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");