret = r5l_start(conf->log);
8276		if (ret)
8277			return ret;
8278
8279		return 0;
8280	}
8281	if (mddev->recovery_disabled == conf->recovery_disabled)
8282		return -EBUSY;
8283
8284	if (rdev->saved_raid_disk < 0 && has_failed(conf))
8285		/* no point adding a device */
8286		return -EINVAL;
8287
8288	if (rdev->raid_disk >= 0)
8289		first = last = rdev->raid_disk;
8290
8291	/*
8292	 * find the disk ... but prefer rdev->saved_raid_disk
8293	 * if possible.
8294	 */
8295	if (rdev->saved_raid_disk >= first &&
8296	    rdev->saved_raid_disk <= last &&
8297	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
8298		first = rdev->saved_raid_disk;
8299
8300	for (disk = first; disk <= last; disk++) {
8301		p = conf->disks + disk;
8302		if (p->rdev == NULL) {
8303			clear_bit(In_sync, &rdev->flags);
8304			rdev->raid_disk = disk;
8305			if (rdev->saved_raid_disk != disk)
8306				conf->fullsync = 1;
8307			WRITE_ONCE(p->rdev, rdev);
8308
8309			err = log_modify(conf, rdev, true);
8310
8311			goto out;
8312		}
8313	}
8314	for (disk = first; disk <= last; disk++) {
8315		p = conf->disks + disk;
8316		tmp = p->rdev;
8317		if (test_bit(WantReplacement, &tmp->flags) &&
8318		    mddev->reshape_position == MaxSector &&
8319		    p->replacement == NULL) {
8320			clear_bit(In_sync, &rdev->flags);
8321			set_bit(Replacement, &rdev->flags);
8322			rdev->raid_disk = disk;
8323			err = 0;
8324			conf->fullsync = 1;
8325			WRITE_ONCE(p->replacement, rdev);
8326			break;
8327		}
8328	}
8329out:
8330	print_raid5_conf(conf);
8331	return err;
8332}
8333
8334static int raid5_resize(struct mddev *mddev, sector_t sectors)
8335{
8336	/* no resync is happening, and there is enough space
8337	 * on all devices, so we can resize.
8338	 * We need to make sure resync covers any new space.
8339	 * If the array is shrinking we should possibly wait until
8340	 * any io in the removed space completes, but it hardly seems
8341	 * worth it.
8342	 */
8343	sector_t newsize;
8344	struct r5conf *conf = mddev->private;
8345
8346	if (raid5_has_log(conf) || raid5_has_ppl(conf))
8347		return -EINVAL;
8348	sectors &= ~((sector_t)conf->chunk_sectors - 1);
8349	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8350	if (mddev->external_size &&
8351	    mddev->array_sectors > newsize)
8352		return -EINVAL;
8353
8354	if (md_bitmap_enabled(mddev, false)) {
8355		int ret = mddev->bitmap_ops->resize(mddev, sectors, 0);
8356
8357		if (ret)
8358			return ret;
8359	}
8360
8361	md_set_array_sectors(mddev, newsize);
8362	if (sectors > mddev->dev_sectors &&
8363	    mddev->resync_offset > mddev->dev_sectors) {
8364		mddev->resync_offset = mddev->dev_sectors;
8365		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8366	}
8367	mddev->dev_sectors = sectors;
8368	mddev->resync_max_sectors = sectors;
8369	return 0;
8370}
8371
8372static int check_stripe_cache(struct mddev *mddev)
8373{
8374	/* Can only proceed if there are plenty of stripe_heads.
8375	 * We need a minimum of one full stripe,, and for sensible progress
8376	 * it is best to have about 4 times that.
8377	 * If we require 4 times, then the default 256 4K stripe_heads will
8378	 * allow for chunk sizes up to 256K, which is probably OK.
8379	 * If the chunk size is greater, user-space should request more
8380	 * stripe_heads first.
8381	 */
8382	struct r5conf *conf = mddev->private;
8383	if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8384	    > conf->min_nr_stripes ||
8385	    ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8386	    > conf->min_nr_stripes) {
8387		pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
8388			mdname(mddev),
8389			((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8390			 / RAID5_STRIPE_SIZE(conf))*4);
8391		return 0;
8392	}
8393	return 1;
8394}
8395
8396static int check_reshape(struct mddev *mddev)
8397{
8398	struct r5conf *conf = mddev->private;
8399
8400	if (raid5_has_log(conf) || raid5_has_ppl(conf))
8401		return -EINVAL;
8402	if (mddev->delta_disks == 0 &&
8403	    mddev->new_layout == mddev->layout &&
8404	    mddev->new_chunk_sectors == mddev->chunk_sectors)
8405		return 0; /* nothing to do */
8406	if (has_failed(conf))
8407		return -EINVAL;
8408	if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8409		/* We might be able to shrink, but the devices must
8410		 * be made bigger first.
8411		 * For raid6, 4 is the minimum size.
8412		 * Otherwise 2 is the minimum
8413		 */
8414		int min = 2;
8415		if (mddev->level == 6)
8416			min = 4;
8417		if (mddev->raid_disks + mddev->delta_disks < min)
8418			return -EINVAL;
8419	}
8420
8421	if (!check_stripe_cache(mddev))
8422		return -ENOSPC;
8423
8424	if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8425	    mddev->delta_disks > 0)
8426		if (resize_chunks(conf,
8427				  conf->previous_raid_disks
8428				  + max(0, mddev->delta_disks),
8429				  max(mddev->new_chunk_sectors,
8430				      mddev->chunk_sectors)
8431			    ) < 0)
8432			return -ENOMEM;
8433
8434	if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8435		return 0; /* never bother to shrink */
8436	return resize_stripes(conf, (conf->previous_raid_disks
8437				     + mddev->delta_disks));
8438}
8439
8440static int raid5_start_reshape(struct mddev *mddev)
8441{
8442	struct r5conf *conf = mddev->private;
8443	struct md_rdev *rdev;
8444	int spares = 0;
8445	int i;
8446	unsigned long flags;
8447
8448	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8449		return -EBUSY;
8450
8451	if (!check_stripe_cache(mddev))
8452		return -ENOSPC;
8453
8454	if (has_failed(conf))
8455		return -EINVAL;
8456
8457	/* raid5 can't handle concurrent reshape and recovery */
8458	if (mddev->resync_offset < MaxSector)
8459		return -EBUSY;
8460	for (i = 0; i < conf->raid_disks; i++)
8461		if (conf->disks[i].replacement)
8462			return -EBUSY;
8463
8464	rdev_for_each(rdev, mddev) {
8465		if (!test_bit(In_sync, &rdev->flags)
8466		    && !test_bit(Faulty, &rdev->flags))
8467			spares++;
8468	}
8469
8470	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8471		/* Not enough devices even to make a degraded array
8472		 * of that size
8473		 */
8474		return -EINVAL;
8475
8476	/* Refuse to reduce size of the array.  Any reductions in
8477	 * array size must be through explicit setting of array_size
8478	 * attribute.
8479	 */
8480	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8481	    < mddev->array_sectors) {
8482		pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8483			mdname(mddev));
8484		return -EINVAL;
8485	}
8486
8487	atomic_set(&conf->reshape_stripes, 0);
8488	spin_lock_irq(&conf->device_lock);
8489	write_seqcount_begin(&conf->gen_lock);
8490	conf->previous_raid_disks = conf->raid_disks;
8491	conf->raid_disks += mddev->delta_disks;
8492	conf->prev_chunk_sectors = conf->chunk_sectors;
8493	conf->chunk_sectors = mddev->new_chunk_sectors;
8494	conf->prev_algo = conf->algorithm;
8495	conf->algorithm = mddev->new_layout;
8496	conf->generation++;
8497	/* Code that selects data_offset needs to see the generation update
8498	 * if reshape_progress has been set - so a memory barrier needed.
8499	 */
8500	smp_mb();
8501	if (mddev->reshape_backwards)
8502		conf->reshape_progress = raid5_size(mddev, 0, 0);
8503	else
8504		conf->reshape_progress = 0;
8505	conf->reshape_safe = conf->reshape_progress;
8506	write_seqcount_end(&conf->gen_lock);
8507	spin_unlock_irq(&conf->device_lock);
8508
8509	/* Now make sure any requests that proceeded on the assumption
8510	 * the reshape wasn't running - like Discard or Read - have
8511	 * completed.
8512	 */
8513	raid5_quiesce(mddev, true);
8514	raid5_quiesce(mddev, false);
8515
8516	/* Add some new drives, as many as will fit.
8517	 * We know there are enough to make the newly sized array work.
8518	 * Don't add devices if we are reducing the number of
8519	 * devices in the array.  This is because it is not possible
8520	 * to correctly record the "partially reconstructed" state of
8521	 * such devices during the reshape and confusion could result.
8522	 */
8523	if (mddev->delta_disks >= 0) {
8524		rdev_for_each(rdev, mddev)
8525			if (rdev->raid_disk < 0 &&
8526			    !test_bit(Faulty, &rdev->flags)) {
8527				if (raid5_add_disk(mddev, rdev) == 0) {
8528					if (rdev->raid_disk
8529					    >= conf->previous_raid_disks)
8530						set_bit(In_sync, &rdev->flags);
8531					else
8532						rdev->recovery_offset = 0;
8533
8534					/* Failure here is OK */
8535					sysfs_link_rdev(mddev, rdev);
8536				}
8537			} else if (rdev->raid_disk >= conf->previous_raid_disks
8538				   && !test_bit(Faulty, &rdev->flags)) {
8539				/* This is a spare that was manually added */
8540				set_bit(In_sync, &rdev->flags);
8541			}
8542
8543		/* When a reshape changes the number of devices,
8544		 * ->degraded is measured against the larger of the
8545		 * pre and post number of devices.
8546		 */
8547		spin_lock_irqsave(&conf->device_lock, flags);
8548		mddev->degraded = raid5_calc_degraded(conf);
8549		spin_unlock_irqrestore(&conf->device_lock, flags);
8550	}
8551	mddev->raid_disks = conf->raid_disks;
8552	mddev->reshape_position = conf->reshape_progress;
8553	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8554
8555	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8556	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8557	clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8558	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8559	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8560	conf->reshape_checkpoint = jiffies;
8561	md_new_event();
8562	return 0;
8563}
8564
8565/* This is called from the reshape thread and should make any
8566 * changes needed in 'conf'
8567 */
8568static void end_reshape(struct r5conf *conf)
8569{
8570
8571	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8572		struct md_rdev *rdev;
8573
8574		spin_lock_irq(&conf->device_lock);
8575		conf->previous_raid_disks = conf->raid_disks;
8576		md_finish_reshape(conf->mddev);
8577		smp_wmb();
8578		conf->reshape_progress = MaxSector;
8579		conf->mddev->reshape_position = MaxSector;
8580		rdev_for_each(rdev, conf->mddev)
8581			if (rdev->raid_disk >= 0 &&
8582			    !test_bit(Journal, &rdev->flags) &&
8583			    !test_bit(In_sync, &rdev->flags))
8584				rdev->recovery_offset = MaxSector;
8585		spin_unlock_irq(&conf->device_lock);
8586		wake_up(&conf->wait_for_reshape);
8587
8588		mddev_update_io_opt(conf->mddev,
8589			conf->raid_disks - conf->max_degraded);
8590	}
8591}
8592
8593/* This is called from the raid5d thread with mddev_lock held.
8594 * It makes config changes to the device.
8595 */
8596static void raid5_finish_reshape(struct mddev *mddev)
8597{
8598	struct r5conf *conf = mddev->private;
8599	struct md_rdev *rdev;
8600
8601	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8602
8603		if (mddev->delta_disks <= 0) {
8604			int d;
8605			spin_lock_irq(&conf->device_lock);
8606			mddev->degraded = raid5_calc_degraded(conf);
8607			spin_unlock_irq(&conf->device_lock);
8608			for (d = conf->raid_disks ;
8609			     d < conf->raid_disks - mddev->delta_disks;
8610			     d++) {
8611				rdev = conf->disks[d].rdev;
8612				if (rdev)
8613					clear_bit(In_sync, &rdev->flags);
8614				rdev = conf->disks[d].replacement;
8615				if (rdev)
8616					clear_bit(In_sync, &rdev->flags);
8617			}
8618		}
8619		mddev->layout = conf->algorithm;
8620		mddev->chunk_sectors = conf->chunk_sectors;
8621		mddev->reshape_position = MaxSector;
8622		mddev->delta_disks = 0;
8623		mddev->reshape_backwards = 0;
8624	}
8625}
8626
8627static void raid5_quiesce(struct mddev *mddev, int quiesce)
8628{
8629	struct r5conf *conf = mddev->private;
8630
8631	if (quiesce) {
8632		/* stop all writes */
8633		lock_all_device_hash_locks_irq(conf);
8634		/* '2' tells resync/reshape to pause so that all
8635		 * active stripes can drain
8636		 */
8637		r5c_flush_cache(conf, INT_MAX);
8638		/* need a memory barrier to make sure read_one_chunk() sees
8639		 * quiesce started and reverts to slow (locked) path.
8640		 */
8641		smp_store_release(&conf->quiesce, 2);
8642		wait_event_cmd(conf->wait_for_quiescent,
8643				    atomic_read(&conf->active_stripes) == 0 &&
8644				    atomic_read(&conf->active_aligned_reads) == 0,
8645				    unlock_all_device_hash_locks_irq(conf),
8646				    lock_all_device_hash_locks_irq(conf));
8647		conf->quiesce = 1;
8648		unlock_all_device_hash_locks_irq(conf);
8649		/* allow reshape to continue */
8650		wake_up(&conf->wait_for_reshape);
8651	} else {
8652		/* re-enable writes */
8653		lock_all_device_hash_locks_irq(conf);
8654		conf->quiesce = 0;
8655		wake_up(&conf->wait_for_quiescent);
8656		wake_up(&conf->wait_for_reshape);
8657		unlock_all_device_hash_locks_irq(conf);
8658	}
8659	log_quiesce(conf, quiesce);
8660}
8661
8662static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8663{
8664	struct r0conf *raid0_conf = mddev->private;
8665	sector_t sectors;
8666
8667	/* for raid0 takeover only one zone is supported */
8668	if (raid0_conf->nr_strip_zones > 1) {
8669		pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8670			mdname(mddev));
8671		return ERR_PTR(-EINVAL);
8672	}
8673
8674	sectors = raid0_conf->strip_zone[0].zone_end;
8675	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8676	mddev->dev_sectors = sectors;
8677	mddev->new_level = level;
8678	mddev->new_layout = ALGORITHM_PARITY_N;
8679	mddev->new_chunk_sectors = mddev->chunk_sectors;
8680	mddev->raid_disks += 1;
8681	mddev->delta_disks = 1;
8682	/* make sure it will be not marked as dirty */
8683	mddev->resync_offset = MaxSector;
8684
8685	return setup_conf(mddev);
8686}
8687
8688static void *raid5_takeover_raid1(struct mddev *mddev)
8689{
8690	int chunksect;
8691	void *ret;
8692
8693	if (mddev->raid_disks != 2 ||
8694	    mddev->degraded > 1)
8695		return ERR_PTR(-EINVAL);
8696
8697	/* Should check if there are write-behind devices? */
8698
8699	chunksect = 64*2; /* 64K by default */
8700
8701	/* The array must be an exact multiple of chunksize */
8702	while (chunksect && (mddev->array_sectors & (chunksect-1)))
8703		chunksect >>= 1;
8704
8705	if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8706		/* array size does not allow a suitable chunk size */
8707		return ERR_PTR(-EINVAL);
8708
8709	mddev->new_level = 5;
8710	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8711	mddev->new_chunk_sectors = chunksect;
8712
8713	ret = setup_conf(mddev);
8714	if (!IS_ERR(ret))
8715		mddev_clear_unsupported_flags(mddev,
8716			UNSUPPORTED_MDDEV_FLAGS);
8717	return ret;
8718}
8719
8720static void *raid5_takeover_raid6(struct mddev *mddev)
8721{
8722	int new_layout;
8723
8724	switch (mddev->layout) {
8725	case ALGORITHM_LEFT_ASYMMETRIC_6:
8726		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8727		break;
8728	case ALGORITHM_RIGHT_ASYMMETRIC_6:
8729		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8730		break;
8731	case ALGORITHM_LEFT_SYMMETRIC_6:
8732		new_layout = ALGORITHM_LEFT_SYMMETRIC;
8733		break;
8734	case ALGORITHM_RIGHT_SYMMETRIC_6:
8735		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8736		break;
8737	case ALGORITHM_PARITY_0_6:
8738		new_layout = ALGORITHM_PARITY_0;
8739		break;
8740	case ALGORITHM_PARITY_N:
8741		new_layout = ALGORITHM_PARITY_N;
8742		break;
8743	default:
8744		return ERR_PTR(-EINVAL);
8745	}
8746	mddev->new_level = 5;
8747	mddev->new_layout = new_layout;
8748	mddev->delta_disks = -1;
8749	mddev->raid_disks -= 1;
8750	return setup_conf(mddev);
8751}
8752
8753static int raid5_check_reshape(struct mddev *mddev)
8754{
8755	/* For a 2-drive array, the layout and chunk size can be changed
8756	 * immediately as not restriping is needed.
8757	 * For larger arrays we record the new value - after validation
8758	 * to be used by a reshape pass.
8759	 */
8760	struct r5conf *conf = mddev->private;
8761	int new_chunk = mddev->new_chunk_sectors;
8762
8763	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8764		return -EINVAL;
8765	if (new_chunk > 0) {
8766		if (!is_power_of_2(new_chunk))
8767			return -EINVAL;
8768		if (new_chunk < (PAGE_SIZE>>9))
8769			return -EINVAL;
8770		if (mddev->array_sectors & (new_chunk-1))
8771			/* not factor of array size */
8772			return -EINVAL;
8773	}
8774
8775	/* They look valid */
8776
8777	if (mddev->raid_disks == 2) {
8778		/* can make the change immediately */
8779		if (mddev->new_layout >= 0) {
8780			conf->algorithm = mddev->new_layout;
8781			mddev->layout = mddev->new_layout;
8782		}
8783		if (new_chunk > 0) {
8784			conf->chunk_sectors = new_chunk ;
8785			mddev->chunk_sectors = new_chunk;
8786		}
8787		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8788		md_wakeup_thread(mddev->thread);
8789	}
8790	return check_reshape(mddev);
8791}
8792
8793static int raid6_check_reshape(struct mddev *mddev)
8794{
8795	int new_chunk = mddev->new_chunk_sectors;
8796
8797	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8798		return -EINVAL;
8799	if (new_chunk > 0) {
8800		if (!is_power_of_2(new_chunk))
8801			return -EINVAL;
8802		if (new_chunk < (PAGE_SIZE >> 9))
8803			return -EINVAL;
8804		if (mddev->array_sectors & (new_chunk-1))
8805			/* not factor of array size */
8806			return -EINVAL;
8807	}
8808
8809	/* They look valid */
8810	return check_reshape(mddev);
8811}
8812
8813static void *raid5_takeover(struct mddev *mddev)
8814{
8815	/* raid5 can take over:
8816	 *  raid0 - if there is only one strip zone - make it a raid4 layout
8817	 *  raid1 - if there are two drives.  We need to know the chunk size
8818	 *  raid4 - trivial - just use a raid4 layout.
8819	 *  raid6 - Providing it is a *_6 layout
8820	 */
8821	if (mddev->level == 0)
8822		return raid45_takeover_raid0(mddev, 5);
8823	if (mddev->level == 1)
8824		return raid5_takeover_raid1(mddev);
8825	if (mddev->level == 4) {
8826		mddev->new_layout = ALGORITHM_PARITY_N;
8827		mddev->new_level = 5;
8828		return setup_conf(mddev);
8829	}
8830	if (mddev->level == 6)
8831		return raid5_takeover_raid6(mddev);
8832
8833	return ERR_PTR(-EINVAL);
8834}
8835
8836static void *raid4_takeover(struct mddev *mddev)
8837{
8838	/* raid4 can take over:
8839	 *  raid0 - if there is only one strip zone
8840	 *  raid5 - if layout is right
8841	 */
8842	if (mddev->level == 0)
8843		return raid45_takeover_raid0(mddev, 4);
8844	if (mddev->level == 5 &&
8845	    mddev->layout == ALGORITHM_PARITY_N) {
8846		mddev->new_layout = 0;
8847		mddev->new_level = 4;
8848		return setup_conf(mddev);
8849	}
8850	return ERR_PTR(-EINVAL);
8851}
8852
8853static struct md_personality raid5_personality;
8854
8855static void *raid6_takeover(struct mddev *mddev)
8856{
8857	/* Currently can only take over a raid5.  We map the
8858	 * personality to an equivalent raid6 personality
8859	 * with the Q block at the end.
8860	 */
8861	int new_layout;
8862
8863	if (mddev->pers != &raid5_personality)
8864		return ERR_PTR(-EINVAL);
8865	if (mddev->degraded > 1)
8866		return ERR_PTR(-EINVAL);
8867	if (mddev->raid_disks > 253)
8868		return ERR_PTR(-EINVAL);
8869	if (mddev->raid_disks < 3)
8870		return ERR_PTR(-EINVAL);
8871
8872	switch (mddev->layout) {
8873	case ALGORITHM_LEFT_ASYMMETRIC:
8874		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8875		break;
8876	case ALGORITHM_RIGHT_ASYMMETRIC:
8877		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8878		break;
8879	case ALGORITHM_LEFT_SYMMETRIC:
8880		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8881		break;
8882	case ALGORITHM_RIGHT_SYMMETRIC:
8883		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8884		break;
8885	case ALGORITHM_PARITY_0:
8886		new_layout = ALGORITHM_PARITY_0_6;
8887		break;
8888	case ALGORITHM_PARITY_N:
8889		new_layout = ALGORITHM_PARITY_N;
8890		break;
8891	default:
8892		return ERR_PTR(-EINVAL);
8893	}
8894	mddev->new_level = 6;
8895	mddev->new_layout = new_layout;
8896	mddev->delta_disks = 1;
8897	mddev->raid_disks += 1;
8898	return setup_conf(mddev);
8899}
8900
8901static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8902{
8903	struct r5conf *conf;
8904	int err;
8905
8906	err = mddev_suspend_and_lock(mddev);
8907	if (err)
8908		return err;
8909	conf = mddev->private;
8910	if (!conf) {
8911		mddev_unlock_and_resume(mddev);
8912		return -ENODEV;
8913	}
8914
8915	if (strncmp(buf, "ppl", 3) == 0) {
8916		/* ppl only works with RAID 5 */
8917		if (!raid5_has_ppl(conf) && conf->level == 5) {
8918			err = log_init(conf, NULL, true);
8919			if (!err) {
8920				err = resize_stripes(conf, conf->pool_size);
8921				if (err)
8922					log_exit(conf);
8923			}
8924		} else
8925			err = -EINVAL;
8926	} else if (strncmp(buf, "resync", 6) == 0) {
8927		if (raid5_has_ppl(conf)) {
8928			log_exit(conf);
8929			err = resize_stripes(conf, conf->pool_size);
8930		} else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8931			   r5l_log_disk_error(conf)) {
8932			bool journal_dev_exists = false;
8933			struct md_rdev *rdev;
8934
8935			rdev_for_each(rdev, mddev)
8936				if (test_bit(Journal, &rdev->flags)) {
8937					journal_dev_exists = true;
8938					break;
8939				}
8940
8941			if (!journal_dev_exists)
8942				clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8943			else  /* need remove journal device first */
8944				err = -EBUSY;
8945		} else
8946			err = -EINVAL;
8947	} else {
8948		err = -EINVAL;
8949	}
8950
8951	if (!err)
8952		md_update_sb(mddev, 1);
8953
8954	mddev_unlock_and_resume(mddev);
8955
8956	return err;
8957}
8958
8959static int raid5_start(struct mddev *mddev)
8960{
8961	struct r5conf *conf = mddev->private;
8962
8963	return r5l_start(conf->log);
8964}
8965
8966/*
8967 * This is only used for dm-raid456, caller already frozen sync_thread, hence
8968 * if rehsape is still in progress, io that is waiting for reshape can never be
8969 * done now, hence wake up and handle those IO.
8970 */
8971static void raid5_prepare_suspend(struct mddev *mddev)
8972{
8973	struct r5conf *conf = mddev->private;
8974
8975	wake_up(&conf->wait_for_reshape);
8976}
8977
8978static struct md_personality raid6_personality =
8979{
8980	.head = {
8981		.type	= MD_PERSONALITY,
8982		.id	= ID_RAID6,
8983		.name	= "raid6",
8984		.owner	= THIS_MODULE,
8985	},
8986
8987	.make_request	= raid5_make_request,
8988	.run		= raid5_run,
8989	.start		= raid5_start,
8990	.free		= raid5_free,
8991	.status		= raid5_status,
8992	.error_handler	= raid5_error,
8993	.hot_add_disk	= raid5_add_disk,
8994	.hot_remove_disk= raid5_remove_disk,
8995	.spare_active	= raid5_spare_active,
8996	.sync_request	= raid5_sync_request,
8997	.resize		= raid5_resize,
8998	.size		= raid5_size,
8999	.check_reshape	= raid6_check_reshape,
9000	.start_reshape  = raid5_start_reshape,
9001	.finish_reshape = raid5_finish_reshape,
9002	.quiesce	= raid5_quiesce,
9003	.takeover	= raid6_takeover,
9004	.change_consistency_policy = raid5_change_consistency_policy,
9005	.prepare_suspend = raid5_prepare_suspend,
9006	.bitmap_sector	= raid5_bitmap_sector,
9007};
9008static struct md_personality raid5_personality =
9009{
9010	.head = {
9011		.type	= MD_PERSONALITY,
9012		.id	= ID_RAID5,
9013		.name	= "raid5",
9014		.owner	= THIS_MODULE,
9015	},
9016
9017	.make_request	= raid5_make_request,
9018	.run		= raid5_run,
9019	.start		= raid5_start,
9020	.free		= raid5_free,
9021	.status		= raid5_status,
9022	.error_handler	= raid5_error,
9023	.hot_add_disk	= raid5_add_disk,
9024	.hot_remove_disk= raid5_remove_disk,
9025	.spare_active	= raid5_spare_active,
9026	.sync_request	= raid5_sync_request,
9027	.resize		= raid5_resize,
9028	.size		= raid5_size,
9029	.check_reshape	= raid5_check_reshape,
9030	.start_reshape  = raid5_start_reshape,
9031	.finish_reshape = raid5_finish_reshape,
9032	.quiesce	= raid5_quiesce,
9033	.takeover	= raid5_takeover,
9034	.change_consistency_policy = raid5_change_consistency_policy,
9035	.prepare_suspend = raid5_prepare_suspend,
9036	.bitmap_sector	= raid5_bitmap_sector,
9037};
9038
9039static struct md_personality raid4_personality =
9040{
9041	.head = {
9042		.type	= MD_PERSONALITY,
9043		.id	= ID_RAID4,
9044		.name	= "raid4",
9045		.owner	= THIS_MODULE,
9046	},
9047
9048	.make_request	= raid5_make_request,
9049	.run		= raid5_run,
9050	.start		= raid5_start,
9051	.free		= raid5_free,
9052	.status		= raid5_status,
9053	.error_handler	= raid5_error,
9054	.hot_add_disk	= raid5_add_disk,
9055	.hot_remove_disk= raid5_remove_disk,
9056	.spare_active	= raid5_spare_active,
9057	.sync_request	= raid5_sync_request,
9058	.resize		= raid5_resize,
9059	.size		= raid5_size,
9060	.check_reshape	= raid5_check_reshape,
9061	.start_reshape  = raid5_start_reshape,
9062	.finish_reshape = raid5_finish_reshape,
9063	.quiesce	= raid5_quiesce,
9064	.takeover	= raid4_takeover,
9065	.change_consistency_policy = raid5_change_consistency_policy,
9066	.prepare_suspend = raid5_prepare_suspend,
9067	.bitmap_sector	= raid5_bitmap_sector,
9068};
9069
9070static int __init raid5_init(void)
9071{
9072	int ret;
9073
9074	raid5_wq = alloc_workqueue("raid5wq",
9075		WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_SYSFS, 0);
9076	if (!raid5_wq)
9077		return -ENOMEM;
9078
9079	ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
9080				      "md/raid5:prepare",
9081				      raid456_cpu_up_prepare,
9082				      raid456_cpu_dead);
9083	if (ret)
9084		goto err_destroy_wq;
9085
9086	ret = register_md_submodule(&raid6_personality.head);
9087	if (ret)
9088		goto err_cpuhp_remove;
9089
9090	ret = register_md_submodule(&raid5_personality.head);
9091	if (ret)
9092		goto err_unregister_raid6;
9093
9094	ret = register_md_submodule(&raid4_personality.head);
9095	if (ret)
9096		goto err_unregister_raid5;
9097
9098	return 0;
9099
9100err_unregister_raid5:
9101	unregister_md_submodule(&raid5_personality.head);
9102err_unregister_raid6:
9103	unregister_md_submodule(&raid6_personality.head);
9104err_cpuhp_remove:
9105	cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
9106err_destroy_wq:
9107	destroy_workqueue(raid5_wq);
9108	return ret;
9109}
9110
9111static void __exit raid5_exit(void)
9112{
9113	unregister_md_submodule(&raid6_personality.head);
9114	unregister_md_submodule(&raid5_personality.head);
9115	unregister_md_submodule(&raid4_personality.head);
9116	cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
9117	destroy_workqueue(raid5_wq);
9118}
9119
9120module_init(raid5_init);
9121module_exit(raid5_exit);
9122MODULE_LICENSE("GPL");
9123MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
9124MODULE_ALIAS("md-personality-4"); /* RAID5 */
9125MODULE_ALIAS("md-raid5");
9126MODULE_ALIAS("md-raid4");
9127MODULE_ALIAS("md-level-5");
9128MODULE_ALIAS("md-level-4");
9129MODULE_ALIAS("md-personality-8"); /* RAID6 */
9130MODULE_ALIAS("md-raid6");
9131MODULE_ALIAS("md-level-6");
9132
9133/* This used to be two separate modules, they were: */
9134MODULE_ALIAS("raid5");
9135MODULE_ALIAS("raid6");