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