Linux Audio

Check our new training course

Yocto distribution development and maintenance

Need a Yocto distribution for your embedded project?
Loading...
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2009-2011 Red Hat, Inc.
   4 *
   5 * Author: Mikulas Patocka <mpatocka@redhat.com>
   6 *
   7 * This file is released under the GPL.
   8 */
   9
  10#include <linux/dm-bufio.h>
  11
  12#include <linux/device-mapper.h>
  13#include <linux/dm-io.h>
  14#include <linux/slab.h>
  15#include <linux/sched/mm.h>
  16#include <linux/jiffies.h>
  17#include <linux/vmalloc.h>
  18#include <linux/shrinker.h>
  19#include <linux/module.h>
  20#include <linux/rbtree.h>
  21#include <linux/stacktrace.h>
  22#include <linux/jump_label.h>
  23
  24#include "dm.h"
  25
  26#define DM_MSG_PREFIX "bufio"
  27
  28/*
  29 * Memory management policy:
  30 *	Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
  31 *	or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
  32 *	Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
  33 *	Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
  34 *	dirty buffers.
  35 */
  36#define DM_BUFIO_MIN_BUFFERS		8
  37
  38#define DM_BUFIO_MEMORY_PERCENT		2
  39#define DM_BUFIO_VMALLOC_PERCENT	25
  40#define DM_BUFIO_WRITEBACK_RATIO	3
  41#define DM_BUFIO_LOW_WATERMARK_RATIO	16
  42
  43/*
  44 * Check buffer ages in this interval (seconds)
  45 */
  46#define DM_BUFIO_WORK_TIMER_SECS	30
  47
  48/*
  49 * Free buffers when they are older than this (seconds)
  50 */
  51#define DM_BUFIO_DEFAULT_AGE_SECS	300
  52
  53/*
  54 * The nr of bytes of cached data to keep around.
  55 */
  56#define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
  57
  58/*
  59 * Align buffer writes to this boundary.
  60 * Tests show that SSDs have the highest IOPS when using 4k writes.
  61 */
  62#define DM_BUFIO_WRITE_ALIGN		4096
  63
  64/*
  65 * dm_buffer->list_mode
  66 */
  67#define LIST_CLEAN	0
  68#define LIST_DIRTY	1
  69#define LIST_SIZE	2
  70
  71/*--------------------------------------------------------------*/
  72
  73/*
  74 * Rather than use an LRU list, we use a clock algorithm where entries
  75 * are held in a circular list.  When an entry is 'hit' a reference bit
  76 * is set.  The least recently used entry is approximated by running a
  77 * cursor around the list selecting unreferenced entries. Referenced
  78 * entries have their reference bit cleared as the cursor passes them.
  79 */
  80struct lru_entry {
  81	struct list_head list;
  82	atomic_t referenced;
  83};
  84
  85struct lru_iter {
  86	struct lru *lru;
  87	struct list_head list;
  88	struct lru_entry *stop;
  89	struct lru_entry *e;
  90};
  91
  92struct lru {
  93	struct list_head *cursor;
  94	unsigned long count;
  95
  96	struct list_head iterators;
  97};
  98
  99/*--------------*/
 100
 101static void lru_init(struct lru *lru)
 102{
 103	lru->cursor = NULL;
 104	lru->count = 0;
 105	INIT_LIST_HEAD(&lru->iterators);
 106}
 107
 108static void lru_destroy(struct lru *lru)
 109{
 110	WARN_ON_ONCE(lru->cursor);
 111	WARN_ON_ONCE(!list_empty(&lru->iterators));
 112}
 113
 114/*
 115 * Insert a new entry into the lru.
 116 */
 117static void lru_insert(struct lru *lru, struct lru_entry *le)
 118{
 119	/*
 120	 * Don't be tempted to set to 1, makes the lru aspect
 121	 * perform poorly.
 122	 */
 123	atomic_set(&le->referenced, 0);
 124
 125	if (lru->cursor) {
 126		list_add_tail(&le->list, lru->cursor);
 127	} else {
 128		INIT_LIST_HEAD(&le->list);
 129		lru->cursor = &le->list;
 130	}
 131	lru->count++;
 132}
 133
 134/*--------------*/
 135
 136/*
 137 * Convert a list_head pointer to an lru_entry pointer.
 138 */
 139static inline struct lru_entry *to_le(struct list_head *l)
 140{
 141	return container_of(l, struct lru_entry, list);
 142}
 143
 144/*
 145 * Initialize an lru_iter and add it to the list of cursors in the lru.
 146 */
 147static void lru_iter_begin(struct lru *lru, struct lru_iter *it)
 148{
 149	it->lru = lru;
 150	it->stop = lru->cursor ? to_le(lru->cursor->prev) : NULL;
 151	it->e = lru->cursor ? to_le(lru->cursor) : NULL;
 152	list_add(&it->list, &lru->iterators);
 153}
 154
 155/*
 156 * Remove an lru_iter from the list of cursors in the lru.
 157 */
 158static inline void lru_iter_end(struct lru_iter *it)
 159{
 160	list_del(&it->list);
 161}
 162
 163/* Predicate function type to be used with lru_iter_next */
 164typedef bool (*iter_predicate)(struct lru_entry *le, void *context);
 165
 166/*
 167 * Advance the cursor to the next entry that passes the
 168 * predicate, and return that entry.  Returns NULL if the
 169 * iteration is complete.
 170 */
 171static struct lru_entry *lru_iter_next(struct lru_iter *it,
 172				       iter_predicate pred, void *context)
 173{
 174	struct lru_entry *e;
 175
 176	while (it->e) {
 177		e = it->e;
 178
 179		/* advance the cursor */
 180		if (it->e == it->stop)
 181			it->e = NULL;
 182		else
 183			it->e = to_le(it->e->list.next);
 184
 185		if (pred(e, context))
 186			return e;
 187	}
 188
 189	return NULL;
 190}
 191
 192/*
 193 * Invalidate a specific lru_entry and update all cursors in
 194 * the lru accordingly.
 195 */
 196static void lru_iter_invalidate(struct lru *lru, struct lru_entry *e)
 197{
 198	struct lru_iter *it;
 199
 200	list_for_each_entry(it, &lru->iterators, list) {
 201		/* Move c->e forwards if necc. */
 202		if (it->e == e) {
 203			it->e = to_le(it->e->list.next);
 204			if (it->e == e)
 205				it->e = NULL;
 206		}
 207
 208		/* Move it->stop backwards if necc. */
 209		if (it->stop == e) {
 210			it->stop = to_le(it->stop->list.prev);
 211			if (it->stop == e)
 212				it->stop = NULL;
 213		}
 214	}
 215}
 216
 217/*--------------*/
 218
 219/*
 220 * Remove a specific entry from the lru.
 221 */
 222static void lru_remove(struct lru *lru, struct lru_entry *le)
 223{
 224	lru_iter_invalidate(lru, le);
 225	if (lru->count == 1) {
 226		lru->cursor = NULL;
 227	} else {
 228		if (lru->cursor == &le->list)
 229			lru->cursor = lru->cursor->next;
 230		list_del(&le->list);
 231	}
 232	lru->count--;
 233}
 234
 235/*
 236 * Mark as referenced.
 237 */
 238static inline void lru_reference(struct lru_entry *le)
 239{
 240	atomic_set(&le->referenced, 1);
 241}
 242
 243/*--------------*/
 244
 245/*
 246 * Remove the least recently used entry (approx), that passes the predicate.
 247 * Returns NULL on failure.
 248 */
 249enum evict_result {
 250	ER_EVICT,
 251	ER_DONT_EVICT,
 252	ER_STOP, /* stop looking for something to evict */
 253};
 254
 255typedef enum evict_result (*le_predicate)(struct lru_entry *le, void *context);
 256
 257static struct lru_entry *lru_evict(struct lru *lru, le_predicate pred, void *context, bool no_sleep)
 258{
 259	unsigned long tested = 0;
 260	struct list_head *h = lru->cursor;
 261	struct lru_entry *le;
 262
 263	if (!h)
 264		return NULL;
 265	/*
 266	 * In the worst case we have to loop around twice. Once to clear
 267	 * the reference flags, and then again to discover the predicate
 268	 * fails for all entries.
 269	 */
 270	while (tested < lru->count) {
 271		le = container_of(h, struct lru_entry, list);
 272
 273		if (atomic_read(&le->referenced)) {
 274			atomic_set(&le->referenced, 0);
 275		} else {
 276			tested++;
 277			switch (pred(le, context)) {
 278			case ER_EVICT:
 279				/*
 280				 * Adjust the cursor, so we start the next
 281				 * search from here.
 282				 */
 283				lru->cursor = le->list.next;
 284				lru_remove(lru, le);
 285				return le;
 286
 287			case ER_DONT_EVICT:
 288				break;
 289
 290			case ER_STOP:
 291				lru->cursor = le->list.next;
 292				return NULL;
 293			}
 294		}
 295
 296		h = h->next;
 297
 298		if (!no_sleep)
 299			cond_resched();
 300	}
 301
 302	return NULL;
 303}
 304
 305/*--------------------------------------------------------------*/
 306
 307/*
 308 * Buffer state bits.
 309 */
 310#define B_READING	0
 311#define B_WRITING	1
 312#define B_DIRTY		2
 313
 314/*
 315 * Describes how the block was allocated:
 316 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
 317 * See the comment at alloc_buffer_data.
 318 */
 319enum data_mode {
 320	DATA_MODE_SLAB = 0,
 321	DATA_MODE_GET_FREE_PAGES = 1,
 322	DATA_MODE_VMALLOC = 2,
 323	DATA_MODE_LIMIT = 3
 324};
 325
 326struct dm_buffer {
 327	/* protected by the locks in dm_buffer_cache */
 328	struct rb_node node;
 329
 330	/* immutable, so don't need protecting */
 331	sector_t block;
 332	void *data;
 333	unsigned char data_mode;		/* DATA_MODE_* */
 334
 335	/*
 336	 * These two fields are used in isolation, so do not need
 337	 * a surrounding lock.
 338	 */
 339	atomic_t hold_count;
 340	unsigned long last_accessed;
 341
 342	/*
 343	 * Everything else is protected by the mutex in
 344	 * dm_bufio_client
 345	 */
 346	unsigned long state;
 347	struct lru_entry lru;
 348	unsigned char list_mode;		/* LIST_* */
 349	blk_status_t read_error;
 350	blk_status_t write_error;
 351	unsigned int dirty_start;
 352	unsigned int dirty_end;
 353	unsigned int write_start;
 354	unsigned int write_end;
 355	struct list_head write_list;
 356	struct dm_bufio_client *c;
 357	void (*end_io)(struct dm_buffer *b, blk_status_t bs);
 358#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 359#define MAX_STACK 10
 360	unsigned int stack_len;
 361	unsigned long stack_entries[MAX_STACK];
 362#endif
 363};
 364
 365/*--------------------------------------------------------------*/
 366
 367/*
 368 * The buffer cache manages buffers, particularly:
 369 *  - inc/dec of holder count
 370 *  - setting the last_accessed field
 371 *  - maintains clean/dirty state along with lru
 372 *  - selecting buffers that match predicates
 373 *
 374 * It does *not* handle:
 375 *  - allocation/freeing of buffers.
 376 *  - IO
 377 *  - Eviction or cache sizing.
 378 *
 379 * cache_get() and cache_put() are threadsafe, you do not need to
 380 * protect these calls with a surrounding mutex.  All the other
 381 * methods are not threadsafe; they do use locking primitives, but
 382 * only enough to ensure get/put are threadsafe.
 383 */
 384
 385struct buffer_tree {
 386	union {
 387		struct rw_semaphore lock;
 388		rwlock_t spinlock;
 389	} u;
 390	struct rb_root root;
 391} ____cacheline_aligned_in_smp;
 392
 393struct dm_buffer_cache {
 394	struct lru lru[LIST_SIZE];
 395	/*
 396	 * We spread entries across multiple trees to reduce contention
 397	 * on the locks.
 398	 */
 399	unsigned int num_locks;
 400	bool no_sleep;
 401	struct buffer_tree trees[];
 402};
 403
 404static DEFINE_STATIC_KEY_FALSE(no_sleep_enabled);
 405
 406static inline unsigned int cache_index(sector_t block, unsigned int num_locks)
 407{
 408	return dm_hash_locks_index(block, num_locks);
 409}
 410
 411static inline void cache_read_lock(struct dm_buffer_cache *bc, sector_t block)
 412{
 413	if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
 414		read_lock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
 415	else
 416		down_read(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
 417}
 418
 419static inline void cache_read_unlock(struct dm_buffer_cache *bc, sector_t block)
 420{
 421	if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
 422		read_unlock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
 423	else
 424		up_read(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
 425}
 426
 427static inline void cache_write_lock(struct dm_buffer_cache *bc, sector_t block)
 428{
 429	if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
 430		write_lock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
 431	else
 432		down_write(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
 433}
 434
 435static inline void cache_write_unlock(struct dm_buffer_cache *bc, sector_t block)
 436{
 437	if (static_branch_unlikely(&no_sleep_enabled) && bc->no_sleep)
 438		write_unlock_bh(&bc->trees[cache_index(block, bc->num_locks)].u.spinlock);
 439	else
 440		up_write(&bc->trees[cache_index(block, bc->num_locks)].u.lock);
 441}
 442
 443/*
 444 * Sometimes we want to repeatedly get and drop locks as part of an iteration.
 445 * This struct helps avoid redundant drop and gets of the same lock.
 446 */
 447struct lock_history {
 448	struct dm_buffer_cache *cache;
 449	bool write;
 450	unsigned int previous;
 451	unsigned int no_previous;
 452};
 453
 454static void lh_init(struct lock_history *lh, struct dm_buffer_cache *cache, bool write)
 455{
 456	lh->cache = cache;
 457	lh->write = write;
 458	lh->no_previous = cache->num_locks;
 459	lh->previous = lh->no_previous;
 460}
 461
 462static void __lh_lock(struct lock_history *lh, unsigned int index)
 463{
 464	if (lh->write) {
 465		if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
 466			write_lock_bh(&lh->cache->trees[index].u.spinlock);
 467		else
 468			down_write(&lh->cache->trees[index].u.lock);
 469	} else {
 470		if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
 471			read_lock_bh(&lh->cache->trees[index].u.spinlock);
 472		else
 473			down_read(&lh->cache->trees[index].u.lock);
 474	}
 475}
 476
 477static void __lh_unlock(struct lock_history *lh, unsigned int index)
 478{
 479	if (lh->write) {
 480		if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
 481			write_unlock_bh(&lh->cache->trees[index].u.spinlock);
 482		else
 483			up_write(&lh->cache->trees[index].u.lock);
 484	} else {
 485		if (static_branch_unlikely(&no_sleep_enabled) && lh->cache->no_sleep)
 486			read_unlock_bh(&lh->cache->trees[index].u.spinlock);
 487		else
 488			up_read(&lh->cache->trees[index].u.lock);
 489	}
 490}
 491
 492/*
 493 * Make sure you call this since it will unlock the final lock.
 494 */
 495static void lh_exit(struct lock_history *lh)
 496{
 497	if (lh->previous != lh->no_previous) {
 498		__lh_unlock(lh, lh->previous);
 499		lh->previous = lh->no_previous;
 500	}
 501}
 502
 503/*
 504 * Named 'next' because there is no corresponding
 505 * 'up/unlock' call since it's done automatically.
 506 */
 507static void lh_next(struct lock_history *lh, sector_t b)
 508{
 509	unsigned int index = cache_index(b, lh->no_previous); /* no_previous is num_locks */
 510
 511	if (lh->previous != lh->no_previous) {
 512		if (lh->previous != index) {
 513			__lh_unlock(lh, lh->previous);
 514			__lh_lock(lh, index);
 515			lh->previous = index;
 516		}
 517	} else {
 518		__lh_lock(lh, index);
 519		lh->previous = index;
 520	}
 521}
 522
 523static inline struct dm_buffer *le_to_buffer(struct lru_entry *le)
 524{
 525	return container_of(le, struct dm_buffer, lru);
 526}
 527
 528static struct dm_buffer *list_to_buffer(struct list_head *l)
 529{
 530	struct lru_entry *le = list_entry(l, struct lru_entry, list);
 531
 532	if (!le)
 533		return NULL;
 534
 535	return le_to_buffer(le);
 536}
 537
 538static void cache_init(struct dm_buffer_cache *bc, unsigned int num_locks, bool no_sleep)
 539{
 540	unsigned int i;
 541
 542	bc->num_locks = num_locks;
 543	bc->no_sleep = no_sleep;
 544
 545	for (i = 0; i < bc->num_locks; i++) {
 546		if (no_sleep)
 547			rwlock_init(&bc->trees[i].u.spinlock);
 548		else
 549			init_rwsem(&bc->trees[i].u.lock);
 550		bc->trees[i].root = RB_ROOT;
 551	}
 552
 553	lru_init(&bc->lru[LIST_CLEAN]);
 554	lru_init(&bc->lru[LIST_DIRTY]);
 555}
 556
 557static void cache_destroy(struct dm_buffer_cache *bc)
 558{
 559	unsigned int i;
 560
 561	for (i = 0; i < bc->num_locks; i++)
 562		WARN_ON_ONCE(!RB_EMPTY_ROOT(&bc->trees[i].root));
 563
 564	lru_destroy(&bc->lru[LIST_CLEAN]);
 565	lru_destroy(&bc->lru[LIST_DIRTY]);
 566}
 567
 568/*--------------*/
 569
 570/*
 571 * not threadsafe, or racey depending how you look at it
 572 */
 573static inline unsigned long cache_count(struct dm_buffer_cache *bc, int list_mode)
 574{
 575	return bc->lru[list_mode].count;
 576}
 577
 578static inline unsigned long cache_total(struct dm_buffer_cache *bc)
 579{
 580	return cache_count(bc, LIST_CLEAN) + cache_count(bc, LIST_DIRTY);
 581}
 582
 583/*--------------*/
 584
 585/*
 586 * Gets a specific buffer, indexed by block.
 587 * If the buffer is found then its holder count will be incremented and
 588 * lru_reference will be called.
 589 *
 590 * threadsafe
 591 */
 592static struct dm_buffer *__cache_get(const struct rb_root *root, sector_t block)
 593{
 594	struct rb_node *n = root->rb_node;
 595	struct dm_buffer *b;
 596
 597	while (n) {
 598		b = container_of(n, struct dm_buffer, node);
 599
 600		if (b->block == block)
 601			return b;
 602
 603		n = block < b->block ? n->rb_left : n->rb_right;
 604	}
 605
 606	return NULL;
 607}
 608
 609static void __cache_inc_buffer(struct dm_buffer *b)
 610{
 611	atomic_inc(&b->hold_count);
 612	WRITE_ONCE(b->last_accessed, jiffies);
 613}
 614
 615static struct dm_buffer *cache_get(struct dm_buffer_cache *bc, sector_t block)
 616{
 617	struct dm_buffer *b;
 618
 619	cache_read_lock(bc, block);
 620	b = __cache_get(&bc->trees[cache_index(block, bc->num_locks)].root, block);
 621	if (b) {
 622		lru_reference(&b->lru);
 623		__cache_inc_buffer(b);
 624	}
 625	cache_read_unlock(bc, block);
 626
 627	return b;
 628}
 629
 630/*--------------*/
 631
 632/*
 633 * Returns true if the hold count hits zero.
 634 * threadsafe
 635 */
 636static bool cache_put(struct dm_buffer_cache *bc, struct dm_buffer *b)
 637{
 638	bool r;
 639
 640	cache_read_lock(bc, b->block);
 641	BUG_ON(!atomic_read(&b->hold_count));
 642	r = atomic_dec_and_test(&b->hold_count);
 643	cache_read_unlock(bc, b->block);
 644
 645	return r;
 646}
 647
 648/*--------------*/
 649
 650typedef enum evict_result (*b_predicate)(struct dm_buffer *, void *);
 651
 652/*
 653 * Evicts a buffer based on a predicate.  The oldest buffer that
 654 * matches the predicate will be selected.  In addition to the
 655 * predicate the hold_count of the selected buffer will be zero.
 656 */
 657struct evict_wrapper {
 658	struct lock_history *lh;
 659	b_predicate pred;
 660	void *context;
 661};
 662
 663/*
 664 * Wraps the buffer predicate turning it into an lru predicate.  Adds
 665 * extra test for hold_count.
 666 */
 667static enum evict_result __evict_pred(struct lru_entry *le, void *context)
 668{
 669	struct evict_wrapper *w = context;
 670	struct dm_buffer *b = le_to_buffer(le);
 671
 672	lh_next(w->lh, b->block);
 673
 674	if (atomic_read(&b->hold_count))
 675		return ER_DONT_EVICT;
 676
 677	return w->pred(b, w->context);
 678}
 679
 680static struct dm_buffer *__cache_evict(struct dm_buffer_cache *bc, int list_mode,
 681				       b_predicate pred, void *context,
 682				       struct lock_history *lh)
 683{
 684	struct evict_wrapper w = {.lh = lh, .pred = pred, .context = context};
 685	struct lru_entry *le;
 686	struct dm_buffer *b;
 687
 688	le = lru_evict(&bc->lru[list_mode], __evict_pred, &w, bc->no_sleep);
 689	if (!le)
 690		return NULL;
 691
 692	b = le_to_buffer(le);
 693	/* __evict_pred will have locked the appropriate tree. */
 694	rb_erase(&b->node, &bc->trees[cache_index(b->block, bc->num_locks)].root);
 695
 696	return b;
 697}
 698
 699static struct dm_buffer *cache_evict(struct dm_buffer_cache *bc, int list_mode,
 700				     b_predicate pred, void *context)
 701{
 702	struct dm_buffer *b;
 703	struct lock_history lh;
 704
 705	lh_init(&lh, bc, true);
 706	b = __cache_evict(bc, list_mode, pred, context, &lh);
 707	lh_exit(&lh);
 708
 709	return b;
 710}
 711
 712/*--------------*/
 713
 714/*
 715 * Mark a buffer as clean or dirty. Not threadsafe.
 716 */
 717static void cache_mark(struct dm_buffer_cache *bc, struct dm_buffer *b, int list_mode)
 718{
 719	cache_write_lock(bc, b->block);
 720	if (list_mode != b->list_mode) {
 721		lru_remove(&bc->lru[b->list_mode], &b->lru);
 722		b->list_mode = list_mode;
 723		lru_insert(&bc->lru[b->list_mode], &b->lru);
 724	}
 725	cache_write_unlock(bc, b->block);
 726}
 727
 728/*--------------*/
 729
 730/*
 731 * Runs through the lru associated with 'old_mode', if the predicate matches then
 732 * it moves them to 'new_mode'.  Not threadsafe.
 733 */
 734static void __cache_mark_many(struct dm_buffer_cache *bc, int old_mode, int new_mode,
 735			      b_predicate pred, void *context, struct lock_history *lh)
 736{
 737	struct lru_entry *le;
 738	struct dm_buffer *b;
 739	struct evict_wrapper w = {.lh = lh, .pred = pred, .context = context};
 740
 741	while (true) {
 742		le = lru_evict(&bc->lru[old_mode], __evict_pred, &w, bc->no_sleep);
 743		if (!le)
 744			break;
 745
 746		b = le_to_buffer(le);
 747		b->list_mode = new_mode;
 748		lru_insert(&bc->lru[b->list_mode], &b->lru);
 749	}
 750}
 751
 752static void cache_mark_many(struct dm_buffer_cache *bc, int old_mode, int new_mode,
 753			    b_predicate pred, void *context)
 754{
 755	struct lock_history lh;
 756
 757	lh_init(&lh, bc, true);
 758	__cache_mark_many(bc, old_mode, new_mode, pred, context, &lh);
 759	lh_exit(&lh);
 760}
 761
 762/*--------------*/
 763
 764/*
 765 * Iterates through all clean or dirty entries calling a function for each
 766 * entry.  The callback may terminate the iteration early.  Not threadsafe.
 767 */
 768
 769/*
 770 * Iterator functions should return one of these actions to indicate
 771 * how the iteration should proceed.
 772 */
 773enum it_action {
 774	IT_NEXT,
 775	IT_COMPLETE,
 776};
 777
 778typedef enum it_action (*iter_fn)(struct dm_buffer *b, void *context);
 779
 780static void __cache_iterate(struct dm_buffer_cache *bc, int list_mode,
 781			    iter_fn fn, void *context, struct lock_history *lh)
 782{
 783	struct lru *lru = &bc->lru[list_mode];
 784	struct lru_entry *le, *first;
 785
 786	if (!lru->cursor)
 787		return;
 788
 789	first = le = to_le(lru->cursor);
 790	do {
 791		struct dm_buffer *b = le_to_buffer(le);
 792
 793		lh_next(lh, b->block);
 794
 795		switch (fn(b, context)) {
 796		case IT_NEXT:
 797			break;
 798
 799		case IT_COMPLETE:
 800			return;
 801		}
 802		cond_resched();
 803
 804		le = to_le(le->list.next);
 805	} while (le != first);
 806}
 807
 808static void cache_iterate(struct dm_buffer_cache *bc, int list_mode,
 809			  iter_fn fn, void *context)
 810{
 811	struct lock_history lh;
 812
 813	lh_init(&lh, bc, false);
 814	__cache_iterate(bc, list_mode, fn, context, &lh);
 815	lh_exit(&lh);
 816}
 817
 818/*--------------*/
 819
 820/*
 821 * Passes ownership of the buffer to the cache. Returns false if the
 822 * buffer was already present (in which case ownership does not pass).
 823 * eg, a race with another thread.
 824 *
 825 * Holder count should be 1 on insertion.
 826 *
 827 * Not threadsafe.
 828 */
 829static bool __cache_insert(struct rb_root *root, struct dm_buffer *b)
 830{
 831	struct rb_node **new = &root->rb_node, *parent = NULL;
 832	struct dm_buffer *found;
 833
 834	while (*new) {
 835		found = container_of(*new, struct dm_buffer, node);
 836
 837		if (found->block == b->block)
 838			return false;
 839
 840		parent = *new;
 841		new = b->block < found->block ?
 842			&found->node.rb_left : &found->node.rb_right;
 843	}
 844
 845	rb_link_node(&b->node, parent, new);
 846	rb_insert_color(&b->node, root);
 847
 848	return true;
 849}
 850
 851static bool cache_insert(struct dm_buffer_cache *bc, struct dm_buffer *b)
 852{
 853	bool r;
 854
 855	if (WARN_ON_ONCE(b->list_mode >= LIST_SIZE))
 856		return false;
 857
 858	cache_write_lock(bc, b->block);
 859	BUG_ON(atomic_read(&b->hold_count) != 1);
 860	r = __cache_insert(&bc->trees[cache_index(b->block, bc->num_locks)].root, b);
 861	if (r)
 862		lru_insert(&bc->lru[b->list_mode], &b->lru);
 863	cache_write_unlock(bc, b->block);
 864
 865	return r;
 866}
 867
 868/*--------------*/
 869
 870/*
 871 * Removes buffer from cache, ownership of the buffer passes back to the caller.
 872 * Fails if the hold_count is not one (ie. the caller holds the only reference).
 873 *
 874 * Not threadsafe.
 875 */
 876static bool cache_remove(struct dm_buffer_cache *bc, struct dm_buffer *b)
 877{
 878	bool r;
 879
 880	cache_write_lock(bc, b->block);
 881
 882	if (atomic_read(&b->hold_count) != 1) {
 883		r = false;
 884	} else {
 885		r = true;
 886		rb_erase(&b->node, &bc->trees[cache_index(b->block, bc->num_locks)].root);
 887		lru_remove(&bc->lru[b->list_mode], &b->lru);
 888	}
 889
 890	cache_write_unlock(bc, b->block);
 891
 892	return r;
 893}
 894
 895/*--------------*/
 896
 897typedef void (*b_release)(struct dm_buffer *);
 898
 899static struct dm_buffer *__find_next(struct rb_root *root, sector_t block)
 900{
 901	struct rb_node *n = root->rb_node;
 902	struct dm_buffer *b;
 903	struct dm_buffer *best = NULL;
 904
 905	while (n) {
 906		b = container_of(n, struct dm_buffer, node);
 907
 908		if (b->block == block)
 909			return b;
 910
 911		if (block <= b->block) {
 912			n = n->rb_left;
 913			best = b;
 914		} else {
 915			n = n->rb_right;
 916		}
 917	}
 918
 919	return best;
 920}
 921
 922static void __remove_range(struct dm_buffer_cache *bc,
 923			   struct rb_root *root,
 924			   sector_t begin, sector_t end,
 925			   b_predicate pred, b_release release)
 926{
 927	struct dm_buffer *b;
 928
 929	while (true) {
 930		cond_resched();
 931
 932		b = __find_next(root, begin);
 933		if (!b || (b->block >= end))
 934			break;
 935
 936		begin = b->block + 1;
 937
 938		if (atomic_read(&b->hold_count))
 939			continue;
 940
 941		if (pred(b, NULL) == ER_EVICT) {
 942			rb_erase(&b->node, root);
 943			lru_remove(&bc->lru[b->list_mode], &b->lru);
 944			release(b);
 945		}
 946	}
 947}
 948
 949static void cache_remove_range(struct dm_buffer_cache *bc,
 950			       sector_t begin, sector_t end,
 951			       b_predicate pred, b_release release)
 952{
 953	unsigned int i;
 954
 955	BUG_ON(bc->no_sleep);
 956	for (i = 0; i < bc->num_locks; i++) {
 957		down_write(&bc->trees[i].u.lock);
 958		__remove_range(bc, &bc->trees[i].root, begin, end, pred, release);
 959		up_write(&bc->trees[i].u.lock);
 960	}
 961}
 962
 963/*----------------------------------------------------------------*/
 964
 965/*
 966 * Linking of buffers:
 967 *	All buffers are linked to buffer_cache with their node field.
 968 *
 969 *	Clean buffers that are not being written (B_WRITING not set)
 970 *	are linked to lru[LIST_CLEAN] with their lru_list field.
 971 *
 972 *	Dirty and clean buffers that are being written are linked to
 973 *	lru[LIST_DIRTY] with their lru_list field. When the write
 974 *	finishes, the buffer cannot be relinked immediately (because we
 975 *	are in an interrupt context and relinking requires process
 976 *	context), so some clean-not-writing buffers can be held on
 977 *	dirty_lru too.  They are later added to lru in the process
 978 *	context.
 979 */
 980struct dm_bufio_client {
 981	struct block_device *bdev;
 982	unsigned int block_size;
 983	s8 sectors_per_block_bits;
 984
 985	bool no_sleep;
 986	struct mutex lock;
 987	spinlock_t spinlock;
 988
 989	int async_write_error;
 990
 991	void (*alloc_callback)(struct dm_buffer *buf);
 992	void (*write_callback)(struct dm_buffer *buf);
 993	struct kmem_cache *slab_buffer;
 994	struct kmem_cache *slab_cache;
 995	struct dm_io_client *dm_io;
 996
 997	struct list_head reserved_buffers;
 998	unsigned int need_reserved_buffers;
 999
1000	unsigned int minimum_buffers;
1001
1002	sector_t start;
1003
1004	struct shrinker *shrinker;
1005	struct work_struct shrink_work;
1006	atomic_long_t need_shrink;
1007
1008	wait_queue_head_t free_buffer_wait;
1009
1010	struct list_head client_list;
1011
1012	/*
1013	 * Used by global_cleanup to sort the clients list.
1014	 */
1015	unsigned long oldest_buffer;
1016
1017	struct dm_buffer_cache cache; /* must be last member */
1018};
1019
1020/*----------------------------------------------------------------*/
1021
1022#define dm_bufio_in_request()	(!!current->bio_list)
1023
1024static void dm_bufio_lock(struct dm_bufio_client *c)
1025{
1026	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1027		spin_lock_bh(&c->spinlock);
1028	else
1029		mutex_lock_nested(&c->lock, dm_bufio_in_request());
1030}
1031
1032static void dm_bufio_unlock(struct dm_bufio_client *c)
1033{
1034	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1035		spin_unlock_bh(&c->spinlock);
1036	else
1037		mutex_unlock(&c->lock);
1038}
1039
1040/*----------------------------------------------------------------*/
1041
1042/*
1043 * Default cache size: available memory divided by the ratio.
1044 */
1045static unsigned long dm_bufio_default_cache_size;
1046
1047/*
1048 * Total cache size set by the user.
1049 */
1050static unsigned long dm_bufio_cache_size;
1051
1052/*
1053 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
1054 * at any time.  If it disagrees, the user has changed cache size.
1055 */
1056static unsigned long dm_bufio_cache_size_latch;
1057
1058static DEFINE_SPINLOCK(global_spinlock);
1059
1060/*
1061 * Buffers are freed after this timeout
1062 */
1063static unsigned int dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
1064static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
1065
1066static unsigned long dm_bufio_peak_allocated;
1067static unsigned long dm_bufio_allocated_kmem_cache;
1068static unsigned long dm_bufio_allocated_get_free_pages;
1069static unsigned long dm_bufio_allocated_vmalloc;
1070static unsigned long dm_bufio_current_allocated;
1071
1072/*----------------------------------------------------------------*/
1073
1074/*
1075 * The current number of clients.
1076 */
1077static int dm_bufio_client_count;
1078
1079/*
1080 * The list of all clients.
1081 */
1082static LIST_HEAD(dm_bufio_all_clients);
1083
1084/*
1085 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
1086 */
1087static DEFINE_MUTEX(dm_bufio_clients_lock);
1088
1089static struct workqueue_struct *dm_bufio_wq;
1090static struct delayed_work dm_bufio_cleanup_old_work;
1091static struct work_struct dm_bufio_replacement_work;
1092
1093
1094#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1095static void buffer_record_stack(struct dm_buffer *b)
1096{
1097	b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
1098}
1099#endif
1100
1101/*----------------------------------------------------------------*/
1102
1103static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
1104{
1105	unsigned char data_mode;
1106	long diff;
1107
1108	static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
1109		&dm_bufio_allocated_kmem_cache,
1110		&dm_bufio_allocated_get_free_pages,
1111		&dm_bufio_allocated_vmalloc,
1112	};
1113
1114	data_mode = b->data_mode;
1115	diff = (long)b->c->block_size;
1116	if (unlink)
1117		diff = -diff;
1118
1119	spin_lock(&global_spinlock);
1120
1121	*class_ptr[data_mode] += diff;
1122
1123	dm_bufio_current_allocated += diff;
1124
1125	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
1126		dm_bufio_peak_allocated = dm_bufio_current_allocated;
1127
1128	if (!unlink) {
1129		if (dm_bufio_current_allocated > dm_bufio_cache_size)
1130			queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
1131	}
1132
1133	spin_unlock(&global_spinlock);
1134}
1135
1136/*
1137 * Change the number of clients and recalculate per-client limit.
1138 */
1139static void __cache_size_refresh(void)
1140{
1141	if (WARN_ON(!mutex_is_locked(&dm_bufio_clients_lock)))
1142		return;
1143	if (WARN_ON(dm_bufio_client_count < 0))
1144		return;
1145
1146	dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
1147
1148	/*
1149	 * Use default if set to 0 and report the actual cache size used.
1150	 */
1151	if (!dm_bufio_cache_size_latch) {
1152		(void)cmpxchg(&dm_bufio_cache_size, 0,
1153			      dm_bufio_default_cache_size);
1154		dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
1155	}
1156}
1157
1158/*
1159 * Allocating buffer data.
1160 *
1161 * Small buffers are allocated with kmem_cache, to use space optimally.
1162 *
1163 * For large buffers, we choose between get_free_pages and vmalloc.
1164 * Each has advantages and disadvantages.
1165 *
1166 * __get_free_pages can randomly fail if the memory is fragmented.
1167 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
1168 * as low as 128M) so using it for caching is not appropriate.
1169 *
1170 * If the allocation may fail we use __get_free_pages. Memory fragmentation
1171 * won't have a fatal effect here, but it just causes flushes of some other
1172 * buffers and more I/O will be performed. Don't use __get_free_pages if it
1173 * always fails (i.e. order > MAX_PAGE_ORDER).
1174 *
1175 * If the allocation shouldn't fail we use __vmalloc. This is only for the
1176 * initial reserve allocation, so there's no risk of wasting all vmalloc
1177 * space.
1178 */
1179static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
1180			       unsigned char *data_mode)
1181{
1182	if (unlikely(c->slab_cache != NULL)) {
1183		*data_mode = DATA_MODE_SLAB;
1184		return kmem_cache_alloc(c->slab_cache, gfp_mask);
1185	}
1186
1187	if (c->block_size <= KMALLOC_MAX_SIZE &&
1188	    gfp_mask & __GFP_NORETRY) {
1189		*data_mode = DATA_MODE_GET_FREE_PAGES;
1190		return (void *)__get_free_pages(gfp_mask,
1191						c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
1192	}
1193
1194	*data_mode = DATA_MODE_VMALLOC;
1195
1196	return __vmalloc(c->block_size, gfp_mask);
1197}
1198
1199/*
1200 * Free buffer's data.
1201 */
1202static void free_buffer_data(struct dm_bufio_client *c,
1203			     void *data, unsigned char data_mode)
1204{
1205	switch (data_mode) {
1206	case DATA_MODE_SLAB:
1207		kmem_cache_free(c->slab_cache, data);
1208		break;
1209
1210	case DATA_MODE_GET_FREE_PAGES:
1211		free_pages((unsigned long)data,
1212			   c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
1213		break;
1214
1215	case DATA_MODE_VMALLOC:
1216		vfree(data);
1217		break;
1218
1219	default:
1220		DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
1221		       data_mode);
1222		BUG();
1223	}
1224}
1225
1226/*
1227 * Allocate buffer and its data.
1228 */
1229static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
1230{
1231	struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
1232
1233	if (!b)
1234		return NULL;
1235
1236	b->c = c;
1237
1238	b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
1239	if (!b->data) {
1240		kmem_cache_free(c->slab_buffer, b);
1241		return NULL;
1242	}
1243	adjust_total_allocated(b, false);
1244
1245#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1246	b->stack_len = 0;
1247#endif
1248	return b;
1249}
1250
1251/*
1252 * Free buffer and its data.
1253 */
1254static void free_buffer(struct dm_buffer *b)
1255{
1256	struct dm_bufio_client *c = b->c;
1257
1258	adjust_total_allocated(b, true);
1259	free_buffer_data(c, b->data, b->data_mode);
1260	kmem_cache_free(c->slab_buffer, b);
1261}
1262
1263/*
1264 *--------------------------------------------------------------------------
1265 * Submit I/O on the buffer.
1266 *
1267 * Bio interface is faster but it has some problems:
1268 *	the vector list is limited (increasing this limit increases
1269 *	memory-consumption per buffer, so it is not viable);
1270 *
1271 *	the memory must be direct-mapped, not vmalloced;
1272 *
1273 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
1274 * it is not vmalloced, try using the bio interface.
1275 *
1276 * If the buffer is big, if it is vmalloced or if the underlying device
1277 * rejects the bio because it is too large, use dm-io layer to do the I/O.
1278 * The dm-io layer splits the I/O into multiple requests, avoiding the above
1279 * shortcomings.
1280 *--------------------------------------------------------------------------
1281 */
1282
1283/*
1284 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
1285 * that the request was handled directly with bio interface.
1286 */
1287static void dmio_complete(unsigned long error, void *context)
1288{
1289	struct dm_buffer *b = context;
1290
1291	b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
1292}
1293
1294static void use_dmio(struct dm_buffer *b, enum req_op op, sector_t sector,
1295		     unsigned int n_sectors, unsigned int offset,
1296		     unsigned short ioprio)
1297{
1298	int r;
1299	struct dm_io_request io_req = {
1300		.bi_opf = op,
1301		.notify.fn = dmio_complete,
1302		.notify.context = b,
1303		.client = b->c->dm_io,
1304	};
1305	struct dm_io_region region = {
1306		.bdev = b->c->bdev,
1307		.sector = sector,
1308		.count = n_sectors,
1309	};
1310
1311	if (b->data_mode != DATA_MODE_VMALLOC) {
1312		io_req.mem.type = DM_IO_KMEM;
1313		io_req.mem.ptr.addr = (char *)b->data + offset;
1314	} else {
1315		io_req.mem.type = DM_IO_VMA;
1316		io_req.mem.ptr.vma = (char *)b->data + offset;
1317	}
1318
1319	r = dm_io(&io_req, 1, &region, NULL, ioprio);
1320	if (unlikely(r))
1321		b->end_io(b, errno_to_blk_status(r));
1322}
1323
1324static void bio_complete(struct bio *bio)
1325{
1326	struct dm_buffer *b = bio->bi_private;
1327	blk_status_t status = bio->bi_status;
1328
1329	bio_uninit(bio);
1330	kfree(bio);
1331	b->end_io(b, status);
1332}
1333
1334static void use_bio(struct dm_buffer *b, enum req_op op, sector_t sector,
1335		    unsigned int n_sectors, unsigned int offset,
1336		    unsigned short ioprio)
1337{
1338	struct bio *bio;
1339	char *ptr;
1340	unsigned int len;
1341
1342	bio = bio_kmalloc(1, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN);
1343	if (!bio) {
1344		use_dmio(b, op, sector, n_sectors, offset, ioprio);
1345		return;
1346	}
1347	bio_init(bio, b->c->bdev, bio->bi_inline_vecs, 1, op);
1348	bio->bi_iter.bi_sector = sector;
1349	bio->bi_end_io = bio_complete;
1350	bio->bi_private = b;
1351	bio->bi_ioprio = ioprio;
1352
1353	ptr = (char *)b->data + offset;
1354	len = n_sectors << SECTOR_SHIFT;
1355
1356	__bio_add_page(bio, virt_to_page(ptr), len, offset_in_page(ptr));
1357
1358	submit_bio(bio);
1359}
1360
1361static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
1362{
1363	sector_t sector;
1364
1365	if (likely(c->sectors_per_block_bits >= 0))
1366		sector = block << c->sectors_per_block_bits;
1367	else
1368		sector = block * (c->block_size >> SECTOR_SHIFT);
1369	sector += c->start;
1370
1371	return sector;
1372}
1373
1374static void submit_io(struct dm_buffer *b, enum req_op op, unsigned short ioprio,
1375		      void (*end_io)(struct dm_buffer *, blk_status_t))
1376{
1377	unsigned int n_sectors;
1378	sector_t sector;
1379	unsigned int offset, end;
1380
1381	b->end_io = end_io;
1382
1383	sector = block_to_sector(b->c, b->block);
1384
1385	if (op != REQ_OP_WRITE) {
1386		n_sectors = b->c->block_size >> SECTOR_SHIFT;
1387		offset = 0;
1388	} else {
1389		if (b->c->write_callback)
1390			b->c->write_callback(b);
1391		offset = b->write_start;
1392		end = b->write_end;
1393		offset &= -DM_BUFIO_WRITE_ALIGN;
1394		end += DM_BUFIO_WRITE_ALIGN - 1;
1395		end &= -DM_BUFIO_WRITE_ALIGN;
1396		if (unlikely(end > b->c->block_size))
1397			end = b->c->block_size;
1398
1399		sector += offset >> SECTOR_SHIFT;
1400		n_sectors = (end - offset) >> SECTOR_SHIFT;
1401	}
1402
1403	if (b->data_mode != DATA_MODE_VMALLOC)
1404		use_bio(b, op, sector, n_sectors, offset, ioprio);
1405	else
1406		use_dmio(b, op, sector, n_sectors, offset, ioprio);
1407}
1408
1409/*
1410 *--------------------------------------------------------------
1411 * Writing dirty buffers
1412 *--------------------------------------------------------------
1413 */
1414
1415/*
1416 * The endio routine for write.
1417 *
1418 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
1419 * it.
1420 */
1421static void write_endio(struct dm_buffer *b, blk_status_t status)
1422{
1423	b->write_error = status;
1424	if (unlikely(status)) {
1425		struct dm_bufio_client *c = b->c;
1426
1427		(void)cmpxchg(&c->async_write_error, 0,
1428				blk_status_to_errno(status));
1429	}
1430
1431	BUG_ON(!test_bit(B_WRITING, &b->state));
1432
1433	smp_mb__before_atomic();
1434	clear_bit(B_WRITING, &b->state);
1435	smp_mb__after_atomic();
1436
1437	wake_up_bit(&b->state, B_WRITING);
1438}
1439
1440/*
1441 * Initiate a write on a dirty buffer, but don't wait for it.
1442 *
1443 * - If the buffer is not dirty, exit.
1444 * - If there some previous write going on, wait for it to finish (we can't
1445 *   have two writes on the same buffer simultaneously).
1446 * - Submit our write and don't wait on it. We set B_WRITING indicating
1447 *   that there is a write in progress.
1448 */
1449static void __write_dirty_buffer(struct dm_buffer *b,
1450				 struct list_head *write_list)
1451{
1452	if (!test_bit(B_DIRTY, &b->state))
1453		return;
1454
1455	clear_bit(B_DIRTY, &b->state);
1456	wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
1457
1458	b->write_start = b->dirty_start;
1459	b->write_end = b->dirty_end;
1460
1461	if (!write_list)
1462		submit_io(b, REQ_OP_WRITE, IOPRIO_DEFAULT, write_endio);
1463	else
1464		list_add_tail(&b->write_list, write_list);
1465}
1466
1467static void __flush_write_list(struct list_head *write_list)
1468{
1469	struct blk_plug plug;
1470
1471	blk_start_plug(&plug);
1472	while (!list_empty(write_list)) {
1473		struct dm_buffer *b =
1474			list_entry(write_list->next, struct dm_buffer, write_list);
1475		list_del(&b->write_list);
1476		submit_io(b, REQ_OP_WRITE, IOPRIO_DEFAULT, write_endio);
1477		cond_resched();
1478	}
1479	blk_finish_plug(&plug);
1480}
1481
1482/*
1483 * Wait until any activity on the buffer finishes.  Possibly write the
1484 * buffer if it is dirty.  When this function finishes, there is no I/O
1485 * running on the buffer and the buffer is not dirty.
1486 */
1487static void __make_buffer_clean(struct dm_buffer *b)
1488{
1489	BUG_ON(atomic_read(&b->hold_count));
1490
1491	/* smp_load_acquire() pairs with read_endio()'s smp_mb__before_atomic() */
1492	if (!smp_load_acquire(&b->state))	/* fast case */
1493		return;
1494
1495	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1496	__write_dirty_buffer(b, NULL);
1497	wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
1498}
1499
1500static enum evict_result is_clean(struct dm_buffer *b, void *context)
1501{
1502	struct dm_bufio_client *c = context;
1503
1504	/* These should never happen */
1505	if (WARN_ON_ONCE(test_bit(B_WRITING, &b->state)))
1506		return ER_DONT_EVICT;
1507	if (WARN_ON_ONCE(test_bit(B_DIRTY, &b->state)))
1508		return ER_DONT_EVICT;
1509	if (WARN_ON_ONCE(b->list_mode != LIST_CLEAN))
1510		return ER_DONT_EVICT;
1511
1512	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep &&
1513	    unlikely(test_bit(B_READING, &b->state)))
1514		return ER_DONT_EVICT;
1515
1516	return ER_EVICT;
1517}
1518
1519static enum evict_result is_dirty(struct dm_buffer *b, void *context)
1520{
1521	/* These should never happen */
1522	if (WARN_ON_ONCE(test_bit(B_READING, &b->state)))
1523		return ER_DONT_EVICT;
1524	if (WARN_ON_ONCE(b->list_mode != LIST_DIRTY))
1525		return ER_DONT_EVICT;
1526
1527	return ER_EVICT;
1528}
1529
1530/*
1531 * Find some buffer that is not held by anybody, clean it, unlink it and
1532 * return it.
1533 */
1534static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
1535{
1536	struct dm_buffer *b;
1537
1538	b = cache_evict(&c->cache, LIST_CLEAN, is_clean, c);
1539	if (b) {
1540		/* this also waits for pending reads */
1541		__make_buffer_clean(b);
1542		return b;
1543	}
1544
1545	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
1546		return NULL;
1547
1548	b = cache_evict(&c->cache, LIST_DIRTY, is_dirty, NULL);
1549	if (b) {
1550		__make_buffer_clean(b);
1551		return b;
1552	}
1553
1554	return NULL;
1555}
1556
1557/*
1558 * Wait until some other threads free some buffer or release hold count on
1559 * some buffer.
1560 *
1561 * This function is entered with c->lock held, drops it and regains it
1562 * before exiting.
1563 */
1564static void __wait_for_free_buffer(struct dm_bufio_client *c)
1565{
1566	DECLARE_WAITQUEUE(wait, current);
1567
1568	add_wait_queue(&c->free_buffer_wait, &wait);
1569	set_current_state(TASK_UNINTERRUPTIBLE);
1570	dm_bufio_unlock(c);
1571
1572	/*
1573	 * It's possible to miss a wake up event since we don't always
1574	 * hold c->lock when wake_up is called.  So we have a timeout here,
1575	 * just in case.
1576	 */
1577	io_schedule_timeout(5 * HZ);
1578
1579	remove_wait_queue(&c->free_buffer_wait, &wait);
1580
1581	dm_bufio_lock(c);
1582}
1583
1584enum new_flag {
1585	NF_FRESH = 0,
1586	NF_READ = 1,
1587	NF_GET = 2,
1588	NF_PREFETCH = 3
1589};
1590
1591/*
1592 * Allocate a new buffer. If the allocation is not possible, wait until
1593 * some other thread frees a buffer.
1594 *
1595 * May drop the lock and regain it.
1596 */
1597static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
1598{
1599	struct dm_buffer *b;
1600	bool tried_noio_alloc = false;
1601
1602	/*
1603	 * dm-bufio is resistant to allocation failures (it just keeps
1604	 * one buffer reserved in cases all the allocations fail).
1605	 * So set flags to not try too hard:
1606	 *	GFP_NOWAIT: don't wait; if we need to sleep we'll release our
1607	 *		    mutex and wait ourselves.
1608	 *	__GFP_NORETRY: don't retry and rather return failure
1609	 *	__GFP_NOMEMALLOC: don't use emergency reserves
1610	 *	__GFP_NOWARN: don't print a warning in case of failure
1611	 *
1612	 * For debugging, if we set the cache size to 1, no new buffers will
1613	 * be allocated.
1614	 */
1615	while (1) {
1616		if (dm_bufio_cache_size_latch != 1) {
1617			b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
1618			if (b)
1619				return b;
1620		}
1621
1622		if (nf == NF_PREFETCH)
1623			return NULL;
1624
1625		if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
1626			dm_bufio_unlock(c);
1627			b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
1628			dm_bufio_lock(c);
1629			if (b)
1630				return b;
1631			tried_noio_alloc = true;
1632		}
1633
1634		if (!list_empty(&c->reserved_buffers)) {
1635			b = list_to_buffer(c->reserved_buffers.next);
1636			list_del(&b->lru.list);
1637			c->need_reserved_buffers++;
1638
1639			return b;
1640		}
1641
1642		b = __get_unclaimed_buffer(c);
1643		if (b)
1644			return b;
1645
1646		__wait_for_free_buffer(c);
1647	}
1648}
1649
1650static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
1651{
1652	struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
1653
1654	if (!b)
1655		return NULL;
1656
1657	if (c->alloc_callback)
1658		c->alloc_callback(b);
1659
1660	return b;
1661}
1662
1663/*
1664 * Free a buffer and wake other threads waiting for free buffers.
1665 */
1666static void __free_buffer_wake(struct dm_buffer *b)
1667{
1668	struct dm_bufio_client *c = b->c;
1669
1670	b->block = -1;
1671	if (!c->need_reserved_buffers)
1672		free_buffer(b);
1673	else {
1674		list_add(&b->lru.list, &c->reserved_buffers);
1675		c->need_reserved_buffers--;
1676	}
1677
1678	/*
1679	 * We hold the bufio lock here, so no one can add entries to the
1680	 * wait queue anyway.
1681	 */
1682	if (unlikely(waitqueue_active(&c->free_buffer_wait)))
1683		wake_up(&c->free_buffer_wait);
1684}
1685
1686static enum evict_result cleaned(struct dm_buffer *b, void *context)
1687{
1688	if (WARN_ON_ONCE(test_bit(B_READING, &b->state)))
1689		return ER_DONT_EVICT; /* should never happen */
1690
1691	if (test_bit(B_DIRTY, &b->state) || test_bit(B_WRITING, &b->state))
1692		return ER_DONT_EVICT;
1693	else
1694		return ER_EVICT;
1695}
1696
1697static void __move_clean_buffers(struct dm_bufio_client *c)
1698{
1699	cache_mark_many(&c->cache, LIST_DIRTY, LIST_CLEAN, cleaned, NULL);
1700}
1701
1702struct write_context {
1703	int no_wait;
1704	struct list_head *write_list;
1705};
1706
1707static enum it_action write_one(struct dm_buffer *b, void *context)
1708{
1709	struct write_context *wc = context;
1710
1711	if (wc->no_wait && test_bit(B_WRITING, &b->state))
1712		return IT_COMPLETE;
1713
1714	__write_dirty_buffer(b, wc->write_list);
1715	return IT_NEXT;
1716}
1717
1718static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
1719					struct list_head *write_list)
1720{
1721	struct write_context wc = {.no_wait = no_wait, .write_list = write_list};
1722
1723	__move_clean_buffers(c);
1724	cache_iterate(&c->cache, LIST_DIRTY, write_one, &wc);
1725}
1726
1727/*
1728 * Check if we're over watermark.
1729 * If we are over threshold_buffers, start freeing buffers.
1730 * If we're over "limit_buffers", block until we get under the limit.
1731 */
1732static void __check_watermark(struct dm_bufio_client *c,
1733			      struct list_head *write_list)
1734{
1735	if (cache_count(&c->cache, LIST_DIRTY) >
1736	    cache_count(&c->cache, LIST_CLEAN) * DM_BUFIO_WRITEBACK_RATIO)
1737		__write_dirty_buffers_async(c, 1, write_list);
1738}
1739
1740/*
1741 *--------------------------------------------------------------
1742 * Getting a buffer
1743 *--------------------------------------------------------------
1744 */
1745
1746static void cache_put_and_wake(struct dm_bufio_client *c, struct dm_buffer *b)
1747{
1748	/*
1749	 * Relying on waitqueue_active() is racey, but we sleep
1750	 * with schedule_timeout anyway.
1751	 */
1752	if (cache_put(&c->cache, b) &&
1753	    unlikely(waitqueue_active(&c->free_buffer_wait)))
1754		wake_up(&c->free_buffer_wait);
1755}
1756
1757/*
1758 * This assumes you have already checked the cache to see if the buffer
1759 * is already present (it will recheck after dropping the lock for allocation).
1760 */
1761static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
1762				     enum new_flag nf, int *need_submit,
1763				     struct list_head *write_list)
1764{
1765	struct dm_buffer *b, *new_b = NULL;
1766
1767	*need_submit = 0;
1768
1769	/* This can't be called with NF_GET */
1770	if (WARN_ON_ONCE(nf == NF_GET))
1771		return NULL;
1772
1773	new_b = __alloc_buffer_wait(c, nf);
1774	if (!new_b)
1775		return NULL;
1776
1777	/*
1778	 * We've had a period where the mutex was unlocked, so need to
1779	 * recheck the buffer tree.
1780	 */
1781	b = cache_get(&c->cache, block);
1782	if (b) {
1783		__free_buffer_wake(new_b);
1784		goto found_buffer;
1785	}
1786
1787	__check_watermark(c, write_list);
1788
1789	b = new_b;
1790	atomic_set(&b->hold_count, 1);
1791	WRITE_ONCE(b->last_accessed, jiffies);
1792	b->block = block;
1793	b->read_error = 0;
1794	b->write_error = 0;
1795	b->list_mode = LIST_CLEAN;
1796
1797	if (nf == NF_FRESH)
1798		b->state = 0;
1799	else {
1800		b->state = 1 << B_READING;
1801		*need_submit = 1;
1802	}
1803
1804	/*
1805	 * We mustn't insert into the cache until the B_READING state
1806	 * is set.  Otherwise another thread could get it and use
1807	 * it before it had been read.
1808	 */
1809	cache_insert(&c->cache, b);
1810
1811	return b;
1812
1813found_buffer:
1814	if (nf == NF_PREFETCH) {
1815		cache_put_and_wake(c, b);
1816		return NULL;
1817	}
1818
1819	/*
1820	 * Note: it is essential that we don't wait for the buffer to be
1821	 * read if dm_bufio_get function is used. Both dm_bufio_get and
1822	 * dm_bufio_prefetch can be used in the driver request routine.
1823	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1824	 * the same buffer, it would deadlock if we waited.
1825	 */
1826	if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state))) {
1827		cache_put_and_wake(c, b);
1828		return NULL;
1829	}
1830
1831	return b;
1832}
1833
1834/*
1835 * The endio routine for reading: set the error, clear the bit and wake up
1836 * anyone waiting on the buffer.
1837 */
1838static void read_endio(struct dm_buffer *b, blk_status_t status)
1839{
1840	b->read_error = status;
1841
1842	BUG_ON(!test_bit(B_READING, &b->state));
1843
1844	smp_mb__before_atomic();
1845	clear_bit(B_READING, &b->state);
1846	smp_mb__after_atomic();
1847
1848	wake_up_bit(&b->state, B_READING);
1849}
1850
1851/*
1852 * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1853 * functions is similar except that dm_bufio_new doesn't read the
1854 * buffer from the disk (assuming that the caller overwrites all the data
1855 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1856 */
1857static void *new_read(struct dm_bufio_client *c, sector_t block,
1858		      enum new_flag nf, struct dm_buffer **bp,
1859		      unsigned short ioprio)
1860{
1861	int need_submit = 0;
1862	struct dm_buffer *b;
1863
1864	LIST_HEAD(write_list);
1865
1866	*bp = NULL;
1867
1868	/*
1869	 * Fast path, hopefully the block is already in the cache.  No need
1870	 * to get the client lock for this.
1871	 */
1872	b = cache_get(&c->cache, block);
1873	if (b) {
1874		if (nf == NF_PREFETCH) {
1875			cache_put_and_wake(c, b);
1876			return NULL;
1877		}
1878
1879		/*
1880		 * Note: it is essential that we don't wait for the buffer to be
1881		 * read if dm_bufio_get function is used. Both dm_bufio_get and
1882		 * dm_bufio_prefetch can be used in the driver request routine.
1883		 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1884		 * the same buffer, it would deadlock if we waited.
1885		 */
1886		if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state))) {
1887			cache_put_and_wake(c, b);
1888			return NULL;
1889		}
1890	}
1891
1892	if (!b) {
1893		if (nf == NF_GET)
1894			return NULL;
1895
1896		dm_bufio_lock(c);
1897		b = __bufio_new(c, block, nf, &need_submit, &write_list);
1898		dm_bufio_unlock(c);
1899	}
1900
1901#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1902	if (b && (atomic_read(&b->hold_count) == 1))
1903		buffer_record_stack(b);
1904#endif
1905
1906	__flush_write_list(&write_list);
1907
1908	if (!b)
1909		return NULL;
1910
1911	if (need_submit)
1912		submit_io(b, REQ_OP_READ, ioprio, read_endio);
1913
1914	if (nf != NF_GET)	/* we already tested this condition above */
1915		wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1916
1917	if (b->read_error) {
1918		int error = blk_status_to_errno(b->read_error);
1919
1920		dm_bufio_release(b);
1921
1922		return ERR_PTR(error);
1923	}
1924
1925	*bp = b;
1926
1927	return b->data;
1928}
1929
1930void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1931		   struct dm_buffer **bp)
1932{
1933	return new_read(c, block, NF_GET, bp, IOPRIO_DEFAULT);
1934}
1935EXPORT_SYMBOL_GPL(dm_bufio_get);
1936
1937static void *__dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1938			struct dm_buffer **bp, unsigned short ioprio)
1939{
1940	if (WARN_ON_ONCE(dm_bufio_in_request()))
1941		return ERR_PTR(-EINVAL);
1942
1943	return new_read(c, block, NF_READ, bp, ioprio);
1944}
1945
1946void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1947		    struct dm_buffer **bp)
1948{
1949	return __dm_bufio_read(c, block, bp, IOPRIO_DEFAULT);
1950}
1951EXPORT_SYMBOL_GPL(dm_bufio_read);
1952
1953void *dm_bufio_read_with_ioprio(struct dm_bufio_client *c, sector_t block,
1954				struct dm_buffer **bp, unsigned short ioprio)
1955{
1956	return __dm_bufio_read(c, block, bp, ioprio);
1957}
1958EXPORT_SYMBOL_GPL(dm_bufio_read_with_ioprio);
1959
1960void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1961		   struct dm_buffer **bp)
1962{
1963	if (WARN_ON_ONCE(dm_bufio_in_request()))
1964		return ERR_PTR(-EINVAL);
1965
1966	return new_read(c, block, NF_FRESH, bp, IOPRIO_DEFAULT);
1967}
1968EXPORT_SYMBOL_GPL(dm_bufio_new);
1969
1970static void __dm_bufio_prefetch(struct dm_bufio_client *c,
1971			sector_t block, unsigned int n_blocks,
1972			unsigned short ioprio)
1973{
1974	struct blk_plug plug;
1975
1976	LIST_HEAD(write_list);
1977
1978	if (WARN_ON_ONCE(dm_bufio_in_request()))
1979		return; /* should never happen */
1980
1981	blk_start_plug(&plug);
1982
1983	for (; n_blocks--; block++) {
1984		int need_submit;
1985		struct dm_buffer *b;
1986
1987		b = cache_get(&c->cache, block);
1988		if (b) {
1989			/* already in cache */
1990			cache_put_and_wake(c, b);
1991			continue;
1992		}
1993
1994		dm_bufio_lock(c);
1995		b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1996				&write_list);
1997		if (unlikely(!list_empty(&write_list))) {
1998			dm_bufio_unlock(c);
1999			blk_finish_plug(&plug);
2000			__flush_write_list(&write_list);
2001			blk_start_plug(&plug);
2002			dm_bufio_lock(c);
2003		}
2004		if (unlikely(b != NULL)) {
2005			dm_bufio_unlock(c);
2006
2007			if (need_submit)
2008				submit_io(b, REQ_OP_READ, ioprio, read_endio);
2009			dm_bufio_release(b);
2010
2011			cond_resched();
2012
2013			if (!n_blocks)
2014				goto flush_plug;
2015			dm_bufio_lock(c);
2016		}
2017		dm_bufio_unlock(c);
2018	}
2019
2020flush_plug:
2021	blk_finish_plug(&plug);
2022}
2023
2024void dm_bufio_prefetch(struct dm_bufio_client *c, sector_t block, unsigned int n_blocks)
2025{
2026	return __dm_bufio_prefetch(c, block, n_blocks, IOPRIO_DEFAULT);
2027}
2028EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
2029
2030void dm_bufio_prefetch_with_ioprio(struct dm_bufio_client *c, sector_t block,
2031				unsigned int n_blocks, unsigned short ioprio)
2032{
2033	return __dm_bufio_prefetch(c, block, n_blocks, ioprio);
2034}
2035EXPORT_SYMBOL_GPL(dm_bufio_prefetch_with_ioprio);
2036
2037void dm_bufio_release(struct dm_buffer *b)
2038{
2039	struct dm_bufio_client *c = b->c;
2040
2041	/*
2042	 * If there were errors on the buffer, and the buffer is not
2043	 * to be written, free the buffer. There is no point in caching
2044	 * invalid buffer.
2045	 */
2046	if ((b->read_error || b->write_error) &&
2047	    !test_bit_acquire(B_READING, &b->state) &&
2048	    !test_bit(B_WRITING, &b->state) &&
2049	    !test_bit(B_DIRTY, &b->state)) {
2050		dm_bufio_lock(c);
2051
2052		/* cache remove can fail if there are other holders */
2053		if (cache_remove(&c->cache, b)) {
2054			__free_buffer_wake(b);
2055			dm_bufio_unlock(c);
2056			return;
2057		}
2058
2059		dm_bufio_unlock(c);
2060	}
2061
2062	cache_put_and_wake(c, b);
2063}
2064EXPORT_SYMBOL_GPL(dm_bufio_release);
2065
2066void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
2067					unsigned int start, unsigned int end)
2068{
2069	struct dm_bufio_client *c = b->c;
2070
2071	BUG_ON(start >= end);
2072	BUG_ON(end > b->c->block_size);
2073
2074	dm_bufio_lock(c);
2075
2076	BUG_ON(test_bit(B_READING, &b->state));
2077
2078	if (!test_and_set_bit(B_DIRTY, &b->state)) {
2079		b->dirty_start = start;
2080		b->dirty_end = end;
2081		cache_mark(&c->cache, b, LIST_DIRTY);
2082	} else {
2083		if (start < b->dirty_start)
2084			b->dirty_start = start;
2085		if (end > b->dirty_end)
2086			b->dirty_end = end;
2087	}
2088
2089	dm_bufio_unlock(c);
2090}
2091EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
2092
2093void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
2094{
2095	dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
2096}
2097EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
2098
2099void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
2100{
2101	LIST_HEAD(write_list);
2102
2103	if (WARN_ON_ONCE(dm_bufio_in_request()))
2104		return; /* should never happen */
2105
2106	dm_bufio_lock(c);
2107	__write_dirty_buffers_async(c, 0, &write_list);
2108	dm_bufio_unlock(c);
2109	__flush_write_list(&write_list);
2110}
2111EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
2112
2113/*
2114 * For performance, it is essential that the buffers are written asynchronously
2115 * and simultaneously (so that the block layer can merge the writes) and then
2116 * waited upon.
2117 *
2118 * Finally, we flush hardware disk cache.
2119 */
2120static bool is_writing(struct lru_entry *e, void *context)
2121{
2122	struct dm_buffer *b = le_to_buffer(e);
2123
2124	return test_bit(B_WRITING, &b->state);
2125}
2126
2127int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
2128{
2129	int a, f;
2130	unsigned long nr_buffers;
2131	struct lru_entry *e;
2132	struct lru_iter it;
2133
2134	LIST_HEAD(write_list);
2135
2136	dm_bufio_lock(c);
2137	__write_dirty_buffers_async(c, 0, &write_list);
2138	dm_bufio_unlock(c);
2139	__flush_write_list(&write_list);
2140	dm_bufio_lock(c);
2141
2142	nr_buffers = cache_count(&c->cache, LIST_DIRTY);
2143	lru_iter_begin(&c->cache.lru[LIST_DIRTY], &it);
2144	while ((e = lru_iter_next(&it, is_writing, c))) {
2145		struct dm_buffer *b = le_to_buffer(e);
2146		__cache_inc_buffer(b);
2147
2148		BUG_ON(test_bit(B_READING, &b->state));
2149
2150		if (nr_buffers) {
2151			nr_buffers--;
2152			dm_bufio_unlock(c);
2153			wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
2154			dm_bufio_lock(c);
2155		} else {
2156			wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
2157		}
2158
2159		if (!test_bit(B_DIRTY, &b->state) && !test_bit(B_WRITING, &b->state))
2160			cache_mark(&c->cache, b, LIST_CLEAN);
2161
2162		cache_put_and_wake(c, b);
2163
2164		cond_resched();
2165	}
2166	lru_iter_end(&it);
2167
2168	wake_up(&c->free_buffer_wait);
2169	dm_bufio_unlock(c);
2170
2171	a = xchg(&c->async_write_error, 0);
2172	f = dm_bufio_issue_flush(c);
2173	if (a)
2174		return a;
2175
2176	return f;
2177}
2178EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
2179
2180/*
2181 * Use dm-io to send an empty barrier to flush the device.
2182 */
2183int dm_bufio_issue_flush(struct dm_bufio_client *c)
2184{
2185	struct dm_io_request io_req = {
2186		.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC,
2187		.mem.type = DM_IO_KMEM,
2188		.mem.ptr.addr = NULL,
2189		.client = c->dm_io,
2190	};
2191	struct dm_io_region io_reg = {
2192		.bdev = c->bdev,
2193		.sector = 0,
2194		.count = 0,
2195	};
2196
2197	if (WARN_ON_ONCE(dm_bufio_in_request()))
2198		return -EINVAL;
2199
2200	return dm_io(&io_req, 1, &io_reg, NULL, IOPRIO_DEFAULT);
2201}
2202EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
2203
2204/*
2205 * Use dm-io to send a discard request to flush the device.
2206 */
2207int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
2208{
2209	struct dm_io_request io_req = {
2210		.bi_opf = REQ_OP_DISCARD | REQ_SYNC,
2211		.mem.type = DM_IO_KMEM,
2212		.mem.ptr.addr = NULL,
2213		.client = c->dm_io,
2214	};
2215	struct dm_io_region io_reg = {
2216		.bdev = c->bdev,
2217		.sector = block_to_sector(c, block),
2218		.count = block_to_sector(c, count),
2219	};
2220
2221	if (WARN_ON_ONCE(dm_bufio_in_request()))
2222		return -EINVAL; /* discards are optional */
2223
2224	return dm_io(&io_req, 1, &io_reg, NULL, IOPRIO_DEFAULT);
2225}
2226EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
2227
2228static bool forget_buffer(struct dm_bufio_client *c, sector_t block)
2229{
2230	struct dm_buffer *b;
2231
2232	b = cache_get(&c->cache, block);
2233	if (b) {
2234		if (likely(!smp_load_acquire(&b->state))) {
2235			if (cache_remove(&c->cache, b))
2236				__free_buffer_wake(b);
2237			else
2238				cache_put_and_wake(c, b);
2239		} else {
2240			cache_put_and_wake(c, b);
2241		}
2242	}
2243
2244	return b ? true : false;
2245}
2246
2247/*
2248 * Free the given buffer.
2249 *
2250 * This is just a hint, if the buffer is in use or dirty, this function
2251 * does nothing.
2252 */
2253void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
2254{
2255	dm_bufio_lock(c);
2256	forget_buffer(c, block);
2257	dm_bufio_unlock(c);
2258}
2259EXPORT_SYMBOL_GPL(dm_bufio_forget);
2260
2261static enum evict_result idle(struct dm_buffer *b, void *context)
2262{
2263	return b->state ? ER_DONT_EVICT : ER_EVICT;
2264}
2265
2266void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
2267{
2268	dm_bufio_lock(c);
2269	cache_remove_range(&c->cache, block, block + n_blocks, idle, __free_buffer_wake);
2270	dm_bufio_unlock(c);
2271}
2272EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
2273
2274void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned int n)
2275{
2276	c->minimum_buffers = n;
2277}
2278EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
2279
2280unsigned int dm_bufio_get_block_size(struct dm_bufio_client *c)
2281{
2282	return c->block_size;
2283}
2284EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
2285
2286sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
2287{
2288	sector_t s = bdev_nr_sectors(c->bdev);
2289
2290	if (s >= c->start)
2291		s -= c->start;
2292	else
2293		s = 0;
2294	if (likely(c->sectors_per_block_bits >= 0))
2295		s >>= c->sectors_per_block_bits;
2296	else
2297		sector_div(s, c->block_size >> SECTOR_SHIFT);
2298	return s;
2299}
2300EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
2301
2302struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
2303{
2304	return c->dm_io;
2305}
2306EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
2307
2308sector_t dm_bufio_get_block_number(struct dm_buffer *b)
2309{
2310	return b->block;
2311}
2312EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
2313
2314void *dm_bufio_get_block_data(struct dm_buffer *b)
2315{
2316	return b->data;
2317}
2318EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
2319
2320void *dm_bufio_get_aux_data(struct dm_buffer *b)
2321{
2322	return b + 1;
2323}
2324EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
2325
2326struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
2327{
2328	return b->c;
2329}
2330EXPORT_SYMBOL_GPL(dm_bufio_get_client);
2331
2332static enum it_action warn_leak(struct dm_buffer *b, void *context)
2333{
2334	bool *warned = context;
2335
2336	WARN_ON(!(*warned));
2337	*warned = true;
2338	DMERR("leaked buffer %llx, hold count %u, list %d",
2339	      (unsigned long long)b->block, atomic_read(&b->hold_count), b->list_mode);
2340#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
2341	stack_trace_print(b->stack_entries, b->stack_len, 1);
2342	/* mark unclaimed to avoid WARN_ON at end of drop_buffers() */
2343	atomic_set(&b->hold_count, 0);
2344#endif
2345	return IT_NEXT;
2346}
2347
2348static void drop_buffers(struct dm_bufio_client *c)
2349{
2350	int i;
2351	struct dm_buffer *b;
2352
2353	if (WARN_ON(dm_bufio_in_request()))
2354		return; /* should never happen */
2355
2356	/*
2357	 * An optimization so that the buffers are not written one-by-one.
2358	 */
2359	dm_bufio_write_dirty_buffers_async(c);
2360
2361	dm_bufio_lock(c);
2362
2363	while ((b = __get_unclaimed_buffer(c)))
2364		__free_buffer_wake(b);
2365
2366	for (i = 0; i < LIST_SIZE; i++) {
2367		bool warned = false;
2368
2369		cache_iterate(&c->cache, i, warn_leak, &warned);
2370	}
2371
2372#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
2373	while ((b = __get_unclaimed_buffer(c)))
2374		__free_buffer_wake(b);
2375#endif
2376
2377	for (i = 0; i < LIST_SIZE; i++)
2378		WARN_ON(cache_count(&c->cache, i));
2379
2380	dm_bufio_unlock(c);
2381}
2382
2383static unsigned long get_retain_buffers(struct dm_bufio_client *c)
2384{
2385	unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
2386
2387	if (likely(c->sectors_per_block_bits >= 0))
2388		retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
2389	else
2390		retain_bytes /= c->block_size;
2391
2392	return retain_bytes;
2393}
2394
2395static void __scan(struct dm_bufio_client *c)
2396{
2397	int l;
2398	struct dm_buffer *b;
2399	unsigned long freed = 0;
2400	unsigned long retain_target = get_retain_buffers(c);
2401	unsigned long count = cache_total(&c->cache);
2402
2403	for (l = 0; l < LIST_SIZE; l++) {
2404		while (true) {
2405			if (count - freed <= retain_target)
2406				atomic_long_set(&c->need_shrink, 0);
2407			if (!atomic_long_read(&c->need_shrink))
2408				break;
2409
2410			b = cache_evict(&c->cache, l,
2411					l == LIST_CLEAN ? is_clean : is_dirty, c);
2412			if (!b)
2413				break;
2414
2415			__make_buffer_clean(b);
2416			__free_buffer_wake(b);
2417
2418			atomic_long_dec(&c->need_shrink);
2419			freed++;
2420			cond_resched();
2421		}
2422	}
2423}
2424
2425static void shrink_work(struct work_struct *w)
2426{
2427	struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
2428
2429	dm_bufio_lock(c);
2430	__scan(c);
2431	dm_bufio_unlock(c);
2432}
2433
2434static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
2435{
2436	struct dm_bufio_client *c;
2437
2438	c = shrink->private_data;
2439	atomic_long_add(sc->nr_to_scan, &c->need_shrink);
2440	queue_work(dm_bufio_wq, &c->shrink_work);
2441
2442	return sc->nr_to_scan;
2443}
2444
2445static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
2446{
2447	struct dm_bufio_client *c = shrink->private_data;
2448	unsigned long count = cache_total(&c->cache);
2449	unsigned long retain_target = get_retain_buffers(c);
2450	unsigned long queued_for_cleanup = atomic_long_read(&c->need_shrink);
2451
2452	if (unlikely(count < retain_target))
2453		count = 0;
2454	else
2455		count -= retain_target;
2456
2457	if (unlikely(count < queued_for_cleanup))
2458		count = 0;
2459	else
2460		count -= queued_for_cleanup;
2461
2462	return count;
2463}
2464
2465/*
2466 * Create the buffering interface
2467 */
2468struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned int block_size,
2469					       unsigned int reserved_buffers, unsigned int aux_size,
2470					       void (*alloc_callback)(struct dm_buffer *),
2471					       void (*write_callback)(struct dm_buffer *),
2472					       unsigned int flags)
2473{
2474	int r;
2475	unsigned int num_locks;
2476	struct dm_bufio_client *c;
2477	char slab_name[27];
2478
2479	if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
2480		DMERR("%s: block size not specified or is not multiple of 512b", __func__);
2481		r = -EINVAL;
2482		goto bad_client;
2483	}
2484
2485	num_locks = dm_num_hash_locks();
2486	c = kzalloc(sizeof(*c) + (num_locks * sizeof(struct buffer_tree)), GFP_KERNEL);
2487	if (!c) {
2488		r = -ENOMEM;
2489		goto bad_client;
2490	}
2491	cache_init(&c->cache, num_locks, (flags & DM_BUFIO_CLIENT_NO_SLEEP) != 0);
2492
2493	c->bdev = bdev;
2494	c->block_size = block_size;
2495	if (is_power_of_2(block_size))
2496		c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
2497	else
2498		c->sectors_per_block_bits = -1;
2499
2500	c->alloc_callback = alloc_callback;
2501	c->write_callback = write_callback;
2502
2503	if (flags & DM_BUFIO_CLIENT_NO_SLEEP) {
2504		c->no_sleep = true;
2505		static_branch_inc(&no_sleep_enabled);
2506	}
2507
2508	mutex_init(&c->lock);
2509	spin_lock_init(&c->spinlock);
2510	INIT_LIST_HEAD(&c->reserved_buffers);
2511	c->need_reserved_buffers = reserved_buffers;
2512
2513	dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
2514
2515	init_waitqueue_head(&c->free_buffer_wait);
2516	c->async_write_error = 0;
2517
2518	c->dm_io = dm_io_client_create();
2519	if (IS_ERR(c->dm_io)) {
2520		r = PTR_ERR(c->dm_io);
2521		goto bad_dm_io;
2522	}
2523
2524	if (block_size <= KMALLOC_MAX_SIZE &&
2525	    (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
2526		unsigned int align = min(1U << __ffs(block_size), (unsigned int)PAGE_SIZE);
2527
2528		snprintf(slab_name, sizeof(slab_name), "dm_bufio_cache-%u", block_size);
2529		c->slab_cache = kmem_cache_create(slab_name, block_size, align,
2530						  SLAB_RECLAIM_ACCOUNT, NULL);
2531		if (!c->slab_cache) {
2532			r = -ENOMEM;
2533			goto bad;
2534		}
2535	}
2536	if (aux_size)
2537		snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer-%u", aux_size);
2538	else
2539		snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer");
2540	c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
2541					   0, SLAB_RECLAIM_ACCOUNT, NULL);
2542	if (!c->slab_buffer) {
2543		r = -ENOMEM;
2544		goto bad;
2545	}
2546
2547	while (c->need_reserved_buffers) {
2548		struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
2549
2550		if (!b) {
2551			r = -ENOMEM;
2552			goto bad;
2553		}
2554		__free_buffer_wake(b);
2555	}
2556
2557	INIT_WORK(&c->shrink_work, shrink_work);
2558	atomic_long_set(&c->need_shrink, 0);
2559
2560	c->shrinker = shrinker_alloc(0, "dm-bufio:(%u:%u)",
2561				     MAJOR(bdev->bd_dev), MINOR(bdev->bd_dev));
2562	if (!c->shrinker) {
2563		r = -ENOMEM;
2564		goto bad;
2565	}
2566
2567	c->shrinker->count_objects = dm_bufio_shrink_count;
2568	c->shrinker->scan_objects = dm_bufio_shrink_scan;
2569	c->shrinker->seeks = 1;
2570	c->shrinker->batch = 0;
2571	c->shrinker->private_data = c;
2572
2573	shrinker_register(c->shrinker);
2574
2575	mutex_lock(&dm_bufio_clients_lock);
2576	dm_bufio_client_count++;
2577	list_add(&c->client_list, &dm_bufio_all_clients);
2578	__cache_size_refresh();
2579	mutex_unlock(&dm_bufio_clients_lock);
2580
2581	return c;
2582
2583bad:
2584	while (!list_empty(&c->reserved_buffers)) {
2585		struct dm_buffer *b = list_to_buffer(c->reserved_buffers.next);
2586
2587		list_del(&b->lru.list);
2588		free_buffer(b);
2589	}
2590	kmem_cache_destroy(c->slab_cache);
2591	kmem_cache_destroy(c->slab_buffer);
2592	dm_io_client_destroy(c->dm_io);
2593bad_dm_io:
2594	mutex_destroy(&c->lock);
2595	if (c->no_sleep)
2596		static_branch_dec(&no_sleep_enabled);
2597	kfree(c);
2598bad_client:
2599	return ERR_PTR(r);
2600}
2601EXPORT_SYMBOL_GPL(dm_bufio_client_create);
2602
2603/*
2604 * Free the buffering interface.
2605 * It is required that there are no references on any buffers.
2606 */
2607void dm_bufio_client_destroy(struct dm_bufio_client *c)
2608{
2609	unsigned int i;
2610
2611	drop_buffers(c);
2612
2613	shrinker_free(c->shrinker);
2614	flush_work(&c->shrink_work);
2615
2616	mutex_lock(&dm_bufio_clients_lock);
2617
2618	list_del(&c->client_list);
2619	dm_bufio_client_count--;
2620	__cache_size_refresh();
2621
2622	mutex_unlock(&dm_bufio_clients_lock);
2623
2624	WARN_ON(c->need_reserved_buffers);
2625
2626	while (!list_empty(&c->reserved_buffers)) {
2627		struct dm_buffer *b = list_to_buffer(c->reserved_buffers.next);
2628
2629		list_del(&b->lru.list);
2630		free_buffer(b);
2631	}
2632
2633	for (i = 0; i < LIST_SIZE; i++)
2634		if (cache_count(&c->cache, i))
2635			DMERR("leaked buffer count %d: %lu", i, cache_count(&c->cache, i));
2636
2637	for (i = 0; i < LIST_SIZE; i++)
2638		WARN_ON(cache_count(&c->cache, i));
2639
2640	cache_destroy(&c->cache);
2641	kmem_cache_destroy(c->slab_cache);
2642	kmem_cache_destroy(c->slab_buffer);
2643	dm_io_client_destroy(c->dm_io);
2644	mutex_destroy(&c->lock);
2645	if (c->no_sleep)
2646		static_branch_dec(&no_sleep_enabled);
2647	kfree(c);
2648}
2649EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
2650
2651void dm_bufio_client_reset(struct dm_bufio_client *c)
2652{
2653	drop_buffers(c);
2654	flush_work(&c->shrink_work);
2655}
2656EXPORT_SYMBOL_GPL(dm_bufio_client_reset);
2657
2658void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
2659{
2660	c->start = start;
2661}
2662EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
2663
2664/*--------------------------------------------------------------*/
2665
2666static unsigned int get_max_age_hz(void)
2667{
2668	unsigned int max_age = READ_ONCE(dm_bufio_max_age);
2669
2670	if (max_age > UINT_MAX / HZ)
2671		max_age = UINT_MAX / HZ;
2672
2673	return max_age * HZ;
2674}
2675
2676static bool older_than(struct dm_buffer *b, unsigned long age_hz)
2677{
2678	return time_after_eq(jiffies, READ_ONCE(b->last_accessed) + age_hz);
2679}
2680
2681struct evict_params {
2682	gfp_t gfp;
2683	unsigned long age_hz;
2684
2685	/*
2686	 * This gets updated with the largest last_accessed (ie. most
2687	 * recently used) of the evicted buffers.  It will not be reinitialised
2688	 * by __evict_many(), so you can use it across multiple invocations.
2689	 */
2690	unsigned long last_accessed;
2691};
2692
2693/*
2694 * We may not be able to evict this buffer if IO pending or the client
2695 * is still using it.
2696 *
2697 * And if GFP_NOFS is used, we must not do any I/O because we hold
2698 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
2699 * rerouted to different bufio client.
2700 */
2701static enum evict_result select_for_evict(struct dm_buffer *b, void *context)
2702{
2703	struct evict_params *params = context;
2704
2705	if (!(params->gfp & __GFP_FS) ||
2706	    (static_branch_unlikely(&no_sleep_enabled) && b->c->no_sleep)) {
2707		if (test_bit_acquire(B_READING, &b->state) ||
2708		    test_bit(B_WRITING, &b->state) ||
2709		    test_bit(B_DIRTY, &b->state))
2710			return ER_DONT_EVICT;
2711	}
2712
2713	return older_than(b, params->age_hz) ? ER_EVICT : ER_STOP;
2714}
2715
2716static unsigned long __evict_many(struct dm_bufio_client *c,
2717				  struct evict_params *params,
2718				  int list_mode, unsigned long max_count)
2719{
2720	unsigned long count;
2721	unsigned long last_accessed;
2722	struct dm_buffer *b;
2723
2724	for (count = 0; count < max_count; count++) {
2725		b = cache_evict(&c->cache, list_mode, select_for_evict, params);
2726		if (!b)
2727			break;
2728
2729		last_accessed = READ_ONCE(b->last_accessed);
2730		if (time_after_eq(params->last_accessed, last_accessed))
2731			params->last_accessed = last_accessed;
2732
2733		__make_buffer_clean(b);
2734		__free_buffer_wake(b);
2735
2736		cond_resched();
2737	}
2738
2739	return count;
2740}
2741
2742static void evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
2743{
2744	struct evict_params params = {.gfp = 0, .age_hz = age_hz, .last_accessed = 0};
2745	unsigned long retain = get_retain_buffers(c);
2746	unsigned long count;
2747	LIST_HEAD(write_list);
2748
2749	dm_bufio_lock(c);
2750
2751	__check_watermark(c, &write_list);
2752	if (unlikely(!list_empty(&write_list))) {
2753		dm_bufio_unlock(c);
2754		__flush_write_list(&write_list);
2755		dm_bufio_lock(c);
2756	}
2757
2758	count = cache_total(&c->cache);
2759	if (count > retain)
2760		__evict_many(c, &params, LIST_CLEAN, count - retain);
2761
2762	dm_bufio_unlock(c);
2763}
2764
2765static void cleanup_old_buffers(void)
2766{
2767	unsigned long max_age_hz = get_max_age_hz();
2768	struct dm_bufio_client *c;
2769
2770	mutex_lock(&dm_bufio_clients_lock);
2771
2772	__cache_size_refresh();
2773
2774	list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2775		evict_old_buffers(c, max_age_hz);
2776
2777	mutex_unlock(&dm_bufio_clients_lock);
2778}
2779
2780static void work_fn(struct work_struct *w)
2781{
2782	cleanup_old_buffers();
2783
2784	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2785			   DM_BUFIO_WORK_TIMER_SECS * HZ);
2786}
2787
2788/*--------------------------------------------------------------*/
2789
2790/*
2791 * Global cleanup tries to evict the oldest buffers from across _all_
2792 * the clients.  It does this by repeatedly evicting a few buffers from
2793 * the client that holds the oldest buffer.  It's approximate, but hopefully
2794 * good enough.
2795 */
2796static struct dm_bufio_client *__pop_client(void)
2797{
2798	struct list_head *h;
2799
2800	if (list_empty(&dm_bufio_all_clients))
2801		return NULL;
2802
2803	h = dm_bufio_all_clients.next;
2804	list_del(h);
2805	return container_of(h, struct dm_bufio_client, client_list);
2806}
2807
2808/*
2809 * Inserts the client in the global client list based on its
2810 * 'oldest_buffer' field.
2811 */
2812static void __insert_client(struct dm_bufio_client *new_client)
2813{
2814	struct dm_bufio_client *c;
2815	struct list_head *h = dm_bufio_all_clients.next;
2816
2817	while (h != &dm_bufio_all_clients) {
2818		c = container_of(h, struct dm_bufio_client, client_list);
2819		if (time_after_eq(c->oldest_buffer, new_client->oldest_buffer))
2820			break;
2821		h = h->next;
2822	}
2823
2824	list_add_tail(&new_client->client_list, h);
2825}
2826
2827static unsigned long __evict_a_few(unsigned long nr_buffers)
2828{
2829	unsigned long count;
2830	struct dm_bufio_client *c;
2831	struct evict_params params = {
2832		.gfp = GFP_KERNEL,
2833		.age_hz = 0,
2834		/* set to jiffies in case there are no buffers in this client */
2835		.last_accessed = jiffies
2836	};
2837
2838	c = __pop_client();
2839	if (!c)
2840		return 0;
2841
2842	dm_bufio_lock(c);
2843	count = __evict_many(c, &params, LIST_CLEAN, nr_buffers);
2844	dm_bufio_unlock(c);
2845
2846	if (count)
2847		c->oldest_buffer = params.last_accessed;
2848	__insert_client(c);
2849
2850	return count;
2851}
2852
2853static void check_watermarks(void)
2854{
2855	LIST_HEAD(write_list);
2856	struct dm_bufio_client *c;
2857
2858	mutex_lock(&dm_bufio_clients_lock);
2859	list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
2860		dm_bufio_lock(c);
2861		__check_watermark(c, &write_list);
2862		dm_bufio_unlock(c);
2863	}
2864	mutex_unlock(&dm_bufio_clients_lock);
2865
2866	__flush_write_list(&write_list);
2867}
2868
2869static void evict_old(void)
2870{
2871	unsigned long threshold = dm_bufio_cache_size -
2872		dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
2873
2874	mutex_lock(&dm_bufio_clients_lock);
2875	while (dm_bufio_current_allocated > threshold) {
2876		if (!__evict_a_few(64))
2877			break;
2878		cond_resched();
2879	}
2880	mutex_unlock(&dm_bufio_clients_lock);
2881}
2882
2883static void do_global_cleanup(struct work_struct *w)
2884{
2885	check_watermarks();
2886	evict_old();
2887}
2888
2889/*
2890 *--------------------------------------------------------------
2891 * Module setup
2892 *--------------------------------------------------------------
2893 */
2894
2895/*
2896 * This is called only once for the whole dm_bufio module.
2897 * It initializes memory limit.
2898 */
2899static int __init dm_bufio_init(void)
2900{
2901	__u64 mem;
2902
2903	dm_bufio_allocated_kmem_cache = 0;
2904	dm_bufio_allocated_get_free_pages = 0;
2905	dm_bufio_allocated_vmalloc = 0;
2906	dm_bufio_current_allocated = 0;
2907
2908	mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2909			       DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2910
2911	if (mem > ULONG_MAX)
2912		mem = ULONG_MAX;
2913
2914#ifdef CONFIG_MMU
2915	if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2916		mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2917#endif
2918
2919	dm_bufio_default_cache_size = mem;
2920
2921	mutex_lock(&dm_bufio_clients_lock);
2922	__cache_size_refresh();
2923	mutex_unlock(&dm_bufio_clients_lock);
2924
2925	dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2926	if (!dm_bufio_wq)
2927		return -ENOMEM;
2928
2929	INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
2930	INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
2931	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2932			   DM_BUFIO_WORK_TIMER_SECS * HZ);
2933
2934	return 0;
2935}
2936
2937/*
2938 * This is called once when unloading the dm_bufio module.
2939 */
2940static void __exit dm_bufio_exit(void)
2941{
2942	int bug = 0;
2943
2944	cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2945	destroy_workqueue(dm_bufio_wq);
2946
2947	if (dm_bufio_client_count) {
2948		DMCRIT("%s: dm_bufio_client_count leaked: %d",
2949			__func__, dm_bufio_client_count);
2950		bug = 1;
2951	}
2952
2953	if (dm_bufio_current_allocated) {
2954		DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2955			__func__, dm_bufio_current_allocated);
2956		bug = 1;
2957	}
2958
2959	if (dm_bufio_allocated_get_free_pages) {
2960		DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
2961		       __func__, dm_bufio_allocated_get_free_pages);
2962		bug = 1;
2963	}
2964
2965	if (dm_bufio_allocated_vmalloc) {
2966		DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2967		       __func__, dm_bufio_allocated_vmalloc);
2968		bug = 1;
2969	}
2970
2971	WARN_ON(bug); /* leaks are not worth crashing the system */
2972}
2973
2974module_init(dm_bufio_init)
2975module_exit(dm_bufio_exit)
2976
2977module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, 0644);
2978MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2979
2980module_param_named(max_age_seconds, dm_bufio_max_age, uint, 0644);
2981MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2982
2983module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, 0644);
2984MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2985
2986module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, 0644);
2987MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2988
2989module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, 0444);
2990MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2991
2992module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, 0444);
2993MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2994
2995module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, 0444);
2996MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2997
2998module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, 0444);
2999MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
3000
3001MODULE_AUTHOR("Mikulas Patocka <dm-devel@lists.linux.dev>");
3002MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
3003MODULE_LICENSE("GPL");
   1/*
   2 * Copyright (C) 2009-2011 Red Hat, Inc.
   3 *
   4 * Author: Mikulas Patocka <mpatocka@redhat.com>
   5 *
   6 * This file is released under the GPL.
   7 */
   8
   9#include <linux/dm-bufio.h>
  10
  11#include <linux/device-mapper.h>
  12#include <linux/dm-io.h>
  13#include <linux/slab.h>
  14#include <linux/sched/mm.h>
  15#include <linux/jiffies.h>
  16#include <linux/vmalloc.h>
  17#include <linux/shrinker.h>
  18#include <linux/module.h>
  19#include <linux/rbtree.h>
  20#include <linux/stacktrace.h>
  21#include <linux/jump_label.h>
  22
  23#define DM_MSG_PREFIX "bufio"
  24
  25/*
  26 * Memory management policy:
  27 *	Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
  28 *	or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
  29 *	Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
  30 *	Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
  31 *	dirty buffers.
  32 */
  33#define DM_BUFIO_MIN_BUFFERS		8
  34
  35#define DM_BUFIO_MEMORY_PERCENT		2
  36#define DM_BUFIO_VMALLOC_PERCENT	25
  37#define DM_BUFIO_WRITEBACK_RATIO	3
  38#define DM_BUFIO_LOW_WATERMARK_RATIO	16
  39
  40/*
  41 * Check buffer ages in this interval (seconds)
  42 */
  43#define DM_BUFIO_WORK_TIMER_SECS	30
  44
  45/*
  46 * Free buffers when they are older than this (seconds)
  47 */
  48#define DM_BUFIO_DEFAULT_AGE_SECS	300
  49
  50/*
  51 * The nr of bytes of cached data to keep around.
  52 */
  53#define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
  54
  55/*
  56 * Align buffer writes to this boundary.
  57 * Tests show that SSDs have the highest IOPS when using 4k writes.
  58 */
  59#define DM_BUFIO_WRITE_ALIGN		4096
  60
  61/*
  62 * dm_buffer->list_mode
  63 */
  64#define LIST_CLEAN	0
  65#define LIST_DIRTY	1
  66#define LIST_SIZE	2
  67
  68/*
  69 * Linking of buffers:
  70 *	All buffers are linked to buffer_tree with their node field.
  71 *
  72 *	Clean buffers that are not being written (B_WRITING not set)
  73 *	are linked to lru[LIST_CLEAN] with their lru_list field.
  74 *
  75 *	Dirty and clean buffers that are being written are linked to
  76 *	lru[LIST_DIRTY] with their lru_list field. When the write
  77 *	finishes, the buffer cannot be relinked immediately (because we
  78 *	are in an interrupt context and relinking requires process
  79 *	context), so some clean-not-writing buffers can be held on
  80 *	dirty_lru too.  They are later added to lru in the process
  81 *	context.
  82 */
  83struct dm_bufio_client {
  84	struct mutex lock;
  85	spinlock_t spinlock;
  86	bool no_sleep;
  87
  88	struct list_head lru[LIST_SIZE];
  89	unsigned long n_buffers[LIST_SIZE];
  90
  91	struct block_device *bdev;
  92	unsigned block_size;
  93	s8 sectors_per_block_bits;
  94	void (*alloc_callback)(struct dm_buffer *);
  95	void (*write_callback)(struct dm_buffer *);
  96	struct kmem_cache *slab_buffer;
  97	struct kmem_cache *slab_cache;
  98	struct dm_io_client *dm_io;
  99
 100	struct list_head reserved_buffers;
 101	unsigned need_reserved_buffers;
 102
 103	unsigned minimum_buffers;
 104
 105	struct rb_root buffer_tree;
 106	wait_queue_head_t free_buffer_wait;
 107
 108	sector_t start;
 109
 110	int async_write_error;
 111
 112	struct list_head client_list;
 113
 114	struct shrinker shrinker;
 115	struct work_struct shrink_work;
 116	atomic_long_t need_shrink;
 117};
 118
 119/*
 120 * Buffer state bits.
 121 */
 122#define B_READING	0
 123#define B_WRITING	1
 124#define B_DIRTY		2
 125
 126/*
 127 * Describes how the block was allocated:
 128 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
 129 * See the comment at alloc_buffer_data.
 130 */
 131enum data_mode {
 132	DATA_MODE_SLAB = 0,
 133	DATA_MODE_GET_FREE_PAGES = 1,
 134	DATA_MODE_VMALLOC = 2,
 135	DATA_MODE_LIMIT = 3
 136};
 137
 138struct dm_buffer {
 139	struct rb_node node;
 140	struct list_head lru_list;
 141	struct list_head global_list;
 142	sector_t block;
 143	void *data;
 144	unsigned char data_mode;		/* DATA_MODE_* */
 145	unsigned char list_mode;		/* LIST_* */
 146	blk_status_t read_error;
 147	blk_status_t write_error;
 148	unsigned accessed;
 149	unsigned hold_count;
 150	unsigned long state;
 151	unsigned long last_accessed;
 152	unsigned dirty_start;
 153	unsigned dirty_end;
 154	unsigned write_start;
 155	unsigned write_end;
 156	struct dm_bufio_client *c;
 157	struct list_head write_list;
 158	void (*end_io)(struct dm_buffer *, blk_status_t);
 159#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 160#define MAX_STACK 10
 161	unsigned int stack_len;
 162	unsigned long stack_entries[MAX_STACK];
 163#endif
 164};
 165
 166static DEFINE_STATIC_KEY_FALSE(no_sleep_enabled);
 167
 168/*----------------------------------------------------------------*/
 169
 170#define dm_bufio_in_request()	(!!current->bio_list)
 171
 172static void dm_bufio_lock(struct dm_bufio_client *c)
 173{
 174	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
 175		spin_lock_bh(&c->spinlock);
 176	else
 177		mutex_lock_nested(&c->lock, dm_bufio_in_request());
 178}
 179
 180static int dm_bufio_trylock(struct dm_bufio_client *c)
 181{
 182	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
 183		return spin_trylock_bh(&c->spinlock);
 184	else
 185		return mutex_trylock(&c->lock);
 186}
 187
 188static void dm_bufio_unlock(struct dm_bufio_client *c)
 189{
 190	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
 191		spin_unlock_bh(&c->spinlock);
 192	else
 193		mutex_unlock(&c->lock);
 194}
 195
 196/*----------------------------------------------------------------*/
 197
 198/*
 199 * Default cache size: available memory divided by the ratio.
 200 */
 201static unsigned long dm_bufio_default_cache_size;
 202
 203/*
 204 * Total cache size set by the user.
 205 */
 206static unsigned long dm_bufio_cache_size;
 207
 208/*
 209 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
 210 * at any time.  If it disagrees, the user has changed cache size.
 211 */
 212static unsigned long dm_bufio_cache_size_latch;
 213
 214static DEFINE_SPINLOCK(global_spinlock);
 215
 216static LIST_HEAD(global_queue);
 217
 218static unsigned long global_num = 0;
 219
 220/*
 221 * Buffers are freed after this timeout
 222 */
 223static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
 224static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
 225
 226static unsigned long dm_bufio_peak_allocated;
 227static unsigned long dm_bufio_allocated_kmem_cache;
 228static unsigned long dm_bufio_allocated_get_free_pages;
 229static unsigned long dm_bufio_allocated_vmalloc;
 230static unsigned long dm_bufio_current_allocated;
 231
 232/*----------------------------------------------------------------*/
 233
 234/*
 235 * The current number of clients.
 236 */
 237static int dm_bufio_client_count;
 238
 239/*
 240 * The list of all clients.
 241 */
 242static LIST_HEAD(dm_bufio_all_clients);
 243
 244/*
 245 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
 246 */
 247static DEFINE_MUTEX(dm_bufio_clients_lock);
 248
 249static struct workqueue_struct *dm_bufio_wq;
 250static struct delayed_work dm_bufio_cleanup_old_work;
 251static struct work_struct dm_bufio_replacement_work;
 252
 253
 254#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 255static void buffer_record_stack(struct dm_buffer *b)
 256{
 257	b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
 258}
 259#endif
 260
 261/*----------------------------------------------------------------
 262 * A red/black tree acts as an index for all the buffers.
 263 *--------------------------------------------------------------*/
 264static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
 265{
 266	struct rb_node *n = c->buffer_tree.rb_node;
 267	struct dm_buffer *b;
 268
 269	while (n) {
 270		b = container_of(n, struct dm_buffer, node);
 271
 272		if (b->block == block)
 273			return b;
 274
 275		n = block < b->block ? n->rb_left : n->rb_right;
 276	}
 277
 278	return NULL;
 279}
 280
 281static struct dm_buffer *__find_next(struct dm_bufio_client *c, sector_t block)
 282{
 283	struct rb_node *n = c->buffer_tree.rb_node;
 284	struct dm_buffer *b;
 285	struct dm_buffer *best = NULL;
 286
 287	while (n) {
 288		b = container_of(n, struct dm_buffer, node);
 289
 290		if (b->block == block)
 291			return b;
 292
 293		if (block <= b->block) {
 294			n = n->rb_left;
 295			best = b;
 296		} else {
 297			n = n->rb_right;
 298		}
 299	}
 300
 301	return best;
 302}
 303
 304static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
 305{
 306	struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
 307	struct dm_buffer *found;
 308
 309	while (*new) {
 310		found = container_of(*new, struct dm_buffer, node);
 311
 312		if (found->block == b->block) {
 313			BUG_ON(found != b);
 314			return;
 315		}
 316
 317		parent = *new;
 318		new = b->block < found->block ?
 319			&found->node.rb_left : &found->node.rb_right;
 320	}
 321
 322	rb_link_node(&b->node, parent, new);
 323	rb_insert_color(&b->node, &c->buffer_tree);
 324}
 325
 326static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
 327{
 328	rb_erase(&b->node, &c->buffer_tree);
 329}
 330
 331/*----------------------------------------------------------------*/
 332
 333static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
 334{
 335	unsigned char data_mode;
 336	long diff;
 337
 338	static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
 339		&dm_bufio_allocated_kmem_cache,
 340		&dm_bufio_allocated_get_free_pages,
 341		&dm_bufio_allocated_vmalloc,
 342	};
 343
 344	data_mode = b->data_mode;
 345	diff = (long)b->c->block_size;
 346	if (unlink)
 347		diff = -diff;
 348
 349	spin_lock(&global_spinlock);
 350
 351	*class_ptr[data_mode] += diff;
 352
 353	dm_bufio_current_allocated += diff;
 354
 355	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
 356		dm_bufio_peak_allocated = dm_bufio_current_allocated;
 357
 358	b->accessed = 1;
 359
 360	if (!unlink) {
 361		list_add(&b->global_list, &global_queue);
 362		global_num++;
 363		if (dm_bufio_current_allocated > dm_bufio_cache_size)
 364			queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
 365	} else {
 366		list_del(&b->global_list);
 367		global_num--;
 368	}
 369
 370	spin_unlock(&global_spinlock);
 371}
 372
 373/*
 374 * Change the number of clients and recalculate per-client limit.
 375 */
 376static void __cache_size_refresh(void)
 377{
 378	BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
 379	BUG_ON(dm_bufio_client_count < 0);
 380
 381	dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
 382
 383	/*
 384	 * Use default if set to 0 and report the actual cache size used.
 385	 */
 386	if (!dm_bufio_cache_size_latch) {
 387		(void)cmpxchg(&dm_bufio_cache_size, 0,
 388			      dm_bufio_default_cache_size);
 389		dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
 390	}
 391}
 392
 393/*
 394 * Allocating buffer data.
 395 *
 396 * Small buffers are allocated with kmem_cache, to use space optimally.
 397 *
 398 * For large buffers, we choose between get_free_pages and vmalloc.
 399 * Each has advantages and disadvantages.
 400 *
 401 * __get_free_pages can randomly fail if the memory is fragmented.
 402 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
 403 * as low as 128M) so using it for caching is not appropriate.
 404 *
 405 * If the allocation may fail we use __get_free_pages. Memory fragmentation
 406 * won't have a fatal effect here, but it just causes flushes of some other
 407 * buffers and more I/O will be performed. Don't use __get_free_pages if it
 408 * always fails (i.e. order >= MAX_ORDER).
 409 *
 410 * If the allocation shouldn't fail we use __vmalloc. This is only for the
 411 * initial reserve allocation, so there's no risk of wasting all vmalloc
 412 * space.
 413 */
 414static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
 415			       unsigned char *data_mode)
 416{
 417	if (unlikely(c->slab_cache != NULL)) {
 418		*data_mode = DATA_MODE_SLAB;
 419		return kmem_cache_alloc(c->slab_cache, gfp_mask);
 420	}
 421
 422	if (c->block_size <= KMALLOC_MAX_SIZE &&
 423	    gfp_mask & __GFP_NORETRY) {
 424		*data_mode = DATA_MODE_GET_FREE_PAGES;
 425		return (void *)__get_free_pages(gfp_mask,
 426						c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
 427	}
 428
 429	*data_mode = DATA_MODE_VMALLOC;
 430
 431	/*
 432	 * __vmalloc allocates the data pages and auxiliary structures with
 433	 * gfp_flags that were specified, but pagetables are always allocated
 434	 * with GFP_KERNEL, no matter what was specified as gfp_mask.
 435	 *
 436	 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
 437	 * all allocations done by this process (including pagetables) are done
 438	 * as if GFP_NOIO was specified.
 439	 */
 440	if (gfp_mask & __GFP_NORETRY) {
 441		unsigned noio_flag = memalloc_noio_save();
 442		void *ptr = __vmalloc(c->block_size, gfp_mask);
 443
 444		memalloc_noio_restore(noio_flag);
 445		return ptr;
 446	}
 447
 448	return __vmalloc(c->block_size, gfp_mask);
 449}
 450
 451/*
 452 * Free buffer's data.
 453 */
 454static void free_buffer_data(struct dm_bufio_client *c,
 455			     void *data, unsigned char data_mode)
 456{
 457	switch (data_mode) {
 458	case DATA_MODE_SLAB:
 459		kmem_cache_free(c->slab_cache, data);
 460		break;
 461
 462	case DATA_MODE_GET_FREE_PAGES:
 463		free_pages((unsigned long)data,
 464			   c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
 465		break;
 466
 467	case DATA_MODE_VMALLOC:
 468		vfree(data);
 469		break;
 470
 471	default:
 472		DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
 473		       data_mode);
 474		BUG();
 475	}
 476}
 477
 478/*
 479 * Allocate buffer and its data.
 480 */
 481static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
 482{
 483	struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
 484
 485	if (!b)
 486		return NULL;
 487
 488	b->c = c;
 489
 490	b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
 491	if (!b->data) {
 492		kmem_cache_free(c->slab_buffer, b);
 493		return NULL;
 494	}
 495
 496#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 497	b->stack_len = 0;
 498#endif
 499	return b;
 500}
 501
 502/*
 503 * Free buffer and its data.
 504 */
 505static void free_buffer(struct dm_buffer *b)
 506{
 507	struct dm_bufio_client *c = b->c;
 508
 509	free_buffer_data(c, b->data, b->data_mode);
 510	kmem_cache_free(c->slab_buffer, b);
 511}
 512
 513/*
 514 * Link buffer to the buffer tree and clean or dirty queue.
 515 */
 516static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
 517{
 518	struct dm_bufio_client *c = b->c;
 519
 520	c->n_buffers[dirty]++;
 521	b->block = block;
 522	b->list_mode = dirty;
 523	list_add(&b->lru_list, &c->lru[dirty]);
 524	__insert(b->c, b);
 525	b->last_accessed = jiffies;
 526
 527	adjust_total_allocated(b, false);
 528}
 529
 530/*
 531 * Unlink buffer from the buffer tree and dirty or clean queue.
 532 */
 533static void __unlink_buffer(struct dm_buffer *b)
 534{
 535	struct dm_bufio_client *c = b->c;
 536
 537	BUG_ON(!c->n_buffers[b->list_mode]);
 538
 539	c->n_buffers[b->list_mode]--;
 540	__remove(b->c, b);
 541	list_del(&b->lru_list);
 542
 543	adjust_total_allocated(b, true);
 544}
 545
 546/*
 547 * Place the buffer to the head of dirty or clean LRU queue.
 548 */
 549static void __relink_lru(struct dm_buffer *b, int dirty)
 550{
 551	struct dm_bufio_client *c = b->c;
 552
 553	b->accessed = 1;
 554
 555	BUG_ON(!c->n_buffers[b->list_mode]);
 556
 557	c->n_buffers[b->list_mode]--;
 558	c->n_buffers[dirty]++;
 559	b->list_mode = dirty;
 560	list_move(&b->lru_list, &c->lru[dirty]);
 561	b->last_accessed = jiffies;
 562}
 563
 564/*----------------------------------------------------------------
 565 * Submit I/O on the buffer.
 566 *
 567 * Bio interface is faster but it has some problems:
 568 *	the vector list is limited (increasing this limit increases
 569 *	memory-consumption per buffer, so it is not viable);
 570 *
 571 *	the memory must be direct-mapped, not vmalloced;
 572 *
 573 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
 574 * it is not vmalloced, try using the bio interface.
 575 *
 576 * If the buffer is big, if it is vmalloced or if the underlying device
 577 * rejects the bio because it is too large, use dm-io layer to do the I/O.
 578 * The dm-io layer splits the I/O into multiple requests, avoiding the above
 579 * shortcomings.
 580 *--------------------------------------------------------------*/
 581
 582/*
 583 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
 584 * that the request was handled directly with bio interface.
 585 */
 586static void dmio_complete(unsigned long error, void *context)
 587{
 588	struct dm_buffer *b = context;
 589
 590	b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
 591}
 592
 593static void use_dmio(struct dm_buffer *b, enum req_op op, sector_t sector,
 594		     unsigned n_sectors, unsigned offset)
 595{
 596	int r;
 597	struct dm_io_request io_req = {
 598		.bi_opf = op,
 599		.notify.fn = dmio_complete,
 600		.notify.context = b,
 601		.client = b->c->dm_io,
 602	};
 603	struct dm_io_region region = {
 604		.bdev = b->c->bdev,
 605		.sector = sector,
 606		.count = n_sectors,
 607	};
 608
 609	if (b->data_mode != DATA_MODE_VMALLOC) {
 610		io_req.mem.type = DM_IO_KMEM;
 611		io_req.mem.ptr.addr = (char *)b->data + offset;
 612	} else {
 613		io_req.mem.type = DM_IO_VMA;
 614		io_req.mem.ptr.vma = (char *)b->data + offset;
 615	}
 616
 617	r = dm_io(&io_req, 1, &region, NULL);
 618	if (unlikely(r))
 619		b->end_io(b, errno_to_blk_status(r));
 620}
 621
 622static void bio_complete(struct bio *bio)
 623{
 624	struct dm_buffer *b = bio->bi_private;
 625	blk_status_t status = bio->bi_status;
 626	bio_uninit(bio);
 627	kfree(bio);
 628	b->end_io(b, status);
 629}
 630
 631static void use_bio(struct dm_buffer *b, enum req_op op, sector_t sector,
 632		    unsigned n_sectors, unsigned offset)
 633{
 634	struct bio *bio;
 635	char *ptr;
 636	unsigned vec_size, len;
 637
 638	vec_size = b->c->block_size >> PAGE_SHIFT;
 639	if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
 640		vec_size += 2;
 641
 642	bio = bio_kmalloc(vec_size, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN);
 643	if (!bio) {
 644dmio:
 645		use_dmio(b, op, sector, n_sectors, offset);
 646		return;
 647	}
 648	bio_init(bio, b->c->bdev, bio->bi_inline_vecs, vec_size, op);
 649	bio->bi_iter.bi_sector = sector;
 650	bio->bi_end_io = bio_complete;
 651	bio->bi_private = b;
 652
 653	ptr = (char *)b->data + offset;
 654	len = n_sectors << SECTOR_SHIFT;
 655
 656	do {
 657		unsigned this_step = min((unsigned)(PAGE_SIZE - offset_in_page(ptr)), len);
 658		if (!bio_add_page(bio, virt_to_page(ptr), this_step,
 659				  offset_in_page(ptr))) {
 660			bio_put(bio);
 661			goto dmio;
 662		}
 663
 664		len -= this_step;
 665		ptr += this_step;
 666	} while (len > 0);
 667
 668	submit_bio(bio);
 669}
 670
 671static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
 672{
 673	sector_t sector;
 674
 675	if (likely(c->sectors_per_block_bits >= 0))
 676		sector = block << c->sectors_per_block_bits;
 677	else
 678		sector = block * (c->block_size >> SECTOR_SHIFT);
 679	sector += c->start;
 680
 681	return sector;
 682}
 683
 684static void submit_io(struct dm_buffer *b, enum req_op op,
 685		      void (*end_io)(struct dm_buffer *, blk_status_t))
 686{
 687	unsigned n_sectors;
 688	sector_t sector;
 689	unsigned offset, end;
 690
 691	b->end_io = end_io;
 692
 693	sector = block_to_sector(b->c, b->block);
 694
 695	if (op != REQ_OP_WRITE) {
 696		n_sectors = b->c->block_size >> SECTOR_SHIFT;
 697		offset = 0;
 698	} else {
 699		if (b->c->write_callback)
 700			b->c->write_callback(b);
 701		offset = b->write_start;
 702		end = b->write_end;
 703		offset &= -DM_BUFIO_WRITE_ALIGN;
 704		end += DM_BUFIO_WRITE_ALIGN - 1;
 705		end &= -DM_BUFIO_WRITE_ALIGN;
 706		if (unlikely(end > b->c->block_size))
 707			end = b->c->block_size;
 708
 709		sector += offset >> SECTOR_SHIFT;
 710		n_sectors = (end - offset) >> SECTOR_SHIFT;
 711	}
 712
 713	if (b->data_mode != DATA_MODE_VMALLOC)
 714		use_bio(b, op, sector, n_sectors, offset);
 715	else
 716		use_dmio(b, op, sector, n_sectors, offset);
 717}
 718
 719/*----------------------------------------------------------------
 720 * Writing dirty buffers
 721 *--------------------------------------------------------------*/
 722
 723/*
 724 * The endio routine for write.
 725 *
 726 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
 727 * it.
 728 */
 729static void write_endio(struct dm_buffer *b, blk_status_t status)
 730{
 731	b->write_error = status;
 732	if (unlikely(status)) {
 733		struct dm_bufio_client *c = b->c;
 734
 735		(void)cmpxchg(&c->async_write_error, 0,
 736				blk_status_to_errno(status));
 737	}
 738
 739	BUG_ON(!test_bit(B_WRITING, &b->state));
 740
 741	smp_mb__before_atomic();
 742	clear_bit(B_WRITING, &b->state);
 743	smp_mb__after_atomic();
 744
 745	wake_up_bit(&b->state, B_WRITING);
 746}
 747
 748/*
 749 * Initiate a write on a dirty buffer, but don't wait for it.
 750 *
 751 * - If the buffer is not dirty, exit.
 752 * - If there some previous write going on, wait for it to finish (we can't
 753 *   have two writes on the same buffer simultaneously).
 754 * - Submit our write and don't wait on it. We set B_WRITING indicating
 755 *   that there is a write in progress.
 756 */
 757static void __write_dirty_buffer(struct dm_buffer *b,
 758				 struct list_head *write_list)
 759{
 760	if (!test_bit(B_DIRTY, &b->state))
 761		return;
 762
 763	clear_bit(B_DIRTY, &b->state);
 764	wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 765
 766	b->write_start = b->dirty_start;
 767	b->write_end = b->dirty_end;
 768
 769	if (!write_list)
 770		submit_io(b, REQ_OP_WRITE, write_endio);
 771	else
 772		list_add_tail(&b->write_list, write_list);
 773}
 774
 775static void __flush_write_list(struct list_head *write_list)
 776{
 777	struct blk_plug plug;
 778	blk_start_plug(&plug);
 779	while (!list_empty(write_list)) {
 780		struct dm_buffer *b =
 781			list_entry(write_list->next, struct dm_buffer, write_list);
 782		list_del(&b->write_list);
 783		submit_io(b, REQ_OP_WRITE, write_endio);
 784		cond_resched();
 785	}
 786	blk_finish_plug(&plug);
 787}
 788
 789/*
 790 * Wait until any activity on the buffer finishes.  Possibly write the
 791 * buffer if it is dirty.  When this function finishes, there is no I/O
 792 * running on the buffer and the buffer is not dirty.
 793 */
 794static void __make_buffer_clean(struct dm_buffer *b)
 795{
 796	BUG_ON(b->hold_count);
 797
 798	/* smp_load_acquire() pairs with read_endio()'s smp_mb__before_atomic() */
 799	if (!smp_load_acquire(&b->state))	/* fast case */
 800		return;
 801
 802	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
 803	__write_dirty_buffer(b, NULL);
 804	wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 805}
 806
 807/*
 808 * Find some buffer that is not held by anybody, clean it, unlink it and
 809 * return it.
 810 */
 811static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
 812{
 813	struct dm_buffer *b;
 814
 815	list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
 816		BUG_ON(test_bit(B_WRITING, &b->state));
 817		BUG_ON(test_bit(B_DIRTY, &b->state));
 818
 819		if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep &&
 820		    unlikely(test_bit_acquire(B_READING, &b->state)))
 821			continue;
 822
 823		if (!b->hold_count) {
 824			__make_buffer_clean(b);
 825			__unlink_buffer(b);
 826			return b;
 827		}
 828		cond_resched();
 829	}
 830
 831	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
 832		return NULL;
 833
 834	list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
 835		BUG_ON(test_bit(B_READING, &b->state));
 836
 837		if (!b->hold_count) {
 838			__make_buffer_clean(b);
 839			__unlink_buffer(b);
 840			return b;
 841		}
 842		cond_resched();
 843	}
 844
 845	return NULL;
 846}
 847
 848/*
 849 * Wait until some other threads free some buffer or release hold count on
 850 * some buffer.
 851 *
 852 * This function is entered with c->lock held, drops it and regains it
 853 * before exiting.
 854 */
 855static void __wait_for_free_buffer(struct dm_bufio_client *c)
 856{
 857	DECLARE_WAITQUEUE(wait, current);
 858
 859	add_wait_queue(&c->free_buffer_wait, &wait);
 860	set_current_state(TASK_UNINTERRUPTIBLE);
 861	dm_bufio_unlock(c);
 862
 863	io_schedule();
 864
 865	remove_wait_queue(&c->free_buffer_wait, &wait);
 866
 867	dm_bufio_lock(c);
 868}
 869
 870enum new_flag {
 871	NF_FRESH = 0,
 872	NF_READ = 1,
 873	NF_GET = 2,
 874	NF_PREFETCH = 3
 875};
 876
 877/*
 878 * Allocate a new buffer. If the allocation is not possible, wait until
 879 * some other thread frees a buffer.
 880 *
 881 * May drop the lock and regain it.
 882 */
 883static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
 884{
 885	struct dm_buffer *b;
 886	bool tried_noio_alloc = false;
 887
 888	/*
 889	 * dm-bufio is resistant to allocation failures (it just keeps
 890	 * one buffer reserved in cases all the allocations fail).
 891	 * So set flags to not try too hard:
 892	 *	GFP_NOWAIT: don't wait; if we need to sleep we'll release our
 893	 *		    mutex and wait ourselves.
 894	 *	__GFP_NORETRY: don't retry and rather return failure
 895	 *	__GFP_NOMEMALLOC: don't use emergency reserves
 896	 *	__GFP_NOWARN: don't print a warning in case of failure
 897	 *
 898	 * For debugging, if we set the cache size to 1, no new buffers will
 899	 * be allocated.
 900	 */
 901	while (1) {
 902		if (dm_bufio_cache_size_latch != 1) {
 903			b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 904			if (b)
 905				return b;
 906		}
 907
 908		if (nf == NF_PREFETCH)
 909			return NULL;
 910
 911		if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
 912			dm_bufio_unlock(c);
 913			b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 914			dm_bufio_lock(c);
 915			if (b)
 916				return b;
 917			tried_noio_alloc = true;
 918		}
 919
 920		if (!list_empty(&c->reserved_buffers)) {
 921			b = list_entry(c->reserved_buffers.next,
 922				       struct dm_buffer, lru_list);
 923			list_del(&b->lru_list);
 924			c->need_reserved_buffers++;
 925
 926			return b;
 927		}
 928
 929		b = __get_unclaimed_buffer(c);
 930		if (b)
 931			return b;
 932
 933		__wait_for_free_buffer(c);
 934	}
 935}
 936
 937static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
 938{
 939	struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
 940
 941	if (!b)
 942		return NULL;
 943
 944	if (c->alloc_callback)
 945		c->alloc_callback(b);
 946
 947	return b;
 948}
 949
 950/*
 951 * Free a buffer and wake other threads waiting for free buffers.
 952 */
 953static void __free_buffer_wake(struct dm_buffer *b)
 954{
 955	struct dm_bufio_client *c = b->c;
 956
 957	if (!c->need_reserved_buffers)
 958		free_buffer(b);
 959	else {
 960		list_add(&b->lru_list, &c->reserved_buffers);
 961		c->need_reserved_buffers--;
 962	}
 963
 964	wake_up(&c->free_buffer_wait);
 965}
 966
 967static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
 968					struct list_head *write_list)
 969{
 970	struct dm_buffer *b, *tmp;
 971
 972	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
 973		BUG_ON(test_bit(B_READING, &b->state));
 974
 975		if (!test_bit(B_DIRTY, &b->state) &&
 976		    !test_bit(B_WRITING, &b->state)) {
 977			__relink_lru(b, LIST_CLEAN);
 978			continue;
 979		}
 980
 981		if (no_wait && test_bit(B_WRITING, &b->state))
 982			return;
 983
 984		__write_dirty_buffer(b, write_list);
 985		cond_resched();
 986	}
 987}
 988
 989/*
 990 * Check if we're over watermark.
 991 * If we are over threshold_buffers, start freeing buffers.
 992 * If we're over "limit_buffers", block until we get under the limit.
 993 */
 994static void __check_watermark(struct dm_bufio_client *c,
 995			      struct list_head *write_list)
 996{
 997	if (c->n_buffers[LIST_DIRTY] > c->n_buffers[LIST_CLEAN] * DM_BUFIO_WRITEBACK_RATIO)
 998		__write_dirty_buffers_async(c, 1, write_list);
 999}
1000
1001/*----------------------------------------------------------------
1002 * Getting a buffer
1003 *--------------------------------------------------------------*/
1004
1005static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
1006				     enum new_flag nf, int *need_submit,
1007				     struct list_head *write_list)
1008{
1009	struct dm_buffer *b, *new_b = NULL;
1010
1011	*need_submit = 0;
1012
1013	b = __find(c, block);
1014	if (b)
1015		goto found_buffer;
1016
1017	if (nf == NF_GET)
1018		return NULL;
1019
1020	new_b = __alloc_buffer_wait(c, nf);
1021	if (!new_b)
1022		return NULL;
1023
1024	/*
1025	 * We've had a period where the mutex was unlocked, so need to
1026	 * recheck the buffer tree.
1027	 */
1028	b = __find(c, block);
1029	if (b) {
1030		__free_buffer_wake(new_b);
1031		goto found_buffer;
1032	}
1033
1034	__check_watermark(c, write_list);
1035
1036	b = new_b;
1037	b->hold_count = 1;
1038	b->read_error = 0;
1039	b->write_error = 0;
1040	__link_buffer(b, block, LIST_CLEAN);
1041
1042	if (nf == NF_FRESH) {
1043		b->state = 0;
1044		return b;
1045	}
1046
1047	b->state = 1 << B_READING;
1048	*need_submit = 1;
1049
1050	return b;
1051
1052found_buffer:
1053	if (nf == NF_PREFETCH)
1054		return NULL;
1055	/*
1056	 * Note: it is essential that we don't wait for the buffer to be
1057	 * read if dm_bufio_get function is used. Both dm_bufio_get and
1058	 * dm_bufio_prefetch can be used in the driver request routine.
1059	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1060	 * the same buffer, it would deadlock if we waited.
1061	 */
1062	if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state)))
1063		return NULL;
1064
1065	b->hold_count++;
1066	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1067		     test_bit(B_WRITING, &b->state));
1068	return b;
1069}
1070
1071/*
1072 * The endio routine for reading: set the error, clear the bit and wake up
1073 * anyone waiting on the buffer.
1074 */
1075static void read_endio(struct dm_buffer *b, blk_status_t status)
1076{
1077	b->read_error = status;
1078
1079	BUG_ON(!test_bit(B_READING, &b->state));
1080
1081	smp_mb__before_atomic();
1082	clear_bit(B_READING, &b->state);
1083	smp_mb__after_atomic();
1084
1085	wake_up_bit(&b->state, B_READING);
1086}
1087
1088/*
1089 * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1090 * functions is similar except that dm_bufio_new doesn't read the
1091 * buffer from the disk (assuming that the caller overwrites all the data
1092 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1093 */
1094static void *new_read(struct dm_bufio_client *c, sector_t block,
1095		      enum new_flag nf, struct dm_buffer **bp)
1096{
1097	int need_submit;
1098	struct dm_buffer *b;
1099
1100	LIST_HEAD(write_list);
1101
1102	dm_bufio_lock(c);
1103	b = __bufio_new(c, block, nf, &need_submit, &write_list);
1104#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1105	if (b && b->hold_count == 1)
1106		buffer_record_stack(b);
1107#endif
1108	dm_bufio_unlock(c);
1109
1110	__flush_write_list(&write_list);
1111
1112	if (!b)
1113		return NULL;
1114
1115	if (need_submit)
1116		submit_io(b, REQ_OP_READ, read_endio);
1117
1118	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1119
1120	if (b->read_error) {
1121		int error = blk_status_to_errno(b->read_error);
1122
1123		dm_bufio_release(b);
1124
1125		return ERR_PTR(error);
1126	}
1127
1128	*bp = b;
1129
1130	return b->data;
1131}
1132
1133void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1134		   struct dm_buffer **bp)
1135{
1136	return new_read(c, block, NF_GET, bp);
1137}
1138EXPORT_SYMBOL_GPL(dm_bufio_get);
1139
1140void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1141		    struct dm_buffer **bp)
1142{
1143	BUG_ON(dm_bufio_in_request());
1144
1145	return new_read(c, block, NF_READ, bp);
1146}
1147EXPORT_SYMBOL_GPL(dm_bufio_read);
1148
1149void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1150		   struct dm_buffer **bp)
1151{
1152	BUG_ON(dm_bufio_in_request());
1153
1154	return new_read(c, block, NF_FRESH, bp);
1155}
1156EXPORT_SYMBOL_GPL(dm_bufio_new);
1157
1158void dm_bufio_prefetch(struct dm_bufio_client *c,
1159		       sector_t block, unsigned n_blocks)
1160{
1161	struct blk_plug plug;
1162
1163	LIST_HEAD(write_list);
1164
1165	BUG_ON(dm_bufio_in_request());
1166
1167	blk_start_plug(&plug);
1168	dm_bufio_lock(c);
1169
1170	for (; n_blocks--; block++) {
1171		int need_submit;
1172		struct dm_buffer *b;
1173		b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1174				&write_list);
1175		if (unlikely(!list_empty(&write_list))) {
1176			dm_bufio_unlock(c);
1177			blk_finish_plug(&plug);
1178			__flush_write_list(&write_list);
1179			blk_start_plug(&plug);
1180			dm_bufio_lock(c);
1181		}
1182		if (unlikely(b != NULL)) {
1183			dm_bufio_unlock(c);
1184
1185			if (need_submit)
1186				submit_io(b, REQ_OP_READ, read_endio);
1187			dm_bufio_release(b);
1188
1189			cond_resched();
1190
1191			if (!n_blocks)
1192				goto flush_plug;
1193			dm_bufio_lock(c);
1194		}
1195	}
1196
1197	dm_bufio_unlock(c);
1198
1199flush_plug:
1200	blk_finish_plug(&plug);
1201}
1202EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1203
1204void dm_bufio_release(struct dm_buffer *b)
1205{
1206	struct dm_bufio_client *c = b->c;
1207
1208	dm_bufio_lock(c);
1209
1210	BUG_ON(!b->hold_count);
1211
1212	b->hold_count--;
1213	if (!b->hold_count) {
1214		wake_up(&c->free_buffer_wait);
1215
1216		/*
1217		 * If there were errors on the buffer, and the buffer is not
1218		 * to be written, free the buffer. There is no point in caching
1219		 * invalid buffer.
1220		 */
1221		if ((b->read_error || b->write_error) &&
1222		    !test_bit_acquire(B_READING, &b->state) &&
1223		    !test_bit(B_WRITING, &b->state) &&
1224		    !test_bit(B_DIRTY, &b->state)) {
1225			__unlink_buffer(b);
1226			__free_buffer_wake(b);
1227		}
1228	}
1229
1230	dm_bufio_unlock(c);
1231}
1232EXPORT_SYMBOL_GPL(dm_bufio_release);
1233
1234void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1235					unsigned start, unsigned end)
1236{
1237	struct dm_bufio_client *c = b->c;
1238
1239	BUG_ON(start >= end);
1240	BUG_ON(end > b->c->block_size);
1241
1242	dm_bufio_lock(c);
1243
1244	BUG_ON(test_bit(B_READING, &b->state));
1245
1246	if (!test_and_set_bit(B_DIRTY, &b->state)) {
1247		b->dirty_start = start;
1248		b->dirty_end = end;
1249		__relink_lru(b, LIST_DIRTY);
1250	} else {
1251		if (start < b->dirty_start)
1252			b->dirty_start = start;
1253		if (end > b->dirty_end)
1254			b->dirty_end = end;
1255	}
1256
1257	dm_bufio_unlock(c);
1258}
1259EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1260
1261void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1262{
1263	dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1264}
1265EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1266
1267void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1268{
1269	LIST_HEAD(write_list);
1270
1271	BUG_ON(dm_bufio_in_request());
1272
1273	dm_bufio_lock(c);
1274	__write_dirty_buffers_async(c, 0, &write_list);
1275	dm_bufio_unlock(c);
1276	__flush_write_list(&write_list);
1277}
1278EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1279
1280/*
1281 * For performance, it is essential that the buffers are written asynchronously
1282 * and simultaneously (so that the block layer can merge the writes) and then
1283 * waited upon.
1284 *
1285 * Finally, we flush hardware disk cache.
1286 */
1287int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1288{
1289	int a, f;
1290	unsigned long buffers_processed = 0;
1291	struct dm_buffer *b, *tmp;
1292
1293	LIST_HEAD(write_list);
1294
1295	dm_bufio_lock(c);
1296	__write_dirty_buffers_async(c, 0, &write_list);
1297	dm_bufio_unlock(c);
1298	__flush_write_list(&write_list);
1299	dm_bufio_lock(c);
1300
1301again:
1302	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1303		int dropped_lock = 0;
1304
1305		if (buffers_processed < c->n_buffers[LIST_DIRTY])
1306			buffers_processed++;
1307
1308		BUG_ON(test_bit(B_READING, &b->state));
1309
1310		if (test_bit(B_WRITING, &b->state)) {
1311			if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1312				dropped_lock = 1;
1313				b->hold_count++;
1314				dm_bufio_unlock(c);
1315				wait_on_bit_io(&b->state, B_WRITING,
1316					       TASK_UNINTERRUPTIBLE);
1317				dm_bufio_lock(c);
1318				b->hold_count--;
1319			} else
1320				wait_on_bit_io(&b->state, B_WRITING,
1321					       TASK_UNINTERRUPTIBLE);
1322		}
1323
1324		if (!test_bit(B_DIRTY, &b->state) &&
1325		    !test_bit(B_WRITING, &b->state))
1326			__relink_lru(b, LIST_CLEAN);
1327
1328		cond_resched();
1329
1330		/*
1331		 * If we dropped the lock, the list is no longer consistent,
1332		 * so we must restart the search.
1333		 *
1334		 * In the most common case, the buffer just processed is
1335		 * relinked to the clean list, so we won't loop scanning the
1336		 * same buffer again and again.
1337		 *
1338		 * This may livelock if there is another thread simultaneously
1339		 * dirtying buffers, so we count the number of buffers walked
1340		 * and if it exceeds the total number of buffers, it means that
1341		 * someone is doing some writes simultaneously with us.  In
1342		 * this case, stop, dropping the lock.
1343		 */
1344		if (dropped_lock)
1345			goto again;
1346	}
1347	wake_up(&c->free_buffer_wait);
1348	dm_bufio_unlock(c);
1349
1350	a = xchg(&c->async_write_error, 0);
1351	f = dm_bufio_issue_flush(c);
1352	if (a)
1353		return a;
1354
1355	return f;
1356}
1357EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1358
1359/*
1360 * Use dm-io to send an empty barrier to flush the device.
1361 */
1362int dm_bufio_issue_flush(struct dm_bufio_client *c)
1363{
1364	struct dm_io_request io_req = {
1365		.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC,
1366		.mem.type = DM_IO_KMEM,
1367		.mem.ptr.addr = NULL,
1368		.client = c->dm_io,
1369	};
1370	struct dm_io_region io_reg = {
1371		.bdev = c->bdev,
1372		.sector = 0,
1373		.count = 0,
1374	};
1375
1376	BUG_ON(dm_bufio_in_request());
1377
1378	return dm_io(&io_req, 1, &io_reg, NULL);
1379}
1380EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1381
1382/*
1383 * Use dm-io to send a discard request to flush the device.
1384 */
1385int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
1386{
1387	struct dm_io_request io_req = {
1388		.bi_opf = REQ_OP_DISCARD | REQ_SYNC,
1389		.mem.type = DM_IO_KMEM,
1390		.mem.ptr.addr = NULL,
1391		.client = c->dm_io,
1392	};
1393	struct dm_io_region io_reg = {
1394		.bdev = c->bdev,
1395		.sector = block_to_sector(c, block),
1396		.count = block_to_sector(c, count),
1397	};
1398
1399	BUG_ON(dm_bufio_in_request());
1400
1401	return dm_io(&io_req, 1, &io_reg, NULL);
1402}
1403EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
1404
1405/*
1406 * We first delete any other buffer that may be at that new location.
1407 *
1408 * Then, we write the buffer to the original location if it was dirty.
1409 *
1410 * Then, if we are the only one who is holding the buffer, relink the buffer
1411 * in the buffer tree for the new location.
1412 *
1413 * If there was someone else holding the buffer, we write it to the new
1414 * location but not relink it, because that other user needs to have the buffer
1415 * at the same place.
1416 */
1417void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1418{
1419	struct dm_bufio_client *c = b->c;
1420	struct dm_buffer *new;
1421
1422	BUG_ON(dm_bufio_in_request());
1423
1424	dm_bufio_lock(c);
1425
1426retry:
1427	new = __find(c, new_block);
1428	if (new) {
1429		if (new->hold_count) {
1430			__wait_for_free_buffer(c);
1431			goto retry;
1432		}
1433
1434		/*
1435		 * FIXME: Is there any point waiting for a write that's going
1436		 * to be overwritten in a bit?
1437		 */
1438		__make_buffer_clean(new);
1439		__unlink_buffer(new);
1440		__free_buffer_wake(new);
1441	}
1442
1443	BUG_ON(!b->hold_count);
1444	BUG_ON(test_bit(B_READING, &b->state));
1445
1446	__write_dirty_buffer(b, NULL);
1447	if (b->hold_count == 1) {
1448		wait_on_bit_io(&b->state, B_WRITING,
1449			       TASK_UNINTERRUPTIBLE);
1450		set_bit(B_DIRTY, &b->state);
1451		b->dirty_start = 0;
1452		b->dirty_end = c->block_size;
1453		__unlink_buffer(b);
1454		__link_buffer(b, new_block, LIST_DIRTY);
1455	} else {
1456		sector_t old_block;
1457		wait_on_bit_lock_io(&b->state, B_WRITING,
1458				    TASK_UNINTERRUPTIBLE);
1459		/*
1460		 * Relink buffer to "new_block" so that write_callback
1461		 * sees "new_block" as a block number.
1462		 * After the write, link the buffer back to old_block.
1463		 * All this must be done in bufio lock, so that block number
1464		 * change isn't visible to other threads.
1465		 */
1466		old_block = b->block;
1467		__unlink_buffer(b);
1468		__link_buffer(b, new_block, b->list_mode);
1469		submit_io(b, REQ_OP_WRITE, write_endio);
1470		wait_on_bit_io(&b->state, B_WRITING,
1471			       TASK_UNINTERRUPTIBLE);
1472		__unlink_buffer(b);
1473		__link_buffer(b, old_block, b->list_mode);
1474	}
1475
1476	dm_bufio_unlock(c);
1477	dm_bufio_release(b);
1478}
1479EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1480
1481static void forget_buffer_locked(struct dm_buffer *b)
1482{
1483	if (likely(!b->hold_count) && likely(!smp_load_acquire(&b->state))) {
1484		__unlink_buffer(b);
1485		__free_buffer_wake(b);
1486	}
1487}
1488
1489/*
1490 * Free the given buffer.
1491 *
1492 * This is just a hint, if the buffer is in use or dirty, this function
1493 * does nothing.
1494 */
1495void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1496{
1497	struct dm_buffer *b;
1498
1499	dm_bufio_lock(c);
1500
1501	b = __find(c, block);
1502	if (b)
1503		forget_buffer_locked(b);
1504
1505	dm_bufio_unlock(c);
1506}
1507EXPORT_SYMBOL_GPL(dm_bufio_forget);
1508
1509void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
1510{
1511	struct dm_buffer *b;
1512	sector_t end_block = block + n_blocks;
1513
1514	while (block < end_block) {
1515		dm_bufio_lock(c);
1516
1517		b = __find_next(c, block);
1518		if (b) {
1519			block = b->block + 1;
1520			forget_buffer_locked(b);
1521		}
1522
1523		dm_bufio_unlock(c);
1524
1525		if (!b)
1526			break;
1527	}
1528
1529}
1530EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
1531
1532void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1533{
1534	c->minimum_buffers = n;
1535}
1536EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1537
1538unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1539{
1540	return c->block_size;
1541}
1542EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1543
1544sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1545{
1546	sector_t s = bdev_nr_sectors(c->bdev);
1547	if (s >= c->start)
1548		s -= c->start;
1549	else
1550		s = 0;
1551	if (likely(c->sectors_per_block_bits >= 0))
1552		s >>= c->sectors_per_block_bits;
1553	else
1554		sector_div(s, c->block_size >> SECTOR_SHIFT);
1555	return s;
1556}
1557EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1558
1559struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
1560{
1561	return c->dm_io;
1562}
1563EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
1564
1565sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1566{
1567	return b->block;
1568}
1569EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1570
1571void *dm_bufio_get_block_data(struct dm_buffer *b)
1572{
1573	return b->data;
1574}
1575EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1576
1577void *dm_bufio_get_aux_data(struct dm_buffer *b)
1578{
1579	return b + 1;
1580}
1581EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1582
1583struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1584{
1585	return b->c;
1586}
1587EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1588
1589static void drop_buffers(struct dm_bufio_client *c)
1590{
1591	struct dm_buffer *b;
1592	int i;
1593	bool warned = false;
1594
1595	BUG_ON(dm_bufio_in_request());
1596
1597	/*
1598	 * An optimization so that the buffers are not written one-by-one.
1599	 */
1600	dm_bufio_write_dirty_buffers_async(c);
1601
1602	dm_bufio_lock(c);
1603
1604	while ((b = __get_unclaimed_buffer(c)))
1605		__free_buffer_wake(b);
1606
1607	for (i = 0; i < LIST_SIZE; i++)
1608		list_for_each_entry(b, &c->lru[i], lru_list) {
1609			WARN_ON(!warned);
1610			warned = true;
1611			DMERR("leaked buffer %llx, hold count %u, list %d",
1612			      (unsigned long long)b->block, b->hold_count, i);
1613#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1614			stack_trace_print(b->stack_entries, b->stack_len, 1);
1615			/* mark unclaimed to avoid BUG_ON below */
1616			b->hold_count = 0;
1617#endif
1618		}
1619
1620#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1621	while ((b = __get_unclaimed_buffer(c)))
1622		__free_buffer_wake(b);
1623#endif
1624
1625	for (i = 0; i < LIST_SIZE; i++)
1626		BUG_ON(!list_empty(&c->lru[i]));
1627
1628	dm_bufio_unlock(c);
1629}
1630
1631/*
1632 * We may not be able to evict this buffer if IO pending or the client
1633 * is still using it.  Caller is expected to know buffer is too old.
1634 *
1635 * And if GFP_NOFS is used, we must not do any I/O because we hold
1636 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1637 * rerouted to different bufio client.
1638 */
1639static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1640{
1641	if (!(gfp & __GFP_FS) ||
1642	    (static_branch_unlikely(&no_sleep_enabled) && b->c->no_sleep)) {
1643		if (test_bit_acquire(B_READING, &b->state) ||
1644		    test_bit(B_WRITING, &b->state) ||
1645		    test_bit(B_DIRTY, &b->state))
1646			return false;
1647	}
1648
1649	if (b->hold_count)
1650		return false;
1651
1652	__make_buffer_clean(b);
1653	__unlink_buffer(b);
1654	__free_buffer_wake(b);
1655
1656	return true;
1657}
1658
1659static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1660{
1661	unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1662	if (likely(c->sectors_per_block_bits >= 0))
1663		retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1664	else
1665		retain_bytes /= c->block_size;
1666	return retain_bytes;
1667}
1668
1669static void __scan(struct dm_bufio_client *c)
1670{
1671	int l;
1672	struct dm_buffer *b, *tmp;
1673	unsigned long freed = 0;
1674	unsigned long count = c->n_buffers[LIST_CLEAN] +
1675			      c->n_buffers[LIST_DIRTY];
1676	unsigned long retain_target = get_retain_buffers(c);
1677
1678	for (l = 0; l < LIST_SIZE; l++) {
1679		list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1680			if (count - freed <= retain_target)
1681				atomic_long_set(&c->need_shrink, 0);
1682			if (!atomic_long_read(&c->need_shrink))
1683				return;
1684			if (__try_evict_buffer(b, GFP_KERNEL)) {
1685				atomic_long_dec(&c->need_shrink);
1686				freed++;
1687			}
1688			cond_resched();
1689		}
1690	}
1691}
1692
1693static void shrink_work(struct work_struct *w)
1694{
1695	struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
1696
1697	dm_bufio_lock(c);
1698	__scan(c);
1699	dm_bufio_unlock(c);
1700}
1701
1702static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1703{
1704	struct dm_bufio_client *c;
1705
1706	c = container_of(shrink, struct dm_bufio_client, shrinker);
1707	atomic_long_add(sc->nr_to_scan, &c->need_shrink);
1708	queue_work(dm_bufio_wq, &c->shrink_work);
1709
1710	return sc->nr_to_scan;
1711}
1712
1713static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1714{
1715	struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1716	unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1717			      READ_ONCE(c->n_buffers[LIST_DIRTY]);
1718	unsigned long retain_target = get_retain_buffers(c);
1719	unsigned long queued_for_cleanup = atomic_long_read(&c->need_shrink);
1720
1721	if (unlikely(count < retain_target))
1722		count = 0;
1723	else
1724		count -= retain_target;
1725
1726	if (unlikely(count < queued_for_cleanup))
1727		count = 0;
1728	else
1729		count -= queued_for_cleanup;
1730
1731	return count;
1732}
1733
1734/*
1735 * Create the buffering interface
1736 */
1737struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1738					       unsigned reserved_buffers, unsigned aux_size,
1739					       void (*alloc_callback)(struct dm_buffer *),
1740					       void (*write_callback)(struct dm_buffer *),
1741					       unsigned int flags)
1742{
1743	int r;
1744	struct dm_bufio_client *c;
1745	unsigned i;
1746	char slab_name[27];
1747
1748	if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1749		DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1750		r = -EINVAL;
1751		goto bad_client;
1752	}
1753
1754	c = kzalloc(sizeof(*c), GFP_KERNEL);
1755	if (!c) {
1756		r = -ENOMEM;
1757		goto bad_client;
1758	}
1759	c->buffer_tree = RB_ROOT;
1760
1761	c->bdev = bdev;
1762	c->block_size = block_size;
1763	if (is_power_of_2(block_size))
1764		c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1765	else
1766		c->sectors_per_block_bits = -1;
1767
1768	c->alloc_callback = alloc_callback;
1769	c->write_callback = write_callback;
1770
1771	if (flags & DM_BUFIO_CLIENT_NO_SLEEP) {
1772		c->no_sleep = true;
1773		static_branch_inc(&no_sleep_enabled);
1774	}
1775
1776	for (i = 0; i < LIST_SIZE; i++) {
1777		INIT_LIST_HEAD(&c->lru[i]);
1778		c->n_buffers[i] = 0;
1779	}
1780
1781	mutex_init(&c->lock);
1782	spin_lock_init(&c->spinlock);
1783	INIT_LIST_HEAD(&c->reserved_buffers);
1784	c->need_reserved_buffers = reserved_buffers;
1785
1786	dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1787
1788	init_waitqueue_head(&c->free_buffer_wait);
1789	c->async_write_error = 0;
1790
1791	c->dm_io = dm_io_client_create();
1792	if (IS_ERR(c->dm_io)) {
1793		r = PTR_ERR(c->dm_io);
1794		goto bad_dm_io;
1795	}
1796
1797	if (block_size <= KMALLOC_MAX_SIZE &&
1798	    (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1799		unsigned align = min(1U << __ffs(block_size), (unsigned)PAGE_SIZE);
1800		snprintf(slab_name, sizeof slab_name, "dm_bufio_cache-%u", block_size);
1801		c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1802						  SLAB_RECLAIM_ACCOUNT, NULL);
1803		if (!c->slab_cache) {
1804			r = -ENOMEM;
1805			goto bad;
1806		}
1807	}
1808	if (aux_size)
1809		snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1810	else
1811		snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer");
1812	c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1813					   0, SLAB_RECLAIM_ACCOUNT, NULL);
1814	if (!c->slab_buffer) {
1815		r = -ENOMEM;
1816		goto bad;
1817	}
1818
1819	while (c->need_reserved_buffers) {
1820		struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1821
1822		if (!b) {
1823			r = -ENOMEM;
1824			goto bad;
1825		}
1826		__free_buffer_wake(b);
1827	}
1828
1829	INIT_WORK(&c->shrink_work, shrink_work);
1830	atomic_long_set(&c->need_shrink, 0);
1831
1832	c->shrinker.count_objects = dm_bufio_shrink_count;
1833	c->shrinker.scan_objects = dm_bufio_shrink_scan;
1834	c->shrinker.seeks = 1;
1835	c->shrinker.batch = 0;
1836	r = register_shrinker(&c->shrinker, "md-%s:(%u:%u)", slab_name,
1837			      MAJOR(bdev->bd_dev), MINOR(bdev->bd_dev));
1838	if (r)
1839		goto bad;
1840
1841	mutex_lock(&dm_bufio_clients_lock);
1842	dm_bufio_client_count++;
1843	list_add(&c->client_list, &dm_bufio_all_clients);
1844	__cache_size_refresh();
1845	mutex_unlock(&dm_bufio_clients_lock);
1846
1847	return c;
1848
1849bad:
1850	while (!list_empty(&c->reserved_buffers)) {
1851		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1852						 struct dm_buffer, lru_list);
1853		list_del(&b->lru_list);
1854		free_buffer(b);
1855	}
1856	kmem_cache_destroy(c->slab_cache);
1857	kmem_cache_destroy(c->slab_buffer);
1858	dm_io_client_destroy(c->dm_io);
1859bad_dm_io:
1860	mutex_destroy(&c->lock);
1861	if (c->no_sleep)
1862		static_branch_dec(&no_sleep_enabled);
1863	kfree(c);
1864bad_client:
1865	return ERR_PTR(r);
1866}
1867EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1868
1869/*
1870 * Free the buffering interface.
1871 * It is required that there are no references on any buffers.
1872 */
1873void dm_bufio_client_destroy(struct dm_bufio_client *c)
1874{
1875	unsigned i;
1876
1877	drop_buffers(c);
1878
1879	unregister_shrinker(&c->shrinker);
1880	flush_work(&c->shrink_work);
1881
1882	mutex_lock(&dm_bufio_clients_lock);
1883
1884	list_del(&c->client_list);
1885	dm_bufio_client_count--;
1886	__cache_size_refresh();
1887
1888	mutex_unlock(&dm_bufio_clients_lock);
1889
1890	BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1891	BUG_ON(c->need_reserved_buffers);
1892
1893	while (!list_empty(&c->reserved_buffers)) {
1894		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1895						 struct dm_buffer, lru_list);
1896		list_del(&b->lru_list);
1897		free_buffer(b);
1898	}
1899
1900	for (i = 0; i < LIST_SIZE; i++)
1901		if (c->n_buffers[i])
1902			DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1903
1904	for (i = 0; i < LIST_SIZE; i++)
1905		BUG_ON(c->n_buffers[i]);
1906
1907	kmem_cache_destroy(c->slab_cache);
1908	kmem_cache_destroy(c->slab_buffer);
1909	dm_io_client_destroy(c->dm_io);
1910	mutex_destroy(&c->lock);
1911	if (c->no_sleep)
1912		static_branch_dec(&no_sleep_enabled);
1913	kfree(c);
1914}
1915EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1916
1917void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1918{
1919	c->start = start;
1920}
1921EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1922
1923static unsigned get_max_age_hz(void)
1924{
1925	unsigned max_age = READ_ONCE(dm_bufio_max_age);
1926
1927	if (max_age > UINT_MAX / HZ)
1928		max_age = UINT_MAX / HZ;
1929
1930	return max_age * HZ;
1931}
1932
1933static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1934{
1935	return time_after_eq(jiffies, b->last_accessed + age_hz);
1936}
1937
1938static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1939{
1940	struct dm_buffer *b, *tmp;
1941	unsigned long retain_target = get_retain_buffers(c);
1942	unsigned long count;
1943	LIST_HEAD(write_list);
1944
1945	dm_bufio_lock(c);
1946
1947	__check_watermark(c, &write_list);
1948	if (unlikely(!list_empty(&write_list))) {
1949		dm_bufio_unlock(c);
1950		__flush_write_list(&write_list);
1951		dm_bufio_lock(c);
1952	}
1953
1954	count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1955	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1956		if (count <= retain_target)
1957			break;
1958
1959		if (!older_than(b, age_hz))
1960			break;
1961
1962		if (__try_evict_buffer(b, 0))
1963			count--;
1964
1965		cond_resched();
1966	}
1967
1968	dm_bufio_unlock(c);
1969}
1970
1971static void do_global_cleanup(struct work_struct *w)
1972{
1973	struct dm_bufio_client *locked_client = NULL;
1974	struct dm_bufio_client *current_client;
1975	struct dm_buffer *b;
1976	unsigned spinlock_hold_count;
1977	unsigned long threshold = dm_bufio_cache_size -
1978		dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
1979	unsigned long loops = global_num * 2;
1980
1981	mutex_lock(&dm_bufio_clients_lock);
1982
1983	while (1) {
1984		cond_resched();
1985
1986		spin_lock(&global_spinlock);
1987		if (unlikely(dm_bufio_current_allocated <= threshold))
1988			break;
1989
1990		spinlock_hold_count = 0;
1991get_next:
1992		if (!loops--)
1993			break;
1994		if (unlikely(list_empty(&global_queue)))
1995			break;
1996		b = list_entry(global_queue.prev, struct dm_buffer, global_list);
1997
1998		if (b->accessed) {
1999			b->accessed = 0;
2000			list_move(&b->global_list, &global_queue);
2001			if (likely(++spinlock_hold_count < 16))
2002				goto get_next;
2003			spin_unlock(&global_spinlock);
2004			continue;
2005		}
2006
2007		current_client = b->c;
2008		if (unlikely(current_client != locked_client)) {
2009			if (locked_client)
2010				dm_bufio_unlock(locked_client);
2011
2012			if (!dm_bufio_trylock(current_client)) {
2013				spin_unlock(&global_spinlock);
2014				dm_bufio_lock(current_client);
2015				locked_client = current_client;
2016				continue;
2017			}
2018
2019			locked_client = current_client;
2020		}
2021
2022		spin_unlock(&global_spinlock);
2023
2024		if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
2025			spin_lock(&global_spinlock);
2026			list_move(&b->global_list, &global_queue);
2027			spin_unlock(&global_spinlock);
2028		}
2029	}
2030
2031	spin_unlock(&global_spinlock);
2032
2033	if (locked_client)
2034		dm_bufio_unlock(locked_client);
2035
2036	mutex_unlock(&dm_bufio_clients_lock);
2037}
2038
2039static void cleanup_old_buffers(void)
2040{
2041	unsigned long max_age_hz = get_max_age_hz();
2042	struct dm_bufio_client *c;
2043
2044	mutex_lock(&dm_bufio_clients_lock);
2045
2046	__cache_size_refresh();
2047
2048	list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2049		__evict_old_buffers(c, max_age_hz);
2050
2051	mutex_unlock(&dm_bufio_clients_lock);
2052}
2053
2054static void work_fn(struct work_struct *w)
2055{
2056	cleanup_old_buffers();
2057
2058	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2059			   DM_BUFIO_WORK_TIMER_SECS * HZ);
2060}
2061
2062/*----------------------------------------------------------------
2063 * Module setup
2064 *--------------------------------------------------------------*/
2065
2066/*
2067 * This is called only once for the whole dm_bufio module.
2068 * It initializes memory limit.
2069 */
2070static int __init dm_bufio_init(void)
2071{
2072	__u64 mem;
2073
2074	dm_bufio_allocated_kmem_cache = 0;
2075	dm_bufio_allocated_get_free_pages = 0;
2076	dm_bufio_allocated_vmalloc = 0;
2077	dm_bufio_current_allocated = 0;
2078
2079	mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2080			       DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2081
2082	if (mem > ULONG_MAX)
2083		mem = ULONG_MAX;
2084
2085#ifdef CONFIG_MMU
2086	if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2087		mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2088#endif
2089
2090	dm_bufio_default_cache_size = mem;
2091
2092	mutex_lock(&dm_bufio_clients_lock);
2093	__cache_size_refresh();
2094	mutex_unlock(&dm_bufio_clients_lock);
2095
2096	dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2097	if (!dm_bufio_wq)
2098		return -ENOMEM;
2099
2100	INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
2101	INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
2102	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2103			   DM_BUFIO_WORK_TIMER_SECS * HZ);
2104
2105	return 0;
2106}
2107
2108/*
2109 * This is called once when unloading the dm_bufio module.
2110 */
2111static void __exit dm_bufio_exit(void)
2112{
2113	int bug = 0;
2114
2115	cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2116	destroy_workqueue(dm_bufio_wq);
2117
2118	if (dm_bufio_client_count) {
2119		DMCRIT("%s: dm_bufio_client_count leaked: %d",
2120			__func__, dm_bufio_client_count);
2121		bug = 1;
2122	}
2123
2124	if (dm_bufio_current_allocated) {
2125		DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2126			__func__, dm_bufio_current_allocated);
2127		bug = 1;
2128	}
2129
2130	if (dm_bufio_allocated_get_free_pages) {
2131		DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
2132		       __func__, dm_bufio_allocated_get_free_pages);
2133		bug = 1;
2134	}
2135
2136	if (dm_bufio_allocated_vmalloc) {
2137		DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2138		       __func__, dm_bufio_allocated_vmalloc);
2139		bug = 1;
2140	}
2141
2142	BUG_ON(bug);
2143}
2144
2145module_init(dm_bufio_init)
2146module_exit(dm_bufio_exit)
2147
2148module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
2149MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2150
2151module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
2152MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2153
2154module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
2155MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2156
2157module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
2158MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2159
2160module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
2161MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2162
2163module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
2164MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2165
2166module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
2167MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2168
2169module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
2170MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2171
2172MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2173MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2174MODULE_LICENSE("GPL");