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};
 764
 765struct stripe_request_ctx {
 766	/* a reference to the last stripe_head for batching */
 767	struct stripe_head *batch_last;
 768
 769	/* first sector in the request */
 770	sector_t first_sector;
 771
 772	/* last sector in the request */
 773	sector_t last_sector;
 774
 775	/*
 776	 * bitmap to track stripe sectors that have been added to stripes
 777	 * add one to account for unaligned requests
 778	 */
 779	DECLARE_BITMAP(sectors_to_do, RAID5_MAX_REQ_STRIPES + 1);
 780
 781	/* the request had REQ_PREFLUSH, cleared after the first stripe_head */
 782	bool do_flush;
 783};
 784
 785/*
 786 * Block until another thread clears R5_INACTIVE_BLOCKED or
 787 * there are fewer than 3/4 the maximum number of active stripes
 788 * and there is an inactive stripe available.
 789 */
 790static bool is_inactive_blocked(struct r5conf *conf, int hash)
 791{
 792	if (list_empty(conf->inactive_list + hash))
 793		return false;
 794
 795	if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
 796		return true;
 797
 798	return (atomic_read(&conf->active_stripes) <
 799		(conf->max_nr_stripes * 3 / 4));
 800}
 801
 802struct stripe_head *raid5_get_active_stripe(struct r5conf *conf,
 803		struct stripe_request_ctx *ctx, sector_t sector,
 804		unsigned int flags)
 805{
 806	struct stripe_head *sh;
 807	int hash = stripe_hash_locks_hash(conf, sector);
 808	int previous = !!(flags & R5_GAS_PREVIOUS);
 809
 810	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 811
 812	spin_lock_irq(conf->hash_locks + hash);
 813
 814	for (;;) {
 815		if (!(flags & R5_GAS_NOQUIESCE) && conf->quiesce) {
 816			/*
 817			 * Must release the reference to batch_last before
 818			 * waiting, on quiesce, otherwise the batch_last will
 819			 * hold a reference to a stripe and raid5_quiesce()
 820			 * will deadlock waiting for active_stripes to go to
 821			 * zero.
 822			 */
 823			if (ctx && ctx->batch_last) {
 824				raid5_release_stripe(ctx->batch_last);
 825				ctx->batch_last = NULL;
 826			}
 827
 828			wait_event_lock_irq(conf->wait_for_quiescent,
 829					    !conf->quiesce,
 830					    *(conf->hash_locks + hash));
 831		}
 832
 833		sh = find_get_stripe(conf, sector, conf->generation - previous,
 834				     hash);
 835		if (sh)
 836			break;
 837
 838		if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 839			sh = get_free_stripe(conf, hash);
 840			if (sh) {
 841				r5c_check_stripe_cache_usage(conf);
 842				init_stripe(sh, sector, previous);
 843				atomic_inc(&sh->count);
 
 
 
 844				break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 845			}
 846
 847			if (!test_bit(R5_DID_ALLOC, &conf->cache_state))
 848				set_bit(R5_ALLOC_MORE, &conf->cache_state);
 849		}
 
 850
 851		if (flags & R5_GAS_NOBLOCK)
 852			break;
 853
 854		set_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 855		r5l_wake_reclaim(conf->log, 0);
 856
 857		/* release batch_last before wait to avoid risk of deadlock */
 858		if (ctx && ctx->batch_last) {
 859			raid5_release_stripe(ctx->batch_last);
 860			ctx->batch_last = NULL;
 861		}
 862
 863		wait_event_lock_irq(conf->wait_for_stripe,
 864				    is_inactive_blocked(conf, hash),
 865				    *(conf->hash_locks + hash));
 866		clear_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 867	}
 868
 869	spin_unlock_irq(conf->hash_locks + hash);
 870	return sh;
 871}
 872
 873static bool is_full_stripe_write(struct stripe_head *sh)
 874{
 875	BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 876	return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 877}
 878
 879static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 880		__acquires(&sh1->stripe_lock)
 881		__acquires(&sh2->stripe_lock)
 882{
 883	if (sh1 > sh2) {
 884		spin_lock_irq(&sh2->stripe_lock);
 885		spin_lock_nested(&sh1->stripe_lock, 1);
 886	} else {
 887		spin_lock_irq(&sh1->stripe_lock);
 888		spin_lock_nested(&sh2->stripe_lock, 1);
 889	}
 890}
 891
 892static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 893		__releases(&sh1->stripe_lock)
 894		__releases(&sh2->stripe_lock)
 895{
 896	spin_unlock(&sh1->stripe_lock);
 897	spin_unlock_irq(&sh2->stripe_lock);
 898}
 899
 900/* Only freshly new full stripe normal write stripe can be added to a batch list */
 901static bool stripe_can_batch(struct stripe_head *sh)
 902{
 903	struct r5conf *conf = sh->raid_conf;
 904
 905	if (raid5_has_log(conf) || raid5_has_ppl(conf))
 906		return false;
 907	return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 908		!test_bit(STRIPE_BITMAP_PENDING, &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			sector_t first_bad;
1214			int bad_sectors;
1215			int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1216					      &first_bad, &bad_sectors);
1217			if (!bad)
1218				break;
1219
1220			if (bad < 0) {
1221				set_bit(BlockedBadBlocks, &rdev->flags);
1222				if (!conf->mddev->external &&
1223				    conf->mddev->sb_flags) {
1224					/* It is very unlikely, but we might
1225					 * still need to write out the
1226					 * bad block log - better give it
1227					 * a chance*/
1228					md_check_recovery(conf->mddev);
1229				}
1230				/*
1231				 * Because md_wait_for_blocked_rdev
1232				 * will dec nr_pending, we must
1233				 * increment it first.
1234				 */
1235				atomic_inc(&rdev->nr_pending);
1236				md_wait_for_blocked_rdev(rdev, conf->mddev);
1237			} else {
1238				/* Acknowledged bad block - skip the write */
1239				rdev_dec_pending(rdev, conf->mddev);
1240				rdev = NULL;
1241			}
1242		}
1243
1244		if (rdev) {
1245			if (s->syncing || s->expanding || s->expanded
1246			    || s->replacing)
1247				md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1248
1249			set_bit(STRIPE_IO_STARTED, &sh->state);
1250
1251			bio_init(bi, rdev->bdev, &dev->vec, 1, op | op_flags);
1252			bi->bi_end_io = op_is_write(op)
1253				? raid5_end_write_request
1254				: raid5_end_read_request;
1255			bi->bi_private = sh;
1256
1257			pr_debug("%s: for %llu schedule op %d on disc %d\n",
1258				__func__, (unsigned long long)sh->sector,
1259				bi->bi_opf, i);
1260			atomic_inc(&sh->count);
1261			if (sh != head_sh)
1262				atomic_inc(&head_sh->count);
1263			if (use_new_offset(conf, sh))
1264				bi->bi_iter.bi_sector = (sh->sector
1265						 + rdev->new_data_offset);
1266			else
1267				bi->bi_iter.bi_sector = (sh->sector
1268						 + rdev->data_offset);
1269			if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1270				bi->bi_opf |= REQ_NOMERGE;
1271
1272			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1273				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1274
1275			if (!op_is_write(op) &&
1276			    test_bit(R5_InJournal, &sh->dev[i].flags))
1277				/*
1278				 * issuing read for a page in journal, this
1279				 * must be preparing for prexor in rmw; read
1280				 * the data into orig_page
1281				 */
1282				sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1283			else
1284				sh->dev[i].vec.bv_page = sh->dev[i].page;
1285			bi->bi_vcnt = 1;
1286			bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1287			bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1288			bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1289			/*
1290			 * If this is discard request, set bi_vcnt 0. We don't
1291			 * want to confuse SCSI because SCSI will replace payload
1292			 */
1293			if (op == REQ_OP_DISCARD)
1294				bi->bi_vcnt = 0;
1295			if (rrdev)
1296				set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1297
1298			if (conf->mddev->gendisk)
1299				trace_block_bio_remap(bi,
1300						disk_devt(conf->mddev->gendisk),
1301						sh->dev[i].sector);
1302			if (should_defer && op_is_write(op))
1303				bio_list_add(&pending_bios, bi);
1304			else
1305				submit_bio_noacct(bi);
1306		}
1307		if (rrdev) {
1308			if (s->syncing || s->expanding || s->expanded
1309			    || s->replacing)
1310				md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1311
1312			set_bit(STRIPE_IO_STARTED, &sh->state);
1313
1314			bio_init(rbi, rrdev->bdev, &dev->rvec, 1, op | op_flags);
1315			BUG_ON(!op_is_write(op));
1316			rbi->bi_end_io = raid5_end_write_request;
1317			rbi->bi_private = sh;
1318
1319			pr_debug("%s: for %llu schedule op %d on "
1320				 "replacement disc %d\n",
1321				__func__, (unsigned long long)sh->sector,
1322				rbi->bi_opf, i);
1323			atomic_inc(&sh->count);
1324			if (sh != head_sh)
1325				atomic_inc(&head_sh->count);
1326			if (use_new_offset(conf, sh))
1327				rbi->bi_iter.bi_sector = (sh->sector
1328						  + rrdev->new_data_offset);
1329			else
1330				rbi->bi_iter.bi_sector = (sh->sector
1331						  + rrdev->data_offset);
1332			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1333				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1334			sh->dev[i].rvec.bv_page = sh->dev[i].page;
1335			rbi->bi_vcnt = 1;
1336			rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1337			rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1338			rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1339			/*
1340			 * If this is discard request, set bi_vcnt 0. We don't
1341			 * want to confuse SCSI because SCSI will replace payload
1342			 */
1343			if (op == REQ_OP_DISCARD)
1344				rbi->bi_vcnt = 0;
1345			if (conf->mddev->gendisk)
1346				trace_block_bio_remap(rbi,
1347						disk_devt(conf->mddev->gendisk),
1348						sh->dev[i].sector);
1349			if (should_defer && op_is_write(op))
1350				bio_list_add(&pending_bios, rbi);
1351			else
1352				submit_bio_noacct(rbi);
1353		}
1354		if (!rdev && !rrdev) {
1355			if (op_is_write(op))
1356				set_bit(STRIPE_DEGRADED, &sh->state);
1357			pr_debug("skip op %d on disc %d for sector %llu\n",
1358				bi->bi_opf, i, (unsigned long long)sh->sector);
1359			clear_bit(R5_LOCKED, &sh->dev[i].flags);
1360			set_bit(STRIPE_HANDLE, &sh->state);
1361		}
1362
1363		if (!head_sh->batch_head)
1364			continue;
1365		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1366				      batch_list);
1367		if (sh != head_sh)
1368			goto again;
1369	}
1370
1371	if (should_defer && !bio_list_empty(&pending_bios))
1372		defer_issue_bios(conf, head_sh->sector, &pending_bios);
1373}
1374
1375static struct dma_async_tx_descriptor *
1376async_copy_data(int frombio, struct bio *bio, struct page **page,
1377	unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1378	struct stripe_head *sh, int no_skipcopy)
1379{
1380	struct bio_vec bvl;
1381	struct bvec_iter iter;
1382	struct page *bio_page;
 
1383	int page_offset;
1384	struct async_submit_ctl submit;
1385	enum async_tx_flags flags = 0;
1386	struct r5conf *conf = sh->raid_conf;
1387
1388	if (bio->bi_iter.bi_sector >= sector)
1389		page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1390	else
1391		page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1392
1393	if (frombio)
1394		flags |= ASYNC_TX_FENCE;
1395	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1396
1397	bio_for_each_segment(bvl, bio, iter) {
1398		int len = bvl.bv_len;
1399		int clen;
1400		int b_offset = 0;
1401
1402		if (page_offset < 0) {
1403			b_offset = -page_offset;
1404			page_offset += b_offset;
1405			len -= b_offset;
1406		}
1407
1408		if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1409			clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1410		else
1411			clen = len;
1412
1413		if (clen > 0) {
1414			b_offset += bvl.bv_offset;
1415			bio_page = bvl.bv_page;
1416			if (frombio) {
1417				if (conf->skip_copy &&
1418				    b_offset == 0 && page_offset == 0 &&
1419				    clen == RAID5_STRIPE_SIZE(conf) &&
1420				    !no_skipcopy)
1421					*page = bio_page;
1422				else
1423					tx = async_memcpy(*page, bio_page, page_offset + poff,
1424						  b_offset, clen, &submit);
1425			} else
1426				tx = async_memcpy(bio_page, *page, b_offset,
1427						  page_offset + poff, clen, &submit);
1428		}
1429		/* chain the operations */
1430		submit.depend_tx = tx;
1431
1432		if (clen < len) /* hit end of page */
1433			break;
1434		page_offset +=  len;
1435	}
1436
1437	return tx;
1438}
1439
1440static void ops_complete_biofill(void *stripe_head_ref)
1441{
1442	struct stripe_head *sh = stripe_head_ref;
 
 
1443	int i;
1444	struct r5conf *conf = sh->raid_conf;
1445
1446	pr_debug("%s: stripe %llu\n", __func__,
1447		(unsigned long long)sh->sector);
1448
1449	/* clear completed biofills */
 
1450	for (i = sh->disks; i--; ) {
1451		struct r5dev *dev = &sh->dev[i];
1452
1453		/* acknowledge completion of a biofill operation */
1454		/* and check if we need to reply to a read request,
1455		 * new R5_Wantfill requests are held off until
1456		 * !STRIPE_BIOFILL_RUN
1457		 */
1458		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1459			struct bio *rbi, *rbi2;
1460
1461			BUG_ON(!dev->read);
1462			rbi = dev->read;
1463			dev->read = NULL;
1464			while (rbi && rbi->bi_iter.bi_sector <
1465				dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1466				rbi2 = r5_next_bio(conf, rbi, dev->sector);
1467				bio_endio(rbi);
 
 
 
1468				rbi = rbi2;
1469			}
1470		}
1471	}
 
1472	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1473
 
 
1474	set_bit(STRIPE_HANDLE, &sh->state);
1475	raid5_release_stripe(sh);
1476}
1477
1478static void ops_run_biofill(struct stripe_head *sh)
1479{
1480	struct dma_async_tx_descriptor *tx = NULL;
 
1481	struct async_submit_ctl submit;
1482	int i;
1483	struct r5conf *conf = sh->raid_conf;
1484
1485	BUG_ON(sh->batch_head);
1486	pr_debug("%s: stripe %llu\n", __func__,
1487		(unsigned long long)sh->sector);
1488
1489	for (i = sh->disks; i--; ) {
1490		struct r5dev *dev = &sh->dev[i];
1491		if (test_bit(R5_Wantfill, &dev->flags)) {
1492			struct bio *rbi;
1493			spin_lock_irq(&sh->stripe_lock);
1494			dev->read = rbi = dev->toread;
1495			dev->toread = NULL;
1496			spin_unlock_irq(&sh->stripe_lock);
1497			while (rbi && rbi->bi_iter.bi_sector <
1498				dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1499				tx = async_copy_data(0, rbi, &dev->page,
1500						     dev->offset,
1501						     dev->sector, tx, sh, 0);
1502				rbi = r5_next_bio(conf, rbi, dev->sector);
1503			}
1504		}
1505	}
1506
1507	atomic_inc(&sh->count);
1508	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1509	async_trigger_callback(&submit);
1510}
1511
1512static void mark_target_uptodate(struct stripe_head *sh, int target)
1513{
1514	struct r5dev *tgt;
1515
1516	if (target < 0)
1517		return;
1518
1519	tgt = &sh->dev[target];
1520	set_bit(R5_UPTODATE, &tgt->flags);
1521	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1522	clear_bit(R5_Wantcompute, &tgt->flags);
1523}
1524
1525static void ops_complete_compute(void *stripe_head_ref)
1526{
1527	struct stripe_head *sh = stripe_head_ref;
1528
1529	pr_debug("%s: stripe %llu\n", __func__,
1530		(unsigned long long)sh->sector);
1531
1532	/* mark the computed target(s) as uptodate */
1533	mark_target_uptodate(sh, sh->ops.target);
1534	mark_target_uptodate(sh, sh->ops.target2);
1535
1536	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1537	if (sh->check_state == check_state_compute_run)
1538		sh->check_state = check_state_compute_result;
1539	set_bit(STRIPE_HANDLE, &sh->state);
1540	raid5_release_stripe(sh);
1541}
1542
1543/* return a pointer to the address conversion region of the scribble buffer */
1544static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1545{
1546	return percpu->scribble + i * percpu->scribble_obj_size;
1547}
1548
1549/* return a pointer to the address conversion region of the scribble buffer */
1550static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1551				 struct raid5_percpu *percpu, int i)
1552{
1553	return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1554}
1555
1556/*
1557 * Return a pointer to record offset address.
1558 */
1559static unsigned int *
1560to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1561{
1562	return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1563}
1564
1565static struct dma_async_tx_descriptor *
1566ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1567{
1568	int disks = sh->disks;
1569	struct page **xor_srcs = to_addr_page(percpu, 0);
1570	unsigned int *off_srcs = to_addr_offs(sh, percpu);
1571	int target = sh->ops.target;
1572	struct r5dev *tgt = &sh->dev[target];
1573	struct page *xor_dest = tgt->page;
1574	unsigned int off_dest = tgt->offset;
1575	int count = 0;
1576	struct dma_async_tx_descriptor *tx;
1577	struct async_submit_ctl submit;
1578	int i;
1579
1580	BUG_ON(sh->batch_head);
1581
1582	pr_debug("%s: stripe %llu block: %d\n",
1583		__func__, (unsigned long long)sh->sector, target);
1584	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1585
1586	for (i = disks; i--; ) {
1587		if (i != target) {
1588			off_srcs[count] = sh->dev[i].offset;
1589			xor_srcs[count++] = sh->dev[i].page;
1590		}
1591	}
1592
1593	atomic_inc(&sh->count);
1594
1595	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1596			  ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1597	if (unlikely(count == 1))
1598		tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1599				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1600	else
1601		tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1602				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1603
1604	return tx;
1605}
1606
1607/* set_syndrome_sources - populate source buffers for gen_syndrome
1608 * @srcs - (struct page *) array of size sh->disks
1609 * @offs - (unsigned int) array of offset for each page
1610 * @sh - stripe_head to parse
1611 *
1612 * Populates srcs in proper layout order for the stripe and returns the
1613 * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1614 * destination buffer is recorded in srcs[count] and the Q destination
1615 * is recorded in srcs[count+1]].
1616 */
1617static int set_syndrome_sources(struct page **srcs,
1618				unsigned int *offs,
1619				struct stripe_head *sh,
1620				int srctype)
1621{
1622	int disks = sh->disks;
1623	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1624	int d0_idx = raid6_d0(sh);
1625	int count;
1626	int i;
1627
1628	for (i = 0; i < disks; i++)
1629		srcs[i] = NULL;
1630
1631	count = 0;
1632	i = d0_idx;
1633	do {
1634		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1635		struct r5dev *dev = &sh->dev[i];
1636
1637		if (i == sh->qd_idx || i == sh->pd_idx ||
1638		    (srctype == SYNDROME_SRC_ALL) ||
1639		    (srctype == SYNDROME_SRC_WANT_DRAIN &&
1640		     (test_bit(R5_Wantdrain, &dev->flags) ||
1641		      test_bit(R5_InJournal, &dev->flags))) ||
1642		    (srctype == SYNDROME_SRC_WRITTEN &&
1643		     (dev->written ||
1644		      test_bit(R5_InJournal, &dev->flags)))) {
1645			if (test_bit(R5_InJournal, &dev->flags))
1646				srcs[slot] = sh->dev[i].orig_page;
1647			else
1648				srcs[slot] = sh->dev[i].page;
1649			/*
1650			 * For R5_InJournal, PAGE_SIZE must be 4KB and will
1651			 * not shared page. In that case, dev[i].offset
1652			 * is 0.
1653			 */
1654			offs[slot] = sh->dev[i].offset;
1655		}
1656		i = raid6_next_disk(i, disks);
1657	} while (i != d0_idx);
1658
1659	return syndrome_disks;
1660}
1661
1662static struct dma_async_tx_descriptor *
1663ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1664{
1665	int disks = sh->disks;
1666	struct page **blocks = to_addr_page(percpu, 0);
1667	unsigned int *offs = to_addr_offs(sh, percpu);
1668	int target;
1669	int qd_idx = sh->qd_idx;
1670	struct dma_async_tx_descriptor *tx;
1671	struct async_submit_ctl submit;
1672	struct r5dev *tgt;
1673	struct page *dest;
1674	unsigned int dest_off;
1675	int i;
1676	int count;
1677
1678	BUG_ON(sh->batch_head);
1679	if (sh->ops.target < 0)
1680		target = sh->ops.target2;
1681	else if (sh->ops.target2 < 0)
1682		target = sh->ops.target;
1683	else
1684		/* we should only have one valid target */
1685		BUG();
1686	BUG_ON(target < 0);
1687	pr_debug("%s: stripe %llu block: %d\n",
1688		__func__, (unsigned long long)sh->sector, target);
1689
1690	tgt = &sh->dev[target];
1691	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1692	dest = tgt->page;
1693	dest_off = tgt->offset;
1694
1695	atomic_inc(&sh->count);
1696
1697	if (target == qd_idx) {
1698		count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1699		blocks[count] = NULL; /* regenerating p is not necessary */
1700		BUG_ON(blocks[count+1] != dest); /* q should already be set */
1701		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1702				  ops_complete_compute, sh,
1703				  to_addr_conv(sh, percpu, 0));
1704		tx = async_gen_syndrome(blocks, offs, count+2,
1705				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1706	} else {
1707		/* Compute any data- or p-drive using XOR */
1708		count = 0;
1709		for (i = disks; i-- ; ) {
1710			if (i == target || i == qd_idx)
1711				continue;
1712			offs[count] = sh->dev[i].offset;
1713			blocks[count++] = sh->dev[i].page;
1714		}
1715
1716		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1717				  NULL, ops_complete_compute, sh,
1718				  to_addr_conv(sh, percpu, 0));
1719		tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1720				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1721	}
1722
1723	return tx;
1724}
1725
1726static struct dma_async_tx_descriptor *
1727ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1728{
1729	int i, count, disks = sh->disks;
1730	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1731	int d0_idx = raid6_d0(sh);
1732	int faila = -1, failb = -1;
1733	int target = sh->ops.target;
1734	int target2 = sh->ops.target2;
1735	struct r5dev *tgt = &sh->dev[target];
1736	struct r5dev *tgt2 = &sh->dev[target2];
1737	struct dma_async_tx_descriptor *tx;
1738	struct page **blocks = to_addr_page(percpu, 0);
1739	unsigned int *offs = to_addr_offs(sh, percpu);
1740	struct async_submit_ctl submit;
1741
1742	BUG_ON(sh->batch_head);
1743	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1744		 __func__, (unsigned long long)sh->sector, target, target2);
1745	BUG_ON(target < 0 || target2 < 0);
1746	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1747	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1748
1749	/* we need to open-code set_syndrome_sources to handle the
1750	 * slot number conversion for 'faila' and 'failb'
1751	 */
1752	for (i = 0; i < disks ; i++) {
1753		offs[i] = 0;
1754		blocks[i] = NULL;
1755	}
1756	count = 0;
1757	i = d0_idx;
1758	do {
1759		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1760
1761		offs[slot] = sh->dev[i].offset;
1762		blocks[slot] = sh->dev[i].page;
1763
1764		if (i == target)
1765			faila = slot;
1766		if (i == target2)
1767			failb = slot;
1768		i = raid6_next_disk(i, disks);
1769	} while (i != d0_idx);
1770
1771	BUG_ON(faila == failb);
1772	if (failb < faila)
1773		swap(faila, failb);
1774	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1775		 __func__, (unsigned long long)sh->sector, faila, failb);
1776
1777	atomic_inc(&sh->count);
1778
1779	if (failb == syndrome_disks+1) {
1780		/* Q disk is one of the missing disks */
1781		if (faila == syndrome_disks) {
1782			/* Missing P+Q, just recompute */
1783			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1784					  ops_complete_compute, sh,
1785					  to_addr_conv(sh, percpu, 0));
1786			return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1787						  RAID5_STRIPE_SIZE(sh->raid_conf),
1788						  &submit);
1789		} else {
1790			struct page *dest;
1791			unsigned int dest_off;
1792			int data_target;
1793			int qd_idx = sh->qd_idx;
1794
1795			/* Missing D+Q: recompute D from P, then recompute Q */
1796			if (target == qd_idx)
1797				data_target = target2;
1798			else
1799				data_target = target;
1800
1801			count = 0;
1802			for (i = disks; i-- ; ) {
1803				if (i == data_target || i == qd_idx)
1804					continue;
1805				offs[count] = sh->dev[i].offset;
1806				blocks[count++] = sh->dev[i].page;
1807			}
1808			dest = sh->dev[data_target].page;
1809			dest_off = sh->dev[data_target].offset;
1810			init_async_submit(&submit,
1811					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1812					  NULL, NULL, NULL,
1813					  to_addr_conv(sh, percpu, 0));
1814			tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1815				       RAID5_STRIPE_SIZE(sh->raid_conf),
1816				       &submit);
1817
1818			count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1819			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1820					  ops_complete_compute, sh,
1821					  to_addr_conv(sh, percpu, 0));
1822			return async_gen_syndrome(blocks, offs, count+2,
1823						  RAID5_STRIPE_SIZE(sh->raid_conf),
1824						  &submit);
1825		}
1826	} else {
1827		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1828				  ops_complete_compute, sh,
1829				  to_addr_conv(sh, percpu, 0));
1830		if (failb == syndrome_disks) {
1831			/* We're missing D+P. */
1832			return async_raid6_datap_recov(syndrome_disks+2,
1833						RAID5_STRIPE_SIZE(sh->raid_conf),
1834						faila,
1835						blocks, offs, &submit);
1836		} else {
1837			/* We're missing D+D. */
1838			return async_raid6_2data_recov(syndrome_disks+2,
1839						RAID5_STRIPE_SIZE(sh->raid_conf),
1840						faila, failb,
1841						blocks, offs, &submit);
1842		}
1843	}
1844}
1845
 
1846static void ops_complete_prexor(void *stripe_head_ref)
1847{
1848	struct stripe_head *sh = stripe_head_ref;
1849
1850	pr_debug("%s: stripe %llu\n", __func__,
1851		(unsigned long long)sh->sector);
1852
1853	if (r5c_is_writeback(sh->raid_conf->log))
1854		/*
1855		 * raid5-cache write back uses orig_page during prexor.
1856		 * After prexor, it is time to free orig_page
1857		 */
1858		r5c_release_extra_page(sh);
1859}
1860
1861static struct dma_async_tx_descriptor *
1862ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1863		struct dma_async_tx_descriptor *tx)
1864{
1865	int disks = sh->disks;
1866	struct page **xor_srcs = to_addr_page(percpu, 0);
1867	unsigned int *off_srcs = to_addr_offs(sh, percpu);
1868	int count = 0, pd_idx = sh->pd_idx, i;
1869	struct async_submit_ctl submit;
1870
1871	/* existing parity data subtracted */
1872	unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1873	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1874
1875	BUG_ON(sh->batch_head);
1876	pr_debug("%s: stripe %llu\n", __func__,
1877		(unsigned long long)sh->sector);
1878
1879	for (i = disks; i--; ) {
1880		struct r5dev *dev = &sh->dev[i];
1881		/* Only process blocks that are known to be uptodate */
1882		if (test_bit(R5_InJournal, &dev->flags)) {
1883			/*
1884			 * For this case, PAGE_SIZE must be equal to 4KB and
1885			 * page offset is zero.
1886			 */
1887			off_srcs[count] = dev->offset;
1888			xor_srcs[count++] = dev->orig_page;
1889		} else if (test_bit(R5_Wantdrain, &dev->flags)) {
1890			off_srcs[count] = dev->offset;
1891			xor_srcs[count++] = dev->page;
1892		}
1893	}
1894
1895	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1896			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1897	tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1898			RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1899
1900	return tx;
1901}
1902
1903static struct dma_async_tx_descriptor *
1904ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1905		struct dma_async_tx_descriptor *tx)
1906{
1907	struct page **blocks = to_addr_page(percpu, 0);
1908	unsigned int *offs = to_addr_offs(sh, percpu);
1909	int count;
1910	struct async_submit_ctl submit;
1911
1912	pr_debug("%s: stripe %llu\n", __func__,
1913		(unsigned long long)sh->sector);
1914
1915	count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1916
1917	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1918			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1919	tx = async_gen_syndrome(blocks, offs, count+2,
1920			RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1921
1922	return tx;
1923}
1924
1925static struct dma_async_tx_descriptor *
1926ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1927{
1928	struct r5conf *conf = sh->raid_conf;
1929	int disks = sh->disks;
1930	int i;
1931	struct stripe_head *head_sh = sh;
1932
1933	pr_debug("%s: stripe %llu\n", __func__,
1934		(unsigned long long)sh->sector);
1935
1936	for (i = disks; i--; ) {
1937		struct r5dev *dev;
1938		struct bio *chosen;
1939
1940		sh = head_sh;
1941		if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1942			struct bio *wbi;
1943
1944again:
1945			dev = &sh->dev[i];
1946			/*
1947			 * clear R5_InJournal, so when rewriting a page in
1948			 * journal, it is not skipped by r5l_log_stripe()
1949			 */
1950			clear_bit(R5_InJournal, &dev->flags);
1951			spin_lock_irq(&sh->stripe_lock);
1952			chosen = dev->towrite;
1953			dev->towrite = NULL;
1954			sh->overwrite_disks = 0;
1955			BUG_ON(dev->written);
1956			wbi = dev->written = chosen;
1957			spin_unlock_irq(&sh->stripe_lock);
1958			WARN_ON(dev->page != dev->orig_page);
1959
1960			while (wbi && wbi->bi_iter.bi_sector <
1961				dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1962				if (wbi->bi_opf & REQ_FUA)
1963					set_bit(R5_WantFUA, &dev->flags);
1964				if (wbi->bi_opf & REQ_SYNC)
1965					set_bit(R5_SyncIO, &dev->flags);
1966				if (bio_op(wbi) == REQ_OP_DISCARD)
1967					set_bit(R5_Discard, &dev->flags);
1968				else {
1969					tx = async_copy_data(1, wbi, &dev->page,
1970							     dev->offset,
1971							     dev->sector, tx, sh,
1972							     r5c_is_writeback(conf->log));
1973					if (dev->page != dev->orig_page &&
1974					    !r5c_is_writeback(conf->log)) {
1975						set_bit(R5_SkipCopy, &dev->flags);
1976						clear_bit(R5_UPTODATE, &dev->flags);
1977						clear_bit(R5_OVERWRITE, &dev->flags);
1978					}
1979				}
1980				wbi = r5_next_bio(conf, wbi, dev->sector);
1981			}
1982
1983			if (head_sh->batch_head) {
1984				sh = list_first_entry(&sh->batch_list,
1985						      struct stripe_head,
1986						      batch_list);
1987				if (sh == head_sh)
1988					continue;
1989				goto again;
1990			}
1991		}
1992	}
1993
1994	return tx;
1995}
1996
1997static void ops_complete_reconstruct(void *stripe_head_ref)
1998{
1999	struct stripe_head *sh = stripe_head_ref;
2000	int disks = sh->disks;
2001	int pd_idx = sh->pd_idx;
2002	int qd_idx = sh->qd_idx;
2003	int i;
2004	bool fua = false, sync = false, discard = false;
2005
2006	pr_debug("%s: stripe %llu\n", __func__,
2007		(unsigned long long)sh->sector);
2008
2009	for (i = disks; i--; ) {
2010		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
2011		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
2012		discard |= test_bit(R5_Discard, &sh->dev[i].flags);
2013	}
2014
2015	for (i = disks; i--; ) {
2016		struct r5dev *dev = &sh->dev[i];
2017
2018		if (dev->written || i == pd_idx || i == qd_idx) {
2019			if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
2020				set_bit(R5_UPTODATE, &dev->flags);
2021				if (test_bit(STRIPE_EXPAND_READY, &sh->state))
2022					set_bit(R5_Expanded, &dev->flags);
2023			}
2024			if (fua)
2025				set_bit(R5_WantFUA, &dev->flags);
2026			if (sync)
2027				set_bit(R5_SyncIO, &dev->flags);
2028		}
2029	}
2030
2031	if (sh->reconstruct_state == reconstruct_state_drain_run)
2032		sh->reconstruct_state = reconstruct_state_drain_result;
2033	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
2034		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
2035	else {
2036		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
2037		sh->reconstruct_state = reconstruct_state_result;
2038	}
2039
2040	set_bit(STRIPE_HANDLE, &sh->state);
2041	raid5_release_stripe(sh);
2042}
2043
2044static void
2045ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
2046		     struct dma_async_tx_descriptor *tx)
2047{
2048	int disks = sh->disks;
2049	struct page **xor_srcs;
2050	unsigned int *off_srcs;
2051	struct async_submit_ctl submit;
2052	int count, pd_idx = sh->pd_idx, i;
2053	struct page *xor_dest;
2054	unsigned int off_dest;
2055	int prexor = 0;
2056	unsigned long flags;
2057	int j = 0;
2058	struct stripe_head *head_sh = sh;
2059	int last_stripe;
2060
2061	pr_debug("%s: stripe %llu\n", __func__,
2062		(unsigned long long)sh->sector);
2063
2064	for (i = 0; i < sh->disks; i++) {
2065		if (pd_idx == i)
2066			continue;
2067		if (!test_bit(R5_Discard, &sh->dev[i].flags))
2068			break;
2069	}
2070	if (i >= sh->disks) {
2071		atomic_inc(&sh->count);
2072		set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2073		ops_complete_reconstruct(sh);
2074		return;
2075	}
2076again:
2077	count = 0;
2078	xor_srcs = to_addr_page(percpu, j);
2079	off_srcs = to_addr_offs(sh, percpu);
2080	/* check if prexor is active which means only process blocks
2081	 * that are part of a read-modify-write (written)
2082	 */
2083	if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2084		prexor = 1;
2085		off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2086		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2087		for (i = disks; i--; ) {
2088			struct r5dev *dev = &sh->dev[i];
2089			if (head_sh->dev[i].written ||
2090			    test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2091				off_srcs[count] = dev->offset;
2092				xor_srcs[count++] = dev->page;
2093			}
2094		}
2095	} else {
2096		xor_dest = sh->dev[pd_idx].page;
2097		off_dest = sh->dev[pd_idx].offset;
2098		for (i = disks; i--; ) {
2099			struct r5dev *dev = &sh->dev[i];
2100			if (i != pd_idx) {
2101				off_srcs[count] = dev->offset;
2102				xor_srcs[count++] = dev->page;
2103			}
2104		}
2105	}
2106
2107	/* 1/ if we prexor'd then the dest is reused as a source
2108	 * 2/ if we did not prexor then we are redoing the parity
2109	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2110	 * for the synchronous xor case
2111	 */
2112	last_stripe = !head_sh->batch_head ||
2113		list_first_entry(&sh->batch_list,
2114				 struct stripe_head, batch_list) == head_sh;
2115	if (last_stripe) {
2116		flags = ASYNC_TX_ACK |
2117			(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2118
2119		atomic_inc(&head_sh->count);
2120		init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2121				  to_addr_conv(sh, percpu, j));
2122	} else {
2123		flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2124		init_async_submit(&submit, flags, tx, NULL, NULL,
2125				  to_addr_conv(sh, percpu, j));
2126	}
2127
 
 
 
 
2128	if (unlikely(count == 1))
2129		tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2130				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2131	else
2132		tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2133				RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2134	if (!last_stripe) {
2135		j++;
2136		sh = list_first_entry(&sh->batch_list, struct stripe_head,
2137				      batch_list);
2138		goto again;
2139	}
2140}
2141
2142static void
2143ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2144		     struct dma_async_tx_descriptor *tx)
2145{
2146	struct async_submit_ctl submit;
2147	struct page **blocks;
2148	unsigned int *offs;
2149	int count, i, j = 0;
2150	struct stripe_head *head_sh = sh;
2151	int last_stripe;
2152	int synflags;
2153	unsigned long txflags;
2154
2155	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2156
2157	for (i = 0; i < sh->disks; i++) {
2158		if (sh->pd_idx == i || sh->qd_idx == i)
2159			continue;
2160		if (!test_bit(R5_Discard, &sh->dev[i].flags))
2161			break;
2162	}
2163	if (i >= sh->disks) {
2164		atomic_inc(&sh->count);
2165		set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2166		set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2167		ops_complete_reconstruct(sh);
2168		return;
2169	}
2170
2171again:
2172	blocks = to_addr_page(percpu, j);
2173	offs = to_addr_offs(sh, percpu);
2174
2175	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2176		synflags = SYNDROME_SRC_WRITTEN;
2177		txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2178	} else {
2179		synflags = SYNDROME_SRC_ALL;
2180		txflags = ASYNC_TX_ACK;
2181	}
2182
2183	count = set_syndrome_sources(blocks, offs, sh, synflags);
2184	last_stripe = !head_sh->batch_head ||
2185		list_first_entry(&sh->batch_list,
2186				 struct stripe_head, batch_list) == head_sh;
2187
2188	if (last_stripe) {
2189		atomic_inc(&head_sh->count);
2190		init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2191				  head_sh, to_addr_conv(sh, percpu, j));
2192	} else
2193		init_async_submit(&submit, 0, tx, NULL, NULL,
2194				  to_addr_conv(sh, percpu, j));
2195	tx = async_gen_syndrome(blocks, offs, count+2,
2196			RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2197	if (!last_stripe) {
2198		j++;
2199		sh = list_first_entry(&sh->batch_list, struct stripe_head,
2200				      batch_list);
2201		goto again;
2202	}
2203}
2204
2205static void ops_complete_check(void *stripe_head_ref)
2206{
2207	struct stripe_head *sh = stripe_head_ref;
2208
2209	pr_debug("%s: stripe %llu\n", __func__,
2210		(unsigned long long)sh->sector);
2211
2212	sh->check_state = check_state_check_result;
2213	set_bit(STRIPE_HANDLE, &sh->state);
2214	raid5_release_stripe(sh);
2215}
2216
2217static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2218{
2219	int disks = sh->disks;
2220	int pd_idx = sh->pd_idx;
2221	int qd_idx = sh->qd_idx;
2222	struct page *xor_dest;
2223	unsigned int off_dest;
2224	struct page **xor_srcs = to_addr_page(percpu, 0);
2225	unsigned int *off_srcs = to_addr_offs(sh, percpu);
2226	struct dma_async_tx_descriptor *tx;
2227	struct async_submit_ctl submit;
2228	int count;
2229	int i;
2230
2231	pr_debug("%s: stripe %llu\n", __func__,
2232		(unsigned long long)sh->sector);
2233
2234	BUG_ON(sh->batch_head);
2235	count = 0;
2236	xor_dest = sh->dev[pd_idx].page;
2237	off_dest = sh->dev[pd_idx].offset;
2238	off_srcs[count] = off_dest;
2239	xor_srcs[count++] = xor_dest;
2240	for (i = disks; i--; ) {
2241		if (i == pd_idx || i == qd_idx)
2242			continue;
2243		off_srcs[count] = sh->dev[i].offset;
2244		xor_srcs[count++] = sh->dev[i].page;
2245	}
2246
2247	init_async_submit(&submit, 0, NULL, NULL, NULL,
2248			  to_addr_conv(sh, percpu, 0));
2249	tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2250			   RAID5_STRIPE_SIZE(sh->raid_conf),
2251			   &sh->ops.zero_sum_result, &submit);
2252
2253	atomic_inc(&sh->count);
2254	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2255	tx = async_trigger_callback(&submit);
2256}
2257
2258static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2259{
2260	struct page **srcs = to_addr_page(percpu, 0);
2261	unsigned int *offs = to_addr_offs(sh, percpu);
2262	struct async_submit_ctl submit;
2263	int count;
2264
2265	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2266		(unsigned long long)sh->sector, checkp);
2267
2268	BUG_ON(sh->batch_head);
2269	count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2270	if (!checkp)
2271		srcs[count] = NULL;
2272
2273	atomic_inc(&sh->count);
2274	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2275			  sh, to_addr_conv(sh, percpu, 0));
2276	async_syndrome_val(srcs, offs, count+2,
2277			   RAID5_STRIPE_SIZE(sh->raid_conf),
2278			   &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2279}
2280
2281static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2282{
2283	int overlap_clear = 0, i, disks = sh->disks;
2284	struct dma_async_tx_descriptor *tx = NULL;
2285	struct r5conf *conf = sh->raid_conf;
2286	int level = conf->level;
2287	struct raid5_percpu *percpu;
 
2288
2289	local_lock(&conf->percpu->lock);
2290	percpu = this_cpu_ptr(conf->percpu);
2291	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2292		ops_run_biofill(sh);
2293		overlap_clear++;
2294	}
2295
2296	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2297		if (level < 6)
2298			tx = ops_run_compute5(sh, percpu);
2299		else {
2300			if (sh->ops.target2 < 0 || sh->ops.target < 0)
2301				tx = ops_run_compute6_1(sh, percpu);
2302			else
2303				tx = ops_run_compute6_2(sh, percpu);
2304		}
2305		/* terminate the chain if reconstruct is not set to be run */
2306		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2307			async_tx_ack(tx);
2308	}
2309
2310	if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2311		if (level < 6)
2312			tx = ops_run_prexor5(sh, percpu, tx);
2313		else
2314			tx = ops_run_prexor6(sh, percpu, tx);
2315	}
2316
2317	if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2318		tx = ops_run_partial_parity(sh, percpu, tx);
2319
2320	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2321		tx = ops_run_biodrain(sh, tx);
2322		overlap_clear++;
2323	}
2324
2325	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2326		if (level < 6)
2327			ops_run_reconstruct5(sh, percpu, tx);
2328		else
2329			ops_run_reconstruct6(sh, percpu, tx);
2330	}
2331
2332	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2333		if (sh->check_state == check_state_run)
2334			ops_run_check_p(sh, percpu);
2335		else if (sh->check_state == check_state_run_q)
2336			ops_run_check_pq(sh, percpu, 0);
2337		else if (sh->check_state == check_state_run_pq)
2338			ops_run_check_pq(sh, percpu, 1);
2339		else
2340			BUG();
2341	}
2342
2343	if (overlap_clear && !sh->batch_head) {
2344		for (i = disks; i--; ) {
2345			struct r5dev *dev = &sh->dev[i];
2346			if (test_and_clear_bit(R5_Overlap, &dev->flags))
2347				wake_up(&sh->raid_conf->wait_for_overlap);
2348		}
2349	}
2350	local_unlock(&conf->percpu->lock);
2351}
2352
2353static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
 
2354{
2355#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2356	kfree(sh->pages);
2357#endif
2358	if (sh->ppl_page)
2359		__free_page(sh->ppl_page);
2360	kmem_cache_free(sc, sh);
 
 
2361}
2362
2363static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2364	int disks, struct r5conf *conf)
2365{
2366	struct stripe_head *sh;
 
 
 
 
 
 
2367
2368	sh = kmem_cache_zalloc(sc, gfp);
2369	if (sh) {
2370		spin_lock_init(&sh->stripe_lock);
2371		spin_lock_init(&sh->batch_lock);
2372		INIT_LIST_HEAD(&sh->batch_list);
2373		INIT_LIST_HEAD(&sh->lru);
2374		INIT_LIST_HEAD(&sh->r5c);
2375		INIT_LIST_HEAD(&sh->log_list);
2376		atomic_set(&sh->count, 1);
2377		sh->raid_conf = conf;
2378		sh->log_start = MaxSector;
2379
2380		if (raid5_has_ppl(conf)) {
2381			sh->ppl_page = alloc_page(gfp);
2382			if (!sh->ppl_page) {
2383				free_stripe(sc, sh);
2384				return NULL;
2385			}
2386		}
2387#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2388		if (init_stripe_shared_pages(sh, conf, disks)) {
2389			free_stripe(sc, sh);
2390			return NULL;
2391		}
2392#endif
2393	}
2394	return sh;
2395}
2396static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
 
 
 
 
2397{
2398	struct stripe_head *sh;
2399
2400	sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2401	if (!sh)
2402		return 0;
2403
2404	if (grow_buffers(sh, gfp)) {
 
 
 
 
 
2405		shrink_buffers(sh);
2406		free_stripe(conf->slab_cache, sh);
2407		return 0;
2408	}
2409	sh->hash_lock_index =
2410		conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2411	/* we just created an active stripe so... */
 
2412	atomic_inc(&conf->active_stripes);
2413
2414	raid5_release_stripe(sh);
2415	conf->max_nr_stripes++;
2416	return 1;
2417}
2418
2419static int grow_stripes(struct r5conf *conf, int num)
2420{
2421	struct kmem_cache *sc;
2422	size_t namelen = sizeof(conf->cache_name[0]);
2423	int devs = max(conf->raid_disks, conf->previous_raid_disks);
2424
2425	if (conf->mddev->gendisk)
2426		snprintf(conf->cache_name[0], namelen,
2427			"raid%d-%s", conf->level, mdname(conf->mddev));
2428	else
2429		snprintf(conf->cache_name[0], namelen,
2430			"raid%d-%p", conf->level, conf->mddev);
2431	snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2432
2433	conf->active_name = 0;
2434	sc = kmem_cache_create(conf->cache_name[conf->active_name],
2435			       struct_size_t(struct stripe_head, dev, devs),
2436			       0, 0, NULL);
2437	if (!sc)
2438		return 1;
2439	conf->slab_cache = sc;
2440	conf->pool_size = devs;
2441	while (num--)
2442		if (!grow_one_stripe(conf, GFP_KERNEL))
2443			return 1;
2444
2445	return 0;
2446}
2447
2448/**
2449 * scribble_alloc - allocate percpu scribble buffer for required size
2450 *		    of the scribble region
2451 * @percpu: from for_each_present_cpu() of the caller
2452 * @num: total number of disks in the array
2453 * @cnt: scribble objs count for required size of the scribble region
2454 *
2455 * The scribble buffer size must be enough to contain:
2456 * 1/ a struct page pointer for each device in the array +2
2457 * 2/ room to convert each entry in (1) to its corresponding dma
2458 *    (dma_map_page()) or page (page_address()) address.
2459 *
2460 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2461 * calculate over all devices (not just the data blocks), using zeros in place
2462 * of the P and Q blocks.
2463 */
2464static int scribble_alloc(struct raid5_percpu *percpu,
2465			  int num, int cnt)
2466{
2467	size_t obj_size =
2468		sizeof(struct page *) * (num + 2) +
2469		sizeof(addr_conv_t) * (num + 2) +
2470		sizeof(unsigned int) * (num + 2);
2471	void *scribble;
2472
2473	/*
2474	 * If here is in raid array suspend context, it is in memalloc noio
2475	 * context as well, there is no potential recursive memory reclaim
2476	 * I/Os with the GFP_KERNEL flag.
2477	 */
2478	scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2479	if (!scribble)
2480		return -ENOMEM;
2481
2482	kvfree(percpu->scribble);
2483
2484	percpu->scribble = scribble;
2485	percpu->scribble_obj_size = obj_size;
2486	return 0;
2487}
2488
2489static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2490{
2491	unsigned long cpu;
2492	int err = 0;
2493
2494	/* Never shrink. */
2495	if (conf->scribble_disks >= new_disks &&
2496	    conf->scribble_sectors >= new_sectors)
2497		return 0;
2498
2499	raid5_quiesce(conf->mddev, true);
2500	cpus_read_lock();
2501
2502	for_each_present_cpu(cpu) {
2503		struct raid5_percpu *percpu;
2504
2505		percpu = per_cpu_ptr(conf->percpu, cpu);
2506		err = scribble_alloc(percpu, new_disks,
2507				     new_sectors / RAID5_STRIPE_SECTORS(conf));
2508		if (err)
2509			break;
2510	}
2511
2512	cpus_read_unlock();
2513	raid5_quiesce(conf->mddev, false);
2514
2515	if (!err) {
2516		conf->scribble_disks = new_disks;
2517		conf->scribble_sectors = new_sectors;
2518	}
2519	return err;
2520}
2521
2522static int resize_stripes(struct r5conf *conf, int newsize)
2523{
2524	/* Make all the stripes able to hold 'newsize' devices.
2525	 * New slots in each stripe get 'page' set to a new page.
2526	 *
2527	 * This happens in stages:
2528	 * 1/ create a new kmem_cache and allocate the required number of
2529	 *    stripe_heads.
2530	 * 2/ gather all the old stripe_heads and transfer the pages across
2531	 *    to the new stripe_heads.  This will have the side effect of
2532	 *    freezing the array as once all stripe_heads have been collected,
2533	 *    no IO will be possible.  Old stripe heads are freed once their
2534	 *    pages have been transferred over, and the old kmem_cache is
2535	 *    freed when all stripes are done.
2536	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2537	 *    we simple return a failure status - no need to clean anything up.
2538	 * 4/ allocate new pages for the new slots in the new stripe_heads.
2539	 *    If this fails, we don't bother trying the shrink the
2540	 *    stripe_heads down again, we just leave them as they are.
2541	 *    As each stripe_head is processed the new one is released into
2542	 *    active service.
2543	 *
2544	 * Once step2 is started, we cannot afford to wait for a write,
2545	 * so we use GFP_NOIO allocations.
2546	 */
2547	struct stripe_head *osh, *nsh;
2548	LIST_HEAD(newstripes);
2549	struct disk_info *ndisks;
2550	int err = 0;
 
2551	struct kmem_cache *sc;
2552	int i;
2553	int hash, cnt;
2554
2555	md_allow_write(conf->mddev);
 
 
 
 
 
2556
2557	/* Step 1 */
2558	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2559			       struct_size_t(struct stripe_head, dev, newsize),
2560			       0, 0, NULL);
2561	if (!sc)
2562		return -ENOMEM;
2563
2564	/* Need to ensure auto-resizing doesn't interfere */
2565	mutex_lock(&conf->cache_size_mutex);
2566
2567	for (i = conf->max_nr_stripes; i; i--) {
2568		nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2569		if (!nsh)
2570			break;
2571
 
 
 
 
 
2572		list_add(&nsh->lru, &newstripes);
2573	}
2574	if (i) {
2575		/* didn't get enough, give up */
2576		while (!list_empty(&newstripes)) {
2577			nsh = list_entry(newstripes.next, struct stripe_head, lru);
2578			list_del(&nsh->lru);
2579			free_stripe(sc, nsh);
2580		}
2581		kmem_cache_destroy(sc);
2582		mutex_unlock(&conf->cache_size_mutex);
2583		return -ENOMEM;
2584	}
2585	/* Step 2 - Must use GFP_NOIO now.
2586	 * OK, we have enough stripes, start collecting inactive
2587	 * stripes and copying them over
2588	 */
2589	hash = 0;
2590	cnt = 0;
2591	list_for_each_entry(nsh, &newstripes, lru) {
2592		lock_device_hash_lock(conf, hash);
2593		wait_event_cmd(conf->wait_for_stripe,
2594				    !list_empty(conf->inactive_list + hash),
2595				    unlock_device_hash_lock(conf, hash),
2596				    lock_device_hash_lock(conf, hash));
2597		osh = get_free_stripe(conf, hash);
2598		unlock_device_hash_lock(conf, hash);
2599
2600#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2601	for (i = 0; i < osh->nr_pages; i++) {
2602		nsh->pages[i] = osh->pages[i];
2603		osh->pages[i] = NULL;
2604	}
2605#endif
2606		for(i=0; i<conf->pool_size; i++) {
2607			nsh->dev[i].page = osh->dev[i].page;
2608			nsh->dev[i].orig_page = osh->dev[i].page;
2609			nsh->dev[i].offset = osh->dev[i].offset;
2610		}
2611		nsh->hash_lock_index = hash;
2612		free_stripe(conf->slab_cache, osh);
2613		cnt++;
2614		if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2615		    !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2616			hash++;
2617			cnt = 0;
2618		}
2619	}
2620	kmem_cache_destroy(conf->slab_cache);
2621
2622	/* Step 3.
2623	 * At this point, we are holding all the stripes so the array
2624	 * is completely stalled, so now is a good time to resize
2625	 * conf->disks and the scribble region
2626	 */
2627	ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2628	if (ndisks) {
2629		for (i = 0; i < conf->pool_size; i++)
2630			ndisks[i] = conf->disks[i];
 
 
 
 
2631
2632		for (i = conf->pool_size; i < newsize; i++) {
2633			ndisks[i].extra_page = alloc_page(GFP_NOIO);
2634			if (!ndisks[i].extra_page)
2635				err = -ENOMEM;
2636		}
 
 
 
2637
2638		if (err) {
2639			for (i = conf->pool_size; i < newsize; i++)
2640				if (ndisks[i].extra_page)
2641					put_page(ndisks[i].extra_page);
2642			kfree(ndisks);
2643		} else {
2644			kfree(conf->disks);
2645			conf->disks = ndisks;
2646		}
2647	} else
2648		err = -ENOMEM;
2649
2650	conf->slab_cache = sc;
2651	conf->active_name = 1-conf->active_name;
2652
2653	/* Step 4, return new stripes to service */
2654	while(!list_empty(&newstripes)) {
2655		nsh = list_entry(newstripes.next, struct stripe_head, lru);
2656		list_del_init(&nsh->lru);
2657
2658#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2659		for (i = 0; i < nsh->nr_pages; i++) {
2660			if (nsh->pages[i])
2661				continue;
2662			nsh->pages[i] = alloc_page(GFP_NOIO);
2663			if (!nsh->pages[i])
2664				err = -ENOMEM;
2665		}
2666
2667		for (i = conf->raid_disks; i < newsize; i++) {
2668			if (nsh->dev[i].page)
2669				continue;
2670			nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2671			nsh->dev[i].orig_page = nsh->dev[i].page;
2672			nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2673		}
2674#else
2675		for (i=conf->raid_disks; i < newsize; i++)
2676			if (nsh->dev[i].page == NULL) {
2677				struct page *p = alloc_page(GFP_NOIO);
2678				nsh->dev[i].page = p;
2679				nsh->dev[i].orig_page = p;
2680				nsh->dev[i].offset = 0;
2681				if (!p)
2682					err = -ENOMEM;
2683			}
2684#endif
2685		raid5_release_stripe(nsh);
2686	}
2687	/* critical section pass, GFP_NOIO no longer needed */
2688
2689	if (!err)
2690		conf->pool_size = newsize;
2691	mutex_unlock(&conf->cache_size_mutex);
2692
2693	return err;
2694}
2695
2696static int drop_one_stripe(struct r5conf *conf)
2697{
2698	struct stripe_head *sh;
2699	int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2700
2701	spin_lock_irq(conf->hash_locks + hash);
2702	sh = get_free_stripe(conf, hash);
2703	spin_unlock_irq(conf->hash_locks + hash);
2704	if (!sh)
2705		return 0;
2706	BUG_ON(atomic_read(&sh->count));
2707	shrink_buffers(sh);
2708	free_stripe(conf->slab_cache, sh);
2709	atomic_dec(&conf->active_stripes);
2710	conf->max_nr_stripes--;
2711	return 1;
2712}
2713
2714static void shrink_stripes(struct r5conf *conf)
2715{
2716	while (conf->max_nr_stripes &&
2717	       drop_one_stripe(conf))
2718		;
2719
2720	kmem_cache_destroy(conf->slab_cache);
 
2721	conf->slab_cache = NULL;
2722}
2723
2724static void raid5_end_read_request(struct bio * bi)
2725{
2726	struct stripe_head *sh = bi->bi_private;
2727	struct r5conf *conf = sh->raid_conf;
2728	int disks = sh->disks, i;
2729	struct md_rdev *rdev = NULL;
2730	sector_t s;
 
 
2731
2732	for (i=0 ; i<disks; i++)
2733		if (bi == &sh->dev[i].req)
2734			break;
2735
2736	pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2737		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2738		bi->bi_status);
2739	if (i == disks) {
2740		BUG();
2741		return;
2742	}
2743	if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2744		/* If replacement finished while this request was outstanding,
2745		 * 'replacement' might be NULL already.
2746		 * In that case it moved down to 'rdev'.
2747		 * rdev is not removed until all requests are finished.
2748		 */
2749		rdev = conf->disks[i].replacement;
2750	if (!rdev)
2751		rdev = conf->disks[i].rdev;
2752
2753	if (use_new_offset(conf, sh))
2754		s = sh->sector + rdev->new_data_offset;
2755	else
2756		s = sh->sector + rdev->data_offset;
2757	if (!bi->bi_status) {
2758		set_bit(R5_UPTODATE, &sh->dev[i].flags);
2759		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2760			/* Note that this cannot happen on a
2761			 * replacement device.  We just fail those on
2762			 * any error
2763			 */
2764			pr_info_ratelimited(
2765				"md/raid:%s: read error corrected (%lu sectors at %llu on %pg)\n",
2766				mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2767				(unsigned long long)s,
2768				rdev->bdev);
2769			atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2770			clear_bit(R5_ReadError, &sh->dev[i].flags);
2771			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2772		} else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2773			clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2774
2775		if (test_bit(R5_InJournal, &sh->dev[i].flags))
2776			/*
2777			 * end read for a page in journal, this
2778			 * must be preparing for prexor in rmw
2779			 */
2780			set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2781
2782		if (atomic_read(&rdev->read_errors))
2783			atomic_set(&rdev->read_errors, 0);
2784	} else {
 
2785		int retry = 0;
2786		int set_bad = 0;
2787
2788		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2789		if (!(bi->bi_status == BLK_STS_PROTECTION))
2790			atomic_inc(&rdev->read_errors);
2791		if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2792			pr_warn_ratelimited(
2793				"md/raid:%s: read error on replacement device (sector %llu on %pg).\n",
 
2794				mdname(conf->mddev),
2795				(unsigned long long)s,
2796				rdev->bdev);
2797		else if (conf->mddev->degraded >= conf->max_degraded) {
2798			set_bad = 1;
2799			pr_warn_ratelimited(
2800				"md/raid:%s: read error not correctable (sector %llu on %pg).\n",
2801				mdname(conf->mddev),
2802				(unsigned long long)s,
2803				rdev->bdev);
2804		} else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2805			/* Oh, no!!! */
2806			set_bad = 1;
2807			pr_warn_ratelimited(
2808				"md/raid:%s: read error NOT corrected!! (sector %llu on %pg).\n",
 
2809				mdname(conf->mddev),
2810				(unsigned long long)s,
2811				rdev->bdev);
2812		} else if (atomic_read(&rdev->read_errors)
2813			 > conf->max_nr_stripes) {
2814			if (!test_bit(Faulty, &rdev->flags)) {
2815				pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2816				    mdname(conf->mddev),
2817				    atomic_read(&rdev->read_errors),
2818				    conf->max_nr_stripes);
2819				pr_warn("md/raid:%s: Too many read errors, failing device %pg.\n",
2820				    mdname(conf->mddev), rdev->bdev);
2821			}
2822		} else
2823			retry = 1;
2824		if (set_bad && test_bit(In_sync, &rdev->flags)
2825		    && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2826			retry = 1;
2827		if (retry)
2828			if (sh->qd_idx >= 0 && sh->pd_idx == i)
2829				set_bit(R5_ReadError, &sh->dev[i].flags);
2830			else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2831				set_bit(R5_ReadError, &sh->dev[i].flags);
2832				clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2833			} else
2834				set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2835		else {
2836			clear_bit(R5_ReadError, &sh->dev[i].flags);
2837			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2838			if (!(set_bad
2839			      && test_bit(In_sync, &rdev->flags)
2840			      && rdev_set_badblocks(
2841				      rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2842				md_error(conf->mddev, rdev);
2843		}
2844	}
2845	rdev_dec_pending(rdev, conf->mddev);
2846	bio_uninit(bi);
2847	clear_bit(R5_LOCKED, &sh->dev[i].flags);
2848	set_bit(STRIPE_HANDLE, &sh->state);
2849	raid5_release_stripe(sh);
2850}
2851
2852static void raid5_end_write_request(struct bio *bi)
2853{
2854	struct stripe_head *sh = bi->bi_private;
2855	struct r5conf *conf = sh->raid_conf;
2856	int disks = sh->disks, i;
2857	struct md_rdev *rdev;
2858	sector_t first_bad;
2859	int bad_sectors;
2860	int replacement = 0;
2861
2862	for (i = 0 ; i < disks; i++) {
2863		if (bi == &sh->dev[i].req) {
2864			rdev = conf->disks[i].rdev;
2865			break;
2866		}
2867		if (bi == &sh->dev[i].rreq) {
2868			rdev = conf->disks[i].replacement;
2869			if (rdev)
2870				replacement = 1;
2871			else
2872				/* rdev was removed and 'replacement'
2873				 * replaced it.  rdev is not removed
2874				 * until all requests are finished.
2875				 */
2876				rdev = conf->disks[i].rdev;
2877			break;
2878		}
2879	}
2880	pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2881		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2882		bi->bi_status);
2883	if (i == disks) {
2884		BUG();
2885		return;
2886	}
2887
2888	if (replacement) {
2889		if (bi->bi_status)
2890			md_error(conf->mddev, rdev);
2891		else if (is_badblock(rdev, sh->sector,
2892				     RAID5_STRIPE_SECTORS(conf),
2893				     &first_bad, &bad_sectors))
2894			set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2895	} else {
2896		if (bi->bi_status) {
2897			set_bit(STRIPE_DEGRADED, &sh->state);
2898			set_bit(WriteErrorSeen, &rdev->flags);
2899			set_bit(R5_WriteError, &sh->dev[i].flags);
2900			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2901				set_bit(MD_RECOVERY_NEEDED,
2902					&rdev->mddev->recovery);
2903		} else if (is_badblock(rdev, sh->sector,
2904				       RAID5_STRIPE_SECTORS(conf),
2905				       &first_bad, &bad_sectors)) {
2906			set_bit(R5_MadeGood, &sh->dev[i].flags);
2907			if (test_bit(R5_ReadError, &sh->dev[i].flags))
2908				/* That was a successful write so make
2909				 * sure it looks like we already did
2910				 * a re-write.
2911				 */
2912				set_bit(R5_ReWrite, &sh->dev[i].flags);
2913		}
2914	}
2915	rdev_dec_pending(rdev, conf->mddev);
2916
2917	if (sh->batch_head && bi->bi_status && !replacement)
2918		set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2919
2920	bio_uninit(bi);
2921	if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2922		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2923	set_bit(STRIPE_HANDLE, &sh->state);
2924
2925	if (sh->batch_head && sh != sh->batch_head)
2926		raid5_release_stripe(sh->batch_head);
2927	raid5_release_stripe(sh);
2928}
2929
2930static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
 
 
 
2931{
2932	struct r5conf *conf = mddev->private;
2933	unsigned long flags;
2934	pr_debug("raid456: error called\n");
2935
2936	pr_crit("md/raid:%s: Disk failure on %pg, disabling device.\n",
2937		mdname(mddev), rdev->bdev);
 
 
 
 
 
2938
2939	spin_lock_irqsave(&conf->device_lock, flags);
2940	set_bit(Faulty, &rdev->flags);
2941	clear_bit(In_sync, &rdev->flags);
2942	mddev->degraded = raid5_calc_degraded(conf);
2943
2944	if (has_failed(conf)) {
2945		set_bit(MD_BROKEN, &conf->mddev->flags);
2946		conf->recovery_disabled = mddev->recovery_disabled;
2947
2948		pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2949			mdname(mddev), mddev->degraded, conf->raid_disks);
2950	} else {
2951		pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2952			mdname(mddev), conf->raid_disks - mddev->degraded);
2953	}
2954
2955	spin_unlock_irqrestore(&conf->device_lock, flags);
2956	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
 
 
 
2957
 
 
 
 
 
 
 
 
 
 
2958	set_bit(Blocked, &rdev->flags);
2959	set_mask_bits(&mddev->sb_flags, 0,
2960		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2961	r5c_update_on_rdev_error(mddev, rdev);
 
 
 
 
 
 
2962}
2963
2964/*
2965 * Input: a 'big' sector number,
2966 * Output: index of the data and parity disk, and the sector # in them.
2967 */
2968sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2969			      int previous, int *dd_idx,
2970			      struct stripe_head *sh)
2971{
2972	sector_t stripe, stripe2;
2973	sector_t chunk_number;
2974	unsigned int chunk_offset;
2975	int pd_idx, qd_idx;
2976	int ddf_layout = 0;
2977	sector_t new_sector;
2978	int algorithm = previous ? conf->prev_algo
2979				 : conf->algorithm;
2980	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2981					 : conf->chunk_sectors;
2982	int raid_disks = previous ? conf->previous_raid_disks
2983				  : conf->raid_disks;
2984	int data_disks = raid_disks - conf->max_degraded;
2985
2986	/* First compute the information on this sector */
2987
2988	/*
2989	 * Compute the chunk number and the sector offset inside the chunk
2990	 */
2991	chunk_offset = sector_div(r_sector, sectors_per_chunk);
2992	chunk_number = r_sector;
2993
2994	/*
2995	 * Compute the stripe number
2996	 */
2997	stripe = chunk_number;
2998	*dd_idx = sector_div(stripe, data_disks);
2999	stripe2 = stripe;
3000	/*
3001	 * Select the parity disk based on the user selected algorithm.
3002	 */
3003	pd_idx = qd_idx = -1;
3004	switch(conf->level) {
3005	case 4:
3006		pd_idx = data_disks;
3007		break;
3008	case 5:
3009		switch (algorithm) {
3010		case ALGORITHM_LEFT_ASYMMETRIC:
3011			pd_idx = data_disks - sector_div(stripe2, raid_disks);
3012			if (*dd_idx >= pd_idx)
3013				(*dd_idx)++;
3014			break;
3015		case ALGORITHM_RIGHT_ASYMMETRIC:
3016			pd_idx = sector_div(stripe2, raid_disks);
3017			if (*dd_idx >= pd_idx)
3018				(*dd_idx)++;
3019			break;
3020		case ALGORITHM_LEFT_SYMMETRIC:
3021			pd_idx = data_disks - sector_div(stripe2, raid_disks);
3022			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3023			break;
3024		case ALGORITHM_RIGHT_SYMMETRIC:
3025			pd_idx = sector_div(stripe2, raid_disks);
3026			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3027			break;
3028		case ALGORITHM_PARITY_0:
3029			pd_idx = 0;
3030			(*dd_idx)++;
3031			break;
3032		case ALGORITHM_PARITY_N:
3033			pd_idx = data_disks;
3034			break;
3035		default:
3036			BUG();
3037		}
3038		break;
3039	case 6:
3040
3041		switch (algorithm) {
3042		case ALGORITHM_LEFT_ASYMMETRIC:
3043			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3044			qd_idx = pd_idx + 1;
3045			if (pd_idx == raid_disks-1) {
3046				(*dd_idx)++;	/* Q D D D P */
3047				qd_idx = 0;
3048			} else if (*dd_idx >= pd_idx)
3049				(*dd_idx) += 2; /* D D P Q D */
3050			break;
3051		case ALGORITHM_RIGHT_ASYMMETRIC:
3052			pd_idx = sector_div(stripe2, raid_disks);
3053			qd_idx = pd_idx + 1;
3054			if (pd_idx == raid_disks-1) {
3055				(*dd_idx)++;	/* Q D D D P */
3056				qd_idx = 0;
3057			} else if (*dd_idx >= pd_idx)
3058				(*dd_idx) += 2; /* D D P Q D */
3059			break;
3060		case ALGORITHM_LEFT_SYMMETRIC:
3061			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3062			qd_idx = (pd_idx + 1) % raid_disks;
3063			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3064			break;
3065		case ALGORITHM_RIGHT_SYMMETRIC:
3066			pd_idx = sector_div(stripe2, raid_disks);
3067			qd_idx = (pd_idx + 1) % raid_disks;
3068			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3069			break;
3070
3071		case ALGORITHM_PARITY_0:
3072			pd_idx = 0;
3073			qd_idx = 1;
3074			(*dd_idx) += 2;
3075			break;
3076		case ALGORITHM_PARITY_N:
3077			pd_idx = data_disks;
3078			qd_idx = data_disks + 1;
3079			break;
3080
3081		case ALGORITHM_ROTATING_ZERO_RESTART:
3082			/* Exactly the same as RIGHT_ASYMMETRIC, but or
3083			 * of blocks for computing Q is different.
3084			 */
3085			pd_idx = sector_div(stripe2, raid_disks);
3086			qd_idx = pd_idx + 1;
3087			if (pd_idx == raid_disks-1) {
3088				(*dd_idx)++;	/* Q D D D P */
3089				qd_idx = 0;
3090			} else if (*dd_idx >= pd_idx)
3091				(*dd_idx) += 2; /* D D P Q D */
3092			ddf_layout = 1;
3093			break;
3094
3095		case ALGORITHM_ROTATING_N_RESTART:
3096			/* Same a left_asymmetric, by first stripe is
3097			 * D D D P Q  rather than
3098			 * Q D D D P
3099			 */
3100			stripe2 += 1;
3101			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3102			qd_idx = pd_idx + 1;
3103			if (pd_idx == raid_disks-1) {
3104				(*dd_idx)++;	/* Q D D D P */
3105				qd_idx = 0;
3106			} else if (*dd_idx >= pd_idx)
3107				(*dd_idx) += 2; /* D D P Q D */
3108			ddf_layout = 1;
3109			break;
3110
3111		case ALGORITHM_ROTATING_N_CONTINUE:
3112			/* Same as left_symmetric but Q is before P */
3113			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3114			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3115			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3116			ddf_layout = 1;
3117			break;
3118
3119		case ALGORITHM_LEFT_ASYMMETRIC_6:
3120			/* RAID5 left_asymmetric, with Q on last device */
3121			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3122			if (*dd_idx >= pd_idx)
3123				(*dd_idx)++;
3124			qd_idx = raid_disks - 1;
3125			break;
3126
3127		case ALGORITHM_RIGHT_ASYMMETRIC_6:
3128			pd_idx = sector_div(stripe2, raid_disks-1);
3129			if (*dd_idx >= pd_idx)
3130				(*dd_idx)++;
3131			qd_idx = raid_disks - 1;
3132			break;
3133
3134		case ALGORITHM_LEFT_SYMMETRIC_6:
3135			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3136			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3137			qd_idx = raid_disks - 1;
3138			break;
3139
3140		case ALGORITHM_RIGHT_SYMMETRIC_6:
3141			pd_idx = sector_div(stripe2, raid_disks-1);
3142			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3143			qd_idx = raid_disks - 1;
3144			break;
3145
3146		case ALGORITHM_PARITY_0_6:
3147			pd_idx = 0;
3148			(*dd_idx)++;
3149			qd_idx = raid_disks - 1;
3150			break;
3151
3152		default:
3153			BUG();
3154		}
3155		break;
3156	}
3157
3158	if (sh) {
3159		sh->pd_idx = pd_idx;
3160		sh->qd_idx = qd_idx;
3161		sh->ddf_layout = ddf_layout;
3162	}
3163	/*
3164	 * Finally, compute the new sector number
3165	 */
3166	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3167	return new_sector;
3168}
3169
3170sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
 
3171{
3172	struct r5conf *conf = sh->raid_conf;
3173	int raid_disks = sh->disks;
3174	int data_disks = raid_disks - conf->max_degraded;
3175	sector_t new_sector = sh->sector, check;
3176	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3177					 : conf->chunk_sectors;
3178	int algorithm = previous ? conf->prev_algo
3179				 : conf->algorithm;
3180	sector_t stripe;
3181	int chunk_offset;
3182	sector_t chunk_number;
3183	int dummy1, dd_idx = i;
3184	sector_t r_sector;
3185	struct stripe_head sh2;
3186
 
3187	chunk_offset = sector_div(new_sector, sectors_per_chunk);
3188	stripe = new_sector;
3189
3190	if (i == sh->pd_idx)
3191		return 0;
3192	switch(conf->level) {
3193	case 4: break;
3194	case 5:
3195		switch (algorithm) {
3196		case ALGORITHM_LEFT_ASYMMETRIC:
3197		case ALGORITHM_RIGHT_ASYMMETRIC:
3198			if (i > sh->pd_idx)
3199				i--;
3200			break;
3201		case ALGORITHM_LEFT_SYMMETRIC:
3202		case ALGORITHM_RIGHT_SYMMETRIC:
3203			if (i < sh->pd_idx)
3204				i += raid_disks;
3205			i -= (sh->pd_idx + 1);
3206			break;
3207		case ALGORITHM_PARITY_0:
3208			i -= 1;
3209			break;
3210		case ALGORITHM_PARITY_N:
3211			break;
3212		default:
3213			BUG();
3214		}
3215		break;
3216	case 6:
3217		if (i == sh->qd_idx)
3218			return 0; /* It is the Q disk */
3219		switch (algorithm) {
3220		case ALGORITHM_LEFT_ASYMMETRIC:
3221		case ALGORITHM_RIGHT_ASYMMETRIC:
3222		case ALGORITHM_ROTATING_ZERO_RESTART:
3223		case ALGORITHM_ROTATING_N_RESTART:
3224			if (sh->pd_idx == raid_disks-1)
3225				i--;	/* Q D D D P */
3226			else if (i > sh->pd_idx)
3227				i -= 2; /* D D P Q D */
3228			break;
3229		case ALGORITHM_LEFT_SYMMETRIC:
3230		case ALGORITHM_RIGHT_SYMMETRIC:
3231			if (sh->pd_idx == raid_disks-1)
3232				i--; /* Q D D D P */
3233			else {
3234				/* D D P Q D */
3235				if (i < sh->pd_idx)
3236					i += raid_disks;
3237				i -= (sh->pd_idx + 2);
3238			}
3239			break;
3240		case ALGORITHM_PARITY_0:
3241			i -= 2;
3242			break;
3243		case ALGORITHM_PARITY_N:
3244			break;
3245		case ALGORITHM_ROTATING_N_CONTINUE:
3246			/* Like left_symmetric, but P is before Q */
3247			if (sh->pd_idx == 0)
3248				i--;	/* P D D D Q */
3249			else {
3250				/* D D Q P D */
3251				if (i < sh->pd_idx)
3252					i += raid_disks;
3253				i -= (sh->pd_idx + 1);
3254			}
3255			break;
3256		case ALGORITHM_LEFT_ASYMMETRIC_6:
3257		case ALGORITHM_RIGHT_ASYMMETRIC_6:
3258			if (i > sh->pd_idx)
3259				i--;
3260			break;
3261		case ALGORITHM_LEFT_SYMMETRIC_6:
3262		case ALGORITHM_RIGHT_SYMMETRIC_6:
3263			if (i < sh->pd_idx)
3264				i += data_disks + 1;
3265			i -= (sh->pd_idx + 1);
3266			break;
3267		case ALGORITHM_PARITY_0_6:
3268			i -= 1;
3269			break;
3270		default:
3271			BUG();
3272		}
3273		break;
3274	}
3275
3276	chunk_number = stripe * data_disks + i;
3277	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3278
3279	check = raid5_compute_sector(conf, r_sector,
3280				     previous, &dummy1, &sh2);
3281	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3282		|| sh2.qd_idx != sh->qd_idx) {
3283		pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3284			mdname(conf->mddev));
3285		return 0;
3286	}
3287	return r_sector;
3288}
3289
3290/*
3291 * There are cases where we want handle_stripe_dirtying() and
3292 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3293 *
3294 * This function checks whether we want to delay the towrite. Specifically,
3295 * we delay the towrite when:
3296 *
3297 *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3298 *      stripe has data in journal (for other devices).
3299 *
3300 *      In this case, when reading data for the non-overwrite dev, it is
3301 *      necessary to handle complex rmw of write back cache (prexor with
3302 *      orig_page, and xor with page). To keep read path simple, we would
3303 *      like to flush data in journal to RAID disks first, so complex rmw
3304 *      is handled in the write patch (handle_stripe_dirtying).
3305 *
3306 *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3307 *
3308 *      It is important to be able to flush all stripes in raid5-cache.
3309 *      Therefore, we need reserve some space on the journal device for
3310 *      these flushes. If flush operation includes pending writes to the
3311 *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3312 *      for the flush out. If we exclude these pending writes from flush
3313 *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3314 *      Therefore, excluding pending writes in these cases enables more
3315 *      efficient use of the journal device.
3316 *
3317 *      Note: To make sure the stripe makes progress, we only delay
3318 *      towrite for stripes with data already in journal (injournal > 0).
3319 *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3320 *      no_space_stripes list.
3321 *
3322 *   3. during journal failure
3323 *      In journal failure, we try to flush all cached data to raid disks
3324 *      based on data in stripe cache. The array is read-only to upper
3325 *      layers, so we would skip all pending writes.
3326 *
3327 */
3328static inline bool delay_towrite(struct r5conf *conf,
3329				 struct r5dev *dev,
3330				 struct stripe_head_state *s)
3331{
3332	/* case 1 above */
3333	if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3334	    !test_bit(R5_Insync, &dev->flags) && s->injournal)
3335		return true;
3336	/* case 2 above */
3337	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3338	    s->injournal > 0)
3339		return true;
3340	/* case 3 above */
3341	if (s->log_failed && s->injournal)
3342		return true;
3343	return false;
3344}
3345
3346static void
3347schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3348			 int rcw, int expand)
3349{
3350	int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3351	struct r5conf *conf = sh->raid_conf;
3352	int level = conf->level;
3353
3354	if (rcw) {
3355		/*
3356		 * In some cases, handle_stripe_dirtying initially decided to
3357		 * run rmw and allocates extra page for prexor. However, rcw is
3358		 * cheaper later on. We need to free the extra page now,
3359		 * because we won't be able to do that in ops_complete_prexor().
3360		 */
3361		r5c_release_extra_page(sh);
3362
3363		for (i = disks; i--; ) {
3364			struct r5dev *dev = &sh->dev[i];
3365
3366			if (dev->towrite && !delay_towrite(conf, dev, s)) {
3367				set_bit(R5_LOCKED, &dev->flags);
3368				set_bit(R5_Wantdrain, &dev->flags);
3369				if (!expand)
3370					clear_bit(R5_UPTODATE, &dev->flags);
3371				s->locked++;
3372			} else if (test_bit(R5_InJournal, &dev->flags)) {
3373				set_bit(R5_LOCKED, &dev->flags);
3374				s->locked++;
3375			}
3376		}
3377		/* if we are not expanding this is a proper write request, and
3378		 * there will be bios with new data to be drained into the
3379		 * stripe cache
3380		 */
3381		if (!expand) {
3382			if (!s->locked)
3383				/* False alarm, nothing to do */
3384				return;
3385			sh->reconstruct_state = reconstruct_state_drain_run;
3386			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3387		} else
3388			sh->reconstruct_state = reconstruct_state_run;
3389
3390		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3391
 
 
 
 
 
 
 
 
 
 
 
3392		if (s->locked + conf->max_degraded == disks)
3393			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3394				atomic_inc(&conf->pending_full_writes);
3395	} else {
 
3396		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3397			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3398		BUG_ON(level == 6 &&
3399			(!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3400			   test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
 
 
3401
3402		for (i = disks; i--; ) {
3403			struct r5dev *dev = &sh->dev[i];
3404			if (i == pd_idx || i == qd_idx)
3405				continue;
3406
3407			if (dev->towrite &&
3408			    (test_bit(R5_UPTODATE, &dev->flags) ||
3409			     test_bit(R5_Wantcompute, &dev->flags))) {
3410				set_bit(R5_Wantdrain, &dev->flags);
3411				set_bit(R5_LOCKED, &dev->flags);
3412				clear_bit(R5_UPTODATE, &dev->flags);
3413				s->locked++;
3414			} else if (test_bit(R5_InJournal, &dev->flags)) {
3415				set_bit(R5_LOCKED, &dev->flags);
3416				s->locked++;
3417			}
3418		}
3419		if (!s->locked)
3420			/* False alarm - nothing to do */
3421			return;
3422		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3423		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3424		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3425		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3426	}
3427
3428	/* keep the parity disk(s) locked while asynchronous operations
3429	 * are in flight
3430	 */
3431	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3432	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3433	s->locked++;
3434
3435	if (level == 6) {
3436		int qd_idx = sh->qd_idx;
3437		struct r5dev *dev = &sh->dev[qd_idx];
3438
3439		set_bit(R5_LOCKED, &dev->flags);
3440		clear_bit(R5_UPTODATE, &dev->flags);
3441		s->locked++;
3442	}
3443
3444	if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3445	    test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3446	    !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3447	    test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3448		set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3449
3450	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3451		__func__, (unsigned long long)sh->sector,
3452		s->locked, s->ops_request);
3453}
3454
3455static bool stripe_bio_overlaps(struct stripe_head *sh, struct bio *bi,
3456				int dd_idx, int forwrite)
 
 
 
 
3457{
3458	struct r5conf *conf = sh->raid_conf;
3459	struct bio **bip;
 
 
3460
3461	pr_debug("checking bi b#%llu to stripe s#%llu\n",
3462		 bi->bi_iter.bi_sector, sh->sector);
3463
3464	/* Don't allow new IO added to stripes in batch list */
3465	if (sh->batch_head)
3466		return true;
3467
3468	if (forwrite)
3469		bip = &sh->dev[dd_idx].towrite;
3470	else
3471		bip = &sh->dev[dd_idx].toread;
3472
3473	while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3474		if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3475			return true;
3476		bip = &(*bip)->bi_next;
3477	}
3478
3479	if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3480		return true;
3481
3482	if (forwrite && raid5_has_ppl(conf)) {
3483		/*
3484		 * With PPL only writes to consecutive data chunks within a
3485		 * stripe are allowed because for a single stripe_head we can
3486		 * only have one PPL entry at a time, which describes one data
3487		 * range. Not really an overlap, but wait_for_overlap can be
3488		 * used to handle this.
3489		 */
3490		sector_t sector;
3491		sector_t first = 0;
3492		sector_t last = 0;
3493		int count = 0;
3494		int i;
3495
3496		for (i = 0; i < sh->disks; i++) {
3497			if (i != sh->pd_idx &&
3498			    (i == dd_idx || sh->dev[i].towrite)) {
3499				sector = sh->dev[i].sector;
3500				if (count == 0 || sector < first)
3501					first = sector;
3502				if (sector > last)
3503					last = sector;
3504				count++;
3505			}
3506		}
3507
3508		if (first + conf->chunk_sectors * (count - 1) != last)
3509			return true;
3510	}
3511
3512	return false;
3513}
3514
3515static void __add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3516			     int dd_idx, int forwrite, int previous)
3517{
3518	struct r5conf *conf = sh->raid_conf;
3519	struct bio **bip;
3520	int firstwrite = 0;
3521
 
3522	if (forwrite) {
3523		bip = &sh->dev[dd_idx].towrite;
3524		if (!*bip)
3525			firstwrite = 1;
3526	} else {
3527		bip = &sh->dev[dd_idx].toread;
 
 
 
 
3528	}
3529
3530	while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector)
3531		bip = &(*bip)->bi_next;
3532
3533	if (!forwrite || previous)
3534		clear_bit(STRIPE_BATCH_READY, &sh->state);
3535
3536	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3537	if (*bip)
3538		bi->bi_next = *bip;
3539	*bip = bi;
3540	bio_inc_remaining(bi);
3541	md_write_inc(conf->mddev, bi);
3542
3543	if (forwrite) {
3544		/* check if page is covered */
3545		sector_t sector = sh->dev[dd_idx].sector;
3546		for (bi=sh->dev[dd_idx].towrite;
3547		     sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3548			     bi && bi->bi_iter.bi_sector <= sector;
3549		     bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3550			if (bio_end_sector(bi) >= sector)
3551				sector = bio_end_sector(bi);
3552		}
3553		if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3554			if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3555				sh->overwrite_disks++;
3556	}
3557
3558	pr_debug("added bi b#%llu to stripe s#%llu, disk %d, logical %llu\n",
3559		 (*bip)->bi_iter.bi_sector, sh->sector, dd_idx,
3560		 sh->dev[dd_idx].sector);
3561
3562	if (conf->mddev->bitmap && firstwrite) {
3563		/* Cannot hold spinlock over bitmap_startwrite,
3564		 * but must ensure this isn't added to a batch until
3565		 * we have added to the bitmap and set bm_seq.
3566		 * So set STRIPE_BITMAP_PENDING to prevent
3567		 * batching.
3568		 * If multiple __add_stripe_bio() calls race here they
3569		 * much all set STRIPE_BITMAP_PENDING.  So only the first one
3570		 * to complete "bitmap_startwrite" gets to set
3571		 * STRIPE_BIT_DELAY.  This is important as once a stripe
3572		 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3573		 * any more.
3574		 */
3575		set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3576		spin_unlock_irq(&sh->stripe_lock);
3577		md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3578				     RAID5_STRIPE_SECTORS(conf), 0);
3579		spin_lock_irq(&sh->stripe_lock);
3580		clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3581		if (!sh->batch_head) {
3582			sh->bm_seq = conf->seq_flush+1;
3583			set_bit(STRIPE_BIT_DELAY, &sh->state);
3584		}
 
 
3585	}
3586}
3587
3588/*
3589 * Each stripe/dev can have one or more bios attached.
3590 * toread/towrite point to the first in a chain.
3591 * The bi_next chain must be in order.
3592 */
3593static bool add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3594			   int dd_idx, int forwrite, int previous)
3595{
3596	spin_lock_irq(&sh->stripe_lock);
3597
3598	if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
3599		set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3600		spin_unlock_irq(&sh->stripe_lock);
3601		return false;
 
3602	}
 
3603
3604	__add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
3605	spin_unlock_irq(&sh->stripe_lock);
3606	return true;
 
3607}
3608
3609static void end_reshape(struct r5conf *conf);
3610
3611static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3612			    struct stripe_head *sh)
3613{
3614	int sectors_per_chunk =
3615		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3616	int dd_idx;
3617	int chunk_offset = sector_div(stripe, sectors_per_chunk);
3618	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3619
3620	raid5_compute_sector(conf,
3621			     stripe * (disks - conf->max_degraded)
3622			     *sectors_per_chunk + chunk_offset,
3623			     previous,
3624			     &dd_idx, sh);
3625}
3626
3627static void
3628handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3629		     struct stripe_head_state *s, int disks)
 
3630{
3631	int i;
3632	BUG_ON(sh->batch_head);
3633	for (i = disks; i--; ) {
3634		struct bio *bi;
3635		int bitmap_end = 0;
3636
3637		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3638			struct md_rdev *rdev = conf->disks[i].rdev;
3639
3640			if (rdev && test_bit(In_sync, &rdev->flags) &&
3641			    !test_bit(Faulty, &rdev->flags))
3642				atomic_inc(&rdev->nr_pending);
3643			else
3644				rdev = NULL;
 
3645			if (rdev) {
3646				if (!rdev_set_badblocks(
3647					    rdev,
3648					    sh->sector,
3649					    RAID5_STRIPE_SECTORS(conf), 0))
3650					md_error(conf->mddev, rdev);
3651				rdev_dec_pending(rdev, conf->mddev);
3652			}
3653		}
3654		spin_lock_irq(&sh->stripe_lock);
3655		/* fail all writes first */
3656		bi = sh->dev[i].towrite;
3657		sh->dev[i].towrite = NULL;
3658		sh->overwrite_disks = 0;
3659		spin_unlock_irq(&sh->stripe_lock);
3660		if (bi)
3661			bitmap_end = 1;
3662
3663		log_stripe_write_finished(sh);
3664
3665		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3666			wake_up(&conf->wait_for_overlap);
3667
3668		while (bi && bi->bi_iter.bi_sector <
3669			sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3670			struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3671
3672			md_write_end(conf->mddev);
3673			bio_io_error(bi);
 
 
 
3674			bi = nextbi;
3675		}
3676		if (bitmap_end)
3677			md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3678					   RAID5_STRIPE_SECTORS(conf), 0, 0);
3679		bitmap_end = 0;
3680		/* and fail all 'written' */
3681		bi = sh->dev[i].written;
3682		sh->dev[i].written = NULL;
3683		if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3684			WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3685			sh->dev[i].page = sh->dev[i].orig_page;
3686		}
3687
3688		if (bi) bitmap_end = 1;
3689		while (bi && bi->bi_iter.bi_sector <
3690		       sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3691			struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3692
3693			md_write_end(conf->mddev);
3694			bio_io_error(bi);
 
 
 
3695			bi = bi2;
3696		}
3697
3698		/* fail any reads if this device is non-operational and
3699		 * the data has not reached the cache yet.
3700		 */
3701		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3702		    s->failed > conf->max_degraded &&
3703		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3704		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
3705			spin_lock_irq(&sh->stripe_lock);
3706			bi = sh->dev[i].toread;
3707			sh->dev[i].toread = NULL;
3708			spin_unlock_irq(&sh->stripe_lock);
3709			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3710				wake_up(&conf->wait_for_overlap);
3711			if (bi)
3712				s->to_read--;
3713			while (bi && bi->bi_iter.bi_sector <
3714			       sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3715				struct bio *nextbi =
3716					r5_next_bio(conf, bi, sh->dev[i].sector);
3717
3718				bio_io_error(bi);
 
 
 
3719				bi = nextbi;
3720			}
3721		}
 
3722		if (bitmap_end)
3723			md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3724					   RAID5_STRIPE_SECTORS(conf), 0, 0);
3725		/* If we were in the middle of a write the parity block might
3726		 * still be locked - so just clear all R5_LOCKED flags
3727		 */
3728		clear_bit(R5_LOCKED, &sh->dev[i].flags);
3729	}
3730	s->to_write = 0;
3731	s->written = 0;
3732
3733	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3734		if (atomic_dec_and_test(&conf->pending_full_writes))
3735			md_wakeup_thread(conf->mddev->thread);
3736}
3737
3738static void
3739handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3740		   struct stripe_head_state *s)
3741{
3742	int abort = 0;
3743	int i;
3744
3745	BUG_ON(sh->batch_head);
3746	clear_bit(STRIPE_SYNCING, &sh->state);
3747	if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3748		wake_up(&conf->wait_for_overlap);
3749	s->syncing = 0;
3750	s->replacing = 0;
3751	/* There is nothing more to do for sync/check/repair.
3752	 * Don't even need to abort as that is handled elsewhere
3753	 * if needed, and not always wanted e.g. if there is a known
3754	 * bad block here.
3755	 * For recover/replace we need to record a bad block on all
3756	 * non-sync devices, or abort the recovery
3757	 */
3758	if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3759		/* During recovery devices cannot be removed, so
3760		 * locking and refcounting of rdevs is not needed
3761		 */
3762		for (i = 0; i < conf->raid_disks; i++) {
3763			struct md_rdev *rdev = conf->disks[i].rdev;
3764
3765			if (rdev
3766			    && !test_bit(Faulty, &rdev->flags)
3767			    && !test_bit(In_sync, &rdev->flags)
3768			    && !rdev_set_badblocks(rdev, sh->sector,
3769						   RAID5_STRIPE_SECTORS(conf), 0))
3770				abort = 1;
3771			rdev = conf->disks[i].replacement;
3772
3773			if (rdev
3774			    && !test_bit(Faulty, &rdev->flags)
3775			    && !test_bit(In_sync, &rdev->flags)
3776			    && !rdev_set_badblocks(rdev, sh->sector,
3777						   RAID5_STRIPE_SECTORS(conf), 0))
3778				abort = 1;
3779		}
3780		if (abort)
3781			conf->recovery_disabled =
3782				conf->mddev->recovery_disabled;
3783	}
3784	md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3785}
3786
3787static int want_replace(struct stripe_head *sh, int disk_idx)
3788{
3789	struct md_rdev *rdev;
3790	int rv = 0;
3791
3792	rdev = sh->raid_conf->disks[disk_idx].replacement;
3793	if (rdev
3794	    && !test_bit(Faulty, &rdev->flags)
3795	    && !test_bit(In_sync, &rdev->flags)
3796	    && (rdev->recovery_offset <= sh->sector
3797		|| rdev->mddev->recovery_cp <= sh->sector))
3798		rv = 1;
3799	return rv;
3800}
3801
3802static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3803			   int disk_idx, int disks)
3804{
3805	struct r5dev *dev = &sh->dev[disk_idx];
3806	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3807				  &sh->dev[s->failed_num[1]] };
3808	int i;
3809	bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3810
3811
3812	if (test_bit(R5_LOCKED, &dev->flags) ||
3813	    test_bit(R5_UPTODATE, &dev->flags))
3814		/* No point reading this as we already have it or have
3815		 * decided to get it.
3816		 */
3817		return 0;
3818
3819	if (dev->toread ||
3820	    (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3821		/* We need this block to directly satisfy a request */
3822		return 1;
3823
3824	if (s->syncing || s->expanding ||
3825	    (s->replacing && want_replace(sh, disk_idx)))
3826		/* When syncing, or expanding we read everything.
3827		 * When replacing, we need the replaced block.
3828		 */
3829		return 1;
3830
3831	if ((s->failed >= 1 && fdev[0]->toread) ||
3832	    (s->failed >= 2 && fdev[1]->toread))
3833		/* If we want to read from a failed device, then
3834		 * we need to actually read every other device.
3835		 */
3836		return 1;
3837
3838	/* Sometimes neither read-modify-write nor reconstruct-write
3839	 * cycles can work.  In those cases we read every block we
3840	 * can.  Then the parity-update is certain to have enough to
3841	 * work with.
3842	 * This can only be a problem when we need to write something,
3843	 * and some device has failed.  If either of those tests
3844	 * fail we need look no further.
3845	 */
3846	if (!s->failed || !s->to_write)
3847		return 0;
3848
3849	if (test_bit(R5_Insync, &dev->flags) &&
3850	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3851		/* Pre-reads at not permitted until after short delay
3852		 * to gather multiple requests.  However if this
3853		 * device is no Insync, the block could only be computed
3854		 * and there is no need to delay that.
3855		 */
3856		return 0;
3857
3858	for (i = 0; i < s->failed && i < 2; i++) {
3859		if (fdev[i]->towrite &&
3860		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3861		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3862			/* If we have a partial write to a failed
3863			 * device, then we will need to reconstruct
3864			 * the content of that device, so all other
3865			 * devices must be read.
3866			 */
3867			return 1;
3868
3869		if (s->failed >= 2 &&
3870		    (fdev[i]->towrite ||
3871		     s->failed_num[i] == sh->pd_idx ||
3872		     s->failed_num[i] == sh->qd_idx) &&
3873		    !test_bit(R5_UPTODATE, &fdev[i]->flags))
3874			/* In max degraded raid6, If the failed disk is P, Q,
3875			 * or we want to read the failed disk, we need to do
3876			 * reconstruct-write.
3877			 */
3878			force_rcw = true;
3879	}
3880
3881	/* If we are forced to do a reconstruct-write, because parity
3882	 * cannot be trusted and we are currently recovering it, there
3883	 * is extra need to be careful.
3884	 * If one of the devices that we would need to read, because
3885	 * it is not being overwritten (and maybe not written at all)
3886	 * is missing/faulty, then we need to read everything we can.
3887	 */
3888	if (!force_rcw &&
3889	    sh->sector < sh->raid_conf->mddev->recovery_cp)
3890		/* reconstruct-write isn't being forced */
3891		return 0;
3892	for (i = 0; i < s->failed && i < 2; i++) {
3893		if (s->failed_num[i] != sh->pd_idx &&
3894		    s->failed_num[i] != sh->qd_idx &&
3895		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3896		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3897			return 1;
3898	}
3899
3900	return 0;
3901}
3902
3903/* fetch_block - checks the given member device to see if its data needs
3904 * to be read or computed to satisfy a request.
3905 *
3906 * Returns 1 when no more member devices need to be checked, otherwise returns
3907 * 0 to tell the loop in handle_stripe_fill to continue
3908 */
3909static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3910		       int disk_idx, int disks)
3911{
3912	struct r5dev *dev = &sh->dev[disk_idx];
 
 
3913
3914	/* is the data in this block needed, and can we get it? */
3915	if (need_this_block(sh, s, disk_idx, disks)) {
 
 
 
 
 
 
 
 
 
3916		/* we would like to get this block, possibly by computing it,
3917		 * otherwise read it if the backing disk is insync
3918		 */
3919		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3920		BUG_ON(test_bit(R5_Wantread, &dev->flags));
3921		BUG_ON(sh->batch_head);
3922
3923		/*
3924		 * In the raid6 case if the only non-uptodate disk is P
3925		 * then we already trusted P to compute the other failed
3926		 * drives. It is safe to compute rather than re-read P.
3927		 * In other cases we only compute blocks from failed
3928		 * devices, otherwise check/repair might fail to detect
3929		 * a real inconsistency.
3930		 */
3931
3932		if ((s->uptodate == disks - 1) &&
3933		    ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3934		    (s->failed && (disk_idx == s->failed_num[0] ||
3935				   disk_idx == s->failed_num[1])))) {
3936			/* have disk failed, and we're requested to fetch it;
3937			 * do compute it
3938			 */
3939			pr_debug("Computing stripe %llu block %d\n",
3940			       (unsigned long long)sh->sector, disk_idx);
3941			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3942			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3943			set_bit(R5_Wantcompute, &dev->flags);
3944			sh->ops.target = disk_idx;
3945			sh->ops.target2 = -1; /* no 2nd target */
3946			s->req_compute = 1;
3947			/* Careful: from this point on 'uptodate' is in the eye
3948			 * of raid_run_ops which services 'compute' operations
3949			 * before writes. R5_Wantcompute flags a block that will
3950			 * be R5_UPTODATE by the time it is needed for a
3951			 * subsequent operation.
3952			 */
3953			s->uptodate++;
3954			return 1;
3955		} else if (s->uptodate == disks-2 && s->failed >= 2) {
3956			/* Computing 2-failure is *very* expensive; only
3957			 * do it if failed >= 2
3958			 */
3959			int other;
3960			for (other = disks; other--; ) {
3961				if (other == disk_idx)
3962					continue;
3963				if (!test_bit(R5_UPTODATE,
3964				      &sh->dev[other].flags))
3965					break;
3966			}
3967			BUG_ON(other < 0);
3968			pr_debug("Computing stripe %llu blocks %d,%d\n",
3969			       (unsigned long long)sh->sector,
3970			       disk_idx, other);
3971			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3972			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3973			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3974			set_bit(R5_Wantcompute, &sh->dev[other].flags);
3975			sh->ops.target = disk_idx;
3976			sh->ops.target2 = other;
3977			s->uptodate += 2;
3978			s->req_compute = 1;
3979			return 1;
3980		} else if (test_bit(R5_Insync, &dev->flags)) {
3981			set_bit(R5_LOCKED, &dev->flags);
3982			set_bit(R5_Wantread, &dev->flags);
3983			s->locked++;
3984			pr_debug("Reading block %d (sync=%d)\n",
3985				disk_idx, s->syncing);
3986		}
3987	}
3988
3989	return 0;
3990}
3991
3992/*
3993 * handle_stripe_fill - read or compute data to satisfy pending requests.
3994 */
3995static void handle_stripe_fill(struct stripe_head *sh,
3996			       struct stripe_head_state *s,
3997			       int disks)
3998{
3999	int i;
4000
4001	/* look for blocks to read/compute, skip this if a compute
4002	 * is already in flight, or if the stripe contents are in the
4003	 * midst of changing due to a write
4004	 */
4005	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
4006	    !sh->reconstruct_state) {
4007
4008		/*
4009		 * For degraded stripe with data in journal, do not handle
4010		 * read requests yet, instead, flush the stripe to raid
4011		 * disks first, this avoids handling complex rmw of write
4012		 * back cache (prexor with orig_page, and then xor with
4013		 * page) in the read path
4014		 */
4015		if (s->to_read && s->injournal && s->failed) {
4016			if (test_bit(STRIPE_R5C_CACHING, &sh->state))
4017				r5c_make_stripe_write_out(sh);
4018			goto out;
4019		}
4020
4021		for (i = disks; i--; )
4022			if (fetch_block(sh, s, i, disks))
4023				break;
4024	}
4025out:
4026	set_bit(STRIPE_HANDLE, &sh->state);
4027}
4028
4029static void break_stripe_batch_list(struct stripe_head *head_sh,
4030				    unsigned long handle_flags);
4031/* handle_stripe_clean_event
4032 * any written block on an uptodate or failed drive can be returned.
4033 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
4034 * never LOCKED, so we don't need to test 'failed' directly.
4035 */
4036static void handle_stripe_clean_event(struct r5conf *conf,
4037	struct stripe_head *sh, int disks)
4038{
4039	int i;
4040	struct r5dev *dev;
4041	int discard_pending = 0;
4042	struct stripe_head *head_sh = sh;
4043	bool do_endio = false;
4044
4045	for (i = disks; i--; )
4046		if (sh->dev[i].written) {
4047			dev = &sh->dev[i];
4048			if (!test_bit(R5_LOCKED, &dev->flags) &&
4049			    (test_bit(R5_UPTODATE, &dev->flags) ||
4050			     test_bit(R5_Discard, &dev->flags) ||
4051			     test_bit(R5_SkipCopy, &dev->flags))) {
4052				/* We can return any write requests */
4053				struct bio *wbi, *wbi2;
 
4054				pr_debug("Return write for disc %d\n", i);
4055				if (test_and_clear_bit(R5_Discard, &dev->flags))
4056					clear_bit(R5_UPTODATE, &dev->flags);
4057				if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
4058					WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
4059				}
4060				do_endio = true;
4061
4062returnbi:
4063				dev->page = dev->orig_page;
4064				wbi = dev->written;
4065				dev->written = NULL;
4066				while (wbi && wbi->bi_iter.bi_sector <
4067					dev->sector + RAID5_STRIPE_SECTORS(conf)) {
4068					wbi2 = r5_next_bio(conf, wbi, dev->sector);
4069					md_write_end(conf->mddev);
4070					bio_endio(wbi);
4071					wbi = wbi2;
4072				}
4073				md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
4074						   RAID5_STRIPE_SECTORS(conf),
4075						   !test_bit(STRIPE_DEGRADED, &sh->state),
4076						   0);
4077				if (head_sh->batch_head) {
4078					sh = list_first_entry(&sh->batch_list,
4079							      struct stripe_head,
4080							      batch_list);
4081					if (sh != head_sh) {
4082						dev = &sh->dev[i];
4083						goto returnbi;
4084					}
 
4085				}
4086				sh = head_sh;
4087				dev = &sh->dev[i];
4088			} else if (test_bit(R5_Discard, &dev->flags))
4089				discard_pending = 1;
4090		}
4091
4092	log_stripe_write_finished(sh);
4093
4094	if (!discard_pending &&
4095	    test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4096		int hash;
4097		clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4098		clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4099		if (sh->qd_idx >= 0) {
4100			clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4101			clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4102		}
4103		/* now that discard is done we can proceed with any sync */
4104		clear_bit(STRIPE_DISCARD, &sh->state);
4105		/*
4106		 * SCSI discard will change some bio fields and the stripe has
4107		 * no updated data, so remove it from hash list and the stripe
4108		 * will be reinitialized
4109		 */
4110unhash:
4111		hash = sh->hash_lock_index;
4112		spin_lock_irq(conf->hash_locks + hash);
4113		remove_hash(sh);
4114		spin_unlock_irq(conf->hash_locks + hash);
4115		if (head_sh->batch_head) {
4116			sh = list_first_entry(&sh->batch_list,
4117					      struct stripe_head, batch_list);
4118			if (sh != head_sh)
4119					goto unhash;
4120		}
4121		sh = head_sh;
4122
4123		if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4124			set_bit(STRIPE_HANDLE, &sh->state);
4125
4126	}
4127
4128	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4129		if (atomic_dec_and_test(&conf->pending_full_writes))
4130			md_wakeup_thread(conf->mddev->thread);
4131
4132	if (head_sh->batch_head && do_endio)
4133		break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4134}
4135
4136/*
4137 * For RMW in write back cache, we need extra page in prexor to store the
4138 * old data. This page is stored in dev->orig_page.
4139 *
4140 * This function checks whether we have data for prexor. The exact logic
4141 * is:
4142 *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4143 */
4144static inline bool uptodate_for_rmw(struct r5dev *dev)
4145{
4146	return (test_bit(R5_UPTODATE, &dev->flags)) &&
4147		(!test_bit(R5_InJournal, &dev->flags) ||
4148		 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4149}
4150
4151static int handle_stripe_dirtying(struct r5conf *conf,
4152				  struct stripe_head *sh,
4153				  struct stripe_head_state *s,
4154				  int disks)
4155{
4156	int rmw = 0, rcw = 0, i;
4157	sector_t recovery_cp = conf->mddev->recovery_cp;
4158
4159	/* Check whether resync is now happening or should start.
4160	 * If yes, then the array is dirty (after unclean shutdown or
4161	 * initial creation), so parity in some stripes might be inconsistent.
4162	 * In this case, we need to always do reconstruct-write, to ensure
4163	 * that in case of drive failure or read-error correction, we
4164	 * generate correct data from the parity.
4165	 */
4166	if (conf->rmw_level == PARITY_DISABLE_RMW ||
4167	    (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4168	     s->failed == 0)) {
4169		/* Calculate the real rcw later - for now make it
4170		 * look like rcw is cheaper
4171		 */
4172		rcw = 1; rmw = 2;
4173		pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4174			 conf->rmw_level, (unsigned long long)recovery_cp,
4175			 (unsigned long long)sh->sector);
4176	} else for (i = disks; i--; ) {
4177		/* would I have to read this buffer for read_modify_write */
4178		struct r5dev *dev = &sh->dev[i];
4179		if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4180		     i == sh->pd_idx || i == sh->qd_idx ||
4181		     test_bit(R5_InJournal, &dev->flags)) &&
4182		    !test_bit(R5_LOCKED, &dev->flags) &&
4183		    !(uptodate_for_rmw(dev) ||
4184		      test_bit(R5_Wantcompute, &dev->flags))) {
4185			if (test_bit(R5_Insync, &dev->flags))
4186				rmw++;
4187			else
4188				rmw += 2*disks;  /* cannot read it */
4189		}
4190		/* Would I have to read this buffer for reconstruct_write */
4191		if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4192		    i != sh->pd_idx && i != sh->qd_idx &&
4193		    !test_bit(R5_LOCKED, &dev->flags) &&
4194		    !(test_bit(R5_UPTODATE, &dev->flags) ||
4195		      test_bit(R5_Wantcompute, &dev->flags))) {
4196			if (test_bit(R5_Insync, &dev->flags))
4197				rcw++;
4198			else
4199				rcw += 2*disks;
4200		}
4201	}
4202
4203	pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4204		 (unsigned long long)sh->sector, sh->state, rmw, rcw);
4205	set_bit(STRIPE_HANDLE, &sh->state);
4206	if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4207		/* prefer read-modify-write, but need to get some data */
4208		if (conf->mddev->queue)
4209			blk_add_trace_msg(conf->mddev->queue,
4210					  "raid5 rmw %llu %d",
4211					  (unsigned long long)sh->sector, rmw);
4212		for (i = disks; i--; ) {
4213			struct r5dev *dev = &sh->dev[i];
4214			if (test_bit(R5_InJournal, &dev->flags) &&
4215			    dev->page == dev->orig_page &&
4216			    !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4217				/* alloc page for prexor */
4218				struct page *p = alloc_page(GFP_NOIO);
4219
4220				if (p) {
4221					dev->orig_page = p;
4222					continue;
4223				}
4224
4225				/*
4226				 * alloc_page() failed, try use
4227				 * disk_info->extra_page
4228				 */
4229				if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4230						      &conf->cache_state)) {
4231					r5c_use_extra_page(sh);
4232					break;
4233				}
4234
4235				/* extra_page in use, add to delayed_list */
4236				set_bit(STRIPE_DELAYED, &sh->state);
4237				s->waiting_extra_page = 1;
4238				return -EAGAIN;
4239			}
4240		}
4241
4242		for (i = disks; i--; ) {
4243			struct r5dev *dev = &sh->dev[i];
4244			if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4245			     i == sh->pd_idx || i == sh->qd_idx ||
4246			     test_bit(R5_InJournal, &dev->flags)) &&
4247			    !test_bit(R5_LOCKED, &dev->flags) &&
4248			    !(uptodate_for_rmw(dev) ||
4249			      test_bit(R5_Wantcompute, &dev->flags)) &&
4250			    test_bit(R5_Insync, &dev->flags)) {
4251				if (test_bit(STRIPE_PREREAD_ACTIVE,
4252					     &sh->state)) {
4253					pr_debug("Read_old block %d for r-m-w\n",
4254						 i);
4255					set_bit(R5_LOCKED, &dev->flags);
4256					set_bit(R5_Wantread, &dev->flags);
4257					s->locked++;
4258				} else
4259					set_bit(STRIPE_DELAYED, &sh->state);
 
 
4260			}
4261		}
4262	}
4263	if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4264		/* want reconstruct write, but need to get some data */
4265		int qread =0;
4266		rcw = 0;
4267		for (i = disks; i--; ) {
4268			struct r5dev *dev = &sh->dev[i];
4269			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4270			    i != sh->pd_idx && i != sh->qd_idx &&
4271			    !test_bit(R5_LOCKED, &dev->flags) &&
4272			    !(test_bit(R5_UPTODATE, &dev->flags) ||
4273			      test_bit(R5_Wantcompute, &dev->flags))) {
4274				rcw++;
4275				if (test_bit(R5_Insync, &dev->flags) &&
4276				    test_bit(STRIPE_PREREAD_ACTIVE,
4277					     &sh->state)) {
 
4278					pr_debug("Read_old block "
4279						"%d for Reconstruct\n", i);
4280					set_bit(R5_LOCKED, &dev->flags);
4281					set_bit(R5_Wantread, &dev->flags);
4282					s->locked++;
4283					qread++;
4284				} else
4285					set_bit(STRIPE_DELAYED, &sh->state);
 
 
4286			}
4287		}
4288		if (rcw && conf->mddev->queue)
4289			blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4290					  (unsigned long long)sh->sector,
4291					  rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4292	}
4293
4294	if (rcw > disks && rmw > disks &&
4295	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4296		set_bit(STRIPE_DELAYED, &sh->state);
4297
4298	/* now if nothing is locked, and if we have enough data,
4299	 * we can start a write request
4300	 */
4301	/* since handle_stripe can be called at any time we need to handle the
4302	 * case where a compute block operation has been submitted and then a
4303	 * subsequent call wants to start a write request.  raid_run_ops only
4304	 * handles the case where compute block and reconstruct are requested
4305	 * simultaneously.  If this is not the case then new writes need to be
4306	 * held off until the compute completes.
4307	 */
4308	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4309	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4310	     !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4311		schedule_reconstruction(sh, s, rcw == 0, 0);
4312	return 0;
4313}
4314
4315static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4316				struct stripe_head_state *s, int disks)
4317{
4318	struct r5dev *dev = NULL;
4319
4320	BUG_ON(sh->batch_head);
4321	set_bit(STRIPE_HANDLE, &sh->state);
4322
4323	switch (sh->check_state) {
4324	case check_state_idle:
4325		/* start a new check operation if there are no failures */
4326		if (s->failed == 0) {
4327			BUG_ON(s->uptodate != disks);
4328			sh->check_state = check_state_run;
4329			set_bit(STRIPE_OP_CHECK, &s->ops_request);
4330			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4331			s->uptodate--;
4332			break;
4333		}
4334		dev = &sh->dev[s->failed_num[0]];
4335		fallthrough;
4336	case check_state_compute_result:
4337		sh->check_state = check_state_idle;
4338		if (!dev)
4339			dev = &sh->dev[sh->pd_idx];
4340
4341		/* check that a write has not made the stripe insync */
4342		if (test_bit(STRIPE_INSYNC, &sh->state))
4343			break;
4344
4345		/* either failed parity check, or recovery is happening */
4346		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4347		BUG_ON(s->uptodate != disks);
4348
4349		set_bit(R5_LOCKED, &dev->flags);
4350		s->locked++;
4351		set_bit(R5_Wantwrite, &dev->flags);
4352
4353		clear_bit(STRIPE_DEGRADED, &sh->state);
4354		set_bit(STRIPE_INSYNC, &sh->state);
4355		break;
4356	case check_state_run:
4357		break; /* we will be called again upon completion */
4358	case check_state_check_result:
4359		sh->check_state = check_state_idle;
4360
4361		/* if a failure occurred during the check operation, leave
4362		 * STRIPE_INSYNC not set and let the stripe be handled again
4363		 */
4364		if (s->failed)
4365			break;
4366
4367		/* handle a successful check operation, if parity is correct
4368		 * we are done.  Otherwise update the mismatch count and repair
4369		 * parity if !MD_RECOVERY_CHECK
4370		 */
4371		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4372			/* parity is correct (on disc,
4373			 * not in buffer any more)
4374			 */
4375			set_bit(STRIPE_INSYNC, &sh->state);
4376		else {
4377			atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4378			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4379				/* don't try to repair!! */
4380				set_bit(STRIPE_INSYNC, &sh->state);
4381				pr_warn_ratelimited("%s: mismatch sector in range "
4382						    "%llu-%llu\n", mdname(conf->mddev),
4383						    (unsigned long long) sh->sector,
4384						    (unsigned long long) sh->sector +
4385						    RAID5_STRIPE_SECTORS(conf));
4386			} else {
4387				sh->check_state = check_state_compute_run;
4388				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4389				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4390				set_bit(R5_Wantcompute,
4391					&sh->dev[sh->pd_idx].flags);
4392				sh->ops.target = sh->pd_idx;
4393				sh->ops.target2 = -1;
4394				s->uptodate++;
4395			}
4396		}
4397		break;
4398	case check_state_compute_run:
4399		break;
4400	default:
4401		pr_err("%s: unknown check_state: %d sector: %llu\n",
4402		       __func__, sh->check_state,
4403		       (unsigned long long) sh->sector);
4404		BUG();
4405	}
4406}
4407
4408static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
 
4409				  struct stripe_head_state *s,
4410				  int disks)
4411{
4412	int pd_idx = sh->pd_idx;
4413	int qd_idx = sh->qd_idx;
4414	struct r5dev *dev;
4415
4416	BUG_ON(sh->batch_head);
4417	set_bit(STRIPE_HANDLE, &sh->state);
4418
4419	BUG_ON(s->failed > 2);
4420
4421	/* Want to check and possibly repair P and Q.
4422	 * However there could be one 'failed' device, in which
4423	 * case we can only check one of them, possibly using the
4424	 * other to generate missing data
4425	 */
4426
4427	switch (sh->check_state) {
4428	case check_state_idle:
4429		/* start a new check operation if there are < 2 failures */
4430		if (s->failed == s->q_failed) {
4431			/* The only possible failed device holds Q, so it
4432			 * makes sense to check P (If anything else were failed,
4433			 * we would have used P to recreate it).
4434			 */
4435			sh->check_state = check_state_run;
4436		}
4437		if (!s->q_failed && s->failed < 2) {
4438			/* Q is not failed, and we didn't use it to generate
4439			 * anything, so it makes sense to check it
4440			 */
4441			if (sh->check_state == check_state_run)
4442				sh->check_state = check_state_run_pq;
4443			else
4444				sh->check_state = check_state_run_q;
4445		}
4446
4447		/* discard potentially stale zero_sum_result */
4448		sh->ops.zero_sum_result = 0;
4449
4450		if (sh->check_state == check_state_run) {
4451			/* async_xor_zero_sum destroys the contents of P */
4452			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4453			s->uptodate--;
4454		}
4455		if (sh->check_state >= check_state_run &&
4456		    sh->check_state <= check_state_run_pq) {
4457			/* async_syndrome_zero_sum preserves P and Q, so
4458			 * no need to mark them !uptodate here
4459			 */
4460			set_bit(STRIPE_OP_CHECK, &s->ops_request);
4461			break;
4462		}
4463
4464		/* we have 2-disk failure */
4465		BUG_ON(s->failed != 2);
4466		fallthrough;
4467	case check_state_compute_result:
4468		sh->check_state = check_state_idle;
4469
4470		/* check that a write has not made the stripe insync */
4471		if (test_bit(STRIPE_INSYNC, &sh->state))
4472			break;
4473
4474		/* now write out any block on a failed drive,
4475		 * or P or Q if they were recomputed
4476		 */
4477		dev = NULL;
4478		if (s->failed == 2) {
4479			dev = &sh->dev[s->failed_num[1]];
4480			s->locked++;
4481			set_bit(R5_LOCKED, &dev->flags);
4482			set_bit(R5_Wantwrite, &dev->flags);
4483		}
4484		if (s->failed >= 1) {
4485			dev = &sh->dev[s->failed_num[0]];
4486			s->locked++;
4487			set_bit(R5_LOCKED, &dev->flags);
4488			set_bit(R5_Wantwrite, &dev->flags);
4489		}
4490		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4491			dev = &sh->dev[pd_idx];
4492			s->locked++;
4493			set_bit(R5_LOCKED, &dev->flags);
4494			set_bit(R5_Wantwrite, &dev->flags);
4495		}
4496		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4497			dev = &sh->dev[qd_idx];
4498			s->locked++;
4499			set_bit(R5_LOCKED, &dev->flags);
4500			set_bit(R5_Wantwrite, &dev->flags);
4501		}
4502		if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4503			      "%s: disk%td not up to date\n",
4504			      mdname(conf->mddev),
4505			      dev - (struct r5dev *) &sh->dev)) {
4506			clear_bit(R5_LOCKED, &dev->flags);
4507			clear_bit(R5_Wantwrite, &dev->flags);
4508			s->locked--;
4509		}
4510		clear_bit(STRIPE_DEGRADED, &sh->state);
4511
4512		set_bit(STRIPE_INSYNC, &sh->state);
4513		break;
4514	case check_state_run:
4515	case check_state_run_q:
4516	case check_state_run_pq:
4517		break; /* we will be called again upon completion */
4518	case check_state_check_result:
4519		sh->check_state = check_state_idle;
4520
4521		/* handle a successful check operation, if parity is correct
4522		 * we are done.  Otherwise update the mismatch count and repair
4523		 * parity if !MD_RECOVERY_CHECK
4524		 */
4525		if (sh->ops.zero_sum_result == 0) {
4526			/* both parities are correct */
4527			if (!s->failed)
4528				set_bit(STRIPE_INSYNC, &sh->state);
4529			else {
4530				/* in contrast to the raid5 case we can validate
4531				 * parity, but still have a failure to write
4532				 * back
4533				 */
4534				sh->check_state = check_state_compute_result;
4535				/* Returning at this point means that we may go
4536				 * off and bring p and/or q uptodate again so
4537				 * we make sure to check zero_sum_result again
4538				 * to verify if p or q need writeback
4539				 */
4540			}
4541		} else {
4542			atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4543			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4544				/* don't try to repair!! */
4545				set_bit(STRIPE_INSYNC, &sh->state);
4546				pr_warn_ratelimited("%s: mismatch sector in range "
4547						    "%llu-%llu\n", mdname(conf->mddev),
4548						    (unsigned long long) sh->sector,
4549						    (unsigned long long) sh->sector +
4550						    RAID5_STRIPE_SECTORS(conf));
4551			} else {
4552				int *target = &sh->ops.target;
4553
4554				sh->ops.target = -1;
4555				sh->ops.target2 = -1;
4556				sh->check_state = check_state_compute_run;
4557				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4558				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4559				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4560					set_bit(R5_Wantcompute,
4561						&sh->dev[pd_idx].flags);
4562					*target = pd_idx;
4563					target = &sh->ops.target2;
4564					s->uptodate++;
4565				}
4566				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4567					set_bit(R5_Wantcompute,
4568						&sh->dev[qd_idx].flags);
4569					*target = qd_idx;
4570					s->uptodate++;
4571				}
4572			}
4573		}
4574		break;
4575	case check_state_compute_run:
4576		break;
4577	default:
4578		pr_warn("%s: unknown check_state: %d sector: %llu\n",
4579			__func__, sh->check_state,
4580			(unsigned long long) sh->sector);
4581		BUG();
4582	}
4583}
4584
4585static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4586{
4587	int i;
4588
4589	/* We have read all the blocks in this stripe and now we need to
4590	 * copy some of them into a target stripe for expand.
4591	 */
4592	struct dma_async_tx_descriptor *tx = NULL;
4593	BUG_ON(sh->batch_head);
4594	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4595	for (i = 0; i < sh->disks; i++)
4596		if (i != sh->pd_idx && i != sh->qd_idx) {
4597			int dd_idx, j;
4598			struct stripe_head *sh2;
4599			struct async_submit_ctl submit;
4600
4601			sector_t bn = raid5_compute_blocknr(sh, i, 1);
4602			sector_t s = raid5_compute_sector(conf, bn, 0,
4603							  &dd_idx, NULL);
4604			sh2 = raid5_get_active_stripe(conf, NULL, s,
4605				R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
4606			if (sh2 == NULL)
4607				/* so far only the early blocks of this stripe
4608				 * have been requested.  When later blocks
4609				 * get requested, we will try again
4610				 */
4611				continue;
4612			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4613			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4614				/* must have already done this block */
4615				raid5_release_stripe(sh2);
4616				continue;
4617			}
4618
4619			/* place all the copies on one channel */
4620			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4621			tx = async_memcpy(sh2->dev[dd_idx].page,
4622					  sh->dev[i].page, sh2->dev[dd_idx].offset,
4623					  sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4624					  &submit);
4625
4626			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4627			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4628			for (j = 0; j < conf->raid_disks; j++)
4629				if (j != sh2->pd_idx &&
4630				    j != sh2->qd_idx &&
4631				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
4632					break;
4633			if (j == conf->raid_disks) {
4634				set_bit(STRIPE_EXPAND_READY, &sh2->state);
4635				set_bit(STRIPE_HANDLE, &sh2->state);
4636			}
4637			raid5_release_stripe(sh2);
4638
4639		}
4640	/* done submitting copies, wait for them to complete */
4641	async_tx_quiesce(&tx);
 
 
 
4642}
4643
 
4644/*
4645 * handle_stripe - do things to a stripe.
4646 *
4647 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4648 * state of various bits to see what needs to be done.
4649 * Possible results:
4650 *    return some read requests which now have data
4651 *    return some write requests which are safely on storage
4652 *    schedule a read on some buffers
4653 *    schedule a write of some buffers
4654 *    return confirmation of parity correctness
4655 *
 
 
 
4656 */
4657
4658static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4659{
4660	struct r5conf *conf = sh->raid_conf;
4661	int disks = sh->disks;
4662	struct r5dev *dev;
4663	int i;
4664	int do_recovery = 0;
4665
4666	memset(s, 0, sizeof(*s));
4667
4668	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4669	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
 
4670	s->failed_num[0] = -1;
4671	s->failed_num[1] = -1;
4672	s->log_failed = r5l_log_disk_error(conf);
4673
4674	/* Now to look around and see what can be done */
 
 
4675	for (i=disks; i--; ) {
4676		struct md_rdev *rdev;
4677		sector_t first_bad;
4678		int bad_sectors;
4679		int is_bad = 0;
4680
4681		dev = &sh->dev[i];
4682
4683		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4684			 i, dev->flags,
4685			 dev->toread, dev->towrite, dev->written);
4686		/* maybe we can reply to a read
4687		 *
4688		 * new wantfill requests are only permitted while
4689		 * ops_complete_biofill is guaranteed to be inactive
4690		 */
4691		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4692		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4693			set_bit(R5_Wantfill, &dev->flags);
4694
4695		/* now count some things */
4696		if (test_bit(R5_LOCKED, &dev->flags))
4697			s->locked++;
4698		if (test_bit(R5_UPTODATE, &dev->flags))
4699			s->uptodate++;
4700		if (test_bit(R5_Wantcompute, &dev->flags)) {
4701			s->compute++;
4702			BUG_ON(s->compute > 2);
4703		}
4704
4705		if (test_bit(R5_Wantfill, &dev->flags))
4706			s->to_fill++;
4707		else if (dev->toread)
4708			s->to_read++;
4709		if (dev->towrite) {
4710			s->to_write++;
4711			if (!test_bit(R5_OVERWRITE, &dev->flags))
4712				s->non_overwrite++;
4713		}
4714		if (dev->written)
4715			s->written++;
4716		/* Prefer to use the replacement for reads, but only
4717		 * if it is recovered enough and has no bad blocks.
4718		 */
4719		rdev = conf->disks[i].replacement;
4720		if (rdev && !test_bit(Faulty, &rdev->flags) &&
4721		    rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4722		    !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4723				 &first_bad, &bad_sectors))
4724			set_bit(R5_ReadRepl, &dev->flags);
4725		else {
4726			if (rdev && !test_bit(Faulty, &rdev->flags))
4727				set_bit(R5_NeedReplace, &dev->flags);
4728			else
4729				clear_bit(R5_NeedReplace, &dev->flags);
4730			rdev = conf->disks[i].rdev;
4731			clear_bit(R5_ReadRepl, &dev->flags);
4732		}
4733		if (rdev && test_bit(Faulty, &rdev->flags))
4734			rdev = NULL;
4735		if (rdev) {
4736			is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4737					     &first_bad, &bad_sectors);
4738			if (s->blocked_rdev == NULL
4739			    && (test_bit(Blocked, &rdev->flags)
4740				|| is_bad < 0)) {
4741				if (is_bad < 0)
4742					set_bit(BlockedBadBlocks,
4743						&rdev->flags);
4744				s->blocked_rdev = rdev;
4745				atomic_inc(&rdev->nr_pending);
4746			}
4747		}
4748		clear_bit(R5_Insync, &dev->flags);
4749		if (!rdev)
4750			/* Not in-sync */;
4751		else if (is_bad) {
4752			/* also not in-sync */
4753			if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4754			    test_bit(R5_UPTODATE, &dev->flags)) {
4755				/* treat as in-sync, but with a read error
4756				 * which we can now try to correct
4757				 */
4758				set_bit(R5_Insync, &dev->flags);
4759				set_bit(R5_ReadError, &dev->flags);
4760			}
4761		} else if (test_bit(In_sync, &rdev->flags))
4762			set_bit(R5_Insync, &dev->flags);
4763		else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4764			/* in sync if before recovery_offset */
4765			set_bit(R5_Insync, &dev->flags);
4766		else if (test_bit(R5_UPTODATE, &dev->flags) &&
4767			 test_bit(R5_Expanded, &dev->flags))
4768			/* If we've reshaped into here, we assume it is Insync.
4769			 * We will shortly update recovery_offset to make
4770			 * it official.
4771			 */
4772			set_bit(R5_Insync, &dev->flags);
4773
4774		if (test_bit(R5_WriteError, &dev->flags)) {
4775			/* This flag does not apply to '.replacement'
4776			 * only to .rdev, so make sure to check that*/
4777			struct md_rdev *rdev2 = conf->disks[i].rdev;
4778
4779			if (rdev2 == rdev)
4780				clear_bit(R5_Insync, &dev->flags);
4781			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4782				s->handle_bad_blocks = 1;
4783				atomic_inc(&rdev2->nr_pending);
4784			} else
4785				clear_bit(R5_WriteError, &dev->flags);
4786		}
4787		if (test_bit(R5_MadeGood, &dev->flags)) {
4788			/* This flag does not apply to '.replacement'
4789			 * only to .rdev, so make sure to check that*/
4790			struct md_rdev *rdev2 = conf->disks[i].rdev;
4791
4792			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4793				s->handle_bad_blocks = 1;
4794				atomic_inc(&rdev2->nr_pending);
4795			} else
4796				clear_bit(R5_MadeGood, &dev->flags);
4797		}
4798		if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4799			struct md_rdev *rdev2 = conf->disks[i].replacement;
4800
4801			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4802				s->handle_bad_blocks = 1;
4803				atomic_inc(&rdev2->nr_pending);
4804			} else
4805				clear_bit(R5_MadeGoodRepl, &dev->flags);
4806		}
4807		if (!test_bit(R5_Insync, &dev->flags)) {
4808			/* The ReadError flag will just be confusing now */
4809			clear_bit(R5_ReadError, &dev->flags);
4810			clear_bit(R5_ReWrite, &dev->flags);
4811		}
4812		if (test_bit(R5_ReadError, &dev->flags))
4813			clear_bit(R5_Insync, &dev->flags);
4814		if (!test_bit(R5_Insync, &dev->flags)) {
4815			if (s->failed < 2)
4816				s->failed_num[s->failed] = i;
4817			s->failed++;
4818			if (rdev && !test_bit(Faulty, &rdev->flags))
4819				do_recovery = 1;
4820			else if (!rdev) {
4821				rdev = conf->disks[i].replacement;
4822				if (rdev && !test_bit(Faulty, &rdev->flags))
4823					do_recovery = 1;
4824			}
4825		}
4826
4827		if (test_bit(R5_InJournal, &dev->flags))
4828			s->injournal++;
4829		if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4830			s->just_cached++;
4831	}
4832	if (test_bit(STRIPE_SYNCING, &sh->state)) {
4833		/* If there is a failed device being replaced,
4834		 *     we must be recovering.
4835		 * else if we are after recovery_cp, we must be syncing
4836		 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4837		 * else we can only be replacing
4838		 * sync and recovery both need to read all devices, and so
4839		 * use the same flag.
4840		 */
4841		if (do_recovery ||
4842		    sh->sector >= conf->mddev->recovery_cp ||
4843		    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4844			s->syncing = 1;
4845		else
4846			s->replacing = 1;
4847	}
4848}
4849
4850/*
4851 * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4852 * a head which can now be handled.
4853 */
4854static int clear_batch_ready(struct stripe_head *sh)
4855{
4856	struct stripe_head *tmp;
4857	if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4858		return (sh->batch_head && sh->batch_head != sh);
4859	spin_lock(&sh->stripe_lock);
4860	if (!sh->batch_head) {
4861		spin_unlock(&sh->stripe_lock);
4862		return 0;
4863	}
4864
4865	/*
4866	 * this stripe could be added to a batch list before we check
4867	 * BATCH_READY, skips it
4868	 */
4869	if (sh->batch_head != sh) {
4870		spin_unlock(&sh->stripe_lock);
4871		return 1;
4872	}
4873	spin_lock(&sh->batch_lock);
4874	list_for_each_entry(tmp, &sh->batch_list, batch_list)
4875		clear_bit(STRIPE_BATCH_READY, &tmp->state);
4876	spin_unlock(&sh->batch_lock);
4877	spin_unlock(&sh->stripe_lock);
4878
4879	/*
4880	 * BATCH_READY is cleared, no new stripes can be added.
4881	 * batch_list can be accessed without lock
4882	 */
4883	return 0;
4884}
4885
4886static void break_stripe_batch_list(struct stripe_head *head_sh,
4887				    unsigned long handle_flags)
4888{
4889	struct stripe_head *sh, *next;
4890	int i;
4891	int do_wakeup = 0;
4892
4893	list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4894
4895		list_del_init(&sh->batch_list);
4896
4897		WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4898					  (1 << STRIPE_SYNCING) |
4899					  (1 << STRIPE_REPLACED) |
4900					  (1 << STRIPE_DELAYED) |
4901					  (1 << STRIPE_BIT_DELAY) |
4902					  (1 << STRIPE_FULL_WRITE) |
4903					  (1 << STRIPE_BIOFILL_RUN) |
4904					  (1 << STRIPE_COMPUTE_RUN)  |
4905					  (1 << STRIPE_DISCARD) |
4906					  (1 << STRIPE_BATCH_READY) |
4907					  (1 << STRIPE_BATCH_ERR) |
4908					  (1 << STRIPE_BITMAP_PENDING)),
4909			"stripe state: %lx\n", sh->state);
4910		WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4911					      (1 << STRIPE_REPLACED)),
4912			"head stripe state: %lx\n", head_sh->state);
4913
4914		set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4915					    (1 << STRIPE_PREREAD_ACTIVE) |
4916					    (1 << STRIPE_DEGRADED) |
4917					    (1 << STRIPE_ON_UNPLUG_LIST)),
4918			      head_sh->state & (1 << STRIPE_INSYNC));
4919
4920		sh->check_state = head_sh->check_state;
4921		sh->reconstruct_state = head_sh->reconstruct_state;
4922		spin_lock_irq(&sh->stripe_lock);
4923		sh->batch_head = NULL;
4924		spin_unlock_irq(&sh->stripe_lock);
4925		for (i = 0; i < sh->disks; i++) {
4926			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4927				do_wakeup = 1;
4928			sh->dev[i].flags = head_sh->dev[i].flags &
4929				(~((1 << R5_WriteError) | (1 << R5_Overlap)));
4930		}
4931		if (handle_flags == 0 ||
4932		    sh->state & handle_flags)
4933			set_bit(STRIPE_HANDLE, &sh->state);
4934		raid5_release_stripe(sh);
4935	}
4936	spin_lock_irq(&head_sh->stripe_lock);
4937	head_sh->batch_head = NULL;
4938	spin_unlock_irq(&head_sh->stripe_lock);
4939	for (i = 0; i < head_sh->disks; i++)
4940		if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4941			do_wakeup = 1;
4942	if (head_sh->state & handle_flags)
4943		set_bit(STRIPE_HANDLE, &head_sh->state);
4944
4945	if (do_wakeup)
4946		wake_up(&head_sh->raid_conf->wait_for_overlap);
4947}
4948
4949static void handle_stripe(struct stripe_head *sh)
4950{
4951	struct stripe_head_state s;
4952	struct r5conf *conf = sh->raid_conf;
4953	int i;
4954	int prexor;
4955	int disks = sh->disks;
4956	struct r5dev *pdev, *qdev;
4957
4958	clear_bit(STRIPE_HANDLE, &sh->state);
4959
4960	/*
4961	 * handle_stripe should not continue handle the batched stripe, only
4962	 * the head of batch list or lone stripe can continue. Otherwise we
4963	 * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4964	 * is set for the batched stripe.
4965	 */
4966	if (clear_batch_ready(sh))
4967		return;
4968
4969	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4970		/* already being handled, ensure it gets handled
4971		 * again when current action finishes */
4972		set_bit(STRIPE_HANDLE, &sh->state);
4973		return;
4974	}
4975
4976	if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4977		break_stripe_batch_list(sh, 0);
4978
4979	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4980		spin_lock(&sh->stripe_lock);
4981		/*
4982		 * Cannot process 'sync' concurrently with 'discard'.
4983		 * Flush data in r5cache before 'sync'.
4984		 */
4985		if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4986		    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4987		    !test_bit(STRIPE_DISCARD, &sh->state) &&
4988		    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4989			set_bit(STRIPE_SYNCING, &sh->state);
4990			clear_bit(STRIPE_INSYNC, &sh->state);
4991			clear_bit(STRIPE_REPLACED, &sh->state);
4992		}
4993		spin_unlock(&sh->stripe_lock);
4994	}
4995	clear_bit(STRIPE_DELAYED, &sh->state);
4996
4997	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4998		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4999	       (unsigned long long)sh->sector, sh->state,
5000	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
5001	       sh->check_state, sh->reconstruct_state);
5002
5003	analyse_stripe(sh, &s);
5004
5005	if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
5006		goto finish;
5007
5008	if (s.handle_bad_blocks ||
5009	    test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
5010		set_bit(STRIPE_HANDLE, &sh->state);
5011		goto finish;
5012	}
5013
5014	if (unlikely(s.blocked_rdev)) {
5015		if (s.syncing || s.expanding || s.expanded ||
5016		    s.replacing || s.to_write || s.written) {
5017			set_bit(STRIPE_HANDLE, &sh->state);
5018			goto finish;
5019		}
5020		/* There is nothing for the blocked_rdev to block */
5021		rdev_dec_pending(s.blocked_rdev, conf->mddev);
5022		s.blocked_rdev = NULL;
5023	}
5024
5025	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
5026		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
5027		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
5028	}
5029
5030	pr_debug("locked=%d uptodate=%d to_read=%d"
5031	       " to_write=%d failed=%d failed_num=%d,%d\n",
5032	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
5033	       s.failed_num[0], s.failed_num[1]);
5034	/*
5035	 * check if the array has lost more than max_degraded devices and,
5036	 * if so, some requests might need to be failed.
5037	 *
5038	 * When journal device failed (log_failed), we will only process
5039	 * the stripe if there is data need write to raid disks
5040	 */
5041	if (s.failed > conf->max_degraded ||
5042	    (s.log_failed && s.injournal == 0)) {
5043		sh->check_state = 0;
5044		sh->reconstruct_state = 0;
5045		break_stripe_batch_list(sh, 0);
5046		if (s.to_read+s.to_write+s.written)
5047			handle_failed_stripe(conf, sh, &s, disks);
5048		if (s.syncing + s.replacing)
5049			handle_failed_sync(conf, sh, &s);
5050	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5051
5052	/* Now we check to see if any write operations have recently
5053	 * completed
5054	 */
5055	prexor = 0;
5056	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
5057		prexor = 1;
5058	if (sh->reconstruct_state == reconstruct_state_drain_result ||
5059	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
5060		sh->reconstruct_state = reconstruct_state_idle;
5061
5062		/* All the 'written' buffers and the parity block are ready to
5063		 * be written back to disk
5064		 */
5065		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
5066		       !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
5067		BUG_ON(sh->qd_idx >= 0 &&
5068		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
5069		       !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
5070		for (i = disks; i--; ) {
5071			struct r5dev *dev = &sh->dev[i];
5072			if (test_bit(R5_LOCKED, &dev->flags) &&
5073				(i == sh->pd_idx || i == sh->qd_idx ||
5074				 dev->written || test_bit(R5_InJournal,
5075							  &dev->flags))) {
5076				pr_debug("Writing block %d\n", i);
5077				set_bit(R5_Wantwrite, &dev->flags);
5078				if (prexor)
5079					continue;
5080				if (s.failed > 1)
5081					continue;
5082				if (!test_bit(R5_Insync, &dev->flags) ||
5083				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
5084				     s.failed == 0))
5085					set_bit(STRIPE_INSYNC, &sh->state);
5086			}
5087		}
5088		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5089			s.dec_preread_active = 1;
5090	}
5091
5092	/*
5093	 * might be able to return some write requests if the parity blocks
5094	 * are safe, or on a failed drive
5095	 */
5096	pdev = &sh->dev[sh->pd_idx];
5097	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5098		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5099	qdev = &sh->dev[sh->qd_idx];
5100	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5101		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5102		|| conf->level < 6;
5103
5104	if (s.written &&
5105	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5106			     && !test_bit(R5_LOCKED, &pdev->flags)
5107			     && (test_bit(R5_UPTODATE, &pdev->flags) ||
5108				 test_bit(R5_Discard, &pdev->flags))))) &&
5109	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5110			     && !test_bit(R5_LOCKED, &qdev->flags)
5111			     && (test_bit(R5_UPTODATE, &qdev->flags) ||
5112				 test_bit(R5_Discard, &qdev->flags))))))
5113		handle_stripe_clean_event(conf, sh, disks);
5114
5115	if (s.just_cached)
5116		r5c_handle_cached_data_endio(conf, sh, disks);
5117	log_stripe_write_finished(sh);
5118
5119	/* Now we might consider reading some blocks, either to check/generate
5120	 * parity, or to satisfy requests
5121	 * or to load a block that is being partially written.
5122	 */
5123	if (s.to_read || s.non_overwrite
5124	    || (s.to_write && s.failed)
5125	    || (s.syncing && (s.uptodate + s.compute < disks))
5126	    || s.replacing
5127	    || s.expanding)
5128		handle_stripe_fill(sh, &s, disks);
5129
5130	/*
5131	 * When the stripe finishes full journal write cycle (write to journal
5132	 * and raid disk), this is the clean up procedure so it is ready for
5133	 * next operation.
5134	 */
5135	r5c_finish_stripe_write_out(conf, sh, &s);
5136
5137	/*
5138	 * Now to consider new write requests, cache write back and what else,
5139	 * if anything should be read.  We do not handle new writes when:
5140	 * 1/ A 'write' operation (copy+xor) is already in flight.
5141	 * 2/ A 'check' operation is in flight, as it may clobber the parity
5142	 *    block.
5143	 * 3/ A r5c cache log write is in flight.
5144	 */
5145
5146	if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5147		if (!r5c_is_writeback(conf->log)) {
5148			if (s.to_write)
5149				handle_stripe_dirtying(conf, sh, &s, disks);
5150		} else { /* write back cache */
5151			int ret = 0;
5152
5153			/* First, try handle writes in caching phase */
5154			if (s.to_write)
5155				ret = r5c_try_caching_write(conf, sh, &s,
5156							    disks);
5157			/*
5158			 * If caching phase failed: ret == -EAGAIN
5159			 *    OR
5160			 * stripe under reclaim: !caching && injournal
5161			 *
5162			 * fall back to handle_stripe_dirtying()
5163			 */
5164			if (ret == -EAGAIN ||
5165			    /* stripe under reclaim: !caching && injournal */
5166			    (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5167			     s.injournal > 0)) {
5168				ret = handle_stripe_dirtying(conf, sh, &s,
5169							     disks);
5170				if (ret == -EAGAIN)
5171					goto finish;
5172			}
5173		}
5174	}
5175
5176	/* maybe we need to check and possibly fix the parity for this stripe
5177	 * Any reads will already have been scheduled, so we just see if enough
5178	 * data is available.  The parity check is held off while parity
5179	 * dependent operations are in flight.
5180	 */
5181	if (sh->check_state ||
5182	    (s.syncing && s.locked == 0 &&
5183	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5184	     !test_bit(STRIPE_INSYNC, &sh->state))) {
5185		if (conf->level == 6)
5186			handle_parity_checks6(conf, sh, &s, disks);
5187		else
5188			handle_parity_checks5(conf, sh, &s, disks);
5189	}
5190
5191	if ((s.replacing || s.syncing) && s.locked == 0
5192	    && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5193	    && !test_bit(STRIPE_REPLACED, &sh->state)) {
5194		/* Write out to replacement devices where possible */
5195		for (i = 0; i < conf->raid_disks; i++)
5196			if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5197				WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5198				set_bit(R5_WantReplace, &sh->dev[i].flags);
5199				set_bit(R5_LOCKED, &sh->dev[i].flags);
5200				s.locked++;
5201			}
5202		if (s.replacing)
5203			set_bit(STRIPE_INSYNC, &sh->state);
5204		set_bit(STRIPE_REPLACED, &sh->state);
5205	}
5206	if ((s.syncing || s.replacing) && s.locked == 0 &&
5207	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5208	    test_bit(STRIPE_INSYNC, &sh->state)) {
5209		md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5210		clear_bit(STRIPE_SYNCING, &sh->state);
5211		if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5212			wake_up(&conf->wait_for_overlap);
5213	}
5214
5215	/* If the failed drives are just a ReadError, then we might need
5216	 * to progress the repair/check process
5217	 */
5218	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5219		for (i = 0; i < s.failed; i++) {
5220			struct r5dev *dev = &sh->dev[s.failed_num[i]];
5221			if (test_bit(R5_ReadError, &dev->flags)
5222			    && !test_bit(R5_LOCKED, &dev->flags)
5223			    && test_bit(R5_UPTODATE, &dev->flags)
5224				) {
5225				if (!test_bit(R5_ReWrite, &dev->flags)) {
5226					set_bit(R5_Wantwrite, &dev->flags);
5227					set_bit(R5_ReWrite, &dev->flags);
5228				} else
 
 
5229					/* let's read it back */
5230					set_bit(R5_Wantread, &dev->flags);
5231				set_bit(R5_LOCKED, &dev->flags);
5232				s.locked++;
 
5233			}
5234		}
5235
 
5236	/* Finish reconstruct operations initiated by the expansion process */
5237	if (sh->reconstruct_state == reconstruct_state_result) {
5238		struct stripe_head *sh_src
5239			= raid5_get_active_stripe(conf, NULL, sh->sector,
5240					R5_GAS_PREVIOUS | R5_GAS_NOBLOCK |
5241					R5_GAS_NOQUIESCE);
5242		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5243			/* sh cannot be written until sh_src has been read.
5244			 * so arrange for sh to be delayed a little
5245			 */
5246			set_bit(STRIPE_DELAYED, &sh->state);
5247			set_bit(STRIPE_HANDLE, &sh->state);
5248			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5249					      &sh_src->state))
5250				atomic_inc(&conf->preread_active_stripes);
5251			raid5_release_stripe(sh_src);
5252			goto finish;
5253		}
5254		if (sh_src)
5255			raid5_release_stripe(sh_src);
5256
5257		sh->reconstruct_state = reconstruct_state_idle;
5258		clear_bit(STRIPE_EXPANDING, &sh->state);
5259		for (i = conf->raid_disks; i--; ) {
5260			set_bit(R5_Wantwrite, &sh->dev[i].flags);
5261			set_bit(R5_LOCKED, &sh->dev[i].flags);
5262			s.locked++;
5263		}
5264	}
5265
5266	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5267	    !sh->reconstruct_state) {
5268		/* Need to write out all blocks after computing parity */
5269		sh->disks = conf->raid_disks;
5270		stripe_set_idx(sh->sector, conf, 0, sh);
5271		schedule_reconstruction(sh, &s, 1, 1);
5272	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5273		clear_bit(STRIPE_EXPAND_READY, &sh->state);
5274		atomic_dec(&conf->reshape_stripes);
5275		wake_up(&conf->wait_for_overlap);
5276		md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5277	}
5278
5279	if (s.expanding && s.locked == 0 &&
5280	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5281		handle_stripe_expansion(conf, sh);
5282
5283finish:
5284	/* wait for this device to become unblocked */
5285	if (unlikely(s.blocked_rdev)) {
5286		if (conf->mddev->external)
5287			md_wait_for_blocked_rdev(s.blocked_rdev,
5288						 conf->mddev);
5289		else
5290			/* Internal metadata will immediately
5291			 * be written by raid5d, so we don't
5292			 * need to wait here.
5293			 */
5294			rdev_dec_pending(s.blocked_rdev,
5295					 conf->mddev);
5296	}
5297
5298	if (s.handle_bad_blocks)
5299		for (i = disks; i--; ) {
5300			struct md_rdev *rdev;
5301			struct r5dev *dev = &sh->dev[i];
5302			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5303				/* We own a safe reference to the rdev */
5304				rdev = conf->disks[i].rdev;
5305				if (!rdev_set_badblocks(rdev, sh->sector,
5306							RAID5_STRIPE_SECTORS(conf), 0))
5307					md_error(conf->mddev, rdev);
5308				rdev_dec_pending(rdev, conf->mddev);
5309			}
5310			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5311				rdev = conf->disks[i].rdev;
5312				rdev_clear_badblocks(rdev, sh->sector,
5313						     RAID5_STRIPE_SECTORS(conf), 0);
5314				rdev_dec_pending(rdev, conf->mddev);
5315			}
5316			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5317				rdev = conf->disks[i].replacement;
5318				if (!rdev)
5319					/* rdev have been moved down */
5320					rdev = conf->disks[i].rdev;
5321				rdev_clear_badblocks(rdev, sh->sector,
5322						     RAID5_STRIPE_SECTORS(conf), 0);
5323				rdev_dec_pending(rdev, conf->mddev);
5324			}
5325		}
5326
5327	if (s.ops_request)
5328		raid_run_ops(sh, s.ops_request);
5329
5330	ops_run_io(sh, &s);
5331
5332	if (s.dec_preread_active) {
5333		/* We delay this until after ops_run_io so that if make_request
5334		 * is waiting on a flush, it won't continue until the writes
5335		 * have actually been submitted.
5336		 */
5337		atomic_dec(&conf->preread_active_stripes);
5338		if (atomic_read(&conf->preread_active_stripes) <
5339		    IO_THRESHOLD)
5340			md_wakeup_thread(conf->mddev->thread);
5341	}
5342
5343	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
 
 
5344}
5345
5346static void raid5_activate_delayed(struct r5conf *conf)
5347	__must_hold(&conf->device_lock)
5348{
5349	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5350		while (!list_empty(&conf->delayed_list)) {
5351			struct list_head *l = conf->delayed_list.next;
5352			struct stripe_head *sh;
5353			sh = list_entry(l, struct stripe_head, lru);
5354			list_del_init(l);
5355			clear_bit(STRIPE_DELAYED, &sh->state);
5356			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5357				atomic_inc(&conf->preread_active_stripes);
5358			list_add_tail(&sh->lru, &conf->hold_list);
5359			raid5_wakeup_stripe_thread(sh);
5360		}
5361	}
5362}
5363
5364static void activate_bit_delay(struct r5conf *conf,
5365		struct list_head *temp_inactive_list)
5366	__must_hold(&conf->device_lock)
5367{
 
5368	struct list_head head;
5369	list_add(&head, &conf->bitmap_list);
5370	list_del_init(&conf->bitmap_list);
5371	while (!list_empty(&head)) {
5372		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5373		int hash;
5374		list_del_init(&sh->lru);
5375		atomic_inc(&sh->count);
5376		hash = sh->hash_lock_index;
5377		__release_stripe(conf, sh, &temp_inactive_list[hash]);
5378	}
5379}
5380
5381static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5382{
5383	struct r5conf *conf = mddev->private;
5384	sector_t sector = bio->bi_iter.bi_sector;
5385	unsigned int chunk_sectors;
5386	unsigned int bio_sectors = bio_sectors(bio);
5387
5388	chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5389	return  chunk_sectors >=
5390		((sector & (chunk_sectors - 1)) + bio_sectors);
5391}
5392
5393/*
5394 *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5395 *  later sampled by raid5d.
5396 */
5397static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5398{
5399	unsigned long flags;
5400
5401	spin_lock_irqsave(&conf->device_lock, flags);
5402
5403	bi->bi_next = conf->retry_read_aligned_list;
5404	conf->retry_read_aligned_list = bi;
5405
5406	spin_unlock_irqrestore(&conf->device_lock, flags);
5407	md_wakeup_thread(conf->mddev->thread);
5408}
5409
5410static struct bio *remove_bio_from_retry(struct r5conf *conf,
5411					 unsigned int *offset)
5412{
5413	struct bio *bi;
5414
5415	bi = conf->retry_read_aligned;
5416	if (bi) {
5417		*offset = conf->retry_read_offset;
5418		conf->retry_read_aligned = NULL;
5419		return bi;
5420	}
5421	bi = conf->retry_read_aligned_list;
5422	if(bi) {
5423		conf->retry_read_aligned_list = bi->bi_next;
5424		bi->bi_next = NULL;
5425		*offset = 0;
 
 
 
 
5426	}
5427
5428	return bi;
5429}
5430
 
5431/*
5432 *  The "raid5_align_endio" should check if the read succeeded and if it
5433 *  did, call bio_endio on the original bio (having bio_put the new bio
5434 *  first).
5435 *  If the read failed..
5436 */
5437static void raid5_align_endio(struct bio *bi)
5438{
5439	struct bio *raid_bi = bi->bi_private;
5440	struct md_rdev *rdev = (void *)raid_bi->bi_next;
5441	struct mddev *mddev = rdev->mddev;
5442	struct r5conf *conf = mddev->private;
5443	blk_status_t error = bi->bi_status;
5444
5445	bio_put(bi);
 
 
5446	raid_bi->bi_next = NULL;
 
 
 
5447	rdev_dec_pending(rdev, conf->mddev);
5448
5449	if (!error) {
5450		bio_endio(raid_bi);
5451		if (atomic_dec_and_test(&conf->active_aligned_reads))
5452			wake_up(&conf->wait_for_quiescent);
5453		return;
5454	}
5455
 
5456	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5457
5458	add_bio_to_retry(raid_bi, conf);
5459}
5460
5461static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5462{
5463	struct r5conf *conf = mddev->private;
5464	struct bio *align_bio;
5465	struct md_rdev *rdev;
5466	sector_t sector, end_sector, first_bad;
5467	int bad_sectors, dd_idx;
5468	bool did_inc;
5469
5470	if (!in_chunk_boundary(mddev, raid_bio)) {
5471		pr_debug("%s: non aligned\n", __func__);
 
 
5472		return 0;
5473	}
5474
5475	sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5476				      &dd_idx, NULL);
5477	end_sector = sector + bio_sectors(raid_bio);
5478
5479	if (r5c_big_stripe_cached(conf, sector))
 
 
 
5480		return 0;
5481
5482	rdev = conf->disks[dd_idx].replacement;
5483	if (!rdev || test_bit(Faulty, &rdev->flags) ||
5484	    rdev->recovery_offset < end_sector) {
5485		rdev = conf->disks[dd_idx].rdev;
5486		if (!rdev)
5487			return 0;
5488		if (test_bit(Faulty, &rdev->flags) ||
5489		    !(test_bit(In_sync, &rdev->flags) ||
5490		      rdev->recovery_offset >= end_sector))
5491			return 0;
5492	}
5493
5494	atomic_inc(&rdev->nr_pending);
5495
5496	if (is_badblock(rdev, sector, bio_sectors(raid_bio), &first_bad,
5497			&bad_sectors)) {
5498		rdev_dec_pending(rdev, mddev);
 
 
 
 
 
 
5499		return 0;
5500	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5501
5502	md_account_bio(mddev, &raid_bio);
5503	raid_bio->bi_next = (void *)rdev;
5504
5505	align_bio = bio_alloc_clone(rdev->bdev, raid_bio, GFP_NOIO,
5506				    &mddev->bio_set);
5507	align_bio->bi_end_io = raid5_align_endio;
5508	align_bio->bi_private = raid_bio;
5509	align_bio->bi_iter.bi_sector = sector;
5510
5511	/* No reshape active, so we can trust rdev->data_offset */
5512	align_bio->bi_iter.bi_sector += rdev->data_offset;
 
 
 
 
 
 
 
 
 
 
 
 
 
5513
5514	did_inc = false;
5515	if (conf->quiesce == 0) {
5516		atomic_inc(&conf->active_aligned_reads);
5517		did_inc = true;
5518	}
5519	/* need a memory barrier to detect the race with raid5_quiesce() */
5520	if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5521		/* quiesce is in progress, so we need to undo io activation and wait
5522		 * for it to finish
5523		 */
5524		if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5525			wake_up(&conf->wait_for_quiescent);
5526		spin_lock_irq(&conf->device_lock);
5527		wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5528				    conf->device_lock);
 
5529		atomic_inc(&conf->active_aligned_reads);
5530		spin_unlock_irq(&conf->device_lock);
5531	}
5532
5533	if (mddev->gendisk)
5534		trace_block_bio_remap(align_bio, disk_devt(mddev->gendisk),
5535				      raid_bio->bi_iter.bi_sector);
5536	submit_bio_noacct(align_bio);
5537	return 1;
5538}
5539
5540static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5541{
5542	struct bio *split;
5543	sector_t sector = raid_bio->bi_iter.bi_sector;
5544	unsigned chunk_sects = mddev->chunk_sectors;
5545	unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5546
5547	if (sectors < bio_sectors(raid_bio)) {
5548		struct r5conf *conf = mddev->private;
5549		split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5550		bio_chain(split, raid_bio);
5551		submit_bio_noacct(raid_bio);
5552		raid_bio = split;
5553	}
5554
5555	if (!raid5_read_one_chunk(mddev, raid_bio))
5556		return raid_bio;
5557
5558	return NULL;
5559}
5560
5561/* __get_priority_stripe - get the next stripe to process
5562 *
5563 * Full stripe writes are allowed to pass preread active stripes up until
5564 * the bypass_threshold is exceeded.  In general the bypass_count
5565 * increments when the handle_list is handled before the hold_list; however, it
5566 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5567 * stripe with in flight i/o.  The bypass_count will be reset when the
5568 * head of the hold_list has changed, i.e. the head was promoted to the
5569 * handle_list.
5570 */
5571static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5572	__must_hold(&conf->device_lock)
5573{
5574	struct stripe_head *sh, *tmp;
5575	struct list_head *handle_list = NULL;
5576	struct r5worker_group *wg;
5577	bool second_try = !r5c_is_writeback(conf->log) &&
5578		!r5l_log_disk_error(conf);
5579	bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5580		r5l_log_disk_error(conf);
5581
5582again:
5583	wg = NULL;
5584	sh = NULL;
5585	if (conf->worker_cnt_per_group == 0) {
5586		handle_list = try_loprio ? &conf->loprio_list :
5587					&conf->handle_list;
5588	} else if (group != ANY_GROUP) {
5589		handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5590				&conf->worker_groups[group].handle_list;
5591		wg = &conf->worker_groups[group];
5592	} else {
5593		int i;
5594		for (i = 0; i < conf->group_cnt; i++) {
5595			handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5596				&conf->worker_groups[i].handle_list;
5597			wg = &conf->worker_groups[i];
5598			if (!list_empty(handle_list))
5599				break;
5600		}
5601	}
5602
5603	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5604		  __func__,
5605		  list_empty(handle_list) ? "empty" : "busy",
5606		  list_empty(&conf->hold_list) ? "empty" : "busy",
5607		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
5608
5609	if (!list_empty(handle_list)) {
5610		sh = list_entry(handle_list->next, typeof(*sh), lru);
5611
5612		if (list_empty(&conf->hold_list))
5613			conf->bypass_count = 0;
5614		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5615			if (conf->hold_list.next == conf->last_hold)
5616				conf->bypass_count++;
5617			else {
5618				conf->last_hold = conf->hold_list.next;
5619				conf->bypass_count -= conf->bypass_threshold;
5620				if (conf->bypass_count < 0)
5621					conf->bypass_count = 0;
5622			}
5623		}
5624	} else if (!list_empty(&conf->hold_list) &&
5625		   ((conf->bypass_threshold &&
5626		     conf->bypass_count > conf->bypass_threshold) ||
5627		    atomic_read(&conf->pending_full_writes) == 0)) {
 
 
 
 
 
 
 
5628
5629		list_for_each_entry(tmp, &conf->hold_list,  lru) {
5630			if (conf->worker_cnt_per_group == 0 ||
5631			    group == ANY_GROUP ||
5632			    !cpu_online(tmp->cpu) ||
5633			    cpu_to_group(tmp->cpu) == group) {
5634				sh = tmp;
5635				break;
5636			}
5637		}
5638
5639		if (sh) {
5640			conf->bypass_count -= conf->bypass_threshold;
5641			if (conf->bypass_count < 0)
5642				conf->bypass_count = 0;
5643		}
5644		wg = NULL;
5645	}
5646
5647	if (!sh) {
5648		if (second_try)
5649			return NULL;
5650		second_try = true;
5651		try_loprio = !try_loprio;
5652		goto again;
5653	}
5654
5655	if (wg) {
5656		wg->stripes_cnt--;
5657		sh->group = NULL;
5658	}
5659	list_del_init(&sh->lru);
5660	BUG_ON(atomic_inc_return(&sh->count) != 1);
 
5661	return sh;
5662}
5663
5664struct raid5_plug_cb {
5665	struct blk_plug_cb	cb;
5666	struct list_head	list;
5667	struct list_head	temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5668};
5669
5670static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5671{
5672	struct raid5_plug_cb *cb = container_of(
5673		blk_cb, struct raid5_plug_cb, cb);
5674	struct stripe_head *sh;
5675	struct mddev *mddev = cb->cb.data;
5676	struct r5conf *conf = mddev->private;
5677	int cnt = 0;
5678	int hash;
5679
5680	if (cb->list.next && !list_empty(&cb->list)) {
5681		spin_lock_irq(&conf->device_lock);
5682		while (!list_empty(&cb->list)) {
5683			sh = list_first_entry(&cb->list, struct stripe_head, lru);
5684			list_del_init(&sh->lru);
5685			/*
5686			 * avoid race release_stripe_plug() sees
5687			 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5688			 * is still in our list
5689			 */
5690			smp_mb__before_atomic();
5691			clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5692			/*
5693			 * STRIPE_ON_RELEASE_LIST could be set here. In that
5694			 * case, the count is always > 1 here
5695			 */
5696			hash = sh->hash_lock_index;
5697			__release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5698			cnt++;
5699		}
5700		spin_unlock_irq(&conf->device_lock);
5701	}
5702	release_inactive_stripe_list(conf, cb->temp_inactive_list,
5703				     NR_STRIPE_HASH_LOCKS);
5704	if (mddev->queue)
5705		trace_block_unplug(mddev->queue, cnt, !from_schedule);
5706	kfree(cb);
5707}
5708
5709static void release_stripe_plug(struct mddev *mddev,
5710				struct stripe_head *sh)
5711{
5712	struct blk_plug_cb *blk_cb = blk_check_plugged(
5713		raid5_unplug, mddev,
5714		sizeof(struct raid5_plug_cb));
5715	struct raid5_plug_cb *cb;
5716
5717	if (!blk_cb) {
5718		raid5_release_stripe(sh);
5719		return;
5720	}
5721
5722	cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5723
5724	if (cb->list.next == NULL) {
5725		int i;
5726		INIT_LIST_HEAD(&cb->list);
5727		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5728			INIT_LIST_HEAD(cb->temp_inactive_list + i);
5729	}
5730
5731	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5732		list_add_tail(&sh->lru, &cb->list);
5733	else
5734		raid5_release_stripe(sh);
5735}
5736
5737static void make_discard_request(struct mddev *mddev, struct bio *bi)
5738{
5739	struct r5conf *conf = mddev->private;
 
 
5740	sector_t logical_sector, last_sector;
5741	struct stripe_head *sh;
5742	int stripe_sectors;
 
 
5743
5744	/* We need to handle this when io_uring supports discard/trim */
5745	if (WARN_ON_ONCE(bi->bi_opf & REQ_NOWAIT))
5746		return;
 
5747
5748	if (mddev->reshape_position != MaxSector)
5749		/* Skip discard while reshape is happening */
5750		return;
5751
5752	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5753	last_sector = bio_end_sector(bi);
 
 
5754
 
 
5755	bi->bi_next = NULL;
 
5756
5757	stripe_sectors = conf->chunk_sectors *
5758		(conf->raid_disks - conf->max_degraded);
5759	logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5760					       stripe_sectors);
5761	sector_div(last_sector, stripe_sectors);
5762
5763	logical_sector *= conf->chunk_sectors;
5764	last_sector *= conf->chunk_sectors;
5765
5766	for (; logical_sector < last_sector;
5767	     logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5768		DEFINE_WAIT(w);
5769		int d;
5770	again:
5771		sh = raid5_get_active_stripe(conf, NULL, logical_sector, 0);
5772		prepare_to_wait(&conf->wait_for_overlap, &w,
5773				TASK_UNINTERRUPTIBLE);
5774		set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5775		if (test_bit(STRIPE_SYNCING, &sh->state)) {
5776			raid5_release_stripe(sh);
5777			schedule();
5778			goto again;
5779		}
5780		clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5781		spin_lock_irq(&sh->stripe_lock);
5782		for (d = 0; d < conf->raid_disks; d++) {
5783			if (d == sh->pd_idx || d == sh->qd_idx)
5784				continue;
5785			if (sh->dev[d].towrite || sh->dev[d].toread) {
5786				set_bit(R5_Overlap, &sh->dev[d].flags);
5787				spin_unlock_irq(&sh->stripe_lock);
5788				raid5_release_stripe(sh);
5789				schedule();
5790				goto again;
5791			}
5792		}
5793		set_bit(STRIPE_DISCARD, &sh->state);
5794		finish_wait(&conf->wait_for_overlap, &w);
5795		sh->overwrite_disks = 0;
5796		for (d = 0; d < conf->raid_disks; d++) {
5797			if (d == sh->pd_idx || d == sh->qd_idx)
5798				continue;
5799			sh->dev[d].towrite = bi;
5800			set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5801			bio_inc_remaining(bi);
5802			md_write_inc(mddev, bi);
5803			sh->overwrite_disks++;
5804		}
5805		spin_unlock_irq(&sh->stripe_lock);
5806		if (conf->mddev->bitmap) {
5807			for (d = 0;
5808			     d < conf->raid_disks - conf->max_degraded;
5809			     d++)
5810				md_bitmap_startwrite(mddev->bitmap,
5811						     sh->sector,
5812						     RAID5_STRIPE_SECTORS(conf),
5813						     0);
5814			sh->bm_seq = conf->seq_flush + 1;
5815			set_bit(STRIPE_BIT_DELAY, &sh->state);
5816		}
5817
5818		set_bit(STRIPE_HANDLE, &sh->state);
5819		clear_bit(STRIPE_DELAYED, &sh->state);
5820		if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5821			atomic_inc(&conf->preread_active_stripes);
5822		release_stripe_plug(mddev, sh);
5823	}
5824
5825	bio_endio(bi);
5826}
5827
5828static bool ahead_of_reshape(struct mddev *mddev, sector_t sector,
5829			     sector_t reshape_sector)
5830{
5831	return mddev->reshape_backwards ? sector < reshape_sector :
5832					  sector >= reshape_sector;
5833}
5834
5835static bool range_ahead_of_reshape(struct mddev *mddev, sector_t min,
5836				   sector_t max, sector_t reshape_sector)
5837{
5838	return mddev->reshape_backwards ? max < reshape_sector :
5839					  min >= reshape_sector;
5840}
5841
5842static bool stripe_ahead_of_reshape(struct mddev *mddev, struct r5conf *conf,
5843				    struct stripe_head *sh)
5844{
5845	sector_t max_sector = 0, min_sector = MaxSector;
5846	bool ret = false;
5847	int dd_idx;
5848
5849	for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5850		if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5851			continue;
5852
5853		min_sector = min(min_sector, sh->dev[dd_idx].sector);
5854		max_sector = max(max_sector, sh->dev[dd_idx].sector);
5855	}
5856
5857	spin_lock_irq(&conf->device_lock);
5858
5859	if (!range_ahead_of_reshape(mddev, min_sector, max_sector,
5860				     conf->reshape_progress))
5861		/* mismatch, need to try again */
5862		ret = true;
5863
5864	spin_unlock_irq(&conf->device_lock);
5865
5866	return ret;
5867}
5868
5869static int add_all_stripe_bios(struct r5conf *conf,
5870		struct stripe_request_ctx *ctx, struct stripe_head *sh,
5871		struct bio *bi, int forwrite, int previous)
5872{
5873	int dd_idx;
5874	int ret = 1;
5875
5876	spin_lock_irq(&sh->stripe_lock);
5877
5878	for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5879		struct r5dev *dev = &sh->dev[dd_idx];
5880
5881		if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5882			continue;
5883
5884		if (dev->sector < ctx->first_sector ||
5885		    dev->sector >= ctx->last_sector)
5886			continue;
5887
5888		if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
5889			set_bit(R5_Overlap, &dev->flags);
5890			ret = 0;
5891			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5892		}
5893	}
5894
5895	if (!ret)
5896		goto out;
5897
5898	for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5899		struct r5dev *dev = &sh->dev[dd_idx];
5900
5901		if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5902			continue;
5903
5904		if (dev->sector < ctx->first_sector ||
5905		    dev->sector >= ctx->last_sector)
5906			continue;
5907
5908		__add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
5909		clear_bit((dev->sector - ctx->first_sector) >>
5910			  RAID5_STRIPE_SHIFT(conf), ctx->sectors_to_do);
5911	}
5912
5913out:
5914	spin_unlock_irq(&sh->stripe_lock);
5915	return ret;
5916}
5917
5918static enum stripe_result make_stripe_request(struct mddev *mddev,
5919		struct r5conf *conf, struct stripe_request_ctx *ctx,
5920		sector_t logical_sector, struct bio *bi)
5921{
5922	const int rw = bio_data_dir(bi);
5923	enum stripe_result ret;
5924	struct stripe_head *sh;
5925	sector_t new_sector;
5926	int previous = 0, flags = 0;
5927	int seq, dd_idx;
5928
5929	seq = read_seqcount_begin(&conf->gen_lock);
 
 
 
 
 
5930
5931	if (unlikely(conf->reshape_progress != MaxSector)) {
5932		/*
5933		 * Spinlock is needed as reshape_progress may be
5934		 * 64bit on a 32bit platform, and so it might be
5935		 * possible to see a half-updated value
5936		 * Of course reshape_progress could change after
5937		 * the lock is dropped, so once we get a reference
5938		 * to the stripe that we think it is, we will have
5939		 * to check again.
5940		 */
5941		spin_lock_irq(&conf->device_lock);
5942		if (ahead_of_reshape(mddev, logical_sector,
5943				     conf->reshape_progress)) {
5944			previous = 1;
5945		} else {
5946			if (ahead_of_reshape(mddev, logical_sector,
5947					     conf->reshape_safe)) {
 
 
5948				spin_unlock_irq(&conf->device_lock);
5949				return STRIPE_SCHEDULE_AND_RETRY;
 
 
 
 
5950			}
5951		}
5952		spin_unlock_irq(&conf->device_lock);
5953	}
5954
5955	new_sector = raid5_compute_sector(conf, logical_sector, previous,
5956					  &dd_idx, NULL);
5957	pr_debug("raid456: %s, sector %llu logical %llu\n", __func__,
5958		 new_sector, logical_sector);
5959
5960	if (previous)
5961		flags |= R5_GAS_PREVIOUS;
5962	if (bi->bi_opf & REQ_RAHEAD)
5963		flags |= R5_GAS_NOBLOCK;
5964	sh = raid5_get_active_stripe(conf, ctx, new_sector, flags);
5965	if (unlikely(!sh)) {
5966		/* cannot get stripe, just give-up */
5967		bi->bi_status = BLK_STS_IOERR;
5968		return STRIPE_FAIL;
5969	}
5970
5971	if (unlikely(previous) &&
5972	    stripe_ahead_of_reshape(mddev, conf, sh)) {
5973		/*
5974		 * Expansion moved on while waiting for a stripe.
5975		 * Expansion could still move past after this
5976		 * test, but as we are holding a reference to
5977		 * 'sh', we know that if that happens,
5978		 *  STRIPE_EXPANDING will get set and the expansion
5979		 * won't proceed until we finish with the stripe.
5980		 */
5981		ret = STRIPE_SCHEDULE_AND_RETRY;
5982		goto out_release;
5983	}
5984
5985	if (read_seqcount_retry(&conf->gen_lock, seq)) {
5986		/* Might have got the wrong stripe_head by accident */
5987		ret = STRIPE_RETRY;
5988		goto out_release;
5989	}
5990
5991	if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5992	    !add_all_stripe_bios(conf, ctx, sh, bi, rw, previous)) {
5993		/*
5994		 * Stripe is busy expanding or add failed due to
5995		 * overlap. Flush everything and wait a while.
5996		 */
5997		md_wakeup_thread(mddev->thread);
5998		ret = STRIPE_SCHEDULE_AND_RETRY;
5999		goto out_release;
6000	}
6001
6002	if (stripe_can_batch(sh)) {
6003		stripe_add_to_batch_list(conf, sh, ctx->batch_last);
6004		if (ctx->batch_last)
6005			raid5_release_stripe(ctx->batch_last);
6006		atomic_inc(&sh->count);
6007		ctx->batch_last = sh;
6008	}
6009
6010	if (ctx->do_flush) {
6011		set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
6012		/* we only need flush for one stripe */
6013		ctx->do_flush = false;
6014	}
6015
6016	set_bit(STRIPE_HANDLE, &sh->state);
6017	clear_bit(STRIPE_DELAYED, &sh->state);
6018	if ((!sh->batch_head || sh == sh->batch_head) &&
6019	    (bi->bi_opf & REQ_SYNC) &&
6020	    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
6021		atomic_inc(&conf->preread_active_stripes);
6022
6023	release_stripe_plug(mddev, sh);
6024	return STRIPE_SUCCESS;
6025
6026out_release:
6027	raid5_release_stripe(sh);
6028	return ret;
6029}
6030
6031/*
6032 * If the bio covers multiple data disks, find sector within the bio that has
6033 * the lowest chunk offset in the first chunk.
6034 */
6035static sector_t raid5_bio_lowest_chunk_sector(struct r5conf *conf,
6036					      struct bio *bi)
6037{
6038	int sectors_per_chunk = conf->chunk_sectors;
6039	int raid_disks = conf->raid_disks;
6040	int dd_idx;
6041	struct stripe_head sh;
6042	unsigned int chunk_offset;
6043	sector_t r_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6044	sector_t sector;
6045
6046	/* We pass in fake stripe_head to get back parity disk numbers */
6047	sector = raid5_compute_sector(conf, r_sector, 0, &dd_idx, &sh);
6048	chunk_offset = sector_div(sector, sectors_per_chunk);
6049	if (sectors_per_chunk - chunk_offset >= bio_sectors(bi))
6050		return r_sector;
6051	/*
6052	 * Bio crosses to the next data disk. Check whether it's in the same
6053	 * chunk.
6054	 */
6055	dd_idx++;
6056	while (dd_idx == sh.pd_idx || dd_idx == sh.qd_idx)
6057		dd_idx++;
6058	if (dd_idx >= raid_disks)
6059		return r_sector;
6060	return r_sector + sectors_per_chunk - chunk_offset;
6061}
6062
6063static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
6064{
6065	DEFINE_WAIT_FUNC(wait, woken_wake_function);
6066	struct r5conf *conf = mddev->private;
6067	sector_t logical_sector;
6068	struct stripe_request_ctx ctx = {};
6069	const int rw = bio_data_dir(bi);
6070	enum stripe_result res;
6071	int s, stripe_cnt;
6072
6073	if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
6074		int ret = log_handle_flush_request(conf, bi);
6075
6076		if (ret == 0)
6077			return true;
6078		if (ret == -ENODEV) {
6079			if (md_flush_request(mddev, bi))
6080				return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6081		}
6082		/* ret == -EAGAIN, fallback */
6083		/*
6084		 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
6085		 * we need to flush journal device
6086		 */
6087		ctx.do_flush = bi->bi_opf & REQ_PREFLUSH;
6088	}
 
 
6089
6090	if (!md_write_start(mddev, bi))
6091		return false;
6092	/*
6093	 * If array is degraded, better not do chunk aligned read because
6094	 * later we might have to read it again in order to reconstruct
6095	 * data on failed drives.
6096	 */
6097	if (rw == READ && mddev->degraded == 0 &&
6098	    mddev->reshape_position == MaxSector) {
6099		bi = chunk_aligned_read(mddev, bi);
6100		if (!bi)
6101			return true;
6102	}
6103
6104	if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
6105		make_discard_request(mddev, bi);
6106		md_write_end(mddev);
6107		return true;
6108	}
6109
6110	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6111	ctx.first_sector = logical_sector;
6112	ctx.last_sector = bio_end_sector(bi);
6113	bi->bi_next = NULL;
6114
6115	stripe_cnt = DIV_ROUND_UP_SECTOR_T(ctx.last_sector - logical_sector,
6116					   RAID5_STRIPE_SECTORS(conf));
6117	bitmap_set(ctx.sectors_to_do, 0, stripe_cnt);
6118
6119	pr_debug("raid456: %s, logical %llu to %llu\n", __func__,
6120		 bi->bi_iter.bi_sector, ctx.last_sector);
6121
6122	/* Bail out if conflicts with reshape and REQ_NOWAIT is set */
6123	if ((bi->bi_opf & REQ_NOWAIT) &&
6124	    (conf->reshape_progress != MaxSector) &&
6125	    !ahead_of_reshape(mddev, logical_sector, conf->reshape_progress) &&
6126	    ahead_of_reshape(mddev, logical_sector, conf->reshape_safe)) {
6127		bio_wouldblock_error(bi);
6128		if (rw == WRITE)
6129			md_write_end(mddev);
6130		return true;
6131	}
6132	md_account_bio(mddev, &bi);
6133
6134	/*
6135	 * Lets start with the stripe with the lowest chunk offset in the first
6136	 * chunk. That has the best chances of creating IOs adjacent to
6137	 * previous IOs in case of sequential IO and thus creates the most
6138	 * sequential IO pattern. We don't bother with the optimization when
6139	 * reshaping as the performance benefit is not worth the complexity.
6140	 */
6141	if (likely(conf->reshape_progress == MaxSector))
6142		logical_sector = raid5_bio_lowest_chunk_sector(conf, bi);
6143	s = (logical_sector - ctx.first_sector) >> RAID5_STRIPE_SHIFT(conf);
6144
6145	add_wait_queue(&conf->wait_for_overlap, &wait);
6146	while (1) {
6147		res = make_stripe_request(mddev, conf, &ctx, logical_sector,
6148					  bi);
6149		if (res == STRIPE_FAIL)
6150			break;
6151
6152		if (res == STRIPE_RETRY)
6153			continue;
6154
6155		if (res == STRIPE_SCHEDULE_AND_RETRY) {
6156			/*
6157			 * Must release the reference to batch_last before
6158			 * scheduling and waiting for work to be done,
6159			 * otherwise the batch_last stripe head could prevent
6160			 * raid5_activate_delayed() from making progress
6161			 * and thus deadlocking.
6162			 */
6163			if (ctx.batch_last) {
6164				raid5_release_stripe(ctx.batch_last);
6165				ctx.batch_last = NULL;
6166			}
6167
6168			wait_woken(&wait, TASK_UNINTERRUPTIBLE,
6169				   MAX_SCHEDULE_TIMEOUT);
6170			continue;
6171		}
6172
6173		s = find_next_bit_wrap(ctx.sectors_to_do, stripe_cnt, s);
6174		if (s == stripe_cnt)
6175			break;
6176
6177		logical_sector = ctx.first_sector +
6178			(s << RAID5_STRIPE_SHIFT(conf));
6179	}
6180	remove_wait_queue(&conf->wait_for_overlap, &wait);
6181
6182	if (ctx.batch_last)
6183		raid5_release_stripe(ctx.batch_last);
6184
6185	if (rw == WRITE)
6186		md_write_end(mddev);
6187	bio_endio(bi);
6188	return true;
6189}
6190
6191static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
6192
6193static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
6194{
6195	/* reshaping is quite different to recovery/resync so it is
6196	 * handled quite separately ... here.
6197	 *
6198	 * On each call to sync_request, we gather one chunk worth of
6199	 * destination stripes and flag them as expanding.
6200	 * Then we find all the source stripes and request reads.
6201	 * As the reads complete, handle_stripe will copy the data
6202	 * into the destination stripe and release that stripe.
6203	 */
6204	struct r5conf *conf = mddev->private;
6205	struct stripe_head *sh;
6206	struct md_rdev *rdev;
6207	sector_t first_sector, last_sector;
6208	int raid_disks = conf->previous_raid_disks;
6209	int data_disks = raid_disks - conf->max_degraded;
6210	int new_data_disks = conf->raid_disks - conf->max_degraded;
6211	int i;
6212	int dd_idx;
6213	sector_t writepos, readpos, safepos;
6214	sector_t stripe_addr;
6215	int reshape_sectors;
6216	struct list_head stripes;
6217	sector_t retn;
6218
6219	if (sector_nr == 0) {
6220		/* If restarting in the middle, skip the initial sectors */
6221		if (mddev->reshape_backwards &&
6222		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
6223			sector_nr = raid5_size(mddev, 0, 0)
6224				- conf->reshape_progress;
6225		} else if (mddev->reshape_backwards &&
6226			   conf->reshape_progress == MaxSector) {
6227			/* shouldn't happen, but just in case, finish up.*/
6228			sector_nr = MaxSector;
6229		} else if (!mddev->reshape_backwards &&
6230			   conf->reshape_progress > 0)
6231			sector_nr = conf->reshape_progress;
6232		sector_div(sector_nr, new_data_disks);
6233		if (sector_nr) {
6234			mddev->curr_resync_completed = sector_nr;
6235			sysfs_notify_dirent_safe(mddev->sysfs_completed);
6236			*skipped = 1;
6237			retn = sector_nr;
6238			goto finish;
6239		}
6240	}
6241
6242	/* We need to process a full chunk at a time.
6243	 * If old and new chunk sizes differ, we need to process the
6244	 * largest of these
6245	 */
 
 
 
 
6246
6247	reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
6248
6249	/* We update the metadata at least every 10 seconds, or when
6250	 * the data about to be copied would over-write the source of
6251	 * the data at the front of the range.  i.e. one new_stripe
6252	 * along from reshape_progress new_maps to after where
6253	 * reshape_safe old_maps to
6254	 */
6255	writepos = conf->reshape_progress;
6256	sector_div(writepos, new_data_disks);
6257	readpos = conf->reshape_progress;
6258	sector_div(readpos, data_disks);
6259	safepos = conf->reshape_safe;
6260	sector_div(safepos, data_disks);
6261	if (mddev->reshape_backwards) {
6262		BUG_ON(writepos < reshape_sectors);
6263		writepos -= reshape_sectors;
6264		readpos += reshape_sectors;
6265		safepos += reshape_sectors;
6266	} else {
6267		writepos += reshape_sectors;
6268		/* readpos and safepos are worst-case calculations.
6269		 * A negative number is overly pessimistic, and causes
6270		 * obvious problems for unsigned storage.  So clip to 0.
6271		 */
6272		readpos -= min_t(sector_t, reshape_sectors, readpos);
6273		safepos -= min_t(sector_t, reshape_sectors, safepos);
6274	}
6275
6276	/* Having calculated the 'writepos' possibly use it
6277	 * to set 'stripe_addr' which is where we will write to.
6278	 */
6279	if (mddev->reshape_backwards) {
6280		BUG_ON(conf->reshape_progress == 0);
6281		stripe_addr = writepos;
6282		BUG_ON((mddev->dev_sectors &
6283			~((sector_t)reshape_sectors - 1))
6284		       - reshape_sectors - stripe_addr
6285		       != sector_nr);
6286	} else {
6287		BUG_ON(writepos != sector_nr + reshape_sectors);
6288		stripe_addr = sector_nr;
6289	}
6290
6291	/* 'writepos' is the most advanced device address we might write.
6292	 * 'readpos' is the least advanced device address we might read.
6293	 * 'safepos' is the least address recorded in the metadata as having
6294	 *     been reshaped.
6295	 * If there is a min_offset_diff, these are adjusted either by
6296	 * increasing the safepos/readpos if diff is negative, or
6297	 * increasing writepos if diff is positive.
6298	 * If 'readpos' is then behind 'writepos', there is no way that we can
6299	 * ensure safety in the face of a crash - that must be done by userspace
6300	 * making a backup of the data.  So in that case there is no particular
6301	 * rush to update metadata.
6302	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
6303	 * update the metadata to advance 'safepos' to match 'readpos' so that
6304	 * we can be safe in the event of a crash.
6305	 * So we insist on updating metadata if safepos is behind writepos and
6306	 * readpos is beyond writepos.
6307	 * In any case, update the metadata every 10 seconds.
6308	 * Maybe that number should be configurable, but I'm not sure it is
6309	 * worth it.... maybe it could be a multiple of safemode_delay???
6310	 */
6311	if (conf->min_offset_diff < 0) {
6312		safepos += -conf->min_offset_diff;
6313		readpos += -conf->min_offset_diff;
6314	} else
6315		writepos += conf->min_offset_diff;
6316
6317	if ((mddev->reshape_backwards
6318	     ? (safepos > writepos && readpos < writepos)
6319	     : (safepos < writepos && readpos > writepos)) ||
6320	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6321		/* Cannot proceed until we've updated the superblock... */
6322		wait_event(conf->wait_for_overlap,
6323			   atomic_read(&conf->reshape_stripes)==0
6324			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6325		if (atomic_read(&conf->reshape_stripes) != 0)
6326			return 0;
6327		mddev->reshape_position = conf->reshape_progress;
6328		mddev->curr_resync_completed = sector_nr;
6329		if (!mddev->reshape_backwards)
6330			/* Can update recovery_offset */
6331			rdev_for_each(rdev, mddev)
6332				if (rdev->raid_disk >= 0 &&
6333				    !test_bit(Journal, &rdev->flags) &&
6334				    !test_bit(In_sync, &rdev->flags) &&
6335				    rdev->recovery_offset < sector_nr)
6336					rdev->recovery_offset = sector_nr;
6337
6338		conf->reshape_checkpoint = jiffies;
6339		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6340		md_wakeup_thread(mddev->thread);
6341		wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6342			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6343		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6344			return 0;
6345		spin_lock_irq(&conf->device_lock);
6346		conf->reshape_safe = mddev->reshape_position;
6347		spin_unlock_irq(&conf->device_lock);
6348		wake_up(&conf->wait_for_overlap);
6349		sysfs_notify_dirent_safe(mddev->sysfs_completed);
6350	}
6351
 
 
 
 
 
 
 
 
 
 
 
6352	INIT_LIST_HEAD(&stripes);
6353	for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6354		int j;
6355		int skipped_disk = 0;
6356		sh = raid5_get_active_stripe(conf, NULL, stripe_addr+i,
6357					     R5_GAS_NOQUIESCE);
6358		set_bit(STRIPE_EXPANDING, &sh->state);
6359		atomic_inc(&conf->reshape_stripes);
6360		/* If any of this stripe is beyond the end of the old
6361		 * array, then we need to zero those blocks
6362		 */
6363		for (j=sh->disks; j--;) {
6364			sector_t s;
6365			if (j == sh->pd_idx)
6366				continue;
6367			if (conf->level == 6 &&
6368			    j == sh->qd_idx)
6369				continue;
6370			s = raid5_compute_blocknr(sh, j, 0);
6371			if (s < raid5_size(mddev, 0, 0)) {
6372				skipped_disk = 1;
6373				continue;
6374			}
6375			memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6376			set_bit(R5_Expanded, &sh->dev[j].flags);
6377			set_bit(R5_UPTODATE, &sh->dev[j].flags);
6378		}
6379		if (!skipped_disk) {
6380			set_bit(STRIPE_EXPAND_READY, &sh->state);
6381			set_bit(STRIPE_HANDLE, &sh->state);
6382		}
6383		list_add(&sh->lru, &stripes);
6384	}
6385	spin_lock_irq(&conf->device_lock);
6386	if (mddev->reshape_backwards)
6387		conf->reshape_progress -= reshape_sectors * new_data_disks;
6388	else
6389		conf->reshape_progress += reshape_sectors * new_data_disks;
6390	spin_unlock_irq(&conf->device_lock);
6391	/* Ok, those stripe are ready. We can start scheduling
6392	 * reads on the source stripes.
6393	 * The source stripes are determined by mapping the first and last
6394	 * block on the destination stripes.
6395	 */
6396	first_sector =
6397		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6398				     1, &dd_idx, NULL);
6399	last_sector =
6400		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6401					    * new_data_disks - 1),
6402				     1, &dd_idx, NULL);
6403	if (last_sector >= mddev->dev_sectors)
6404		last_sector = mddev->dev_sectors - 1;
6405	while (first_sector <= last_sector) {
6406		sh = raid5_get_active_stripe(conf, NULL, first_sector,
6407				R5_GAS_PREVIOUS | R5_GAS_NOQUIESCE);
6408		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6409		set_bit(STRIPE_HANDLE, &sh->state);
6410		raid5_release_stripe(sh);
6411		first_sector += RAID5_STRIPE_SECTORS(conf);
6412	}
6413	/* Now that the sources are clearly marked, we can release
6414	 * the destination stripes
6415	 */
6416	while (!list_empty(&stripes)) {
6417		sh = list_entry(stripes.next, struct stripe_head, lru);
6418		list_del_init(&sh->lru);
6419		raid5_release_stripe(sh);
6420	}
6421	/* If this takes us to the resync_max point where we have to pause,
6422	 * then we need to write out the superblock.
6423	 */
6424	sector_nr += reshape_sectors;
6425	retn = reshape_sectors;
6426finish:
6427	if (mddev->curr_resync_completed > mddev->resync_max ||
6428	    (sector_nr - mddev->curr_resync_completed) * 2
6429	    >= mddev->resync_max - mddev->curr_resync_completed) {
6430		/* Cannot proceed until we've updated the superblock... */
6431		wait_event(conf->wait_for_overlap,
6432			   atomic_read(&conf->reshape_stripes) == 0
6433			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6434		if (atomic_read(&conf->reshape_stripes) != 0)
6435			goto ret;
6436		mddev->reshape_position = conf->reshape_progress;
6437		mddev->curr_resync_completed = sector_nr;
6438		if (!mddev->reshape_backwards)
6439			/* Can update recovery_offset */
6440			rdev_for_each(rdev, mddev)
6441				if (rdev->raid_disk >= 0 &&
6442				    !test_bit(Journal, &rdev->flags) &&
6443				    !test_bit(In_sync, &rdev->flags) &&
6444				    rdev->recovery_offset < sector_nr)
6445					rdev->recovery_offset = sector_nr;
6446		conf->reshape_checkpoint = jiffies;
6447		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6448		md_wakeup_thread(mddev->thread);
6449		wait_event(mddev->sb_wait,
6450			   !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6451			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6452		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6453			goto ret;
6454		spin_lock_irq(&conf->device_lock);
6455		conf->reshape_safe = mddev->reshape_position;
6456		spin_unlock_irq(&conf->device_lock);
6457		wake_up(&conf->wait_for_overlap);
6458		sysfs_notify_dirent_safe(mddev->sysfs_completed);
6459	}
6460ret:
6461	return retn;
6462}
6463
6464static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6465					  int *skipped)
6466{
6467	struct r5conf *conf = mddev->private;
6468	struct stripe_head *sh;
6469	sector_t max_sector = mddev->dev_sectors;
6470	sector_t sync_blocks;
6471	int still_degraded = 0;
6472	int i;
6473
6474	if (sector_nr >= max_sector) {
6475		/* just being told to finish up .. nothing much to do */
6476
6477		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6478			end_reshape(conf);
6479			return 0;
6480		}
6481
6482		if (mddev->curr_resync < max_sector) /* aborted */
6483			md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6484					   &sync_blocks, 1);
6485		else /* completed sync */
6486			conf->fullsync = 0;
6487		md_bitmap_close_sync(mddev->bitmap);
6488
6489		return 0;
6490	}
6491
6492	/* Allow raid5_quiesce to complete */
6493	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6494
6495	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6496		return reshape_request(mddev, sector_nr, skipped);
6497
6498	/* No need to check resync_max as we never do more than one
6499	 * stripe, and as resync_max will always be on a chunk boundary,
6500	 * if the check in md_do_sync didn't fire, there is no chance
6501	 * of overstepping resync_max here
6502	 */
6503
6504	/* if there is too many failed drives and we are trying
6505	 * to resync, then assert that we are finished, because there is
6506	 * nothing we can do.
6507	 */
6508	if (mddev->degraded >= conf->max_degraded &&
6509	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6510		sector_t rv = mddev->dev_sectors - sector_nr;
6511		*skipped = 1;
6512		return rv;
6513	}
6514	if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6515	    !conf->fullsync &&
6516	    !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6517	    sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6518		/* we can skip this block, and probably more */
6519		do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6520		*skipped = 1;
6521		/* keep things rounded to whole stripes */
6522		return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6523	}
6524
6525	md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6526
6527	sh = raid5_get_active_stripe(conf, NULL, sector_nr,
6528				     R5_GAS_NOBLOCK);
 
6529	if (sh == NULL) {
6530		sh = raid5_get_active_stripe(conf, NULL, sector_nr, 0);
6531		/* make sure we don't swamp the stripe cache if someone else
6532		 * is trying to get access
6533		 */
6534		schedule_timeout_uninterruptible(1);
6535	}
6536	/* Need to check if array will still be degraded after recovery/resync
6537	 * Note in case of > 1 drive failures it's possible we're rebuilding
6538	 * one drive while leaving another faulty drive in array.
6539	 */
6540	for (i = 0; i < conf->raid_disks; i++) {
6541		struct md_rdev *rdev = conf->disks[i].rdev;
6542
6543		if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6544			still_degraded = 1;
6545	}
6546
6547	md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6548
6549	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6550	set_bit(STRIPE_HANDLE, &sh->state);
6551
6552	raid5_release_stripe(sh);
 
6553
6554	return RAID5_STRIPE_SECTORS(conf);
6555}
6556
6557static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6558			       unsigned int offset)
6559{
6560	/* We may not be able to submit a whole bio at once as there
6561	 * may not be enough stripe_heads available.
6562	 * We cannot pre-allocate enough stripe_heads as we may need
6563	 * more than exist in the cache (if we allow ever large chunks).
6564	 * So we do one stripe head at a time and record in
6565	 * ->bi_hw_segments how many have been done.
6566	 *
6567	 * We *know* that this entire raid_bio is in one chunk, so
6568	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6569	 */
6570	struct stripe_head *sh;
6571	int dd_idx;
6572	sector_t sector, logical_sector, last_sector;
6573	int scnt = 0;
 
6574	int handled = 0;
6575
6576	logical_sector = raid_bio->bi_iter.bi_sector &
6577		~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6578	sector = raid5_compute_sector(conf, logical_sector,
6579				      0, &dd_idx, NULL);
6580	last_sector = bio_end_sector(raid_bio);
6581
6582	for (; logical_sector < last_sector;
6583	     logical_sector += RAID5_STRIPE_SECTORS(conf),
6584		     sector += RAID5_STRIPE_SECTORS(conf),
6585		     scnt++) {
6586
6587		if (scnt < offset)
6588			/* already done this stripe */
6589			continue;
6590
6591		sh = raid5_get_active_stripe(conf, NULL, sector,
6592				R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
6593		if (!sh) {
6594			/* failed to get a stripe - must wait */
 
6595			conf->retry_read_aligned = raid_bio;
6596			conf->retry_read_offset = scnt;
6597			return handled;
6598		}
6599
6600		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6601			raid5_release_stripe(sh);
 
 
6602			conf->retry_read_aligned = raid_bio;
6603			conf->retry_read_offset = scnt;
6604			return handled;
6605		}
6606
6607		set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6608		handle_stripe(sh);
6609		raid5_release_stripe(sh);
6610		handled++;
6611	}
6612
6613	bio_endio(raid_bio);
6614
 
 
6615	if (atomic_dec_and_test(&conf->active_aligned_reads))
6616		wake_up(&conf->wait_for_quiescent);
6617	return handled;
6618}
6619
6620static int handle_active_stripes(struct r5conf *conf, int group,
6621				 struct r5worker *worker,
6622				 struct list_head *temp_inactive_list)
6623		__must_hold(&conf->device_lock)
6624{
6625	struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6626	int i, batch_size = 0, hash;
6627	bool release_inactive = false;
6628
6629	while (batch_size < MAX_STRIPE_BATCH &&
6630			(sh = __get_priority_stripe(conf, group)) != NULL)
6631		batch[batch_size++] = sh;
6632
6633	if (batch_size == 0) {
6634		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6635			if (!list_empty(temp_inactive_list + i))
6636				break;
6637		if (i == NR_STRIPE_HASH_LOCKS) {
6638			spin_unlock_irq(&conf->device_lock);
6639			log_flush_stripe_to_raid(conf);
6640			spin_lock_irq(&conf->device_lock);
6641			return batch_size;
6642		}
6643		release_inactive = true;
6644	}
6645	spin_unlock_irq(&conf->device_lock);
6646
6647	release_inactive_stripe_list(conf, temp_inactive_list,
6648				     NR_STRIPE_HASH_LOCKS);
6649
6650	r5l_flush_stripe_to_raid(conf->log);
6651	if (release_inactive) {
6652		spin_lock_irq(&conf->device_lock);
6653		return 0;
6654	}
6655
6656	for (i = 0; i < batch_size; i++)
6657		handle_stripe(batch[i]);
6658	log_write_stripe_run(conf);
6659
6660	cond_resched();
6661
6662	spin_lock_irq(&conf->device_lock);
6663	for (i = 0; i < batch_size; i++) {
6664		hash = batch[i]->hash_lock_index;
6665		__release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6666	}
6667	return batch_size;
6668}
6669
6670static void raid5_do_work(struct work_struct *work)
6671{
6672	struct r5worker *worker = container_of(work, struct r5worker, work);
6673	struct r5worker_group *group = worker->group;
6674	struct r5conf *conf = group->conf;
6675	struct mddev *mddev = conf->mddev;
6676	int group_id = group - conf->worker_groups;
6677	int handled;
6678	struct blk_plug plug;
6679
6680	pr_debug("+++ raid5worker active\n");
6681
6682	blk_start_plug(&plug);
6683	handled = 0;
6684	spin_lock_irq(&conf->device_lock);
6685	while (1) {
6686		int batch_size, released;
6687
6688		released = release_stripe_list(conf, worker->temp_inactive_list);
6689
6690		batch_size = handle_active_stripes(conf, group_id, worker,
6691						   worker->temp_inactive_list);
6692		worker->working = false;
6693		if (!batch_size && !released)
6694			break;
6695		handled += batch_size;
6696		wait_event_lock_irq(mddev->sb_wait,
6697			!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6698			conf->device_lock);
6699	}
6700	pr_debug("%d stripes handled\n", handled);
6701
6702	spin_unlock_irq(&conf->device_lock);
6703
6704	flush_deferred_bios(conf);
6705
6706	r5l_flush_stripe_to_raid(conf->log);
6707
6708	async_tx_issue_pending_all();
6709	blk_finish_plug(&plug);
6710
6711	pr_debug("--- raid5worker inactive\n");
6712}
6713
6714/*
6715 * This is our raid5 kernel thread.
6716 *
6717 * We scan the hash table for stripes which can be handled now.
6718 * During the scan, completed stripes are saved for us by the interrupt
6719 * handler, so that they will not have to wait for our next wakeup.
6720 */
6721static void raid5d(struct md_thread *thread)
6722{
6723	struct mddev *mddev = thread->mddev;
6724	struct r5conf *conf = mddev->private;
6725	int handled;
6726	struct blk_plug plug;
6727
6728	pr_debug("+++ raid5d active\n");
6729
6730	md_check_recovery(mddev);
6731
6732	blk_start_plug(&plug);
6733	handled = 0;
6734	spin_lock_irq(&conf->device_lock);
6735	while (1) {
6736		struct bio *bio;
6737		int batch_size, released;
6738		unsigned int offset;
6739
6740		released = release_stripe_list(conf, conf->temp_inactive_list);
6741		if (released)
6742			clear_bit(R5_DID_ALLOC, &conf->cache_state);
6743
6744		if (
6745		    !list_empty(&conf->bitmap_list)) {
6746			/* Now is a good time to flush some bitmap updates */
6747			conf->seq_flush++;
6748			spin_unlock_irq(&conf->device_lock);
6749			md_bitmap_unplug(mddev->bitmap);
6750			spin_lock_irq(&conf->device_lock);
6751			conf->seq_write = conf->seq_flush;
6752			activate_bit_delay(conf, conf->temp_inactive_list);
6753		}
6754		raid5_activate_delayed(conf);
 
6755
6756		while ((bio = remove_bio_from_retry(conf, &offset))) {
6757			int ok;
6758			spin_unlock_irq(&conf->device_lock);
6759			ok = retry_aligned_read(conf, bio, offset);
6760			spin_lock_irq(&conf->device_lock);
6761			if (!ok)
6762				break;
6763			handled++;
6764		}
6765
6766		batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6767						   conf->temp_inactive_list);
6768		if (!batch_size && !released)
6769			break;
6770		handled += batch_size;
 
 
 
 
 
6771
6772		if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6773			spin_unlock_irq(&conf->device_lock);
6774			md_check_recovery(mddev);
6775			spin_lock_irq(&conf->device_lock);
6776		}
6777	}
6778	pr_debug("%d stripes handled\n", handled);
6779
6780	spin_unlock_irq(&conf->device_lock);
6781	if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6782	    mutex_trylock(&conf->cache_size_mutex)) {
6783		grow_one_stripe(conf, __GFP_NOWARN);
6784		/* Set flag even if allocation failed.  This helps
6785		 * slow down allocation requests when mem is short
6786		 */
6787		set_bit(R5_DID_ALLOC, &conf->cache_state);
6788		mutex_unlock(&conf->cache_size_mutex);
6789	}
6790
6791	flush_deferred_bios(conf);
6792
6793	r5l_flush_stripe_to_raid(conf->log);
6794
6795	async_tx_issue_pending_all();
6796	blk_finish_plug(&plug);
6797
6798	pr_debug("--- raid5d inactive\n");
6799}
6800
6801static ssize_t
6802raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6803{
6804	struct r5conf *conf;
6805	int ret = 0;
6806	spin_lock(&mddev->lock);
6807	conf = mddev->private;
6808	if (conf)
6809		ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6810	spin_unlock(&mddev->lock);
6811	return ret;
6812}
6813
6814int
6815raid5_set_cache_size(struct mddev *mddev, int size)
6816{
6817	int result = 0;
6818	struct r5conf *conf = mddev->private;
6819
6820	if (size <= 16 || size > 32768)
6821		return -EINVAL;
6822
6823	conf->min_nr_stripes = size;
6824	mutex_lock(&conf->cache_size_mutex);
6825	while (size < conf->max_nr_stripes &&
6826	       drop_one_stripe(conf))
6827		;
6828	mutex_unlock(&conf->cache_size_mutex);
6829
6830	md_allow_write(mddev);
6831
6832	mutex_lock(&conf->cache_size_mutex);
6833	while (size > conf->max_nr_stripes)
6834		if (!grow_one_stripe(conf, GFP_KERNEL)) {
6835			conf->min_nr_stripes = conf->max_nr_stripes;
6836			result = -ENOMEM;
6837			break;
6838		}
6839	mutex_unlock(&conf->cache_size_mutex);
6840
6841	return result;
 
 
 
 
 
 
6842}
6843EXPORT_SYMBOL(raid5_set_cache_size);
6844
6845static ssize_t
6846raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6847{
6848	struct r5conf *conf;
6849	unsigned long new;
6850	int err;
6851
6852	if (len >= PAGE_SIZE)
6853		return -EINVAL;
6854	if (kstrtoul(page, 10, &new))
 
 
 
6855		return -EINVAL;
6856	err = mddev_lock(mddev);
6857	if (err)
6858		return err;
6859	conf = mddev->private;
6860	if (!conf)
6861		err = -ENODEV;
6862	else
6863		err = raid5_set_cache_size(mddev, new);
6864	mddev_unlock(mddev);
6865
6866	return err ?: len;
6867}
6868
6869static struct md_sysfs_entry
6870raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6871				raid5_show_stripe_cache_size,
6872				raid5_store_stripe_cache_size);
6873
6874static ssize_t
6875raid5_show_rmw_level(struct mddev  *mddev, char *page)
6876{
6877	struct r5conf *conf = mddev->private;
6878	if (conf)
6879		return sprintf(page, "%d\n", conf->rmw_level);
6880	else
6881		return 0;
6882}
6883
6884static ssize_t
6885raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6886{
6887	struct r5conf *conf = mddev->private;
6888	unsigned long new;
6889
6890	if (!conf)
6891		return -ENODEV;
6892
6893	if (len >= PAGE_SIZE)
6894		return -EINVAL;
 
 
6895
6896	if (kstrtoul(page, 10, &new))
6897		return -EINVAL;
6898
6899	if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6900		return -EINVAL;
6901
6902	if (new != PARITY_DISABLE_RMW &&
6903	    new != PARITY_ENABLE_RMW &&
6904	    new != PARITY_PREFER_RMW)
6905		return -EINVAL;
6906
6907	conf->rmw_level = new;
6908	return len;
6909}
6910
6911static struct md_sysfs_entry
6912raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6913			 raid5_show_rmw_level,
6914			 raid5_store_rmw_level);
6915
6916static ssize_t
6917raid5_show_stripe_size(struct mddev  *mddev, char *page)
6918{
6919	struct r5conf *conf;
6920	int ret = 0;
6921
6922	spin_lock(&mddev->lock);
6923	conf = mddev->private;
6924	if (conf)
6925		ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6926	spin_unlock(&mddev->lock);
6927	return ret;
6928}
6929
6930#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6931static ssize_t
6932raid5_store_stripe_size(struct mddev  *mddev, const char *page, size_t len)
6933{
6934	struct r5conf *conf;
6935	unsigned long new;
6936	int err;
6937	int size;
6938
6939	if (len >= PAGE_SIZE)
6940		return -EINVAL;
6941	if (kstrtoul(page, 10, &new))
6942		return -EINVAL;
6943
6944	/*
6945	 * The value should not be bigger than PAGE_SIZE. It requires to
6946	 * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6947	 * of two.
6948	 */
6949	if (new % DEFAULT_STRIPE_SIZE != 0 ||
6950			new > PAGE_SIZE || new == 0 ||
6951			new != roundup_pow_of_two(new))
6952		return -EINVAL;
6953
6954	err = mddev_suspend_and_lock(mddev);
6955	if (err)
6956		return err;
6957
6958	conf = mddev->private;
6959	if (!conf) {
6960		err = -ENODEV;
6961		goto out_unlock;
6962	}
6963
6964	if (new == conf->stripe_size)
6965		goto out_unlock;
6966
6967	pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6968			conf->stripe_size, new);
6969
6970	if (mddev->sync_thread ||
6971		test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6972		mddev->reshape_position != MaxSector ||
6973		mddev->sysfs_active) {
6974		err = -EBUSY;
6975		goto out_unlock;
6976	}
6977
6978	mutex_lock(&conf->cache_size_mutex);
6979	size = conf->max_nr_stripes;
6980
6981	shrink_stripes(conf);
6982
6983	conf->stripe_size = new;
6984	conf->stripe_shift = ilog2(new) - 9;
6985	conf->stripe_sectors = new >> 9;
6986	if (grow_stripes(conf, size)) {
6987		pr_warn("md/raid:%s: couldn't allocate buffers\n",
6988				mdname(mddev));
6989		err = -ENOMEM;
6990	}
6991	mutex_unlock(&conf->cache_size_mutex);
6992
6993out_unlock:
6994	mddev_unlock_and_resume(mddev);
6995	return err ?: len;
6996}
6997
6998static struct md_sysfs_entry
6999raid5_stripe_size = __ATTR(stripe_size, 0644,
7000			 raid5_show_stripe_size,
7001			 raid5_store_stripe_size);
7002#else
7003static struct md_sysfs_entry
7004raid5_stripe_size = __ATTR(stripe_size, 0444,
7005			 raid5_show_stripe_size,
7006			 NULL);
7007#endif
7008
7009static ssize_t
7010raid5_show_preread_threshold(struct mddev *mddev, char *page)
7011{
7012	struct r5conf *conf;
7013	int ret = 0;
7014	spin_lock(&mddev->lock);
7015	conf = mddev->private;
7016	if (conf)
7017		ret = sprintf(page, "%d\n", conf->bypass_threshold);
7018	spin_unlock(&mddev->lock);
7019	return ret;
7020}
7021
7022static ssize_t
7023raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
7024{
7025	struct r5conf *conf;
7026	unsigned long new;
7027	int err;
7028
7029	if (len >= PAGE_SIZE)
7030		return -EINVAL;
7031	if (kstrtoul(page, 10, &new))
7032		return -EINVAL;
7033
7034	err = mddev_lock(mddev);
7035	if (err)
7036		return err;
7037	conf = mddev->private;
7038	if (!conf)
7039		err = -ENODEV;
7040	else if (new > conf->min_nr_stripes)
7041		err = -EINVAL;
7042	else
7043		conf->bypass_threshold = new;
7044	mddev_unlock(mddev);
7045	return err ?: len;
7046}
7047
7048static struct md_sysfs_entry
7049raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
7050					S_IRUGO | S_IWUSR,
7051					raid5_show_preread_threshold,
7052					raid5_store_preread_threshold);
7053
7054static ssize_t
7055raid5_show_skip_copy(struct mddev *mddev, char *page)
7056{
7057	struct r5conf *conf;
7058	int ret = 0;
7059	spin_lock(&mddev->lock);
7060	conf = mddev->private;
7061	if (conf)
7062		ret = sprintf(page, "%d\n", conf->skip_copy);
7063	spin_unlock(&mddev->lock);
7064	return ret;
7065}
7066
7067static ssize_t
7068raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
7069{
7070	struct r5conf *conf;
7071	unsigned long new;
7072	int err;
7073
7074	if (len >= PAGE_SIZE)
7075		return -EINVAL;
7076	if (kstrtoul(page, 10, &new))
7077		return -EINVAL;
7078	new = !!new;
7079
7080	err = mddev_suspend_and_lock(mddev);
7081	if (err)
7082		return err;
7083	conf = mddev->private;
7084	if (!conf)
7085		err = -ENODEV;
7086	else if (new != conf->skip_copy) {
7087		struct request_queue *q = mddev->queue;
7088
7089		conf->skip_copy = new;
7090		if (new)
7091			blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q);
7092		else
7093			blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES, q);
7094	}
7095	mddev_unlock_and_resume(mddev);
7096	return err ?: len;
7097}
7098
7099static struct md_sysfs_entry
7100raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
7101					raid5_show_skip_copy,
7102					raid5_store_skip_copy);
7103
7104static ssize_t
7105stripe_cache_active_show(struct mddev *mddev, char *page)
7106{
7107	struct r5conf *conf = mddev->private;
7108	if (conf)
7109		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
7110	else
7111		return 0;
7112}
7113
7114static struct md_sysfs_entry
7115raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
7116
7117static ssize_t
7118raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
7119{
7120	struct r5conf *conf;
7121	int ret = 0;
7122	spin_lock(&mddev->lock);
7123	conf = mddev->private;
7124	if (conf)
7125		ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
7126	spin_unlock(&mddev->lock);
7127	return ret;
7128}
7129
7130static int alloc_thread_groups(struct r5conf *conf, int cnt,
7131			       int *group_cnt,
7132			       struct r5worker_group **worker_groups);
7133static ssize_t
7134raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
7135{
7136	struct r5conf *conf;
7137	unsigned int new;
7138	int err;
7139	struct r5worker_group *new_groups, *old_groups;
7140	int group_cnt;
7141
7142	if (len >= PAGE_SIZE)
7143		return -EINVAL;
7144	if (kstrtouint(page, 10, &new))
7145		return -EINVAL;
7146	/* 8192 should be big enough */
7147	if (new > 8192)
7148		return -EINVAL;
7149
7150	err = mddev_suspend_and_lock(mddev);
7151	if (err)
7152		return err;
7153	conf = mddev->private;
7154	if (!conf)
7155		err = -ENODEV;
7156	else if (new != conf->worker_cnt_per_group) {
7157		old_groups = conf->worker_groups;
7158		if (old_groups)
7159			flush_workqueue(raid5_wq);
7160
7161		err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
7162		if (!err) {
7163			spin_lock_irq(&conf->device_lock);
7164			conf->group_cnt = group_cnt;
7165			conf->worker_cnt_per_group = new;
7166			conf->worker_groups = new_groups;
7167			spin_unlock_irq(&conf->device_lock);
7168
7169			if (old_groups)
7170				kfree(old_groups[0].workers);
7171			kfree(old_groups);
7172		}
7173	}
7174	mddev_unlock_and_resume(mddev);
7175
7176	return err ?: len;
7177}
7178
7179static struct md_sysfs_entry
7180raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
7181				raid5_show_group_thread_cnt,
7182				raid5_store_group_thread_cnt);
7183
7184static struct attribute *raid5_attrs[] =  {
7185	&raid5_stripecache_size.attr,
7186	&raid5_stripecache_active.attr,
7187	&raid5_preread_bypass_threshold.attr,
7188	&raid5_group_thread_cnt.attr,
7189	&raid5_skip_copy.attr,
7190	&raid5_rmw_level.attr,
7191	&raid5_stripe_size.attr,
7192	&r5c_journal_mode.attr,
7193	&ppl_write_hint.attr,
7194	NULL,
7195};
7196static const struct attribute_group raid5_attrs_group = {
7197	.name = NULL,
7198	.attrs = raid5_attrs,
7199};
7200
7201static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
7202			       struct r5worker_group **worker_groups)
7203{
7204	int i, j, k;
7205	ssize_t size;
7206	struct r5worker *workers;
7207
7208	if (cnt == 0) {
7209		*group_cnt = 0;
7210		*worker_groups = NULL;
7211		return 0;
7212	}
7213	*group_cnt = num_possible_nodes();
7214	size = sizeof(struct r5worker) * cnt;
7215	workers = kcalloc(size, *group_cnt, GFP_NOIO);
7216	*worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
7217				 GFP_NOIO);
7218	if (!*worker_groups || !workers) {
7219		kfree(workers);
7220		kfree(*worker_groups);
7221		return -ENOMEM;
7222	}
7223
7224	for (i = 0; i < *group_cnt; i++) {
7225		struct r5worker_group *group;
7226
7227		group = &(*worker_groups)[i];
7228		INIT_LIST_HEAD(&group->handle_list);
7229		INIT_LIST_HEAD(&group->loprio_list);
7230		group->conf = conf;
7231		group->workers = workers + i * cnt;
7232
7233		for (j = 0; j < cnt; j++) {
7234			struct r5worker *worker = group->workers + j;
7235			worker->group = group;
7236			INIT_WORK(&worker->work, raid5_do_work);
7237
7238			for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
7239				INIT_LIST_HEAD(worker->temp_inactive_list + k);
7240		}
7241	}
7242
7243	return 0;
7244}
7245
7246static void free_thread_groups(struct r5conf *conf)
7247{
7248	if (conf->worker_groups)
7249		kfree(conf->worker_groups[0].workers);
7250	kfree(conf->worker_groups);
7251	conf->worker_groups = NULL;
7252}
7253
7254static sector_t
7255raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7256{
7257	struct r5conf *conf = mddev->private;
7258
7259	if (!sectors)
7260		sectors = mddev->dev_sectors;
7261	if (!raid_disks)
7262		/* size is defined by the smallest of previous and new size */
7263		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7264
7265	sectors &= ~((sector_t)conf->chunk_sectors - 1);
7266	sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7267	return sectors * (raid_disks - conf->max_degraded);
7268}
7269
7270static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7271{
7272	safe_put_page(percpu->spare_page);
7273	percpu->spare_page = NULL;
7274	kvfree(percpu->scribble);
7275	percpu->scribble = NULL;
7276}
7277
7278static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7279{
7280	if (conf->level == 6 && !percpu->spare_page) {
7281		percpu->spare_page = alloc_page(GFP_KERNEL);
7282		if (!percpu->spare_page)
7283			return -ENOMEM;
7284	}
7285
7286	if (scribble_alloc(percpu,
7287			   max(conf->raid_disks,
7288			       conf->previous_raid_disks),
7289			   max(conf->chunk_sectors,
7290			       conf->prev_chunk_sectors)
7291			   / RAID5_STRIPE_SECTORS(conf))) {
7292		free_scratch_buffer(conf, percpu);
7293		return -ENOMEM;
7294	}
7295
7296	local_lock_init(&percpu->lock);
7297	return 0;
7298}
7299
7300static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7301{
7302	struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7303
7304	free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7305	return 0;
7306}
7307
7308static void raid5_free_percpu(struct r5conf *conf)
7309{
7310	if (!conf->percpu)
7311		return;
7312
7313	cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
 
 
 
 
 
 
 
 
 
 
7314	free_percpu(conf->percpu);
7315}
7316
7317static void free_conf(struct r5conf *conf)
7318{
7319	int i;
7320
7321	log_exit(conf);
7322
7323	shrinker_free(conf->shrinker);
7324	free_thread_groups(conf);
7325	shrink_stripes(conf);
7326	raid5_free_percpu(conf);
7327	for (i = 0; i < conf->pool_size; i++)
7328		if (conf->disks[i].extra_page)
7329			put_page(conf->disks[i].extra_page);
7330	kfree(conf->disks);
7331	bioset_exit(&conf->bio_split);
7332	kfree(conf->stripe_hashtbl);
7333	kfree(conf->pending_data);
7334	kfree(conf);
7335}
7336
7337static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
 
 
7338{
7339	struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
 
7340	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7341
7342	if (alloc_scratch_buffer(conf, percpu)) {
7343		pr_warn("%s: failed memory allocation for cpu%u\n",
7344			__func__, cpu);
7345		return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7346	}
7347	return 0;
7348}
 
7349
7350static int raid5_alloc_percpu(struct r5conf *conf)
7351{
7352	int err = 0;
 
 
 
 
7353
7354	conf->percpu = alloc_percpu(struct raid5_percpu);
7355	if (!conf->percpu)
7356		return -ENOMEM;
 
7357
7358	err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7359	if (!err) {
7360		conf->scribble_disks = max(conf->raid_disks,
7361			conf->previous_raid_disks);
7362		conf->scribble_sectors = max(conf->chunk_sectors,
7363			conf->prev_chunk_sectors);
7364	}
7365	return err;
7366}
7367
7368static unsigned long raid5_cache_scan(struct shrinker *shrink,
7369				      struct shrink_control *sc)
7370{
7371	struct r5conf *conf = shrink->private_data;
7372	unsigned long ret = SHRINK_STOP;
7373
7374	if (mutex_trylock(&conf->cache_size_mutex)) {
7375		ret= 0;
7376		while (ret < sc->nr_to_scan &&
7377		       conf->max_nr_stripes > conf->min_nr_stripes) {
7378			if (drop_one_stripe(conf) == 0) {
7379				ret = SHRINK_STOP;
7380				break;
7381			}
7382			ret++;
7383		}
7384		mutex_unlock(&conf->cache_size_mutex);
 
 
 
 
 
7385	}
7386	return ret;
7387}
7388
7389static unsigned long raid5_cache_count(struct shrinker *shrink,
7390				       struct shrink_control *sc)
7391{
7392	struct r5conf *conf = shrink->private_data;
7393
7394	if (conf->max_nr_stripes < conf->min_nr_stripes)
7395		/* unlikely, but not impossible */
7396		return 0;
7397	return conf->max_nr_stripes - conf->min_nr_stripes;
7398}
7399
7400static struct r5conf *setup_conf(struct mddev *mddev)
7401{
7402	struct r5conf *conf;
7403	int raid_disk, memory, max_disks;
7404	struct md_rdev *rdev;
7405	struct disk_info *disk;
7406	char pers_name[6];
7407	int i;
7408	int group_cnt;
7409	struct r5worker_group *new_group;
7410	int ret = -ENOMEM;
7411
7412	if (mddev->new_level != 5
7413	    && mddev->new_level != 4
7414	    && mddev->new_level != 6) {
7415		pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7416			mdname(mddev), mddev->new_level);
7417		return ERR_PTR(-EIO);
7418	}
7419	if ((mddev->new_level == 5
7420	     && !algorithm_valid_raid5(mddev->new_layout)) ||
7421	    (mddev->new_level == 6
7422	     && !algorithm_valid_raid6(mddev->new_layout))) {
7423		pr_warn("md/raid:%s: layout %d not supported\n",
7424			mdname(mddev), mddev->new_layout);
7425		return ERR_PTR(-EIO);
7426	}
7427	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7428		pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7429			mdname(mddev), mddev->raid_disks);
7430		return ERR_PTR(-EINVAL);
7431	}
7432
7433	if (!mddev->new_chunk_sectors ||
7434	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7435	    !is_power_of_2(mddev->new_chunk_sectors)) {
7436		pr_warn("md/raid:%s: invalid chunk size %d\n",
7437			mdname(mddev), mddev->new_chunk_sectors << 9);
7438		return ERR_PTR(-EINVAL);
7439	}
7440
7441	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7442	if (conf == NULL)
7443		goto abort;
7444
7445#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7446	conf->stripe_size = DEFAULT_STRIPE_SIZE;
7447	conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7448	conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7449#endif
7450	INIT_LIST_HEAD(&conf->free_list);
7451	INIT_LIST_HEAD(&conf->pending_list);
7452	conf->pending_data = kcalloc(PENDING_IO_MAX,
7453				     sizeof(struct r5pending_data),
7454				     GFP_KERNEL);
7455	if (!conf->pending_data)
7456		goto abort;
7457	for (i = 0; i < PENDING_IO_MAX; i++)
7458		list_add(&conf->pending_data[i].sibling, &conf->free_list);
7459	/* Don't enable multi-threading by default*/
7460	if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7461		conf->group_cnt = group_cnt;
7462		conf->worker_cnt_per_group = 0;
7463		conf->worker_groups = new_group;
7464	} else
7465		goto abort;
7466	spin_lock_init(&conf->device_lock);
7467	seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7468	mutex_init(&conf->cache_size_mutex);
7469
7470	init_waitqueue_head(&conf->wait_for_quiescent);
7471	init_waitqueue_head(&conf->wait_for_stripe);
7472	init_waitqueue_head(&conf->wait_for_overlap);
7473	INIT_LIST_HEAD(&conf->handle_list);
7474	INIT_LIST_HEAD(&conf->loprio_list);
7475	INIT_LIST_HEAD(&conf->hold_list);
7476	INIT_LIST_HEAD(&conf->delayed_list);
7477	INIT_LIST_HEAD(&conf->bitmap_list);
7478	init_llist_head(&conf->released_stripes);
7479	atomic_set(&conf->active_stripes, 0);
7480	atomic_set(&conf->preread_active_stripes, 0);
7481	atomic_set(&conf->active_aligned_reads, 0);
7482	spin_lock_init(&conf->pending_bios_lock);
7483	conf->batch_bio_dispatch = true;
7484	rdev_for_each(rdev, mddev) {
7485		if (test_bit(Journal, &rdev->flags))
7486			continue;
7487		if (bdev_nonrot(rdev->bdev)) {
7488			conf->batch_bio_dispatch = false;
7489			break;
7490		}
7491	}
7492
7493	conf->bypass_threshold = BYPASS_THRESHOLD;
7494	conf->recovery_disabled = mddev->recovery_disabled - 1;
7495
7496	conf->raid_disks = mddev->raid_disks;
7497	if (mddev->reshape_position == MaxSector)
7498		conf->previous_raid_disks = mddev->raid_disks;
7499	else
7500		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7501	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
 
7502
7503	conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7504			      GFP_KERNEL);
7505
7506	if (!conf->disks)
7507		goto abort;
7508
7509	for (i = 0; i < max_disks; i++) {
7510		conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7511		if (!conf->disks[i].extra_page)
7512			goto abort;
7513	}
7514
7515	ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7516	if (ret)
7517		goto abort;
7518	conf->mddev = mddev;
7519
7520	ret = -ENOMEM;
7521	conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL);
7522	if (!conf->stripe_hashtbl)
7523		goto abort;
7524
7525	/* We init hash_locks[0] separately to that it can be used
7526	 * as the reference lock in the spin_lock_nest_lock() call
7527	 * in lock_all_device_hash_locks_irq in order to convince
7528	 * lockdep that we know what we are doing.
7529	 */
7530	spin_lock_init(conf->hash_locks);
7531	for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7532		spin_lock_init(conf->hash_locks + i);
7533
7534	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7535		INIT_LIST_HEAD(conf->inactive_list + i);
7536
7537	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7538		INIT_LIST_HEAD(conf->temp_inactive_list + i);
7539
7540	atomic_set(&conf->r5c_cached_full_stripes, 0);
7541	INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7542	atomic_set(&conf->r5c_cached_partial_stripes, 0);
7543	INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7544	atomic_set(&conf->r5c_flushing_full_stripes, 0);
7545	atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7546
7547	conf->level = mddev->new_level;
7548	conf->chunk_sectors = mddev->new_chunk_sectors;
7549	ret = raid5_alloc_percpu(conf);
7550	if (ret)
7551		goto abort;
7552
7553	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7554
7555	ret = -EIO;
7556	rdev_for_each(rdev, mddev) {
7557		raid_disk = rdev->raid_disk;
7558		if (raid_disk >= max_disks
7559		    || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7560			continue;
7561		disk = conf->disks + raid_disk;
7562
7563		if (test_bit(Replacement, &rdev->flags)) {
7564			if (disk->replacement)
7565				goto abort;
7566			RCU_INIT_POINTER(disk->replacement, rdev);
7567		} else {
7568			if (disk->rdev)
7569				goto abort;
7570			RCU_INIT_POINTER(disk->rdev, rdev);
7571		}
7572
7573		if (test_bit(In_sync, &rdev->flags)) {
7574			pr_info("md/raid:%s: device %pg operational as raid disk %d\n",
7575				mdname(mddev), rdev->bdev, raid_disk);
 
 
7576		} else if (rdev->saved_raid_disk != raid_disk)
7577			/* Cannot rely on bitmap to complete recovery */
7578			conf->fullsync = 1;
7579	}
7580
 
7581	conf->level = mddev->new_level;
7582	if (conf->level == 6) {
7583		conf->max_degraded = 2;
7584		if (raid6_call.xor_syndrome)
7585			conf->rmw_level = PARITY_ENABLE_RMW;
7586		else
7587			conf->rmw_level = PARITY_DISABLE_RMW;
7588	} else {
7589		conf->max_degraded = 1;
7590		conf->rmw_level = PARITY_ENABLE_RMW;
7591	}
7592	conf->algorithm = mddev->new_layout;
 
7593	conf->reshape_progress = mddev->reshape_position;
7594	if (conf->reshape_progress != MaxSector) {
7595		conf->prev_chunk_sectors = mddev->chunk_sectors;
7596		conf->prev_algo = mddev->layout;
7597	} else {
7598		conf->prev_chunk_sectors = conf->chunk_sectors;
7599		conf->prev_algo = conf->algorithm;
7600	}
7601
7602	conf->min_nr_stripes = NR_STRIPES;
7603	if (mddev->reshape_position != MaxSector) {
7604		int stripes = max_t(int,
7605			((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7606			((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7607		conf->min_nr_stripes = max(NR_STRIPES, stripes);
7608		if (conf->min_nr_stripes != NR_STRIPES)
7609			pr_info("md/raid:%s: force stripe size %d for reshape\n",
7610				mdname(mddev), conf->min_nr_stripes);
7611	}
7612	memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7613		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7614	atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7615	if (grow_stripes(conf, conf->min_nr_stripes)) {
7616		pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7617			mdname(mddev), memory);
7618		ret = -ENOMEM;
7619		goto abort;
7620	} else
7621		pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7622	/*
7623	 * Losing a stripe head costs more than the time to refill it,
7624	 * it reduces the queue depth and so can hurt throughput.
7625	 * So set it rather large, scaled by number of devices.
7626	 */
7627	conf->shrinker = shrinker_alloc(0, "md-raid5:%s", mdname(mddev));
7628	if (!conf->shrinker) {
7629		ret = -ENOMEM;
7630		pr_warn("md/raid:%s: couldn't allocate shrinker.\n",
7631			mdname(mddev));
7632		goto abort;
7633	}
7634
7635	conf->shrinker->seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7636	conf->shrinker->scan_objects = raid5_cache_scan;
7637	conf->shrinker->count_objects = raid5_cache_count;
7638	conf->shrinker->batch = 128;
7639	conf->shrinker->private_data = conf;
7640
7641	shrinker_register(conf->shrinker);
7642
7643	sprintf(pers_name, "raid%d", mddev->new_level);
7644	rcu_assign_pointer(conf->thread,
7645			   md_register_thread(raid5d, mddev, pers_name));
7646	if (!conf->thread) {
7647		pr_warn("md/raid:%s: couldn't allocate thread.\n",
7648			mdname(mddev));
7649		ret = -ENOMEM;
7650		goto abort;
7651	}
7652
7653	return conf;
7654
7655 abort:
7656	if (conf)
7657		free_conf(conf);
7658	return ERR_PTR(ret);
 
 
7659}
7660
 
7661static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7662{
7663	switch (algo) {
7664	case ALGORITHM_PARITY_0:
7665		if (raid_disk < max_degraded)
7666			return 1;
7667		break;
7668	case ALGORITHM_PARITY_N:
7669		if (raid_disk >= raid_disks - max_degraded)
7670			return 1;
7671		break;
7672	case ALGORITHM_PARITY_0_6:
7673		if (raid_disk == 0 ||
7674		    raid_disk == raid_disks - 1)
7675			return 1;
7676		break;
7677	case ALGORITHM_LEFT_ASYMMETRIC_6:
7678	case ALGORITHM_RIGHT_ASYMMETRIC_6:
7679	case ALGORITHM_LEFT_SYMMETRIC_6:
7680	case ALGORITHM_RIGHT_SYMMETRIC_6:
7681		if (raid_disk == raid_disks - 1)
7682			return 1;
7683	}
7684	return 0;
7685}
7686
7687static void raid5_set_io_opt(struct r5conf *conf)
7688{
7689	blk_queue_io_opt(conf->mddev->queue, (conf->chunk_sectors << 9) *
7690			 (conf->raid_disks - conf->max_degraded));
7691}
7692
7693static int raid5_run(struct mddev *mddev)
7694{
7695	struct r5conf *conf;
7696	int dirty_parity_disks = 0;
7697	struct md_rdev *rdev;
7698	struct md_rdev *journal_dev = NULL;
7699	sector_t reshape_offset = 0;
7700	int i;
7701	long long min_offset_diff = 0;
7702	int first = 1;
7703
7704	if (mddev->recovery_cp != MaxSector)
7705		pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7706			  mdname(mddev));
7707
7708	rdev_for_each(rdev, mddev) {
7709		long long diff;
7710
7711		if (test_bit(Journal, &rdev->flags)) {
7712			journal_dev = rdev;
7713			continue;
7714		}
7715		if (rdev->raid_disk < 0)
7716			continue;
7717		diff = (rdev->new_data_offset - rdev->data_offset);
7718		if (first) {
7719			min_offset_diff = diff;
7720			first = 0;
7721		} else if (mddev->reshape_backwards &&
7722			 diff < min_offset_diff)
7723			min_offset_diff = diff;
7724		else if (!mddev->reshape_backwards &&
7725			 diff > min_offset_diff)
7726			min_offset_diff = diff;
7727	}
7728
7729	if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7730	    (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7731		pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7732			  mdname(mddev));
7733		return -EINVAL;
7734	}
7735
7736	if (mddev->reshape_position != MaxSector) {
7737		/* Check that we can continue the reshape.
7738		 * Difficulties arise if the stripe we would write to
7739		 * next is at or after the stripe we would read from next.
7740		 * For a reshape that changes the number of devices, this
7741		 * is only possible for a very short time, and mdadm makes
7742		 * sure that time appears to have past before assembling
7743		 * the array.  So we fail if that time hasn't passed.
7744		 * For a reshape that keeps the number of devices the same
7745		 * mdadm must be monitoring the reshape can keeping the
7746		 * critical areas read-only and backed up.  It will start
7747		 * the array in read-only mode, so we check for that.
7748		 */
7749		sector_t here_new, here_old;
7750		int old_disks;
7751		int max_degraded = (mddev->level == 6 ? 2 : 1);
7752		int chunk_sectors;
7753		int new_data_disks;
7754
7755		if (journal_dev) {
7756			pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7757				mdname(mddev));
7758			return -EINVAL;
7759		}
7760
7761		if (mddev->new_level != mddev->level) {
7762			pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7763				mdname(mddev));
 
7764			return -EINVAL;
7765		}
7766		old_disks = mddev->raid_disks - mddev->delta_disks;
7767		/* reshape_position must be on a new-stripe boundary, and one
7768		 * further up in new geometry must map after here in old
7769		 * geometry.
7770		 * If the chunk sizes are different, then as we perform reshape
7771		 * in units of the largest of the two, reshape_position needs
7772		 * be a multiple of the largest chunk size times new data disks.
7773		 */
7774		here_new = mddev->reshape_position;
7775		chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7776		new_data_disks = mddev->raid_disks - max_degraded;
7777		if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7778			pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7779				mdname(mddev));
7780			return -EINVAL;
7781		}
7782		reshape_offset = here_new * chunk_sectors;
7783		/* here_new is the stripe we will write to */
7784		here_old = mddev->reshape_position;
7785		sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
 
7786		/* here_old is the first stripe that we might need to read
7787		 * from */
7788		if (mddev->delta_disks == 0) {
7789			/* We cannot be sure it is safe to start an in-place
7790			 * reshape.  It is only safe if user-space is monitoring
7791			 * and taking constant backups.
7792			 * mdadm always starts a situation like this in
7793			 * readonly mode so it can take control before
7794			 * allowing any writes.  So just check for that.
7795			 */
7796			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7797			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
7798				/* not really in-place - so OK */;
7799			else if (mddev->ro == 0) {
7800				pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7801					mdname(mddev));
7802				return -EINVAL;
7803			}
7804		} else if (mddev->reshape_backwards
7805		    ? (here_new * chunk_sectors + min_offset_diff <=
7806		       here_old * chunk_sectors)
7807		    : (here_new * chunk_sectors >=
7808		       here_old * chunk_sectors + (-min_offset_diff))) {
7809			/* Reading from the same stripe as writing to - bad */
7810			pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7811				mdname(mddev));
 
7812			return -EINVAL;
7813		}
7814		pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
 
7815		/* OK, we should be able to continue; */
7816	} else {
7817		BUG_ON(mddev->level != mddev->new_level);
7818		BUG_ON(mddev->layout != mddev->new_layout);
7819		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7820		BUG_ON(mddev->delta_disks != 0);
7821	}
7822
7823	if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7824	    test_bit(MD_HAS_PPL, &mddev->flags)) {
7825		pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7826			mdname(mddev));
7827		clear_bit(MD_HAS_PPL, &mddev->flags);
7828		clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7829	}
7830
7831	if (mddev->private == NULL)
7832		conf = setup_conf(mddev);
7833	else
7834		conf = mddev->private;
7835
7836	if (IS_ERR(conf))
7837		return PTR_ERR(conf);
7838
7839	if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7840		if (!journal_dev) {
7841			pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7842				mdname(mddev));
7843			mddev->ro = 1;
7844			set_disk_ro(mddev->gendisk, 1);
7845		} else if (mddev->recovery_cp == MaxSector)
7846			set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7847	}
7848
7849	conf->min_offset_diff = min_offset_diff;
7850	rcu_assign_pointer(mddev->thread, conf->thread);
7851	rcu_assign_pointer(conf->thread, NULL);
7852	mddev->private = conf;
7853
7854	for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7855	     i++) {
7856		rdev = conf->disks[i].rdev;
7857		if (!rdev)
 
7858			continue;
7859		if (conf->disks[i].replacement &&
7860		    conf->reshape_progress != MaxSector) {
7861			/* replacements and reshape simply do not mix. */
7862			pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7863			goto abort;
7864		}
7865		if (test_bit(In_sync, &rdev->flags))
7866			continue;
 
7867		/* This disc is not fully in-sync.  However if it
7868		 * just stored parity (beyond the recovery_offset),
7869		 * when we don't need to be concerned about the
7870		 * array being dirty.
7871		 * When reshape goes 'backwards', we never have
7872		 * partially completed devices, so we only need
7873		 * to worry about reshape going forwards.
7874		 */
7875		/* Hack because v0.91 doesn't store recovery_offset properly. */
7876		if (mddev->major_version == 0 &&
7877		    mddev->minor_version > 90)
7878			rdev->recovery_offset = reshape_offset;
7879
7880		if (rdev->recovery_offset < reshape_offset) {
7881			/* We need to check old and new layout */
7882			if (!only_parity(rdev->raid_disk,
7883					 conf->algorithm,
7884					 conf->raid_disks,
7885					 conf->max_degraded))
7886				continue;
7887		}
7888		if (!only_parity(rdev->raid_disk,
7889				 conf->prev_algo,
7890				 conf->previous_raid_disks,
7891				 conf->max_degraded))
7892			continue;
7893		dirty_parity_disks++;
7894	}
7895
7896	/*
7897	 * 0 for a fully functional array, 1 or 2 for a degraded array.
7898	 */
7899	mddev->degraded = raid5_calc_degraded(conf);
7900
7901	if (has_failed(conf)) {
7902		pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
 
7903			mdname(mddev), mddev->degraded, conf->raid_disks);
7904		goto abort;
7905	}
7906
7907	/* device size must be a multiple of chunk size */
7908	mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7909	mddev->resync_max_sectors = mddev->dev_sectors;
7910
7911	if (mddev->degraded > dirty_parity_disks &&
7912	    mddev->recovery_cp != MaxSector) {
7913		if (test_bit(MD_HAS_PPL, &mddev->flags))
7914			pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7915				mdname(mddev));
7916		else if (mddev->ok_start_degraded)
7917			pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7918				mdname(mddev));
7919		else {
7920			pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7921				mdname(mddev));
 
7922			goto abort;
7923		}
7924	}
7925
7926	pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7927		mdname(mddev), conf->level,
7928		mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7929		mddev->new_layout);
 
 
 
 
 
 
 
7930
7931	print_raid5_conf(conf);
7932
7933	if (conf->reshape_progress != MaxSector) {
7934		conf->reshape_safe = conf->reshape_progress;
7935		atomic_set(&conf->reshape_stripes, 0);
7936		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7937		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7938		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7939		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 
 
7940	}
7941
 
7942	/* Ok, everything is just fine now */
7943	if (mddev->to_remove == &raid5_attrs_group)
7944		mddev->to_remove = NULL;
7945	else if (mddev->kobj.sd &&
7946	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7947		pr_warn("raid5: failed to create sysfs attributes for %s\n",
7948			mdname(mddev));
 
7949	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7950
7951	if (mddev->queue) {
7952		int chunk_size;
7953		/* read-ahead size must cover two whole stripes, which
7954		 * is 2 * (datadisks) * chunksize where 'n' is the
7955		 * number of raid devices
7956		 */
7957		int data_disks = conf->previous_raid_disks - conf->max_degraded;
7958		int stripe = data_disks *
7959			((mddev->chunk_sectors << 9) / PAGE_SIZE);
 
 
 
 
 
 
 
7960
7961		chunk_size = mddev->chunk_sectors << 9;
7962		blk_queue_io_min(mddev->queue, chunk_size);
7963		raid5_set_io_opt(conf);
7964		mddev->queue->limits.raid_partial_stripes_expensive = 1;
7965		/*
7966		 * We can only discard a whole stripe. It doesn't make sense to
7967		 * discard data disk but write parity disk
7968		 */
7969		stripe = stripe * PAGE_SIZE;
7970		stripe = roundup_pow_of_two(stripe);
7971		mddev->queue->limits.discard_granularity = stripe;
7972
7973		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7974
7975		rdev_for_each(rdev, mddev) {
7976			disk_stack_limits(mddev->gendisk, rdev->bdev,
7977					  rdev->data_offset << 9);
7978			disk_stack_limits(mddev->gendisk, rdev->bdev,
7979					  rdev->new_data_offset << 9);
7980		}
7981
7982		/*
7983		 * zeroing is required, otherwise data
7984		 * could be lost. Consider a scenario: discard a stripe
7985		 * (the stripe could be inconsistent if
7986		 * discard_zeroes_data is 0); write one disk of the
7987		 * stripe (the stripe could be inconsistent again
7988		 * depending on which disks are used to calculate
7989		 * parity); the disk is broken; The stripe data of this
7990		 * disk is lost.
7991		 *
7992		 * We only allow DISCARD if the sysadmin has confirmed that
7993		 * only safe devices are in use by setting a module parameter.
7994		 * A better idea might be to turn DISCARD into WRITE_ZEROES
7995		 * requests, as that is required to be safe.
7996		 */
7997		if (!devices_handle_discard_safely ||
7998		    mddev->queue->limits.max_discard_sectors < (stripe >> 9) ||
7999		    mddev->queue->limits.discard_granularity < stripe)
8000			blk_queue_max_discard_sectors(mddev->queue, 0);
8001
8002		/*
8003		 * Requests require having a bitmap for each stripe.
8004		 * Limit the max sectors based on this.
8005		 */
8006		blk_queue_max_hw_sectors(mddev->queue,
8007			RAID5_MAX_REQ_STRIPES << RAID5_STRIPE_SHIFT(conf));
8008
8009		/* No restrictions on the number of segments in the request */
8010		blk_queue_max_segments(mddev->queue, USHRT_MAX);
8011	}
8012
8013	if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
8014		goto abort;
8015
8016	return 0;
8017abort:
8018	md_unregister_thread(mddev, &mddev->thread);
8019	print_raid5_conf(conf);
8020	free_conf(conf);
 
 
8021	mddev->private = NULL;
8022	pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
8023	return -EIO;
8024}
8025
8026static void raid5_free(struct mddev *mddev, void *priv)
8027{
8028	struct r5conf *conf = priv;
8029
 
 
 
8030	free_conf(conf);
 
8031	mddev->to_remove = &raid5_attrs_group;
 
8032}
8033
8034static void raid5_status(struct seq_file *seq, struct mddev *mddev)
 
8035{
8036	struct r5conf *conf = mddev->private;
8037	int i;
8038
8039	lockdep_assert_held(&mddev->lock);
 
 
 
 
 
 
 
 
 
 
8040
8041	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
8042		conf->chunk_sectors / 2, mddev->layout);
8043	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
8044	for (i = 0; i < conf->raid_disks; i++) {
8045		struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
8046
8047		seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 
 
 
 
 
 
8048	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8049	seq_printf (seq, "]");
 
 
 
 
8050}
8051
8052static void print_raid5_conf (struct r5conf *conf)
8053{
8054	struct md_rdev *rdev;
8055	int i;
 
8056
8057	pr_debug("RAID conf printout:\n");
8058	if (!conf) {
8059		pr_debug("(conf==NULL)\n");
8060		return;
8061	}
8062	pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
8063	       conf->raid_disks,
8064	       conf->raid_disks - conf->mddev->degraded);
8065
8066	rcu_read_lock();
8067	for (i = 0; i < conf->raid_disks; i++) {
8068		rdev = rcu_dereference(conf->disks[i].rdev);
8069		if (rdev)
8070			pr_debug(" disk %d, o:%d, dev:%pg\n",
8071			       i, !test_bit(Faulty, &rdev->flags),
8072			       rdev->bdev);
 
8073	}
8074	rcu_read_unlock();
8075}
8076
8077static int raid5_spare_active(struct mddev *mddev)
8078{
8079	int i;
8080	struct r5conf *conf = mddev->private;
8081	struct md_rdev *rdev, *replacement;
8082	int count = 0;
8083	unsigned long flags;
8084
8085	for (i = 0; i < conf->raid_disks; i++) {
8086		rdev = conf->disks[i].rdev;
8087		replacement = conf->disks[i].replacement;
8088		if (replacement
8089		    && replacement->recovery_offset == MaxSector
8090		    && !test_bit(Faulty, &replacement->flags)
8091		    && !test_and_set_bit(In_sync, &replacement->flags)) {
8092			/* Replacement has just become active. */
8093			if (!rdev
8094			    || !test_and_clear_bit(In_sync, &rdev->flags))
8095				count++;
8096			if (rdev) {
8097				/* Replaced device not technically faulty,
8098				 * but we need to be sure it gets removed
8099				 * and never re-added.
8100				 */
8101				set_bit(Faulty, &rdev->flags);
8102				sysfs_notify_dirent_safe(
8103					rdev->sysfs_state);
8104			}
8105			sysfs_notify_dirent_safe(replacement->sysfs_state);
8106		} else if (rdev
8107		    && rdev->recovery_offset == MaxSector
8108		    && !test_bit(Faulty, &rdev->flags)
8109		    && !test_and_set_bit(In_sync, &rdev->flags)) {
8110			count++;
8111			sysfs_notify_dirent_safe(rdev->sysfs_state);
8112		}
8113	}
8114	spin_lock_irqsave(&conf->device_lock, flags);
8115	mddev->degraded = raid5_calc_degraded(conf);
8116	spin_unlock_irqrestore(&conf->device_lock, flags);
8117	print_raid5_conf(conf);
8118	return count;
8119}
8120
8121static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
8122{
8123	struct r5conf *conf = mddev->private;
8124	int err = 0;
8125	int number = rdev->raid_disk;
8126	struct md_rdev **rdevp;
8127	struct disk_info *p;
8128	struct md_rdev *tmp;
8129
8130	print_raid5_conf(conf);
8131	if (test_bit(Journal, &rdev->flags) && conf->log) {
8132		/*
8133		 * we can't wait pending write here, as this is called in
8134		 * raid5d, wait will deadlock.
8135		 * neilb: there is no locking about new writes here,
8136		 * so this cannot be safe.
8137		 */
8138		if (atomic_read(&conf->active_stripes) ||
8139		    atomic_read(&conf->r5c_cached_full_stripes) ||
8140		    atomic_read(&conf->r5c_cached_partial_stripes)) {
8141			return -EBUSY;
8142		}
8143		log_exit(conf);
8144		return 0;
8145	}
8146	if (unlikely(number >= conf->pool_size))
8147		return 0;
8148	p = conf->disks + number;
8149	if (rdev == p->rdev)
8150		rdevp = &p->rdev;
8151	else if (rdev == p->replacement)
8152		rdevp = &p->replacement;
8153	else
8154		return 0;
8155
8156	if (number >= conf->raid_disks &&
8157	    conf->reshape_progress == MaxSector)
8158		clear_bit(In_sync, &rdev->flags);
8159
8160	if (test_bit(In_sync, &rdev->flags) ||
8161	    atomic_read(&rdev->nr_pending)) {
8162		err = -EBUSY;
8163		goto abort;
8164	}
8165	/* Only remove non-faulty devices if recovery
8166	 * isn't possible.
8167	 */
8168	if (!test_bit(Faulty, &rdev->flags) &&
8169	    mddev->recovery_disabled != conf->recovery_disabled &&
8170	    !has_failed(conf) &&
8171	    (!p->replacement || p->replacement == rdev) &&
8172	    number < conf->raid_disks) {
8173		err = -EBUSY;
8174		goto abort;
8175	}
8176	WRITE_ONCE(*rdevp, NULL);
8177	if (!err) {
8178		err = log_modify(conf, rdev, false);
8179		if (err)
8180			goto abort;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8181	}
8182
8183	tmp = p->replacement;
8184	if (tmp) {
8185		/* We must have just cleared 'rdev' */
8186		WRITE_ONCE(p->rdev, tmp);
8187		clear_bit(Replacement, &tmp->flags);
8188		WRITE_ONCE(p->replacement, NULL);
8189
8190		if (!err)
8191			err = log_modify(conf, tmp, true);
8192	}
8193
8194	clear_bit(WantReplacement, &rdev->flags);
8195abort:
8196
8197	print_raid5_conf(conf);
8198	return err;
8199}
8200
8201static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
8202{
8203	struct r5conf *conf = mddev->private;
8204	int ret, err = -EEXIST;
8205	int disk;
8206	struct disk_info *p;
8207	struct md_rdev *tmp;
8208	int first = 0;
8209	int last = conf->raid_disks - 1;
8210
8211	if (test_bit(Journal, &rdev->flags)) {
8212		if (conf->log)
8213			return -EBUSY;
8214
8215		rdev->raid_disk = 0;
8216		/*
8217		 * The array is in readonly mode if journal is missing, so no
8218		 * write requests running. We should be safe
8219		 */
8220		ret = log_init(conf, rdev, false);
8221		if (ret)
8222			return ret;
8223
8224		ret = r5l_start(conf->log);
8225		if (ret)
8226			return ret;
8227
8228		return 0;
8229	}
8230	if (mddev->recovery_disabled == conf->recovery_disabled)
8231		return -EBUSY;
8232
8233	if (rdev->saved_raid_disk < 0 && has_failed(conf))
8234		/* no point adding a device */
8235		return -EINVAL;
8236
8237	if (rdev->raid_disk >= 0)
8238		first = last = rdev->raid_disk;
8239
8240	/*
8241	 * find the disk ... but prefer rdev->saved_raid_disk
8242	 * if possible.
8243	 */
8244	if (rdev->saved_raid_disk >= first &&
8245	    rdev->saved_raid_disk <= last &&
8246	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
8247		first = rdev->saved_raid_disk;
8248
8249	for (disk = first; disk <= last; disk++) {
8250		p = conf->disks + disk;
8251		if (p->rdev == NULL) {
8252			clear_bit(In_sync, &rdev->flags);
8253			rdev->raid_disk = disk;
 
8254			if (rdev->saved_raid_disk != disk)
8255				conf->fullsync = 1;
8256			WRITE_ONCE(p->rdev, rdev);
8257
8258			err = log_modify(conf, rdev, true);
8259
8260			goto out;
8261		}
8262	}
8263	for (disk = first; disk <= last; disk++) {
8264		p = conf->disks + disk;
8265		tmp = p->rdev;
8266		if (test_bit(WantReplacement, &tmp->flags) &&
8267		    mddev->reshape_position == MaxSector &&
8268		    p->replacement == NULL) {
8269			clear_bit(In_sync, &rdev->flags);
8270			set_bit(Replacement, &rdev->flags);
8271			rdev->raid_disk = disk;
8272			err = 0;
8273			conf->fullsync = 1;
8274			WRITE_ONCE(p->replacement, rdev);
8275			break;
8276		}
8277	}
8278out:
8279	print_raid5_conf(conf);
8280	return err;
8281}
8282
8283static int raid5_resize(struct mddev *mddev, sector_t sectors)
8284{
8285	/* no resync is happening, and there is enough space
8286	 * on all devices, so we can resize.
8287	 * We need to make sure resync covers any new space.
8288	 * If the array is shrinking we should possibly wait until
8289	 * any io in the removed space completes, but it hardly seems
8290	 * worth it.
8291	 */
8292	sector_t newsize;
8293	struct r5conf *conf = mddev->private;
8294
8295	if (raid5_has_log(conf) || raid5_has_ppl(conf))
8296		return -EINVAL;
8297	sectors &= ~((sector_t)conf->chunk_sectors - 1);
8298	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8299	if (mddev->external_size &&
8300	    mddev->array_sectors > newsize)
8301		return -EINVAL;
8302	if (mddev->bitmap) {
8303		int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
8304		if (ret)
8305			return ret;
8306	}
8307	md_set_array_sectors(mddev, newsize);
8308	if (sectors > mddev->dev_sectors &&
8309	    mddev->recovery_cp > mddev->dev_sectors) {
8310		mddev->recovery_cp = mddev->dev_sectors;
8311		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8312	}
8313	mddev->dev_sectors = sectors;
8314	mddev->resync_max_sectors = sectors;
8315	return 0;
8316}
8317
8318static int check_stripe_cache(struct mddev *mddev)
8319{
8320	/* Can only proceed if there are plenty of stripe_heads.
8321	 * We need a minimum of one full stripe,, and for sensible progress
8322	 * it is best to have about 4 times that.
8323	 * If we require 4 times, then the default 256 4K stripe_heads will
8324	 * allow for chunk sizes up to 256K, which is probably OK.
8325	 * If the chunk size is greater, user-space should request more
8326	 * stripe_heads first.
8327	 */
8328	struct r5conf *conf = mddev->private;
8329	if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8330	    > conf->min_nr_stripes ||
8331	    ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8332	    > conf->min_nr_stripes) {
8333		pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
8334			mdname(mddev),
8335			((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8336			 / RAID5_STRIPE_SIZE(conf))*4);
8337		return 0;
8338	}
8339	return 1;
8340}
8341
8342static int check_reshape(struct mddev *mddev)
8343{
8344	struct r5conf *conf = mddev->private;
8345
8346	if (raid5_has_log(conf) || raid5_has_ppl(conf))
8347		return -EINVAL;
8348	if (mddev->delta_disks == 0 &&
8349	    mddev->new_layout == mddev->layout &&
8350	    mddev->new_chunk_sectors == mddev->chunk_sectors)
8351		return 0; /* nothing to do */
 
 
 
8352	if (has_failed(conf))
8353		return -EINVAL;
8354	if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8355		/* We might be able to shrink, but the devices must
8356		 * be made bigger first.
8357		 * For raid6, 4 is the minimum size.
8358		 * Otherwise 2 is the minimum
8359		 */
8360		int min = 2;
8361		if (mddev->level == 6)
8362			min = 4;
8363		if (mddev->raid_disks + mddev->delta_disks < min)
8364			return -EINVAL;
8365	}
8366
8367	if (!check_stripe_cache(mddev))
8368		return -ENOSPC;
8369
8370	if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8371	    mddev->delta_disks > 0)
8372		if (resize_chunks(conf,
8373				  conf->previous_raid_disks
8374				  + max(0, mddev->delta_disks),
8375				  max(mddev->new_chunk_sectors,
8376				      mddev->chunk_sectors)
8377			    ) < 0)
8378			return -ENOMEM;
8379
8380	if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8381		return 0; /* never bother to shrink */
8382	return resize_stripes(conf, (conf->previous_raid_disks
8383				     + mddev->delta_disks));
8384}
8385
8386static int raid5_start_reshape(struct mddev *mddev)
8387{
8388	struct r5conf *conf = mddev->private;
8389	struct md_rdev *rdev;
8390	int spares = 0;
8391	int i;
8392	unsigned long flags;
8393
8394	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8395		return -EBUSY;
8396
8397	if (!check_stripe_cache(mddev))
8398		return -ENOSPC;
8399
8400	if (has_failed(conf))
8401		return -EINVAL;
8402
8403	/* raid5 can't handle concurrent reshape and recovery */
8404	if (mddev->recovery_cp < MaxSector)
8405		return -EBUSY;
8406	for (i = 0; i < conf->raid_disks; i++)
8407		if (conf->disks[i].replacement)
8408			return -EBUSY;
8409
8410	rdev_for_each(rdev, mddev) {
8411		if (!test_bit(In_sync, &rdev->flags)
8412		    && !test_bit(Faulty, &rdev->flags))
8413			spares++;
8414	}
8415
8416	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8417		/* Not enough devices even to make a degraded array
8418		 * of that size
8419		 */
8420		return -EINVAL;
8421
8422	/* Refuse to reduce size of the array.  Any reductions in
8423	 * array size must be through explicit setting of array_size
8424	 * attribute.
8425	 */
8426	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8427	    < mddev->array_sectors) {
8428		pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8429			mdname(mddev));
8430		return -EINVAL;
8431	}
8432
8433	atomic_set(&conf->reshape_stripes, 0);
8434	spin_lock_irq(&conf->device_lock);
8435	write_seqcount_begin(&conf->gen_lock);
8436	conf->previous_raid_disks = conf->raid_disks;
8437	conf->raid_disks += mddev->delta_disks;
8438	conf->prev_chunk_sectors = conf->chunk_sectors;
8439	conf->chunk_sectors = mddev->new_chunk_sectors;
8440	conf->prev_algo = conf->algorithm;
8441	conf->algorithm = mddev->new_layout;
8442	conf->generation++;
8443	/* Code that selects data_offset needs to see the generation update
8444	 * if reshape_progress has been set - so a memory barrier needed.
8445	 */
8446	smp_mb();
8447	if (mddev->reshape_backwards)
8448		conf->reshape_progress = raid5_size(mddev, 0, 0);
8449	else
8450		conf->reshape_progress = 0;
8451	conf->reshape_safe = conf->reshape_progress;
8452	write_seqcount_end(&conf->gen_lock);
8453	spin_unlock_irq(&conf->device_lock);
8454
8455	/* Now make sure any requests that proceeded on the assumption
8456	 * the reshape wasn't running - like Discard or Read - have
8457	 * completed.
8458	 */
8459	raid5_quiesce(mddev, true);
8460	raid5_quiesce(mddev, false);
8461
8462	/* Add some new drives, as many as will fit.
8463	 * We know there are enough to make the newly sized array work.
8464	 * Don't add devices if we are reducing the number of
8465	 * devices in the array.  This is because it is not possible
8466	 * to correctly record the "partially reconstructed" state of
8467	 * such devices during the reshape and confusion could result.
8468	 */
8469	if (mddev->delta_disks >= 0) {
8470		rdev_for_each(rdev, mddev)
 
8471			if (rdev->raid_disk < 0 &&
8472			    !test_bit(Faulty, &rdev->flags)) {
8473				if (raid5_add_disk(mddev, rdev) == 0) {
8474					if (rdev->raid_disk
8475					    >= conf->previous_raid_disks)
8476						set_bit(In_sync, &rdev->flags);
8477					else
 
8478						rdev->recovery_offset = 0;
8479
8480					/* Failure here is OK */
8481					sysfs_link_rdev(mddev, rdev);
8482				}
8483			} else if (rdev->raid_disk >= conf->previous_raid_disks
8484				   && !test_bit(Faulty, &rdev->flags)) {
8485				/* This is a spare that was manually added */
8486				set_bit(In_sync, &rdev->flags);
 
8487			}
8488
8489		/* When a reshape changes the number of devices,
8490		 * ->degraded is measured against the larger of the
8491		 * pre and post number of devices.
8492		 */
8493		spin_lock_irqsave(&conf->device_lock, flags);
8494		mddev->degraded = raid5_calc_degraded(conf);
 
8495		spin_unlock_irqrestore(&conf->device_lock, flags);
8496	}
8497	mddev->raid_disks = conf->raid_disks;
8498	mddev->reshape_position = conf->reshape_progress;
8499	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8500
8501	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8502	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8503	clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8504	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8505	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 
 
 
 
 
 
 
 
 
 
8506	conf->reshape_checkpoint = jiffies;
8507	md_new_event();
 
8508	return 0;
8509}
8510
8511/* This is called from the reshape thread and should make any
8512 * changes needed in 'conf'
8513 */
8514static void end_reshape(struct r5conf *conf)
8515{
8516
8517	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8518		struct md_rdev *rdev;
8519
8520		spin_lock_irq(&conf->device_lock);
8521		conf->previous_raid_disks = conf->raid_disks;
8522		md_finish_reshape(conf->mddev);
8523		smp_wmb();
8524		conf->reshape_progress = MaxSector;
8525		conf->mddev->reshape_position = MaxSector;
8526		rdev_for_each(rdev, conf->mddev)
8527			if (rdev->raid_disk >= 0 &&
8528			    !test_bit(Journal, &rdev->flags) &&
8529			    !test_bit(In_sync, &rdev->flags))
8530				rdev->recovery_offset = MaxSector;
8531		spin_unlock_irq(&conf->device_lock);
8532		wake_up(&conf->wait_for_overlap);
8533
8534		if (conf->mddev->queue)
8535			raid5_set_io_opt(conf);
 
 
 
 
 
 
 
 
8536	}
8537}
8538
8539/* This is called from the raid5d thread with mddev_lock held.
8540 * It makes config changes to the device.
8541 */
8542static void raid5_finish_reshape(struct mddev *mddev)
8543{
8544	struct r5conf *conf = mddev->private;
8545	struct md_rdev *rdev;
8546
8547	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8548
8549		if (mddev->delta_disks <= 0) {
 
 
 
 
8550			int d;
8551			spin_lock_irq(&conf->device_lock);
8552			mddev->degraded = raid5_calc_degraded(conf);
8553			spin_unlock_irq(&conf->device_lock);
 
 
 
8554			for (d = conf->raid_disks ;
8555			     d < conf->raid_disks - mddev->delta_disks;
8556			     d++) {
8557				rdev = conf->disks[d].rdev;
8558				if (rdev)
8559					clear_bit(In_sync, &rdev->flags);
8560				rdev = conf->disks[d].replacement;
8561				if (rdev)
8562					clear_bit(In_sync, &rdev->flags);
8563			}
8564		}
8565		mddev->layout = conf->algorithm;
8566		mddev->chunk_sectors = conf->chunk_sectors;
8567		mddev->reshape_position = MaxSector;
8568		mddev->delta_disks = 0;
8569		mddev->reshape_backwards = 0;
8570	}
8571}
8572
8573static void raid5_quiesce(struct mddev *mddev, int quiesce)
8574{
8575	struct r5conf *conf = mddev->private;
8576
8577	if (quiesce) {
8578		/* stop all writes */
8579		lock_all_device_hash_locks_irq(conf);
 
 
 
 
8580		/* '2' tells resync/reshape to pause so that all
8581		 * active stripes can drain
8582		 */
8583		r5c_flush_cache(conf, INT_MAX);
8584		/* need a memory barrier to make sure read_one_chunk() sees
8585		 * quiesce started and reverts to slow (locked) path.
8586		 */
8587		smp_store_release(&conf->quiesce, 2);
8588		wait_event_cmd(conf->wait_for_quiescent,
8589				    atomic_read(&conf->active_stripes) == 0 &&
8590				    atomic_read(&conf->active_aligned_reads) == 0,
8591				    unlock_all_device_hash_locks_irq(conf),
8592				    lock_all_device_hash_locks_irq(conf));
8593		conf->quiesce = 1;
8594		unlock_all_device_hash_locks_irq(conf);
8595		/* allow reshape to continue */
8596		wake_up(&conf->wait_for_overlap);
8597	} else {
8598		/* re-enable writes */
8599		lock_all_device_hash_locks_irq(conf);
 
8600		conf->quiesce = 0;
8601		wake_up(&conf->wait_for_quiescent);
8602		wake_up(&conf->wait_for_overlap);
8603		unlock_all_device_hash_locks_irq(conf);
 
8604	}
8605	log_quiesce(conf, quiesce);
8606}
8607
8608static void *raid45_takeover_raid0(struct mddev *mddev, int level)
 
8609{
8610	struct r0conf *raid0_conf = mddev->private;
8611	sector_t sectors;
8612
8613	/* for raid0 takeover only one zone is supported */
8614	if (raid0_conf->nr_strip_zones > 1) {
8615		pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8616			mdname(mddev));
8617		return ERR_PTR(-EINVAL);
8618	}
8619
8620	sectors = raid0_conf->strip_zone[0].zone_end;
8621	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8622	mddev->dev_sectors = sectors;
8623	mddev->new_level = level;
8624	mddev->new_layout = ALGORITHM_PARITY_N;
8625	mddev->new_chunk_sectors = mddev->chunk_sectors;
8626	mddev->raid_disks += 1;
8627	mddev->delta_disks = 1;
8628	/* make sure it will be not marked as dirty */
8629	mddev->recovery_cp = MaxSector;
8630
8631	return setup_conf(mddev);
8632}
8633
8634static void *raid5_takeover_raid1(struct mddev *mddev)
 
8635{
8636	int chunksect;
8637	void *ret;
8638
8639	if (mddev->raid_disks != 2 ||
8640	    mddev->degraded > 1)
8641		return ERR_PTR(-EINVAL);
8642
8643	/* Should check if there are write-behind devices? */
8644
8645	chunksect = 64*2; /* 64K by default */
8646
8647	/* The array must be an exact multiple of chunksize */
8648	while (chunksect && (mddev->array_sectors & (chunksect-1)))
8649		chunksect >>= 1;
8650
8651	if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8652		/* array size does not allow a suitable chunk size */
8653		return ERR_PTR(-EINVAL);
8654
8655	mddev->new_level = 5;
8656	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8657	mddev->new_chunk_sectors = chunksect;
8658
8659	ret = setup_conf(mddev);
8660	if (!IS_ERR(ret))
8661		mddev_clear_unsupported_flags(mddev,
8662			UNSUPPORTED_MDDEV_FLAGS);
8663	return ret;
8664}
8665
8666static void *raid5_takeover_raid6(struct mddev *mddev)
8667{
8668	int new_layout;
8669
8670	switch (mddev->layout) {
8671	case ALGORITHM_LEFT_ASYMMETRIC_6:
8672		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8673		break;
8674	case ALGORITHM_RIGHT_ASYMMETRIC_6:
8675		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8676		break;
8677	case ALGORITHM_LEFT_SYMMETRIC_6:
8678		new_layout = ALGORITHM_LEFT_SYMMETRIC;
8679		break;
8680	case ALGORITHM_RIGHT_SYMMETRIC_6:
8681		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8682		break;
8683	case ALGORITHM_PARITY_0_6:
8684		new_layout = ALGORITHM_PARITY_0;
8685		break;
8686	case ALGORITHM_PARITY_N:
8687		new_layout = ALGORITHM_PARITY_N;
8688		break;
8689	default:
8690		return ERR_PTR(-EINVAL);
8691	}
8692	mddev->new_level = 5;
8693	mddev->new_layout = new_layout;
8694	mddev->delta_disks = -1;
8695	mddev->raid_disks -= 1;
8696	return setup_conf(mddev);
8697}
8698
8699static int raid5_check_reshape(struct mddev *mddev)
 
8700{
8701	/* For a 2-drive array, the layout and chunk size can be changed
8702	 * immediately as not restriping is needed.
8703	 * For larger arrays we record the new value - after validation
8704	 * to be used by a reshape pass.
8705	 */
8706	struct r5conf *conf = mddev->private;
8707	int new_chunk = mddev->new_chunk_sectors;
8708
8709	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8710		return -EINVAL;
8711	if (new_chunk > 0) {
8712		if (!is_power_of_2(new_chunk))
8713			return -EINVAL;
8714		if (new_chunk < (PAGE_SIZE>>9))
8715			return -EINVAL;
8716		if (mddev->array_sectors & (new_chunk-1))
8717			/* not factor of array size */
8718			return -EINVAL;
8719	}
8720
8721	/* They look valid */
8722
8723	if (mddev->raid_disks == 2) {
8724		/* can make the change immediately */
8725		if (mddev->new_layout >= 0) {
8726			conf->algorithm = mddev->new_layout;
8727			mddev->layout = mddev->new_layout;
8728		}
8729		if (new_chunk > 0) {
8730			conf->chunk_sectors = new_chunk ;
8731			mddev->chunk_sectors = new_chunk;
8732		}
8733		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8734		md_wakeup_thread(mddev->thread);
8735	}
8736	return check_reshape(mddev);
8737}
8738
8739static int raid6_check_reshape(struct mddev *mddev)
8740{
8741	int new_chunk = mddev->new_chunk_sectors;
8742
8743	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8744		return -EINVAL;
8745	if (new_chunk > 0) {
8746		if (!is_power_of_2(new_chunk))
8747			return -EINVAL;
8748		if (new_chunk < (PAGE_SIZE >> 9))
8749			return -EINVAL;
8750		if (mddev->array_sectors & (new_chunk-1))
8751			/* not factor of array size */
8752			return -EINVAL;
8753	}
8754
8755	/* They look valid */
8756	return check_reshape(mddev);
8757}
8758
8759static void *raid5_takeover(struct mddev *mddev)
8760{
8761	/* raid5 can take over:
8762	 *  raid0 - if there is only one strip zone - make it a raid4 layout
8763	 *  raid1 - if there are two drives.  We need to know the chunk size
8764	 *  raid4 - trivial - just use a raid4 layout.
8765	 *  raid6 - Providing it is a *_6 layout
8766	 */
8767	if (mddev->level == 0)
8768		return raid45_takeover_raid0(mddev, 5);
8769	if (mddev->level == 1)
8770		return raid5_takeover_raid1(mddev);
8771	if (mddev->level == 4) {
8772		mddev->new_layout = ALGORITHM_PARITY_N;
8773		mddev->new_level = 5;
8774		return setup_conf(mddev);
8775	}
8776	if (mddev->level == 6)
8777		return raid5_takeover_raid6(mddev);
8778
8779	return ERR_PTR(-EINVAL);
8780}
8781
8782static void *raid4_takeover(struct mddev *mddev)
8783{
8784	/* raid4 can take over:
8785	 *  raid0 - if there is only one strip zone
8786	 *  raid5 - if layout is right
8787	 */
8788	if (mddev->level == 0)
8789		return raid45_takeover_raid0(mddev, 4);
8790	if (mddev->level == 5 &&
8791	    mddev->layout == ALGORITHM_PARITY_N) {
8792		mddev->new_layout = 0;
8793		mddev->new_level = 4;
8794		return setup_conf(mddev);
8795	}
8796	return ERR_PTR(-EINVAL);
8797}
8798
8799static struct md_personality raid5_personality;
8800
8801static void *raid6_takeover(struct mddev *mddev)
8802{
8803	/* Currently can only take over a raid5.  We map the
8804	 * personality to an equivalent raid6 personality
8805	 * with the Q block at the end.
8806	 */
8807	int new_layout;
8808
8809	if (mddev->pers != &raid5_personality)
8810		return ERR_PTR(-EINVAL);
8811	if (mddev->degraded > 1)
8812		return ERR_PTR(-EINVAL);
8813	if (mddev->raid_disks > 253)
8814		return ERR_PTR(-EINVAL);
8815	if (mddev->raid_disks < 3)
8816		return ERR_PTR(-EINVAL);
8817
8818	switch (mddev->layout) {
8819	case ALGORITHM_LEFT_ASYMMETRIC:
8820		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8821		break;
8822	case ALGORITHM_RIGHT_ASYMMETRIC:
8823		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8824		break;
8825	case ALGORITHM_LEFT_SYMMETRIC:
8826		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8827		break;
8828	case ALGORITHM_RIGHT_SYMMETRIC:
8829		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8830		break;
8831	case ALGORITHM_PARITY_0:
8832		new_layout = ALGORITHM_PARITY_0_6;
8833		break;
8834	case ALGORITHM_PARITY_N:
8835		new_layout = ALGORITHM_PARITY_N;
8836		break;
8837	default:
8838		return ERR_PTR(-EINVAL);
8839	}
8840	mddev->new_level = 6;
8841	mddev->new_layout = new_layout;
8842	mddev->delta_disks = 1;
8843	mddev->raid_disks += 1;
8844	return setup_conf(mddev);
8845}
8846
8847static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8848{
8849	struct r5conf *conf;
8850	int err;
8851
8852	err = mddev_suspend_and_lock(mddev);
8853	if (err)
8854		return err;
8855	conf = mddev->private;
8856	if (!conf) {
8857		mddev_unlock_and_resume(mddev);
8858		return -ENODEV;
8859	}
8860
8861	if (strncmp(buf, "ppl", 3) == 0) {
8862		/* ppl only works with RAID 5 */
8863		if (!raid5_has_ppl(conf) && conf->level == 5) {
8864			err = log_init(conf, NULL, true);
8865			if (!err) {
8866				err = resize_stripes(conf, conf->pool_size);
8867				if (err)
8868					log_exit(conf);
8869			}
8870		} else
8871			err = -EINVAL;
8872	} else if (strncmp(buf, "resync", 6) == 0) {
8873		if (raid5_has_ppl(conf)) {
8874			log_exit(conf);
8875			err = resize_stripes(conf, conf->pool_size);
8876		} else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8877			   r5l_log_disk_error(conf)) {
8878			bool journal_dev_exists = false;
8879			struct md_rdev *rdev;
8880
8881			rdev_for_each(rdev, mddev)
8882				if (test_bit(Journal, &rdev->flags)) {
8883					journal_dev_exists = true;
8884					break;
8885				}
8886
8887			if (!journal_dev_exists)
8888				clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8889			else  /* need remove journal device first */
8890				err = -EBUSY;
8891		} else
8892			err = -EINVAL;
8893	} else {
8894		err = -EINVAL;
8895	}
8896
8897	if (!err)
8898		md_update_sb(mddev, 1);
8899
8900	mddev_unlock_and_resume(mddev);
8901
8902	return err;
8903}
8904
8905static int raid5_start(struct mddev *mddev)
8906{
8907	struct r5conf *conf = mddev->private;
8908
8909	return r5l_start(conf->log);
8910}
8911
8912static struct md_personality raid6_personality =
8913{
8914	.name		= "raid6",
8915	.level		= 6,
8916	.owner		= THIS_MODULE,
8917	.make_request	= raid5_make_request,
8918	.run		= raid5_run,
8919	.start		= raid5_start,
8920	.free		= raid5_free,
8921	.status		= raid5_status,
8922	.error_handler	= raid5_error,
8923	.hot_add_disk	= raid5_add_disk,
8924	.hot_remove_disk= raid5_remove_disk,
8925	.spare_active	= raid5_spare_active,
8926	.sync_request	= raid5_sync_request,
8927	.resize		= raid5_resize,
8928	.size		= raid5_size,
8929	.check_reshape	= raid6_check_reshape,
8930	.start_reshape  = raid5_start_reshape,
8931	.finish_reshape = raid5_finish_reshape,
8932	.quiesce	= raid5_quiesce,
8933	.takeover	= raid6_takeover,
8934	.change_consistency_policy = raid5_change_consistency_policy,
8935};
8936static struct md_personality raid5_personality =
8937{
8938	.name		= "raid5",
8939	.level		= 5,
8940	.owner		= THIS_MODULE,
8941	.make_request	= raid5_make_request,
8942	.run		= raid5_run,
8943	.start		= raid5_start,
8944	.free		= raid5_free,
8945	.status		= raid5_status,
8946	.error_handler	= raid5_error,
8947	.hot_add_disk	= raid5_add_disk,
8948	.hot_remove_disk= raid5_remove_disk,
8949	.spare_active	= raid5_spare_active,
8950	.sync_request	= raid5_sync_request,
8951	.resize		= raid5_resize,
8952	.size		= raid5_size,
8953	.check_reshape	= raid5_check_reshape,
8954	.start_reshape  = raid5_start_reshape,
8955	.finish_reshape = raid5_finish_reshape,
8956	.quiesce	= raid5_quiesce,
8957	.takeover	= raid5_takeover,
8958	.change_consistency_policy = raid5_change_consistency_policy,
8959};
8960
8961static struct md_personality raid4_personality =
8962{
8963	.name		= "raid4",
8964	.level		= 4,
8965	.owner		= THIS_MODULE,
8966	.make_request	= raid5_make_request,
8967	.run		= raid5_run,
8968	.start		= raid5_start,
8969	.free		= raid5_free,
8970	.status		= raid5_status,
8971	.error_handler	= raid5_error,
8972	.hot_add_disk	= raid5_add_disk,
8973	.hot_remove_disk= raid5_remove_disk,
8974	.spare_active	= raid5_spare_active,
8975	.sync_request	= raid5_sync_request,
8976	.resize		= raid5_resize,
8977	.size		= raid5_size,
8978	.check_reshape	= raid5_check_reshape,
8979	.start_reshape  = raid5_start_reshape,
8980	.finish_reshape = raid5_finish_reshape,
8981	.quiesce	= raid5_quiesce,
8982	.takeover	= raid4_takeover,
8983	.change_consistency_policy = raid5_change_consistency_policy,
8984};
8985
8986static int __init raid5_init(void)
8987{
8988	int ret;
8989
8990	raid5_wq = alloc_workqueue("raid5wq",
8991		WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8992	if (!raid5_wq)
8993		return -ENOMEM;
8994
8995	ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8996				      "md/raid5:prepare",
8997				      raid456_cpu_up_prepare,
8998				      raid456_cpu_dead);
8999	if (ret) {
9000		destroy_workqueue(raid5_wq);
9001		return ret;
9002	}
9003	register_md_personality(&raid6_personality);
9004	register_md_personality(&raid5_personality);
9005	register_md_personality(&raid4_personality);
9006	return 0;
9007}
9008
9009static void raid5_exit(void)
9010{
9011	unregister_md_personality(&raid6_personality);
9012	unregister_md_personality(&raid5_personality);
9013	unregister_md_personality(&raid4_personality);
9014	cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
9015	destroy_workqueue(raid5_wq);
9016}
9017
9018module_init(raid5_init);
9019module_exit(raid5_exit);
9020MODULE_LICENSE("GPL");
9021MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
9022MODULE_ALIAS("md-personality-4"); /* RAID5 */
9023MODULE_ALIAS("md-raid5");
9024MODULE_ALIAS("md-raid4");
9025MODULE_ALIAS("md-level-5");
9026MODULE_ALIAS("md-level-4");
9027MODULE_ALIAS("md-personality-8"); /* RAID6 */
9028MODULE_ALIAS("md-raid6");
9029MODULE_ALIAS("md-level-6");
9030
9031/* This used to be two separate modules, they were: */
9032MODULE_ALIAS("raid5");
9033MODULE_ALIAS("raid6");